model: fix poolside parser/renderer

The superbuild from the llama-server work changed paths but missed updating the MLX library resolution code to match.

.github/workflows/release.yaml

@@ -16,7 +16,7 @@ jobs:
     outputs:
       GOFLAGS: ${{ steps.goflags.outputs.GOFLAGS }}
       VERSION: ${{ steps.goflags.outputs.VERSION }}
-      vendorsha: ${{ steps.changes.outputs.vendorsha }}
+      vendorsha: ${{ steps.goflags.outputs.vendorsha }}
     steps:
       - uses: actions/checkout@v4
       - name: Set environment
@@ -24,7 +24,7 @@ jobs:
         run: |
           echo GOFLAGS="'-ldflags=-w -s \"-X=github.com/ollama/ollama/version.Version=${GITHUB_REF_NAME#v}\" \"-X=github.com/ollama/ollama/server.mode=release\"'" | tee -a $GITHUB_OUTPUT
           echo VERSION="${GITHUB_REF_NAME#v}" | tee -a $GITHUB_OUTPUT
-          echo vendorsha=$(make -f Makefile.sync print-base) | tee -a $GITHUB_OUTPUT
+          echo vendorsha=$(cat LLAMA_CPP_VERSION)-$(cat MLX_VERSION)-$(cat MLX_C_VERSION) | tee -a $GITHUB_OUTPUT
 
   darwin-build:
     runs-on: macos-26-xlarge
@@ -57,7 +57,9 @@ jobs:
           go-version-file: go.mod
           cache-dependency-path: |
             go.sum
-            Makefile.sync
+            LLAMA_CPP_VERSION
+            MLX_VERSION
+            MLX_C_VERSION
       - run: |
           ./scripts/build_darwin.sh
       - name: Log build results
@@ -73,15 +75,18 @@ jobs:
             dist/*.dmg
 
   windows-depends:
+    needs: setup-environment
     strategy:
       matrix:
         os: [windows]
         arch: [amd64]
         preset: ['CPU']
+        build-steps: ['cpu cpuArm64']
         include:
           - os: windows
             arch: amd64
             preset: 'CUDA 12'
+            build-steps: cuda12
             install: https://developer.download.nvidia.com/compute/cuda/12.8.0/local_installers/cuda_12.8.0_571.96_windows.exe
             cuda-components:
               - '"cudart"'
@@ -89,10 +94,10 @@ jobs:
               - '"cublas"'
               - '"cublas_dev"'
             cuda-version: '12.8'
-            flags: ''
           - os: windows
             arch: amd64
             preset: 'CUDA 13'
+            build-steps: cuda13
             install: https://developer.download.nvidia.com/compute/cuda/13.0.0/local_installers/cuda_13.0.0_windows.exe
             cuda-components:
               - '"cudart"'
@@ -103,23 +108,23 @@ jobs:
               - '"nvvm"'
               - '"nvptxcompiler"'
             cuda-version: '13.0'
-            flags: ''
           - os: windows
             arch: amd64
-            preset: 'ROCm 6'
-            install: https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q4-WinSvr2022-For-HIP.exe
-            rocm-version: '6.2'
-            flags: '-DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_C_FLAGS="-parallel-jobs=4 -Wno-ignored-attributes -Wno-deprecated-pragma" -DCMAKE_CXX_FLAGS="-parallel-jobs=4 -Wno-ignored-attributes -Wno-deprecated-pragma"'
-            runner_dir: 'rocm'
+            preset: 'ROCm 7'
+            build-steps: rocm7
+            install: https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-26.Q1-Win11-For-HIP.exe
+            rocm-version: '7.1'
           - os: windows
             arch: amd64
             preset: Vulkan
+            build-steps: vulkan
             install: https://sdk.lunarg.com/sdk/download/1.4.321.1/windows/vulkansdk-windows-X64-1.4.321.1.exe
-            flags: ''
-            runner_dir: 'vulkan'
           - os: windows
             arch: amd64
             preset: 'MLX CUDA 13'
+            build-steps: mlxCuda13
+            build-parallel: '16'
+            cmake-cuda-flags: '-t 6'
             install: https://developer.download.nvidia.com/compute/cuda/13.0.0/local_installers/cuda_13.0.0_windows.exe
             cudnn-install: https://developer.download.nvidia.com/compute/cudnn/redist/cudnn/windows-x86_64/cudnn-windows-x86_64-9.18.1.3_cuda13-archive.zip
             cuda-components:
@@ -135,18 +140,34 @@ jobs:
               - '"nvvm"'
               - '"nvptxcompiler"'
             cuda-version: '13.0'
-            flags: ''
     runs-on: ${{ matrix.arch == 'arm64' && format('{0}-{1}', matrix.os, matrix.arch) || matrix.os }}
     environment: release
     env:
       GOFLAGS: ${{ needs.setup-environment.outputs.GOFLAGS }}
+      VERSION: ${{ needs.setup-environment.outputs.VERSION }}
     steps:
+      - if: startsWith(matrix.preset, 'MLX ')
+        name: Increase pagefile to 200 GB
+        uses: al-cheb/configure-pagefile-action@v1.5
+        with:
+          minimum-size: 16GB
+          maximum-size: 200GB
+          disk-root: "D:"
       - name: Install system dependencies
         run: |
           choco install -y --no-progress ccache ninja
           if (Get-Command ccache -ErrorAction SilentlyContinue) {
             ccache -o cache_dir=${{ github.workspace }}\.ccache
           }
+      - if: matrix.preset == 'CPU'
+        name: Install Windows ARM64 cross compiler
+        run: |
+          Invoke-WebRequest -Uri "https://github.com/mstorsjo/llvm-mingw/releases/download/20240619/llvm-mingw-20240619-ucrt-x86_64.zip" -OutFile "${{ runner.temp }}\llvm-mingw-ucrt.zip"
+          Expand-Archive -Path ${{ runner.temp }}\llvm-mingw-ucrt.zip -DestinationPath "C:\Program Files\"
+          $installPath=(Resolve-Path -Path "C:\Program Files\llvm-mingw-*-ucrt-x86_64").path
+          if (!(Test-Path "$installPath\bin\aarch64-w64-mingw32-gcc.exe")) {
+            throw "llvm-mingw x86_64 package is missing the aarch64 cross compiler"
+          }
       - if: startsWith(matrix.preset, 'CUDA ') || startsWith(matrix.preset, 'ROCm ') || startsWith(matrix.preset, 'Vulkan') || startsWith(matrix.preset, 'MLX ')
         id: cache-install
         uses: actions/cache/restore@v4
@@ -195,12 +216,12 @@ jobs:
           }
           
           $vulkanPath = (Resolve-Path "C:\VulkanSDK\*").path
+          $vulkanRuntime = Join-Path $vulkanPath "Helpers\VulkanRT.exe"
+          if (Test-Path $vulkanRuntime) {
+            Start-Process -FilePath $vulkanRuntime -ArgumentList "/s" -NoNewWindow -Wait
+          }
           echo "$vulkanPath\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
           echo "VULKAN_SDK=$vulkanPath" >> $env:GITHUB_ENV
-      - if: matrix.preset == 'CPU'
-        run: |
-          echo "CC=clang.exe" | Out-File -FilePath $env:GITHUB_ENV -Append
-          echo "CXX=clang++.exe" | Out-File -FilePath $env:GITHUB_ENV -Append
       - if: startsWith(matrix.preset, 'MLX ')
         name: Install cuDNN for MLX
         run: |
@@ -232,72 +253,63 @@ jobs:
         with:
           path: ${{ github.workspace }}\.ccache
           key: ccache-${{ matrix.os }}-${{ matrix.arch }}-${{ matrix.preset }}-${{ needs.setup-environment.outputs.vendorsha }}
-      - name: Build target "${{ matrix.preset }}"
+      - name: Build Windows dependencies
         run: |
           Import-Module 'C:\Program Files\Microsoft Visual Studio\2022\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
           Enter-VsDevShell -VsInstallPath 'C:\Program Files\Microsoft Visual Studio\2022\Enterprise' -SkipAutomaticLocation  -DevCmdArguments '-arch=x64 -no_logo'
-          cmake --preset "${{ matrix.preset }}" ${{ matrix.flags }} --install-prefix "$((pwd).Path)\dist\${{ matrix.os }}-${{ matrix.arch }}"
-          cmake --build --parallel ([Environment]::ProcessorCount) --preset "${{ matrix.preset }}"
-          cmake --install build --component "${{ startsWith(matrix.preset, 'MLX ') && 'MLX' || startsWith(matrix.preset, 'CUDA ') && 'CUDA' || startsWith(matrix.preset, 'ROCm ') && 'HIP' || startsWith(matrix.preset, 'Vulkan') && 'Vulkan' || 'CPU' }}" --strip
-          Remove-Item -Path dist\lib\ollama\rocm\rocblas\library\*gfx906* -ErrorAction SilentlyContinue
+          $steps = "${{ matrix.build-steps }}".Split(' ', [System.StringSplitOptions]::RemoveEmptyEntries)
+          ./scripts/build_windows.ps1 @steps
         env:
           CMAKE_GENERATOR: Ninja
+          OLLAMA_BUILD_PARALLEL: ${{ matrix.build-parallel || '' }}
+          OLLAMA_CMAKE_CUDA_FLAGS: ${{ matrix.cmake-cuda-flags || '' }}
       - name: Log build results
         run: |
           gci -path .\dist -Recurse -File | ForEach-Object { get-filehash -path $_.FullName -Algorithm SHA256 } | format-list
+      - if: matrix.preset == 'CPU'
+        name: Verify Windows CPU payloads
+        shell: bash
+        run: |
+          set -euo pipefail
+          for payload in \
+            dist/windows-amd64/lib/ollama/llama-server.exe \
+            dist/windows-arm64/lib/ollama/llama-server.exe
+          do
+            [ -f "$payload" ] || { echo "missing $payload"; exit 1; }
+          done
       - uses: actions/upload-artifact@v4
         with:
           name: depends-${{ matrix.os }}-${{ matrix.arch }}-${{ matrix.preset }}
           path: dist\*
 
   windows-build:
-    strategy:
-      matrix:
-        os: [windows]
-        arch: [amd64, arm64]
-        include:
-        - os: windows
-          arch: amd64
-          llvmarch: x86_64
-        - os: windows
-          arch: arm64
-          llvmarch: aarch64
-    runs-on: ${{ matrix.arch == 'arm64' && format('{0}-{1}', matrix.os, matrix.arch) || matrix.os }}
+    runs-on: windows
     environment: release
     needs: [setup-environment]
     env:
       GOFLAGS: ${{ needs.setup-environment.outputs.GOFLAGS }}
       VERSION: ${{ needs.setup-environment.outputs.VERSION }}
     steps:
-      - name: Install ARM64 system dependencies
-        if: matrix.arch == 'arm64'
-        run: |
-          $ErrorActionPreference = "Stop"
-          Set-ExecutionPolicy Bypass -Scope Process -Force
-          [System.Net.ServicePointManager]::SecurityProtocol = [System.Net.ServicePointManager]::SecurityProtocol -bor 3072
-          iex ((New-Object System.Net.WebClient).DownloadString('https://community.chocolatey.org/install.ps1'))
-          echo "C:\ProgramData\chocolatey\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
-
-          Invoke-WebRequest -Uri https://aka.ms/vs/17/release/vc_redist.arm64.exe -OutFile "${{ runner.temp }}\vc_redist.arm64.exe"
-          Start-Process -FilePath "${{ runner.temp }}\vc_redist.arm64.exe" -ArgumentList @("/install", "/quiet", "/norestart") -NoNewWindow -Wait
-
-          choco install -y --no-progress git gzip
-          echo "C:\Program Files\Git\cmd" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
       - name: Install clang and gcc-compat
         run: |
           $ErrorActionPreference = "Stop"
           Set-ExecutionPolicy Bypass -Scope Process -Force
-          Invoke-WebRequest -Uri "https://github.com/mstorsjo/llvm-mingw/releases/download/20240619/llvm-mingw-20240619-ucrt-${{ matrix.llvmarch }}.zip" -OutFile "${{ runner.temp }}\llvm-mingw-ucrt.zip"
+          Invoke-WebRequest -Uri "https://github.com/mstorsjo/llvm-mingw/releases/download/20240619/llvm-mingw-20240619-ucrt-x86_64.zip" -OutFile "${{ runner.temp }}\llvm-mingw-ucrt.zip"
           Expand-Archive -Path ${{ runner.temp }}\llvm-mingw-ucrt.zip -DestinationPath "C:\Program Files\"
-          $installPath=(Resolve-Path -Path "C:\Program Files\llvm-mingw-*-ucrt*").path
+          $installPath=(Resolve-Path -Path "C:\Program Files\llvm-mingw-*-ucrt-x86_64").path
           echo "$installPath\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+          if (!(Test-Path "$installPath\bin\aarch64-w64-mingw32-gcc.exe")) {
+            throw "llvm-mingw x86_64 package is missing the aarch64 cross compiler"
+          }
       - uses: actions/checkout@v4
       - uses: actions/setup-go@v5
         with:
           go-version-file: go.mod
           cache-dependency-path: |
             go.sum
-            Makefile.sync
+            LLAMA_CPP_VERSION
+            MLX_VERSION
+            MLX_C_VERSION
       - name: Verify gcc is actually clang
         run: |
           $ErrorActionPreference='Continue'
@@ -314,20 +326,30 @@ jobs:
         with:
           node-version: "20"
       - run: |
-          ./scripts/build_windows ollama app
+          ./scripts/build_windows ollama ollamaArm64 app appArm64
+      - name: Verify Windows build payloads
+        shell: bash
+        run: |
+          set -euo pipefail
+          for payload in \
+            dist/windows-amd64/ollama.exe \
+            dist/windows-arm64/ollama.exe
+          do
+            [ -f "$payload" ] || { echo "missing $payload"; exit 1; }
+          done
       - name: Log build results
         run: |
           gci -path .\dist -Recurse -File | ForEach-Object { get-filehash -path $_.FullName -Algorithm SHA256 } | format-list
       - uses: actions/upload-artifact@v4
         with:
-          name: build-${{ matrix.os }}-${{ matrix.arch }}
+          name: build-windows-amd64
           path: |
             dist\*
 
   windows-app:
     runs-on: windows
     environment: release
-    needs: [windows-build, windows-depends]
+    needs: [setup-environment, windows-build, windows-depends]
     env:
       GOFLAGS: ${{ needs.setup-environment.outputs.GOFLAGS }}
       VERSION: ${{ needs.setup-environment.outputs.VERSION }}
@@ -353,7 +375,9 @@ jobs:
           go-version-file: go.mod
           cache-dependency-path: |
             go.sum
-            Makefile.sync
+            LLAMA_CPP_VERSION
+            MLX_VERSION
+            MLX_C_VERSION
       - uses: actions/download-artifact@v4
         with:
           pattern: depends-windows*
@@ -367,6 +391,18 @@ jobs:
       - name: Log dist contents after download
         run: |
           gci -path .\dist -recurse
+      - name: Verify Windows package inputs
+        shell: bash
+        run: |
+          set -euo pipefail
+          for payload in \
+            dist/windows-amd64/ollama.exe \
+            dist/windows-amd64/lib/ollama/llama-server.exe \
+            dist/windows-arm64/ollama.exe \
+            dist/windows-arm64/lib/ollama/llama-server.exe
+          do
+            [ -f "$payload" ] || { echo "missing $payload"; exit 1; }
+          done
       - run: |
           ./scripts/build_windows.ps1 deps sign installer zip
       - name: Log contents after build
@@ -380,20 +416,33 @@ jobs:
             dist/*.ps1
             dist/OllamaSetup.exe
 
-  linux-build:
+  linux-depends:
     strategy:
       matrix:
         include:
-          - os: linux
-            arch: amd64
-            target: archive
-          - os: linux
-            arch: amd64
-            target: rocm
-          - os: linux
-            arch: arm64
-            target: archive
-    runs-on: ${{ matrix.arch == 'arm64' && format('{0}-{1}', matrix.os, matrix.arch) || matrix.os }}
+          - arch: amd64
+            target: llama-server-cpu
+          - arch: amd64
+            target: llama-server-cuda_v12
+          - arch: amd64
+            target: llama-server-cuda_v13
+          - arch: amd64
+            target: mlx
+          - arch: amd64
+            target: llama-server-rocm_v7_2
+          - arch: amd64
+            target: llama-server-vulkan
+          - arch: arm64
+            target: llama-server-cpu
+          - arch: arm64
+            target: llama-server-cuda_v12
+          - arch: arm64
+            target: llama-server-cuda_v13
+          - arch: arm64
+            target: jetpack-5
+          - arch: arm64
+            target: jetpack-6
+    runs-on: ${{ matrix.arch == 'arm64' && 'linux-arm64' || 'linux' }}
     environment: release
     needs: setup-environment
     env:
@@ -401,83 +450,114 @@ jobs:
     steps:
       - uses: actions/checkout@v4
       - uses: docker/setup-buildx-action@v3
+      - uses: docker/login-action@v3
+        with:
+          username: ${{ vars.DOCKER_USER }}
+          password: ${{ secrets.DOCKER_ACCESS_TOKEN }}
+      - if: matrix.target == 'mlx'
+        name: Increase Linux swap to 200 GB
+        shell: bash
+        run: |
+          set -e
+          SWAP_PATH=/swapfile-mlx
+          SWAP_SIZE_GB=200
+          if [ -f "$SWAP_PATH" ]; then
+              sudo swapoff "$SWAP_PATH" 2>/dev/null || true
+              sudo rm -f "$SWAP_PATH"
+          fi
+          if ! sudo fallocate -l ${SWAP_SIZE_GB}G "$SWAP_PATH" 2>/dev/null; then
+              echo "fallocate unsupported, falling back to dd"
+              sudo dd if=/dev/zero of="$SWAP_PATH" bs=1M count=$((SWAP_SIZE_GB * 1024))
+          fi
+          sudo chmod 600 "$SWAP_PATH"
+          sudo mkswap "$SWAP_PATH"
+          sudo swapon "$SWAP_PATH"
+          swapon --show
+          free -h
       - uses: docker/build-push-action@v6
         with:
           context: .
-          platforms: ${{ matrix.os }}/${{ matrix.arch }}
+          platforms: linux/${{ matrix.arch }}
           target: ${{ matrix.target }}
+          provenance: false
+          sbom: false
           build-args: |
             GOFLAGS=${{ env.GOFLAGS }}
             CGO_CFLAGS=${{ env.CGO_CFLAGS }}
             CGO_CXXFLAGS=${{ env.CGO_CXXFLAGS }}
-          outputs: type=local,dest=dist/${{ matrix.os }}-${{ matrix.arch }}
-          cache-from: type=registry,ref=${{ vars.DOCKER_REPO }}:latest
-          cache-to: type=inline
-      - name: Deduplicate CUDA libraries
-        run: |
-          ./scripts/deduplicate_cuda_libs.sh dist/${{ matrix.os }}-${{ matrix.arch }}
-      - run: |
-          for COMPONENT in bin/* lib/ollama/*; do
-            case "$COMPONENT" in
-              bin/ollama*)               echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
-              lib/ollama/*.so*)          echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
-              lib/ollama/cuda_v*)        echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
-              lib/ollama/vulkan*)        echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
-              lib/ollama/mlx*)           echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}-mlx.tar.in ;;
-              lib/ollama/include*)       echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}-mlx.tar.in ;;
-              lib/ollama/cuda_jetpack5)  echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}-jetpack5.tar.in ;;
-              lib/ollama/cuda_jetpack6)  echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}-jetpack6.tar.in ;;
-              lib/ollama/rocm)           echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}-rocm.tar.in ;;
-            esac
-          done
-        working-directory: dist/${{ matrix.os }}-${{ matrix.arch }}
-      - run: |
-          echo "Manifests"
-          for ARCHIVE in dist/${{ matrix.os }}-${{ matrix.arch }}/*.tar.in ; do
-            echo $ARCHIVE
-            cat $ARCHIVE
-          done
-      - run: |
-          for ARCHIVE in dist/${{ matrix.os }}-${{ matrix.arch }}/*.tar.in; do
-            tar c -C dist/${{ matrix.os }}-${{ matrix.arch }} -T $ARCHIVE --owner 0 --group 0 | zstd --ultra -22 -T0 >$(basename ${ARCHIVE//.*/}.tar.zst);
-          done
-      - uses: actions/upload-artifact@v4
-        with:
-          name: bundles-${{ matrix.os }}-${{ matrix.arch }}-${{ matrix.target }}
-          path: |
-            *.tar.zst
+            OLLAMA_MLX_BUILD_JOBS=16
+            OLLAMA_MLX_NVCC_THREADS=6
+            APT_MIRROR=http://azure.archive.ubuntu.com/ubuntu
+            APT_PORTS_MIRROR=http://azure.ports.ubuntu.com/ubuntu-ports
+          cache-from: |
+            type=registry,ref=ollama/release:cache-${{ matrix.arch }}-${{ matrix.target }}
+            type=registry,ref=${{ vars.DOCKER_REPO }}:latest
+          cache-to: type=registry,ref=ollama/release:cache-${{ matrix.arch }}-${{ matrix.target }},mode=max
 
   # Build each Docker variant (OS, arch, and flavor) separately. Using QEMU is unreliable and slower.
+  # Heavy stages were pre-built by linux-depends; this job is cache-hit-only for those layers
+  # and just assembles, runs the Go build, pushes the final image, and extracts release bundles.
   docker-build-push:
     strategy:
       matrix:
         include:
           - os: linux
             arch: arm64
+            archive-target: archive
             build-args: |
               CGO_CFLAGS
               CGO_CXXFLAGS
               GOFLAGS
               APT_MIRROR=http://azure.archive.ubuntu.com/ubuntu
+              APT_PORTS_MIRROR=http://azure.ports.ubuntu.com/ubuntu-ports
+              OLLAMA_MLX_BUILD_JOBS=16
+              OLLAMA_MLX_NVCC_THREADS=6
+            cache-from: |
+              type=registry,ref=ollama/release:cache-arm64-llama-server-cpu
+              type=registry,ref=ollama/release:cache-arm64-llama-server-cuda_v12
+              type=registry,ref=ollama/release:cache-arm64-llama-server-cuda_v13
+              type=registry,ref=ollama/release:cache-arm64-jetpack-5
+              type=registry,ref=ollama/release:cache-arm64-jetpack-6
+              type=registry,ref=${{ vars.DOCKER_REPO }}:latest
           - os: linux
             arch: amd64
+            archive-target: archive
             build-args: |
               CGO_CFLAGS
               CGO_CXXFLAGS
               GOFLAGS
               APT_MIRROR=http://azure.archive.ubuntu.com/ubuntu
+              APT_PORTS_MIRROR=http://azure.ports.ubuntu.com/ubuntu-ports
+              OLLAMA_MLX_BUILD_JOBS=16
+              OLLAMA_MLX_NVCC_THREADS=6
+            cache-from: |
+              type=registry,ref=ollama/release:cache-amd64-llama-server-cpu
+              type=registry,ref=ollama/release:cache-amd64-llama-server-cuda_v12
+              type=registry,ref=ollama/release:cache-amd64-llama-server-cuda_v13
+              type=registry,ref=ollama/release:cache-amd64-mlx
+              type=registry,ref=ollama/release:cache-amd64-llama-server-rocm_v7_2
+              type=registry,ref=ollama/release:cache-amd64-llama-server-vulkan
+              type=registry,ref=${{ vars.DOCKER_REPO }}:latest
           - os: linux
             arch: amd64
             suffix: '-rocm'
+            archive-target: image-archive
             build-args: |
               CGO_CFLAGS
               CGO_CXXFLAGS
               GOFLAGS
               FLAVOR=rocm
               APT_MIRROR=http://azure.archive.ubuntu.com/ubuntu
+              APT_PORTS_MIRROR=http://azure.ports.ubuntu.com/ubuntu-ports
+              OLLAMA_MLX_BUILD_JOBS=16
+              OLLAMA_MLX_NVCC_THREADS=6
+            cache-from: |
+              type=registry,ref=ollama/release:cache-amd64-llama-server-cpu
+              type=registry,ref=ollama/release:cache-amd64-llama-server-rocm_v7_2
+              type=registry,ref=${{ vars.DOCKER_REPO }}:latest
     runs-on: ${{ matrix.arch == 'arm64' && format('{0}-{1}', matrix.os, matrix.arch) || matrix.os }}
     environment: release
-    needs: setup-environment
+    needs: [setup-environment, linux-depends]
     env:
       GOFLAGS: ${{ needs.setup-environment.outputs.GOFLAGS }}
     steps:
@@ -492,9 +572,11 @@ jobs:
         with:
           context: .
           platforms: ${{ matrix.os }}/${{ matrix.arch }}
+          provenance: false
+          sbom: false
           build-args: ${{ matrix.build-args }}
           outputs: type=image,name=${{ vars.DOCKER_REPO }},push-by-digest=true,name-canonical=true,push=true
-          cache-from: type=registry,ref=${{ vars.DOCKER_REPO }}:latest
+          cache-from: ${{ matrix.cache-from }}
           cache-to: type=inline
       - run: |
           mkdir -p ${{ matrix.os }}-${{ matrix.arch }}
@@ -505,6 +587,64 @@ jobs:
           name: digest-${{ matrix.os }}-${{ matrix.arch }}-${{ matrix.suffix }}
           path: |
             ${{ runner.temp }}/${{ matrix.os }}-${{ matrix.arch }}-${{ matrix.suffix }}.txt
+      - uses: docker/build-push-action@v6
+        with:
+          context: .
+          platforms: ${{ matrix.os }}/${{ matrix.arch }}
+          target: ${{ matrix.archive-target }}
+          provenance: false
+          sbom: false
+          build-args: ${{ matrix.build-args }}
+          outputs: type=local,dest=dist/${{ matrix.os }}-${{ matrix.arch }}
+          cache-from: ${{ matrix.cache-from }}
+      - name: Deduplicate CUDA libraries
+        run: |
+          ./scripts/deduplicate_cuda_libs.sh dist/${{ matrix.os }}-${{ matrix.arch }}
+      - name: Verify Linux build payloads
+        shell: bash
+        run: |
+          set -euo pipefail
+          base="dist/${{ matrix.os }}-${{ matrix.arch }}"
+          for payload in \
+            "$base/bin/ollama" \
+            "$base/lib/ollama/llama-server"
+          do
+            [ -f "$payload" ] || { echo "missing $payload"; exit 1; }
+          done
+      - run: |
+          for COMPONENT in bin/* lib/ollama/*; do
+            case "$COMPONENT" in
+              bin/ollama*)               echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
+              lib/ollama/*.so*)          echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
+              lib/ollama/llama-server*|lib/ollama/llama-quantize*)  echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
+              lib/ollama/cuda_v*)        echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
+              lib/ollama/vulkan*)        echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
+              lib/ollama/mlx*)           echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}-mlx.tar.in ;;
+              lib/ollama/include*)       echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}-mlx.tar.in ;;
+              lib/ollama/cuda_jetpack5)  echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}-jetpack5.tar.in ;;
+              lib/ollama/cuda_jetpack6)  echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}-jetpack6.tar.in ;;
+              lib/ollama/rocm_v*)        echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}-rocm.tar.in ;;
+            esac
+          done
+        working-directory: dist/${{ matrix.os }}-${{ matrix.arch }}
+      - if: matrix.suffix == '-rocm'
+        run: rm -f dist/${{ matrix.os }}-${{ matrix.arch }}/ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in
+      - run: |
+          echo "Manifests"
+          for ARCHIVE in dist/${{ matrix.os }}-${{ matrix.arch }}/*.tar.in ; do
+            echo $ARCHIVE
+            cat $ARCHIVE
+          done
+      - run: |
+          for ARCHIVE in dist/${{ matrix.os }}-${{ matrix.arch }}/*.tar.in; do
+            tar c -C dist/${{ matrix.os }}-${{ matrix.arch }} -T $ARCHIVE --owner 0 --group 0 | zstd -19 -T0 >$(basename ${ARCHIVE//.*/}.tar.zst) &
+          done
+          wait
+      - uses: actions/upload-artifact@v4
+        with:
+          name: bundles-${{ matrix.os }}-${{ matrix.arch }}${{ matrix.suffix }}
+          path: |
+            *.tar.zst
 
   # Merge Docker images for the same flavor into a single multi-arch manifest
   docker-merge-push:
@@ -544,7 +684,7 @@ jobs:
   release:
     runs-on: ubuntu-latest
     environment: release
-    needs: [darwin-build, windows-app, linux-build]
+    needs: [darwin-build, windows-app, docker-build-push]
     permissions:
       contents: write
     env:
@@ -562,6 +702,21 @@ jobs:
       - name: Copy install scripts to dist
         run: |
           cp scripts/install.sh dist/install.sh
+      - name: Verify release artifacts
+        run: |
+          required=(
+            dist/OllamaSetup.exe
+            dist/install.ps1
+            dist/install.sh
+            dist/ollama-windows-amd64.zip
+            dist/ollama-windows-arm64.zip
+          )
+          for payload in "${required[@]}"; do
+            if [ ! -f "$payload" ]; then
+              echo "::error::Missing expected release artifact: $payload"
+              exit 1
+            fi
+          done
       - name: Generate checksum file
         run: find . -type f -not -name 'sha256sum.txt' | xargs sha256sum | tee sha256sum.txt
         working-directory: dist

.github/workflows/test.yaml

@@ -22,7 +22,8 @@ jobs:
     runs-on: ubuntu-latest
     outputs:
       changed: ${{ steps.changes.outputs.changed }}
-      vendorsha: ${{ steps.changes.outputs.vendorsha }}
+      app_changed: ${{ steps.changes.outputs.app_changed }}
+      enginehash: ${{ steps.changes.outputs.enginehash }}
     steps:
       - uses: actions/checkout@v4
         with:
@@ -37,8 +38,42 @@ jobs:
               | xargs python3 -c "import sys; from pathlib import Path; print(any(Path(x).match(glob) for x in sys.argv[1:] for glob in '$*'.split(' ')))"
           }
 
-          echo changed=$(changed 'llama/llama.cpp/**/*' 'ml/backend/ggml/ggml/**/*' '.github/**/*') | tee -a $GITHUB_OUTPUT
-          echo vendorsha=$(make -f Makefile.sync print-base) | tee -a $GITHUB_OUTPUT
+          echo changed=$(changed \
+            'CMakeLists.txt' \
+            'CMakePresets.json' \
+            'cmake/**' \
+            'cmake/**/*' \
+            'llama/server/**/*' \
+            'llama/compat/**/*' \
+            'LLAMA_CPP_VERSION' \
+            'MLX_VERSION' \
+            'MLX_C_VERSION' \
+            'llama/llama.cpp/**/*' \
+            'ml/backend/ggml/ggml/**/*' \
+            'x/imagegen/mlx/**' \
+            'x/imagegen/mlx/**/*' \
+            '.github/**/*') | tee -a $GITHUB_OUTPUT
+          echo app_changed=$(changed 'app/**' 'app/**/*') | tee -a $GITHUB_OUTPUT
+          echo enginehash=$(cat LLAMA_CPP_VERSION)-$(cat MLX_VERSION)-$(cat MLX_C_VERSION) | tee -a $GITHUB_OUTPUT
+
+  patches:
+    strategy:
+      matrix:
+        os: [ubuntu-latest, windows-latest]
+    runs-on: ${{ matrix.os }}
+    steps:
+      - uses: actions/checkout@v4
+      - name: Verify patches apply cleanly
+        shell: bash
+        run: |
+          cmake -S llama/server -B "$RUNNER_TEMP/llama-server-patch-check" \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DBUILD_SHARED_LIBS=ON \
+            -DGGML_BACKEND_DL=ON \
+            -DGGML_NATIVE=OFF \
+            -DGGML_OPENMP=OFF \
+            -DGGML_CPU_ALL_VARIANTS=ON \
+            -DOLLAMA_RUNNER_DIR=
 
   linux:
     needs: [changes]
@@ -47,23 +82,41 @@ jobs:
       matrix:
         include:
           - preset: CPU
+            superbuild_target: ollama-local
+            superbuild_dir: build/local-superbuild
+            superbuild_args: ''
+            expected_payload: lib/ollama/llama-server
+            install-go: true
           - preset: CUDA
             container: nvidia/cuda:13.0.0-devel-ubuntu22.04
-            flags: '-DCMAKE_CUDA_ARCHITECTURES=87'
+            superbuild_target: ollama-llama-server-cuda_v13
+            superbuild_dir: build/local-superbuild-cuda_v13
+            superbuild_args: '-DOLLAMA_LLAMA_BACKENDS=cuda_v13 -DCMAKE_CUDA_ARCHITECTURES=87'
+            expected_payload: lib/ollama/cuda_v13/libggml-cuda.so
           - preset: ROCm
             container: rocm/dev-ubuntu-22.04:7.2.1
             extra-packages: rocm-libs
-            flags: '-DAMDGPU_TARGETS=gfx1010 -DCMAKE_PREFIX_PATH=/opt/rocm'
+            superbuild_target: ollama-llama-server-rocm_v7_2
+            superbuild_dir: build/local-superbuild-rocm_v7_2
+            superbuild_args: '-DOLLAMA_LLAMA_BACKENDS=rocm_v7_2 -DAMDGPU_TARGETS=gfx1010 -DCMAKE_PREFIX_PATH=/opt/rocm'
+            expected_payload: lib/ollama/rocm_v7_2/libggml-hip.so
           - preset: Vulkan
             container: ubuntu:22.04
             extra-packages: >
               mesa-vulkan-drivers vulkan-tools
               libvulkan1 libvulkan-dev
-              vulkan-sdk cmake ccache g++ make
+              vulkan-sdk spirv-headers cmake ccache g++ make
+            superbuild_target: ollama-llama-server-vulkan
+            superbuild_dir: build/local-superbuild-vulkan
+            superbuild_args: '-DOLLAMA_LLAMA_BACKENDS=vulkan'
+            expected_payload: lib/ollama/vulkan/libggml-vulkan.so
           - preset: 'MLX CUDA 13'
             container: nvidia/cuda:13.0.0-devel-ubuntu22.04
             extra-packages: libcudnn9-dev-cuda-13 libopenblas-dev liblapack-dev liblapacke-dev git curl
-            flags: '-DCMAKE_CUDA_ARCHITECTURES=87 -DBLAS_INCLUDE_DIRS=/usr/include/x86_64-linux-gnu -DLAPACK_INCLUDE_DIRS=/usr/include/x86_64-linux-gnu'
+            superbuild_target: ollama-mlx-cuda_v13
+            superbuild_dir: build/local-superbuild-mlx-cuda_v13
+            superbuild_args: '-DOLLAMA_MLX_BACKENDS=cuda_v13 -DCMAKE_CUDA_ARCHITECTURES=87 -DMLX_CUDA_ARCHITECTURES=80-virtual -DBLAS_INCLUDE_DIRS=/usr/include/x86_64-linux-gnu -DLAPACK_INCLUDE_DIRS=/usr/include/x86_64-linux-gnu'
+            expected_payload: lib/ollama/mlx_cuda_v13/libmlx.so
             install-go: true
     runs-on: linux
     container: ${{ matrix.container }}
@@ -80,11 +133,9 @@ jobs:
             echo "deb [signed-by=/usr/share/keyrings/lunarg-archive-keyring.gpg]  https://packages.lunarg.com/vulkan/1.4.313 jammy main" | $sudo tee /etc/apt/sources.list.d/lunarg-vulkan-1.4.313-jammy.list > /dev/null
             $sudo apt-get update
           fi
-          $sudo apt-get install -y cmake ccache ${{ matrix.extra-packages }}
-          # MLX requires CMake 3.25+, install from official releases
-          if [ "${{ matrix.preset }}" = "MLX CUDA 13" ]; then
-            curl -fsSL https://github.com/Kitware/CMake/releases/download/v3.31.2/cmake-3.31.2-linux-$(uname -m).tar.gz | $sudo tar xz -C /usr/local --strip-components 1
-          fi
+          $sudo apt-get install -y cmake ccache curl git ${{ matrix.extra-packages }}
+          # Use a current CMake for upstream llama.cpp and Vulkan dependency discovery.
+          curl -fsSL https://github.com/Kitware/CMake/releases/download/v3.31.2/cmake-3.31.2-linux-$(uname -m).tar.gz | $sudo tar xz -C /usr/local --strip-components 1
           # Export VULKAN_SDK if provided by LunarG package (defensive)
           if [ -d "/usr/lib/x86_64-linux-gnu/vulkan" ] && [ "${{ matrix.preset }}" = "Vulkan" ]; then
             echo "VULKAN_SDK=/usr" >> $GITHUB_ENV
@@ -94,17 +145,30 @@ jobs:
       - if: matrix.install-go
         name: Install Go
         run: |
+          [ -n "${{ matrix.container }}" ] || sudo=sudo
           GO_VERSION=$(awk '/^go / { print $2 }' go.mod)
-          curl -fsSL "https://golang.org/dl/go${GO_VERSION}.linux-$(dpkg --print-architecture).tar.gz" | tar xz -C /usr/local
+          curl -fsSL "https://golang.org/dl/go${GO_VERSION}.linux-$(dpkg --print-architecture).tar.gz" | $sudo tar xz -C /usr/local
           echo "/usr/local/go/bin" >> $GITHUB_PATH
       - uses: actions/cache@v4
         with:
           path: /github/home/.cache/ccache
-          key: ccache-${{ runner.os }}-${{ runner.arch }}-${{ matrix.preset }}-${{ needs.changes.outputs.vendorsha }}
-      - run: |
-          cmake --preset "${{ matrix.preset }}" ${{ matrix.flags }}
-          cmake --build --preset "${{ matrix.preset }}" --parallel
-
+          key: ccache-${{ runner.os }}-${{ runner.arch }}-${{ matrix.preset }}-${{ needs.changes.outputs.enginehash }}
+      - name: Build native superbuild
+        if: matrix.superbuild_target
+        run: |
+          cmake -S . -B "${{ matrix.superbuild_dir }}" ${{ matrix.superbuild_args }}
+          CMAKE_BUILD_PARALLEL_LEVEL=$(nproc) cmake --build "${{ matrix.superbuild_dir }}" --target "${{ matrix.superbuild_target }}" -- -l $(nproc)
+          test -e "${{ matrix.superbuild_dir }}/${{ matrix.expected_payload }}"
+      - name: Verify local superbuild install
+        if: matrix.superbuild_target == 'ollama-local'
+        run: |
+          ./ollama --version
+          "${{ matrix.superbuild_dir }}/lib/ollama/llama-server" --version
+          test -x "${{ matrix.superbuild_dir }}/lib/ollama/llama-quantize"
+          cmake --install "${{ matrix.superbuild_dir }}" --component ollama-local --prefix "$RUNNER_TEMP/ollama-local"
+          "$RUNNER_TEMP/ollama-local/bin/ollama" --version
+          "$RUNNER_TEMP/ollama-local/lib/ollama/llama-server" --version
+          test -x "$RUNNER_TEMP/ollama-local/lib/ollama/llama-quantize"
   windows:
     needs: [changes]
     if: needs.changes.outputs.changed == 'True'
@@ -112,9 +176,16 @@ jobs:
       matrix:
         include:
           - preset: CPU
+            superbuild_target: ollama-local
+            superbuild_dir: build\local-superbuild
+            superbuild_args: ''
+            expected_payload: lib\ollama\llama-server.exe
           - preset: CUDA
             install: https://developer.download.nvidia.com/compute/cuda/13.0.0/local_installers/cuda_13.0.0_windows.exe
-            flags: '-DCMAKE_CUDA_ARCHITECTURES=80'
+            superbuild_target: ollama-llama-server-cuda_v13
+            superbuild_dir: build\local-superbuild-cuda_v13
+            superbuild_args: '-DOLLAMA_LLAMA_BACKENDS=cuda_v13 -DCMAKE_CUDA_ARCHITECTURES=80'
+            expected_payload: lib\ollama\cuda_v13\ggml-cuda.dll
             cuda-components:
               - '"cudart"'
               - '"nvcc"'
@@ -125,14 +196,26 @@ jobs:
               - '"nvptxcompiler"'
             cuda-version: '13.0'
           - preset: ROCm
-            install: https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q4-WinSvr2022-For-HIP.exe
-            flags: '-DAMDGPU_TARGETS=gfx1010 -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_C_FLAGS="-parallel-jobs=4 -Wno-ignored-attributes -Wno-deprecated-pragma" -DCMAKE_CXX_FLAGS="-parallel-jobs=4 -Wno-ignored-attributes -Wno-deprecated-pragma"'
+            install: https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-26.Q1-Win11-For-HIP.exe
+            rocm-version: '7.1'
+            superbuild_target: ollama-llama-server-rocm_v7_1
+            superbuild_dir: build\local-superbuild-rocm_v7_1
+            superbuild_args: '-DOLLAMA_LLAMA_BACKENDS=rocm_v7_1 -DAMDGPU_TARGETS=gfx1010'
+            expected_payload: lib\ollama\rocm_v7_1\ggml-hip.dll
           - preset: Vulkan
             install: https://sdk.lunarg.com/sdk/download/1.4.321.1/windows/vulkansdk-windows-X64-1.4.321.1.exe
+            superbuild_target: ollama-llama-server-vulkan
+            superbuild_dir: build\local-superbuild-vulkan
+            superbuild_args: '-DOLLAMA_LLAMA_BACKENDS=vulkan'
+            expected_payload: lib\ollama\vulkan\ggml-vulkan.dll
           - preset: 'MLX CUDA 13'
             install: https://developer.download.nvidia.com/compute/cuda/13.0.0/local_installers/cuda_13.0.0_windows.exe
             cudnn-install: https://developer.download.nvidia.com/compute/cudnn/redist/cudnn/windows-x86_64/cudnn-windows-x86_64-9.18.1.3_cuda13-archive.zip
-            flags: '-DCMAKE_CUDA_ARCHITECTURES=80'
+            superbuild_target: ollama-mlx-cuda_v13
+            superbuild_dir: build\local-superbuild-mlx-cuda_v13
+            superbuild_args: '-DOLLAMA_MLX_BACKENDS=cuda_v13 -DCMAKE_CUDA_ARCHITECTURES=80 -DMLX_CUDA_ARCHITECTURES=80-virtual'
+            expected_payload: lib\ollama\mlx_cuda_v13\mlx.dll
+            install-go: true
             cuda-components:
               - '"cudart"'
               - '"nvcc"'
@@ -201,6 +284,10 @@ jobs:
           }
 
           $vulkanPath = (Resolve-Path "C:\VulkanSDK\*").path
+          $vulkanRuntime = Join-Path $vulkanPath "Helpers\VulkanRT.exe"
+          if (Test-Path $vulkanRuntime) {
+            Start-Process -FilePath $vulkanRuntime -ArgumentList "/s" -NoNewWindow -Wait
+          }
           echo "$vulkanPath\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
           echo "VULKAN_SDK=$vulkanPath" >> $env:GITHUB_ENV
       - if: matrix.preset == 'MLX CUDA 13'
@@ -230,18 +317,44 @@ jobs:
             C:\Program Files\NVIDIA\CUDNN
           key: ${{ matrix.install }}-${{ matrix.cudnn-install }}
       - uses: actions/checkout@v4
+      - if: matrix.superbuild_target == 'ollama-local' || matrix.install-go
+        uses: actions/setup-go@v5
+        with:
+          go-version-file: 'go.mod'
       - uses: actions/cache@v4
         with:
           path: ${{ github.workspace }}\.ccache
-          key: ccache-${{ runner.os }}-${{ runner.arch }}-${{ matrix.preset }}-${{ needs.changes.outputs.vendorsha }}
-      - run: |
+          key: ccache-${{ runner.os }}-${{ runner.arch }}-${{ matrix.preset }}-${{ needs.changes.outputs.enginehash }}
+      - name: Build native superbuild
+        if: matrix.superbuild_target
+        run: |
+          $ErrorActionPreference = "Stop"
           Import-Module 'C:\Program Files\Microsoft Visual Studio\2022\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
           Enter-VsDevShell -VsInstallPath 'C:\Program Files\Microsoft Visual Studio\2022\Enterprise' -SkipAutomaticLocation  -DevCmdArguments '-arch=x64 -no_logo'
-          cmake --preset "${{ matrix.preset }}" ${{ matrix.flags }}
-          cmake --build --parallel --preset "${{ matrix.preset }}"
+          cmake -S . -B "${{ matrix.superbuild_dir }}" ${{ matrix.superbuild_args }}
+          $env:CMAKE_BUILD_PARALLEL_LEVEL = [Environment]::ProcessorCount
+          cmake --build "${{ matrix.superbuild_dir }}" --target "${{ matrix.superbuild_target }}" -- -l $([Environment]::ProcessorCount)
+          if (!(Test-Path "${{ matrix.superbuild_dir }}\${{ matrix.expected_payload }}")) {
+            throw "missing ${{ matrix.expected_payload }}"
+          }
         env:
           CMAKE_GENERATOR: Ninja
-
+      - name: Verify local superbuild install
+        if: matrix.superbuild_target == 'ollama-local'
+        run: |
+          $ErrorActionPreference = "Stop"
+          & ".\ollama.exe" --version
+          & "${{ matrix.superbuild_dir }}\lib\ollama\llama-server.exe" --version
+          if (!(Test-Path "${{ matrix.superbuild_dir }}\lib\ollama\llama-quantize.exe")) {
+            throw "missing llama-quantize.exe"
+          }
+          $installPrefix = Join-Path $env:RUNNER_TEMP "ollama-local"
+          cmake --install "${{ matrix.superbuild_dir }}" --component ollama-local --prefix "$installPrefix"
+          & "$installPrefix\bin\ollama.exe" --version
+          & "$installPrefix\lib\ollama\llama-server.exe" --version
+          if (!(Test-Path "$installPrefix\lib\ollama\llama-quantize.exe")) {
+            throw "missing installed llama-quantize.exe"
+          }
   go_mod_tidy:
     runs-on: ubuntu-latest
     steps:
@@ -250,6 +363,7 @@ jobs:
         run: go mod tidy --diff || (echo "Please run 'go mod tidy'." && exit 1)
 
   test:
+    needs: [changes]
     strategy:
       matrix:
         os: [ubuntu-latest, macos-latest, windows-latest]
@@ -263,7 +377,9 @@ jobs:
           go-version-file: 'go.mod'
           cache-dependency-path: |
             go.sum
-            Makefile.sync
+            LLAMA_CPP_VERSION
+            MLX_VERSION
+            MLX_C_VERSION
       - uses: actions/setup-node@v4
         with:
           node-version: '20'
@@ -277,6 +393,17 @@ jobs:
         if: ${{ startsWith(matrix.os, 'ubuntu') }}
         working-directory: ./app/ui/app
         run: npm test
+      - name: Verify MLX generated files are current
+        if: ${{ startsWith(matrix.os, 'ubuntu') }}
+        run: |
+          cmake -S . -B build/mlx-generate -DOLLAMA_MLX_BACKENDS=cuda_v13
+          cmake --build build/mlx-generate --target ollama-mlx-generate-wrappers
+          git diff --exit-code -- \
+            x/imagegen/mlx/mlx.h \
+            x/imagegen/mlx/mlx.c \
+            x/mlxrunner/mlx/generated.h \
+            x/mlxrunner/mlx/generated.c \
+            x/mlxrunner/mlx/include/mlx/c
       - name: Run go generate
         run: go generate ./...
 
@@ -284,15 +411,10 @@ jobs:
         if: always()
         run: go test -count=1 -benchtime=1x ./...
 
+      - name: go test app with live updater tag
+        if: ${{ needs.changes.outputs.app_changed == 'True' && contains(fromJSON('["macos-latest","windows-latest"]'), matrix.os) }}
+        run: go test -count=1 -tags updater_live ./app/...
+
       - uses: golangci/golangci-lint-action@v9
         with:
           only-new-issues: true
-
-  patches:
-    runs-on: ubuntu-latest
-    steps:
-      - uses: actions/checkout@v4
-      - name: Verify patches apply cleanly and do not change files
-        run: |
-          make -f Makefile.sync clean checkout apply-patches sync
-          git diff --compact-summary --exit-code

CMakeLists.txt

@@ -1,4 +1,4 @@
-cmake_minimum_required(VERSION 3.21)
+cmake_minimum_required(VERSION 3.24)
 
 project(Ollama C CXX)
 
@@ -23,30 +23,23 @@ include(GNUInstallDirs)
 
 find_package(Threads REQUIRED)
 
-set(CMAKE_BUILD_TYPE Release)
-set(BUILD_SHARED_LIBS ON)
+if(NOT CMAKE_CONFIGURATION_TYPES AND NOT CMAKE_BUILD_TYPE)
+    set(CMAKE_BUILD_TYPE Release CACHE STRING "Build type" FORCE)
+endif()
+
+# These defaults can be overridden by presets (e.g., for static macOS llama-server builds)
+if(NOT DEFINED BUILD_SHARED_LIBS)
+    set(BUILD_SHARED_LIBS ON)
+endif()
 
 set(CMAKE_CXX_STANDARD 17)
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
-set(CMAKE_CXX_EXTENSIONS ON) # Recent versions of MLX Requires gnu++17 extensions to compile properly
+set(CMAKE_CXX_EXTENSIONS ON) # Recent versions of MLX require gnu++17 extensions to compile properly
 
-set(GGML_BUILD ON)
-set(GGML_SHARED ON)
-set(GGML_CCACHE ON)
-set(GGML_BACKEND_DL ON)
-set(GGML_BACKEND_SHARED ON)
-set(GGML_SCHED_MAX_COPIES 4)
-
-set(GGML_LLAMAFILE ON)
-set(GGML_CUDA_PEER_MAX_BATCH_SIZE 128)
-set(GGML_CUDA_GRAPHS ON)
-set(GGML_CUDA_FA ON)
-set(GGML_CUDA_COMPRESSION_MODE default)
-
-if((CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_OSX_ARCHITECTURES MATCHES "arm64")
-    OR (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_SYSTEM_PROCESSOR MATCHES "arm|aarch64|ARM64|ARMv[0-9]+"))
-    set(GGML_CPU_ALL_VARIANTS ON)
-endif()
+# GGML backend for inference is provided by llama-server (built separately via
+# llama/server/CMakeLists.txt using FetchContent from the pinned llama.cpp source).
+# The root CMake project is the orchestration entrypoint; backend-specific
+# build rules live in subprojects under cmake/.
 
 if(APPLE)
     set(CMAKE_BUILD_RPATH "@loader_path")
@@ -55,7 +48,8 @@ if(APPLE)
 endif()
 
 set(OLLAMA_BUILD_DIR ${CMAKE_BINARY_DIR}/lib/ollama)
-set(OLLAMA_INSTALL_DIR ${CMAKE_INSTALL_PREFIX}/lib/ollama/${OLLAMA_RUNNER_DIR})
+set(OLLAMA_LIB_DIR "lib/ollama" CACHE STRING "Install destination for Ollama runtime payloads")
+set(OLLAMA_INSTALL_DIR ${OLLAMA_LIB_DIR}/${OLLAMA_RUNNER_DIR})
 
 set(CMAKE_RUNTIME_OUTPUT_DIRECTORY         ${OLLAMA_BUILD_DIR})
 set(CMAKE_RUNTIME_OUTPUT_DIRECTORY_DEBUG   ${OLLAMA_BUILD_DIR})
@@ -64,309 +58,9 @@ set(CMAKE_LIBRARY_OUTPUT_DIRECTORY         ${OLLAMA_BUILD_DIR})
 set(CMAKE_LIBRARY_OUTPUT_DIRECTORY_DEBUG   ${OLLAMA_BUILD_DIR})
 set(CMAKE_LIBRARY_OUTPUT_DIRECTORY_RELEASE ${OLLAMA_BUILD_DIR})
 
-# Store ggml include paths for use with target_include_directories later.
-# We avoid global include_directories() to prevent polluting the include path
-# for other projects like MLX (whose openblas dependency has its own common.h).
-set(GGML_INCLUDE_DIRS
-    ${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src
-    ${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/include
-    ${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-cpu
-    ${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-cpu/amx
-)
-
-add_compile_definitions(NDEBUG GGML_VERSION=0x0 GGML_COMMIT=0x0)
-
-# Define GGML version variables for shared library SOVERSION
-# These are required by ggml/src/CMakeLists.txt for proper library versioning
-set(GGML_VERSION_MAJOR 0)
-set(GGML_VERSION_MINOR 0)
-set(GGML_VERSION_PATCH 0)
-set(GGML_VERSION "${GGML_VERSION_MAJOR}.${GGML_VERSION_MINOR}.${GGML_VERSION_PATCH}")
-
-set(GGML_CPU ON)
-add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src)
-set_property(TARGET ggml PROPERTY EXCLUDE_FROM_ALL TRUE)
-
-get_target_property(CPU_VARIANTS ggml-cpu MANUALLY_ADDED_DEPENDENCIES)
-if(NOT CPU_VARIANTS)
-    set(CPU_VARIANTS "ggml-cpu")
-endif()
-
-# Apply ggml include directories to ggml targets only (not globally)
-target_include_directories(ggml-base PRIVATE ${GGML_INCLUDE_DIRS})
-foreach(variant ${CPU_VARIANTS})
-    if(TARGET ${variant})
-        target_include_directories(${variant} PRIVATE ${GGML_INCLUDE_DIRS})
-    endif()
-endforeach()
-
-install(TARGETS ggml-base ${CPU_VARIANTS}
-    RUNTIME_DEPENDENCIES
-        PRE_EXCLUDE_REGEXES ".*"
-    RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT CPU
-    LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT CPU
-    FRAMEWORK DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT CPU
-)
-
-check_language(CUDA)
-if(CMAKE_CUDA_COMPILER)
-    if(CMAKE_VERSION VERSION_GREATER_EQUAL "3.24" AND NOT CMAKE_CUDA_ARCHITECTURES)
-        set(CMAKE_CUDA_ARCHITECTURES "native")
-    endif()
-
-    find_package(CUDAToolkit)
-    add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-cuda)
-    target_include_directories(ggml-cuda PRIVATE ${GGML_INCLUDE_DIRS})
-    install(TARGETS ggml-cuda
-        RUNTIME_DEPENDENCIES
-            DIRECTORIES ${CUDAToolkit_BIN_DIR} ${CUDAToolkit_BIN_DIR}/x64 ${CUDAToolkit_LIBRARY_DIR}
-            PRE_INCLUDE_REGEXES cublas cublasLt cudart
-            PRE_EXCLUDE_REGEXES ".*"
-        RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT CUDA
-        LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT CUDA
-    )
-endif()
-
-set(WINDOWS_AMDGPU_TARGETS_EXCLUDE_REGEX "^gfx(908|90a|1200|1201):xnack[+-]$"
-    CACHE STRING
-    "Regular expression describing AMDGPU_TARGETS not supported on Windows. Override to force building these targets. Default \"^gfx(908|90a|1200|1201):xnack[+-]$\"."
-)
-
-check_language(HIP)
-if(CMAKE_HIP_COMPILER)
-    set(HIP_PLATFORM "amd")
-
-    if(NOT AMDGPU_TARGETS)
-        find_package(hip REQUIRED)
-        list(FILTER AMDGPU_TARGETS INCLUDE REGEX "^gfx(94[012]|101[02]|1030|110[012]|120[01])$")
-    endif()
-
-    if(WIN32 AND WINDOWS_AMDGPU_TARGETS_EXCLUDE_REGEX)
-        list(FILTER AMDGPU_TARGETS EXCLUDE REGEX ${WINDOWS_AMDGPU_TARGETS_EXCLUDE_REGEX})
-    endif()
-
-    if(AMDGPU_TARGETS)
-        find_package(hip REQUIRED)
-        add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-hip)
-        target_include_directories(ggml-hip PRIVATE ${GGML_INCLUDE_DIRS})
-
-        if (WIN32)
-            target_compile_definitions(ggml-hip PRIVATE GGML_CUDA_NO_PEER_COPY)
-        endif()
-
-        target_compile_definitions(ggml-hip PRIVATE GGML_HIP_NO_VMM)
-
-        install(TARGETS ggml-hip
-            RUNTIME_DEPENDENCY_SET rocm
-            RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT HIP
-            LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT HIP
-        )
-        install(RUNTIME_DEPENDENCY_SET rocm
-                DIRECTORIES ${HIP_BIN_INSTALL_DIR} ${HIP_LIB_INSTALL_DIR}
-                PRE_INCLUDE_REGEXES hipblas rocblas amdhip64 rocsolver amd_comgr hsa-runtime64 rocsparse tinfo rocprofiler-register roctx64 rocroller drm drm_amdgpu numa elf
-                PRE_EXCLUDE_REGEXES ".*"
-                POST_EXCLUDE_REGEXES "system32"
-            RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT HIP
-            LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT HIP
-        )
-
-        foreach(HIP_LIB_BIN_INSTALL_DIR IN ITEMS ${HIP_BIN_INSTALL_DIR} ${HIP_LIB_INSTALL_DIR})
-            if(EXISTS ${HIP_LIB_BIN_INSTALL_DIR}/rocblas)
-                install(DIRECTORY ${HIP_LIB_BIN_INSTALL_DIR}/rocblas DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT HIP)
-                break()
-            endif()
-        endforeach()
-    endif()
-endif()
-
-if(NOT APPLE)
-    find_package(Vulkan)
-    if(Vulkan_FOUND)
-        add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-vulkan)
-        target_include_directories(ggml-vulkan PRIVATE ${GGML_INCLUDE_DIRS})
-        install(TARGETS ggml-vulkan
-            RUNTIME_DEPENDENCIES
-                PRE_INCLUDE_REGEXES vulkan
-                PRE_EXCLUDE_REGEXES ".*"
-            RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT Vulkan
-            LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT Vulkan
-        )
-    endif()
-endif()
-
-option(MLX_ENGINE "Enable MLX backend" OFF)
-if(MLX_ENGINE)
-    message(STATUS "Setting up MLX (this takes a while...)")
-    add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/x/imagegen/mlx)
-
-    # Find CUDA toolkit if MLX is built with CUDA support
-    find_package(CUDAToolkit)
-
-    # Build list of directories for runtime dependency resolution
-    set(MLX_RUNTIME_DIRS ${CUDAToolkit_BIN_DIR} ${CUDAToolkit_BIN_DIR}/x64 ${CUDAToolkit_LIBRARY_DIR})
-    # Add cuDNN bin paths for DLLs (Windows MLX CUDA builds)
-    # CUDNN_ROOT_DIR is the standard CMake variable for cuDNN location
-    if(DEFINED ENV{CUDNN_ROOT_DIR})
-        # cuDNN 9.x has versioned subdirectories under bin/ (e.g., bin/13.0/)
-        file(GLOB CUDNN_BIN_SUBDIRS "$ENV{CUDNN_ROOT_DIR}/bin/*")
-        list(APPEND MLX_RUNTIME_DIRS ${CUDNN_BIN_SUBDIRS})
-    endif()
-    # Add build output directory and MLX dependency build directories
-    list(APPEND MLX_RUNTIME_DIRS ${OLLAMA_BUILD_DIR})
-    # OpenBLAS DLL location (pre-built zip extracts into openblas-src/bin/)
-    list(APPEND MLX_RUNTIME_DIRS ${CMAKE_BINARY_DIR}/_deps/openblas-src/bin)
-    # NCCL: on Linux, if real NCCL is found, cmake bundles libnccl.so via the
-    # regex below. If NCCL is not found, MLX links a static stub (OBJECT lib)
-    # so there is no runtime dependency. This path covers the stub build dir
-    # for windows so we include the DLL in our dependencies.
-    list(APPEND MLX_RUNTIME_DIRS ${CMAKE_BINARY_DIR}/_deps/mlx-build/mlx/distributed/nccl/nccl_stub-prefix/src/nccl_stub-build/Release)
-
-    # Base regexes for runtime dependencies (cross-platform)
-    set(MLX_INCLUDE_REGEXES cublas cublasLt cudart cufft nvrtc nvrtc-builtins cudnn nccl openblas gfortran)
-    # On Windows, also include dl.dll (dlfcn-win32 POSIX emulation layer)
-    if(WIN32)
-        list(APPEND MLX_INCLUDE_REGEXES "^dl\\.dll$")
-    endif()
-
-    install(TARGETS mlx mlxc
-        RUNTIME_DEPENDENCIES
-            DIRECTORIES ${MLX_RUNTIME_DIRS}
-            PRE_INCLUDE_REGEXES ${MLX_INCLUDE_REGEXES}
-            PRE_EXCLUDE_REGEXES ".*"
-        RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT MLX
-        LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT MLX
-        FRAMEWORK DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT MLX
-    )
-
-    if(TARGET jaccl)
-        install(TARGETS jaccl
-            RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT MLX
-            LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT MLX
-            FRAMEWORK DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT MLX
-        )
-    endif()
-
-    # Install the Metal library for macOS arm64 (must be colocated with the binary)
-    # Metal backend is only built for arm64, not x86_64
-    if(APPLE AND CMAKE_SYSTEM_PROCESSOR STREQUAL "arm64")
-        install(FILES ${CMAKE_BINARY_DIR}/_deps/mlx-build/mlx/backend/metal/kernels/mlx.metallib
-            DESTINATION ${OLLAMA_INSTALL_DIR}
-            COMPONENT MLX)
-    endif()
-
-    # Install headers for NVRTC JIT compilation at runtime.
-    # MLX's own install rules use the default component so they get skipped by
-    # --component MLX. Headers are installed alongside libmlx in OLLAMA_INSTALL_DIR.
-    #
-    # Layout:
-    #   ${OLLAMA_INSTALL_DIR}/include/cccl/{cuda,nv}/  — CCCL headers
-    #   ${OLLAMA_INSTALL_DIR}/include/*.h               — CUDA toolkit headers
-    #
-    # MLX's jit_module.cpp resolves CCCL via
-    #   current_binary_dir()[.parent_path()] / "include" / "cccl"
-    # On Linux, MLX's jit_module.cpp resolves CCCL via
-    #   current_binary_dir().parent_path() / "include" / "cccl", so we create a
-    #   symlink from lib/ollama/include -> ${OLLAMA_RUNNER_DIR}/include
-    #   This will need refinement if we add multiple CUDA versions for MLX in the future.
-    # CUDA runtime headers are found via CUDA_PATH env var (set by mlxrunner).
-    if(EXISTS ${CMAKE_BINARY_DIR}/_deps/cccl-src/include/cuda)
-        install(DIRECTORY ${CMAKE_BINARY_DIR}/_deps/cccl-src/include/cuda
-            DESTINATION ${OLLAMA_INSTALL_DIR}/include/cccl
-            COMPONENT MLX)
-        install(DIRECTORY ${CMAKE_BINARY_DIR}/_deps/cccl-src/include/nv
-            DESTINATION ${OLLAMA_INSTALL_DIR}/include/cccl
-            COMPONENT MLX)
-        if(NOT WIN32 AND NOT APPLE)
-            install(CODE "
-                set(_link \"${CMAKE_INSTALL_PREFIX}/lib/ollama/include\")
-                set(_target \"${OLLAMA_RUNNER_DIR}/include\")
-                if(NOT EXISTS \${_link})
-                    execute_process(COMMAND \${CMAKE_COMMAND} -E create_symlink \${_target} \${_link})
-                endif()
-            " COMPONENT MLX)
-        endif()
-    endif()
-
-    # Install minimal CUDA toolkit headers needed by MLX JIT kernels.
-    # These are the transitive closure of includes from mlx/backend/cuda/device/*.cuh.
-    # The Go mlxrunner sets CUDA_PATH to OLLAMA_INSTALL_DIR so MLX finds them at
-    # $CUDA_PATH/include/*.h via NVRTC --include-path.
-    if(CUDAToolkit_FOUND)
-        # CUDAToolkit_INCLUDE_DIRS may be a semicolon-separated list
-        # (e.g. ".../include;.../include/cccl"). Find the entry that
-        # contains the CUDA runtime headers we need.
-        set(_cuda_inc "")
-        foreach(_dir ${CUDAToolkit_INCLUDE_DIRS})
-            if(EXISTS "${_dir}/cuda_runtime_api.h")
-                set(_cuda_inc "${_dir}")
-                break()
-            endif()
-        endforeach()
-        if(NOT _cuda_inc)
-            message(WARNING "Could not find cuda_runtime_api.h in CUDAToolkit_INCLUDE_DIRS: ${CUDAToolkit_INCLUDE_DIRS}")
-        else()
-            set(_dst "${OLLAMA_INSTALL_DIR}/include")
-            set(_MLX_JIT_CUDA_HEADERS
-                builtin_types.h
-                cooperative_groups.h
-                cuda_bf16.h
-                cuda_bf16.hpp
-                cuda_device_runtime_api.h
-                cuda_fp16.h
-                cuda_fp16.hpp
-                cuda_fp8.h
-                cuda_fp8.hpp
-                cuda_runtime_api.h
-                device_types.h
-                driver_types.h
-                math_constants.h
-                surface_types.h
-                texture_types.h
-                vector_functions.h
-                vector_functions.hpp
-                vector_types.h
-            )
-            foreach(_hdr ${_MLX_JIT_CUDA_HEADERS})
-                install(FILES "${_cuda_inc}/${_hdr}"
-                    DESTINATION ${_dst}
-                    COMPONENT MLX)
-            endforeach()
-            # Subdirectory headers
-            install(DIRECTORY "${_cuda_inc}/cooperative_groups"
-                DESTINATION ${_dst}
-                COMPONENT MLX
-                FILES_MATCHING PATTERN "*.h")
-            install(FILES "${_cuda_inc}/crt/host_defines.h"
-                DESTINATION "${_dst}/crt"
-                COMPONENT MLX)
-        endif()
-    endif()
-
-    # On Windows, explicitly install dl.dll (dlfcn-win32 POSIX dlopen emulation)
-    # RUNTIME_DEPENDENCIES auto-excludes it via POST_EXCLUDE_FILES_STRICT because
-    # dlfcn-win32 is a known CMake target with its own install rules (which install
-    # to the wrong destination). We must install it explicitly here.
-    if(WIN32)
-        install(FILES ${OLLAMA_BUILD_DIR}/dl.dll
-            DESTINATION ${OLLAMA_INSTALL_DIR}
-            COMPONENT MLX)
-    endif()
-
-    # Manually install CUDA runtime libraries that MLX loads via dlopen
-    # (not detected by RUNTIME_DEPENDENCIES since they aren't link-time deps)
-    if(CUDAToolkit_FOUND)
-        file(GLOB MLX_CUDA_LIBS
-            "${CUDAToolkit_LIBRARY_DIR}/libcudart.so*"
-            "${CUDAToolkit_LIBRARY_DIR}/libcublas.so*"
-            "${CUDAToolkit_LIBRARY_DIR}/libcublasLt.so*"
-            "${CUDAToolkit_LIBRARY_DIR}/libnvrtc.so*"
-            "${CUDAToolkit_LIBRARY_DIR}/libnvrtc-builtins.so*"
-            "${CUDAToolkit_LIBRARY_DIR}/libcufft.so*"
-            "${CUDAToolkit_LIBRARY_DIR}/libcudnn.so*")
-        if(MLX_CUDA_LIBS)
-            install(FILES ${MLX_CUDA_LIBS}
-                DESTINATION ${OLLAMA_INSTALL_DIR}
-                COMPONENT MLX)
-        endif()
-    endif()
+if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/llama/server/CMakeLists.txt")
+    set(OLLAMA_HAVE_LLAMA_SERVER TRUE)
+else()
+    set(OLLAMA_HAVE_LLAMA_SERVER FALSE)
 endif()
+include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/local.cmake)

CMakePresets.json

@@ -11,109 +11,10 @@
       }
     },
     {
-      "name": "CPU",
-      "inherits": [ "Default" ]
-    },
-    {
-      "name": "CUDA",
-      "inherits": [ "Default" ]
-    },
-    {
-      "name": "CUDA 11",
-      "inherits": [ "CUDA" ],
-      "cacheVariables": {
-        "CMAKE_CUDA_ARCHITECTURES": "50-virtual;60-virtual;61-virtual;70-virtual;75-virtual;80-virtual;86-virtual;87-virtual;89-virtual;90-virtual",
-        "CMAKE_CUDA_FLAGS": "-Wno-deprecated-gpu-targets -t 2",
-        "OLLAMA_RUNNER_DIR": "cuda_v11"
-      }
-    },
-    {
-      "name": "CUDA 12",
-      "inherits": [ "CUDA" ],
-      "cacheVariables": {
-        "CMAKE_CUDA_ARCHITECTURES": "50;52;60;61;70;75;80;86;89;90;90a;120",
-        "CMAKE_CUDA_FLAGS": "-Wno-deprecated-gpu-targets -t 2",
-        "OLLAMA_RUNNER_DIR": "cuda_v12"
-      }
-    },
-    {
-      "name": "CUDA 13",
-      "inherits": [ "CUDA" ],
-      "cacheVariables": {
-        "CMAKE_CUDA_ARCHITECTURES": "75-virtual;80-virtual;86-virtual;87-virtual;89-virtual;90-virtual;90a-virtual;100-virtual;103-virtual;110-virtual;120-virtual;121-virtual",
-        "CMAKE_CUDA_FLAGS": "-t 2",
-        "OLLAMA_RUNNER_DIR": "cuda_v13"
-      }
-    },
-    {
-      "name": "JetPack 5",
-      "inherits": [ "CUDA" ],
-      "cacheVariables": {
-        "CMAKE_CUDA_ARCHITECTURES": "72;87",
-        "OLLAMA_RUNNER_DIR": "cuda_jetpack5"
-      }
-    },
-    {
-      "name": "JetPack 6",
-      "inherits": [ "CUDA" ],
-      "cacheVariables": {
-        "CMAKE_CUDA_ARCHITECTURES": "87",
-        "OLLAMA_RUNNER_DIR": "cuda_jetpack6"
-      }
-    },
-    {
-      "name": "ROCm",
+      "name": "MLX Metal",
       "inherits": [ "Default" ],
       "cacheVariables": {
-        "CMAKE_HIP_PLATFORM": "amd"
-      }
-    },
-    {
-      "name": "ROCm 6",
-      "inherits": [ "ROCm" ],
-      "cacheVariables": {
-        "CMAKE_HIP_FLAGS": "-parallel-jobs=4",
-        "AMDGPU_TARGETS": "gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102;gfx1151;gfx1200;gfx1201;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-",
-        "OLLAMA_RUNNER_DIR": "rocm"
-      }
-    },
-    {
-      "name": "ROCm 7",
-      "inherits": [ "ROCm" ],
-      "cacheVariables": {
-        "CMAKE_HIP_FLAGS": "-parallel-jobs=4",
-        "AMDGPU_TARGETS": "gfx942;gfx950;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102;gfx1103;gfx1150;gfx1151;gfx1200;gfx1201;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-",
-        "OLLAMA_RUNNER_DIR": "rocm"
-      }
-    },
-    {
-      "name": "Vulkan",
-      "inherits": [ "Default" ],
-      "cacheVariables": {
-        "OLLAMA_RUNNER_DIR": "vulkan"
-      }
-    },
-    {
-      "name": "MLX",
-      "inherits": [ "Default" ],
-      "cacheVariables": {
-        "MLX_ENGINE": "ON",
-        "OLLAMA_RUNNER_DIR": "mlx"
-      }
-    },
-    {
-      "name": "MLX CUDA 12",
-      "inherits": [ "MLX", "CUDA 12" ],
-      "cacheVariables": {
-        "OLLAMA_RUNNER_DIR": "mlx_cuda_v12"
-      }
-    },
-    {
-      "name": "MLX CUDA 13",
-      "inherits": [ "MLX", "CUDA 13" ],
-      "cacheVariables": {
-        "MLX_CUDA_ARCHITECTURES": "86;89;90;90a;100;103;75-virtual;80-virtual;110-virtual;120-virtual;121-virtual",
-        "OLLAMA_RUNNER_DIR": "mlx_cuda_v13"
+        "OLLAMA_MLX_BACKENDS": "metal_v3;metal_v4"
       }
     }
   ],
@@ -124,74 +25,9 @@
       "configuration": "Release"
     },
     {
-      "name": "CPU",
-      "configurePreset": "Default",
-      "targets": [ "ggml-cpu" ]
-    },
-    {
-      "name": "CUDA",
-      "configurePreset": "CUDA",
-      "targets": [ "ggml-cuda" ]
-    },
-    {
-      "name": "CUDA 11",
-      "inherits": [ "CUDA" ],
-      "configurePreset": "CUDA 11"
-    },
-    {
-      "name": "CUDA 12",
-      "inherits": [ "CUDA" ],
-      "configurePreset": "CUDA 12"
-    },
-    {
-      "name": "CUDA 13",
-      "inherits": [ "CUDA" ],
-      "configurePreset": "CUDA 13"
-    },
-    {
-      "name": "JetPack 5",
-      "inherits": [ "CUDA" ],
-      "configurePreset": "JetPack 5"
-    },
-    {
-      "name": "JetPack 6",
-      "inherits": [ "CUDA" ],
-      "configurePreset": "JetPack 6"
-    },
-    {
-      "name": "ROCm",
-      "configurePreset": "ROCm",
-      "targets": [ "ggml-hip" ]
-    },
-    {
-      "name": "ROCm 6",
-      "inherits": [ "ROCm" ],
-      "configurePreset": "ROCm 6"
-    },
-    {
-      "name": "ROCm 7",
-      "inherits": [ "ROCm" ],
-      "configurePreset": "ROCm 7"
-    },
-    {
-      "name": "Vulkan",
-      "targets": [ "ggml-vulkan" ],
-      "configurePreset": "Vulkan"
-    },
-    {
-      "name": "MLX",
-      "targets": [ "mlx", "mlxc" ],
-      "configurePreset": "MLX"
-    },
-    {
-      "name": "MLX CUDA 12",
-      "targets": [ "mlx", "mlxc" ],
-      "configurePreset": "MLX CUDA 12"
-    },
-    {
-      "name": "MLX CUDA 13",
-      "targets": [ "mlx", "mlxc" ],
-      "configurePreset": "MLX CUDA 13"
+      "name": "MLX Metal",
+      "targets": [ "ollama-mlx-backends" ],
+      "configurePreset": "MLX Metal"
     }
   ]
 }

Dockerfile

@@ -37,113 +37,150 @@ RUN dnf install -y unzip \
 ENV CMAKE_GENERATOR=Ninja
 ENV LDFLAGS=-s
 
-FROM base AS cpu
+#
+# GPU toolchain stages — provide compilers for llama-server GPU builds
+#
+
+FROM base AS cpu-deps
 RUN dnf install -y gcc-toolset-11-gcc gcc-toolset-11-gcc-c++
 ENV PATH=/opt/rh/gcc-toolset-11/root/usr/bin:$PATH
-COPY CMakeLists.txt CMakePresets.json .
-COPY ml/backend/ggml/ggml ml/backend/ggml/ggml
-RUN --mount=type=cache,target=/root/.ccache \
-    cmake --preset 'CPU' \
-        && cmake --build --preset 'CPU' -- -l $(nproc) \
-        && cmake --install build --component CPU --strip
 
-FROM base AS cuda-11
-ARG CUDA11VERSION=11.8
-RUN dnf install -y cuda-toolkit-${CUDA11VERSION//./-}
-ENV PATH=/usr/local/cuda-11/bin:$PATH
-COPY CMakeLists.txt CMakePresets.json .
-COPY ml/backend/ggml/ggml ml/backend/ggml/ggml
-RUN --mount=type=cache,target=/root/.ccache \
-    cmake --preset 'CUDA 11' \
-        && cmake --build --preset 'CUDA 11' -- -l $(nproc) \
-        && cmake --install build --component CUDA --strip
-
-FROM base AS cuda-12
+FROM base AS cuda-12-deps
 ARG CUDA12VERSION=12.8
 RUN dnf install -y cuda-toolkit-${CUDA12VERSION//./-}
 ENV PATH=/usr/local/cuda-12/bin:$PATH
-COPY CMakeLists.txt CMakePresets.json .
-COPY ml/backend/ggml/ggml ml/backend/ggml/ggml
-RUN --mount=type=cache,target=/root/.ccache \
-    cmake --preset 'CUDA 12' \
-        && cmake --build --preset 'CUDA 12' -- -l $(nproc) \
-        && cmake --install build --component CUDA --strip
 
-
-FROM base AS cuda-13
+FROM base AS cuda-13-deps
 ARG CUDA13VERSION=13.0
 RUN dnf install -y cuda-toolkit-${CUDA13VERSION//./-}
 ENV PATH=/usr/local/cuda-13/bin:$PATH
-COPY CMakeLists.txt CMakePresets.json .
-COPY ml/backend/ggml/ggml ml/backend/ggml/ggml
-RUN --mount=type=cache,target=/root/.ccache \
-    cmake --preset 'CUDA 13' \
-        && cmake --build --preset 'CUDA 13' -- -l $(nproc) \
-        && cmake --install build --component CUDA --strip
 
+FROM base AS rocm-7-deps
+ENV PATH=/opt/rocm/llvm/bin:/opt/rocm/hcc/bin:/opt/rocm/hip/bin:/opt/rocm/bin:$PATH
 
-FROM base AS rocm-7
-ENV PATH=/opt/rocm/hcc/bin:/opt/rocm/hip/bin:/opt/rocm/bin:/opt/rocm/hcc/bin:$PATH
-COPY CMakeLists.txt CMakePresets.json .
-COPY ml/backend/ggml/ggml ml/backend/ggml/ggml
-RUN --mount=type=cache,target=/root/.ccache \
-    cmake --preset 'ROCm 7' \
-        && cmake --build --preset 'ROCm 7' -- -l $(nproc) \
-        && cmake --install build --component HIP --strip
-RUN rm -f dist/lib/ollama/rocm/rocblas/library/*gfx90[06]*
-
-FROM --platform=linux/arm64 nvcr.io/nvidia/l4t-jetpack:${JETPACK5VERSION} AS jetpack-5
-ARG CMAKEVERSION
-ARG NINJAVERSION
-RUN apt-get update && apt-get install -y curl ccache unzip \
-    && curl -fsSL https://github.com/Kitware/CMake/releases/download/v${CMAKEVERSION}/cmake-${CMAKEVERSION}-linux-$(uname -m).tar.gz | tar xz -C /usr/local --strip-components 1 \
-    && curl -fsSL -o /tmp/ninja.zip https://github.com/ninja-build/ninja/releases/download/v${NINJAVERSION}/ninja-linux-aarch64.zip \
-    && unzip /tmp/ninja.zip -d /usr/local/bin \
-    && rm /tmp/ninja.zip
-ENV CMAKE_GENERATOR=Ninja
-COPY CMakeLists.txt CMakePresets.json .
-COPY ml/backend/ggml/ggml ml/backend/ggml/ggml
-RUN --mount=type=cache,target=/root/.ccache \
-    cmake --preset 'JetPack 5' \
-        && cmake --build --preset 'JetPack 5' -- -l $(nproc) \
-        && cmake --install build --component CUDA --strip
-
-FROM --platform=linux/arm64 nvcr.io/nvidia/l4t-jetpack:${JETPACK6VERSION} AS jetpack-6
-ARG CMAKEVERSION
-ARG NINJAVERSION
-RUN apt-get update && apt-get install -y curl ccache unzip \
-    && curl -fsSL https://github.com/Kitware/CMake/releases/download/v${CMAKEVERSION}/cmake-${CMAKEVERSION}-linux-$(uname -m).tar.gz | tar xz -C /usr/local --strip-components 1 \
-    && curl -fsSL -o /tmp/ninja.zip https://github.com/ninja-build/ninja/releases/download/v${NINJAVERSION}/ninja-linux-aarch64.zip \
-    && unzip /tmp/ninja.zip -d /usr/local/bin \
-    && rm /tmp/ninja.zip
-ENV CMAKE_GENERATOR=Ninja
-COPY CMakeLists.txt CMakePresets.json .
-COPY ml/backend/ggml/ggml ml/backend/ggml/ggml
-RUN --mount=type=cache,target=/root/.ccache \
-    cmake --preset 'JetPack 6' \
-        && cmake --build --preset 'JetPack 6' -- -l $(nproc) \
-        && cmake --install build --component CUDA --strip
-
-FROM base AS vulkan
+FROM base AS vulkan-deps
 ARG VULKANVERSION
 RUN ln -s /usr/bin/python3 /usr/bin/python \
     && wget https://sdk.lunarg.com/sdk/download/${VULKANVERSION}/linux/vulkansdk-linux-x86_64-${VULKANVERSION}.tar.xz -O /tmp/vulkansdk.tar.xz \
     && tar xvf /tmp/vulkansdk.tar.xz -C /tmp \
     && /tmp/${VULKANVERSION}/vulkansdk -j 8 vulkan-headers \
+    && /tmp/${VULKANVERSION}/vulkansdk -j 8 spirv-headers \
     && /tmp/${VULKANVERSION}/vulkansdk -j 8 shaderc \
     && cp -r /tmp/${VULKANVERSION}/x86_64/include/* /usr/local/include/ \
     && cp -r /tmp/${VULKANVERSION}/x86_64/lib/* /usr/local/lib \
+    && cp -r /tmp/${VULKANVERSION}/x86_64/share/* /usr/local/share/ \
     && cp -r /tmp/${VULKANVERSION}/x86_64/bin/* /usr/local/bin/ \
     && rm -rf /tmp/${VULKANVERSION} /tmp/vulkansdk.tar.xz
-COPY CMakeLists.txt CMakePresets.json .
-COPY ml/backend/ggml/ggml ml/backend/ggml/ggml
+ENV VULKAN_SDK=/usr/local
+
+#
+# llama-server stages — rebuild when LLAMA_CPP_VERSION, llama/server/, or llama/compat/ changes.
+#
+# CPU stage: llama-server + ggml-base + ggml-cpu variants → lib/ollama/
+# GPU stages: GPU backend .so only → lib/ollama/<variant>/
+#
+
+FROM cpu-deps AS llama-server-cpu
+COPY LLAMA_CPP_VERSION .
+COPY llama/server llama/server
+COPY llama/compat llama/compat
 RUN --mount=type=cache,target=/root/.ccache \
-    cmake --preset 'Vulkan' \
-        && cmake --build --preset 'Vulkan' -- -l $(nproc) \
-        && cmake --install build --component Vulkan --strip
+    cmake -S llama/server --preset cpu \
+        && cmake --build build/llama-server-cpu -- -l $(nproc) \
+        && cmake --install build/llama-server-cpu --component llama-server --strip \
+        && for lib in \
+            /usr/lib64/libgomp.so* \
+            /usr/lib64/libomp.so* \
+            /opt/rh/gcc-toolset-11/root/usr/lib64/libgomp.so* \
+            /opt/rh/gcc-toolset-11/root/usr/lib64/libomp.so*; do \
+                [ -e "$lib" ] && cp -a "$lib" dist/lib/ollama/ || true; \
+            done
+
+FROM cuda-12-deps AS llama-server-cuda_v12
+COPY LLAMA_CPP_VERSION .
+COPY llama/server llama/server
+COPY llama/compat llama/compat
+RUN --mount=type=cache,target=/root/.ccache \
+    cmake -S llama/server --preset llama_cuda_v12_linux \
+        && cmake --build build/llama-server-cuda_v12 -- -l $(nproc) \
+        && cmake --install build/llama-server-cuda_v12 --component llama-server --strip
+
+FROM cuda-13-deps AS llama-server-cuda_v13
+COPY LLAMA_CPP_VERSION .
+COPY llama/server llama/server
+COPY llama/compat llama/compat
+RUN --mount=type=cache,target=/root/.ccache \
+    cmake -S llama/server --preset llama_cuda_v13_linux \
+        && cmake --build build/llama-server-cuda_v13 -- -l $(nproc) \
+        && cmake --install build/llama-server-cuda_v13 --component llama-server --strip
+
+FROM rocm-7-deps AS llama-server-rocm_v7_2
+ENV CC=clang CXX=clang++
+COPY LLAMA_CPP_VERSION .
+COPY llama/server llama/server
+COPY llama/compat llama/compat
+RUN --mount=type=cache,target=/root/.ccache \
+    cmake -S llama/server --preset rocm_v7_2_linux \
+        && cmake --build build/llama-server-rocm_v7_2 -- -l $(nproc) \
+        && cmake --install build/llama-server-rocm_v7_2 --component llama-server --strip
+RUN rm -f dist/lib/ollama/rocm_v7_2/rocblas/library/*gfx90[06]*
+
+FROM vulkan-deps AS llama-server-vulkan
+COPY LLAMA_CPP_VERSION .
+COPY llama/server llama/server
+COPY llama/compat llama/compat
+RUN --mount=type=cache,target=/root/.ccache \
+    cmake -S llama/server --preset vulkan \
+        && cmake --build build/llama-server-vulkan -- -l $(nproc) \
+        && cmake --install build/llama-server-vulkan --component llama-server --strip
+
+#
+# JetPack stages — self-contained with their own base images
+#
+
+FROM --platform=linux/arm64 nvcr.io/nvidia/l4t-jetpack:${JETPACK5VERSION} AS jetpack-5
+ARG CMAKEVERSION
+ARG NINJAVERSION
+RUN apt-get update && apt-get install -y curl ccache git unzip \
+    && curl -fsSL https://github.com/Kitware/CMake/releases/download/v${CMAKEVERSION}/cmake-${CMAKEVERSION}-linux-$(uname -m).tar.gz | tar xz -C /usr/local --strip-components 1 \
+    && curl -fsSL -o /tmp/ninja.zip https://github.com/ninja-build/ninja/releases/download/v${NINJAVERSION}/ninja-linux-aarch64.zip \
+    && unzip /tmp/ninja.zip -d /usr/local/bin \
+    && rm /tmp/ninja.zip
+ENV CMAKE_GENERATOR=Ninja
+COPY LLAMA_CPP_VERSION .
+COPY llama/server llama/server
+COPY llama/compat llama/compat
+RUN --mount=type=cache,target=/root/.ccache \
+    cmake -S llama/server --preset llama_cuda_jetpack5 \
+        && cmake --build build/llama-server-cuda_jetpack5 -- -l $(nproc) \
+        && cmake --install build/llama-server-cuda_jetpack5 --component llama-server --strip
+
+FROM --platform=linux/arm64 nvcr.io/nvidia/l4t-jetpack:${JETPACK6VERSION} AS jetpack-6
+ARG CMAKEVERSION
+ARG NINJAVERSION
+RUN apt-get update && apt-get install -y curl ccache git unzip \
+    && curl -fsSL https://github.com/Kitware/CMake/releases/download/v${CMAKEVERSION}/cmake-${CMAKEVERSION}-linux-$(uname -m).tar.gz | tar xz -C /usr/local --strip-components 1 \
+    && curl -fsSL -o /tmp/ninja.zip https://github.com/ninja-build/ninja/releases/download/v${NINJAVERSION}/ninja-linux-aarch64.zip \
+    && unzip /tmp/ninja.zip -d /usr/local/bin \
+    && rm /tmp/ninja.zip
+ENV CMAKE_GENERATOR=Ninja
+COPY LLAMA_CPP_VERSION .
+COPY llama/server llama/server
+COPY llama/compat llama/compat
+RUN --mount=type=cache,target=/root/.ccache \
+    cmake -S llama/server --preset llama_cuda_jetpack6 \
+        && cmake --build build/llama-server-cuda_jetpack6 -- -l $(nproc) \
+        && cmake --install build/llama-server-cuda_jetpack6 --component llama-server --strip
+
+#
+# MLX stage
+#
 
 FROM base AS mlx
 ARG CUDA13VERSION=13.0
+ARG OLLAMA_MLX_BUILD_JOBS=
+ARG OLLAMA_MLX_NVCC_THREADS=2
+ARG MLX_CUDA_RAM_MB=
 RUN dnf install -y cuda-toolkit-${CUDA13VERSION//./-} \
     && dnf install -y openblas-devel lapack-devel \
     && dnf install -y libcudnn9-cuda-13 libcudnn9-devel-cuda-13 \
@@ -154,7 +191,7 @@ ENV LAPACK_INCLUDE_DIRS=/usr/include/openblas
 ENV CGO_LDFLAGS="-L/usr/local/cuda-13/lib64 -L/usr/local/cuda-13/targets/x86_64-linux/lib/stubs"
 WORKDIR /go/src/github.com/ollama/ollama
 COPY CMakeLists.txt CMakePresets.json .
-COPY ml/backend/ggml/ggml ml/backend/ggml/ggml
+COPY cmake cmake
 COPY x/imagegen/mlx x/imagegen/mlx
 COPY go.mod go.sum .
 COPY MLX_VERSION MLX_C_VERSION .
@@ -170,9 +207,12 @@ RUN --mount=type=cache,target=/root/.ccache \
     && if [ -f /tmp/local-mlx-c/CMakeLists.txt ]; then \
         export OLLAMA_MLX_C_SOURCE=/tmp/local-mlx-c; \
     fi \
-    && cmake --preset 'MLX CUDA 13' -DBLAS_INCLUDE_DIRS=/usr/include/openblas -DLAPACK_INCLUDE_DIRS=/usr/include/openblas \
-        && cmake --build --preset 'MLX CUDA 13' -- -l $(nproc) \
-        && cmake --install build --component MLX --strip
+    && cmake -S . -B build/mlx_cuda_v13 -DOLLAMA_MLX_BACKENDS=cuda_v13 -DBLAS_INCLUDE_DIRS=/usr/include/openblas -DLAPACK_INCLUDE_DIRS=/usr/include/openblas -DCMAKE_CUDA_FLAGS="-t ${OLLAMA_MLX_NVCC_THREADS}" ${MLX_CUDA_RAM_MB:+-DMLX_CUDA_RAM_MB=${MLX_CUDA_RAM_MB}} -DOLLAMA_PAYLOAD_INSTALL_PREFIX=/go/src/github.com/ollama/ollama/dist \
+        && cmake --build build/mlx_cuda_v13 --target ollama-mlx-cuda_v13 -- -l $(nproc) ${OLLAMA_MLX_BUILD_JOBS:+-j ${OLLAMA_MLX_BUILD_JOBS}}
+
+#
+# Go build
+#
 
 FROM base AS build
 WORKDIR /go/src/github.com/ollama/ollama
@@ -190,38 +230,59 @@ ENV CGO_CXXFLAGS="${CGO_CXXFLAGS}"
 RUN --mount=type=cache,target=/root/.cache/go-build \
     go build -trimpath -buildmode=pie -o /bin/ollama .
 
+#
+# Assembly stages — combine llama-server variants + GPU runtime libs
+#
+
 FROM --platform=linux/amd64 scratch AS amd64
-# COPY --from=cuda-11 dist/lib/ollama/ /lib/ollama/
-COPY --from=cuda-12 dist/lib/ollama /lib/ollama/
-COPY --from=cuda-13 dist/lib/ollama /lib/ollama/
-COPY --from=vulkan  dist/lib/ollama  /lib/ollama/
+COPY --from=llama-server-cpu      dist/lib/ollama /lib/ollama/
+COPY --from=llama-server-cuda_v12 dist/lib/ollama /lib/ollama/
+COPY --from=llama-server-cuda_v13 dist/lib/ollama /lib/ollama/
+COPY --from=llama-server-vulkan   dist/lib/ollama /lib/ollama/
 COPY --from=mlx     /go/src/github.com/ollama/ollama/dist/lib/ollama /lib/ollama/
 
 FROM --platform=linux/arm64 scratch AS arm64
-# COPY --from=cuda-11 dist/lib/ollama/ /lib/ollama/
-COPY --from=cuda-12 dist/lib/ollama /lib/ollama/
-COPY --from=cuda-13 dist/lib/ollama/ /lib/ollama/
+COPY --from=llama-server-cpu dist/lib/ollama /lib/ollama/
+COPY --from=llama-server-cuda_v12 dist/lib/ollama /lib/ollama/
+COPY --from=llama-server-cuda_v13 dist/lib/ollama /lib/ollama/
 COPY --from=jetpack-5 dist/lib/ollama/ /lib/ollama/
 COPY --from=jetpack-6 dist/lib/ollama/ /lib/ollama/
 
 FROM scratch AS rocm
-COPY --from=rocm-7 dist/lib/ollama /lib/ollama
+COPY --from=llama-server-cpu  dist/lib/ollama /lib/ollama
+COPY --from=llama-server-rocm_v7_2 dist/lib/ollama /lib/ollama
 
-FROM ${FLAVOR} AS archive
-COPY --from=cpu dist/lib/ollama /lib/ollama
+FROM --platform=linux/amd64 scratch AS amd64-archive
+COPY --from=amd64 /lib/ollama /lib/ollama/
+COPY --from=llama-server-rocm_v7_2 dist/lib/ollama /lib/ollama/
+
+FROM --platform=linux/arm64 scratch AS arm64-archive
+COPY --from=arm64 /lib/ollama /lib/ollama/
+
+FROM ${TARGETARCH}-archive AS archive
+COPY --from=build /bin/ollama /bin/ollama
+
+FROM ${FLAVOR} AS image-archive
 COPY --from=build /bin/ollama /bin/ollama
 
 FROM ubuntu:24.04
 ARG APT_MIRROR=http://archive.ubuntu.com/ubuntu
-RUN sed -i "s|http://archive.ubuntu.com/ubuntu|$APT_MIRROR|g" /etc/apt/sources.list.d/ubuntu.sources \
+ARG APT_PORTS_MIRROR=http://ports.ubuntu.com/ubuntu-ports
+RUN sed -i \
+        -e "s|http://archive.ubuntu.com/ubuntu|$APT_MIRROR|g" \
+        -e "s|http://ports.ubuntu.com/ubuntu-ports|$APT_PORTS_MIRROR|g" \
+        /etc/apt/sources.list.d/ubuntu.sources \
     && apt-get update \
     && apt-get install -y ca-certificates libvulkan1 libopenblas0 \
-    && sed -i "s|$APT_MIRROR|http://archive.ubuntu.com/ubuntu|g" /etc/apt/sources.list.d/ubuntu.sources \
+    && sed -i \
+        -e "s|$APT_MIRROR|http://archive.ubuntu.com/ubuntu|g" \
+        -e "s|$APT_PORTS_MIRROR|http://ports.ubuntu.com/ubuntu-ports|g" \
+        /etc/apt/sources.list.d/ubuntu.sources \
     && apt-get clean \
     && rm -rf /var/lib/apt/lists/*
-COPY --from=archive /bin /usr/bin
+COPY --from=image-archive /bin /usr/bin
 ENV PATH=/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin
-COPY --from=archive /lib/ollama /usr/lib/ollama
+COPY --from=image-archive /lib/ollama /usr/lib/ollama
 ENV LD_LIBRARY_PATH=/usr/local/nvidia/lib:/usr/local/nvidia/lib64
 ENV NVIDIA_DRIVER_CAPABILITIES=compute,utility
 ENV NVIDIA_VISIBLE_DEVICES=all

LLAMA_CPP_VERSION

@@ -0,0 +1 @@
+b9409

MLX_VERSION

@@ -1 +1 @@
-e8ebdebeeb655feaa85a51f6b24ece5b6d5518d1
+2165dc08d7b33258260aa849d39f087d50e62962

Makefile.sync

@@ -1,76 +0,0 @@
-UPSTREAM=https://github.com/ggml-org/llama.cpp.git
-WORKDIR=llama/vendor
-FETCH_HEAD=ec98e2002
-
-.PHONY: help
-help:
-	@echo "Available targets:"
-	@echo "    sync                 Sync with upstream repositories"
-	@echo "    checkout             Checkout upstream repository"
-	@echo "    apply-patches        Apply patches to local repository"
-	@echo "    format-patches       Format patches from local repository"
-	@echo "    clean                Clean local repository"
-	@echo
-	@echo "Example:"
-	@echo "    make -f $(lastword $(MAKEFILE_LIST)) clean apply-patches sync"
-
-.PHONY: sync
-sync: llama/build-info.cpp ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-embed.metal
-
-llama/build-info.cpp: llama/build-info.cpp.in llama/llama.cpp
-	sed -e 's|@FETCH_HEAD@|$(FETCH_HEAD)|' <$< >$@
-
-ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-embed.metal: ml/backend/ggml/ggml
-	go generate ./$(@D)
-
-.PHONY: llama/llama.cpp
-llama/llama.cpp: llama/vendor
-	rsync -arvzc --delete -f "include LICENSE" -f "merge $@/.rsync-filter" $(addprefix $<,/LICENSE /) $@
-
-.PHONY: ml/backend/ggml/ggml
-ml/backend/ggml/ggml: llama/vendor
-	rsync -arvzc --delete -f "include LICENSE" -f "merge $@/.rsync-filter" $(addprefix $<,/LICENSE /ggml/) $@
-
-PATCHES=$(wildcard llama/patches/*.patch)
-PATCHED=$(join $(dir $(PATCHES)), $(addsuffix ed, $(addprefix ., $(notdir $(PATCHES)))))
-
-.PHONY: apply-patches
-.NOTPARALLEL:
-apply-patches: $(PATCHED)
-
-llama/patches/.%.patched: llama/patches/%.patch
-	@if git -c user.name=nobody -c 'user.email=<>' -C $(WORKDIR) am -3 $(realpath $<); then \
-		touch $@;                                                                           \
-	else                                                                                    \
-		echo "Patch failed. Resolve any conflicts then continue.";                          \
-		echo "1. Run 'git -C $(WORKDIR) am --continue'";                                    \
-		echo "2. Run 'make -f $(lastword $(MAKEFILE_LIST)) format-patches'";                \
-		echo "3. Run 'make -f $(lastword $(MAKEFILE_LIST)) clean apply-patches'";           \
-		exit 1;                                                                             \
-	fi
-
-.PHONY: checkout
-checkout: $(WORKDIR)
-	git -C $(WORKDIR) fetch
-	git -C $(WORKDIR) checkout -f $(FETCH_HEAD)
-
-$(WORKDIR):
-	git clone $(UPSTREAM) $(WORKDIR)
-
-.PHONY: format-patches
-format-patches: llama/patches
-	git -C $(WORKDIR) format-patch \
-		--no-signature \
-		--no-numbered \
-		--zero-commit \
-		-o $(realpath $<) \
-		$(FETCH_HEAD)
-
-.PHONY: clean
-clean: checkout
-	@git -C $(WORKDIR) am --abort || true
-	$(RM) llama/patches/.*.patched
-
-.PHONY: print-base
-print-base:
-	@echo $(FETCH_HEAD)

anthropic/anthropic.go

@@ -78,6 +78,11 @@ type MessagesRequest struct {
 	ToolChoice    *ToolChoice     `json:"tool_choice,omitempty"`
 	Thinking      *ThinkingConfig `json:"thinking,omitempty"`
 	Metadata      *Metadata       `json:"metadata,omitempty"`
+	OutputConfig  *OutputConfig   `json:"output_config,omitempty"`
+}
+
+type OutputConfig struct {
+	Effort string `json:"effort,omitempty"`
 }
 
 // MessageParam represents a message in the request
@@ -161,7 +166,7 @@ type WebSearchToolResultError struct {
 
 // ImageSource represents the source of an image
 type ImageSource struct {
-	Type      string `json:"type"` // "base64" or "url"
+	Type      string `json:"type"` // "base64"
 	MediaType string `json:"media_type,omitempty"`
 	Data      string `json:"data,omitempty"`
 	URL       string `json:"url,omitempty"`
@@ -373,9 +378,26 @@ func FromMessagesRequest(r MessagesRequest) (*api.ChatRequest, error) {
 	}
 
 	var think *api.ThinkValue
+	normalizedEffort := ""
+	if r.OutputConfig != nil {
+		normalizedEffort = strings.ToLower(strings.TrimSpace(r.OutputConfig.Effort))
+		if normalizedEffort == "xhigh" {
+			normalizedEffort = "high"
+		}
+	}
+
 	if r.Thinking != nil && r.Thinking.Type == "enabled" {
 		think = &api.ThinkValue{Value: true}
 	}
+	if r.Thinking != nil && r.Thinking.Type == "disabled" {
+		think = &api.ThinkValue{Value: false}
+	}
+	if think == nil && r.OutputConfig != nil {
+		switch normalizedEffort {
+		case "high", "medium", "low", "max":
+			think = &api.ThinkValue{Value: normalizedEffort}
+		}
+	}
 
 	stream := r.Stream
 	convertedRequest := &api.ChatRequest{
@@ -425,17 +447,12 @@ func convertMessage(msg MessageParam) ([]api.Message, error) {
 				return nil, errors.New("invalid image source")
 			}
 
-			if block.Source.Type == "base64" {
-				decoded, err := base64.StdEncoding.DecodeString(block.Source.Data)
-				if err != nil {
-					logutil.Trace("anthropic: invalid base64 image data", "role", role, "error", err)
-					return nil, fmt.Errorf("invalid base64 image data: %w", err)
-				}
-				images = append(images, decoded)
-			} else {
-				logutil.Trace("anthropic: unsupported image source type", "role", role, "source_type", block.Source.Type)
-				return nil, fmt.Errorf("invalid image source type: %s. Only base64 images are supported.", block.Source.Type)
+			decoded, err := resolveImageSource(block.Source)
+			if err != nil {
+				logutil.Trace("anthropic: unsupported image source", "role", role, "source_type", block.Source.Type, "error", err)
+				return nil, err
 			}
+			images = append(images, decoded)
 
 		case "tool_use":
 			toolUseBlocks++
@@ -457,26 +474,16 @@ func convertMessage(msg MessageParam) ([]api.Message, error) {
 
 		case "tool_result":
 			toolResultBlocks++
-			var resultContent string
-
-			switch c := block.Content.(type) {
-			case string:
-				resultContent = c
-			case []any:
-				for _, cb := range c {
-					if cbMap, ok := cb.(map[string]any); ok {
-						if cbMap["type"] == "text" {
-							if text, ok := cbMap["text"].(string); ok {
-								resultContent += text
-							}
-						}
-					}
-				}
+			resultContent, resultImages, err := convertToolResultContent(block.Content)
+			if err != nil {
+				logutil.Trace("anthropic: invalid tool_result content", "role", role, "error", err)
+				return nil, err
 			}
 
 			toolResults = append(toolResults, api.Message{
 				Role:       "tool",
 				Content:    resultContent,
+				Images:     resultImages,
 				ToolCallID: block.ToolUseID,
 			})
 
@@ -508,6 +515,10 @@ func convertMessage(msg MessageParam) ([]api.Message, error) {
 		}
 	}
 
+	if role == "user" && len(toolResults) > 0 {
+		messages = append(messages, toolResults...)
+	}
+
 	if textContent.Len() > 0 || len(images) > 0 || len(toolCalls) > 0 || thinking != "" {
 		m := api.Message{
 			Role:      role,
@@ -519,8 +530,10 @@ func convertMessage(msg MessageParam) ([]api.Message, error) {
 		messages = append(messages, m)
 	}
 
-	// Add tool results as separate messages
-	messages = append(messages, toolResults...)
+	// Add tool results as separate messages.
+	if role != "user" || len(toolResults) == 0 {
+		messages = append(messages, toolResults...)
+	}
 	logutil.Trace("anthropic: converted block message",
 		"role", role,
 		"blocks", len(msg.Content),
@@ -969,6 +982,71 @@ func GenerateMessageID() string {
 	return generateID("msg")
 }
 
+func resolveImageSource(source *ImageSource) (api.ImageData, error) {
+	if source.Type != "base64" {
+		return nil, fmt.Errorf("invalid image source type: %s. Only base64 images are supported.", source.Type)
+	}
+
+	decoded, err := base64.StdEncoding.DecodeString(source.Data)
+	if err != nil {
+		return nil, fmt.Errorf("invalid base64 image data: %w", err)
+	}
+
+	return decoded, nil
+}
+
+func convertToolResultContent(content any) (string, []api.ImageData, error) {
+	switch c := content.(type) {
+	case nil:
+		return "", nil, nil
+	case string:
+		return c, nil, nil
+	case []any:
+		var text strings.Builder
+		var images []api.ImageData
+
+		for _, cb := range c {
+			cbMap, ok := cb.(map[string]any)
+			if !ok {
+				continue
+			}
+
+			switch cbMap["type"] {
+			case "text":
+				if t, ok := cbMap["text"].(string); ok {
+					text.WriteString(t)
+				}
+			case "image":
+				rawSource, ok := cbMap["source"].(map[string]any)
+				if !ok {
+					return "", nil, errors.New("invalid tool_result image source")
+				}
+
+				var source ImageSource
+				if rawType, ok := rawSource["type"].(string); ok {
+					source.Type = rawType
+				}
+				if rawMediaType, ok := rawSource["media_type"].(string); ok {
+					source.MediaType = rawMediaType
+				}
+				if rawData, ok := rawSource["data"].(string); ok {
+					source.Data = rawData
+				}
+
+				img, err := resolveImageSource(&source)
+				if err != nil {
+					return "", nil, err
+				}
+				images = append(images, img)
+			}
+		}
+
+		return text.String(), images, nil
+	default:
+		return "", nil, nil
+	}
+}
+
 // ptr returns a pointer to the given string value
 func ptr(s string) *string {
 	return &s

anthropic/anthropic_test.go

@@ -271,6 +271,241 @@ func TestFromMessagesRequest_WithToolResult(t *testing.T) {
 	}
 }
 
+func TestFromMessagesRequest_WithToolResultImage(t *testing.T) {
+	imgData, _ := base64.StdEncoding.DecodeString(testImage)
+
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		Messages: []MessageParam{
+			{
+				Role: "user",
+				Content: []ContentBlock{
+					{
+						Type:      "tool_result",
+						ToolUseID: "call_img",
+						Content: []any{
+							map[string]any{"type": "text", "text": "Attached image"},
+							map[string]any{
+								"type": "image",
+								"source": map[string]any{
+									"type":       "base64",
+									"media_type": "image/png",
+									"data":       testImage,
+								},
+							},
+						},
+					},
+				},
+			},
+		},
+	}
+
+	result, err := FromMessagesRequest(req)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if len(result.Messages) != 1 {
+		t.Fatalf("expected 1 message, got %d", len(result.Messages))
+	}
+
+	msg := result.Messages[0]
+	if msg.Role != "tool" {
+		t.Errorf("expected role 'tool', got %q", msg.Role)
+	}
+	if msg.ToolCallID != "call_img" {
+		t.Errorf("expected tool_call_id 'call_img', got %q", msg.ToolCallID)
+	}
+	if msg.Content != "Attached image" {
+		t.Errorf("unexpected content: %q", msg.Content)
+	}
+	if len(msg.Images) != 1 {
+		t.Fatalf("expected 1 image, got %d", len(msg.Images))
+	}
+	if string(msg.Images[0]) != string(imgData) {
+		t.Error("image data mismatch")
+	}
+}
+
+func TestFromMessagesRequest_WithToolResultFollowedByUserText(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		Messages: []MessageParam{
+			{
+				Role: "assistant",
+				Content: []ContentBlock{
+					{
+						Type:  "tool_use",
+						ID:    "call_read",
+						Name:  "Read",
+						Input: makeArgs("file_path", "/Users/hoyyeva/Desktop/aaa.png"),
+					},
+				},
+			},
+			{
+				Role: "user",
+				Content: []ContentBlock{
+					{
+						Type:      "tool_result",
+						ToolUseID: "call_read",
+						Content:   "Read image (311.5KB)",
+					},
+					{
+						Type: "text",
+						Text: ptr("Please describe it."),
+					},
+				},
+			},
+		},
+	}
+
+	result, err := FromMessagesRequest(req)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if len(result.Messages) != 3 {
+		t.Fatalf("expected 3 messages, got %d", len(result.Messages))
+	}
+
+	if result.Messages[1].Role != "tool" {
+		t.Fatalf("expected second message to be tool, got %q", result.Messages[1].Role)
+	}
+	if result.Messages[1].ToolCallID != "call_read" {
+		t.Fatalf("expected tool_call_id 'call_read', got %q", result.Messages[1].ToolCallID)
+	}
+	if result.Messages[2].Role != "user" {
+		t.Fatalf("expected third message to be user, got %q", result.Messages[2].Role)
+	}
+	if result.Messages[2].Content != "Please describe it." {
+		t.Fatalf("unexpected user content: %q", result.Messages[2].Content)
+	}
+}
+
+func TestFromMessagesRequest_WithOutputConfigEffort(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "gemma4",
+		MaxTokens: 32000,
+		Messages: []MessageParam{
+			{
+				Role:    "user",
+				Content: textContent("Describe the image."),
+			},
+		},
+		OutputConfig: &OutputConfig{
+			Effort: "high",
+		},
+	}
+
+	result, err := FromMessagesRequest(req)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if result.Think == nil {
+		t.Fatal("expected think to be set from output_config.effort")
+	}
+
+	if got := result.Think.String(); got != "high" {
+		t.Fatalf("expected think level 'high', got %q", got)
+	}
+}
+
+func TestFromMessagesRequest_WithOutputConfigEffortXHighMapsToHigh(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "gemma4",
+		MaxTokens: 32000,
+		Messages: []MessageParam{
+			{
+				Role:    "user",
+				Content: textContent("Describe the image."),
+			},
+		},
+		OutputConfig: &OutputConfig{
+			Effort: "xhigh",
+		},
+	}
+
+	result, err := FromMessagesRequest(req)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if result.Think == nil {
+		t.Fatal("expected think to be set from output_config.effort")
+	}
+
+	if got := result.Think.String(); got != "high" {
+		t.Fatalf("expected think level 'high' for xhigh effort, got %q", got)
+	}
+}
+
+func TestFromMessagesRequest_ThinkingDisabledOverridesOutputConfigEffort(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "gemma4",
+		MaxTokens: 32000,
+		Messages: []MessageParam{
+			{
+				Role:    "user",
+				Content: textContent("Describe the image."),
+			},
+		},
+		Thinking: &ThinkingConfig{
+			Type: "disabled",
+		},
+		OutputConfig: &OutputConfig{
+			Effort: "high",
+		},
+	}
+
+	result, err := FromMessagesRequest(req)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if result.Think == nil {
+		t.Fatal("expected think to be set")
+	}
+
+	if got := result.Think.Value; got != false {
+		t.Fatalf("expected think=false when thinking is disabled, got %v", got)
+	}
+}
+
+func TestFromMessagesRequest_ThinkingAdaptiveUsesOutputConfigEffort(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "gemma4",
+		MaxTokens: 32000,
+		Messages: []MessageParam{
+			{
+				Role:    "user",
+				Content: textContent("Describe the image."),
+			},
+		},
+		Thinking: &ThinkingConfig{
+			Type: "adaptive",
+		},
+		OutputConfig: &OutputConfig{
+			Effort: "high",
+		},
+	}
+
+	result, err := FromMessagesRequest(req)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if result.Think == nil {
+		t.Fatal("expected think to be set from output_config.effort")
+	}
+
+	if got := result.Think.String(); got != "high" {
+		t.Fatalf("expected think level 'high' for adaptive thinking, got %q", got)
+	}
+}
+
 func TestFromMessagesRequest_WithTools(t *testing.T) {
 	req := MessagesRequest{
 		Model:     "test-model",

api/client.go

@@ -259,6 +259,10 @@ func (c *Client) stream(ctx context.Context, method, path string, data any, fn f
 		}
 	}
 
+	if err := scanner.Err(); err != nil {
+		return err
+	}
+
 	return nil
 }
 

api/client_test.go

@@ -3,6 +3,7 @@ package api
 import (
 	"encoding/json"
 	"fmt"
+	"io"
 	"net/http"
 	"net/http/httptest"
 	"net/url"
@@ -192,6 +193,35 @@ func TestClientStream(t *testing.T) {
 	}
 }
 
+func TestClientStreamReportsReadErrors(t *testing.T) {
+	client := NewClient(
+		&url.URL{Scheme: "http", Host: "example.com"},
+		&http.Client{Transport: roundTripFunc(func(*http.Request) (*http.Response, error) {
+			body := failingReader{
+				data: []byte(`{"message":{"content":"partial"}}` + "\n"),
+				err:  io.ErrUnexpectedEOF,
+			}
+
+			return &http.Response{
+				StatusCode: http.StatusOK,
+				Status:     "200 OK",
+				Body:       io.NopCloser(&body),
+				Header:     make(http.Header),
+			}, nil
+		})},
+	)
+
+	err := client.stream(t.Context(), http.MethodPost, "/api/chat", nil, func([]byte) error {
+		return nil
+	})
+	if err == nil {
+		t.Fatal("expected stream read error")
+	}
+	if !strings.Contains(err.Error(), io.ErrUnexpectedEOF.Error()) {
+		t.Fatalf("expected unexpected EOF, got %v", err)
+	}
+}
+
 func TestClientDo(t *testing.T) {
 	testCases := []struct {
 		name           string
@@ -320,3 +350,23 @@ func TestClientDo(t *testing.T) {
 		})
 	}
 }
+
+type roundTripFunc func(*http.Request) (*http.Response, error)
+
+func (f roundTripFunc) RoundTrip(req *http.Request) (*http.Response, error) {
+	return f(req)
+}
+
+type failingReader struct {
+	data []byte
+	err  error
+}
+
+func (r *failingReader) Read(p []byte) (int, error) {
+	if len(r.data) > 0 {
+		n := copy(p, r.data)
+		r.data = r.data[n:]
+		return n, nil
+	}
+	return 0, r.err
+}

api/types.go

@@ -600,12 +600,13 @@ type Options struct {
 
 // Runner options which must be set when the model is loaded into memory
 type Runner struct {
-	NumCtx    int   `json:"num_ctx,omitempty"`
-	NumBatch  int   `json:"num_batch,omitempty"`
-	NumGPU    int   `json:"num_gpu,omitempty"`
-	MainGPU   int   `json:"main_gpu,omitempty"`
-	UseMMap   *bool `json:"use_mmap,omitempty"`
-	NumThread int   `json:"num_thread,omitempty"`
+	NumCtx          int   `json:"num_ctx,omitempty"`
+	NumBatch        int   `json:"num_batch,omitempty"`
+	NumGPU          int   `json:"num_gpu,omitempty"`
+	MainGPU         *int  `json:"main_gpu,omitempty"`
+	UseMMap         *bool `json:"use_mmap,omitempty"`
+	NumThread       int   `json:"num_thread,omitempty"`
+	DraftNumPredict int   `json:"draft_num_predict,omitempty"`
 }
 
 // EmbedRequest is the request passed to [Client.Embed].
@@ -672,6 +673,9 @@ type CreateRequest struct {
 	// Quantize is the quantization format for the model; leave blank to not change the quantization level.
 	Quantize string `json:"quantize,omitempty"`
 
+	// DraftQuantize is the quantization format for the draft model.
+	DraftQuantize string `json:"draft_quantize,omitempty"`
+
 	// From is the name of the model or file to use as the source.
 	From string `json:"from,omitempty"`
 
@@ -681,6 +685,9 @@ type CreateRequest struct {
 	// Files is a map of files include when creating the model.
 	Files map[string]string `json:"files,omitempty"`
 
+	// DraftFiles is a map of draft model files to include when creating the model.
+	DraftFiles map[string]string `json:"draft_files,omitempty"`
+
 	// Adapters is a map of LoRA adapters to include when creating the model.
 	Adapters map[string]string `json:"adapters,omitempty"`
 
@@ -824,14 +831,15 @@ type ProcessResponse struct {
 
 // ListModelResponse is a single model description in [ListResponse].
 type ListModelResponse struct {
-	Name        string       `json:"name"`
-	Model       string       `json:"model"`
-	RemoteModel string       `json:"remote_model,omitempty"`
-	RemoteHost  string       `json:"remote_host,omitempty"`
-	ModifiedAt  time.Time    `json:"modified_at"`
-	Size        int64        `json:"size"`
-	Digest      string       `json:"digest"`
-	Details     ModelDetails `json:"details,omitempty"`
+	Name         string             `json:"name"`
+	Model        string             `json:"model"`
+	RemoteModel  string             `json:"remote_model,omitempty"`
+	RemoteHost   string             `json:"remote_host,omitempty"`
+	ModifiedAt   time.Time          `json:"modified_at"`
+	Size         int64              `json:"size"`
+	Digest       string             `json:"digest"`
+	Details      ModelDetails       `json:"details,omitempty"`
+	Capabilities []model.Capability `json:"capabilities,omitempty"`
 }
 
 // ProcessModelResponse is a single model description in [ProcessResponse].
@@ -924,6 +932,8 @@ type ModelDetails struct {
 	Families          []string `json:"families"`
 	ParameterSize     string   `json:"parameter_size"`
 	QuantizationLevel string   `json:"quantization_level"`
+	ContextLength     int      `json:"context_length,omitempty"`
+	EmbeddingLength   int      `json:"embedding_length,omitempty"`
 }
 
 // UserResponse provides information about a user.
@@ -1046,14 +1056,25 @@ func (opts *Options) FromMap(m map[string]any) error {
 				}
 				field.Set(reflect.ValueOf(slice))
 			case reflect.Pointer:
-				var b bool
-				if field.Type() == reflect.TypeOf(&b) {
+				switch field.Type().Elem().Kind() {
+				case reflect.Bool:
 					val, ok := val.(bool)
 					if !ok {
 						return fmt.Errorf("option %q must be of type boolean", key)
 					}
 					field.Set(reflect.ValueOf(&val))
-				} else {
+				case reflect.Int:
+					var i int
+					switch t := val.(type) {
+					case int64:
+						i = int(t)
+					case float64:
+						i = int(t)
+					default:
+						return fmt.Errorf("option %q must be of type integer", key)
+					}
+					field.Set(reflect.ValueOf(&i))
+				default:
 					return fmt.Errorf("unknown type loading config params: %v %v", field.Kind(), field.Type())
 				}
 			default:
@@ -1086,11 +1107,12 @@ func DefaultOptions() Options {
 
 		Runner: Runner{
 			// options set when the model is loaded
-			NumCtx:    int(envconfig.ContextLength()),
-			NumBatch:  512,
-			NumGPU:    -1, // -1 here indicates that NumGPU should be set dynamically
-			NumThread: 0,  // let the runtime decide
-			UseMMap:   nil,
+			NumCtx:          int(envconfig.ContextLength()),
+			NumBatch:        512,
+			NumGPU:          -1, // -1 here indicates that NumGPU should be set dynamically
+			NumThread:       0,  // let the runtime decide
+			DraftNumPredict: 4,
+			UseMMap:         nil,
 		},
 	}
 }
@@ -1294,14 +1316,20 @@ func FormatParams(params map[string][]string) (map[string]any, error) {
 					// TODO: only string slices are supported right now
 					out[key] = vals
 				case reflect.Pointer:
-					var b bool
-					if field.Type() == reflect.TypeOf(&b) {
+					switch field.Type().Elem().Kind() {
+					case reflect.Bool:
 						boolVal, err := strconv.ParseBool(vals[0])
 						if err != nil {
 							return nil, fmt.Errorf("invalid bool value %s", vals)
 						}
 						out[key] = &boolVal
-					} else {
+					case reflect.Int:
+						intVal, err := strconv.ParseInt(vals[0], 10, 64)
+						if err != nil {
+							return nil, fmt.Errorf("invalid int value %s", vals)
+						}
+						out[key] = intVal
+					default:
 						return nil, fmt.Errorf("unknown type %s for %s", field.Kind(), key)
 					}
 				default:

api/types_test.go

@@ -20,6 +20,10 @@ func testPropsMap(m map[string]ToolProperty) *ToolPropertiesMap {
 	return props
 }
 
+func testIntPtr(v int) *int {
+	return &v
+}
+
 // testArgs creates ToolCallFunctionArguments from a map (convenience function for tests, order not preserved)
 func testArgs(m map[string]any) ToolCallFunctionArguments {
 	args := NewToolCallFunctionArguments()
@@ -168,6 +172,47 @@ func TestUseMmapParsingFromJSON(t *testing.T) {
 	}
 }
 
+func TestMainGPUParsingFromJSON(t *testing.T) {
+	tests := []struct {
+		name    string
+		req     string
+		wantGPU *int
+	}{
+		{
+			name: "Undefined",
+			req:  `{}`,
+		},
+		{
+			name:    "Zero",
+			req:     `{ "main_gpu": 0 }`,
+			wantGPU: testIntPtr(0),
+		},
+		{
+			name:    "Nonzero",
+			req:     `{ "main_gpu": 1 }`,
+			wantGPU: testIntPtr(1),
+		},
+	}
+
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			var oMap map[string]any
+			err := json.Unmarshal([]byte(test.req), &oMap)
+			require.NoError(t, err)
+
+			opts := DefaultOptions()
+			err = opts.FromMap(oMap)
+			require.NoError(t, err)
+
+			if test.wantGPU == nil {
+				assert.Nil(t, opts.MainGPU)
+			} else if assert.NotNil(t, opts.MainGPU) {
+				assert.Equal(t, *test.wantGPU, *opts.MainGPU)
+			}
+		})
+	}
+}
+
 func TestUseMmapFormatParams(t *testing.T) {
 	tr := true
 	fa := false
@@ -232,6 +277,12 @@ func TestUseMmapFormatParams(t *testing.T) {
 	}
 }
 
+func TestMainGPUFormatParams(t *testing.T) {
+	resp, err := FormatParams(map[string][]string{"main_gpu": {"0"}})
+	require.NoError(t, err)
+	assert.Equal(t, int64(0), resp["main_gpu"])
+}
+
 func TestMessage_UnmarshalJSON(t *testing.T) {
 	tests := []struct {
 		input    string

app/ollama.iss

@@ -90,9 +90,8 @@ DialogFontSize=12
 [Files]
 #if FileExists("..\dist\windows-ollama-app-amd64.exe")
 Source: "..\dist\windows-ollama-app-amd64.exe"; DestDir: "{app}"; DestName: "{#MyAppExeName}" ;Check: not IsArm64();  Flags: ignoreversion 64bit; BeforeInstall: TaskKill('{#MyAppExeName}')
-Source: "..\dist\windows-amd64\vc_redist.x64.exe"; DestDir: "{tmp}"; Check: not IsArm64() and vc_redist_needed(); Flags: deleteafterinstall
 Source: "..\dist\windows-amd64\ollama.exe"; DestDir: "{app}"; Check: not IsArm64(); Flags: ignoreversion 64bit; BeforeInstall: TaskKill('ollama.exe')
-Source: "..\dist\windows-amd64\lib\ollama\*"; DestDir: "{app}\lib\ollama\"; Check: not IsArm64(); Flags: ignoreversion 64bit recursesubdirs
+Source: "..\dist\windows-amd64\lib\ollama\*"; Excludes: "\mlx_*\*"; DestDir: "{app}\lib\ollama\"; Check: not IsArm64(); Flags: ignoreversion 64bit recursesubdirs
 #endif
 
 ; For local development, rely on binary compatibility at runtime since we can't cross compile
@@ -103,9 +102,11 @@ Source: "..\dist\windows-ollama-app-amd64.exe"; DestDir: "{app}"; DestName: "{#M
 #endif
 
 #if FileExists("..\dist\windows-arm64\ollama.exe")
-Source: "..\dist\windows-arm64\vc_redist.arm64.exe"; DestDir: "{tmp}"; Check: IsArm64() and vc_redist_needed(); Flags: deleteafterinstall
 Source: "..\dist\windows-arm64\ollama.exe"; DestDir: "{app}"; Check: IsArm64(); Flags: ignoreversion 64bit; BeforeInstall: TaskKill('ollama.exe')
 #endif
+#if DirExists("..\dist\windows-arm64\lib\ollama")
+Source: "..\dist\windows-arm64\lib\ollama\*"; DestDir: "{app}\lib\ollama\"; Check: IsArm64(); Flags: ignoreversion 64bit recursesubdirs
+#endif
 
 Source: ".\assets\app.ico"; DestDir: "{app}"; Flags: ignoreversion
 
@@ -118,12 +119,6 @@ Name: "{userprograms}\{#MyAppName}"; Filename: "{app}\{#MyAppExeName}"; IconFile
 Type: files; Name: "{%LOCALAPPDATA}\Ollama\updates"
 
 [Run]
-#if DirExists("..\dist\windows-arm64")
-Filename: "{tmp}\vc_redist.arm64.exe"; Parameters: "/install /passive /norestart"; Check: IsArm64() and vc_redist_needed(); StatusMsg: "Installing VC++ Redistributables..."; Flags: waituntilterminated
-#endif
-#if DirExists("..\dist\windows-amd64")
-Filename: "{tmp}\vc_redist.x64.exe"; Parameters: "/install /passive /norestart"; Check: not IsArm64() and vc_redist_needed(); StatusMsg: "Installing VC++ Redistributables..."; Flags: waituntilterminated
-#endif
 Filename: "{cmd}"; Parameters: "/C set PATH={app};%PATH% & ""{app}\{#MyAppExeName}"""; Flags: postinstall nowait runhidden
 
 [UninstallRun]
@@ -184,46 +179,6 @@ begin
   Result := Pos(';' + ExpandConstant(Param) + ';', ';' + OrigPath + ';') = 0;
 end;
 
-{ --- VC Runtime libraries discovery code - Only install vc_redist if it isn't already installed ----- }
-const VCRTL_MIN_V1 = 14;
-const VCRTL_MIN_V2 = 40;
-const VCRTL_MIN_V3 = 33807;
-const VCRTL_MIN_V4 = 0;
-
- // check if the minimum required vc redist is installed (by looking the registry)
-function vc_redist_needed (): Boolean;
-var
-  sRegKey: string;
-  v1: Cardinal;
-  v2: Cardinal;
-  v3: Cardinal;
-  v4: Cardinal;
-begin
-  if (IsArm64()) then begin
-    sRegKey := 'SOFTWARE\WOW6432Node\Microsoft\VisualStudio\14.0\VC\Runtimes\arm64';
-  end else begin
-    sRegKey := 'SOFTWARE\Microsoft\VisualStudio\14.0\VC\Runtimes\x64';
-  end;
-  if (RegQueryDWordValue (HKEY_LOCAL_MACHINE, sRegKey, 'Major', v1)  and
-      RegQueryDWordValue (HKEY_LOCAL_MACHINE, sRegKey, 'Minor', v2) and
-      RegQueryDWordValue (HKEY_LOCAL_MACHINE, sRegKey, 'Bld', v3) and
-      RegQueryDWordValue (HKEY_LOCAL_MACHINE, sRegKey, 'RBld', v4)) then
-  begin
-    Log ('VC Redist version: ' + IntToStr (v1) +
-        '.' + IntToStr (v2) + '.' + IntToStr (v3) +
-        '.' + IntToStr (v4));
-    { Version info was found. Return true if later or equal to our
-       minimal required version RTL_MIN_Vx }
-    Result := not (
-        (v1 > VCRTL_MIN_V1) or ((v1 = VCRTL_MIN_V1) and
-         ((v2 > VCRTL_MIN_V2) or ((v2 = VCRTL_MIN_V2) and
-          ((v3 > VCRTL_MIN_V3) or ((v3 = VCRTL_MIN_V3) and
-           (v4 >= VCRTL_MIN_V4)))))));
-  end
-  else
-    Result := TRUE;
-end;
-
 function GetDirSize(Path: String): Int64;
 var
   FindRec: TFindRec;

app/store/database.go

@@ -1201,15 +1201,16 @@ func (db *database) getSettings() (Settings, error) {
 func (db *database) setSettings(s Settings) error {
 	lastHomeView := strings.ToLower(strings.TrimSpace(s.LastHomeView))
 	validLaunchView := map[string]struct{}{
-		"launch":   {},
-		"openclaw": {},
-		"claude":   {},
-		"hermes":   {},
-		"codex":    {},
-		"copilot":  {},
-		"opencode": {},
-		"droid":    {},
-		"pi":       {},
+		"launch":    {},
+		"openclaw":  {},
+		"claude":    {},
+		"hermes":    {},
+		"codex":     {},
+		"codex-app": {},
+		"copilot":   {},
+		"opencode":  {},
+		"droid":     {},
+		"pi":        {},
 	}
 	if lastHomeView != "chat" {
 		if _, ok := validLaunchView[lastHomeView]; !ok {

app/store/store_test.go

@@ -122,6 +122,21 @@ func TestStore(t *testing.T) {
 		}
 	})
 
+	t.Run("settings codex app home view is accepted", func(t *testing.T) {
+		if err := s.SetSettings(Settings{LastHomeView: "codex-app"}); err != nil {
+			t.Fatal(err)
+		}
+
+		loaded, err := s.Settings()
+		if err != nil {
+			t.Fatal(err)
+		}
+
+		if loaded.LastHomeView != "codex-app" {
+			t.Fatalf("expected codex-app LastHomeView to be preserved, got %q", loaded.LastHomeView)
+		}
+	})
+
 	t.Run("window size", func(t *testing.T) {
 		if err := s.SetWindowSize(1024, 768); err != nil {
 			t.Fatal(err)

app/ui/app/src/components/LaunchCommands.tsx

@@ -22,11 +22,12 @@ const LAUNCH_COMMANDS: LaunchCommand[] = [
     iconClassName: "h-7 w-7",
   },
   {
-    id: "openclaw",
-    name: "OpenClaw",
-    command: "ollama launch openclaw",
-    description: "Personal AI with 100+ skills",
-    icon: "/launch-icons/openclaw.svg",
+    id: "codex-app",
+    name: "Codex App",
+    command: "ollama launch codex-app",
+    description: "An AI agent you can delegate real work to, by OpenAI",
+    icon: "/launch-icons/codex-app.png",
+    iconClassName: "h-full w-full",
   },
   {
     id: "hermes",
@@ -36,6 +37,13 @@ const LAUNCH_COMMANDS: LaunchCommand[] = [
     icon: "/launch-icons/hermes-agent.svg",
     iconClassName: "h-7 w-7",
   },
+  {
+    id: "openclaw",
+    name: "OpenClaw",
+    command: "ollama launch openclaw",
+    description: "Personal AI with 100+ skills",
+    icon: "/launch-icons/openclaw.svg",
+  },
   {
     id: "opencode",
     name: "OpenCode",

app/updater/updater.go

@@ -5,6 +5,8 @@ package updater
 import (
 	"context"
 	"crypto/rand"
+	"crypto/sha256"
+	"encoding/hex"
 	"encoding/json"
 	"errors"
 	"fmt"
@@ -169,22 +171,20 @@ func (u *Updater) DownloadNewRelease(ctx context.Context, updateResp UpdateRespo
 	if err != nil {
 		return fmt.Errorf("error checking update: %w", err)
 	}
+	resp.Body.Close()
 	if resp.StatusCode != http.StatusOK {
 		return fmt.Errorf("unexpected status attempting to download update %d", resp.StatusCode)
 	}
-	resp.Body.Close()
-	etag := strings.Trim(resp.Header.Get("etag"), "\"")
-	if etag == "" {
-		slog.Debug("no etag detected, falling back to filename based dedup")
-		etag = "_"
-	}
 	filename := Installer
 	_, params, err := mime.ParseMediaType(resp.Header.Get("content-disposition"))
-	if err == nil {
+	if err == nil && params["filename"] != "" {
 		filename = params["filename"]
 	}
 
-	stageFilename := filepath.Join(UpdateStageDir, etag, filename)
+	stageFilename, err := updateStagePath(UpdateStageDir, resp.Header.Get("etag"), filename)
+	if err != nil {
+		return err
+	}
 
 	// Check to see if we already have it downloaded
 	_, err = os.Stat(stageFilename)
@@ -202,13 +202,14 @@ func (u *Updater) DownloadNewRelease(ctx context.Context, updateResp UpdateRespo
 		return fmt.Errorf("error checking update: %w", err)
 	}
 	defer resp.Body.Close()
-	etag = strings.Trim(resp.Header.Get("etag"), "\"")
-	if etag == "" {
-		slog.Debug("no etag detected, falling back to filename based dedup") // TODO probably can get rid of this redundant log
-		etag = "_"
+	if resp.StatusCode != http.StatusOK {
+		return fmt.Errorf("unexpected status attempting to download update %d", resp.StatusCode)
 	}
 
-	stageFilename = filepath.Join(UpdateStageDir, etag, filename)
+	stageFilename, err = updateStagePath(UpdateStageDir, resp.Header.Get("etag"), filename)
+	if err != nil {
+		return err
+	}
 
 	_, err = os.Stat(filepath.Dir(stageFilename))
 	if errors.Is(err, os.ErrNotExist) {
@@ -225,10 +226,13 @@ func (u *Updater) DownloadNewRelease(ctx context.Context, updateResp UpdateRespo
 	if err != nil {
 		return fmt.Errorf("write payload %s: %w", stageFilename, err)
 	}
-	defer fp.Close()
 	if n, err := fp.Write(payload); err != nil || n != len(payload) {
+		_ = fp.Close()
 		return fmt.Errorf("write payload %s: %d vs %d -- %w", stageFilename, n, len(payload), err)
 	}
+	if err := fp.Close(); err != nil {
+		return fmt.Errorf("close payload %s: %w", stageFilename, err)
+	}
 	slog.Info("new update downloaded " + stageFilename)
 
 	if err := VerifyDownload(); err != nil {
@@ -239,6 +243,61 @@ func (u *Updater) DownloadNewRelease(ctx context.Context, updateResp UpdateRespo
 	return nil
 }
 
+func updateStagePath(stageDir, etag, filename string) (string, error) {
+	filename, err := safeUpdateFilename(filename)
+	if err != nil {
+		return "", err
+	}
+
+	stageDir, err = filepath.Abs(stageDir)
+	if err != nil {
+		return "", fmt.Errorf("resolve update stage dir: %w", err)
+	}
+
+	stageFilename := filepath.Join(stageDir, updateStageETagDir(etag), filename)
+	if err := ensurePathInDir(stageDir, stageFilename); err != nil {
+		return "", err
+	}
+
+	return stageFilename, nil
+}
+
+func safeUpdateFilename(filename string) (string, error) {
+	filename = strings.TrimSpace(filename)
+	if filename == "" {
+		return "", errors.New("missing update filename")
+	}
+	if filename == "." || filename == ".." ||
+		filepath.IsAbs(filename) || path.IsAbs(filename) ||
+		strings.ContainsAny(filename, `/\:`) ||
+		filepath.Base(filename) != filename || path.Base(filename) != filename {
+		return "", fmt.Errorf("unsafe update filename %q", filename)
+	}
+	return filename, nil
+}
+
+func updateStageETagDir(etag string) string {
+	etag = strings.Trim(strings.TrimSpace(etag), "\"")
+	if etag == "" {
+		slog.Debug("no etag detected, falling back to filename based dedup")
+		return "_"
+	}
+
+	sum := sha256.Sum256([]byte(etag))
+	return hex.EncodeToString(sum[:])
+}
+
+func ensurePathInDir(dir, name string) error {
+	rel, err := filepath.Rel(dir, name)
+	if err != nil {
+		return fmt.Errorf("resolve update staging path: %w", err)
+	}
+	if rel == ".." || strings.HasPrefix(rel, ".."+string(filepath.Separator)) || filepath.IsAbs(rel) {
+		return fmt.Errorf("update staging path escapes stage dir: %s", name)
+	}
+	return nil
+}
+
 func cleanupOldDownloads(stageDir string) {
 	files, err := os.ReadDir(stageDir)
 	if err != nil && errors.Is(err, os.ErrNotExist) {

app/updater/updater_darwin.go

@@ -22,6 +22,15 @@ import (
 	"golang.org/x/sys/unix"
 )
 
+const updateArchiveRoot = "Ollama.app"
+
+type bundleEntryScope int
+
+const (
+	bundleEntryRelative bundleEntryScope = iota
+	bundleEntryWithArchiveRoot
+)
+
 var (
 	appBackupDir   string
 	SystemWidePath = "/Applications/Ollama.app"
@@ -167,8 +176,12 @@ func DoUpgrade(interactive bool) error {
 		}
 		name := s[1]
 		if strings.HasSuffix(name, "/") {
-			d := filepath.Join(BundlePath, name)
-			err := os.MkdirAll(d, 0o755)
+			d, err := bundleEntryPath(BundlePath, name, bundleEntryRelative)
+			if err != nil {
+				anyFailures = true
+				return err
+			}
+			err = os.MkdirAll(d, 0o755)
 			if err != nil {
 				anyFailures = true
 				return fmt.Errorf("failed to mkdir %s: %w", d, err)
@@ -181,30 +194,14 @@ func DoUpgrade(interactive bool) error {
 			continue
 		}
 
-		src, err := f.Open()
+		destName, err := bundleEntryPath(BundlePath, name, bundleEntryRelative)
 		if err != nil {
 			anyFailures = true
-			return fmt.Errorf("failed to open bundle file %s: %w", name, err)
+			return err
 		}
-		destName := filepath.Join(BundlePath, name)
-		// Verify directory first
-		d := filepath.Dir(destName)
-		if _, err := os.Stat(d); err != nil {
-			err := os.MkdirAll(d, 0o755)
-			if err != nil {
-				anyFailures = true
-				return fmt.Errorf("failed to mkdir %s: %w", d, err)
-			}
-		}
-		destFile, err := os.OpenFile(destName, os.O_WRONLY|os.O_CREATE|os.O_TRUNC, 0o755)
-		if err != nil {
+		if err := extractBundleFile(f, destName, name); err != nil {
 			anyFailures = true
-			return fmt.Errorf("failed to open output file %s: %w", destName, err)
-		}
-		defer destFile.Close()
-		if _, err := io.Copy(destFile, src); err != nil {
-			anyFailures = true
-			return fmt.Errorf("failed to open extract file %s: %w", destName, err)
+			return err
 		}
 	}
 	for _, f := range links {
@@ -225,16 +222,24 @@ func DoUpgrade(interactive bool) error {
 			return err
 		}
 		link := string(buf)
-		if link[0] == '/' {
+		if link == "" {
+			anyFailures = true
+			return fmt.Errorf("bundle contains empty symlink %s", f.Name)
+		}
+		if filepath.IsAbs(link) {
 			anyFailures = true
 			return fmt.Errorf("bundle contains absolute symlink %s -> %s", f.Name, link)
 		}
-		// Don't allow links outside of Ollama.app
-		if strings.HasPrefix(filepath.Join(filepath.Dir(name), link), "..") {
+		if !validBundleLinkTarget(name, link, bundleEntryRelative) {
 			anyFailures = true
-			return fmt.Errorf("bundle contains link outside of contents %s -> %s", f.Name, link)
+			return fmt.Errorf("bundle contains invalid symlink %s -> %s", f.Name, link)
 		}
-		if err = os.Symlink(link, filepath.Join(BundlePath, name)); err != nil {
+		destName, err := bundleEntryPath(BundlePath, name, bundleEntryRelative)
+		if err != nil {
+			anyFailures = true
+			return err
+		}
+		if err = os.Symlink(link, destName); err != nil {
 			anyFailures = true
 			return err
 		}
@@ -282,8 +287,11 @@ func verifyDownload() error {
 	links := []*zip.File{}
 	for _, f := range r.File {
 		if strings.HasSuffix(f.Name, "/") {
-			d := filepath.Join(dir, f.Name)
-			err := os.MkdirAll(d, 0o755)
+			d, err := bundleEntryPath(dir, f.Name, bundleEntryWithArchiveRoot)
+			if err != nil {
+				return err
+			}
+			err = os.MkdirAll(d, 0o755)
 			if err != nil {
 				return fmt.Errorf("failed to mkdir %s: %w", d, err)
 			}
@@ -294,26 +302,12 @@ func verifyDownload() error {
 			links = append(links, f)
 			continue
 		}
-		src, err := f.Open()
+		destName, err := bundleEntryPath(dir, f.Name, bundleEntryWithArchiveRoot)
 		if err != nil {
-			return fmt.Errorf("failed to open bundle file %s: %w", f.Name, err)
+			return err
 		}
-		destName := filepath.Join(dir, f.Name)
-		// Verify directory first
-		d := filepath.Dir(destName)
-		if _, err := os.Stat(d); err != nil {
-			err := os.MkdirAll(d, 0o755)
-			if err != nil {
-				return fmt.Errorf("failed to mkdir %s: %w", d, err)
-			}
-		}
-		destFile, err := os.OpenFile(destName, os.O_WRONLY|os.O_CREATE|os.O_TRUNC, 0o755)
-		if err != nil {
-			return fmt.Errorf("failed to open output file %s: %w", destName, err)
-		}
-		defer destFile.Close()
-		if _, err := io.Copy(destFile, src); err != nil {
-			return fmt.Errorf("failed to open extract file %s: %w", destName, err)
+		if err := extractBundleFile(f, destName, f.Name); err != nil {
+			return err
 		}
 	}
 	for _, f := range links {
@@ -326,13 +320,20 @@ func verifyDownload() error {
 			return err
 		}
 		link := string(buf)
-		if link[0] == '/' {
+		if link == "" {
+			return fmt.Errorf("bundle contains empty symlink %s", f.Name)
+		}
+		if filepath.IsAbs(link) {
 			return fmt.Errorf("bundle contains absolute symlink %s -> %s", f.Name, link)
 		}
-		if strings.HasPrefix(filepath.Join(filepath.Dir(f.Name), link), "..") {
-			return fmt.Errorf("bundle contains link outside of contents %s -> %s", f.Name, link)
+		if !validBundleLinkTarget(f.Name, link, bundleEntryWithArchiveRoot) {
+			return fmt.Errorf("bundle contains invalid symlink %s -> %s", f.Name, link)
 		}
-		if err = os.Symlink(link, filepath.Join(dir, f.Name)); err != nil {
+		destName, err := bundleEntryPath(dir, f.Name, bundleEntryWithArchiveRoot)
+		if err != nil {
+			return err
+		}
+		if err = os.Symlink(link, destName); err != nil {
 			return err
 		}
 	}
@@ -343,6 +344,53 @@ func verifyDownload() error {
 	return nil
 }
 
+func bundleEntryPath(root, name string, scope bundleEntryScope) (string, error) {
+	cleanName := filepath.Clean(filepath.FromSlash(name))
+	if !filepath.IsLocal(cleanName) {
+		return "", fmt.Errorf("bundle contains invalid path: %s", name)
+	}
+	if scope == bundleEntryWithArchiveRoot && cleanName != updateArchiveRoot &&
+		!strings.HasPrefix(cleanName, updateArchiveRoot+string(os.PathSeparator)) {
+		return "", fmt.Errorf("bundle contains invalid path: %s", name)
+	}
+	return filepath.Join(root, cleanName), nil
+}
+
+func extractBundleFile(f *zip.File, destName, name string) error {
+	src, err := f.Open()
+	if err != nil {
+		return fmt.Errorf("failed to open bundle file %s: %w", name, err)
+	}
+	defer src.Close()
+
+	d := filepath.Dir(destName)
+	if _, err := os.Stat(d); err != nil {
+		if err := os.MkdirAll(d, 0o755); err != nil {
+			return fmt.Errorf("failed to mkdir %s: %w", d, err)
+		}
+	}
+
+	destFile, err := os.OpenFile(destName, os.O_WRONLY|os.O_CREATE|os.O_TRUNC, 0o755)
+	if err != nil {
+		return fmt.Errorf("failed to open output file %s: %w", destName, err)
+	}
+	defer destFile.Close()
+
+	if _, err := io.Copy(destFile, src); err != nil {
+		return fmt.Errorf("failed to open extract file %s: %w", destName, err)
+	}
+	return nil
+}
+
+func validBundleLinkTarget(name, link string, scope bundleEntryScope) bool {
+	cleanTarget := filepath.Clean(filepath.Join(filepath.Dir(filepath.FromSlash(name)), filepath.FromSlash(link)))
+	if !filepath.IsLocal(cleanTarget) {
+		return false
+	}
+	return scope == bundleEntryRelative || cleanTarget == updateArchiveRoot ||
+		strings.HasPrefix(cleanTarget, updateArchiveRoot+string(os.PathSeparator))
+}
+
 // If we detect an upgrade bundle, attempt to upgrade at startup
 func DoUpgradeAtStartup() error {
 	bundle := getStagedUpdate()

app/updater/updater_darwin_test.go

@@ -2,6 +2,7 @@ package updater
 
 import (
 	"archive/zip"
+	"errors"
 	"io/fs"
 	"os"
 	"path/filepath"
@@ -146,6 +147,46 @@ func TestDoUpgrade(t *testing.T) {
 	}
 }
 
+func TestDoUpgradeRejectsInvalidBundlePath(t *testing.T) {
+	tmpDir := t.TempDir()
+	BundlePath = filepath.Join(tmpDir, "Ollama.app")
+	appBackupDir = filepath.Join(tmpDir, "backup")
+	UpdateStageDir = filepath.Join(tmpDir, "updates")
+	UpgradeMarkerFile = filepath.Join(tmpDir, "upgraded")
+	bundle := filepath.Join(UpdateStageDir, "foo", "ollama-darwin.zip")
+	invalidTarget := filepath.Join(tmpDir, "invalid-entry")
+
+	if err := os.MkdirAll(filepath.Join(BundlePath, "Contents", "MacOS"), 0o755); err != nil {
+		t.Fatal("failed to create empty dirs")
+	}
+	if err := os.WriteFile(filepath.Join(BundlePath, "Contents", "MacOS", "Ollama"), []byte("old app"), 0o755); err != nil {
+		t.Fatal("failed to create old app")
+	}
+	if err := os.MkdirAll(filepath.Dir(bundle), 0o755); err != nil {
+		t.Fatal("failed to create empty dirs")
+	}
+	if err := zipCreationHelper(bundle, []testPayload{{
+		Name: "Ollama.app/../invalid-entry",
+		Body: []byte("payload"),
+	}}); err != nil {
+		t.Fatal(err)
+	}
+
+	if err := DoUpgrade(false); err == nil {
+		t.Fatal("expected failure with invalid bundle path")
+	} else if !strings.Contains(err.Error(), "bundle contains invalid path") {
+		t.Fatalf("unexpected error with invalid bundle path: %s", err)
+	}
+	if _, err := os.Stat(invalidTarget); err == nil {
+		t.Fatalf("invalid bundle path wrote %s", invalidTarget)
+	} else if !errors.Is(err, os.ErrNotExist) {
+		t.Fatalf("unexpected stat error for %s: %s", invalidTarget, err)
+	}
+	if _, err := os.Stat(filepath.Join(BundlePath, "Contents", "MacOS", "Ollama")); err != nil {
+		t.Fatalf("old app was not restored: %s", err)
+	}
+}
+
 func TestDoUpgradeAtStartup(t *testing.T) {
 	tmpDir := t.TempDir()
 	BundlePath = filepath.Join(tmpDir, "Ollama.app")
@@ -203,7 +244,7 @@ func TestVerifyDownloadFailures(t *testing.T) {
 		in       []testPayload
 		expected string
 	}{
-		{"breakout", []testPayload{
+		{"invalid symlink target", []testPayload{
 			{
 				Name: "Ollama.app/",
 				Body: []byte{},
@@ -212,15 +253,34 @@ func TestVerifyDownloadFailures(t *testing.T) {
 				Body: []byte("cli payload here"),
 			}, {
 				Name: "Ollama.app/Contents/MacOS/Ollama",
-				Body: []byte("../../../../breakout"),
+				Body: []byte("../../../../invalid-target"),
 				Mode: os.ModeSymlink,
 			},
-		}, "bundle contains link outside"},
+		}, "bundle contains invalid symlink"},
+		{"invalid archive symlink target", []testPayload{
+			{
+				Name: "Ollama.app/Contents/MacOS/Ollama",
+				Body: []byte("../../../invalid-target"),
+				Mode: os.ModeSymlink,
+			},
+		}, "bundle contains invalid symlink"},
 		{"absolute", []testPayload{{
 			Name: "Ollama.app/Contents/MacOS/Ollama",
 			Body: []byte("/etc/foo"),
 			Mode: os.ModeSymlink,
 		}}, "bundle contains absolute"},
+		{"invalid relative file", []testPayload{{
+			Name: "Ollama.app/../invalid-entry",
+			Body: []byte("payload"),
+		}}, "bundle contains invalid path"},
+		{"invalid relative directory", []testPayload{{
+			Name: "Ollama.app/../invalid-entry/",
+			Body: []byte{},
+		}}, "bundle contains invalid path"},
+		{"absolute file", []testPayload{{
+			Name: filepath.Join(tmpDir, "invalid-entry"),
+			Body: []byte("payload"),
+		}}, "bundle contains invalid path"},
 		{"missing", []testPayload{{
 			Name: "Ollama.app/Contents/MacOS/Ollama",
 			Body: []byte("../nothere"),
@@ -242,6 +302,11 @@ func TestVerifyDownloadFailures(t *testing.T) {
 			if err == nil || !strings.Contains(err.Error(), tt.expected) {
 				t.Fatalf("expected \"%s\" got %s", tt.expected, err)
 			}
+			if _, err := os.Stat(filepath.Join(tmpDir, "invalid-entry")); err == nil {
+				t.Fatal("invalid bundle path wrote unexpected file")
+			} else if !errors.Is(err, os.ErrNotExist) {
+				t.Fatalf("unexpected stat error for invalid file: %s", err)
+			}
 		})
 	}
 }

app/updater/updater_live_test.go

@@ -0,0 +1,127 @@
+//go:build (windows || darwin) && updater_live
+
+package updater
+
+import (
+	"context"
+	"os"
+	"path/filepath"
+	"runtime"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/ollama/ollama/app/store"
+	"github.com/ollama/ollama/app/version"
+)
+
+// TestLiveAppUpdate exercises the production update endpoint and downloads the
+// current OS update artifact. It is intentionally excluded from normal test
+// runs because it depends on ollama.com and downloads a release artifact.
+//
+// Run with:
+//
+//	go test -tags updater_live -run TestLiveAppUpdate ./app/updater
+func TestLiveAppUpdate(t *testing.T) {
+	const spoofedVersion = "0.20.0"
+
+	ctx, cancel := context.WithTimeout(context.Background(), 10*time.Minute)
+	defer cancel()
+
+	oldUpdateStageDir := UpdateStageDir
+	oldUpdateDownloaded := UpdateDownloaded
+	oldVerifyDownload := VerifyDownload
+	oldVersion := version.Version
+	defer func() {
+		UpdateStageDir = oldUpdateStageDir
+		UpdateDownloaded = oldUpdateDownloaded
+		VerifyDownload = oldVerifyDownload
+		version.Version = oldVersion
+	}()
+
+	version.Version = spoofedVersion
+
+	expectedFilename := ""
+	switch runtime.GOOS {
+	case "windows":
+		t.Setenv("LOCALAPPDATA", t.TempDir())
+		expectedFilename = "OllamaSetup.exe"
+	case "darwin":
+		expectedFilename = "Ollama-darwin.zip"
+	default:
+		t.Fatalf("unsupported updater live test OS %q", runtime.GOOS)
+	}
+
+	UpdateStageDir = filepath.Join(t.TempDir(), "updates")
+	UpdateDownloaded = false
+	verifyCalled := false
+	VerifyDownload = func() error {
+		verifyCalled = true
+		return verifyDownload()
+	}
+
+	updater := &Updater{Store: &store.Store{DBPath: filepath.Join(t.TempDir(), "db.sqlite")}}
+	defer updater.Store.Close()
+
+	available, updateResp := updater.checkForUpdate(ctx)
+	if !available {
+		t.Fatalf("expected production update check to offer an update for spoofed version %s", spoofedVersion)
+	}
+	if updateResp.UpdateURL == "" {
+		t.Fatal("production update response did not include a download URL")
+	}
+	t.Logf("production update version=%q url=%q", updateResp.UpdateVersion, updateResp.UpdateURL)
+
+	if err := updater.DownloadNewRelease(ctx, updateResp); err != nil {
+		t.Fatalf("download production update: %v", err)
+	}
+
+	staged := getStagedUpdate()
+	if staged == "" {
+		t.Fatal("production update was not staged")
+	}
+	t.Logf("staged production update at %s", staged)
+
+	assertPathInsideDir(t, UpdateStageDir, staged)
+	if filepath.Base(staged) != expectedFilename {
+		t.Fatalf("expected staged %s update filename to be %q, got %q", runtime.GOOS, expectedFilename, filepath.Base(staged))
+	}
+	expectedExt := filepath.Ext(expectedFilename)
+	if filepath.Ext(staged) != expectedExt {
+		t.Fatalf("expected staged %s update to be a %s artifact, got %s", runtime.GOOS, expectedExt, staged)
+	}
+
+	info, err := os.Stat(staged)
+	if err != nil {
+		t.Fatalf("stat staged update: %v", err)
+	}
+	if info.Size() == 0 {
+		t.Fatal("staged production update is empty")
+	}
+
+	if !verifyCalled {
+		t.Fatal("DownloadNewRelease did not call VerifyDownload")
+	}
+	t.Logf("production updater download path verified staged %s update", runtime.GOOS)
+}
+
+func assertPathInsideDir(t *testing.T, dir, name string) {
+	t.Helper()
+
+	dir, err := filepath.Abs(dir)
+	if err != nil {
+		t.Fatal(err)
+	}
+	name, err = filepath.Abs(name)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	rel, err := filepath.Rel(dir, name)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if rel == ".." || strings.HasPrefix(rel, ".."+string(filepath.Separator)) || filepath.IsAbs(rel) {
+		t.Fatalf("staged update escaped update stage dir: %s", name)
+	}
+}

app/updater/updater_test.go

@@ -11,7 +11,9 @@ import (
 	"log/slog"
 	"net/http"
 	"net/http/httptest"
+	"os"
 	"path/filepath"
+	"strings"
 	"sync/atomic"
 	"testing"
 	"time"
@@ -19,6 +21,52 @@ import (
 	"github.com/ollama/ollama/app/store"
 )
 
+func TestUpdateStagePathRejectsUnsafeFilename(t *testing.T) {
+	stageDir := t.TempDir()
+	for _, tt := range []struct {
+		name     string
+		filename string
+	}{
+		{"empty", ""},
+		{"dot", "."},
+		{"dotdot", ".."},
+		{"posix_parent", "../OllamaSetup.exe"},
+		{"windows_parent", `..\OllamaSetup.exe`},
+		{"posix_absolute_tmp", "/tmp/OllamaSetup.exe"},
+		{"darwin_absolute_app", "/Applications/Ollama.app"},
+		{"darwin_bundle_path", "Ollama.app/Contents/MacOS/Ollama"},
+		{"darwin_user_download", "~/Downloads/Ollama-darwin.zip"},
+		{"windows_absolute", `C:\Users\Public\OllamaSetup.exe`},
+		{"colon", "Ollama:Setup.exe"},
+	} {
+		t.Run(tt.name, func(t *testing.T) {
+			if _, err := updateStagePath(stageDir, "etag", tt.filename); err == nil {
+				t.Fatal("expected unsafe filename to be rejected")
+			}
+		})
+	}
+}
+
+func TestUpdateStagePathHashesETag(t *testing.T) {
+	stageDir := t.TempDir()
+	stageFilename, err := updateStagePath(stageDir, `../escaped`, "OllamaSetup.exe")
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	rel, err := filepath.Rel(stageDir, stageFilename)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if rel == ".." || strings.HasPrefix(rel, ".."+string(filepath.Separator)) || filepath.IsAbs(rel) {
+		t.Fatalf("stage filename escaped stage dir: %s", stageFilename)
+	}
+	etagDir := filepath.Base(filepath.Dir(stageFilename))
+	if etagDir == ".." || etagDir == "escaped" || strings.ContainsAny(etagDir, `/\`) {
+		t.Fatalf("stage filename used raw etag path component: %s", stageFilename)
+	}
+}
+
 func TestIsNewReleaseAvailable(t *testing.T) {
 	slog.SetLogLoggerLevel(slog.LevelDebug)
 	var server *httptest.Server
@@ -47,6 +95,223 @@ func TestIsNewReleaseAvailable(t *testing.T) {
 	}
 }
 
+func TestDownloadNewReleaseRejectsUnsafeHeaderFilename(t *testing.T) {
+	UpdateStageDir = t.TempDir()
+	oldInstaller := Installer
+	oldVerifyDownload := VerifyDownload
+	oldUpdateDownloaded := UpdateDownloaded
+	defer func() {
+		Installer = oldInstaller
+		VerifyDownload = oldVerifyDownload
+		UpdateDownloaded = oldUpdateDownloaded
+	}()
+	Installer = "OllamaSetup.exe"
+	UpdateDownloaded = false
+	VerifyDownload = func() error {
+		t.Fatal("verification should not run for rejected downloads")
+		return nil
+	}
+
+	var getAttempted atomic.Bool
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.Method == http.MethodHead {
+			w.Header().Set("ETag", `"safe"`)
+			w.Header().Set("Content-Disposition", `attachment; filename="../OllamaSetup.exe"`)
+			w.WriteHeader(http.StatusOK)
+			return
+		}
+		getAttempted.Store(true)
+		w.WriteHeader(http.StatusOK)
+	}))
+	defer server.Close()
+
+	updater := &Updater{}
+	err := updater.DownloadNewRelease(t.Context(), UpdateResponse{UpdateURL: server.URL + "/download"})
+	if err == nil || !strings.Contains(err.Error(), "unsafe update filename") {
+		t.Fatalf("expected unsafe filename error, got %v", err)
+	}
+	if getAttempted.Load() {
+		t.Fatal("download should not continue after unsafe filename")
+	}
+	if _, err := os.Stat(filepath.Join(filepath.Dir(UpdateStageDir), "OllamaSetup.exe")); err == nil {
+		t.Fatal("download escaped update stage dir")
+	}
+}
+
+func TestDownloadNewReleaseDoesNotUseRawETagAsPathComponent(t *testing.T) {
+	UpdateStageDir = t.TempDir()
+	oldInstaller := Installer
+	oldVerifyDownload := VerifyDownload
+	oldUpdateDownloaded := UpdateDownloaded
+	defer func() {
+		Installer = oldInstaller
+		VerifyDownload = oldVerifyDownload
+		UpdateDownloaded = oldUpdateDownloaded
+	}()
+	Installer = "OllamaSetup.exe"
+	UpdateDownloaded = false
+	VerifyDownload = func() error {
+		return nil
+	}
+
+	payload := []byte("payload")
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("ETag", `"../escaped"`)
+		w.WriteHeader(http.StatusOK)
+		if r.Method == http.MethodGet {
+			_, _ = w.Write(payload)
+		}
+	}))
+	defer server.Close()
+
+	updater := &Updater{}
+	if err := updater.DownloadNewRelease(t.Context(), UpdateResponse{UpdateURL: server.URL + "/download"}); err != nil {
+		t.Fatal(err)
+	}
+
+	if _, err := os.Stat(filepath.Join(filepath.Dir(UpdateStageDir), "escaped", Installer)); err == nil {
+		t.Fatal("download escaped update stage dir via etag")
+	}
+
+	entries, err := os.ReadDir(UpdateStageDir)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(entries) != 1 {
+		t.Fatalf("expected one staged update dir, got %d", len(entries))
+	}
+	stageFilename := filepath.Join(UpdateStageDir, entries[0].Name(), Installer)
+	got, err := os.ReadFile(stageFilename)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if string(got) != string(payload) {
+		t.Fatalf("unexpected staged payload %q", got)
+	}
+}
+
+func TestBackgroundCheckerSkipsAlreadyStagedETagDownload(t *testing.T) {
+	UpdateStageDir = t.TempDir()
+	oldInstaller := Installer
+	oldVerifyDownload := VerifyDownload
+	oldUpdateDownloaded := UpdateDownloaded
+	oldUpdateCheckInitialDelay := UpdateCheckInitialDelay
+	oldUpdateCheckInterval := UpdateCheckInterval
+	oldUpdateCheckURLBase := UpdateCheckURLBase
+	defer func() {
+		Installer = oldInstaller
+		VerifyDownload = oldVerifyDownload
+		UpdateDownloaded = oldUpdateDownloaded
+		UpdateCheckInitialDelay = oldUpdateCheckInitialDelay
+		UpdateCheckInterval = oldUpdateCheckInterval
+		UpdateCheckURLBase = oldUpdateCheckURLBase
+	}()
+	Installer = "OllamaSetup.exe"
+	UpdateDownloaded = false
+	UpdateCheckInitialDelay = time.Millisecond
+	UpdateCheckInterval = 5 * time.Millisecond
+
+	var verifyCount atomic.Int32
+	VerifyDownload = func() error {
+		verifyCount.Add(1)
+		return nil
+	}
+
+	headETag := `"old-update"`
+	getETag := `"download-response-etag"`
+	payload := []byte("payload")
+	var headCount atomic.Int32
+	var getCount atomic.Int32
+	var server *httptest.Server
+	server = httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		switch r.URL.Path {
+		case "/update.json":
+			w.Write([]byte(
+				fmt.Sprintf(`{"version": "9.9.9", "url": "%s"}`,
+					server.URL+"/9.9.9/"+Installer)))
+		case "/9.9.9/" + Installer:
+			w.Header().Set("Content-Disposition", `attachment; filename="OllamaSetup.exe"`)
+			switch r.Method {
+			case http.MethodHead:
+				etag := headETag
+				if getCount.Load() > 0 {
+					etag = getETag
+				}
+				w.Header().Set("ETag", etag)
+				headCount.Add(1)
+				w.WriteHeader(http.StatusOK)
+			case http.MethodGet:
+				w.Header().Set("ETag", getETag)
+				getCount.Add(1)
+				w.WriteHeader(http.StatusOK)
+				_, _ = w.Write(payload)
+			default:
+				t.Errorf("unexpected request method %s", r.Method)
+				w.WriteHeader(http.StatusMethodNotAllowed)
+			}
+		default:
+			t.Errorf("unexpected request path %s", r.URL.Path)
+			w.WriteHeader(http.StatusNotFound)
+		}
+	}))
+	defer server.Close()
+	UpdateCheckURLBase = server.URL + "/update.json"
+
+	updater := &Updater{Store: &store.Store{DBPath: filepath.Join(t.TempDir(), "test.db")}}
+	defer updater.Store.Close()
+	settings, err := updater.Store.Settings()
+	if err != nil {
+		t.Fatal(err)
+	}
+	settings.AutoUpdateEnabled = true
+	if err := updater.Store.SetSettings(settings); err != nil {
+		t.Fatal(err)
+	}
+
+	ctx, cancel := context.WithCancel(t.Context())
+	defer cancel()
+
+	callbacks := make(chan string, 4)
+	updater.StartBackgroundUpdaterChecker(ctx, func(ver string) error {
+		callbacks <- ver
+		return nil
+	})
+
+	for range 2 {
+		select {
+		case <-callbacks:
+		case <-time.After(5 * time.Second):
+			t.Fatal("timed out waiting for repeated update checks")
+		}
+	}
+	cancel()
+
+	stageFilename, err := updateStagePath(UpdateStageDir, getETag, Installer)
+	if err != nil {
+		t.Fatal(err)
+	}
+	got, err := os.ReadFile(stageFilename)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if string(got) != string(payload) {
+		t.Fatalf("unexpected staged payload %q", got)
+	}
+
+	if headCount.Load() < 2 {
+		t.Fatalf("HEAD count = %d, want at least 2", headCount.Load())
+	}
+	if getCount.Load() != 1 {
+		t.Fatalf("GET count = %d, want 1", getCount.Load())
+	}
+	if verifyCount.Load() != 1 {
+		t.Fatalf("verification count = %d, want 1", verifyCount.Load())
+	}
+	if !UpdateDownloaded {
+		t.Fatal("UpdateDownloaded should stay true for already staged update")
+	}
+}
+
 func TestBackgoundChecker(t *testing.T) {
 	UpdateStageDir = t.TempDir()
 	haveUpdate := false

app/updater/updater_windows.go

@@ -1,6 +1,7 @@
 package updater
 
 import (
+	"crypto/x509"
 	"errors"
 	"fmt"
 	"log/slog"
@@ -18,6 +19,30 @@ import (
 
 var runningInstaller string
 
+var (
+	crypt32              = windows.NewLazySystemDLL("crypt32.dll")
+	procCryptMsgGetParam = crypt32.NewProc("CryptMsgGetParam")
+	procCryptMsgClose    = crypt32.NewProc("CryptMsgClose")
+)
+
+const cmsgSignerInfoParam = 6
+
+type cmsgSignerInfo struct {
+	Version                 uint32
+	Issuer                  windows.CertNameBlob
+	SerialNumber            windows.CryptIntegerBlob
+	HashAlgorithm           windows.CryptAlgorithmIdentifier
+	HashEncryptionAlgorithm windows.CryptAlgorithmIdentifier
+	EncryptedHash           windows.CryptDataBlob
+	AuthAttrs               cryptAttributes
+	UnauthAttrs             cryptAttributes
+}
+
+type cryptAttributes struct {
+	Count      uint32
+	Attributes unsafe.Pointer
+}
+
 type OSVERSIONINFOEXW struct {
 	dwOSVersionInfoSize uint32
 	dwMajorVersion      uint32
@@ -99,6 +124,12 @@ func DoUpgrade(interactive bool) error {
 		return fmt.Errorf("failed to lookup downloads")
 	}
 
+	if err := VerifyDownload(); err != nil {
+		_ = os.Remove(bundle)
+		slog.Warn("verification failure", "bundle", bundle, "error", err)
+		return fmt.Errorf("staged update verification failed: %w", err)
+	}
+
 	// We move the installer to ensure we don't race with multiple apps starting in quick succession
 	if err := os.Rename(bundle, runningInstaller); err != nil {
 		return fmt.Errorf("unable to rename %s -> %s : %w", bundle, runningInstaller, err)
@@ -184,6 +215,150 @@ func DoPostUpgradeCleanup() error {
 }
 
 func verifyDownload() error {
+	bundle := getStagedUpdate()
+	if bundle == "" {
+		return fmt.Errorf("failed to lookup downloads")
+	}
+	slog.Debug("verifying update", "bundle", bundle)
+
+	if err := verifyWindowsInstallerSignature(bundle); err != nil {
+		return fmt.Errorf("signature verification failed: %w", err)
+	}
+	return nil
+}
+
+func verifyWindowsInstallerSignature(filename string) error {
+	filename16, err := windows.UTF16PtrFromString(filename)
+	if err != nil {
+		return err
+	}
+
+	data := &windows.WinTrustData{
+		Size:             uint32(unsafe.Sizeof(windows.WinTrustData{})),
+		UIChoice:         windows.WTD_UI_NONE,
+		RevocationChecks: windows.WTD_REVOKE_WHOLECHAIN,
+		UnionChoice:      windows.WTD_CHOICE_FILE,
+		StateAction:      windows.WTD_STATEACTION_VERIFY,
+		UIContext:        windows.WTD_UICONTEXT_INSTALL,
+		FileOrCatalogOrBlobOrSgnrOrCert: unsafe.Pointer(&windows.WinTrustFileInfo{
+			Size:     uint32(unsafe.Sizeof(windows.WinTrustFileInfo{})),
+			FilePath: filename16,
+		}),
+	}
+
+	verifyErr := windows.WinVerifyTrustEx(windows.InvalidHWND, &windows.WINTRUST_ACTION_GENERIC_VERIFY_V2, data)
+	data.StateAction = windows.WTD_STATEACTION_CLOSE
+	closeErr := windows.WinVerifyTrustEx(windows.InvalidHWND, &windows.WINTRUST_ACTION_GENERIC_VERIFY_V2, data)
+	if verifyErr != nil {
+		return verifyErr
+	}
+	if closeErr != nil {
+		return fmt.Errorf("close WinVerifyTrust state: %w", closeErr)
+	}
+
+	subject, err := windowsInstallerSignerSubject(filename)
+	if err != nil {
+		return err
+	}
+	slog.Debug("verified update signature", "subject", subject)
+	return nil
+}
+
+func windowsInstallerSignerSubject(filename string) (string, error) {
+	filename16, err := windows.UTF16PtrFromString(filename)
+	if err != nil {
+		return "", err
+	}
+
+	var certStore windows.Handle
+	var msg windows.Handle
+	if err := windows.CryptQueryObject(
+		windows.CERT_QUERY_OBJECT_FILE,
+		unsafe.Pointer(filename16),
+		windows.CERT_QUERY_CONTENT_FLAG_PKCS7_SIGNED_EMBED,
+		windows.CERT_QUERY_FORMAT_FLAG_BINARY,
+		0,
+		nil,
+		nil,
+		nil,
+		&certStore,
+		&msg,
+		nil,
+	); err != nil {
+		return "", err
+	}
+	defer windows.CertCloseStore(certStore, 0) //nolint:errcheck
+	defer cryptMsgClose(msg)                   //nolint:errcheck
+
+	var signerInfoSize uint32
+	if err := cryptMsgGetParam(msg, cmsgSignerInfoParam, 0, nil, &signerInfoSize); err != nil {
+		return "", err
+	}
+	if signerInfoSize == 0 {
+		return "", fmt.Errorf("missing signer info")
+	}
+
+	signerInfoBuf := make([]byte, signerInfoSize)
+	if err := cryptMsgGetParam(msg, cmsgSignerInfoParam, 0, unsafe.Pointer(&signerInfoBuf[0]), &signerInfoSize); err != nil {
+		return "", err
+	}
+	signerInfo := (*cmsgSignerInfo)(unsafe.Pointer(&signerInfoBuf[0]))
+	certInfo := windows.CertInfo{
+		Issuer:       signerInfo.Issuer,
+		SerialNumber: signerInfo.SerialNumber,
+	}
+
+	cert, err := windows.CertFindCertificateInStore(
+		certStore,
+		windows.X509_ASN_ENCODING|windows.PKCS_7_ASN_ENCODING,
+		0,
+		windows.CERT_FIND_SUBJECT_CERT,
+		unsafe.Pointer(&certInfo),
+		nil,
+	)
+	if err != nil {
+		return "", err
+	}
+	defer windows.CertFreeCertificateContext(cert) //nolint:errcheck
+
+	parsed, err := x509.ParseCertificate(unsafe.Slice(cert.EncodedCert, cert.Length))
+	if err != nil {
+		return "", err
+	}
+
+	for _, org := range parsed.Subject.Organization {
+		if org == "Ollama Inc." {
+			return parsed.Subject.String(), nil
+		}
+	}
+	return "", fmt.Errorf("unexpected signer: %s", parsed.Subject.String())
+}
+
+func cryptMsgGetParam(msg windows.Handle, paramType, index uint32, data unsafe.Pointer, size *uint32) error {
+	r1, _, e1 := procCryptMsgGetParam.Call(
+		uintptr(msg),
+		uintptr(paramType),
+		uintptr(index),
+		uintptr(data),
+		uintptr(unsafe.Pointer(size)),
+	)
+	if r1 == 0 {
+		if e1 != syscall.Errno(0) {
+			return e1
+		}
+		return syscall.EINVAL
+	}
+	return nil
+}
+
+func cryptMsgClose(msg windows.Handle) error {
+	r1, _, e1 := procCryptMsgClose.Call(uintptr(msg))
+	if r1 == 0 {
+		if e1 != syscall.Errno(0) {
+			return e1
+		}
+		return syscall.EINVAL
+	}
 	return nil
 }
 

app/updater/updater_windows_test.go

@@ -1,13 +1,85 @@
-//go:build windows || darwin
+//go:build windows
 
 package updater
 
 import (
 	"log/slog"
+	"os"
+	"path/filepath"
+	"strings"
 	"testing"
 )
 
+func TestVerifyDownloadRejectsUnsignedWindowsInstaller(t *testing.T) {
+	oldUpdateStageDir := UpdateStageDir
+	defer func() {
+		UpdateStageDir = oldUpdateStageDir
+	}()
+
+	t.Setenv("LOCALAPPDATA", t.TempDir())
+	UpdateStageDir = t.TempDir()
+	bundle := filepath.Join(UpdateStageDir, "etag", "OllamaSetup.exe")
+	if err := os.MkdirAll(filepath.Dir(bundle), 0o755); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(bundle, []byte("not a signed installer"), 0o755); err != nil {
+		t.Fatal(err)
+	}
+
+	err := verifyDownload()
+	if err == nil || !strings.Contains(err.Error(), "signature verification failed") {
+		t.Fatalf("expected signature verification failure, got %v", err)
+	}
+}
+
+func TestDoUpgradeAtStartupRejectsUnsignedWindowsInstaller(t *testing.T) {
+	oldUpdateStageDir := UpdateStageDir
+	oldRunningInstaller := runningInstaller
+	oldUpgradeLogFile := UpgradeLogFile
+	oldUpgradeMarkerFile := UpgradeMarkerFile
+	oldVerifyDownload := VerifyDownload
+	defer func() {
+		UpdateStageDir = oldUpdateStageDir
+		runningInstaller = oldRunningInstaller
+		UpgradeLogFile = oldUpgradeLogFile
+		UpgradeMarkerFile = oldUpgradeMarkerFile
+		VerifyDownload = oldVerifyDownload
+	}()
+
+	t.Setenv("LOCALAPPDATA", t.TempDir())
+	UpdateStageDir = t.TempDir()
+	runDir := t.TempDir()
+	runningInstaller = filepath.Join(runDir, "OllamaSetup.exe")
+	UpgradeLogFile = filepath.Join(runDir, "upgrade.log")
+	UpgradeMarkerFile = filepath.Join(runDir, "upgraded")
+	VerifyDownload = verifyDownload
+
+	bundle := filepath.Join(UpdateStageDir, "etag", "OllamaSetup.exe")
+	if err := os.MkdirAll(filepath.Dir(bundle), 0o755); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(bundle, []byte("not a signed installer"), 0o755); err != nil {
+		t.Fatal(err)
+	}
+
+	err := DoUpgradeAtStartup()
+	if err == nil || !strings.Contains(err.Error(), "signature verification failed") {
+		t.Fatalf("expected signature verification failure, got %v", err)
+	}
+	if _, err := os.Stat(runningInstaller); !os.IsNotExist(err) {
+		t.Fatalf("unsigned installer was moved before verification failed: %v", err)
+	}
+	if _, err := os.Stat(bundle); !os.IsNotExist(err) {
+		t.Fatalf("unsigned staged installer was not removed after verification failure: %v", err)
+	}
+}
+
 func TestIsInstallerRunning(t *testing.T) {
+	oldInstaller := Installer
+	defer func() {
+		Installer = oldInstaller
+	}()
+
 	slog.SetLogLoggerLevel(slog.LevelDebug)
 	Installer = "go.exe"
 	if !isInstallerRunning() {

cmake/local.cmake

@@ -0,0 +1,691 @@
+# Local Ollama superbuild targets.
+#
+# This file keeps the repository-root CMake project focused on orchestration:
+# it builds a runnable local Ollama payload by delegating llama.cpp work to the
+# llama/server CMake project and building the Go binary into a matching layout.
+
+include(ExternalProject)
+
+set(OLLAMA_LLAMA_BACKENDS "" CACHE STRING
+    "Semicolon-separated llama-server GPU backends to build: cuda_v12;cuda_v13;rocm_v7_1;rocm_v7_2;vulkan;cuda_jetpack5;cuda_jetpack6")
+set(_ollama_mlx_backends_doc "Semicolon-separated MLX backends to build: cuda_v13;metal_v3;metal_v4")
+set(OLLAMA_VERSION "0.0.0" CACHE STRING "Ollama version embedded in the local Go binary")
+set(OLLAMA_PAYLOAD_INSTALL_PREFIX "${CMAKE_BINARY_DIR}" CACHE PATH
+    "Build-time staging prefix for nested Ollama native payloads")
+
+string(REGEX REPLACE "^v" "" OLLAMA_VERSION "${OLLAMA_VERSION}")
+
+set(OLLAMA_NATIVE_CONFIG_ARG)
+if(CMAKE_CONFIGURATION_TYPES)
+    set(OLLAMA_NATIVE_CONFIG_ARG --config Release)
+endif()
+
+set(OLLAMA_NATIVE_EXTERNAL_OPTIONS)
+if(CMAKE_VERSION VERSION_GREATER_EQUAL 3.28)
+    list(APPEND OLLAMA_NATIVE_EXTERNAL_OPTIONS BUILD_JOB_SERVER_AWARE TRUE)
+endif()
+
+function(ollama_check_metal_toolchain output_version)
+    find_program(_ollama_xcrun xcrun)
+    if(NOT _ollama_xcrun)
+        message(FATAL_ERROR
+            "MLX Metal requires Xcode command line tools. Install Xcode, run "
+            "`sudo xcode-select -s /Applications/Xcode.app/Contents/Developer`, "
+            "then install the Metal toolchain with "
+            "`xcodebuild -downloadComponent MetalToolchain`.")
+    endif()
+
+    execute_process(
+        COMMAND zsh "-c"
+            "echo \"__METAL_VERSION__\" | \"${_ollama_xcrun}\" -sdk macosx metal -E -x metal -P - 2>/dev/null | tail -1 | tr -d '\n'"
+        OUTPUT_VARIABLE _metal_version
+        RESULT_VARIABLE _metal_result)
+    if(NOT _metal_result EQUAL 0 OR NOT _metal_version MATCHES "^[0-9]+$")
+        message(FATAL_ERROR
+            "MLX Metal requires Xcode's Metal toolchain. Install Xcode, run "
+            "`sudo xcode-select -s /Applications/Xcode.app/Contents/Developer`, "
+            "then install the Metal toolchain with "
+            "`xcodebuild -downloadComponent MetalToolchain`.")
+    endif()
+
+    set(${output_version} "${_metal_version}" PARENT_SCOPE)
+endfunction()
+
+function(ollama_macos_major_version output)
+    execute_process(
+        COMMAND sw_vers -productVersion
+        OUTPUT_VARIABLE _macos_version
+        OUTPUT_STRIP_TRAILING_WHITESPACE
+        RESULT_VARIABLE _macos_result
+        ERROR_QUIET)
+    if(_macos_result EQUAL 0)
+        string(REGEX MATCH "^[0-9]+" _macos_major "${_macos_version}")
+    endif()
+    set(${output} "${_macos_major}" PARENT_SCOPE)
+endfunction()
+
+function(ollama_macos_sdk_major_version output)
+    execute_process(
+        COMMAND xcrun --sdk macosx --show-sdk-version
+        OUTPUT_VARIABLE _sdk_version
+        OUTPUT_STRIP_TRAILING_WHITESPACE
+        RESULT_VARIABLE _sdk_result
+        ERROR_QUIET)
+    if(_sdk_result EQUAL 0)
+        string(REGEX MATCH "^[0-9]+" _sdk_major "${_sdk_version}")
+    endif()
+    set(${output} "${_sdk_major}" PARENT_SCOPE)
+endfunction()
+
+function(ollama_default_mlx_backends output)
+    set(_backends "")
+    if(APPLE AND CMAKE_SYSTEM_PROCESSOR STREQUAL "arm64")
+        ollama_check_metal_toolchain(_metal_version)
+        ollama_macos_major_version(_macos_major)
+        ollama_macos_sdk_major_version(_sdk_major)
+        if(_macos_major AND _sdk_major AND _macos_major GREATER_EQUAL 26 AND _sdk_major GREATER_EQUAL 26)
+            set(_backends "metal_v4")
+        else()
+            set(_backends "metal_v3")
+        endif()
+        message(STATUS "Defaulting OLLAMA_MLX_BACKENDS=${_backends} for macOS arm64")
+    endif()
+    set(${output} "${_backends}" PARENT_SCOPE)
+endfunction()
+
+if(NOT DEFINED OLLAMA_MLX_BACKENDS)
+    ollama_default_mlx_backends(_ollama_default_mlx_backends)
+    set(OLLAMA_MLX_BACKENDS "${_ollama_default_mlx_backends}" CACHE STRING "${_ollama_mlx_backends_doc}")
+else()
+    set(OLLAMA_MLX_BACKENDS "${OLLAMA_MLX_BACKENDS}" CACHE STRING "${_ollama_mlx_backends_doc}")
+endif()
+
+if(NOT OLLAMA_HAVE_LLAMA_SERVER)
+    if(OLLAMA_LLAMA_BACKENDS)
+        message(FATAL_ERROR "llama/server is required when OLLAMA_LLAMA_BACKENDS is set")
+    endif()
+    if(NOT OLLAMA_MLX_BACKENDS)
+        message(FATAL_ERROR "llama/server is required for local Ollama builds")
+    endif()
+else()
+    file(READ "${CMAKE_SOURCE_DIR}/LLAMA_CPP_VERSION" OLLAMA_LLAMA_CPP_GIT_TAG)
+    string(STRIP "${OLLAMA_LLAMA_CPP_GIT_TAG}" OLLAMA_LLAMA_CPP_GIT_TAG)
+    include(${CMAKE_SOURCE_DIR}/llama/compat/compat.cmake)
+    if(DEFINED FETCHCONTENT_SOURCE_DIR_LLAMA_CPP AND NOT "${FETCHCONTENT_SOURCE_DIR_LLAMA_CPP}" STREQUAL "")
+        get_filename_component(OLLAMA_LLAMA_CPP_SOURCE_DIR
+            "${FETCHCONTENT_SOURCE_DIR_LLAMA_CPP}" ABSOLUTE BASE_DIR "${CMAKE_SOURCE_DIR}")
+        message(STATUS "Using llama.cpp source override: ${OLLAMA_LLAMA_CPP_SOURCE_DIR}")
+        add_custom_target(ollama-llama-cpp-source)
+    elseif(DEFINED ENV{OLLAMA_LLAMA_CPP_SOURCE})
+        get_filename_component(OLLAMA_LLAMA_CPP_SOURCE_DIR
+            "$ENV{OLLAMA_LLAMA_CPP_SOURCE}" ABSOLUTE BASE_DIR "${CMAKE_SOURCE_DIR}")
+        message(STATUS "Using local llama.cpp source: ${OLLAMA_LLAMA_CPP_SOURCE_DIR}")
+        add_custom_target(ollama-llama-cpp-source)
+    else()
+        set(OLLAMA_LLAMA_CPP_SOURCE_DIR "${CMAKE_BINARY_DIR}/_deps/llama_cpp-src")
+        ExternalProject_Add(ollama-llama-cpp-source
+            GIT_REPOSITORY "https://github.com/ggml-org/llama.cpp.git"
+            GIT_TAG ${OLLAMA_LLAMA_CPP_GIT_TAG}
+            GIT_SHALLOW TRUE
+            SOURCE_DIR ${OLLAMA_LLAMA_CPP_SOURCE_DIR}
+            CONFIGURE_COMMAND ""
+            BUILD_COMMAND ""
+            INSTALL_COMMAND ""
+            PATCH_COMMAND ${OLLAMA_LLAMA_CPP_COMPAT_PATCH_COMMAND}
+            USES_TERMINAL_DOWNLOAD TRUE
+            USES_TERMINAL_PATCH TRUE)
+    endif()
+endif()
+
+set(_mlx_source_targets)
+if(OLLAMA_MLX_BACKENDS)
+    file(READ "${CMAKE_SOURCE_DIR}/MLX_VERSION" OLLAMA_MLX_GIT_TAG)
+    string(STRIP "${OLLAMA_MLX_GIT_TAG}" OLLAMA_MLX_GIT_TAG)
+    file(READ "${CMAKE_SOURCE_DIR}/MLX_C_VERSION" OLLAMA_MLX_C_GIT_TAG)
+    string(STRIP "${OLLAMA_MLX_C_GIT_TAG}" OLLAMA_MLX_C_GIT_TAG)
+
+    if(DEFINED FETCHCONTENT_SOURCE_DIR_MLX AND NOT "${FETCHCONTENT_SOURCE_DIR_MLX}" STREQUAL "")
+        get_filename_component(OLLAMA_MLX_SOURCE_DIR
+            "${FETCHCONTENT_SOURCE_DIR_MLX}" ABSOLUTE BASE_DIR "${CMAKE_SOURCE_DIR}")
+        message(STATUS "Using MLX source override: ${OLLAMA_MLX_SOURCE_DIR}")
+    elseif(DEFINED ENV{OLLAMA_MLX_SOURCE})
+        get_filename_component(OLLAMA_MLX_SOURCE_DIR
+            "$ENV{OLLAMA_MLX_SOURCE}" ABSOLUTE BASE_DIR "${CMAKE_SOURCE_DIR}")
+        message(STATUS "Using local MLX source: ${OLLAMA_MLX_SOURCE_DIR}")
+    else()
+        set(OLLAMA_MLX_SOURCE_DIR "${CMAKE_BINARY_DIR}/_deps/mlx-src")
+        ExternalProject_Add(ollama-mlx-source
+            GIT_REPOSITORY "https://github.com/ml-explore/mlx.git"
+            GIT_TAG ${OLLAMA_MLX_GIT_TAG}
+            # MLX uses commit hashes while we track closely; switch to shallow when MLX pins move to tags.
+            GIT_SHALLOW FALSE
+            SOURCE_DIR ${OLLAMA_MLX_SOURCE_DIR}
+            CONFIGURE_COMMAND ""
+            BUILD_COMMAND ""
+            INSTALL_COMMAND ""
+            USES_TERMINAL_DOWNLOAD TRUE)
+        list(APPEND _mlx_source_targets ollama-mlx-source)
+    endif()
+
+    if(DEFINED "FETCHCONTENT_SOURCE_DIR_MLX-C" AND NOT "${FETCHCONTENT_SOURCE_DIR_MLX-C}" STREQUAL "")
+        get_filename_component(OLLAMA_MLX_C_SOURCE_DIR
+            "${FETCHCONTENT_SOURCE_DIR_MLX-C}" ABSOLUTE BASE_DIR "${CMAKE_SOURCE_DIR}")
+        message(STATUS "Using MLX-C source override: ${OLLAMA_MLX_C_SOURCE_DIR}")
+    elseif(DEFINED ENV{OLLAMA_MLX_C_SOURCE})
+        get_filename_component(OLLAMA_MLX_C_SOURCE_DIR
+            "$ENV{OLLAMA_MLX_C_SOURCE}" ABSOLUTE BASE_DIR "${CMAKE_SOURCE_DIR}")
+        message(STATUS "Using local MLX-C source: ${OLLAMA_MLX_C_SOURCE_DIR}")
+    else()
+        set(OLLAMA_MLX_C_SOURCE_DIR "${CMAKE_BINARY_DIR}/_deps/mlx-c-src")
+        ExternalProject_Add(ollama-mlx-c-source
+            GIT_REPOSITORY "https://github.com/ml-explore/mlx-c.git"
+            GIT_TAG ${OLLAMA_MLX_C_GIT_TAG}
+            # MLX-C uses commit hashes while we track closely; switch to shallow when MLX-C pins move to tags.
+            GIT_SHALLOW FALSE
+            SOURCE_DIR ${OLLAMA_MLX_C_SOURCE_DIR}
+            CONFIGURE_COMMAND ""
+            BUILD_COMMAND ""
+            INSTALL_COMMAND ""
+            USES_TERMINAL_DOWNLOAD TRUE)
+        list(APPEND _mlx_source_targets ollama-mlx-c-source)
+    endif()
+    add_custom_target(ollama-mlx-sources DEPENDS ${_mlx_source_targets})
+endif()
+
+set(OLLAMA_NATIVE_BUILD_TOOL_COMMAND
+    ${CMAKE_COMMAND} --build <BINARY_DIR>)
+set(OLLAMA_NATIVE_BUILD_TARGET_ARG --target)
+if(CMAKE_GENERATOR MATCHES "Makefiles")
+    set(OLLAMA_NATIVE_BUILD_TOOL_COMMAND
+        "$(MAKE)" -C <BINARY_DIR>)
+    set(OLLAMA_NATIVE_BUILD_TARGET_ARG)
+endif()
+
+function(ollama_escape_cmake_list input output)
+    string(REPLACE ";" "|" _escaped "${input}")
+    set(${output} "${_escaped}" PARENT_SCOPE)
+endfunction()
+
+function(ollama_collect_cache_args_with_prefix prefix output)
+    get_cmake_property(_cache_variables CACHE_VARIABLES)
+    list(SORT _cache_variables)
+
+    set(_args)
+    foreach(_var IN LISTS _cache_variables)
+        if(_var MATCHES "^${prefix}")
+            ollama_escape_cmake_list("${${_var}}" _value)
+            list(APPEND _args "-D${_var}=${_value}")
+        endif()
+    endforeach()
+
+    set(${output} "${_args}" PARENT_SCOPE)
+endfunction()
+
+function(ollama_append_cache_arg_if_set output name)
+    if(DEFINED ${name} AND NOT "${${name}}" STREQUAL "")
+        ollama_escape_cmake_list("${${name}}" _value)
+        set(${output} ${${output}} "-D${name}=${_value}" PARENT_SCOPE)
+    endif()
+endfunction()
+
+function(ollama_cache_arg_is_set name output)
+    if(DEFINED ${name} AND NOT "${${name}}" STREQUAL "")
+        set(${output} TRUE PARENT_SCOPE)
+    else()
+        set(${output} FALSE PARENT_SCOPE)
+    endif()
+endfunction()
+
+function(ollama_llama_cuda_preset backend output)
+    ollama_cache_arg_is_set(CMAKE_CUDA_ARCHITECTURES _has_cuda_arch)
+    if(_has_cuda_arch)
+        set(_preset "llama_${backend}_user_arch")
+    elseif(WIN32)
+        set(_preset "llama_${backend}_windows")
+    else()
+        set(_preset "llama_${backend}_linux")
+    endif()
+    set(${output} "${_preset}" PARENT_SCOPE)
+endfunction()
+
+function(ollama_mlx_cuda_preset output)
+    ollama_cache_arg_is_set(MLX_CUDA_ARCHITECTURES _has_mlx_arch)
+    ollama_cache_arg_is_set(CMAKE_CUDA_ARCHITECTURES _has_cuda_arch)
+    if(_has_mlx_arch OR _has_cuda_arch)
+        set(_preset "mlx_cuda_v13_user_arch")
+    elseif(WIN32)
+        set(_preset "mlx_cuda_v13_windows")
+    else()
+        set(_preset "mlx_cuda_v13_linux")
+    endif()
+    set(${output} "${_preset}" PARENT_SCOPE)
+endfunction()
+
+function(ollama_rocm_preset backend output)
+    ollama_cache_arg_is_set(AMDGPU_TARGETS _has_amdgpu_targets)
+    ollama_cache_arg_is_set(CMAKE_HIP_ARCHITECTURES _has_hip_arch)
+    if(_has_amdgpu_targets OR _has_hip_arch)
+        if(backend STREQUAL "rocm_v7_1" AND NOT WIN32)
+            message(FATAL_ERROR "OLLAMA_LLAMA_BACKENDS=rocm_v7_1 is only supported for Windows ROCm builds")
+        elseif(backend STREQUAL "rocm_v7_2" AND WIN32)
+            message(FATAL_ERROR "OLLAMA_LLAMA_BACKENDS=rocm_v7_2 is only supported for Linux ROCm builds")
+        endif()
+    elseif(backend STREQUAL "rocm_v7_1")
+        if(NOT WIN32)
+            message(FATAL_ERROR "OLLAMA_LLAMA_BACKENDS=rocm_v7_1 is only supported for Windows ROCm builds")
+        endif()
+        set(_preset "${backend}_windows")
+    elseif(backend STREQUAL "rocm_v7_2")
+        if(WIN32)
+            message(FATAL_ERROR "OLLAMA_LLAMA_BACKENDS=rocm_v7_2 is only supported for Linux ROCm builds")
+        endif()
+        set(_preset "${backend}_linux")
+    else()
+        message(FATAL_ERROR "Unknown ROCm backend '${backend}'")
+    endif()
+    if(_has_amdgpu_targets OR _has_hip_arch)
+        set(_preset "${backend}_user_arch")
+    endif()
+    set(${output} "${_preset}" PARENT_SCOPE)
+endfunction()
+
+function(ollama_add_llama_server_build name)
+    cmake_parse_arguments(ARG "" "PRESET;RUNNER_DIR" "TARGETS;CMAKE_ARGS" ${ARGN})
+    if(NOT ARG_TARGETS)
+        message(FATAL_ERROR "ollama_add_llama_server_build(${name}) requires TARGETS")
+    endif()
+
+    if(WIN32 AND name STREQUAL "vulkan")
+        # The Vulkan shader generator nests deeply enough to hit Windows MAX_PATH.
+        set(_build_dir ${CMAKE_BINARY_DIR}/ls-vk)
+    else()
+        set(_build_dir ${CMAKE_BINARY_DIR}/llama-server-${name})
+    endif()
+    ollama_collect_cache_args_with_prefix("GGML_" _ggml_cache_args)
+    ollama_collect_cache_args_with_prefix("LLAMA_" _llama_cache_args)
+    set(_cmake_args
+        -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE}
+        -DCMAKE_INSTALL_PREFIX=${OLLAMA_PAYLOAD_INSTALL_PREFIX}
+        -DOLLAMA_LIB_DIR:STRING=${OLLAMA_LIB_DIR}
+        -DOLLAMA_RUNNER_DIR=${ARG_RUNNER_DIR}
+        -DFETCHCONTENT_SOURCE_DIR_LLAMA_CPP=${OLLAMA_LLAMA_CPP_SOURCE_DIR}
+        -DOLLAMA_LLAMA_CPP_SKIP_COMPAT_PATCH=ON
+        -DGGML_NATIVE=OFF
+        -DGGML_OPENMP=OFF
+        ${ARG_CMAKE_ARGS}
+        ${_ggml_cache_args}
+        ${_llama_cache_args}
+    )
+
+    if(APPLE)
+        if(CMAKE_OSX_ARCHITECTURES)
+            list(APPEND _cmake_args
+                -DCMAKE_OSX_ARCHITECTURES=${CMAKE_OSX_ARCHITECTURES})
+        endif()
+        if(CMAKE_OSX_DEPLOYMENT_TARGET)
+            list(APPEND _cmake_args
+                -DCMAKE_OSX_DEPLOYMENT_TARGET=${CMAKE_OSX_DEPLOYMENT_TARGET})
+        endif()
+    endif()
+    set(_configure_command ${CMAKE_COMMAND}
+        -S ${CMAKE_SOURCE_DIR}/llama/server
+        -B <BINARY_DIR>
+        ${_cmake_args})
+    if(ARG_PRESET)
+        set(_configure_command ${CMAKE_COMMAND}
+            -S ${CMAKE_SOURCE_DIR}/llama/server
+            --preset ${ARG_PRESET}
+            -B <BINARY_DIR>
+            ${_cmake_args})
+    endif()
+    ExternalProject_Add(ollama-llama-server-${name}
+        SOURCE_DIR ${CMAKE_SOURCE_DIR}/llama/server
+        BINARY_DIR ${_build_dir}
+        CONFIGURE_COMMAND ${_configure_command}
+        BUILD_COMMAND ${OLLAMA_NATIVE_BUILD_TOOL_COMMAND}
+            ${OLLAMA_NATIVE_CONFIG_ARG}
+            ${OLLAMA_NATIVE_BUILD_TARGET_ARG} ${ARG_TARGETS}
+        INSTALL_COMMAND ${CMAKE_COMMAND} --install <BINARY_DIR>
+            ${OLLAMA_NATIVE_CONFIG_ARG}
+            --component llama-server
+        DEPENDS ollama-llama-cpp-source
+        LIST_SEPARATOR |
+        # ExternalProject cannot reliably infer when nested FetchContent
+        # sources, compat patches, or forwarded GGML/LLAMA cache settings need
+        # a rebuild. Always entering the sub-build keeps direct `cmake --build`
+        # iteration correct; the nested generator still performs incremental
+        # compilation.
+        BUILD_ALWAYS TRUE
+        ${OLLAMA_NATIVE_EXTERNAL_OPTIONS}
+        USES_TERMINAL_CONFIGURE TRUE
+        USES_TERMINAL_BUILD TRUE
+        USES_TERMINAL_INSTALL TRUE)
+endfunction()
+
+function(ollama_add_mlx_build name)
+    cmake_parse_arguments(ARG "" "PRESET;RUNNER_DIR" "CMAKE_ARGS" ${ARGN})
+    if(NOT ARG_RUNNER_DIR)
+        message(FATAL_ERROR "ollama_add_mlx_build(${name}) requires RUNNER_DIR")
+    endif()
+
+    set(_build_dir ${CMAKE_BINARY_DIR}/${ARG_RUNNER_DIR})
+    ollama_collect_cache_args_with_prefix("MLX_" _mlx_cache_args)
+    set(_cmake_args
+        -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE}
+        -DCMAKE_INSTALL_PREFIX=${OLLAMA_PAYLOAD_INSTALL_PREFIX}
+        -DOLLAMA_LIB_DIR:STRING=${OLLAMA_LIB_DIR}
+        -DOLLAMA_RUNNER_DIR=${ARG_RUNNER_DIR}
+        -DOLLAMA_SOURCE_DIR=${CMAKE_SOURCE_DIR}
+        -DFETCHCONTENT_SOURCE_DIR_MLX=${OLLAMA_MLX_SOURCE_DIR}
+        -DFETCHCONTENT_SOURCE_DIR_MLX-C=${OLLAMA_MLX_C_SOURCE_DIR}
+        -DOLLAMA_MLX_GENERATE_WRAPPERS=OFF
+        ${ARG_CMAKE_ARGS}
+        ${_mlx_cache_args}
+    )
+    foreach(_arg IN ITEMS
+            BLAS_INCLUDE_DIRS
+            LAPACK_INCLUDE_DIRS
+            CUDAToolkit_ROOT
+            CUDNN_ROOT_DIR
+            CUDNN_INCLUDE_PATH
+            CUDNN_LIBRARY_PATH
+            CMAKE_CUDA_COMPILER
+            CMAKE_CUDA_HOST_COMPILER
+            CMAKE_INCLUDE_PATH
+            CMAKE_LIBRARY_PATH
+            CMAKE_PREFIX_PATH)
+        ollama_append_cache_arg_if_set(_cmake_args ${_arg})
+    endforeach()
+
+    if(APPLE)
+        if(CMAKE_OSX_ARCHITECTURES)
+            list(APPEND _cmake_args
+                -DCMAKE_OSX_ARCHITECTURES=${CMAKE_OSX_ARCHITECTURES})
+        endif()
+    endif()
+    set(_configure_command ${CMAKE_COMMAND}
+        -S ${CMAKE_SOURCE_DIR}/cmake/mlx
+        -B <BINARY_DIR>
+        ${_cmake_args})
+    if(ARG_PRESET)
+        set(_configure_command ${CMAKE_COMMAND}
+            -S ${CMAKE_SOURCE_DIR}/cmake/mlx
+            --preset ${ARG_PRESET}
+            -B <BINARY_DIR>
+            ${_cmake_args})
+    endif()
+
+    ExternalProject_Add(ollama-mlx-${name}
+        SOURCE_DIR ${CMAKE_SOURCE_DIR}/cmake/mlx
+        BINARY_DIR ${_build_dir}
+        CONFIGURE_COMMAND ${_configure_command}
+        BUILD_COMMAND ${OLLAMA_NATIVE_BUILD_TOOL_COMMAND}
+            ${OLLAMA_NATIVE_CONFIG_ARG}
+            ${OLLAMA_NATIVE_BUILD_TARGET_ARG} mlx
+            ${OLLAMA_NATIVE_BUILD_TARGET_ARG} mlxc
+        INSTALL_COMMAND ${CMAKE_COMMAND} --install <BINARY_DIR>
+            ${OLLAMA_NATIVE_CONFIG_ARG}
+            --component MLX
+            COMMAND ${CMAKE_COMMAND} --install <BINARY_DIR>
+            ${OLLAMA_NATIVE_CONFIG_ARG}
+            --component MLX_VENDOR
+        DEPENDS ollama-mlx-sources
+        LIST_SEPARATOR |
+        BUILD_ALWAYS TRUE
+        ${OLLAMA_NATIVE_EXTERNAL_OPTIONS}
+        USES_TERMINAL_CONFIGURE TRUE
+        USES_TERMINAL_BUILD TRUE
+        USES_TERMINAL_INSTALL TRUE)
+endfunction()
+
+find_program(GO_EXECUTABLE go)
+
+if(OLLAMA_MLX_BACKENDS)
+    set(_mlx_c_headers_dir "${OLLAMA_MLX_C_SOURCE_DIR}/mlx/c")
+    set(_mlx_c_headers_dest "${CMAKE_SOURCE_DIR}/x/mlxrunner/mlx/include/mlx/c")
+
+    if(GO_EXECUTABLE AND (NOT APPLE OR CMAKE_SYSTEM_PROCESSOR STREQUAL CMAKE_HOST_SYSTEM_PROCESSOR))
+        add_custom_target(ollama-mlx-generate-wrappers
+            COMMAND ${CMAKE_COMMAND}
+                -DMLX_C_HEADERS_DIR=${_mlx_c_headers_dir}
+                -DMLX_C_HEADERS_DEST=${_mlx_c_headers_dest}
+                -P "${CMAKE_SOURCE_DIR}/cmake/vendor-mlx-c-headers.cmake"
+            COMMAND ${CMAKE_COMMAND} -E env
+                CC= CGO_CFLAGS= CGO_CXXFLAGS=
+                ${GO_EXECUTABLE} generate ./x/...
+            WORKING_DIRECTORY ${CMAKE_SOURCE_DIR}
+            DEPENDS ollama-mlx-sources
+            COMMENT "Regenerating MLX Go wrappers"
+            VERBATIM)
+    else()
+        add_custom_target(ollama-mlx-generate-wrappers
+            COMMAND ${CMAKE_COMMAND} -E echo
+                "Cannot regenerate MLX wrappers while Go is unavailable or while cross-compiling"
+            COMMAND ${CMAKE_COMMAND} -E false
+            DEPENDS ollama-mlx-sources
+            VERBATIM)
+    endif()
+endif()
+
+if(OLLAMA_HAVE_LLAMA_SERVER)
+    if(NOT OLLAMA_GO_OUTPUT)
+        if(WIN32)
+            set(OLLAMA_GO_OUTPUT ${CMAKE_SOURCE_DIR}/ollama.exe)
+        else()
+            set(OLLAMA_GO_OUTPUT ${CMAKE_SOURCE_DIR}/ollama)
+        endif()
+    endif()
+    if(NOT IS_ABSOLUTE "${OLLAMA_GO_OUTPUT}")
+        set(OLLAMA_GO_OUTPUT "${CMAKE_SOURCE_DIR}/${OLLAMA_GO_OUTPUT}")
+    endif()
+    get_filename_component(OLLAMA_GO_OUTPUT "${OLLAMA_GO_OUTPUT}" ABSOLUTE)
+    set(OLLAMA_GO_OUTPUT "${OLLAMA_GO_OUTPUT}" CACHE FILEPATH "Output path for the local Ollama Go binary")
+    get_filename_component(OLLAMA_GO_OUTPUT_DIR "${OLLAMA_GO_OUTPUT}" DIRECTORY)
+
+    set(OLLAMA_GO_LDFLAGS
+        "-s -w -X=github.com/ollama/ollama/version.Version=${OLLAMA_VERSION} -X=github.com/ollama/ollama/server.mode=release")
+    if(GO_EXECUTABLE)
+        add_custom_target(ollama-go ALL
+            COMMAND ${CMAKE_COMMAND} -E make_directory "${OLLAMA_GO_OUTPUT_DIR}"
+            COMMAND ${CMAKE_COMMAND} -E env CGO_ENABLED=1
+                ${GO_EXECUTABLE} build -trimpath -ldflags "${OLLAMA_GO_LDFLAGS}" -o "${OLLAMA_GO_OUTPUT}" .
+            WORKING_DIRECTORY ${CMAKE_SOURCE_DIR}
+            BYPRODUCTS ${OLLAMA_GO_OUTPUT}
+            COMMENT "Building Ollama Go binary"
+            VERBATIM)
+    else()
+        add_custom_target(ollama-go ALL
+            COMMAND ${CMAKE_COMMAND} -E echo
+                "Go executable not found. Install Go or set GO_EXECUTABLE to build the local Ollama binary."
+            COMMAND ${CMAKE_COMMAND} -E false
+            COMMENT "Building Ollama Go binary"
+            VERBATIM)
+    endif()
+
+    set(_cpu_args)
+    if(APPLE AND CMAKE_SYSTEM_PROCESSOR STREQUAL "arm64")
+        list(APPEND _cpu_args
+            -DBUILD_SHARED_LIBS=OFF
+            -DGGML_BACKEND_DL=OFF
+            -DGGML_METAL=ON
+            -DGGML_METAL_EMBED_LIBRARY=ON)
+    else()
+        list(APPEND _cpu_args
+            -DBUILD_SHARED_LIBS=ON
+            -DGGML_BACKEND_DL=ON
+            -DGGML_CPU_ALL_VARIANTS=ON)
+        if(WIN32)
+            list(APPEND _cpu_args -DGGML_OPENMP=ON)
+        endif()
+        if(APPLE)
+            list(APPEND _cpu_args -DGGML_METAL=OFF)
+        endif()
+    endif()
+
+    ollama_add_llama_server_build(local
+        RUNNER_DIR ""
+        TARGETS llama-server llama-quantize
+        CMAKE_ARGS ${_cpu_args})
+
+    add_custom_target(ollama-local ALL
+        DEPENDS ollama-go ollama-llama-server-local
+        COMMENT "Building local Ollama payload")
+
+    install(PROGRAMS "${OLLAMA_GO_OUTPUT}"
+        DESTINATION "${CMAKE_INSTALL_BINDIR}"
+        COMPONENT ollama-local)
+endif()
+
+set(_backend_targets)
+if(OLLAMA_HAVE_LLAMA_SERVER)
+    foreach(_backend IN LISTS OLLAMA_LLAMA_BACKENDS)
+        if(_backend STREQUAL "cuda_v12")
+            ollama_llama_cuda_preset(${_backend} _cuda_preset)
+            set(_cuda_args)
+            ollama_append_cache_arg_if_set(_cuda_args CMAKE_CUDA_ARCHITECTURES)
+            ollama_append_cache_arg_if_set(_cuda_args CMAKE_CUDA_FLAGS)
+            ollama_add_llama_server_build(${_backend}
+                PRESET ${_cuda_preset}
+                RUNNER_DIR ${_backend}
+                TARGETS ggml-cuda
+                CMAKE_ARGS ${_cuda_args})
+            list(APPEND _backend_targets ollama-llama-server-${_backend})
+        elseif(_backend STREQUAL "cuda_v13")
+            ollama_llama_cuda_preset(${_backend} _cuda_preset)
+            set(_cuda_args)
+            ollama_append_cache_arg_if_set(_cuda_args CMAKE_CUDA_ARCHITECTURES)
+            ollama_append_cache_arg_if_set(_cuda_args CMAKE_CUDA_FLAGS)
+            ollama_add_llama_server_build(${_backend}
+                PRESET ${_cuda_preset}
+                RUNNER_DIR ${_backend}
+                TARGETS ggml-cuda
+                CMAKE_ARGS ${_cuda_args})
+            list(APPEND _backend_targets ollama-llama-server-${_backend})
+        elseif(_backend STREQUAL "rocm_v7_1" OR _backend STREQUAL "rocm_v7_2")
+            # ROCm 7.1 and 7.2 currently share build settings. Keep the backend
+            # names versioned so future packaging can install side-by-side ROCm
+            # payloads without changing the superbuild interface.
+            ollama_rocm_preset(${_backend} _rocm_preset)
+            set(_rocm_args
+                -DBUILD_SHARED_LIBS=ON
+                -DGGML_BACKEND_DL=ON
+                -DGGML_HIP=ON
+                -DCMAKE_HIP_PLATFORM=amd
+                -DOLLAMA_GPU_BACKEND=hip)
+            ollama_append_cache_arg_if_set(_rocm_args AMDGPU_TARGETS)
+            ollama_append_cache_arg_if_set(_rocm_args CMAKE_HIP_ARCHITECTURES)
+            ollama_append_cache_arg_if_set(_rocm_args CMAKE_HIP_FLAGS)
+            ollama_append_cache_arg_if_set(_rocm_args CMAKE_PREFIX_PATH)
+            ollama_add_llama_server_build(${_backend}
+                PRESET ${_rocm_preset}
+                RUNNER_DIR ${_backend}
+                TARGETS ggml-hip
+                CMAKE_ARGS ${_rocm_args})
+            list(APPEND _backend_targets ollama-llama-server-${_backend})
+        elseif(_backend STREQUAL "vulkan")
+            ollama_add_llama_server_build(vulkan
+                RUNNER_DIR vulkan
+                TARGETS ggml-vulkan
+                CMAKE_ARGS
+                    -DBUILD_SHARED_LIBS=ON
+                    -DGGML_BACKEND_DL=ON
+                    -DGGML_VULKAN=ON
+                    -DOLLAMA_GPU_BACKEND=vulkan)
+            list(APPEND _backend_targets ollama-llama-server-vulkan)
+        elseif(_backend STREQUAL "cuda_jetpack5")
+            if(CMAKE_CUDA_ARCHITECTURES)
+                set(_cuda_preset llama_cuda_jetpack5_user_arch)
+            else()
+                set(_cuda_preset llama_cuda_jetpack5)
+            endif()
+            set(_cuda_args)
+            ollama_append_cache_arg_if_set(_cuda_args CMAKE_CUDA_ARCHITECTURES)
+            ollama_append_cache_arg_if_set(_cuda_args CMAKE_CUDA_FLAGS)
+            ollama_add_llama_server_build(${_backend}
+                PRESET ${_cuda_preset}
+                RUNNER_DIR ${_backend}
+                TARGETS ggml-cuda
+                CMAKE_ARGS ${_cuda_args})
+            list(APPEND _backend_targets ollama-llama-server-${_backend})
+        elseif(_backend STREQUAL "cuda_jetpack6")
+            if(CMAKE_CUDA_ARCHITECTURES)
+                set(_cuda_preset llama_cuda_jetpack6_user_arch)
+            else()
+                set(_cuda_preset llama_cuda_jetpack6)
+            endif()
+            set(_cuda_args)
+            ollama_append_cache_arg_if_set(_cuda_args CMAKE_CUDA_ARCHITECTURES)
+            ollama_append_cache_arg_if_set(_cuda_args CMAKE_CUDA_FLAGS)
+            ollama_add_llama_server_build(${_backend}
+                PRESET ${_cuda_preset}
+                RUNNER_DIR ${_backend}
+                TARGETS ggml-cuda
+                CMAKE_ARGS ${_cuda_args})
+            list(APPEND _backend_targets ollama-llama-server-${_backend})
+        else()
+            message(FATAL_ERROR
+                "Unknown OLLAMA_LLAMA_BACKENDS entry '${_backend}'")
+        endif()
+    endforeach()
+endif()
+
+if(_backend_targets)
+    add_custom_target(ollama-llama-server-backends ALL
+        DEPENDS ${_backend_targets}
+        COMMENT "Building llama-server GPU backends")
+endif()
+
+set(_mlx_targets)
+foreach(_backend IN LISTS OLLAMA_MLX_BACKENDS)
+    if(_backend STREQUAL "cuda_v13")
+        ollama_mlx_cuda_preset(_mlx_cuda_preset)
+        set(_mlx_cuda_args)
+        ollama_append_cache_arg_if_set(_mlx_cuda_args CMAKE_CUDA_ARCHITECTURES)
+        ollama_append_cache_arg_if_set(_mlx_cuda_args MLX_CUDA_ARCHITECTURES)
+        ollama_append_cache_arg_if_set(_mlx_cuda_args CMAKE_CUDA_FLAGS)
+        ollama_add_mlx_build(cuda_v13
+            PRESET ${_mlx_cuda_preset}
+            RUNNER_DIR mlx_cuda_v13
+            CMAKE_ARGS ${_mlx_cuda_args})
+        list(APPEND _mlx_targets ollama-mlx-cuda_v13)
+    elseif(_backend STREQUAL "metal_v3")
+        if(NOT APPLE)
+            message(FATAL_ERROR "OLLAMA_MLX_BACKENDS=metal_v3 is only supported on macOS")
+        endif()
+        ollama_check_metal_toolchain(_metal_version)
+        ollama_add_mlx_build(metal_v3
+            PRESET mlx_metal_v3
+            RUNNER_DIR mlx_metal_v3)
+        list(APPEND _mlx_targets ollama-mlx-metal_v3)
+    elseif(_backend STREQUAL "metal_v4")
+        if(NOT APPLE)
+            message(FATAL_ERROR "OLLAMA_MLX_BACKENDS=metal_v4 is only supported on macOS")
+        endif()
+        ollama_check_metal_toolchain(_metal_version)
+        ollama_macos_sdk_major_version(_ollama_mlx_sdk_major)
+        if(_ollama_mlx_sdk_major AND _ollama_mlx_sdk_major GREATER_EQUAL 26)
+            ollama_add_mlx_build(metal_v4
+                PRESET mlx_metal_v4
+                RUNNER_DIR mlx_metal_v4)
+            list(APPEND _mlx_targets ollama-mlx-metal_v4)
+        else()
+            message(FATAL_ERROR
+                "OLLAMA_MLX_BACKENDS=metal_v4 requires the macOS 26 SDK. "
+                "Install a newer Xcode or use OLLAMA_MLX_BACKENDS=metal_v3.")
+        endif()
+    else()
+        message(FATAL_ERROR
+            "Unknown OLLAMA_MLX_BACKENDS entry '${_backend}'")
+    endif()
+endforeach()
+
+if(_mlx_targets)
+    add_custom_target(ollama-mlx-backends ALL
+        DEPENDS ${_mlx_targets}
+        COMMENT "Building MLX backends")
+endif()
+
+install(DIRECTORY "${OLLAMA_PAYLOAD_INSTALL_PREFIX}/${OLLAMA_LIB_DIR}/"
+    DESTINATION "${OLLAMA_LIB_DIR}"
+    COMPONENT ollama-local
+    USE_SOURCE_PERMISSIONS)

cmake/mlx/CMakeLists.txt

@@ -0,0 +1,235 @@
+cmake_minimum_required(VERSION 3.24)
+
+project(OllamaMLX C CXX)
+
+include(CheckLanguage)
+include(GNUInstallDirs)
+
+find_package(Threads REQUIRED)
+
+if(NOT CMAKE_CONFIGURATION_TYPES AND NOT CMAKE_BUILD_TYPE)
+    set(CMAKE_BUILD_TYPE Release CACHE STRING "Build type" FORCE)
+endif()
+
+if(NOT DEFINED BUILD_SHARED_LIBS)
+    set(BUILD_SHARED_LIBS ON)
+endif()
+
+set(CMAKE_CXX_STANDARD 17)
+set(CMAKE_CXX_STANDARD_REQUIRED ON)
+set(CMAKE_CXX_EXTENSIONS ON)
+
+if(APPLE)
+    set(CMAKE_BUILD_RPATH "@loader_path")
+    set(CMAKE_INSTALL_RPATH "@loader_path")
+    set(CMAKE_BUILD_WITH_INSTALL_RPATH ON)
+endif()
+
+if(NOT DEFINED OLLAMA_SOURCE_DIR OR "${OLLAMA_SOURCE_DIR}" STREQUAL "")
+    get_filename_component(OLLAMA_SOURCE_DIR "${CMAKE_CURRENT_LIST_DIR}/../.." ABSOLUTE)
+endif()
+get_filename_component(OLLAMA_SOURCE_DIR "${OLLAMA_SOURCE_DIR}" ABSOLUTE BASE_DIR "${CMAKE_CURRENT_LIST_DIR}")
+set(OLLAMA_SOURCE_DIR "${OLLAMA_SOURCE_DIR}" CACHE PATH "Ollama repository root")
+
+set(OLLAMA_LIB_DIR "lib/ollama" CACHE STRING "Install destination for Ollama runtime payloads")
+set(OLLAMA_RUNNER_DIR "" CACHE STRING "Ollama runtime payload subdirectory")
+set(OLLAMA_BUILD_DIR ${CMAKE_BINARY_DIR}/lib/ollama)
+set(OLLAMA_INSTALL_DIR ${OLLAMA_LIB_DIR}/${OLLAMA_RUNNER_DIR})
+
+set(CMAKE_RUNTIME_OUTPUT_DIRECTORY         ${OLLAMA_BUILD_DIR})
+set(CMAKE_RUNTIME_OUTPUT_DIRECTORY_DEBUG   ${OLLAMA_BUILD_DIR})
+set(CMAKE_RUNTIME_OUTPUT_DIRECTORY_RELEASE ${OLLAMA_BUILD_DIR})
+set(CMAKE_LIBRARY_OUTPUT_DIRECTORY         ${OLLAMA_BUILD_DIR})
+set(CMAKE_LIBRARY_OUTPUT_DIRECTORY_DEBUG   ${OLLAMA_BUILD_DIR})
+set(CMAKE_LIBRARY_OUTPUT_DIRECTORY_RELEASE ${OLLAMA_BUILD_DIR})
+
+if(MLX_CUDA_ARCHITECTURES OR CMAKE_CUDA_ARCHITECTURES)
+    check_language(CUDA)
+endif()
+
+option(OLLAMA_MLX_GENERATE_WRAPPERS "Regenerate MLX Go wrappers" OFF)
+
+message(STATUS "Setting up MLX (this takes a while...)")
+add_subdirectory(${OLLAMA_SOURCE_DIR}/x/imagegen/mlx ${CMAKE_BINARY_DIR}/x/imagegen/mlx)
+
+# Find CUDA toolkit if MLX is built with CUDA support.
+find_package(CUDAToolkit)
+
+# Build list of directories for runtime dependency resolution.
+set(MLX_RUNTIME_DIRS ${CUDAToolkit_BIN_DIR} ${CUDAToolkit_BIN_DIR}/x64 ${CUDAToolkit_LIBRARY_DIR})
+# Add cuDNN bin paths for DLLs (Windows MLX CUDA builds).
+# CUDNN_ROOT_DIR is the standard CMake variable for cuDNN location.
+if(CUDNN_ROOT_DIR)
+    set(_cudnn_root "${CUDNN_ROOT_DIR}")
+elseif(DEFINED ENV{CUDNN_ROOT_DIR})
+    set(_cudnn_root "$ENV{CUDNN_ROOT_DIR}")
+endif()
+if(_cudnn_root)
+    # cuDNN 9.x has versioned subdirectories under bin/ (e.g., bin/13.0/).
+    file(GLOB CUDNN_BIN_SUBDIRS "${_cudnn_root}/bin/*")
+    list(APPEND MLX_RUNTIME_DIRS ${CUDNN_BIN_SUBDIRS})
+endif()
+# Add build output directory and MLX dependency build directories.
+list(APPEND MLX_RUNTIME_DIRS ${OLLAMA_BUILD_DIR})
+# OpenBLAS DLL location (pre-built zip extracts into openblas-src/bin/).
+list(APPEND MLX_RUNTIME_DIRS ${CMAKE_BINARY_DIR}/_deps/openblas-src/bin)
+# NCCL: on Linux, if real NCCL is found, cmake bundles libnccl.so via the
+# regex below. If NCCL is not found, MLX links a static stub (OBJECT lib)
+# so there is no runtime dependency. This path covers the stub build dir
+# for windows so we include the DLL in our dependencies.
+list(APPEND MLX_RUNTIME_DIRS ${CMAKE_BINARY_DIR}/_deps/mlx-build/mlx/distributed/nccl/nccl_stub-prefix/src/nccl_stub-build/Release)
+
+# Base regexes for runtime dependencies (cross-platform).
+set(MLX_INCLUDE_REGEXES cublas cublasLt cudart cufft nvrtc nvrtc-builtins cudnn nccl openblas gfortran)
+# On Windows, also include dl.dll (dlfcn-win32 POSIX emulation layer).
+if(WIN32)
+    list(APPEND MLX_INCLUDE_REGEXES "^dl\\.dll$")
+endif()
+
+# Keep mlx/mlxc targets separate from runtime dependencies so --strip only
+# applies to the binaries we build, not vendor DLLs/libs.
+install(TARGETS mlx mlxc
+    RUNTIME_DEPENDENCY_SET mlx_runtime_deps
+    RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT MLX
+    LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT MLX
+    FRAMEWORK DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT MLX
+)
+install(RUNTIME_DEPENDENCY_SET mlx_runtime_deps
+    DIRECTORIES ${MLX_RUNTIME_DIRS}
+    PRE_INCLUDE_REGEXES ${MLX_INCLUDE_REGEXES}
+    PRE_EXCLUDE_REGEXES ".*"
+    RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT MLX_VENDOR
+    LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT MLX_VENDOR
+)
+
+if(TARGET jaccl)
+    install(TARGETS jaccl
+        RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT MLX
+        LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT MLX
+        FRAMEWORK DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT MLX
+    )
+endif()
+
+# Install the Metal library for macOS arm64 (must be colocated with the binary).
+# Metal backend is only built for arm64, not x86_64.
+if(APPLE AND CMAKE_SYSTEM_PROCESSOR STREQUAL "arm64")
+    install(FILES ${CMAKE_BINARY_DIR}/_deps/mlx-build/mlx/backend/metal/kernels/mlx.metallib
+        DESTINATION ${OLLAMA_INSTALL_DIR}
+        COMPONENT MLX)
+endif()
+
+# Install headers for NVRTC JIT compilation at runtime.
+# MLX's own install rules use the default component so they get skipped by
+# --component MLX. Headers are installed alongside libmlx in OLLAMA_INSTALL_DIR.
+#
+# Layout:
+#   ${OLLAMA_INSTALL_DIR}/include/cccl/{cuda,nv}/  - CCCL headers
+#   ${OLLAMA_INSTALL_DIR}/include/*.h              - CUDA toolkit headers
+#
+# MLX's jit_module.cpp resolves CCCL via
+#   current_binary_dir()[.parent_path()] / "include" / "cccl"
+# On Linux, MLX's jit_module.cpp resolves CCCL via
+#   current_binary_dir().parent_path() / "include" / "cccl", so we create a
+#   symlink from lib/ollama/include -> ${OLLAMA_RUNNER_DIR}/include.
+# This will need refinement if we add multiple CUDA versions for MLX in the future.
+# CUDA runtime headers are found via CUDA_PATH env var (set by mlxrunner).
+if(EXISTS ${CMAKE_BINARY_DIR}/_deps/cccl-src/include/cuda)
+    install(DIRECTORY ${CMAKE_BINARY_DIR}/_deps/cccl-src/include/cuda
+        DESTINATION ${OLLAMA_INSTALL_DIR}/include/cccl
+        COMPONENT MLX)
+    install(DIRECTORY ${CMAKE_BINARY_DIR}/_deps/cccl-src/include/nv
+        DESTINATION ${OLLAMA_INSTALL_DIR}/include/cccl
+        COMPONENT MLX)
+endif()
+
+# Install minimal CUDA toolkit headers needed by MLX JIT kernels.
+# These are the transitive closure of includes from mlx/backend/cuda/device/*.cuh.
+# The Go mlxrunner sets CUDA_PATH to OLLAMA_INSTALL_DIR so MLX finds them at
+# $CUDA_PATH/include/*.h via NVRTC --include-path.
+if(CUDAToolkit_FOUND)
+    # CUDAToolkit_INCLUDE_DIRS may be a semicolon-separated list
+    # (e.g. ".../include;.../include/cccl"). Find the entry that
+    # contains the CUDA runtime headers we need.
+    set(_cuda_inc "")
+    foreach(_dir ${CUDAToolkit_INCLUDE_DIRS})
+        if(EXISTS "${_dir}/cuda_runtime_api.h")
+            set(_cuda_inc "${_dir}")
+            break()
+        endif()
+    endforeach()
+    if(NOT _cuda_inc)
+        message(WARNING "Could not find cuda_runtime_api.h in CUDAToolkit_INCLUDE_DIRS: ${CUDAToolkit_INCLUDE_DIRS}")
+    else()
+        set(_dst "${OLLAMA_INSTALL_DIR}/include")
+        set(_MLX_JIT_CUDA_HEADERS
+            builtin_types.h
+            cooperative_groups.h
+            cuda_bf16.h
+            cuda_bf16.hpp
+            cuda_device_runtime_api.h
+            cuda_fp16.h
+            cuda_fp16.hpp
+            cuda_fp8.h
+            cuda_fp8.hpp
+            cuda_runtime_api.h
+            device_types.h
+            driver_types.h
+            math_constants.h
+            surface_types.h
+            texture_types.h
+            vector_functions.h
+            vector_functions.hpp
+            vector_types.h
+        )
+        foreach(_hdr ${_MLX_JIT_CUDA_HEADERS})
+            install(FILES "${_cuda_inc}/${_hdr}"
+                DESTINATION ${_dst}
+                COMPONENT MLX)
+        endforeach()
+        # Subdirectory headers.
+        install(DIRECTORY "${_cuda_inc}/cooperative_groups"
+            DESTINATION ${_dst}
+            COMPONENT MLX
+            FILES_MATCHING PATTERN "*.h")
+        install(FILES "${_cuda_inc}/crt/host_defines.h"
+            DESTINATION "${_dst}/crt"
+            COMPONENT MLX)
+        if(NOT WIN32 AND NOT APPLE)
+            install(CODE "
+                set(_link \"${CMAKE_INSTALL_PREFIX}/${OLLAMA_LIB_DIR}/include\")
+                set(_target \"${OLLAMA_RUNNER_DIR}/include\")
+                if(NOT EXISTS \${_link})
+                    execute_process(COMMAND \${CMAKE_COMMAND} -E create_symlink \${_target} \${_link})
+                endif()
+            " COMPONENT MLX)
+        endif()
+    endif()
+endif()
+
+# On Windows, explicitly install dl.dll (dlfcn-win32 POSIX dlopen emulation).
+# RUNTIME_DEPENDENCIES auto-excludes it via POST_EXCLUDE_FILES_STRICT because
+# dlfcn-win32 is a known CMake target with its own install rules (which install
+# to the wrong destination). We must install it explicitly here.
+if(WIN32)
+    install(FILES ${OLLAMA_BUILD_DIR}/dl.dll
+        DESTINATION ${OLLAMA_INSTALL_DIR}
+        COMPONENT MLX)
+endif()
+
+# Manually install CUDA runtime libraries that MLX loads via dlopen
+# (not detected by RUNTIME_DEPENDENCIES since they aren't link-time deps).
+if(CUDAToolkit_FOUND)
+    file(GLOB MLX_CUDA_LIBS
+        "${CUDAToolkit_LIBRARY_DIR}/libcudart.so*"
+        "${CUDAToolkit_LIBRARY_DIR}/libcublas.so*"
+        "${CUDAToolkit_LIBRARY_DIR}/libcublasLt.so*"
+        "${CUDAToolkit_LIBRARY_DIR}/libnvrtc.so*"
+        "${CUDAToolkit_LIBRARY_DIR}/libnvrtc-builtins.so*"
+        "${CUDAToolkit_LIBRARY_DIR}/libcufft.so*"
+        "${CUDAToolkit_LIBRARY_DIR}/libcudnn.so*")
+    if(MLX_CUDA_LIBS)
+        install(FILES ${MLX_CUDA_LIBS}
+            DESTINATION ${OLLAMA_INSTALL_DIR}
+            COMPONENT MLX_VENDOR)
+    endif()
+endif()

cmake/mlx/CMakePresets.json

@@ -0,0 +1,90 @@
+{
+  "version": 3,
+  "configurePresets": [
+    {
+      "name": "default",
+      "binaryDir": "${sourceDir}/../../build/mlx",
+      "installDir": "${sourceDir}/../../dist",
+      "cacheVariables": {
+        "CMAKE_BUILD_TYPE": "Release",
+        "CMAKE_MSVC_RUNTIME_LIBRARY": "MultiThreaded",
+        "OLLAMA_SOURCE_DIR": "${sourceDir}/../.."
+      }
+    },
+    {
+      "name": "mlx_cuda_v13_base",
+      "hidden": true,
+      "inherits": [ "default" ],
+      "cacheVariables": {
+        "CMAKE_CUDA_FLAGS": "-t 2",
+        "OLLAMA_RUNNER_DIR": "mlx_cuda_v13"
+      }
+    },
+    {
+      "name": "mlx_cuda_v13_linux",
+      "inherits": [ "mlx_cuda_v13_base" ],
+      "binaryDir": "${sourceDir}/../../build/mlx_cuda_v13",
+      "cacheVariables": {
+        "MLX_CUDA_ARCHITECTURES": "75-virtual;80-virtual;86-virtual;89-virtual;90-virtual;90a-virtual;100-virtual;103-virtual;110-virtual;120-virtual;121-virtual"
+      }
+    },
+    {
+      "name": "mlx_cuda_v13_windows",
+      "inherits": [ "mlx_cuda_v13_base" ],
+      "binaryDir": "${sourceDir}/../../build/mlx_cuda_v13",
+      "cacheVariables": {
+        "MLX_CUDA_ARCHITECTURES": "75-virtual;80-virtual;86-virtual;89-virtual;90-virtual;90a-virtual;100-virtual;103-virtual;110-virtual;120-virtual;121-virtual"
+      }
+    },
+    {
+      "name": "mlx_cuda_v13_user_arch",
+      "inherits": [ "mlx_cuda_v13_base" ],
+      "binaryDir": "${sourceDir}/../../build/mlx_cuda_v13"
+    },
+    {
+      "name": "mlx_metal_v3",
+      "inherits": [ "default" ],
+      "binaryDir": "${sourceDir}/../../build/metal-v3",
+      "cacheVariables": {
+        "CMAKE_OSX_DEPLOYMENT_TARGET": "14.0",
+        "OLLAMA_RUNNER_DIR": "mlx_metal_v3"
+      }
+    },
+    {
+      "name": "mlx_metal_v4",
+      "inherits": [ "default" ],
+      "binaryDir": "${sourceDir}/../../build/metal-v4",
+      "cacheVariables": {
+        "CMAKE_OSX_DEPLOYMENT_TARGET": "26.0",
+        "OLLAMA_RUNNER_DIR": "mlx_metal_v4"
+      }
+    }
+  ],
+  "buildPresets": [
+    {
+      "name": "mlx_cuda_v13_linux",
+      "configurePreset": "mlx_cuda_v13_linux",
+      "targets": [ "mlx", "mlxc" ]
+    },
+    {
+      "name": "mlx_cuda_v13_windows",
+      "configurePreset": "mlx_cuda_v13_windows",
+      "targets": [ "mlx", "mlxc" ]
+    },
+    {
+      "name": "mlx_cuda_v13_user_arch",
+      "configurePreset": "mlx_cuda_v13_user_arch",
+      "targets": [ "mlx", "mlxc" ]
+    },
+    {
+      "name": "mlx_metal_v3",
+      "configurePreset": "mlx_metal_v3",
+      "targets": [ "mlx", "mlxc" ]
+    },
+    {
+      "name": "mlx_metal_v4",
+      "configurePreset": "mlx_metal_v4",
+      "targets": [ "mlx", "mlxc" ]
+    }
+  ]
+}

cmake/vendor-mlx-c-headers.cmake

@@ -0,0 +1,14 @@
+if(NOT DEFINED MLX_C_HEADERS_DIR OR NOT IS_DIRECTORY "${MLX_C_HEADERS_DIR}")
+    message(FATAL_ERROR "MLX_C_HEADERS_DIR does not exist: ${MLX_C_HEADERS_DIR}")
+endif()
+if(NOT DEFINED MLX_C_HEADERS_DEST OR "${MLX_C_HEADERS_DEST}" STREQUAL "")
+    message(FATAL_ERROR "MLX_C_HEADERS_DEST is required")
+endif()
+
+file(GLOB _mlx_c_headers LIST_DIRECTORIES false "${MLX_C_HEADERS_DIR}/*.h")
+if(NOT _mlx_c_headers)
+    message(FATAL_ERROR "No MLX-C headers found in ${MLX_C_HEADERS_DIR}")
+endif()
+
+file(MAKE_DIRECTORY "${MLX_C_HEADERS_DEST}")
+file(COPY ${_mlx_c_headers} DESTINATION "${MLX_C_HEADERS_DEST}")

cmake/windows-arm64-llvm-mingw.cmake

@@ -0,0 +1,69 @@
+set(CMAKE_SYSTEM_NAME Windows)
+set(CMAKE_SYSTEM_PROCESSOR ARM64)
+
+set(_ollama_llvm_mingw_hints)
+if(DEFINED ENV{ProgramFiles})
+    file(GLOB _ollama_program_files_llvm_mingw_bins
+        LIST_DIRECTORIES true
+        "$ENV{ProgramFiles}/llvm-mingw-*-x86_64*/bin")
+    list(SORT _ollama_program_files_llvm_mingw_bins COMPARE NATURAL ORDER DESCENDING)
+    list(APPEND _ollama_llvm_mingw_hints ${_ollama_program_files_llvm_mingw_bins})
+endif()
+if(DEFINED ENV{LOCALAPPDATA})
+    file(GLOB _ollama_winget_llvm_mingw_bins
+        LIST_DIRECTORIES true
+        "$ENV{LOCALAPPDATA}/Microsoft/WinGet/Packages/MartinStorsjo.LLVM-MinGW*/llvm-mingw-*-x86_64*/bin")
+    list(SORT _ollama_winget_llvm_mingw_bins COMPARE NATURAL ORDER DESCENDING)
+    list(APPEND _ollama_llvm_mingw_hints ${_ollama_winget_llvm_mingw_bins})
+endif()
+
+if(NOT CMAKE_C_COMPILER)
+    find_program(CMAKE_C_COMPILER
+        NAMES aarch64-w64-mingw32-gcc
+        HINTS ${_ollama_llvm_mingw_hints}
+        REQUIRED)
+endif()
+
+if(NOT CMAKE_CXX_COMPILER)
+    find_program(CMAKE_CXX_COMPILER
+        NAMES aarch64-w64-mingw32-g++
+        HINTS ${_ollama_llvm_mingw_hints}
+        REQUIRED)
+endif()
+
+get_filename_component(_ollama_llvm_mingw_bin_dir "${CMAKE_CXX_COMPILER}" DIRECTORY)
+
+if(NOT HOST_CXX_COMPILER)
+    find_program(_ollama_path_host_cxx
+        NAMES clang++ g++
+        NO_CMAKE_FIND_ROOT_PATH)
+    if(_ollama_path_host_cxx)
+        set(HOST_CXX_COMPILER "${_ollama_path_host_cxx}")
+    endif()
+endif()
+if(NOT HOST_CXX_COMPILER)
+    find_program(_ollama_mingw_host_cxx
+        NAMES x86_64-w64-mingw32-g++
+        HINTS "${_ollama_llvm_mingw_bin_dir}"
+        REQUIRED)
+    if(CMAKE_HOST_WIN32)
+        # llama.cpp builds a small host-only UI embedding tool during
+        # cross-compiles, but currently models HOST_CXX_COMPILER as only an
+        # executable path and has no companion host flags hook. When the host
+        # compiler is llvm-mingw, the generated host tool otherwise depends on
+        # llvm-mingw runtime DLLs being on PATH. Keep that workaround local and
+        # explicit: wrap the compiler only to add -static for this host tool.
+        set(_ollama_host_cxx_wrapper "${CMAKE_BINARY_DIR}/ollama-host-cxx.cmd")
+        file(TO_NATIVE_PATH "${_ollama_mingw_host_cxx}" _ollama_mingw_host_cxx_native)
+        file(WRITE "${_ollama_host_cxx_wrapper}"
+            "@echo off\r\n"
+            "\"${_ollama_mingw_host_cxx_native}\" -static %*\r\n")
+        set(HOST_CXX_COMPILER "${_ollama_host_cxx_wrapper}")
+    else()
+        set(HOST_CXX_COMPILER "${_ollama_mingw_host_cxx}")
+    endif()
+endif()
+set(HOST_CXX_COMPILER "${HOST_CXX_COMPILER}" CACHE FILEPATH "Host C++ compiler for build-time tools" FORCE)
+
+string(PREPEND CMAKE_C_FLAGS_INIT "-D_WIN32_WINNT=0x0A00 ")
+string(PREPEND CMAKE_CXX_FLAGS_INIT "-D_WIN32_WINNT=0x0A00 ")

cmd/cmd.go

@@ -18,6 +18,7 @@ import (
 	"os"
 	"os/exec"
 	"os/signal"
+	"path"
 	"path/filepath"
 	"runtime"
 	"slices"
@@ -41,6 +42,7 @@ import (
 	"github.com/ollama/ollama/cmd/config"
 	"github.com/ollama/ollama/cmd/launch"
 	"github.com/ollama/ollama/cmd/tui"
+	"github.com/ollama/ollama/discover"
 	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/format"
 	"github.com/ollama/ollama/internal/modelref"
@@ -232,9 +234,6 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 	// This gates both safetensors LLM and imagegen model creation
 	experimental, _ := cmd.Flags().GetBool("experimental")
 	draftQuantize, _ := cmd.Flags().GetString("draft-quantize")
-	if draftQuantize != "" && !experimental {
-		return errors.New("--draft-quantize requires --experimental")
-	}
 	if experimental {
 		if !isLocalhost() {
 			return errors.New("remote safetensor model creation not yet supported")
@@ -329,6 +328,12 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 	if quantize != "" {
 		req.Quantize = quantize
 	}
+	if draftQuantize != "" {
+		if len(req.DraftFiles) == 0 {
+			return errors.New("--draft-quantize requires a DRAFT model")
+		}
+		req.DraftQuantize = draftQuantize
+	}
 
 	client, err := api.ClientFromEnvironment()
 	if err != nil {
@@ -339,29 +344,40 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 	g.SetLimit(max(runtime.GOMAXPROCS(0)-1, 1))
 
 	files := syncmap.NewSyncMap[string, string]()
+	fileNames := createRequestFileNames(req.Files)
 	for f, digest := range req.Files {
 		g.Go(func() error {
 			if _, err := createBlob(cmd, client, f, digest, p); err != nil {
 				return err
 			}
 
-			// TODO: this is incorrect since the file might be in a subdirectory
-			//       instead this should take the path relative to the model directory
-			//       but the current implementation does not allow this
-			files.Store(filepath.Base(f), digest)
+			files.Store(fileNames[f], digest)
 			return nil
 		})
 	}
 
 	adapters := syncmap.NewSyncMap[string, string]()
+	adapterNames := createRequestFileNames(req.Adapters)
 	for f, digest := range req.Adapters {
 		g.Go(func() error {
 			if _, err := createBlob(cmd, client, f, digest, p); err != nil {
 				return err
 			}
 
-			// TODO: same here
-			adapters.Store(filepath.Base(f), digest)
+			adapters.Store(adapterNames[f], digest)
+			return nil
+		})
+	}
+
+	draftFiles := syncmap.NewSyncMap[string, string]()
+	draftFileNames := createRequestFileNames(req.DraftFiles)
+	for f, digest := range req.DraftFiles {
+		g.Go(func() error {
+			if _, err := createBlob(cmd, client, f, digest, p); err != nil {
+				return err
+			}
+
+			draftFiles.Store(draftFileNames[f], digest)
 			return nil
 		})
 	}
@@ -372,6 +388,7 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 
 	req.Files = files.Items()
 	req.Adapters = adapters.Items()
+	req.DraftFiles = draftFiles.Items()
 
 	bars := make(map[string]*progress.Bar)
 	fn := func(resp api.ProgressResponse) error {
@@ -409,6 +426,65 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 	return nil
 }
 
+func createRequestFileNames(files map[string]string) map[string]string {
+	names := make(map[string]string, len(files))
+	root, ok := commonFileRoot(files)
+	for f := range files {
+		name := filepath.Base(f)
+		if ok {
+			abs, err := filepath.Abs(f)
+			if err == nil {
+				if rel, err := filepath.Rel(root, abs); err == nil && rel != "." && rel != ".." && !strings.HasPrefix(rel, ".."+string(filepath.Separator)) {
+					name = rel
+				}
+			}
+		}
+		names[f] = path.Clean(filepath.ToSlash(name))
+	}
+	return names
+}
+
+func commonFileRoot(files map[string]string) (string, bool) {
+	if len(files) < 2 {
+		return "", false
+	}
+
+	var root string
+	var volume string
+	for f := range files {
+		abs, err := filepath.Abs(f)
+		if err != nil {
+			return "", false
+		}
+		if nextVolume := filepath.VolumeName(abs); volume == "" {
+			volume = nextVolume
+		} else if !strings.EqualFold(volume, nextVolume) {
+			return "", false
+		}
+
+		dir := filepath.Dir(abs)
+		if root == "" {
+			root = dir
+			continue
+		}
+
+		for {
+			rel, err := filepath.Rel(root, dir)
+			if err == nil && (rel == "." || (rel != ".." && !strings.HasPrefix(rel, ".."+string(filepath.Separator)))) {
+				break
+			}
+
+			parent := filepath.Dir(root)
+			if parent == root {
+				return "", false
+			}
+			root = parent
+		}
+	}
+
+	return root, root != ""
+}
+
 func createBlob(cmd *cobra.Command, client *api.Client, path string, digest string, p *progress.Progress) (string, error) {
 	realPath, err := filepath.EvalSymlinks(path)
 	if err != nil {
@@ -1277,11 +1353,28 @@ func showInfo(resp *api.ShowResponse, verbose bool, w io.Writer) error {
 
 	if resp.ProjectorInfo != nil {
 		tableRender("Projector", func() (rows [][]string) {
-			arch := resp.ProjectorInfo["general.architecture"].(string)
-			rows = append(rows, []string{"", "architecture", arch})
-			rows = append(rows, []string{"", "parameters", format.HumanNumber(uint64(resp.ProjectorInfo["general.parameter_count"].(float64)))})
-			rows = append(rows, []string{"", "embedding length", strconv.FormatFloat(resp.ProjectorInfo[fmt.Sprintf("%s.vision.embedding_length", arch)].(float64), 'f', -1, 64)})
-			rows = append(rows, []string{"", "dimensions", strconv.FormatFloat(resp.ProjectorInfo[fmt.Sprintf("%s.vision.projection_dim", arch)].(float64), 'f', -1, 64)})
+			arch, _ := resp.ProjectorInfo["general.architecture"].(string)
+			if arch != "" {
+				rows = append(rows, []string{"", "architecture", arch})
+			}
+			if v, ok := resp.ProjectorInfo["general.parameter_count"].(float64); ok {
+				rows = append(rows, []string{"", "parameters", format.HumanNumber(uint64(v))})
+			}
+
+			projectorValue := func(suffix string) (float64, bool) {
+				for _, modality := range []string{"vision", "audio"} {
+					if v, ok := resp.ProjectorInfo[fmt.Sprintf("%s.%s.%s", arch, modality, suffix)].(float64); ok {
+						return v, true
+					}
+				}
+				return 0, false
+			}
+			if v, ok := projectorValue("embedding_length"); ok {
+				rows = append(rows, []string{"", "embedding length", strconv.FormatFloat(v, 'f', -1, 64)})
+			}
+			if v, ok := projectorValue("projection_dim"); ok {
+				rows = append(rows, []string{"", "dimensions", strconv.FormatFloat(v, 'f', -1, 64)})
+			}
 			return
 		})
 	}
@@ -2009,7 +2102,7 @@ func appendEnvDocs(cmd *cobra.Command, envs []envconfig.EnvVar) {
 Environment Variables:
 `
 	for _, e := range envs {
-		envUsage += fmt.Sprintf("      %-24s   %s\n", e.Name, e.Description)
+		envUsage += fmt.Sprintf("      %-27s   %s\n", e.Name, e.Description)
 	}
 
 	cmd.SetUsageTemplate(cmd.UsageTemplate() + envUsage)
@@ -2231,7 +2324,7 @@ func runLauncherAction(cmd *cobra.Command, action tui.TUIAction, deps launcherDe
 
 func launcherActionExitsLoop(integration string) bool {
 	switch integration {
-	case "vscode":
+	case "codex-app", "vscode":
 		return true
 	default:
 		return false
@@ -2277,9 +2370,6 @@ func NewCLI() *cobra.Command {
 			if experimental, _ := cmd.Flags().GetBool("experimental"); experimental {
 				return nil
 			}
-			if draftQuantize, _ := cmd.Flags().GetString("draft-quantize"); draftQuantize != "" {
-				return errors.New("--draft-quantize requires --experimental")
-			}
 			return checkServerHeartbeat(cmd, args)
 		},
 		RunE: CreateHandler,
@@ -2445,6 +2535,16 @@ func NewCLI() *cobra.Command {
 		_ = runner.Execute(args[1:])
 	})
 
+	var gpuDiscoverLibDirs []string
+	gpuDiscoverCmd := &cobra.Command{
+		Use:    "gpu-discover",
+		Hidden: true,
+		RunE: func(cmd *cobra.Command, _ []string) error {
+			return discover.RunNativeProbeCommand(cmd.Context(), gpuDiscoverLibDirs, os.Stdout)
+		},
+	}
+	gpuDiscoverCmd.Flags().StringArrayVar(&gpuDiscoverLibDirs, "lib-dir", nil, "Ollama runtime library directory")
+
 	envVars := envconfig.AsMap()
 
 	envs := []envconfig.EnvVar{envVars["OLLAMA_HOST"]}
@@ -2473,6 +2573,7 @@ func NewCLI() *cobra.Command {
 				envVars["OLLAMA_CONTEXT_LENGTH"],
 				envVars["OLLAMA_KEEP_ALIVE"],
 				envVars["OLLAMA_MAX_LOADED_MODELS"],
+				envVars["OLLAMA_MAX_TRANSFER_STREAMS"],
 				envVars["OLLAMA_MAX_QUEUE"],
 				envVars["OLLAMA_MODELS"],
 				envVars["OLLAMA_NUM_PARALLEL"],
@@ -2484,6 +2585,9 @@ func NewCLI() *cobra.Command {
 				envVars["OLLAMA_KV_CACHE_TYPE"],
 				envVars["OLLAMA_LLM_LIBRARY"],
 				envVars["OLLAMA_GPU_OVERHEAD"],
+				envVars["OLLAMA_IGPU_ENABLE"],
+				envVars["LLAMA_ARG_FIT"],
+				envVars["LLAMA_ARG_FIT_TARGET"],
 				envVars["OLLAMA_LOAD_TIMEOUT"],
 			})
 		default:
@@ -2508,6 +2612,7 @@ func NewCLI() *cobra.Command {
 		copyCmd,
 		deleteCmd,
 		runnerCmd,
+		gpuDiscoverCmd,
 		launch.LaunchCmd(checkServerHeartbeat, runInteractiveTUI),
 	)
 

cmd/cmd_launcher_test.go

@@ -249,7 +249,7 @@ func TestRunLauncherAction_GUIAppsExitTUILoop(t *testing.T) {
 	cmd := &cobra.Command{}
 	cmd.SetContext(context.Background())
 
-	for _, integration := range []string{"vscode"} {
+	for _, integration := range []string{"codex-app", "vscode"} {
 		continueLoop, err := runLauncherAction(cmd, tui.TUIAction{Kind: tui.TUIActionLaunchIntegration, Integration: integration}, launcherDeps{
 			resolveRunModel: unexpectedRunModelResolution(t),
 			launchIntegration: func(ctx context.Context, req launch.IntegrationLaunchRequest) error {

cmd/cmd_test.go

@@ -1525,34 +1525,65 @@ func TestCreateHandler(t *testing.T) {
 	}
 }
 
-func TestCreateHandlerDraftQuantizeRequiresExperimental(t *testing.T) {
-	cmd := &cobra.Command{}
-	cmd.Flags().Bool("experimental", false, "")
-	cmd.Flags().String("draft-quantize", "mxfp8", "")
-	cmd.SetContext(t.Context())
+func TestCreateRequestFileNamesPreservesModelDirectoryLayout(t *testing.T) {
+	root := t.TempDir()
+	files := map[string]string{
+		filepath.Join(root, "model.safetensors"):            "sha256:model",
+		filepath.Join(root, "config.json"):                  "sha256:config",
+		filepath.Join(root, "2_Dense", "config.json"):       "sha256:dense-config",
+		filepath.Join(root, "2_Dense", "model.safetensors"): "sha256:dense-model",
+	}
 
-	err := CreateHandler(cmd, []string{"test-model"})
-	if err == nil || !strings.Contains(err.Error(), "--draft-quantize requires --experimental") {
-		t.Fatalf("error = %v, want draft-quantize requires experimental", err)
+	got := createRequestFileNames(files)
+	want := map[string]string{
+		filepath.Join(root, "model.safetensors"):            "model.safetensors",
+		filepath.Join(root, "config.json"):                  "config.json",
+		filepath.Join(root, "2_Dense", "config.json"):       "2_Dense/config.json",
+		filepath.Join(root, "2_Dense", "model.safetensors"): "2_Dense/model.safetensors",
+	}
+
+	if diff := cmp.Diff(want, got); diff != "" {
+		t.Fatalf("mismatch (-want +got):\n%s", diff)
 	}
 }
 
-func TestCreateHandlerDraftRequiresExperimental(t *testing.T) {
+func TestCreateRequestFileNamesPreservesRelativeModelDirectoryLayout(t *testing.T) {
+	root := t.TempDir()
+	t.Chdir(root)
+
+	files := map[string]string{
+		"model.safetensors":         "sha256:model",
+		"config.json":               "sha256:config",
+		"2_Dense/config.json":       "sha256:dense-config",
+		"2_Dense/model.safetensors": "sha256:dense-model",
+		"3_Dense/config.json":       "sha256:dense-config",
+		"3_Dense/model.safetensors": "sha256:dense-model",
+	}
+
+	got := createRequestFileNames(files)
+	for file := range files {
+		if got[file] != filepath.ToSlash(file) {
+			t.Fatalf("%s = %q, want %q", file, got[file], filepath.ToSlash(file))
+		}
+	}
+}
+
+func TestCreateHandlerDraftQuantizeRequiresDraft(t *testing.T) {
 	dir := t.TempDir()
 	modelfile := filepath.Join(dir, "Modelfile")
-	if err := os.WriteFile(modelfile, []byte("FROM base\nDRAFT ./assistant\n"), 0o644); err != nil {
+	if err := os.WriteFile(modelfile, []byte("FROM base\n"), 0o644); err != nil {
 		t.Fatal(err)
 	}
 
 	cmd := &cobra.Command{}
 	cmd.Flags().Bool("experimental", false, "")
-	cmd.Flags().String("draft-quantize", "", "")
 	cmd.Flags().String("file", modelfile, "")
+	cmd.Flags().String("draft-quantize", "mxfp8", "")
 	cmd.SetContext(t.Context())
 
 	err := CreateHandler(cmd, []string{"test-model"})
-	if err == nil || !strings.Contains(err.Error(), "DRAFT requires --experimental") {
-		t.Fatalf("error = %v, want DRAFT requires --experimental", err)
+	if err == nil || !strings.Contains(err.Error(), "--draft-quantize requires a DRAFT model") {
+		t.Fatalf("error = %v, want draft-quantize requires DRAFT", err)
 	}
 }
 

cmd/launch/claude.go

@@ -44,7 +44,7 @@ func (c *Claude) findPath() (string, error) {
 	return fallback, nil
 }
 
-func (c *Claude) Run(model string, args []string) error {
+func (c *Claude) Run(model string, _ []LaunchModel, args []string) error {
 	claudePath, err := c.findPath()
 	if err != nil {
 		return fmt.Errorf("claude is not installed, install from https://code.claude.com/docs/en/quickstart")

cmd/launch/claude_desktop.go

@@ -130,7 +130,7 @@ func (c *ClaudeDesktop) SkipModelReadiness() bool {
 	return true
 }
 
-func (c *ClaudeDesktop) Run(_ string, _ []string) error {
+func (c *ClaudeDesktop) Run(_ string, _ []LaunchModel, _ []string) error {
 	return errClaudeDesktopUnsupported()
 }
 

cmd/launch/claude_desktop_test.go

@@ -932,7 +932,7 @@ func TestClaudeDesktopRunReturnsUnsupported(t *testing.T) {
 	)
 
 	for _, args := range [][]string{nil, {"--foo"}} {
-		err := (&ClaudeDesktop{}).Run("qwen3.5", args)
+		err := (&ClaudeDesktop{}).Run("qwen3.5", nil, args)
 		if err == nil {
 			t.Fatal("expected Run to fail")
 		}

cmd/launch/cline.go

@@ -16,7 +16,7 @@ type Cline struct{}
 
 func (c *Cline) String() string { return "Cline" }
 
-func (c *Cline) Run(model string, args []string) error {
+func (c *Cline) Run(model string, _ []LaunchModel, args []string) error {
 	if _, err := exec.LookPath("cline"); err != nil {
 		return fmt.Errorf("cline is not installed, install with: npm install -g cline")
 	}
@@ -40,7 +40,7 @@ func (c *Cline) Paths() []string {
 	return nil
 }
 
-func (c *Cline) Edit(models []string) error {
+func (c *Cline) Edit(models []LaunchModel) error {
 	if len(models) == 0 {
 		return nil
 	}
@@ -66,10 +66,10 @@ func (c *Cline) Edit(models []string) error {
 	baseURL := envconfig.Host().String()
 	config["ollamaBaseUrl"] = baseURL
 	config["actModeApiProvider"] = "ollama"
-	config["actModeOllamaModelId"] = models[0]
+	config["actModeOllamaModelId"] = models[0].Name
 	config["actModeOllamaBaseUrl"] = baseURL
 	config["planModeApiProvider"] = "ollama"
-	config["planModeOllamaModelId"] = models[0]
+	config["planModeOllamaModelId"] = models[0].Name
 	config["planModeOllamaBaseUrl"] = baseURL
 
 	config["welcomeViewCompleted"] = true

cmd/launch/cline_test.go

@@ -43,7 +43,7 @@ func TestClineEdit(t *testing.T) {
 	t.Run("creates config from scratch", func(t *testing.T) {
 		os.RemoveAll(filepath.Join(tmpDir, ".cline"))
 
-		if err := c.Edit([]string{"kimi-k2.5:cloud"}); err != nil {
+		if err := c.Edit(testLaunchModels("kimi-k2.5:cloud")); err != nil {
 			t.Fatal(err)
 		}
 
@@ -77,7 +77,7 @@ func TestClineEdit(t *testing.T) {
 		data, _ := json.Marshal(existing)
 		os.WriteFile(configPath, data, 0o644)
 
-		if err := c.Edit([]string{"glm-5:cloud"}); err != nil {
+		if err := c.Edit(testLaunchModels("glm-5:cloud")); err != nil {
 			t.Fatal(err)
 		}
 
@@ -93,10 +93,10 @@ func TestClineEdit(t *testing.T) {
 	t.Run("updates model on re-edit", func(t *testing.T) {
 		os.RemoveAll(filepath.Join(tmpDir, ".cline"))
 
-		if err := c.Edit([]string{"kimi-k2.5:cloud"}); err != nil {
+		if err := c.Edit(testLaunchModels("kimi-k2.5:cloud")); err != nil {
 			t.Fatal(err)
 		}
-		if err := c.Edit([]string{"glm-5:cloud"}); err != nil {
+		if err := c.Edit(testLaunchModels("glm-5:cloud")); err != nil {
 			t.Fatal(err)
 		}
 
@@ -124,7 +124,7 @@ func TestClineEdit(t *testing.T) {
 	t.Run("uses first model as primary", func(t *testing.T) {
 		os.RemoveAll(filepath.Join(tmpDir, ".cline"))
 
-		if err := c.Edit([]string{"kimi-k2.5:cloud", "glm-5:cloud"}); err != nil {
+		if err := c.Edit(testLaunchModels("kimi-k2.5:cloud", "glm-5:cloud")); err != nil {
 			t.Fatal(err)
 		}
 

cmd/launch/codex.go

@@ -1,13 +1,17 @@
 package launch
 
 import (
+	"encoding/json"
 	"fmt"
 	"os"
 	"os/exec"
 	"path/filepath"
 	"strings"
 
+	"github.com/ollama/ollama/cmd/internal/fileutil"
 	"github.com/ollama/ollama/envconfig"
+	"github.com/ollama/ollama/types/model"
+	"github.com/pelletier/go-toml/v2"
 	"golang.org/x/mod/semver"
 )
 
@@ -16,10 +20,22 @@ type Codex struct{}
 
 func (c *Codex) String() string { return "Codex" }
 
-const codexProfileName = "ollama-launch"
+const (
+	codexProfileName           = "ollama-launch"
+	codexProviderName          = "Ollama"
+	codexFallbackContextWindow = 128_000
 
-func (c *Codex) args(model string, extra []string) []string {
+	codexRootProfileKey          = "profile"
+	codexRootModelKey            = "model"
+	codexRootModelProviderKey    = "model_provider"
+	codexRootModelCatalogJSONKey = "model_catalog_json"
+)
+
+func (c *Codex) args(model, modelCatalogPath string, extra []string) []string {
 	args := []string{"--profile", codexProfileName}
+	if modelCatalogPath != "" {
+		args = append(args, "-c", fmt.Sprintf("%s=%q", codexRootModelCatalogJSONKey, modelCatalogPath))
+	}
 	if model != "" {
 		args = append(args, "-m", model)
 	}
@@ -27,16 +43,21 @@ func (c *Codex) args(model string, extra []string) []string {
 	return args
 }
 
-func (c *Codex) Run(model string, args []string) error {
+func (c *Codex) Run(model string, models []LaunchModel, args []string) error {
 	if err := checkCodexVersion(); err != nil {
 		return err
 	}
 
-	if err := ensureCodexConfig(); err != nil {
+	if err := ensureCodexConfig(model, models); err != nil {
 		return fmt.Errorf("failed to configure codex: %w", err)
 	}
 
-	cmd := exec.Command("codex", c.args(model, args)...)
+	catalogPath, err := codexModelCatalogPath()
+	if err != nil {
+		return fmt.Errorf("failed to configure codex: %w", err)
+	}
+
+	cmd := exec.Command("codex", c.args(model, catalogPath, args)...)
 	cmd.Stdin = os.Stdin
 	cmd.Stdout = os.Stdout
 	cmd.Stderr = os.Stderr
@@ -46,79 +67,581 @@ func (c *Codex) Run(model string, args []string) error {
 	return cmd.Run()
 }
 
-// ensureCodexConfig writes a [profiles.ollama-launch] section to ~/.codex/config.toml
-// with openai_base_url pointing to the local Ollama server.
-func ensureCodexConfig() error {
-	home, err := os.UserHomeDir()
+// ensureCodexConfig writes a Codex profile and model catalog so Codex uses the
+// local Ollama server and has model metadata available.
+func ensureCodexConfig(modelName string, models []LaunchModel) error {
+	configPath, err := codexConfigPath()
 	if err != nil {
 		return err
 	}
 
-	codexDir := filepath.Join(home, ".codex")
+	codexDir := filepath.Dir(configPath)
 	if err := os.MkdirAll(codexDir, 0o755); err != nil {
 		return err
 	}
 
-	configPath := filepath.Join(codexDir, "config.toml")
-	return writeCodexProfile(configPath)
+	catalogPath := codexModelCatalogPathForConfig(configPath)
+	if err := writeCodexModelCatalog(catalogPath, codexCatalogModel(modelName, models)); err != nil {
+		return err
+	}
+
+	return writeCodexProfile(configPath, catalogPath)
+}
+
+func codexConfigPath() (string, error) {
+	home, err := os.UserHomeDir()
+	if err != nil {
+		return "", err
+	}
+	return filepath.Join(home, ".codex", "config.toml"), nil
+}
+
+func codexModelCatalogPath() (string, error) {
+	configPath, err := codexConfigPath()
+	if err != nil {
+		return "", err
+	}
+	return codexModelCatalogPathForConfig(configPath), nil
+}
+
+func codexModelCatalogPathForConfig(configPath string) string {
+	return filepath.Join(filepath.Dir(configPath), "model.json")
 }
 
 // writeCodexProfile ensures ~/.codex/config.toml has the ollama-launch profile
 // and model provider sections with the correct base URL.
-func writeCodexProfile(configPath string) error {
-	baseURL := envconfig.Host().String() + "/v1/"
+func writeCodexProfile(configPath string, modelCatalogPath ...string) error {
+	opts := codexLaunchProfileOptions{
+		forceAPIAuth: true,
+	}
+	if len(modelCatalogPath) > 0 {
+		opts.modelCatalogPath = modelCatalogPath[0]
+	}
+	return writeCodexLaunchProfile(configPath, opts)
+}
+
+type codexLaunchProfileOptions struct {
+	activate           bool
+	profileName        string
+	forceAPIAuth       bool
+	setRootModelConfig bool
+	model              string
+	modelCatalogPath   string
+	backupIntegration  string
+}
+
+func writeCodexLaunchProfile(configPath string, opts codexLaunchProfileOptions) error {
+	baseURL := codexBaseURL()
+	profileName := codexLaunchProfileName(opts)
+	profileHeader := codexProfileHeaderFor(profileName)
+	providerHeader := codexProviderHeaderFor(profileName)
+
+	content, readErr := os.ReadFile(configPath)
+	text := ""
+	if readErr == nil {
+		text = string(content)
+	} else if !os.IsNotExist(readErr) {
+		return readErr
+	}
+	parsed, err := codexParseConfig(text)
+	if err != nil {
+		return err
+	}
+
+	model := strings.TrimSpace(opts.model)
+	if model == "" {
+		model = parsed.ProfileString(profileName, codexRootModelKey)
+	}
+	modelCatalogPath := strings.TrimSpace(opts.modelCatalogPath)
+	if modelCatalogPath == "" {
+		modelCatalogPath = parsed.ProfileString(profileName, codexRootModelCatalogJSONKey)
+	}
+
+	profileLines := []string{}
+	if model != "" {
+		profileLines = append(profileLines, fmt.Sprintf("%s = %q", codexRootModelKey, model))
+	}
+	profileLines = append(profileLines,
+		fmt.Sprintf("openai_base_url = %q", baseURL),
+		fmt.Sprintf("%s = %q", codexRootModelProviderKey, profileName),
+	)
+	if opts.forceAPIAuth {
+		profileLines = append(profileLines, `forced_login_method = "api"`)
+	}
+	if modelCatalogPath != "" {
+		profileLines = append(profileLines, fmt.Sprintf("%s = %q", codexRootModelCatalogJSONKey, modelCatalogPath))
+	}
 
 	sections := []struct {
 		header string
 		lines  []string
 	}{
 		{
-			header: fmt.Sprintf("[profiles.%s]", codexProfileName),
-			lines: []string{
-				fmt.Sprintf("openai_base_url = %q", baseURL),
-				`forced_login_method = "api"`,
-				fmt.Sprintf("model_provider = %q", codexProfileName),
-			},
+			header: profileHeader,
+			lines:  profileLines,
 		},
 		{
-			header: fmt.Sprintf("[model_providers.%s]", codexProfileName),
+			header: providerHeader,
 			lines: []string{
-				`name = "Ollama"`,
+				fmt.Sprintf("name = %q", codexProviderName),
 				fmt.Sprintf("base_url = %q", baseURL),
+				`wire_api = "responses"`,
 			},
 		},
 	}
 
-	content, readErr := os.ReadFile(configPath)
-	text := ""
-	if readErr == nil {
-		text = string(content)
+	if opts.activate {
+		text = codexSetRootStringValue(text, codexRootProfileKey, profileName)
 	}
-
-	for _, s := range sections {
-		block := strings.Join(append([]string{s.header}, s.lines...), "\n") + "\n"
-
-		if idx := strings.Index(text, s.header); idx >= 0 {
-			// Replace the existing section up to the next section header.
-			rest := text[idx+len(s.header):]
-			if endIdx := strings.Index(rest, "\n["); endIdx >= 0 {
-				text = text[:idx] + block + rest[endIdx+1:]
-			} else {
-				text = text[:idx] + block
-			}
-		} else {
-			// Append the section.
-			if text != "" && !strings.HasSuffix(text, "\n") {
-				text += "\n"
-			}
-			if text != "" {
-				text += "\n"
-			}
-			text += block
+	if opts.setRootModelConfig {
+		if model != "" {
+			text = codexSetRootStringValue(text, codexRootModelKey, model)
+		}
+		text = codexSetRootStringValue(text, codexRootModelProviderKey, profileName)
+		if modelCatalogPath != "" {
+			text = codexSetRootStringValue(text, codexRootModelCatalogJSONKey, modelCatalogPath)
 		}
 	}
 
-	return os.WriteFile(configPath, []byte(text), 0o644)
+	for _, s := range sections {
+		text = codexUpsertSection(text, s.header, s.lines)
+	}
+	parsed, err = codexParseConfig(text)
+	if err != nil {
+		return err
+	}
+	if err := codexValidateLaunchProfileText(parsed, profileName, opts, model, modelCatalogPath, baseURL); err != nil {
+		return err
+	}
+
+	if err := os.MkdirAll(filepath.Dir(configPath), 0o755); err != nil {
+		return err
+	}
+	return fileutil.WriteWithBackup(configPath, []byte(text), opts.backupIntegration)
+}
+
+func codexLaunchProfileName(opts codexLaunchProfileOptions) string {
+	if name := strings.TrimSpace(opts.profileName); name != "" {
+		return name
+	}
+	return codexProfileName
+}
+
+func codexBaseURL() string {
+	return strings.TrimRight(envconfig.ConnectableHost().String(), "/") + "/v1/"
+}
+
+func codexProfileHeader() string {
+	return codexProfileHeaderFor(codexProfileName)
+}
+
+func codexProviderHeader() string {
+	return codexProviderHeaderFor(codexProfileName)
+}
+
+func codexProfileHeaderFor(profileName string) string {
+	return fmt.Sprintf("[profiles.%s]", profileName)
+}
+
+func codexProviderHeaderFor(profileName string) string {
+	return fmt.Sprintf("[model_providers.%s]", profileName)
+}
+
+func codexValidateLaunchProfileText(config codexParsedConfig, profileName string, opts codexLaunchProfileOptions, model, modelCatalogPath, baseURL string) error {
+	for _, check := range []struct {
+		path []string
+		want string
+	}{
+		{[]string{"profiles", profileName, "openai_base_url"}, baseURL},
+		{[]string{"profiles", profileName, codexRootModelProviderKey}, profileName},
+		{[]string{"model_providers", profileName, "name"}, codexProviderName},
+		{[]string{"model_providers", profileName, "base_url"}, baseURL},
+		{[]string{"model_providers", profileName, "wire_api"}, "responses"},
+	} {
+		if got, ok := config.String(check.path...); !ok || got != check.want {
+			return fmt.Errorf("generated Codex config missing %s = %q", strings.Join(check.path, "."), check.want)
+		}
+	}
+	if opts.forceAPIAuth {
+		if got, ok := config.String("profiles", profileName, "forced_login_method"); !ok || got != "api" {
+			return fmt.Errorf("generated Codex config missing profiles.%s.forced_login_method = %q", profileName, "api")
+		}
+	}
+	if model != "" {
+		if got, ok := config.String("profiles", profileName, codexRootModelKey); !ok || got != model {
+			return fmt.Errorf("generated Codex config missing profiles.%s.model = %q", profileName, model)
+		}
+	}
+	if modelCatalogPath != "" {
+		if got, ok := config.String("profiles", profileName, codexRootModelCatalogJSONKey); !ok || got != modelCatalogPath {
+			return fmt.Errorf("generated Codex config missing profiles.%s.model_catalog_json = %q", profileName, modelCatalogPath)
+		}
+	}
+	if opts.activate {
+		if got := config.RootString(codexRootProfileKey); got != profileName {
+			return fmt.Errorf("generated Codex config missing profile = %q", profileName)
+		}
+	}
+	if opts.setRootModelConfig {
+		if model != "" {
+			if got := config.RootString(codexRootModelKey); got != model {
+				return fmt.Errorf("generated Codex config missing model = %q", model)
+			}
+		}
+		if got := config.RootString(codexRootModelProviderKey); got != profileName {
+			return fmt.Errorf("generated Codex config missing model_provider = %q", profileName)
+		}
+		if modelCatalogPath != "" {
+			if got := config.RootString(codexRootModelCatalogJSONKey); got != modelCatalogPath {
+				return fmt.Errorf("generated Codex config missing model_catalog_json = %q", modelCatalogPath)
+			}
+		}
+	}
+	return nil
+}
+
+func codexUpsertSection(text, header string, lines []string) string {
+	block := strings.Join(append([]string{header}, lines...), "\n") + "\n"
+
+	if targetPath, ok := codexTableHeaderPath(header); ok {
+		if start, end, found := codexSectionRange(text, targetPath); found {
+			return text[:start] + block + text[end:]
+		}
+	}
+
+	if text != "" && !strings.HasSuffix(text, "\n") {
+		text += "\n"
+	}
+	if text != "" {
+		text += "\n"
+	}
+	return text + block
+}
+
+func codexRemoveSection(text, header string) string {
+	targetPath, ok := codexTableHeaderPath(header)
+	if !ok {
+		return text
+	}
+	start, end, found := codexSectionRange(text, targetPath)
+	if !found {
+		return text
+	}
+	return text[:start] + text[end:]
+}
+
+type codexParsedConfig struct {
+	values map[string]any
+}
+
+func (c codexParsedConfig) String(path ...string) (string, bool) {
+	if len(path) == 0 {
+		return "", false
+	}
+	var current any = c.values
+	for _, part := range path {
+		table, ok := current.(map[string]any)
+		if !ok {
+			return "", false
+		}
+		current, ok = table[part]
+		if !ok {
+			return "", false
+		}
+	}
+	value, ok := current.(string)
+	if !ok {
+		return "", false
+	}
+	return value, true
+}
+
+func (c codexParsedConfig) RootString(key string) string {
+	value, _ := c.RootStringOK(key)
+	return value
+}
+
+func (c codexParsedConfig) RootStringOK(key string) (string, bool) {
+	return c.String(key)
+}
+
+func (c codexParsedConfig) ProfileString(profileName, key string) string {
+	value, _ := c.String("profiles", profileName, key)
+	return value
+}
+
+func (c codexParsedConfig) ProviderString(profileName, key string) string {
+	value, _ := c.String("model_providers", profileName, key)
+	return value
+}
+
+func codexRootStringValue(text, key string) string {
+	config, err := codexParseConfig(text)
+	if err != nil {
+		return ""
+	}
+	return config.RootString(key)
+}
+
+func codexRootStringValueOK(text, key string) (string, bool) {
+	config, err := codexParseConfig(text)
+	if err != nil {
+		return "", false
+	}
+	return config.RootStringOK(key)
+}
+
+func codexStringValue(text string, path ...string) (string, bool) {
+	config, err := codexParseConfig(text)
+	if err != nil {
+		return "", false
+	}
+	return config.String(path...)
+}
+
+func codexSectionStringValue(text, header, key string) string {
+	path, ok := codexTableHeaderPath(header)
+	if !ok {
+		return ""
+	}
+	value, _ := codexStringValue(text, append(path, key)...)
+	return value
+}
+
+func codexParseConfig(text string) (codexParsedConfig, error) {
+	values, err := codexParseConfigText(text)
+	if err != nil {
+		return codexParsedConfig{}, err
+	}
+	return codexParsedConfig{values: values}, nil
+}
+
+func codexParseConfigText(text string) (map[string]any, error) {
+	cfg := map[string]any{}
+	if strings.TrimSpace(text) == "" {
+		return cfg, nil
+	}
+	if err := toml.Unmarshal([]byte(text), &cfg); err != nil {
+		return nil, fmt.Errorf("invalid Codex config TOML: %w", err)
+	}
+	return cfg, nil
+}
+
+func codexValidateConfigText(text string) error {
+	_, err := codexParseConfig(text)
+	return err
+}
+
+func codexSectionRange(text string, targetPath []string) (int, int, bool) {
+	lines := strings.SplitAfter(text, "\n")
+	offset := 0
+	start := -1
+	for _, line := range lines {
+		trimmed := strings.TrimSpace(line)
+		if !strings.HasPrefix(trimmed, "[") || strings.HasPrefix(trimmed, "#") {
+			offset += len(line)
+			continue
+		}
+		if start >= 0 {
+			return start, offset, true
+		}
+		if path, ok := codexTableHeaderPath(trimmed); ok && codexSamePath(path, targetPath) {
+			start = offset
+		}
+		offset += len(line)
+	}
+	if start >= 0 {
+		return start, len(text), true
+	}
+	return 0, 0, false
+}
+
+func codexTableHeaderPath(header string) ([]string, bool) {
+	trimmed := strings.TrimSpace(header)
+	if !strings.HasPrefix(trimmed, "[") || strings.HasPrefix(trimmed, "[[") {
+		return nil, false
+	}
+
+	const probeKey = "__ollama_launch_probe"
+	cfg := map[string]any{}
+	if err := toml.Unmarshal([]byte(trimmed+"\n"+probeKey+" = true\n"), &cfg); err != nil {
+		return nil, false
+	}
+	return codexFindProbePath(cfg, probeKey, nil)
+}
+
+func codexFindProbePath(value any, probeKey string, path []string) ([]string, bool) {
+	table, ok := value.(map[string]any)
+	if !ok {
+		return nil, false
+	}
+	if probe, ok := table[probeKey].(bool); ok && probe {
+		return path, true
+	}
+	for key, child := range table {
+		if key == probeKey {
+			continue
+		}
+		if childPath, ok := codexFindProbePath(child, probeKey, append(path, key)); ok {
+			return childPath, true
+		}
+	}
+	return nil, false
+}
+
+func codexSamePath(a, b []string) bool {
+	if len(a) != len(b) {
+		return false
+	}
+	for i := range a {
+		if a[i] != b[i] {
+			return false
+		}
+	}
+	return true
+}
+
+func codexSetRootStringValue(text, key, value string) string {
+	lines := strings.SplitAfter(text, "\n")
+	rootEnd := len(lines)
+	for i, line := range lines {
+		if strings.HasPrefix(strings.TrimSpace(line), "[") {
+			rootEnd = i
+			break
+		}
+	}
+
+	assignment := fmt.Sprintf("%s = %q", key, value)
+	for i := range rootEnd {
+		line := lines[i]
+		trimmed := strings.TrimSpace(line)
+		if trimmed == "" || strings.HasPrefix(trimmed, "#") {
+			continue
+		}
+		if codexRootLineHasKey(trimmed, key) {
+			if strings.HasSuffix(line, "\n") {
+				lines[i] = assignment + "\n"
+			} else {
+				lines[i] = assignment
+			}
+			return strings.Join(lines, "")
+		}
+	}
+
+	insert := assignment + "\n"
+	root := strings.Join(lines[:rootEnd], "")
+	rest := strings.Join(lines[rootEnd:], "")
+	if root != "" && !strings.HasSuffix(root, "\n") {
+		root += "\n"
+	}
+	if rest != "" && !strings.HasSuffix(insert, "\n\n") {
+		insert += "\n"
+	}
+	return root + insert + rest
+}
+
+func codexRemoveRootValue(text, key string) string {
+	lines := strings.SplitAfter(text, "\n")
+	rootEnd := len(lines)
+	for i, line := range lines {
+		if strings.HasPrefix(strings.TrimSpace(line), "[") {
+			rootEnd = i
+			break
+		}
+	}
+
+	out := make([]string, 0, len(lines))
+	for i, line := range lines {
+		if i < rootEnd {
+			trimmed := strings.TrimSpace(line)
+			if trimmed != "" && !strings.HasPrefix(trimmed, "#") && codexRootLineHasKey(trimmed, key) {
+				continue
+			}
+		}
+		out = append(out, line)
+	}
+	return strings.Join(out, "")
+}
+
+func codexRootLineHasKey(line, key string) bool {
+	cfg := map[string]any{}
+	if err := toml.Unmarshal([]byte(line+"\n"), &cfg); err != nil {
+		return false
+	}
+	_, ok := cfg[key]
+	return ok
+}
+
+func codexCatalogModel(modelName string, models []LaunchModel) LaunchModel {
+	if model, ok := findLaunchModel(models, modelName); ok {
+		return model.WithCloudLimits()
+	}
+	return fallbackLaunchModel(modelName)
+}
+
+func writeCodexModelCatalog(catalogPath string, model LaunchModel) error {
+	entry := buildCodexModelEntry(model)
+
+	catalog := map[string]any{
+		"models": []any{entry},
+	}
+
+	data, err := json.MarshalIndent(catalog, "", "  ")
+	if err != nil {
+		return err
+	}
+
+	return os.WriteFile(catalogPath, data, 0o644)
+}
+
+func buildCodexModelEntry(launchModel LaunchModel) map[string]any {
+	modelName := launchModel.Name
+	contextWindow := codexFallbackContextWindow
+	systemPrompt := ""
+
+	if launchModel.ContextLength > 0 {
+		contextWindow = launchModel.ContextLength
+	} else if launchModel.Details.ContextLength > 0 {
+		contextWindow = launchModel.Details.ContextLength
+	}
+	if l, ok := lookupCloudModelLimit(modelName); ok {
+		contextWindow = l.Context
+	}
+
+	if !isCloudModelName(modelName) && launchModel.Details.Format != "safetensors" {
+		if ctxLen := envconfig.ContextLength(); ctxLen > 0 {
+			contextWindow = int(ctxLen)
+		}
+	}
+
+	modalities := []string{"text"}
+	if launchModel.HasCapability(model.CapabilityVision) {
+		modalities = append(modalities, "image")
+	}
+
+	truncationMode := "bytes"
+	if isCloudModelName(modelName) {
+		truncationMode = "tokens"
+	}
+
+	return map[string]any{
+		"slug":                         modelName,
+		"display_name":                 modelName,
+		"context_window":               contextWindow,
+		"shell_type":                   "default",
+		"visibility":                   "list",
+		"supported_in_api":             true,
+		"priority":                     0,
+		"truncation_policy":            map[string]any{"mode": truncationMode, "limit": 10000},
+		"input_modalities":             modalities,
+		"base_instructions":            systemPrompt,
+		"support_verbosity":            true,
+		"default_verbosity":            "low",
+		"supports_parallel_tool_calls": false,
+		"supports_reasoning_summaries": false,
+		"supported_reasoning_levels":   []any{},
+		"experimental_supported_tools": []any{},
+	}
 }
 
 func checkCodexVersion() error {

cmd/launch/codex_test.go

@@ -1,15 +1,23 @@
 package launch
 
 import (
+	"encoding/json"
+	"fmt"
 	"os"
 	"path/filepath"
 	"slices"
 	"strings"
 	"testing"
+
+	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/cmd/internal/fileutil"
+	modelpkg "github.com/ollama/ollama/types/model"
 )
 
 func TestCodexArgs(t *testing.T) {
 	c := &Codex{}
+	catalogPath := filepath.Join("tmp", "model.json")
+	catalogArg := fmt.Sprintf("%s=%q", codexRootModelCatalogJSONKey, catalogPath)
 
 	tests := []struct {
 		name  string
@@ -17,15 +25,15 @@ func TestCodexArgs(t *testing.T) {
 		args  []string
 		want  []string
 	}{
-		{"with model", "llama3.2", nil, []string{"--profile", "ollama-launch", "-m", "llama3.2"}},
-		{"empty model", "", nil, []string{"--profile", "ollama-launch"}},
-		{"with model and extra args", "qwen3.5", []string{"-p", "myprofile"}, []string{"--profile", "ollama-launch", "-m", "qwen3.5", "-p", "myprofile"}},
-		{"with sandbox flag", "llama3.2", []string{"--sandbox", "workspace-write"}, []string{"--profile", "ollama-launch", "-m", "llama3.2", "--sandbox", "workspace-write"}},
+		{"with model", "llama3.2", nil, []string{"--profile", "ollama-launch", "-c", catalogArg, "-m", "llama3.2"}},
+		{"empty model", "", nil, []string{"--profile", "ollama-launch", "-c", catalogArg}},
+		{"with model and extra args", "qwen3.5", []string{"-p", "myprofile"}, []string{"--profile", "ollama-launch", "-c", catalogArg, "-m", "qwen3.5", "-p", "myprofile"}},
+		{"with sandbox flag", "llama3.2", []string{"--sandbox", "workspace-write"}, []string{"--profile", "ollama-launch", "-c", catalogArg, "-m", "llama3.2", "--sandbox", "workspace-write"}},
 	}
 
 	for _, tt := range tests {
 		t.Run(tt.name, func(t *testing.T) {
-			got := c.args(tt.model, tt.args)
+			got := c.args(tt.model, catalogPath, tt.args)
 			if !slices.Equal(got, tt.want) {
 				t.Errorf("args(%q, %v) = %v, want %v", tt.model, tt.args, got, tt.want)
 			}
@@ -37,8 +45,9 @@ func TestWriteCodexProfile(t *testing.T) {
 	t.Run("creates new file when none exists", func(t *testing.T) {
 		tmpDir := t.TempDir()
 		configPath := filepath.Join(tmpDir, "config.toml")
+		catalogPath := filepath.Join(tmpDir, "model.json")
 
-		if err := writeCodexProfile(configPath); err != nil {
+		if err := writeCodexProfile(configPath, catalogPath); err != nil {
 			t.Fatal(err)
 		}
 
@@ -63,21 +72,28 @@ func TestWriteCodexProfile(t *testing.T) {
 		if !strings.Contains(content, `model_provider = "ollama-launch"`) {
 			t.Error("missing model_provider key")
 		}
+		if !strings.Contains(content, fmt.Sprintf("model_catalog_json = %q", catalogPath)) {
+			t.Error("missing model_catalog_json key")
+		}
 		if !strings.Contains(content, "[model_providers.ollama-launch]") {
 			t.Error("missing [model_providers.ollama-launch] section")
 		}
 		if !strings.Contains(content, `name = "Ollama"`) {
 			t.Error("missing model provider name")
 		}
+		if err := codexValidateConfigText(content); err != nil {
+			t.Fatalf("generated config should be valid TOML: %v\n%s", err, content)
+		}
 	})
 
 	t.Run("appends profile to existing file without profile", func(t *testing.T) {
 		tmpDir := t.TempDir()
 		configPath := filepath.Join(tmpDir, "config.toml")
+		catalogPath := filepath.Join(tmpDir, "model.json")
 		existing := "[some_other_section]\nkey = \"value\"\n"
 		os.WriteFile(configPath, []byte(existing), 0o644)
 
-		if err := writeCodexProfile(configPath); err != nil {
+		if err := writeCodexProfile(configPath, catalogPath); err != nil {
 			t.Fatal(err)
 		}
 
@@ -95,10 +111,11 @@ func TestWriteCodexProfile(t *testing.T) {
 	t.Run("replaces existing profile section", func(t *testing.T) {
 		tmpDir := t.TempDir()
 		configPath := filepath.Join(tmpDir, "config.toml")
+		catalogPath := filepath.Join(tmpDir, "model.json")
 		existing := "[profiles.ollama-launch]\nopenai_base_url = \"http://old:1234/v1/\"\n\n[model_providers.ollama-launch]\nname = \"Ollama\"\nbase_url = \"http://old:1234/v1/\"\n"
 		os.WriteFile(configPath, []byte(existing), 0o644)
 
-		if err := writeCodexProfile(configPath); err != nil {
+		if err := writeCodexProfile(configPath, catalogPath); err != nil {
 			t.Fatal(err)
 		}
 
@@ -114,15 +131,160 @@ func TestWriteCodexProfile(t *testing.T) {
 		if strings.Count(content, "[model_providers.ollama-launch]") != 1 {
 			t.Errorf("expected exactly one [model_providers.ollama-launch] section, got %d", strings.Count(content, "[model_providers.ollama-launch]"))
 		}
+		if err := codexValidateConfigText(content); err != nil {
+			t.Fatalf("generated config should be valid TOML: %v\n%s", err, content)
+		}
+	})
+
+	t.Run("replaces equivalent quoted profile table", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		configPath := filepath.Join(tmpDir, "config.toml")
+		existing := "" +
+			`profile = "default"` + "\n\n" +
+			`[profiles."ollama-launch"]` + "\n" +
+			`openai_base_url = "http://old:1234/v1/"` + "\n\n" +
+			`[model_providers."ollama-launch"]` + "\n" +
+			`name = "Old"` + "\n" +
+			`base_url = "http://old:1234/v1/"` + "\n\n" +
+			`[profiles.default]` + "\n" +
+			`model = "gpt-5.5"` + "\n"
+		os.WriteFile(configPath, []byte(existing), 0o644)
+
+		if err := writeCodexProfile(configPath); err != nil {
+			t.Fatal(err)
+		}
+
+		data, _ := os.ReadFile(configPath)
+		content := string(data)
+
+		if strings.Contains(content, `profiles."ollama-launch"`) {
+			t.Fatalf("quoted profile table should be replaced, got:\n%s", content)
+		}
+		if strings.Contains(content, "old:1234") {
+			t.Fatalf("old URL was not replaced, got:\n%s", content)
+		}
+		if got := codexSectionStringValue(content, codexProfileHeader(), "model_provider"); got != codexProfileName {
+			t.Fatalf("profile model_provider = %q, want %q", got, codexProfileName)
+		}
+		if got := codexSectionStringValue(content, codexProviderHeader(), "base_url"); !strings.Contains(got, "/v1/") {
+			t.Fatalf("provider base_url = %q, want /v1/ URL", got)
+		}
+		if err := codexValidateConfigText(content); err != nil {
+			t.Fatalf("generated config should be valid TOML: %v\n%s", err, content)
+		}
+	})
+
+	t.Run("rejects invalid existing toml without writing", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		configPath := filepath.Join(tmpDir, "config.toml")
+		existing := "profile = \n"
+		os.WriteFile(configPath, []byte(existing), 0o644)
+
+		err := writeCodexProfile(configPath)
+		if err == nil || !strings.Contains(err.Error(), "invalid Codex config TOML") {
+			t.Fatalf("writeCodexProfile error = %v, want invalid TOML", err)
+		}
+
+		data, _ := os.ReadFile(configPath)
+		if string(data) != existing {
+			t.Fatalf("invalid config should be left untouched, got:\n%s", data)
+		}
+	})
+
+	t.Run("rejects malformed existing toml variants without writing", func(t *testing.T) {
+		tests := map[string]string{
+			"duplicate root key":  "profile = \"default\"\nprofile = \"other\"\n",
+			"unterminated string": "model = \"gpt-5.5\n",
+			"bad table":           "[profiles.ollama-launch\nmodel = \"llama3.2\"\n",
+			"duplicate table key": "[profiles.ollama-launch]\nmodel = \"a\"\nmodel = \"b\"\n",
+		}
+		for name, existing := range tests {
+			t.Run(name, func(t *testing.T) {
+				tmpDir := t.TempDir()
+				configPath := filepath.Join(tmpDir, "config.toml")
+				if err := os.WriteFile(configPath, []byte(existing), 0o644); err != nil {
+					t.Fatal(err)
+				}
+
+				err := writeCodexProfile(configPath)
+				if err == nil || !strings.Contains(err.Error(), "invalid Codex config TOML") {
+					t.Fatalf("writeCodexProfile error = %v, want invalid TOML", err)
+				}
+
+				data, _ := os.ReadFile(configPath)
+				if string(data) != existing {
+					t.Fatalf("invalid config should be left untouched, got:\n%s", data)
+				}
+			})
+		}
+	})
+
+	t.Run("backs up previous config before overwrite", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+		configPath := filepath.Join(tmpDir, ".codex", "config.toml")
+		if err := os.MkdirAll(filepath.Dir(configPath), 0o755); err != nil {
+			t.Fatal(err)
+		}
+		existing := "# original-codex-backup-marker\n[profiles.default]\nmodel = \"gpt-5.5\"\n"
+		if err := os.WriteFile(configPath, []byte(existing), 0o644); err != nil {
+			t.Fatal(err)
+		}
+
+		if err := writeCodexProfile(configPath); err != nil {
+			t.Fatal(err)
+		}
+
+		assertBackupContains(t, filepath.Join(fileutil.BackupDir(), "config.toml.*"), "original-codex-backup-marker")
+	})
+
+	t.Run("updates equivalent quoted root keys", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		configPath := filepath.Join(tmpDir, "config.toml")
+		existing := "" +
+			`"profile" = "default"` + "\n" +
+			`"model" = "gpt-5.5"` + "\n" +
+			`"model_provider" = "openai"` + "\n\n" +
+			`[profiles.default]` + "\n" +
+			`model = "gpt-5.5"` + "\n"
+		os.WriteFile(configPath, []byte(existing), 0o644)
+
+		err := writeCodexLaunchProfile(configPath, codexLaunchProfileOptions{
+			activate:           true,
+			setRootModelConfig: true,
+			model:              "llama3.2",
+		})
+		if err != nil {
+			t.Fatal(err)
+		}
+
+		data, _ := os.ReadFile(configPath)
+		content := string(data)
+		for key, want := range map[string]string{
+			"profile":        codexProfileName,
+			"model":          "llama3.2",
+			"model_provider": codexProfileName,
+		} {
+			if got := codexRootStringValue(content, key); got != want {
+				t.Fatalf("root %s = %q, want %q in:\n%s", key, got, want, content)
+			}
+		}
+		if strings.Contains(content, `"profile"`) || strings.Contains(content, `"model_provider"`) {
+			t.Fatalf("quoted root keys should be rewritten once, got:\n%s", content)
+		}
+		if err := codexValidateConfigText(content); err != nil {
+			t.Fatalf("generated config should be valid TOML: %v\n%s", err, content)
+		}
 	})
 
 	t.Run("replaces profile while preserving following sections", func(t *testing.T) {
 		tmpDir := t.TempDir()
 		configPath := filepath.Join(tmpDir, "config.toml")
+		catalogPath := filepath.Join(tmpDir, "model.json")
 		existing := "[profiles.ollama-launch]\nopenai_base_url = \"http://old:1234/v1/\"\n[another_section]\nfoo = \"bar\"\n"
 		os.WriteFile(configPath, []byte(existing), 0o644)
 
-		if err := writeCodexProfile(configPath); err != nil {
+		if err := writeCodexProfile(configPath, catalogPath); err != nil {
 			t.Fatal(err)
 		}
 
@@ -143,10 +305,11 @@ func TestWriteCodexProfile(t *testing.T) {
 	t.Run("appends newline to file not ending with newline", func(t *testing.T) {
 		tmpDir := t.TempDir()
 		configPath := filepath.Join(tmpDir, "config.toml")
+		catalogPath := filepath.Join(tmpDir, "model.json")
 		existing := "[other]\nkey = \"val\""
 		os.WriteFile(configPath, []byte(existing), 0o644)
 
-		if err := writeCodexProfile(configPath); err != nil {
+		if err := writeCodexProfile(configPath, catalogPath); err != nil {
 			t.Fatal(err)
 		}
 
@@ -166,8 +329,9 @@ func TestWriteCodexProfile(t *testing.T) {
 		t.Setenv("OLLAMA_HOST", "http://myhost:9999")
 		tmpDir := t.TempDir()
 		configPath := filepath.Join(tmpDir, "config.toml")
+		catalogPath := filepath.Join(tmpDir, "model.json")
 
-		if err := writeCodexProfile(configPath); err != nil {
+		if err := writeCodexProfile(configPath, catalogPath); err != nil {
 			t.Fatal(err)
 		}
 
@@ -178,6 +342,26 @@ func TestWriteCodexProfile(t *testing.T) {
 			t.Errorf("expected custom host in URL, got:\n%s", content)
 		}
 	})
+
+	t.Run("uses connectable host for unspecified bind address", func(t *testing.T) {
+		t.Setenv("OLLAMA_HOST", "http://0.0.0.0:11434")
+		tmpDir := t.TempDir()
+		configPath := filepath.Join(tmpDir, "config.toml")
+
+		if err := writeCodexProfile(configPath); err != nil {
+			t.Fatal(err)
+		}
+
+		data, _ := os.ReadFile(configPath)
+		content := string(data)
+
+		if strings.Contains(content, "0.0.0.0") {
+			t.Fatalf("config should not write bind-only host, got:\n%s", content)
+		}
+		if !strings.Contains(content, "127.0.0.1:11434/v1/") {
+			t.Fatalf("expected connectable loopback URL, got:\n%s", content)
+		}
+	})
 }
 
 func TestEnsureCodexConfig(t *testing.T) {
@@ -185,7 +369,7 @@ func TestEnsureCodexConfig(t *testing.T) {
 		tmpDir := t.TempDir()
 		setTestHome(t, tmpDir)
 
-		if err := ensureCodexConfig(); err != nil {
+		if err := ensureCodexConfig("llama3.2", launchModelsFromNames([]string{"llama3.2"})); err != nil {
 			t.Fatal(err)
 		}
 
@@ -202,16 +386,25 @@ func TestEnsureCodexConfig(t *testing.T) {
 		if !strings.Contains(content, "openai_base_url") {
 			t.Error("missing openai_base_url key")
 		}
+
+		catalogPath := filepath.Join(tmpDir, ".codex", "model.json")
+		data, err = os.ReadFile(catalogPath)
+		if err != nil {
+			t.Fatalf("model.json not created: %v", err)
+		}
+		if !strings.Contains(string(data), `"slug": "llama3.2"`) {
+			t.Error("missing model catalog entry for selected model")
+		}
 	})
 
 	t.Run("is idempotent", func(t *testing.T) {
 		tmpDir := t.TempDir()
 		setTestHome(t, tmpDir)
 
-		if err := ensureCodexConfig(); err != nil {
+		if err := ensureCodexConfig("llama3.2", launchModelsFromNames([]string{"llama3.2"})); err != nil {
 			t.Fatal(err)
 		}
-		if err := ensureCodexConfig(); err != nil {
+		if err := ensureCodexConfig("llama3.2", launchModelsFromNames([]string{"llama3.2"})); err != nil {
 			t.Fatal(err)
 		}
 
@@ -227,3 +420,176 @@ func TestEnsureCodexConfig(t *testing.T) {
 		}
 	})
 }
+
+func assertBackupContains(t *testing.T, pattern, marker string) {
+	t.Helper()
+	backups, err := filepath.Glob(pattern)
+	if err != nil {
+		t.Fatal(err)
+	}
+	for _, backupPath := range backups {
+		data, err := os.ReadFile(backupPath)
+		if err != nil {
+			t.Fatal(err)
+		}
+		if strings.Contains(string(data), marker) {
+			return
+		}
+	}
+	t.Fatalf("backup matching %q with marker %q not found", pattern, marker)
+}
+
+func TestModelInfoContextLength(t *testing.T) {
+	tests := []struct {
+		name      string
+		modelInfo map[string]any
+		want      int
+	}{
+		{"float64 value", map[string]any{"qwen3_5_moe.context_length": float64(262144)}, 262144},
+		{"int value", map[string]any{"llama.context_length": 131072}, 131072},
+		{"no context_length key", map[string]any{"llama.embedding_length": float64(4096)}, 0},
+		{"empty map", map[string]any{}, 0},
+		{"nil map", nil, 0},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			got, _ := modelInfoContextLength(tt.modelInfo)
+			if got != tt.want {
+				t.Errorf("modelInfoContextLength() = %d, want %d", got, tt.want)
+			}
+		})
+	}
+}
+
+func TestBuildCodexModelEntryContextWindow(t *testing.T) {
+	tests := []struct {
+		name          string
+		model         LaunchModel
+		envContextLen string
+		wantContext   int
+	}{
+		{
+			name: "inventory context length as fallback",
+			model: LaunchModel{
+				Name:          "llama3.2",
+				ContextLength: 131072,
+				Details:       api.ModelDetails{Format: "gguf"},
+			},
+			wantContext: 131072,
+		},
+		{
+			name: "details context length is used when model context is empty",
+			model: LaunchModel{
+				Name:    "llama3.2",
+				Details: api.ModelDetails{Format: "gguf", ContextLength: 131072},
+			},
+			wantContext: 131072,
+		},
+		{
+			name: "OLLAMA_CONTEXT_LENGTH overrides local gguf inventory context",
+			model: LaunchModel{
+				Name:          "llama3.2",
+				ContextLength: 131072,
+				Details:       api.ModelDetails{Format: "gguf"},
+			},
+			envContextLen: "64000",
+			wantContext:   64000,
+		},
+		{
+			name: "safetensors uses inventory context only",
+			model: LaunchModel{
+				Name:          "llama3.2",
+				ContextLength: 131072,
+				Details:       api.ModelDetails{Format: "safetensors"},
+			},
+			envContextLen: "64000",
+			wantContext:   131072,
+		},
+		{
+			name: "cloud model uses hardcoded limits",
+			model: LaunchModel{
+				Name:          "qwen3.5:cloud",
+				ContextLength: 131072,
+				Details:       api.ModelDetails{Format: "gguf"},
+			},
+			envContextLen: "64000",
+			wantContext:   262144,
+		},
+		{
+			name: "unknown cloud model without metadata uses fallback context",
+			model: LaunchModel{
+				Name: "deepseek-v4-pro:cloud",
+			},
+			envContextLen: "64000",
+			wantContext:   codexFallbackContextWindow,
+		},
+		{
+			name: "vision capability without reasoning advertisement",
+			model: LaunchModel{
+				Name:          "llama3.2",
+				ContextLength: 131072,
+				Details:       api.ModelDetails{Format: "gguf"},
+				Capabilities:  []modelpkg.Capability{modelpkg.CapabilityVision, modelpkg.CapabilityThinking},
+			},
+			wantContext: 131072,
+		},
+		{
+			name:        "missing metadata uses fallback context",
+			model:       LaunchModel{Name: "llama3.2"},
+			wantContext: codexFallbackContextWindow,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			if tt.envContextLen != "" {
+				t.Setenv("OLLAMA_CONTEXT_LENGTH", tt.envContextLen)
+			} else {
+				t.Setenv("OLLAMA_CONTEXT_LENGTH", "")
+			}
+
+			entry := buildCodexModelEntry(tt.model)
+
+			gotContext, _ := entry["context_window"].(int)
+			if gotContext != tt.wantContext {
+				t.Errorf("context_window = %d, want %d", gotContext, tt.wantContext)
+			}
+
+			if tt.name == "vision capability without reasoning advertisement" {
+				modalities, _ := entry["input_modalities"].([]string)
+				if !slices.Contains(modalities, "image") {
+					t.Error("expected image in input_modalities")
+				}
+				levels, _ := entry["supported_reasoning_levels"].([]any)
+				if len(levels) != 0 {
+					t.Errorf("supported_reasoning_levels length = %d, want 0", len(levels))
+				}
+				if got, _ := entry["supports_reasoning_summaries"].(bool); got {
+					t.Error("supports_reasoning_summaries = true, want false")
+				}
+			}
+
+			if tt.name == "cloud model uses hardcoded limits" {
+				truncationPolicy, _ := entry["truncation_policy"].(map[string]any)
+				if mode, _ := truncationPolicy["mode"].(string); mode != "tokens" {
+					t.Errorf("truncation_policy mode = %q, want %q", mode, "tokens")
+				}
+			}
+
+			requiredKeys := []string{"slug", "display_name", "shell_type"}
+			for _, key := range requiredKeys {
+				if _, ok := entry[key]; !ok {
+					t.Errorf("missing required key %q", key)
+				}
+			}
+			if _, ok := entry["apply_patch_tool_type"]; ok {
+				t.Error("apply_patch_tool_type should be omitted so Codex CLI defaults can handle schema changes")
+			}
+
+			if _, err := json.Marshal(entry); err != nil {
+				t.Errorf("entry is not JSON serializable: %v", err)
+			}
+		})
+	}
+}

cmd/launch/copilot.go

@@ -43,7 +43,7 @@ func (c *Copilot) findPath() (string, error) {
 	return fallback, nil
 }
 
-func (c *Copilot) Run(model string, args []string) error {
+func (c *Copilot) Run(model string, _ []LaunchModel, args []string) error {
 	copilotPath, err := c.findPath()
 	if err != nil {
 		return fmt.Errorf("copilot is not installed, install from https://docs.github.com/en/copilot/how-tos/set-up/install-copilot-cli")

cmd/launch/droid.go

@@ -40,7 +40,7 @@ type modelEntry struct {
 
 func (d *Droid) String() string { return "Droid" }
 
-func (d *Droid) Run(model string, args []string) error {
+func (d *Droid) Run(model string, _ []LaunchModel, args []string) error {
 	if _, err := exec.LookPath("droid"); err != nil {
 		return fmt.Errorf("droid is not installed, install from https://docs.factory.ai/cli/getting-started/quickstart")
 	}
@@ -64,7 +64,7 @@ func (d *Droid) Paths() []string {
 	return nil
 }
 
-func (d *Droid) Edit(models []string) error {
+func (d *Droid) Edit(models []LaunchModel) error {
 	if len(models) == 0 {
 		return nil
 	}
@@ -99,7 +99,7 @@ func (d *Droid) Edit(models []string) error {
 	return fileutil.WriteWithBackup(settingsPath, data, "droid")
 }
 
-func updateDroidSettings(settingsMap map[string]any, settings droidSettings, models []string) map[string]any {
+func updateDroidSettings(settingsMap map[string]any, settings droidSettings, models []LaunchModel) map[string]any {
 	// Keep only non-Ollama models from the raw map (preserves extra fields)
 	// Rebuild Ollama models
 	var nonOllamaModels []any
@@ -119,20 +119,18 @@ func updateDroidSettings(settingsMap map[string]any, settings droidSettings, mod
 	var defaultModelID string
 	for i, model := range models {
 		maxOutput := 64000
-		if isCloudModelName(model) {
-			if l, ok := lookupCloudModelLimit(model); ok {
-				maxOutput = l.Output
-			}
+		if model.MaxOutputTokens > 0 {
+			maxOutput = model.MaxOutputTokens
 		}
-		modelID := fmt.Sprintf("custom:%s-%d", model, i)
+		modelID := fmt.Sprintf("custom:%s-%d", model.Name, i)
 		newModels = append(newModels, modelEntry{
-			Model:           model,
-			DisplayName:     model,
+			Model:           model.Name,
+			DisplayName:     model.Name,
 			BaseURL:         envconfig.Host().String() + "/v1",
 			APIKey:          "ollama",
 			Provider:        "generic-chat-completion-api",
 			MaxOutputTokens: maxOutput,
-			SupportsImages:  false,
+			SupportsImages:  model.HasCapability("vision"),
 			ID:              modelID,
 			Index:           i,
 		})

cmd/launch/droid_test.go

@@ -63,7 +63,7 @@ func TestDroidEdit(t *testing.T) {
 
 	t.Run("fresh install creates models with sequential indices", func(t *testing.T) {
 		cleanup()
-		if err := d.Edit([]string{"model-a", "model-b"}); err != nil {
+		if err := d.Edit(testLaunchModels("model-a", "model-b")); err != nil {
 			t.Fatal(err)
 		}
 
@@ -99,7 +99,7 @@ func TestDroidEdit(t *testing.T) {
 
 	t.Run("sets sessionDefaultSettings.model to first model ID", func(t *testing.T) {
 		cleanup()
-		if err := d.Edit([]string{"model-a", "model-b"}); err != nil {
+		if err := d.Edit(testLaunchModels("model-a", "model-b")); err != nil {
 			t.Fatal(err)
 		}
 
@@ -116,10 +116,10 @@ func TestDroidEdit(t *testing.T) {
 	t.Run("re-indexes when models removed", func(t *testing.T) {
 		cleanup()
 		// Add three models
-		d.Edit([]string{"model-a", "model-b", "model-c"})
+		d.Edit(testLaunchModels("model-a", "model-b", "model-c"))
 
 		// Remove middle model
-		d.Edit([]string{"model-a", "model-c"})
+		d.Edit(testLaunchModels("model-a", "model-c"))
 
 		settings := readSettings()
 		models := getCustomModels(settings)
@@ -155,7 +155,7 @@ func TestDroidEdit(t *testing.T) {
 			]
 		}`), 0o644)
 
-		d.Edit([]string{"model-a"})
+		d.Edit(testLaunchModels("model-a"))
 
 		settings := readSettings()
 		models := getCustomModels(settings)
@@ -184,7 +184,7 @@ func TestDroidEdit(t *testing.T) {
 			"sessionDefaultSettings": {"autonomyMode": "auto-high"}
 		}`), 0o644)
 
-		d.Edit([]string{"model-a"})
+		d.Edit(testLaunchModels("model-a"))
 
 		settings := readSettings()
 
@@ -203,7 +203,7 @@ func TestDroidEdit(t *testing.T) {
 
 	t.Run("required fields present", func(t *testing.T) {
 		cleanup()
-		d.Edit([]string{"test-model"})
+		d.Edit(testLaunchModels("test-model"))
 
 		settings := readSettings()
 		models := getCustomModels(settings)
@@ -239,7 +239,7 @@ func TestDroidEdit(t *testing.T) {
 			"sessionDefaultSettings": {"reasoningEffort": "off"}
 		}`), 0o644)
 
-		d.Edit([]string{"model-a"})
+		d.Edit(testLaunchModels("model-a"))
 
 		settings := readSettings()
 		session := settings["sessionDefaultSettings"].(map[string]any)
@@ -256,7 +256,7 @@ func TestDroidEdit(t *testing.T) {
 			"sessionDefaultSettings": {"reasoningEffort": "high"}
 		}`), 0o644)
 
-		d.Edit([]string{"model-a"})
+		d.Edit(testLaunchModels("model-a"))
 
 		settings := readSettings()
 		session := settings["sessionDefaultSettings"].(map[string]any)
@@ -281,7 +281,7 @@ func TestDroidEdit_CorruptedJSON(t *testing.T) {
 	os.WriteFile(settingsPath, []byte(`{corrupted json content`), 0o644)
 
 	// Corrupted JSON should return an error so user knows something is wrong
-	err := d.Edit([]string{"model-a"})
+	err := d.Edit(testLaunchModels("model-a"))
 	if err == nil {
 		t.Fatal("expected error for corrupted JSON, got nil")
 	}
@@ -306,7 +306,7 @@ func TestDroidEdit_WrongTypeCustomModels(t *testing.T) {
 	os.WriteFile(settingsPath, []byte(`{"customModels": "not an array"}`), 0o644)
 
 	// Should not panic - wrong type should be handled gracefully
-	err := d.Edit([]string{"model-a"})
+	err := d.Edit(testLaunchModels("model-a"))
 	if err != nil {
 		t.Fatalf("Edit failed with wrong type customModels: %v", err)
 	}
@@ -338,7 +338,7 @@ func TestDroidEdit_EmptyModels(t *testing.T) {
 	os.WriteFile(settingsPath, []byte(originalContent), 0o644)
 
 	// Empty models should be no-op
-	err := d.Edit([]string{})
+	err := d.Edit(testLaunchModels())
 	if err != nil {
 		t.Fatalf("Edit with empty models failed: %v", err)
 	}
@@ -359,7 +359,7 @@ func TestDroidEdit_DuplicateModels(t *testing.T) {
 	settingsPath := filepath.Join(settingsDir, "settings.json")
 
 	// Add same model twice
-	err := d.Edit([]string{"model-a", "model-a"})
+	err := d.Edit(testLaunchModels("model-a", "model-a"))
 	if err != nil {
 		t.Fatalf("Edit with duplicates failed: %v", err)
 	}
@@ -388,7 +388,7 @@ func TestDroidEdit_MalformedModelEntry(t *testing.T) {
 	// Model entry is a string instead of a map
 	os.WriteFile(settingsPath, []byte(`{"customModels": ["not a map", 123]}`), 0o644)
 
-	err := d.Edit([]string{"model-a"})
+	err := d.Edit(testLaunchModels("model-a"))
 	if err != nil {
 		t.Fatalf("Edit with malformed entries failed: %v", err)
 	}
@@ -415,7 +415,7 @@ func TestDroidEdit_WrongTypeSessionSettings(t *testing.T) {
 	// sessionDefaultSettings is a string instead of map
 	os.WriteFile(settingsPath, []byte(`{"sessionDefaultSettings": "not a map"}`), 0o644)
 
-	err := d.Edit([]string{"model-a"})
+	err := d.Edit(testLaunchModels("model-a"))
 	if err != nil {
 		t.Fatalf("Edit with wrong type sessionDefaultSettings failed: %v", err)
 	}
@@ -490,7 +490,7 @@ func TestDroidEdit_RoundTrip(t *testing.T) {
 	os.WriteFile(settingsPath, []byte(testDroidSettingsFixture), 0o644)
 
 	// Edit with new models
-	if err := d.Edit([]string{"llama3", "mistral"}); err != nil {
+	if err := d.Edit(testLaunchModels("llama3", "mistral")); err != nil {
 		t.Fatal(err)
 	}
 
@@ -615,7 +615,7 @@ func TestDroidEdit_PreservesUnknownFields(t *testing.T) {
 		}`
 		os.WriteFile(settingsPath, []byte(original), 0o644)
 
-		if err := d.Edit([]string{"model-a"}); err != nil {
+		if err := d.Edit(testLaunchModels("model-a")); err != nil {
 			t.Fatal(err)
 		}
 
@@ -660,7 +660,7 @@ func TestDroidEdit_PreservesUnknownFields(t *testing.T) {
 		}`
 		os.WriteFile(settingsPath, []byte(original), 0o644)
 
-		if err := d.Edit([]string{"llama3"}); err != nil {
+		if err := d.Edit(testLaunchModels("llama3")); err != nil {
 			t.Fatal(err)
 		}
 
@@ -715,10 +715,10 @@ func TestDroidEdit_Idempotent(t *testing.T) {
 	os.WriteFile(settingsPath, []byte(testDroidSettingsFixture), 0o644)
 
 	// Edit twice with same models
-	d.Edit([]string{"llama3", "mistral"})
+	d.Edit(testLaunchModels("llama3", "mistral"))
 	firstData, _ := os.ReadFile(settingsPath)
 
-	d.Edit([]string{"llama3", "mistral"})
+	d.Edit(testLaunchModels("llama3", "mistral"))
 	secondData, _ := os.ReadFile(settingsPath)
 
 	// Results should be identical
@@ -744,7 +744,7 @@ func TestDroidEdit_MultipleConsecutiveEdits(t *testing.T) {
 		if i%2 == 0 {
 			models = []string{"model-x", "model-y", "model-z"}
 		}
-		if err := d.Edit(models); err != nil {
+		if err := d.Edit(launchModelsFromNames(models)); err != nil {
 			t.Fatalf("edit %d failed: %v", i, err)
 		}
 	}
@@ -803,7 +803,7 @@ func TestDroidEdit_UnicodeAndSpecialCharacters(t *testing.T) {
 	}`
 	os.WriteFile(settingsPath, []byte(original), 0o644)
 
-	if err := d.Edit([]string{"model-a"}); err != nil {
+	if err := d.Edit(testLaunchModels("model-a")); err != nil {
 		t.Fatal(err)
 	}
 
@@ -845,7 +845,7 @@ func TestDroidEdit_LargeNumbers(t *testing.T) {
 	}`
 	os.WriteFile(settingsPath, []byte(original), 0o644)
 
-	if err := d.Edit([]string{"model-a"}); err != nil {
+	if err := d.Edit(testLaunchModels("model-a")); err != nil {
 		t.Fatal(err)
 	}
 
@@ -889,7 +889,7 @@ func TestDroidEdit_EmptyAndNullValues(t *testing.T) {
 	}`
 	os.WriteFile(settingsPath, []byte(original), 0o644)
 
-	if err := d.Edit([]string{"model-a"}); err != nil {
+	if err := d.Edit(testLaunchModels("model-a")); err != nil {
 		t.Fatal(err)
 	}
 
@@ -943,7 +943,7 @@ func TestDroidEdit_DeeplyNestedStructures(t *testing.T) {
 	}`
 	os.WriteFile(settingsPath, []byte(original), 0o644)
 
-	if err := d.Edit([]string{"model-a"}); err != nil {
+	if err := d.Edit(testLaunchModels("model-a")); err != nil {
 		t.Fatal(err)
 	}
 
@@ -988,7 +988,7 @@ func TestDroidEdit_ModelNamesWithSpecialCharacters(t *testing.T) {
 		"model_with_underscores",
 	}
 
-	if err := d.Edit(specialModels); err != nil {
+	if err := d.Edit(launchModelsFromNames(specialModels)); err != nil {
 		t.Fatal(err)
 	}
 
@@ -1025,7 +1025,7 @@ func TestDroidEdit_MissingCustomModelsKey(t *testing.T) {
 		t.Fatal(err)
 	}
 
-	settings = updateDroidSettings(settings, settingsStruct, []string{"model-a"})
+	settings = updateDroidSettings(settings, settingsStruct, testLaunchModels("model-a"))
 
 	// Original fields preserved
 	if settings["diffMode"] != "github" {
@@ -1062,7 +1062,7 @@ func TestDroidEdit_NullCustomModels(t *testing.T) {
 	}`
 	os.WriteFile(settingsPath, []byte(original), 0o644)
 
-	if err := d.Edit([]string{"model-a"}); err != nil {
+	if err := d.Edit(testLaunchModels("model-a")); err != nil {
 		t.Fatal(err)
 	}
 
@@ -1090,7 +1090,7 @@ func TestDroidEdit_MinifiedJSON(t *testing.T) {
 	original := `{"diffMode":"github","enableHooks":true,"hooks":{"imported":["cmd1","cmd2"]},"customModels":[],"sessionDefaultSettings":{}}`
 	os.WriteFile(settingsPath, []byte(original), 0o644)
 
-	if err := d.Edit([]string{"model-a"}); err != nil {
+	if err := d.Edit(testLaunchModels("model-a")); err != nil {
 		t.Fatal(err)
 	}
 
@@ -1120,7 +1120,7 @@ func TestDroidEdit_CreatesDirectoryIfMissing(t *testing.T) {
 		t.Fatal("directory should not exist before test")
 	}
 
-	if err := d.Edit([]string{"model-a"}); err != nil {
+	if err := d.Edit(testLaunchModels("model-a")); err != nil {
 		t.Fatal(err)
 	}
 
@@ -1157,7 +1157,7 @@ func TestDroidEdit_PreservesFileAfterError(t *testing.T) {
 	os.WriteFile(settingsPath, []byte(original), 0o644)
 
 	// Empty models list is a no-op, should not modify file
-	d.Edit([]string{})
+	d.Edit(testLaunchModels())
 
 	data, _ := os.ReadFile(settingsPath)
 	if string(data) != original {
@@ -1181,7 +1181,7 @@ func TestDroidEdit_BackupCreated(t *testing.T) {
 	original := fmt.Sprintf(`{"diffMode": "%s", "customModels": [], "sessionDefaultSettings": {}}`, uniqueMarker)
 	os.WriteFile(settingsPath, []byte(original), 0o644)
 
-	if err := d.Edit([]string{"model-a"}); err != nil {
+	if err := d.Edit(testLaunchModels("model-a")); err != nil {
 		t.Fatal(err)
 	}
 
@@ -1231,7 +1231,7 @@ func TestDroidEdit_LargeNumberOfModels(t *testing.T) {
 		models = append(models, fmt.Sprintf("model-%d", i))
 	}
 
-	if err := d.Edit(models); err != nil {
+	if err := d.Edit(launchModelsFromNames(models)); err != nil {
 		t.Fatal(err)
 	}
 
@@ -1261,7 +1261,7 @@ func TestDroidEdit_LocalModelDefaultMaxOutput(t *testing.T) {
 	settingsDir := filepath.Join(tmpDir, ".factory")
 	settingsPath := filepath.Join(settingsDir, "settings.json")
 
-	if err := d.Edit([]string{"llama3.2"}); err != nil {
+	if err := d.Edit(testLaunchModels("llama3.2")); err != nil {
 		t.Fatal(err)
 	}
 
@@ -1312,7 +1312,7 @@ func TestDroidEdit_ArraysWithMixedTypes(t *testing.T) {
 	}`
 	os.WriteFile(settingsPath, []byte(original), 0o644)
 
-	if err := d.Edit([]string{"model-a"}); err != nil {
+	if err := d.Edit(testLaunchModels("model-a")); err != nil {
 		t.Fatal(err)
 	}
 

cmd/launch/hermes.go

@@ -65,7 +65,7 @@ type Hermes struct{}
 
 func (h *Hermes) String() string { return "Hermes Agent" }
 
-func (h *Hermes) Run(_ string, args []string) error {
+func (h *Hermes) Run(_ string, _ []LaunchModel, args []string) error {
 	// Hermes reads its primary model from config.yaml. launch configures that
 	// default model ahead of time so we can keep runtime invocation simple and
 	// still let Hermes discover additional models later via its own UX.

cmd/launch/hermes_test.go

@@ -552,7 +552,7 @@ func TestHermesRunPassthroughArgs(t *testing.T) {
 	}
 
 	h := &Hermes{}
-	if err := h.Run("", []string{"--continue"}); err != nil {
+	if err := h.Run("", nil, []string{"--continue"}); err != nil {
 		t.Fatalf("Run returned error: %v", err)
 	}
 
@@ -603,7 +603,7 @@ fi
 	}
 
 	h := &Hermes{}
-	if err := h.Run("", nil); err != nil {
+	if err := h.Run("", nil, nil); err != nil {
 		t.Fatalf("Run returned error: %v", err)
 	}
 
@@ -655,10 +655,10 @@ func TestHermesRun_SetUpLaterRepromptsOnLaterLaunches(t *testing.T) {
 	}
 
 	h := &Hermes{}
-	if err := h.Run("", nil); err != nil {
+	if err := h.Run("", nil, nil); err != nil {
 		t.Fatalf("first Run returned error: %v", err)
 	}
-	if err := h.Run("", nil); err != nil {
+	if err := h.Run("", nil, nil); err != nil {
 		t.Fatalf("second Run returned error: %v", err)
 	}
 
@@ -713,7 +713,7 @@ func TestHermesRun_SkipsMessagingPromptWhenConfigured(t *testing.T) {
 	}
 
 	h := &Hermes{}
-	if err := h.Run("", nil); err != nil {
+	if err := h.Run("", nil, nil); err != nil {
 		t.Fatalf("Run returned error: %v", err)
 	}
 
@@ -753,7 +753,7 @@ func TestHermesRun_SkipsMessagingPromptWithYesPolicy(t *testing.T) {
 	}
 
 	h := &Hermes{}
-	if err := h.Run("", nil); err != nil {
+	if err := h.Run("", nil, nil); err != nil {
 		t.Fatalf("Run returned error: %v", err)
 	}
 
@@ -798,7 +798,7 @@ fi
 	}
 
 	h := &Hermes{}
-	err := h.Run("", nil)
+	err := h.Run("", nil, nil)
 	if err == nil {
 		t.Fatal("expected messaging setup failure")
 	}

cmd/launch/integrations_test.go

@@ -25,7 +25,7 @@ type stubEditorRunner struct {
 	editErr  error
 }
 
-func (s *stubEditorRunner) Run(model string, args []string) error {
+func (s *stubEditorRunner) Run(model string, _ []LaunchModel, args []string) error {
 	s.ranModel = model
 	return nil
 }
@@ -34,11 +34,11 @@ func (s *stubEditorRunner) String() string { return "StubEditor" }
 
 func (s *stubEditorRunner) Paths() []string { return nil }
 
-func (s *stubEditorRunner) Edit(models []string) error {
+func (s *stubEditorRunner) Edit(models []LaunchModel) error {
 	if s.editErr != nil {
 		return s.editErr
 	}
-	cloned := append([]string(nil), models...)
+	cloned := launchModelNames(models)
 	s.edited = append(s.edited, cloned)
 	return nil
 }
@@ -58,6 +58,9 @@ func TestIntegrationLookup(t *testing.T) {
 		{"claude desktop", "claude-desktop", true, "Claude Desktop"},
 		{"claude desktop alias", "claude-app", true, "Claude Desktop"},
 		{"codex", "codex", true, "Codex"},
+		{"codex app", "codex-app", true, "Codex App"},
+		{"codex app desktop alias", "codex-desktop", true, "Codex App"},
+		{"codex app gui alias", "codex-gui", true, "Codex App"},
 		{"kimi", "kimi", true, "Kimi Code CLI"},
 		{"droid", "droid", true, "Droid"},
 		{"opencode", "opencode", true, "OpenCode"},
@@ -80,7 +83,7 @@ func TestIntegrationLookup(t *testing.T) {
 }
 
 func TestIntegrationRegistry(t *testing.T) {
-	expectedIntegrations := []string{"claude", "claude-desktop", "codex", "kimi", "droid", "opencode", "hermes", "pool"}
+	expectedIntegrations := []string{"claude", "claude-desktop", "codex", "codex-app", "kimi", "droid", "opencode", "hermes", "pool"}
 	for _, name := range expectedIntegrations {
 		t.Run(name, func(t *testing.T) {
 			r, ok := integrations[name]
@@ -203,7 +206,7 @@ func TestAllIntegrations_HaveRequiredMethods(t *testing.T) {
 			if displayName == "" {
 				t.Error("String() should not return empty")
 			}
-			var _ func(string, []string) error = r.Run
+			var _ func(string, []LaunchModel, []string) error = r.Run
 		})
 	}
 }
@@ -478,11 +481,11 @@ func TestBuildModelList_ExistingRecommendedMarked(t *testing.T) {
 func TestBuildModelList_PreservesRecommendationRequiredPlanForExistingCloudModel(t *testing.T) {
 	recommendations := []ModelItem{
 		{
-			Name:          "glm-5:cloud",
-			Description:   "Reasoning and code generation",
-			Recommended:   true,
-			RequiredPlan:  "pro",
-			ContextLength: 202_752,
+			Name:         "glm-5:cloud",
+			Description:  "Reasoning and code generation",
+			Recommended:  true,
+			RequiredPlan: "pro",
+			Details:      api.ModelDetails{ContextLength: 202_752},
 		},
 	}
 	existing := []modelInfo{{Name: "glm-5:cloud", Remote: true}}
@@ -863,7 +866,7 @@ func TestPrepareEditorIntegration_SavesOnlyAfterSuccessfulEdit(t *testing.T) {
 	}
 
 	editor := &stubEditorRunner{editErr: errors.New("boom")}
-	err := prepareEditorIntegration("droid", editor, []string{"new-model"})
+	err := prepareEditorIntegration("droid", editor, testLaunchModels("new-model"))
 	if err == nil || !strings.Contains(err.Error(), "setup failed") {
 		t.Fatalf("expected setup failure, got %v", err)
 	}
@@ -1737,6 +1740,11 @@ func TestIntegration_InstallHint(t *testing.T) {
 			input:   "codex",
 			wantURL: "https://developers.openai.com/codex/cli/",
 		},
+		{
+			name:    "codex app has hint",
+			input:   "codex-app",
+			wantURL: "https://developers.openai.com/codex/quickstart",
+		},
 		{
 			name:    "openclaw has hint",
 			input:   "openclaw",
@@ -1813,11 +1821,38 @@ func TestListIntegrationInfos(t *testing.T) {
 			}
 			want = filtered
 		}
+		if codexAppSupported() != nil {
+			filtered := make([]string, 0, len(want))
+			for _, name := range want {
+				if name != "codex-app" {
+					filtered = append(filtered, name)
+				}
+			}
+			want = filtered
+		}
+
 		if diff := compareStrings(got, want); diff != "" {
 			t.Fatalf("launcher integration order mismatch: %s", diff)
 		}
 	})
 
+	t.Run("prioritizes primary launcher integrations", func(t *testing.T) {
+		got := make([]string, 0, len(infos))
+		for _, info := range infos {
+			got = append(got, info.Name)
+		}
+		wantPrefix := []string{"claude", "codex-app", "hermes", "openclaw"}
+		if codexAppSupported() != nil {
+			wantPrefix = []string{"claude", "hermes", "openclaw", "opencode"}
+		}
+		if len(got) < len(wantPrefix) {
+			t.Fatalf("expected at least %d integrations, got %v", len(wantPrefix), got)
+		}
+		if diff := compareStrings(got[:len(wantPrefix)], wantPrefix); diff != "" {
+			t.Fatalf("unexpected primary launcher order: %s", diff)
+		}
+	})
+
 	t.Run("all fields populated", func(t *testing.T) {
 		for _, info := range infos {
 			if info.Name == "" {
@@ -1831,6 +1866,9 @@ func TestListIntegrationInfos(t *testing.T) {
 
 	t.Run("includes known integrations", func(t *testing.T) {
 		known := map[string]bool{"claude": false, "codex": false, "opencode": false}
+		if codexAppSupported() == nil {
+			known["codex-app"] = false
+		}
 		if poolsideGOOS != "windows" {
 			known["pool"] = false
 		}

cmd/launch/kimi.go

@@ -36,7 +36,7 @@ func (k *Kimi) args(config string, extra []string) []string {
 	return args
 }
 
-func (k *Kimi) Run(model string, args []string) error {
+func (k *Kimi) Run(model string, _ []LaunchModel, args []string) error {
 	if strings.TrimSpace(model) == "" {
 		return fmt.Errorf("model is required")
 	}

cmd/launch/kimi_test.go

@@ -307,7 +307,7 @@ func TestKimiRun_RejectsConflictingArgsBeforeInstall(t *testing.T) {
 	}
 	t.Cleanup(func() { DefaultConfirmPrompt = oldConfirm })
 
-	err := k.Run("llama3.2", []string{"--model", "other"})
+	err := k.Run("llama3.2", nil, []string{"--model", "other"})
 	if err == nil || !strings.Contains(err.Error(), "--model") {
 		t.Fatalf("expected conflict error mentioning --model, got %v", err)
 	}
@@ -337,7 +337,7 @@ exit 0
 	t.Setenv("OLLAMA_HOST", srv.URL)
 
 	k := &Kimi{}
-	if err := k.Run("llama3.2", []string{"--quiet", "--print"}); err != nil {
+	if err := k.Run("llama3.2", nil, []string{"--quiet", "--print"}); err != nil {
 		t.Fatalf("Run() error = %v", err)
 	}
 

cmd/launch/launch.go

@@ -12,6 +12,7 @@ import (
 
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/cmd/config"
+	modelpkg "github.com/ollama/ollama/types/model"
 	"github.com/spf13/cobra"
 	"golang.org/x/term"
 )
@@ -137,16 +138,17 @@ var isInteractiveSession = func() bool {
 	return term.IsTerminal(int(os.Stdin.Fd())) && term.IsTerminal(int(os.Stdout.Fd()))
 }
 
-// Runner executes a model with an integration.
+// Runner executes an integration with the selected model and its resolved
+// launch metadata. models is ordered with the primary model first.
 type Runner interface {
-	Run(model string, args []string) error
+	Run(model string, models []LaunchModel, args []string) error
 	String() string
 }
 
 // Editor can edit config files for integrations that support model configuration.
 type Editor interface {
 	Paths() []string
-	Edit(models []string) error
+	Edit(models []LaunchModel) error
 	Models() []string
 }
 
@@ -165,7 +167,7 @@ type ManagedSingleModel interface {
 // ManagedModelListConfigurer lets managed single-model integrations receive
 // the launcher's model list while still preserving one primary selected model.
 type ManagedModelListConfigurer interface {
-	ConfigureWithModels(primary string, models []string) error
+	ConfigureWithModels(primary string, models []LaunchModel) error
 }
 
 // ManagedAutodiscoveryIntegration is for managed integrations that do not need
@@ -239,24 +241,18 @@ type SupportedIntegration interface {
 	Supported() error
 }
 
-type modelInfo struct {
-	Name        string
-	Remote      bool
-	ToolCapable bool
-}
-
-// ModelInfo re-exports launcher model inventory details for callers.
-type ModelInfo = modelInfo
-
 // ModelItem represents model metadata before selector-only UI state is derived.
 type ModelItem struct {
 	Name            string
 	Description     string
 	Recommended     bool
 	VRAMBytes       int64
-	ContextLength   int
 	MaxOutputTokens int
 	RequiredPlan    string
+	ToolCapable     bool
+	Capabilities    []modelpkg.Capability
+	Size            int64
+	Details         api.ModelDetails
 }
 
 // SelectionItem represents a model row after launch has derived selector-only UI state.
@@ -285,22 +281,25 @@ Flags and extra arguments require an integration name.
 
 Supported integrations:
   claude          Claude Code
-  cline           Cline
+  codex-app       Codex App (aliases: codex-desktop, codex-gui)
+  hermes          Hermes Agent
+  openclaw        OpenClaw (aliases: clawdbot, moltbot)
+  opencode        OpenCode
   codex           Codex
   copilot         Copilot CLI (aliases: copilot-cli)
   droid           Droid
-  hermes          Hermes Agent
   kimi            Kimi Code CLI
-  opencode        OpenCode
-  openclaw        OpenClaw (aliases: clawdbot, moltbot)
   pi              Pi
   pool            Pool
+  cline           Cline
   vscode          VS Code (aliases: code)
 
 Examples:
   ollama launch
   ollama launch claude
   ollama launch claude --model <model>
+  ollama launch codex-app
+  ollama launch codex-app --restore
   ollama launch hermes
   ollama launch droid --config (does not auto-launch)
   ollama launch codex -- -p myprofile (pass extra args to integration)
@@ -406,8 +405,7 @@ func launchCommandIsClaudeDesktop(name string) bool {
 
 type launcherClient struct {
 	apiClient             *api.Client
-	modelInventory        []ModelInfo
-	inventoryLoaded       bool
+	inventory             *modelInventory
 	recommendationsLoaded bool
 	recommendationItems   []ModelItem
 	accountState          *AccountState
@@ -424,10 +422,18 @@ func newLauncherClient(policy LaunchPolicy) (*launcherClient, error) {
 
 	return &launcherClient{
 		apiClient: apiClient,
+		inventory: newModelInventory(apiClient),
 		policy:    policy,
 	}, nil
 }
 
+func (c *launcherClient) modelInventory() *modelInventory {
+	if c.inventory == nil {
+		c.inventory = newModelInventory(c.apiClient)
+	}
+	return c.inventory
+}
+
 // BuildLauncherState returns the launch-owned root launcher menu snapshot.
 func BuildLauncherState(ctx context.Context) (*LauncherState, error) {
 	launchClient, err := newLauncherClient(defaultLaunchPolicy(isInteractiveSession(), false))
@@ -549,7 +555,7 @@ func prepareIntegrationLaunch(name string, policy LaunchPolicy) (*launcherClient
 }
 
 func (c *launcherClient) buildLauncherState(ctx context.Context) (*LauncherState, error) {
-	_ = c.loadModelInventoryOnce(ctx)
+	_, _ = c.modelInventory().Load(ctx)
 
 	state := &LauncherState{
 		LastSelection: config.LastSelection(),
@@ -723,7 +729,7 @@ func (c *launcherClient) launchSingleIntegration(ctx context.Context, name strin
 		}
 	}
 
-	return launchAfterConfiguration(name, runner, target, req)
+	return launchAfterConfiguration(name, runner, target, c.resolveRunModels(ctx, []string{target}), req)
 }
 
 func (c *launcherClient) launchEditorIntegration(ctx context.Context, name string, runner Runner, editor Editor, saved *config.IntegrationConfig, req IntegrationLaunchRequest) error {
@@ -745,13 +751,17 @@ func (c *launcherClient) launchEditorIntegration(ctx context.Context, name strin
 		return nil
 	}
 
+	var launchModels []LaunchModel
 	if (needsConfigure || req.ModelOverride != "") && !savedMatchesModels(saved, models) {
-		if err := prepareEditorIntegration(name, editor, models); err != nil {
+		launchModels = c.modelInventory().Resolve(ctx, models)
+		if err := prepareEditorIntegration(name, editor, launchModels); err != nil {
 			return err
 		}
+	} else {
+		launchModels = c.resolveRunModels(ctx, models)
 	}
 
-	return launchAfterConfiguration(name, runner, models[0], req)
+	return launchAfterConfiguration(name, runner, models[0], launchModels, req)
 }
 
 func (c *launcherClient) launchManagedSingleIntegration(ctx context.Context, name string, runner Runner, managed ManagedSingleModel, saved *config.IntegrationConfig, req IntegrationLaunchRequest) error {
@@ -769,12 +779,18 @@ func (c *launcherClient) launchManagedSingleIntegration(ctx context.Context, nam
 		return nil
 	}
 
-	if needsConfigure || req.ModelOverride != "" || (current != "" && target != current) || !savedMatchesModels(saved, []string{target}) {
+	// current is the live managed app config; target may come from saved launch
+	// state. Rewrite when the live config is missing or has drifted so the app
+	// config converges with the model which launch is about to use.
+	liveConfigMissing := current == ""
+	liveConfigDrifted := current != "" && target != current
+	configured := false
+	if needsConfigure || req.ModelOverride != "" || liveConfigMissing || liveConfigDrifted || !savedMatchesModels(saved, []string{target}) {
 		configureModels, err := c.managedSingleConfigureModels(ctx, managed, target)
 		if err != nil {
 			return err
 		}
-		if err := prepareManagedSingleIntegration(name, managed, target, configureModels); err != nil {
+		if err := prepareManagedSingleIntegration(name, managed, target, c.modelInventory().Resolve(ctx, configureModels)); err != nil {
 			return err
 		}
 		if refresher, ok := managed.(ManagedRuntimeRefresher); ok {
@@ -782,6 +798,7 @@ func (c *launcherClient) launchManagedSingleIntegration(ctx context.Context, nam
 				return err
 			}
 		}
+		configured = true
 	}
 
 	if !managedIntegrationOnboarded(saved, managed) {
@@ -793,11 +810,17 @@ func (c *launcherClient) launchManagedSingleIntegration(ctx context.Context, nam
 		}
 	}
 
+	if configured {
+		if !printConfigurationSuccess(managed) {
+			printRestoreHint(managed)
+		}
+	}
+
 	if req.ConfigureOnly {
 		return nil
 	}
 
-	return runIntegration(runner, target, req.ExtraArgs)
+	return runIntegration(runner, target, c.resolveRunModels(ctx, []string{target}), req.ExtraArgs)
 }
 
 func (c *launcherClient) launchManagedAutodiscoveryIntegration(ctx context.Context, name string, runner Runner, autodiscovery ManagedAutodiscoveryIntegration, saved *config.IntegrationConfig, req IntegrationLaunchRequest) error {
@@ -840,7 +863,7 @@ func (c *launcherClient) launchManagedAutodiscoveryIntegration(ctx context.Conte
 		return nil
 	}
 
-	return runIntegration(runner, target, req.ExtraArgs)
+	return runIntegration(runner, target, c.resolveRunModels(ctx, []string{target}), req.ExtraArgs)
 }
 
 func (c *launcherClient) managedAutodiscoveryUsable(ctx context.Context, autodiscovery ManagedAutodiscoveryIntegration) bool {
@@ -941,7 +964,7 @@ func (c *launcherClient) resolveSingleIntegrationTarget(ctx context.Context, run
 		}
 	}
 
-	if needsConfigure {
+	if needsConfigure && req.ModelOverride == "" {
 		selected, err := c.selectSingleModelWithSelectorReady(ctx, fmt.Sprintf("Select model for %s:", runner), target, DefaultSingleSelector, !skipReadiness)
 		if err != nil {
 			return "", false, err
@@ -1092,13 +1115,14 @@ func runMultiSelector(title string, items []SelectionItem, preChecked []string,
 }
 
 func (c *launcherClient) loadSelectableModels(ctx context.Context, preChecked []string, current, emptyMessage string) ([]ModelItem, []string, error) {
-	if err := c.loadModelInventoryOnce(ctx); err != nil {
+	inventory, err := c.modelInventory().Load(ctx)
+	if err != nil {
 		return nil, nil, err
 	}
 	recommendations := c.recommendations(ctx)
 
 	cloudDisabled, _ := cloudStatusDisabled(ctx, c.apiClient)
-	items, orderedChecked, _, _ := buildModelListWithRecommendations(c.modelInventory, recommendations, preChecked, current)
+	items, orderedChecked, _, _ := buildModelListWithRecommendations(inventory, recommendations, preChecked, current)
 	if cloudDisabled {
 		items = filterCloudItems(items)
 		orderedChecked = c.filterDisabledCloudModels(ctx, orderedChecked)
@@ -1163,9 +1187,11 @@ func (c *launcherClient) requestRecommendations(ctx context.Context) ([]ModelIte
 			Description:     description,
 			Recommended:     true,
 			VRAMBytes:       rec.VRAMBytes,
-			ContextLength:   rec.ContextLength,
 			MaxOutputTokens: rec.MaxOutputTokens,
 			RequiredPlan:    strings.TrimSpace(rec.RequiredPlan),
+			Details: api.ModelDetails{
+				ContextLength: rec.ContextLength,
+			},
 		})
 	}
 
@@ -1286,10 +1312,11 @@ func (c *launcherClient) filterDisabledCloudModels(ctx context.Context, models [
 }
 
 func (c *launcherClient) savedModelUsable(ctx context.Context, name string) (bool, error) {
-	if err := c.loadModelInventoryOnce(ctx); err != nil {
+	inventory, err := c.modelInventory().Load(ctx)
+	if err != nil {
 		return c.showBasedModelUsable(ctx, name)
 	}
-	return c.singleModelUsable(ctx, name), nil
+	return c.singleModelUsable(ctx, name, inventory), nil
 }
 
 func (c *launcherClient) showBasedModelUsable(ctx context.Context, name string) (bool, error) {
@@ -1315,7 +1342,7 @@ func (c *launcherClient) showBasedModelUsable(ctx context.Context, name string)
 	return true, nil
 }
 
-func (c *launcherClient) singleModelUsable(ctx context.Context, name string) bool {
+func (c *launcherClient) singleModelUsable(ctx context.Context, name string, inventory []LaunchModel) bool {
 	if name == "" {
 		return false
 	}
@@ -1323,11 +1350,11 @@ func (c *launcherClient) singleModelUsable(ctx context.Context, name string) boo
 		cloudDisabled, _ := cloudStatusDisabled(ctx, c.apiClient)
 		return !cloudDisabled
 	}
-	return c.hasLocalModel(name)
+	return hasLocalModel(inventory, name)
 }
 
-func (c *launcherClient) hasLocalModel(name string) bool {
-	for _, model := range c.modelInventory {
+func hasLocalModel(inventory []LaunchModel, name string) bool {
+	for _, model := range inventory {
 		if model.Remote {
 			continue
 		}
@@ -1338,37 +1365,18 @@ func (c *launcherClient) hasLocalModel(name string) bool {
 	return false
 }
 
-func (c *launcherClient) loadModelInventoryOnce(ctx context.Context) error {
-	if c.inventoryLoaded {
-		return nil
-	}
-
-	resp, err := c.apiClient.List(ctx)
-	if err != nil {
-		return err
-	}
-
-	c.modelInventory = c.modelInventory[:0]
-	for _, model := range resp.Models {
-		c.modelInventory = append(c.modelInventory, ModelInfo{
-			Name:   model.Name,
-			Remote: model.RemoteModel != "",
-		})
-	}
-
-	cloudDisabled, _ := cloudStatusDisabled(ctx, c.apiClient)
-	if cloudDisabled {
-		c.modelInventory = filterCloudModels(c.modelInventory)
-	}
-	c.inventoryLoaded = true
-	return nil
+func (c *launcherClient) resolveRunModels(ctx context.Context, models []string) []LaunchModel {
+	return c.modelInventory().Resolve(ctx, models)
 }
 
-func runIntegration(runner Runner, modelName string, args []string) error {
-	return runner.Run(modelName, args)
+func runIntegration(runner Runner, modelName string, models []LaunchModel, args []string) error {
+	if len(models) == 0 && modelName != "" {
+		models = launchModelsFromNames([]string{modelName})
+	}
+	return runner.Run(modelName, models, args)
 }
 
-func launchAfterConfiguration(name string, runner Runner, model string, req IntegrationLaunchRequest) error {
+func launchAfterConfiguration(name string, runner Runner, model string, models []LaunchModel, req IntegrationLaunchRequest) error {
 	if req.ConfigureOnly {
 		launch, err := ConfirmPrompt(fmt.Sprintf("Launch %s now?", runner))
 		if err != nil {
@@ -1381,7 +1389,7 @@ func launchAfterConfiguration(name string, runner Runner, model string, req Inte
 	if err := EnsureIntegrationInstalled(name, runner); err != nil {
 		return err
 	}
-	return runIntegration(runner, model, req.ExtraArgs)
+	return runIntegration(runner, model, models, req.ExtraArgs)
 }
 
 func loadStoredIntegrationConfig(name string) (*config.IntegrationConfig, error) {

cmd/launch/launch_test.go

@@ -11,6 +11,7 @@ import (
 	"runtime"
 	"slices"
 	"strings"
+	"sync/atomic"
 	"testing"
 	"time"
 
@@ -24,7 +25,7 @@ type launcherEditorRunner struct {
 	ranModel string
 }
 
-func (r *launcherEditorRunner) Run(model string, args []string) error {
+func (r *launcherEditorRunner) Run(model string, _ []LaunchModel, args []string) error {
 	r.ranModel = model
 	return nil
 }
@@ -33,8 +34,8 @@ func (r *launcherEditorRunner) String() string { return "LauncherEditor" }
 
 func (r *launcherEditorRunner) Paths() []string { return r.paths }
 
-func (r *launcherEditorRunner) Edit(models []string) error {
-	r.edited = append(r.edited, append([]string(nil), models...))
+func (r *launcherEditorRunner) Edit(models []LaunchModel) error {
+	r.edited = append(r.edited, launchModelNames(models))
 	return nil
 }
 
@@ -44,7 +45,7 @@ type launcherSingleRunner struct {
 	ranModel string
 }
 
-func (r *launcherSingleRunner) Run(model string, args []string) error {
+func (r *launcherSingleRunner) Run(model string, _ []LaunchModel, args []string) error {
 	r.ranModel = model
 	return nil
 }
@@ -68,18 +69,21 @@ func (r *launcherRestorableRunner) RestoreSuccessMessage() string {
 }
 
 type launcherManagedRunner struct {
-	paths              []string
-	currentModel       string
-	configured         []string
-	ranModel           string
-	onboarded          bool
-	onboardCalls       int
-	onboardingComplete bool
-	refreshCalls       int
-	refreshErr         error
+	paths                []string
+	currentModel         string
+	configured           []string
+	ranModel             string
+	onboarded            bool
+	onboardCalls         int
+	onboardingComplete   bool
+	refreshCalls         int
+	refreshErr           error
+	restoreHint          string
+	configSuccessMessage string
+	skipModelReadiness   bool
 }
 
-func (r *launcherManagedRunner) Run(model string, args []string) error {
+func (r *launcherManagedRunner) Run(model string, _ []LaunchModel, args []string) error {
 	r.ranModel = model
 	return nil
 }
@@ -110,6 +114,14 @@ func (r *launcherManagedRunner) RefreshRuntimeAfterConfigure() error {
 	return r.refreshErr
 }
 
+func (r *launcherManagedRunner) RestoreHint() string { return r.restoreHint }
+
+func (r *launcherManagedRunner) ConfigurationSuccessMessage() string {
+	return r.configSuccessMessage
+}
+
+func (r *launcherManagedRunner) SkipModelReadiness() bool { return r.skipModelReadiness }
+
 type launcherHeadlessManagedRunner struct {
 	launcherManagedRunner
 }
@@ -121,8 +133,8 @@ type launcherManagedListRunner struct {
 	configuredModelLists [][]string
 }
 
-func (r *launcherManagedListRunner) ConfigureWithModels(primary string, models []string) error {
-	r.configuredModelLists = append(r.configuredModelLists, append([]string(nil), models...))
+func (r *launcherManagedListRunner) ConfigureWithModels(primary string, models []LaunchModel) error {
+	r.configuredModelLists = append(r.configuredModelLists, launchModelNames(models))
 	return r.Configure(primary)
 }
 
@@ -480,6 +492,116 @@ func TestLaunchIntegration_ManagedSingleIntegrationConfigOnlySkipsFinalRun(t *te
 	}
 }
 
+func TestLaunchIntegration_ManagedSingleIntegrationPrintsConfigurationSuccessAfterConfigure(t *testing.T) {
+	tmpDir := t.TempDir()
+	setLaunchTestHome(t, tmpDir)
+	withInteractiveSession(t, true)
+	withLauncherHooks(t)
+
+	runner := &launcherManagedRunner{
+		configSuccessMessage: "configured successfully\nrestore via success message",
+		restoreHint:          "run restore command",
+		skipModelReadiness:   true,
+	}
+	withIntegrationOverride(t, "stubmanaged", runner)
+
+	stderr := captureStderr(t, func() {
+		if err := LaunchIntegration(context.Background(), IntegrationLaunchRequest{
+			Name:           "stubmanaged",
+			ModelOverride:  "gemma4",
+			ForceConfigure: true,
+			ConfigureOnly:  true,
+		}); err != nil {
+			t.Fatalf("LaunchIntegration returned error: %v", err)
+		}
+	})
+
+	if diff := compareStrings(runner.configured, []string{"gemma4"}); diff != "" {
+		t.Fatalf("configured models mismatch: %s", diff)
+	}
+	if !strings.Contains(stderr, "configured successfully") {
+		t.Fatalf("expected configuration success in stderr, got %q", stderr)
+	}
+	if !strings.Contains(stderr, "restore via success message") {
+		t.Fatalf("expected restore guidance in configuration success, got %q", stderr)
+	}
+	if strings.Contains(stderr, "run restore command") {
+		t.Fatalf("restore hint should not print separately after configure, got %q", stderr)
+	}
+}
+
+func TestLaunchIntegration_ManagedSingleIntegrationDoesNotPrintRestoreHintWhenUnchanged(t *testing.T) {
+	tmpDir := t.TempDir()
+	setLaunchTestHome(t, tmpDir)
+	withInteractiveSession(t, true)
+	withLauncherHooks(t)
+
+	runner := &launcherManagedRunner{
+		currentModel:         "gemma4",
+		onboardingComplete:   true,
+		configSuccessMessage: "configured successfully",
+		restoreHint:          "run restore command",
+		skipModelReadiness:   true,
+	}
+	withIntegrationOverride(t, "stubmanaged", runner)
+
+	if err := config.SaveIntegration("stubmanaged", []string{"gemma4"}); err != nil {
+		t.Fatalf("failed to save managed integration config: %v", err)
+	}
+	if err := config.MarkIntegrationOnboarded("stubmanaged"); err != nil {
+		t.Fatalf("failed to mark integration onboarded: %v", err)
+	}
+
+	stderr := captureStderr(t, func() {
+		if err := LaunchIntegration(context.Background(), IntegrationLaunchRequest{Name: "stubmanaged"}); err != nil {
+			t.Fatalf("LaunchIntegration returned error: %v", err)
+		}
+	})
+
+	if len(runner.configured) != 0 {
+		t.Fatalf("expected Configure to be skipped when saved matches, got %v", runner.configured)
+	}
+	if strings.Contains(stderr, "configured successfully") {
+		t.Fatalf("configuration success should not print when config is unchanged, got %q", stderr)
+	}
+	if strings.Contains(stderr, "run restore command") {
+		t.Fatalf("restore hint should not print when config is unchanged, got %q", stderr)
+	}
+}
+
+func TestLaunchIntegration_ManagedSingleIntegrationForceConfigureUsesModelOverride(t *testing.T) {
+	tmpDir := t.TempDir()
+	setLaunchTestHome(t, tmpDir)
+	withInteractiveSession(t, true)
+	withLauncherHooks(t)
+
+	runner := &launcherManagedRunner{
+		paths:              nil,
+		skipModelReadiness: true,
+	}
+	withIntegrationOverride(t, "stubmanaged", runner)
+
+	DefaultSingleSelector = func(title string, items []SelectionItem, current string) (string, error) {
+		return "", fmt.Errorf("selector should not run with an explicit model override")
+	}
+
+	if err := LaunchIntegration(context.Background(), IntegrationLaunchRequest{
+		Name:           "stubmanaged",
+		ModelOverride:  "gemma4",
+		ForceConfigure: true,
+		ConfigureOnly:  true,
+	}); err != nil {
+		t.Fatalf("LaunchIntegration returned error: %v", err)
+	}
+
+	if diff := compareStrings(runner.configured, []string{"gemma4"}); diff != "" {
+		t.Fatalf("configured models mismatch: %s", diff)
+	}
+	if runner.ranModel != "" {
+		t.Fatalf("expected configure-only flow to skip final launch, got %q", runner.ranModel)
+	}
+}
+
 func TestLaunchIntegration_ManagedSingleIntegrationSkipsRewriteWhenSavedMatches(t *testing.T) {
 	tmpDir := t.TempDir()
 	setLaunchTestHome(t, tmpDir)
@@ -505,7 +627,9 @@ func TestLaunchIntegration_ManagedSingleIntegrationSkipsRewriteWhenSavedMatches(
 		t.Fatalf("failed to save managed integration config: %v", err)
 	}
 
-	runner := &launcherManagedRunner{}
+	runner := &launcherManagedRunner{
+		currentModel: "gemma4",
+	}
 	withIntegrationOverride(t, "stubmanaged", runner)
 
 	DefaultSingleSelector = func(title string, items []SelectionItem, current string) (string, error) {
@@ -532,6 +656,53 @@ func TestLaunchIntegration_ManagedSingleIntegrationSkipsRewriteWhenSavedMatches(
 	}
 }
 
+func TestLaunchIntegration_ManagedSingleIntegrationRewritesWhenSavedMatchesButLiveConfigMissing(t *testing.T) {
+	tmpDir := t.TempDir()
+	setLaunchTestHome(t, tmpDir)
+	withInteractiveSession(t, true)
+	withLauncherHooks(t)
+
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		switch r.URL.Path {
+		case "/api/show":
+			fmt.Fprint(w, `{"model_info":{"general.context_length":131072}}`)
+		default:
+			http.NotFound(w, r)
+		}
+	}))
+	defer srv.Close()
+	t.Setenv("OLLAMA_HOST", srv.URL)
+
+	if err := config.SaveIntegration("stubmanaged", []string{"gemma4"}); err != nil {
+		t.Fatalf("failed to save managed integration config: %v", err)
+	}
+
+	runner := &launcherManagedRunner{}
+	withIntegrationOverride(t, "stubmanaged", runner)
+
+	DefaultSingleSelector = func(title string, items []SelectionItem, current string) (string, error) {
+		t.Fatal("selector should not be called when saved model is usable")
+		return "", nil
+	}
+	DefaultConfirmPrompt = func(prompt string, options ConfirmOptions) (bool, error) {
+		return true, nil
+	}
+
+	if err := LaunchIntegration(context.Background(), IntegrationLaunchRequest{Name: "stubmanaged"}); err != nil {
+		t.Fatalf("LaunchIntegration returned error: %v", err)
+	}
+
+	if diff := compareStrings(runner.configured, []string{"gemma4"}); diff != "" {
+		t.Fatalf("expected Configure to rewrite missing live config: %s", diff)
+	}
+	if runner.refreshCalls != 1 {
+		t.Fatalf("expected runtime refresh once after rewrite, got %d", runner.refreshCalls)
+	}
+	if runner.ranModel != "gemma4" {
+		t.Fatalf("expected launch to run saved model, got %q", runner.ranModel)
+	}
+}
+
 func TestLaunchIntegration_ManagedSingleIntegrationRewritesWhenSavedDiffers(t *testing.T) {
 	tmpDir := t.TempDir()
 	setLaunchTestHome(t, tmpDir)
@@ -1292,6 +1463,53 @@ func TestBuildLauncherState_ToleratesInventoryFailure(t *testing.T) {
 	}
 }
 
+func TestBuildLauncherState_UsesTagsInventoryWithoutShow(t *testing.T) {
+	tmpDir := t.TempDir()
+	setLaunchTestHome(t, tmpDir)
+
+	if err := config.SetLastModel("llama3.2"); err != nil {
+		t.Fatalf("failed to seed last model: %v", err)
+	}
+	if err := config.SaveIntegration("codex", []string{"qwen3:8b"}); err != nil {
+		t.Fatalf("failed to seed codex config: %v", err)
+	}
+
+	var showCalls atomic.Int32
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		switch r.URL.Path {
+		case "/api/tags":
+			fmt.Fprint(w, `{"models":[`+
+				`{"name":"llama3.2","capabilities":["completion","tools"],"context_length":131072,"size":3200000000},`+
+				`{"name":"qwen3:8b","capabilities":["completion","tools"],"context_length":65536,"size":4500000000}`+
+				`]}`)
+		case "/api/show":
+			showCalls.Add(1)
+			fmt.Fprint(w, `{"model_info":{"general.context_length":131072}}`)
+		default:
+			http.NotFound(w, r)
+		}
+	}))
+	defer srv.Close()
+	t.Setenv("OLLAMA_HOST", srv.URL)
+
+	state, err := BuildLauncherState(context.Background())
+	if err != nil {
+		t.Fatalf("BuildLauncherState returned error: %v", err)
+	}
+	if !state.RunModelUsable {
+		t.Fatal("expected saved run model to be usable from tags inventory")
+	}
+	if state.Integrations["codex"].CurrentModel != "qwen3:8b" {
+		t.Fatalf("expected codex current model from saved config, got %q", state.Integrations["codex"].CurrentModel)
+	}
+	if !state.Integrations["codex"].ModelUsable {
+		t.Fatal("expected saved codex model to be usable from tags inventory")
+	}
+	if got := showCalls.Load(); got != 0 {
+		t.Fatalf("show calls = %d, want 0 for broad launcher state", got)
+	}
+}
+
 func TestResolveRunModel_UsesSavedModelWithoutSelector(t *testing.T) {
 	tmpDir := t.TempDir()
 	setLaunchTestHome(t, tmpDir)

cmd/launch/model_inventory.go

@@ -0,0 +1,201 @@
+package launch
+
+import (
+	"context"
+	"slices"
+	"strings"
+	"sync"
+
+	"github.com/ollama/ollama/api"
+	modelpkg "github.com/ollama/ollama/types/model"
+)
+
+// LaunchModel is the model metadata Launch passes to integration config
+// writers after resolving selected model names through the per-run inventory.
+type LaunchModel struct {
+	Name            string
+	Remote          bool
+	ToolCapable     bool
+	Capabilities    []modelpkg.Capability
+	ContextLength   int
+	MaxOutputTokens int
+	EmbeddingLength int
+	Size            int64
+	Details         api.ModelDetails
+}
+
+type modelInfo = LaunchModel
+
+// ModelInfo re-exports launcher model inventory details for callers.
+type ModelInfo = LaunchModel
+
+func (m LaunchModel) HasCapability(capability modelpkg.Capability) bool {
+	return slices.Contains(m.Capabilities, capability)
+}
+
+func (m LaunchModel) WithCloudLimits() LaunchModel {
+	if limit, ok := lookupCloudModelLimit(m.Name); ok {
+		if m.ContextLength <= 0 {
+			m.ContextLength = limit.Context
+		}
+		if m.MaxOutputTokens <= 0 {
+			m.MaxOutputTokens = limit.Output
+		}
+	}
+	return m
+}
+
+type modelInventory struct {
+	client *api.Client
+
+	mu     sync.Mutex
+	loaded bool
+	models []LaunchModel
+	err    error
+}
+
+func newModelInventory(client *api.Client) *modelInventory {
+	return &modelInventory{client: client}
+}
+
+func (i *modelInventory) Load(ctx context.Context) ([]LaunchModel, error) {
+	return i.load(ctx, false)
+}
+
+func (i *modelInventory) Refresh(ctx context.Context) ([]LaunchModel, error) {
+	return i.load(ctx, true)
+}
+
+func (i *modelInventory) load(ctx context.Context, force bool) ([]LaunchModel, error) {
+	if i == nil || i.client == nil {
+		return nil, nil
+	}
+
+	i.mu.Lock()
+	defer i.mu.Unlock()
+
+	if i.loaded && !force {
+		return cloneLaunchModels(i.models), i.err
+	}
+
+	resp, err := i.client.List(ctx)
+	if err != nil {
+		i.models = nil
+		i.err = err
+		i.loaded = true
+		return nil, err
+	}
+
+	i.models = make([]LaunchModel, 0, len(resp.Models))
+	for _, model := range resp.Models {
+		i.models = append(i.models, launchModelFromListResponse(model))
+	}
+	i.err = nil
+	i.loaded = true
+
+	return cloneLaunchModels(i.models), i.err
+}
+
+func (i *modelInventory) Resolve(ctx context.Context, names []string) []LaunchModel {
+	names = dedupeModelList(names)
+	if len(names) == 0 {
+		return nil
+	}
+
+	models, err := i.Load(ctx)
+	if err != nil {
+		models = nil
+	}
+
+	resolved, localMiss := resolveLaunchModels(names, models)
+	if localMiss {
+		if refreshed, err := i.Refresh(ctx); err == nil {
+			resolved, _ = resolveLaunchModels(names, refreshed)
+		}
+	}
+	return resolved
+}
+
+func resolveLaunchModels(names []string, models []LaunchModel) ([]LaunchModel, bool) {
+	resolved := make([]LaunchModel, 0, len(names))
+	localMiss := false
+	for _, name := range names {
+		if model, ok := findLaunchModel(models, name); ok {
+			resolved = append(resolved, model.WithCloudLimits())
+			continue
+		}
+		if !isCloudModelName(name) {
+			localMiss = true
+		}
+		resolved = append(resolved, fallbackLaunchModel(name))
+	}
+	return resolved, localMiss
+}
+
+func launchModelFromListResponse(model api.ListModelResponse) LaunchModel {
+	return LaunchModel{
+		Name:            model.Name,
+		Remote:          model.RemoteModel != "",
+		ToolCapable:     slices.Contains(model.Capabilities, modelpkg.CapabilityTools),
+		Capabilities:    append([]modelpkg.Capability(nil), model.Capabilities...),
+		ContextLength:   model.Details.ContextLength,
+		EmbeddingLength: model.Details.EmbeddingLength,
+		Size:            model.Size,
+		Details:         model.Details,
+	}.WithCloudLimits()
+}
+
+func fallbackLaunchModel(name string) LaunchModel {
+	return LaunchModel{Name: name, Remote: isCloudModelName(name)}.WithCloudLimits()
+}
+
+func findLaunchModel(models []LaunchModel, name string) (LaunchModel, bool) {
+	for _, model := range models {
+		if launchModelMatches(model.Name, name) {
+			return cloneLaunchModel(model), true
+		}
+	}
+	return LaunchModel{}, false
+}
+
+func launchModelMatches(candidate, name string) bool {
+	if candidate == name {
+		return true
+	}
+	return strings.TrimSuffix(candidate, ":latest") == name
+}
+
+func cloneLaunchModel(model LaunchModel) LaunchModel {
+	model.Capabilities = append([]modelpkg.Capability(nil), model.Capabilities...)
+	model.Details.Families = append([]string(nil), model.Details.Families...)
+	return model
+}
+
+func cloneLaunchModels(models []LaunchModel) []LaunchModel {
+	cloned := make([]LaunchModel, len(models))
+	for i, model := range models {
+		cloned[i] = cloneLaunchModel(model)
+	}
+	return cloned
+}
+
+func launchModelNames(models []LaunchModel) []string {
+	names := make([]string, 0, len(models))
+	for _, model := range models {
+		if model.Name != "" {
+			names = append(names, model.Name)
+		}
+	}
+	return names
+}
+
+func launchModelsFromNames(names []string) []LaunchModel {
+	models := make([]LaunchModel, 0, len(names))
+	for _, name := range names {
+		if name == "" {
+			continue
+		}
+		models = append(models, fallbackLaunchModel(name))
+	}
+	return models
+}

cmd/launch/model_inventory_test.go

@@ -0,0 +1,80 @@
+package launch
+
+import (
+	"context"
+	"fmt"
+	"net/http"
+	"net/http/httptest"
+	"net/url"
+	"testing"
+
+	"github.com/ollama/ollama/api"
+	modelpkg "github.com/ollama/ollama/types/model"
+)
+
+func TestModelInventoryResolveRefreshesLocalMiss(t *testing.T) {
+	calls := 0
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path != "/api/tags" {
+			http.NotFound(w, r)
+			return
+		}
+		calls++
+		if calls == 1 {
+			fmt.Fprint(w, `{"models":[]}`)
+			return
+		}
+		fmt.Fprint(w, `{"models":[{"name":"new-model","size":123,"details":{"context_length":65536,"embedding_length":1024},"capabilities":["vision","tools"]}]}`)
+	}))
+	defer srv.Close()
+
+	u, _ := url.Parse(srv.URL)
+	inventory := newModelInventory(api.NewClient(u, srv.Client()))
+
+	got := inventory.Resolve(context.Background(), []string{"new-model"})
+	if calls != 2 {
+		t.Fatalf("List calls = %d, want 2", calls)
+	}
+	if len(got) != 1 {
+		t.Fatalf("Resolve returned %d models, want 1", len(got))
+	}
+	if got[0].Name != "new-model" {
+		t.Fatalf("Name = %q, want new-model", got[0].Name)
+	}
+	if got[0].ContextLength != 65_536 || got[0].EmbeddingLength != 1_024 {
+		t.Fatalf("metadata = context %d embedding %d, want refreshed metadata", got[0].ContextLength, got[0].EmbeddingLength)
+	}
+	if !got[0].HasCapability(modelpkg.CapabilityVision) || !got[0].ToolCapable {
+		t.Fatalf("capabilities = %v toolCapable=%v, want refreshed capabilities", got[0].Capabilities, got[0].ToolCapable)
+	}
+}
+
+func TestModelInventoryResolveDoesNotRefreshCloudMiss(t *testing.T) {
+	calls := 0
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path != "/api/tags" {
+			http.NotFound(w, r)
+			return
+		}
+		calls++
+		fmt.Fprint(w, `{"models":[]}`)
+	}))
+	defer srv.Close()
+
+	u, _ := url.Parse(srv.URL)
+	inventory := newModelInventory(api.NewClient(u, srv.Client()))
+
+	got := inventory.Resolve(context.Background(), []string{"glm-5.1:cloud"})
+	if calls != 1 {
+		t.Fatalf("List calls = %d, want 1", calls)
+	}
+	if len(got) != 1 {
+		t.Fatalf("Resolve returned %d models, want 1", len(got))
+	}
+	if got[0].Name != "glm-5.1:cloud" || !got[0].Remote {
+		t.Fatalf("resolved model = %#v, want cloud fallback", got[0])
+	}
+	if got[0].ContextLength <= 0 || got[0].MaxOutputTokens <= 0 {
+		t.Fatalf("cloud limits not applied: %#v", got[0])
+	}
+}

cmd/launch/models.go

@@ -23,10 +23,10 @@ import (
 )
 
 var recommendedModels = []ModelItem{
-	{Name: "kimi-k2.6:cloud", Description: "State-of-the-art coding, long-horizon execution, and multimodal agent swarm capability", Recommended: true, ContextLength: 262_144, MaxOutputTokens: 262_144},
-	{Name: "qwen3.5:cloud", Description: "Reasoning, coding, and agentic tool use with vision", Recommended: true, ContextLength: 262_144, MaxOutputTokens: 32_768},
-	{Name: "glm-5.1:cloud", Description: "Reasoning and code generation", Recommended: true, ContextLength: 202_752, MaxOutputTokens: 131_072},
-	{Name: "minimax-m2.7:cloud", Description: "Fast, efficient coding and real-world productivity", Recommended: true, ContextLength: 204_800, MaxOutputTokens: 128_000},
+	{Name: "kimi-k2.6:cloud", Description: "State-of-the-art coding, long-horizon execution, and multimodal agent swarm capability", Recommended: true, Details: api.ModelDetails{ContextLength: 262_144}, MaxOutputTokens: 262_144},
+	{Name: "qwen3.5:cloud", Description: "Reasoning, coding, and agentic tool use with vision", Recommended: true, Details: api.ModelDetails{ContextLength: 262_144}, MaxOutputTokens: 32_768},
+	{Name: "glm-5.1:cloud", Description: "Reasoning and code generation", Recommended: true, Details: api.ModelDetails{ContextLength: 202_752}, MaxOutputTokens: 131_072},
+	{Name: "minimax-m2.7:cloud", Description: "Fast, efficient coding and real-world productivity", Recommended: true, Details: api.ModelDetails{ContextLength: 204_800}, MaxOutputTokens: 128_000},
 	{Name: "gemma4", Description: "Reasoning and code generation locally", Recommended: true, VRAMBytes: 12 * format.GigaByte},
 	{Name: "qwen3.5", Description: "Reasoning, coding, and visual understanding locally", Recommended: true, VRAMBytes: 14 * format.GigaByte},
 }
@@ -115,7 +115,7 @@ func setDynamicCloudModelLimits(limits map[string]cloudModelLimit) {
 func cloudModelLimitsFromRecommendations(recommendations []ModelItem) map[string]cloudModelLimit {
 	limits := make(map[string]cloudModelLimit, len(recommendations))
 	for _, rec := range recommendations {
-		if !isCloudModelName(rec.Name) || rec.ContextLength <= 0 || rec.MaxOutputTokens <= 0 {
+		if !isCloudModelName(rec.Name) || rec.Details.ContextLength <= 0 || rec.MaxOutputTokens <= 0 {
 			continue
 		}
 		base, stripped := modelref.StripCloudSourceTag(rec.Name)
@@ -123,7 +123,7 @@ func cloudModelLimitsFromRecommendations(recommendations []ModelItem) map[string
 			continue
 		}
 		limits[base] = cloudModelLimit{
-			Context: rec.ContextLength,
+			Context: rec.Details.ContextLength,
 			Output:  rec.MaxOutputTokens,
 		}
 	}
@@ -299,18 +299,17 @@ func pullMissingModel(ctx context.Context, client *api.Client, model string) err
 }
 
 // prepareEditorIntegration persists models and applies editor-managed config files.
-func prepareEditorIntegration(name string, editor Editor, models []string) error {
+func prepareEditorIntegration(name string, editor Editor, models []LaunchModel) error {
 	if err := editor.Edit(models); err != nil {
 		return fmt.Errorf("setup failed: %w", err)
 	}
-	if err := config.SaveIntegration(name, models); err != nil {
+	if err := config.SaveIntegration(name, launchModelNames(models)); err != nil {
 		return fmt.Errorf("failed to save: %w", err)
 	}
 	return nil
 }
 
-func prepareManagedSingleIntegration(name string, managed ManagedSingleModel, model string, models []string) error {
-	models = dedupeModelList(append([]string{model}, models...))
+func prepareManagedSingleIntegration(name string, managed ManagedSingleModel, model string, models []LaunchModel) error {
 	var err error
 	if withModels, ok := managed.(ManagedModelListConfigurer); ok {
 		err = withModels.ConfigureWithModels(model, models)
@@ -365,11 +364,11 @@ func buildModelListWithRecommendations(existing []modelInfo, recommendations []M
 		}
 		displayName := strings.TrimSuffix(m.Name, ":latest")
 		existingModels[displayName] = true
-		item := ModelItem{Name: displayName, Recommended: recommended[displayName], Description: recDesc[displayName]}
 		if rec, ok := recByName[displayName]; ok {
-			item = copyModelRecommendationFields(displayName, rec)
+			items = append(items, modelItemFromInventory(displayName, m, copyModelRecommendationFields(displayName, rec)))
+		} else {
+			items = append(items, modelItemFromInventory(displayName, m, ModelItem{Name: displayName, Recommended: recommended[displayName], Description: recDesc[displayName]}))
 		}
-		items = append(items, item)
 	}
 
 	for _, rec := range recommendations {
@@ -483,22 +482,20 @@ func copyModelRecommendationFields(name string, rec ModelItem) ModelItem {
 	return rec
 }
 
+func modelItemFromInventory(name string, info modelInfo, item ModelItem) ModelItem {
+	item.Name = name
+	item.ToolCapable = info.ToolCapable
+	item.Capabilities = slices.Clone(info.Capabilities)
+	item.Size = info.Size
+	item.Details = info.Details
+	return item
+}
+
 // isCloudModelName reports whether the model name has an explicit cloud source.
 func isCloudModelName(name string) bool {
 	return modelref.HasExplicitCloudSource(name)
 }
 
-// filterCloudModels drops remote-only models from the given inventory.
-func filterCloudModels(existing []modelInfo) []modelInfo {
-	filtered := existing[:0]
-	for _, m := range existing {
-		if !m.Remote {
-			filtered = append(filtered, m)
-		}
-	}
-	return filtered
-}
-
 // filterCloudItems removes cloud models from selection items.
 func filterCloudItems(items []ModelItem) []ModelItem {
 	filtered := items[:0]

cmd/launch/models_test.go

@@ -0,0 +1,83 @@
+package launch
+
+import (
+	"testing"
+
+	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/format"
+	modelpkg "github.com/ollama/ollama/types/model"
+)
+
+func TestBuildModelList_UsesInventoryMetadataForInstalledModels(t *testing.T) {
+	existing := []modelInfo{
+		{
+			Name:         "custom-tools:latest",
+			ToolCapable:  true,
+			Capabilities: []modelpkg.Capability{modelpkg.CapabilityCompletion, modelpkg.CapabilityTools, modelpkg.CapabilityThinking},
+			Size:         7500 * format.MegaByte,
+			Details: api.ModelDetails{
+				ParameterSize:     "8B",
+				QuantizationLevel: "Q4_K_M",
+				ContextLength:     131_072,
+				EmbeddingLength:   4096,
+			},
+		},
+	}
+
+	items, _, _, _ := buildModelList(existing, nil, "")
+	var got ModelItem
+	for _, item := range items {
+		if item.Name == "custom-tools" {
+			got = item
+			break
+		}
+	}
+	if got.Name == "" {
+		t.Fatal("custom-tools not found in items")
+	}
+	if !got.ToolCapable {
+		t.Fatal("expected installed model to preserve tool capability from tags metadata")
+	}
+	if got.Details.ContextLength != 131_072 {
+		t.Fatalf("Details.ContextLength = %d, want 131072", got.Details.ContextLength)
+	}
+	if got.Size != 7500*format.MegaByte {
+		t.Fatalf("Size = %d, want %d", got.Size, 7500*format.MegaByte)
+	}
+	if got.Description != "" {
+		t.Fatalf("Description = %q, want empty for installed model without recommendation copy", got.Description)
+	}
+}
+
+func TestBuildModelList_InstalledRecommendedPreservesRecommendationAndMetadata(t *testing.T) {
+	existing := []modelInfo{
+		{
+			Name:         "qwen3.5",
+			ToolCapable:  true,
+			Capabilities: []modelpkg.Capability{modelpkg.CapabilityCompletion, modelpkg.CapabilityTools, modelpkg.CapabilityVision},
+			Size:         14 * format.GigaByte,
+			Details:      api.ModelDetails{ContextLength: 262_144},
+		},
+	}
+
+	items, _, _, _ := buildModelList(existing, nil, "")
+	var got ModelItem
+	for _, item := range items {
+		if item.Name == "qwen3.5" {
+			got = item
+			break
+		}
+	}
+	if got.Name == "" {
+		t.Fatal("qwen3.5 not found in items")
+	}
+	if !got.Recommended || !got.ToolCapable {
+		t.Fatalf("recommended/tool metadata = %v/%v, want true/true", got.Recommended, got.ToolCapable)
+	}
+	if got.Details.ContextLength != 262_144 {
+		t.Fatalf("Details.ContextLength = %d, want 262144", got.Details.ContextLength)
+	}
+	if got.Description != "Reasoning, coding, and visual understanding locally" {
+		t.Fatalf("Description = %q, want recommendation description", got.Description)
+	}
+}

cmd/launch/openclaw.go

@@ -1,7 +1,6 @@
 package launch
 
 import (
-	"context"
 	"encoding/json"
 	"fmt"
 	"net"
@@ -10,21 +9,15 @@ import (
 	"os/exec"
 	"path/filepath"
 	"runtime"
-	"slices"
 	"strings"
 	"time"
 
-	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/cmd/internal/fileutil"
 	"github.com/ollama/ollama/envconfig"
-	"github.com/ollama/ollama/types/model"
 )
 
 const defaultGatewayPort = 18789
 
-// Bound model capability probing so launch/config cannot hang on slow/unreachable API calls.
-var openclawModelShowTimeout = 5 * time.Second
-
 // openclawFreshInstall is set to true when ensureOpenclawInstalled performs an install
 var openclawFreshInstall bool
 
@@ -34,7 +27,7 @@ type Openclaw struct{}
 
 func (c *Openclaw) String() string { return "OpenClaw" }
 
-func (c *Openclaw) Run(model string, args []string) error {
+func (c *Openclaw) Run(model string, _ []LaunchModel, args []string) error {
 	bin, err := ensureOpenclawInstalled()
 	if err != nil {
 		return err
@@ -656,7 +649,7 @@ func (c *Openclaw) Paths() []string {
 	return nil
 }
 
-func (c *Openclaw) Edit(models []string) error {
+func (c *Openclaw) Edit(models []LaunchModel) error {
 	if len(models) == 0 {
 		return nil
 	}
@@ -710,13 +703,11 @@ func (c *Openclaw) Edit(models []string) error {
 		}
 	}
 
-	client, _ := api.ClientFromEnvironment()
-
 	var newModels []any
 	for _, m := range models {
-		entry, _ := openclawModelConfig(context.Background(), client, m)
+		entry, _ := openclawModelConfig(m)
 		// Merge existing fields (user customizations)
-		if existing, ok := existingByID[m]; ok {
+		if existing, ok := existingByID[m.Name]; ok {
 			for k, v := range existing {
 				if _, isNew := entry[k]; !isNew {
 					entry[k] = v
@@ -744,7 +735,7 @@ func (c *Openclaw) Edit(models []string) error {
 	if modelConfig == nil {
 		modelConfig = make(map[string]any)
 	}
-	modelConfig["primary"] = "ollama/" + models[0]
+	modelConfig["primary"] = "ollama/" + models[0].Name
 	defaults["model"] = modelConfig
 	agents["defaults"] = defaults
 	config["agents"] = agents
@@ -759,7 +750,7 @@ func (c *Openclaw) Edit(models []string) error {
 
 	// Clear any per-session model overrides so the new primary takes effect
 	// immediately rather than being shadowed by a cached modelOverride.
-	clearSessionModelOverride(models[0])
+	clearSessionModelOverride(models[0].Name)
 	return nil
 }
 
@@ -936,10 +927,10 @@ func configureOllamaWebSearch() {
 
 // openclawModelConfig builds an OpenClaw model config entry with capability detection.
 // The second return value indicates whether the model is a cloud (remote) model.
-func openclawModelConfig(ctx context.Context, client *api.Client, modelID string) (map[string]any, bool) {
+func openclawModelConfig(model LaunchModel) (map[string]any, bool) {
 	entry := map[string]any{
-		"id":    modelID,
-		"name":  modelID,
+		"id":    model.Name,
+		"name":  model.Name,
 		"input": []any{"text"},
 		"cost": map[string]any{
 			"input":      0,
@@ -949,53 +940,24 @@ func openclawModelConfig(ctx context.Context, client *api.Client, modelID string
 		},
 	}
 
-	if client == nil {
-		return entry, false
-	}
-
-	showCtx := ctx
-	if _, hasDeadline := ctx.Deadline(); !hasDeadline {
-		var cancel context.CancelFunc
-		showCtx, cancel = context.WithTimeout(ctx, openclawModelShowTimeout)
-		defer cancel()
-	}
-
-	resp, err := client.Show(showCtx, &api.ShowRequest{Model: modelID})
-	if err != nil {
-		return entry, false
-	}
-
 	// Set input types based on vision capability
-	if slices.Contains(resp.Capabilities, model.CapabilityVision) {
+	if model.HasCapability("vision") {
 		entry["input"] = []any{"text", "image"}
 	}
 
 	// Set reasoning based on thinking capability
-	if slices.Contains(resp.Capabilities, model.CapabilityThinking) {
+	if model.HasCapability("thinking") {
 		entry["reasoning"] = true
 	}
 
-	// Cloud models: use hardcoded limits for context/output tokens.
-	// Capability detection above still applies (vision, thinking).
-	if resp.RemoteModel != "" {
-		if l, ok := lookupCloudModelLimit(modelID); ok {
-			entry["contextWindow"] = l.Context
-			entry["maxTokens"] = l.Output
-		}
-		return entry, true
+	if model.ContextLength > 0 {
+		entry["contextWindow"] = model.ContextLength
+	}
+	if model.MaxOutputTokens > 0 {
+		entry["maxTokens"] = model.MaxOutputTokens
 	}
 
-	// Extract context window from ModelInfo (local models only)
-	for key, val := range resp.ModelInfo {
-		if strings.HasSuffix(key, ".context_length") {
-			if ctxLen, ok := val.(float64); ok && ctxLen > 0 {
-				entry["contextWindow"] = int(ctxLen)
-			}
-			break
-		}
-	}
-
-	return entry, false
+	return entry, model.Remote || isCloudModelName(model.Name)
 }
 
 func (c *Openclaw) Models() []string {

cmd/launch/openclaw_test.go

@@ -2,12 +2,9 @@ package launch
 
 import (
 	"bytes"
-	"context"
 	"encoding/json"
 	"fmt"
 	"net"
-	"net/http"
-	"net/http/httptest"
 	"net/url"
 	"os"
 	"path/filepath"
@@ -16,8 +13,8 @@ import (
 	"testing"
 	"time"
 
-	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/cmd/internal/fileutil"
+	"github.com/ollama/ollama/types/model"
 )
 
 func TestOpenclawIntegration(t *testing.T) {
@@ -78,7 +75,7 @@ func TestOpenclawRunPassthroughArgs(t *testing.T) {
 	defer func() { DefaultConfirmPrompt = oldConfirmPrompt }()
 
 	c := &Openclaw{}
-	if err := c.Run("llama3.2", []string{"gateway", "--someflag"}); err != nil {
+	if err := c.Run("llama3.2", nil, []string{"gateway", "--someflag"}); err != nil {
 		t.Fatalf("Run() error = %v", err)
 	}
 
@@ -152,7 +149,7 @@ fi
 	defer func() { DefaultConfirmPrompt = oldConfirmPrompt }()
 
 	c := &Openclaw{}
-	if err := c.Run("llama3.2", nil); err != nil {
+	if err := c.Run("llama3.2", nil, nil); err != nil {
 		t.Fatalf("Run() error = %v", err)
 	}
 
@@ -224,7 +221,7 @@ func TestOpenclawRun_SetupLaterContinuesToGatewayAndTUI(t *testing.T) {
 	defer func() { DefaultConfirmPrompt = oldConfirmPrompt }()
 
 	c := &Openclaw{}
-	if err := c.Run("llama3.2", nil); err != nil {
+	if err := c.Run("llama3.2", nil, nil); err != nil {
 		t.Fatalf("Run() error = %v", err)
 	}
 
@@ -287,7 +284,7 @@ exit 0
 	defer func() { DefaultConfirmPrompt = oldConfirmPrompt }()
 
 	c := &Openclaw{}
-	if err := c.Run("llama3.2", []string{"status"}); err != nil {
+	if err := c.Run("llama3.2", nil, []string{"status"}); err != nil {
 		t.Fatalf("Run() error = %v", err)
 	}
 
@@ -367,7 +364,7 @@ exit 0
 	defer func() { DefaultConfirmPrompt = oldConfirmPrompt }()
 
 	c := &Openclaw{}
-	if err := c.Run("llama3.2", []string{"tui"}); err != nil {
+	if err := c.Run("llama3.2", nil, []string{"tui"}); err != nil {
 		t.Fatalf("Run() error = %v", err)
 	}
 
@@ -617,7 +614,7 @@ func TestOpenclawEdit(t *testing.T) {
 
 	t.Run("fresh install", func(t *testing.T) {
 		cleanup()
-		if err := c.Edit([]string{"llama3.2"}); err != nil {
+		if err := c.Edit(testLaunchModels("llama3.2")); err != nil {
 			t.Fatal(err)
 		}
 		assertOpenclawModelExists(t, configPath, "llama3.2")
@@ -626,7 +623,7 @@ func TestOpenclawEdit(t *testing.T) {
 
 	t.Run("multiple models - first is primary", func(t *testing.T) {
 		cleanup()
-		if err := c.Edit([]string{"llama3.2", "mistral"}); err != nil {
+		if err := c.Edit(testLaunchModels("llama3.2", "mistral")); err != nil {
 			t.Fatal(err)
 		}
 		assertOpenclawModelExists(t, configPath, "llama3.2")
@@ -638,7 +635,7 @@ func TestOpenclawEdit(t *testing.T) {
 		cleanup()
 		os.MkdirAll(configDir, 0o755)
 		os.WriteFile(configPath, []byte(`{"models":{"providers":{"anthropic":{"apiKey":"xxx"}}}}`), 0o644)
-		if err := c.Edit([]string{"llama3.2"}); err != nil {
+		if err := c.Edit(testLaunchModels("llama3.2")); err != nil {
 			t.Fatal(err)
 		}
 		data, _ := os.ReadFile(configPath)
@@ -655,7 +652,7 @@ func TestOpenclawEdit(t *testing.T) {
 		cleanup()
 		os.MkdirAll(configDir, 0o755)
 		os.WriteFile(configPath, []byte(`{"theme":"dark","mcp":{"servers":{}}}`), 0o644)
-		if err := c.Edit([]string{"llama3.2"}); err != nil {
+		if err := c.Edit(testLaunchModels("llama3.2")); err != nil {
 			t.Fatal(err)
 		}
 		data, _ := os.ReadFile(configPath)
@@ -671,7 +668,7 @@ func TestOpenclawEdit(t *testing.T) {
 
 	t.Run("preserve user customizations on models", func(t *testing.T) {
 		cleanup()
-		c.Edit([]string{"llama3.2"})
+		c.Edit(testLaunchModels("llama3.2"))
 
 		// User adds custom field
 		data, _ := os.ReadFile(configPath)
@@ -687,7 +684,7 @@ func TestOpenclawEdit(t *testing.T) {
 		os.WriteFile(configPath, configData, 0o644)
 
 		// Re-run Edit
-		c.Edit([]string{"llama3.2"})
+		c.Edit(testLaunchModels("llama3.2"))
 
 		data, _ = os.ReadFile(configPath)
 		json.Unmarshal(data, &cfg)
@@ -703,8 +700,8 @@ func TestOpenclawEdit(t *testing.T) {
 
 	t.Run("edit replaces models list", func(t *testing.T) {
 		cleanup()
-		c.Edit([]string{"llama3.2", "mistral"})
-		c.Edit([]string{"llama3.2"})
+		c.Edit(testLaunchModels("llama3.2", "mistral"))
+		c.Edit(testLaunchModels("llama3.2"))
 
 		assertOpenclawModelExists(t, configPath, "llama3.2")
 		assertOpenclawModelNotExists(t, configPath, "mistral")
@@ -716,7 +713,7 @@ func TestOpenclawEdit(t *testing.T) {
 		original := `{"existing":"data"}`
 		os.WriteFile(configPath, []byte(original), 0o644)
 
-		c.Edit([]string{})
+		c.Edit(testLaunchModels())
 
 		data, _ := os.ReadFile(configPath)
 		if string(data) != original {
@@ -729,7 +726,7 @@ func TestOpenclawEdit(t *testing.T) {
 		os.MkdirAll(configDir, 0o755)
 		os.WriteFile(configPath, []byte(`{corrupted`), 0o644)
 
-		if err := c.Edit([]string{"llama3.2"}); err != nil {
+		if err := c.Edit(testLaunchModels("llama3.2")); err != nil {
 			t.Fatal(err)
 		}
 
@@ -745,7 +742,7 @@ func TestOpenclawEdit(t *testing.T) {
 		os.MkdirAll(configDir, 0o755)
 		os.WriteFile(configPath, []byte(`{"models":"not a map"}`), 0o644)
 
-		if err := c.Edit([]string{"llama3.2"}); err != nil {
+		if err := c.Edit(testLaunchModels("llama3.2")); err != nil {
 			t.Fatal(err)
 		}
 		assertOpenclawModelExists(t, configPath, "llama3.2")
@@ -925,7 +922,7 @@ func TestOpenclawEditSchemaFields(t *testing.T) {
 	setTestHome(t, tmpDir)
 	configPath := filepath.Join(tmpDir, ".openclaw", "openclaw.json")
 
-	if err := c.Edit([]string{"llama3.2"}); err != nil {
+	if err := c.Edit(testLaunchModels("llama3.2")); err != nil {
 		t.Fatal(err)
 	}
 
@@ -966,7 +963,7 @@ func TestOpenclawEditModelNames(t *testing.T) {
 
 	t.Run("model with colon tag", func(t *testing.T) {
 		cleanup()
-		if err := c.Edit([]string{"llama3.2:70b"}); err != nil {
+		if err := c.Edit(testLaunchModels("llama3.2:70b")); err != nil {
 			t.Fatal(err)
 		}
 		assertOpenclawModelExists(t, configPath, "llama3.2:70b")
@@ -975,7 +972,7 @@ func TestOpenclawEditModelNames(t *testing.T) {
 
 	t.Run("model with slash", func(t *testing.T) {
 		cleanup()
-		if err := c.Edit([]string{"library/model:tag"}); err != nil {
+		if err := c.Edit(testLaunchModels("library/model:tag")); err != nil {
 			t.Fatal(err)
 		}
 		assertOpenclawModelExists(t, configPath, "library/model:tag")
@@ -984,7 +981,7 @@ func TestOpenclawEditModelNames(t *testing.T) {
 
 	t.Run("model with hyphen", func(t *testing.T) {
 		cleanup()
-		if err := c.Edit([]string{"test-model"}); err != nil {
+		if err := c.Edit(testLaunchModels("test-model")); err != nil {
 			t.Fatal(err)
 		}
 		assertOpenclawModelExists(t, configPath, "test-model")
@@ -1004,7 +1001,7 @@ func TestOpenclawEditAgentsPreservation(t *testing.T) {
 		os.MkdirAll(configDir, 0o755)
 		os.WriteFile(configPath, []byte(`{"agents":{"defaults":{"model":{"primary":"old"},"temperature":0.7}}}`), 0o644)
 
-		c.Edit([]string{"llama3.2"})
+		c.Edit(testLaunchModels("llama3.2"))
 
 		data, _ := os.ReadFile(configPath)
 		var cfg map[string]any
@@ -1021,7 +1018,7 @@ func TestOpenclawEditAgentsPreservation(t *testing.T) {
 		os.MkdirAll(configDir, 0o755)
 		os.WriteFile(configPath, []byte(`{"agents":{"defaults":{},"custom-agent":{"foo":"bar"}}}`), 0o644)
 
-		c.Edit([]string{"llama3.2"})
+		c.Edit(testLaunchModels("llama3.2"))
 
 		data, _ := os.ReadFile(configPath)
 		var cfg map[string]any
@@ -1061,7 +1058,7 @@ func TestOpenclawEdit_RoundTrip(t *testing.T) {
 	os.MkdirAll(configDir, 0o755)
 	os.WriteFile(configPath, []byte(testOpenclawFixture), 0o644)
 
-	if err := c.Edit([]string{"llama3.2", "mistral"}); err != nil {
+	if err := c.Edit(testLaunchModels("llama3.2", "mistral")); err != nil {
 		t.Fatal(err)
 	}
 
@@ -1107,10 +1104,10 @@ func TestOpenclawEdit_Idempotent(t *testing.T) {
 	os.MkdirAll(configDir, 0o755)
 	os.WriteFile(configPath, []byte(testOpenclawFixture), 0o644)
 
-	c.Edit([]string{"llama3.2", "mistral"})
+	c.Edit(testLaunchModels("llama3.2", "mistral"))
 	firstData, _ := os.ReadFile(configPath)
 
-	c.Edit([]string{"llama3.2", "mistral"})
+	c.Edit(testLaunchModels("llama3.2", "mistral"))
 	secondData, _ := os.ReadFile(configPath)
 
 	if string(firstData) != string(secondData) {
@@ -1133,7 +1130,7 @@ func TestOpenclawEdit_MultipleConsecutiveEdits(t *testing.T) {
 		if i%2 == 0 {
 			models = []string{"model-x", "model-y", "model-z"}
 		}
-		if err := c.Edit(models); err != nil {
+		if err := c.Edit(launchModelsFromNames(models)); err != nil {
 			t.Fatalf("edit %d failed: %v", i, err)
 		}
 	}
@@ -1162,7 +1159,7 @@ func TestOpenclawEdit_BackupCreated(t *testing.T) {
 	original := fmt.Sprintf(`{"theme": "%s"}`, uniqueMarker)
 	os.WriteFile(configPath, []byte(original), 0o644)
 
-	if err := c.Edit([]string{"model-a"}); err != nil {
+	if err := c.Edit(testLaunchModels("model-a")); err != nil {
 		t.Fatal(err)
 	}
 
@@ -1284,7 +1281,7 @@ func TestOpenclawLegacyPaths(t *testing.T) {
 		os.WriteFile(filepath.Join(newDir, "openclaw.json"), []byte(`{"theme":"new"}`), 0o644)
 		os.WriteFile(filepath.Join(legacyDir, "clawdbot.json"), []byte(`{"theme":"legacy"}`), 0o644)
 
-		if err := c.Edit([]string{"llama3.2"}); err != nil {
+		if err := c.Edit(testLaunchModels("llama3.2")); err != nil {
 			t.Fatal(err)
 		}
 
@@ -1303,7 +1300,7 @@ func TestOpenclawLegacyPaths(t *testing.T) {
 		os.MkdirAll(legacyDir, 0o755)
 		os.WriteFile(filepath.Join(legacyDir, "clawdbot.json"), []byte(`{"theme":"dark"}`), 0o644)
 
-		if err := c.Edit([]string{"llama3.2"}); err != nil {
+		if err := c.Edit(testLaunchModels("llama3.2")); err != nil {
 			t.Fatal(err)
 		}
 
@@ -1331,7 +1328,7 @@ func TestOpenclawEdit_CreatesDirectoryIfMissing(t *testing.T) {
 		t.Fatal("directory should not exist before test")
 	}
 
-	if err := c.Edit([]string{"model-a"}); err != nil {
+	if err := c.Edit(testLaunchModels("model-a")); err != nil {
 		t.Fatal(err)
 	}
 
@@ -2248,8 +2245,8 @@ func TestPrintOpenclawReady(t *testing.T) {
 }
 
 func TestOpenclawModelConfig(t *testing.T) {
-	t.Run("nil client returns base config", func(t *testing.T) {
-		cfg, _ := openclawModelConfig(context.Background(), nil, "llama3.2")
+	t.Run("minimal model returns base config", func(t *testing.T) {
+		cfg, _ := openclawModelConfig(fallbackLaunchModel("llama3.2"))
 
 		if cfg["id"] != "llama3.2" {
 			t.Errorf("id = %v, want llama3.2", cfg["id"])
@@ -2260,29 +2257,17 @@ func TestOpenclawModelConfig(t *testing.T) {
 		if cfg["cost"] == nil {
 			t.Error("cost should be set")
 		}
-		// Should not have capability fields without API
+		// Should not have capability fields without inventory metadata.
 		if _, ok := cfg["reasoning"]; ok {
-			t.Error("reasoning should not be set without API")
+			t.Error("reasoning should not be set without metadata")
 		}
 		if _, ok := cfg["contextWindow"]; ok {
-			t.Error("contextWindow should not be set without API")
+			t.Error("contextWindow should not be set without metadata")
 		}
 	})
 
 	t.Run("sets vision input when model has vision capability", func(t *testing.T) {
-		srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-			if r.URL.Path == "/api/show" {
-				fmt.Fprintf(w, `{"capabilities":["vision"],"model_info":{"llama.context_length":4096}}`)
-				return
-			}
-			w.WriteHeader(http.StatusNotFound)
-		}))
-		defer srv.Close()
-
-		u, _ := url.Parse(srv.URL)
-		client := api.NewClient(u, srv.Client())
-
-		cfg, _ := openclawModelConfig(context.Background(), client, "llava:7b")
+		cfg, _ := openclawModelConfig(LaunchModel{Name: "llava:7b", Capabilities: []model.Capability{"vision"}, ContextLength: 4096})
 
 		input, ok := cfg["input"].([]any)
 		if !ok || len(input) != 2 {
@@ -2291,19 +2276,7 @@ func TestOpenclawModelConfig(t *testing.T) {
 	})
 
 	t.Run("sets text-only input when model lacks vision", func(t *testing.T) {
-		srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-			if r.URL.Path == "/api/show" {
-				fmt.Fprintf(w, `{"capabilities":["completion"],"model_info":{}}`)
-				return
-			}
-			w.WriteHeader(http.StatusNotFound)
-		}))
-		defer srv.Close()
-
-		u, _ := url.Parse(srv.URL)
-		client := api.NewClient(u, srv.Client())
-
-		cfg, _ := openclawModelConfig(context.Background(), client, "llama3.2")
+		cfg, _ := openclawModelConfig(LaunchModel{Name: "llama3.2", Capabilities: []model.Capability{"completion"}})
 
 		input, ok := cfg["input"].([]any)
 		if !ok || len(input) != 1 {
@@ -2315,39 +2288,15 @@ func TestOpenclawModelConfig(t *testing.T) {
 	})
 
 	t.Run("sets reasoning when model has thinking capability", func(t *testing.T) {
-		srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-			if r.URL.Path == "/api/show" {
-				fmt.Fprintf(w, `{"capabilities":["thinking"],"model_info":{}}`)
-				return
-			}
-			w.WriteHeader(http.StatusNotFound)
-		}))
-		defer srv.Close()
-
-		u, _ := url.Parse(srv.URL)
-		client := api.NewClient(u, srv.Client())
-
-		cfg, _ := openclawModelConfig(context.Background(), client, "qwq")
+		cfg, _ := openclawModelConfig(LaunchModel{Name: "qwq", Capabilities: []model.Capability{"thinking"}})
 
 		if cfg["reasoning"] != true {
 			t.Error("expected reasoning = true for thinking model")
 		}
 	})
 
-	t.Run("extracts context window from model info", func(t *testing.T) {
-		srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-			if r.URL.Path == "/api/show" {
-				fmt.Fprintf(w, `{"capabilities":[],"model_info":{"llama.context_length":131072}}`)
-				return
-			}
-			w.WriteHeader(http.StatusNotFound)
-		}))
-		defer srv.Close()
-
-		u, _ := url.Parse(srv.URL)
-		client := api.NewClient(u, srv.Client())
-
-		cfg, _ := openclawModelConfig(context.Background(), client, "llama3.2")
+	t.Run("sets context window from inventory metadata", func(t *testing.T) {
+		cfg, _ := openclawModelConfig(LaunchModel{Name: "llama3.2", ContextLength: 131072})
 
 		if cfg["contextWindow"] != 131072 {
 			t.Errorf("contextWindow = %v, want 131072", cfg["contextWindow"])
@@ -2355,19 +2304,11 @@ func TestOpenclawModelConfig(t *testing.T) {
 	})
 
 	t.Run("handles all capabilities together", func(t *testing.T) {
-		srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-			if r.URL.Path == "/api/show" {
-				fmt.Fprintf(w, `{"capabilities":["vision","thinking"],"model_info":{"qwen3.context_length":32768}}`)
-				return
-			}
-			w.WriteHeader(http.StatusNotFound)
-		}))
-		defer srv.Close()
-
-		u, _ := url.Parse(srv.URL)
-		client := api.NewClient(u, srv.Client())
-
-		cfg, _ := openclawModelConfig(context.Background(), client, "qwen3-vision")
+		cfg, _ := openclawModelConfig(LaunchModel{
+			Name:          "qwen3-vision",
+			Capabilities:  []model.Capability{"vision", "thinking"},
+			ContextLength: 32768,
+		})
 
 		input, ok := cfg["input"].([]any)
 		if !ok || len(input) != 2 {
@@ -2381,17 +2322,8 @@ func TestOpenclawModelConfig(t *testing.T) {
 		}
 	})
 
-	t.Run("returns base config when show fails", func(t *testing.T) {
-		srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-			w.WriteHeader(http.StatusNotFound)
-			fmt.Fprintf(w, `{"error":"model not found"}`)
-		}))
-		defer srv.Close()
-
-		u, _ := url.Parse(srv.URL)
-		client := api.NewClient(u, srv.Client())
-
-		cfg, _ := openclawModelConfig(context.Background(), client, "missing-model")
+	t.Run("returns base config when metadata is unavailable", func(t *testing.T) {
+		cfg, _ := openclawModelConfig(fallbackLaunchModel("missing-model"))
 
 		if cfg["id"] != "missing-model" {
 			t.Errorf("id = %v, want missing-model", cfg["id"])
@@ -2401,62 +2333,15 @@ func TestOpenclawModelConfig(t *testing.T) {
 			t.Error("input should always be set")
 		}
 		if _, ok := cfg["reasoning"]; ok {
-			t.Error("reasoning should not be set when show fails")
+			t.Error("reasoning should not be set when metadata is unavailable")
 		}
 		if _, ok := cfg["contextWindow"]; ok {
-			t.Error("contextWindow should not be set when show fails")
-		}
-	})
-
-	t.Run("times out slow show and returns base config", func(t *testing.T) {
-		oldTimeout := openclawModelShowTimeout
-		openclawModelShowTimeout = 50 * time.Millisecond
-		t.Cleanup(func() { openclawModelShowTimeout = oldTimeout })
-
-		srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-			if r.URL.Path == "/api/show" {
-				time.Sleep(300 * time.Millisecond)
-				fmt.Fprintf(w, `{"capabilities":["thinking"],"model_info":{"llama.context_length":4096}}`)
-				return
-			}
-			w.WriteHeader(http.StatusNotFound)
-		}))
-		defer srv.Close()
-
-		u, _ := url.Parse(srv.URL)
-		client := api.NewClient(u, srv.Client())
-
-		start := time.Now()
-		cfg, _ := openclawModelConfig(context.Background(), client, "slow-model")
-		elapsed := time.Since(start)
-		if elapsed >= 250*time.Millisecond {
-			t.Fatalf("openclawModelConfig took too long: %v", elapsed)
-		}
-		if cfg["id"] != "slow-model" {
-			t.Errorf("id = %v, want slow-model", cfg["id"])
-		}
-		if _, ok := cfg["reasoning"]; ok {
-			t.Error("reasoning should not be set on timeout")
-		}
-		if _, ok := cfg["contextWindow"]; ok {
-			t.Error("contextWindow should not be set on timeout")
+			t.Error("contextWindow should not be set when metadata is unavailable")
 		}
 	})
 
 	t.Run("skips zero context length", func(t *testing.T) {
-		srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-			if r.URL.Path == "/api/show" {
-				fmt.Fprintf(w, `{"capabilities":[],"model_info":{"llama.context_length":0}}`)
-				return
-			}
-			w.WriteHeader(http.StatusNotFound)
-		}))
-		defer srv.Close()
-
-		u, _ := url.Parse(srv.URL)
-		client := api.NewClient(u, srv.Client())
-
-		cfg, _ := openclawModelConfig(context.Background(), client, "test-model")
+		cfg, _ := openclawModelConfig(LaunchModel{Name: "test-model", ContextLength: 0})
 
 		if _, ok := cfg["contextWindow"]; ok {
 			t.Error("contextWindow should not be set for zero value")
@@ -2464,21 +2349,7 @@ func TestOpenclawModelConfig(t *testing.T) {
 	})
 
 	t.Run("cloud model uses hardcoded limits", func(t *testing.T) {
-		// Use a model name that's in cloudModelLimits and make the server
-		// report it as a remote/cloud model
-		srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-			if r.URL.Path == "/api/show" {
-				fmt.Fprintf(w, `{"capabilities":[],"model_info":{},"remote_model":"minimax-m2.7"}`)
-				return
-			}
-			w.WriteHeader(http.StatusNotFound)
-		}))
-		defer srv.Close()
-
-		u, _ := url.Parse(srv.URL)
-		client := api.NewClient(u, srv.Client())
-
-		cfg, isCloud := openclawModelConfig(context.Background(), client, "minimax-m2.7:cloud")
+		cfg, isCloud := openclawModelConfig(fallbackLaunchModel("minimax-m2.7:cloud"))
 
 		if !isCloud {
 			t.Error("expected isCloud = true for cloud model")
@@ -2492,21 +2363,11 @@ func TestOpenclawModelConfig(t *testing.T) {
 	})
 
 	t.Run("cloud model with vision capability gets image input", func(t *testing.T) {
-		// Regression test: cloud models must not skip capability detection.
-		// A cloud model that reports vision capability should have input: [text, image].
-		srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-			if r.URL.Path == "/api/show" {
-				fmt.Fprintf(w, `{"capabilities":["vision"],"model_info":{},"remote_model":"qwen3-vl"}`)
-				return
-			}
-			w.WriteHeader(http.StatusNotFound)
-		}))
-		defer srv.Close()
-
-		u, _ := url.Parse(srv.URL)
-		client := api.NewClient(u, srv.Client())
-
-		cfg, isCloud := openclawModelConfig(context.Background(), client, "qwen3-vl:235b-cloud")
+		cfg, isCloud := openclawModelConfig(LaunchModel{
+			Name:         "qwen3-vl:235b-cloud",
+			Remote:       true,
+			Capabilities: []model.Capability{"vision"},
+		}.WithCloudLimits())
 
 		if !isCloud {
 			t.Error("expected isCloud = true for cloud vision model")
@@ -2518,21 +2379,11 @@ func TestOpenclawModelConfig(t *testing.T) {
 	})
 
 	t.Run("cloud model with thinking capability gets reasoning flag", func(t *testing.T) {
-		// Regression test: cloud models must not skip capability detection.
-		// A cloud model that reports thinking capability should have reasoning: true.
-		srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-			if r.URL.Path == "/api/show" {
-				fmt.Fprintf(w, `{"capabilities":["thinking"],"model_info":{},"remote_model":"qwq-cloud"}`)
-				return
-			}
-			w.WriteHeader(http.StatusNotFound)
-		}))
-		defer srv.Close()
-
-		u, _ := url.Parse(srv.URL)
-		client := api.NewClient(u, srv.Client())
-
-		cfg, isCloud := openclawModelConfig(context.Background(), client, "qwq:cloud")
+		cfg, isCloud := openclawModelConfig(LaunchModel{
+			Name:         "qwq:cloud",
+			Remote:       true,
+			Capabilities: []model.Capability{"thinking"},
+		})
 
 		if !isCloud {
 			t.Error("expected isCloud = true for cloud thinking model")

cmd/launch/opencode.go

@@ -43,7 +43,7 @@ func findOpenCode() (string, bool) {
 	return "", false
 }
 
-func (o *OpenCode) Run(model string, args []string) error {
+func (o *OpenCode) Run(model string, models []LaunchModel, args []string) error {
 	opencodePath, ok := findOpenCode()
 	if !ok {
 		return fmt.Errorf("opencode is not installed, install from https://opencode.ai")
@@ -54,7 +54,7 @@ func (o *OpenCode) Run(model string, args []string) error {
 	cmd.Stdout = os.Stdout
 	cmd.Stderr = os.Stderr
 	cmd.Env = os.Environ()
-	if content := o.resolveContent(model); content != "" {
+	if content := o.resolveContent(model, models); content != "" {
 		cmd.Env = append(cmd.Env, "OPENCODE_CONFIG_CONTENT="+content)
 	}
 	return cmd.Run()
@@ -63,21 +63,57 @@ func (o *OpenCode) Run(model string, args []string) error {
 // resolveContent returns the inline config to send via OPENCODE_CONFIG_CONTENT.
 // Returns content built by Edit if available, otherwise builds from model.json
 // with the requested model as primary (e.g. re-launch with saved config).
-func (o *OpenCode) resolveContent(model string) string {
+func (o *OpenCode) resolveContent(model string, models []LaunchModel) string {
 	if o.configContent != "" {
 		return o.configContent
 	}
-	models := readModelJSONModels()
-	if !slices.Contains(models, model) {
-		models = append([]string{model}, models...)
+	resolvedModels := resolveOpenCodeRunModels(model, models, readModelJSONModels())
+	if len(resolvedModels) == 0 {
+		return ""
 	}
-	content, err := buildInlineConfig(model, models)
+	content, err := buildInlineConfig(resolvedModels[0], resolvedModels)
 	if err != nil {
 		return ""
 	}
 	return content
 }
 
+func resolveOpenCodeRunModels(primary string, models []LaunchModel, stateModels []string) []LaunchModel {
+	if primary == "" {
+		return nil
+	}
+
+	resolved := make([]LaunchModel, 0, 1+len(models)+len(stateModels))
+	appendModel := func(name string) {
+		if name == "" || hasLaunchModel(resolved, name) {
+			return
+		}
+		if model, ok := findLaunchModel(models, name); ok {
+			resolved = append(resolved, model)
+			return
+		}
+		resolved = append(resolved, fallbackLaunchModel(name))
+	}
+
+	appendModel(primary)
+	for _, model := range models {
+		appendModel(model.Name)
+	}
+	for _, model := range stateModels {
+		appendModel(model)
+	}
+	return resolved
+}
+
+func hasLaunchModel(models []LaunchModel, name string) bool {
+	for _, model := range models {
+		if launchModelMatches(model.Name, name) || launchModelMatches(name, model.Name) {
+			return true
+		}
+	}
+	return false
+}
+
 func (o *OpenCode) Paths() []string {
 	sp, err := openCodeStatePath()
 	if err != nil {
@@ -100,12 +136,13 @@ func openCodeStatePath() (string, error) {
 	return filepath.Join(home, ".local", "state", "opencode", "model.json"), nil
 }
 
-func (o *OpenCode) Edit(modelList []string) error {
+func (o *OpenCode) Edit(models []LaunchModel) error {
+	modelList := launchModelNames(models)
 	if len(modelList) == 0 {
 		return nil
 	}
 
-	content, err := buildInlineConfig(modelList[0], modelList)
+	content, err := buildInlineConfig(models[0], models)
 	if err != nil {
 		return err
 	}
@@ -172,10 +209,11 @@ func (o *OpenCode) Models() []string {
 
 // buildInlineConfig produces the JSON string for OPENCODE_CONFIG_CONTENT.
 // primary is the model to launch with, models is the full list of available models.
-func buildInlineConfig(primary string, models []string) (string, error) {
-	if primary == "" || len(models) == 0 {
+func buildInlineConfig(primary LaunchModel, models []LaunchModel) (string, error) {
+	if primary.Name == "" || len(models) == 0 {
 		return "", fmt.Errorf("buildInlineConfig: primary and models are required")
 	}
+
 	config := map[string]any{
 		"$schema": "https://opencode.ai/config.json",
 		"provider": map[string]any{
@@ -188,7 +226,7 @@ func buildInlineConfig(primary string, models []string) (string, error) {
 				"models": buildModelEntries(models),
 			},
 		},
-		"model": "ollama/" + primary,
+		"model": "ollama/" + primary.Name,
 	}
 	data, err := json.Marshal(config)
 	if err != nil {
@@ -228,21 +266,29 @@ func readModelJSONModels() []string {
 	return models
 }
 
-func buildModelEntries(modelList []string) map[string]any {
+func buildModelEntries(modelList []LaunchModel) map[string]any {
 	models := make(map[string]any)
 	for _, model := range modelList {
 		entry := map[string]any{
-			"name": model,
+			"name": model.Name,
 		}
-		if isCloudModelName(model) {
-			if l, ok := lookupCloudModelLimit(model); ok {
-				entry["limit"] = map[string]any{
-					"context": l.Context,
-					"output":  l.Output,
-				}
+		if model.HasCapability("vision") {
+			entry["modalities"] = map[string]any{
+				"input":  []string{"text", "image"},
+				"output": []string{"text"},
 			}
 		}
-		models[model] = entry
+		if model.ContextLength > 0 || model.MaxOutputTokens > 0 {
+			limit := make(map[string]any)
+			if model.ContextLength > 0 {
+				limit["context"] = model.ContextLength
+			}
+			if model.MaxOutputTokens > 0 {
+				limit["output"] = model.MaxOutputTokens
+			}
+			entry["limit"] = limit
+		}
+		models[model.Name] = entry
 	}
 	return models
 }

cmd/launch/opencode_test.go

@@ -7,6 +7,8 @@ import (
 	"path/filepath"
 	"runtime"
 	"testing"
+
+	"github.com/ollama/ollama/types/model"
 )
 
 func TestOpenCodeIntegration(t *testing.T) {
@@ -31,7 +33,7 @@ func TestOpenCodeEdit(t *testing.T) {
 	t.Run("builds config content with provider", func(t *testing.T) {
 		setTestHome(t, t.TempDir())
 		o := &OpenCode{}
-		if err := o.Edit([]string{"llama3.2"}); err != nil {
+		if err := o.Edit(testLaunchModels("llama3.2")); err != nil {
 			t.Fatal(err)
 		}
 
@@ -65,7 +67,7 @@ func TestOpenCodeEdit(t *testing.T) {
 	t.Run("multiple models", func(t *testing.T) {
 		setTestHome(t, t.TempDir())
 		o := &OpenCode{}
-		if err := o.Edit([]string{"llama3.2", "qwen3:32b"}); err != nil {
+		if err := o.Edit(testLaunchModels("llama3.2", "qwen3:32b")); err != nil {
 			t.Fatal(err)
 		}
 
@@ -90,7 +92,7 @@ func TestOpenCodeEdit(t *testing.T) {
 	t.Run("empty models is no-op", func(t *testing.T) {
 		setTestHome(t, t.TempDir())
 		o := &OpenCode{}
-		if err := o.Edit([]string{}); err != nil {
+		if err := o.Edit(testLaunchModels()); err != nil {
 			t.Fatal(err)
 		}
 		if o.configContent != "" {
@@ -102,7 +104,7 @@ func TestOpenCodeEdit(t *testing.T) {
 		tmpDir := t.TempDir()
 		setTestHome(t, tmpDir)
 		o := &OpenCode{}
-		o.Edit([]string{"llama3.2"})
+		o.Edit(testLaunchModels("llama3.2"))
 
 		configDir := filepath.Join(tmpDir, ".config", "opencode")
 
@@ -117,7 +119,7 @@ func TestOpenCodeEdit(t *testing.T) {
 	t.Run("cloud model has limits", func(t *testing.T) {
 		setTestHome(t, t.TempDir())
 		o := &OpenCode{}
-		if err := o.Edit([]string{"glm-4.7:cloud"}); err != nil {
+		if err := o.Edit(testLaunchModels("glm-4.7:cloud")); err != nil {
 			t.Fatal(err)
 		}
 
@@ -144,7 +146,7 @@ func TestOpenCodeEdit(t *testing.T) {
 	t.Run("local model has no limits", func(t *testing.T) {
 		setTestHome(t, t.TempDir())
 		o := &OpenCode{}
-		o.Edit([]string{"llama3.2"})
+		o.Edit(testLaunchModels("llama3.2"))
 
 		var cfg map[string]any
 		json.Unmarshal([]byte(o.configContent), &cfg)
@@ -157,6 +159,43 @@ func TestOpenCodeEdit(t *testing.T) {
 			t.Errorf("local model should not have limit, got %v", entry["limit"])
 		}
 	})
+
+	t.Run("vision model gets image input modalities", func(t *testing.T) {
+		models := buildModelEntries([]LaunchModel{{Name: "gemma4:26b", Capabilities: []model.Capability{"vision"}}})
+		entry, _ := models["gemma4:26b"].(map[string]any)
+		modalities, _ := entry["modalities"].(map[string]any)
+		input, _ := modalities["input"].([]string)
+		output, _ := modalities["output"].([]string)
+
+		if len(input) != 2 || input[0] != "text" || input[1] != "image" {
+			t.Fatalf("modalities.input = %v, want [text image]", input)
+		}
+		if len(output) != 1 || output[0] != "text" {
+			t.Fatalf("modalities.output = %v, want [text]", output)
+		}
+	})
+}
+
+func TestBuildModelEntries(t *testing.T) {
+	t.Run("defaults to model name without capabilities", func(t *testing.T) {
+		models := buildModelEntries(testLaunchModels("llama3.2"))
+		entry, _ := models["llama3.2"].(map[string]any)
+		if entry["name"] != "llama3.2" {
+			t.Fatalf("name = %v, want llama3.2", entry["name"])
+		}
+		if _, ok := entry["modalities"]; ok {
+			t.Fatalf("modalities should not be set without capabilities, got %v", entry["modalities"])
+		}
+	})
+
+	t.Run("uses context and output limits from metadata", func(t *testing.T) {
+		models := buildModelEntries([]LaunchModel{{Name: "glm-5:cloud", ContextLength: 202_752, MaxOutputTokens: 131_072}})
+		entry, _ := models["glm-5:cloud"].(map[string]any)
+		limit, _ := entry["limit"].(map[string]any)
+		if limit["context"] != 202_752 || limit["output"] != 131_072 {
+			t.Fatalf("limit = %v, want context/output", limit)
+		}
+	})
 }
 
 func TestOpenCodeModels_ReturnsNil(t *testing.T) {
@@ -284,7 +323,7 @@ func TestOpenCodeEdit_CloudModelLimitStructure(t *testing.T) {
 
 	expected := cloudModelLimits["glm-4.7"]
 
-	if err := o.Edit([]string{"glm-4.7:cloud"}); err != nil {
+	if err := o.Edit(testLaunchModels("glm-4.7:cloud")); err != nil {
 		t.Fatal(err)
 	}
 
@@ -314,7 +353,7 @@ func TestOpenCodeEdit_SpecialCharsInModelName(t *testing.T) {
 
 	specialModel := `model-with-"quotes"`
 
-	err := o.Edit([]string{specialModel})
+	err := o.Edit(testLaunchModels(specialModel))
 	if err != nil {
 		t.Fatalf("Edit with special chars failed: %v", err)
 	}
@@ -407,7 +446,7 @@ func TestOpenCodeResolveContent(t *testing.T) {
 		setTestHome(t, tmpDir)
 
 		o := &OpenCode{}
-		if err := o.Edit([]string{"gemma4"}); err != nil {
+		if err := o.Edit(testLaunchModels("gemma4")); err != nil {
 			t.Fatal(err)
 		}
 		editContent := o.configContent
@@ -422,7 +461,7 @@ func TestOpenCodeResolveContent(t *testing.T) {
 		data, _ := json.MarshalIndent(state, "", "  ")
 		os.WriteFile(filepath.Join(stateDir, "model.json"), data, 0o644)
 
-		got := o.resolveContent("gemma4")
+		got := o.resolveContent("gemma4", nil)
 		if got != editContent {
 			t.Errorf("resolveContent returned different content than Edit set\ngot:  %s\nwant: %s", got, editContent)
 		}
@@ -444,7 +483,7 @@ func TestOpenCodeResolveContent(t *testing.T) {
 		os.WriteFile(filepath.Join(stateDir, "model.json"), data, 0o644)
 
 		o := &OpenCode{}
-		content := o.resolveContent("llama3.2")
+		content := o.resolveContent("llama3.2", nil)
 		if content == "" {
 			t.Fatal("resolveContent returned empty")
 		}
@@ -478,7 +517,7 @@ func TestOpenCodeResolveContent(t *testing.T) {
 		os.WriteFile(filepath.Join(stateDir, "model.json"), data, 0o644)
 
 		o := &OpenCode{}
-		content := o.resolveContent("qwen3:32b")
+		content := o.resolveContent("qwen3:32b", nil)
 
 		var cfg map[string]any
 		json.Unmarshal([]byte(content), &cfg)
@@ -502,7 +541,7 @@ func TestOpenCodeResolveContent(t *testing.T) {
 		os.WriteFile(filepath.Join(stateDir, "model.json"), data, 0o644)
 
 		o := &OpenCode{}
-		content := o.resolveContent("gemma4")
+		content := o.resolveContent("gemma4", nil)
 
 		var cfg map[string]any
 		json.Unmarshal([]byte(content), &cfg)
@@ -522,11 +561,56 @@ func TestOpenCodeResolveContent(t *testing.T) {
 		setTestHome(t, tmpDir)
 
 		o := &OpenCode{}
-		if got := o.resolveContent(""); got != "" {
+		if got := o.resolveContent("", nil); got != "" {
 			t.Errorf("resolveContent(\"\") = %q, want empty", got)
 		}
 	})
 
+	t.Run("uses run model metadata when Edit was not called", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+
+		stateDir := filepath.Join(tmpDir, ".local", "state", "opencode")
+		os.MkdirAll(stateDir, 0o755)
+		state := map[string]any{
+			"recent": []any{
+				map[string]any{"providerID": "ollama", "modelID": "llama3.2"},
+			},
+		}
+		data, _ := json.MarshalIndent(state, "", "  ")
+		os.WriteFile(filepath.Join(stateDir, "model.json"), data, 0o644)
+
+		o := &OpenCode{}
+		content := o.resolveContent("gemma4", []LaunchModel{
+			{
+				Name:            "gemma4",
+				Capabilities:    []model.Capability{model.CapabilityVision},
+				ContextLength:   65_536,
+				MaxOutputTokens: 8_192,
+			},
+		})
+		if content == "" {
+			t.Fatal("resolveContent returned empty")
+		}
+
+		var cfg map[string]any
+		json.Unmarshal([]byte(content), &cfg)
+		provider, _ := cfg["provider"].(map[string]any)
+		ollama, _ := provider["ollama"].(map[string]any)
+		cfgModels, _ := ollama["models"].(map[string]any)
+		entry, _ := cfgModels["gemma4"].(map[string]any)
+		limit, _ := entry["limit"].(map[string]any)
+		if limit["context"] != float64(65_536) || limit["output"] != float64(8_192) {
+			t.Fatalf("limit = %v, want context/output from launch metadata", limit)
+		}
+		if _, ok := entry["modalities"].(map[string]any); !ok {
+			t.Fatalf("modalities should be set from launch metadata, got %v", entry["modalities"])
+		}
+		if cfgModels["llama3.2"] == nil {
+			t.Fatalf("state model missing from fallback config: %v", cfgModels)
+		}
+	})
+
 	t.Run("does not mutate configContent on fallback", func(t *testing.T) {
 		tmpDir := t.TempDir()
 		setTestHome(t, tmpDir)
@@ -542,7 +626,7 @@ func TestOpenCodeResolveContent(t *testing.T) {
 		os.WriteFile(filepath.Join(stateDir, "model.json"), data, 0o644)
 
 		o := &OpenCode{}
-		_ = o.resolveContent("llama3.2")
+		_ = o.resolveContent("llama3.2", nil)
 		if o.configContent != "" {
 			t.Errorf("resolveContent should not mutate configContent, got %q", o.configContent)
 		}
@@ -551,19 +635,19 @@ func TestOpenCodeResolveContent(t *testing.T) {
 
 func TestBuildInlineConfig(t *testing.T) {
 	t.Run("returns error for empty primary", func(t *testing.T) {
-		if _, err := buildInlineConfig("", []string{"llama3.2"}); err == nil {
+		if _, err := buildInlineConfig(LaunchModel{}, testLaunchModels("llama3.2")); err == nil {
 			t.Error("expected error for empty primary")
 		}
 	})
 
 	t.Run("returns error for empty models", func(t *testing.T) {
-		if _, err := buildInlineConfig("llama3.2", nil); err == nil {
+		if _, err := buildInlineConfig(fallbackLaunchModel("llama3.2"), nil); err == nil {
 			t.Error("expected error for empty models")
 		}
 	})
 
 	t.Run("primary differs from first model in list", func(t *testing.T) {
-		content, err := buildInlineConfig("qwen3:32b", []string{"llama3.2", "qwen3:32b"})
+		content, err := buildInlineConfig(fallbackLaunchModel("qwen3:32b"), testLaunchModels("llama3.2", "qwen3:32b"))
 		if err != nil {
 			t.Fatal(err)
 		}
@@ -592,7 +676,7 @@ func TestOpenCodeEdit_PreservesRecentEntries(t *testing.T) {
 		os.WriteFile(filepath.Join(stateDir, "model.json"), data, 0o644)
 
 		o := &OpenCode{}
-		if err := o.Edit([]string{"new-X"}); err != nil {
+		if err := o.Edit(testLaunchModels("new-X")); err != nil {
 			t.Fatal(err)
 		}
 
@@ -626,7 +710,7 @@ func TestOpenCodeEdit_PreservesRecentEntries(t *testing.T) {
 		os.WriteFile(filepath.Join(stateDir, "model.json"), data, 0o644)
 
 		o := &OpenCode{}
-		if err := o.Edit([]string{"X", "Y", "Z"}); err != nil {
+		if err := o.Edit(testLaunchModels("X", "Y", "Z")); err != nil {
 			t.Fatal(err)
 		}
 
@@ -663,7 +747,7 @@ func TestOpenCodeEdit_PreservesRecentEntries(t *testing.T) {
 		os.WriteFile(filepath.Join(stateDir, "model.json"), data, 0o644)
 
 		o := &OpenCode{}
-		if err := o.Edit([]string{"qwen3:32b"}); err != nil {
+		if err := o.Edit(testLaunchModels("qwen3:32b")); err != nil {
 			t.Fatal(err)
 		}
 
@@ -700,7 +784,7 @@ func TestOpenCodeEdit_PreservesRecentEntries(t *testing.T) {
 		os.WriteFile(filepath.Join(stateDir, "model.json"), data, 0o644)
 
 		o := &OpenCode{}
-		if err := o.Edit([]string{"llama3.2"}); err != nil {
+		if err := o.Edit(testLaunchModels("llama3.2")); err != nil {
 			t.Fatal(err)
 		}
 
@@ -742,7 +826,7 @@ func TestOpenCodeEdit_PreservesRecentEntries(t *testing.T) {
 
 		// Add 5 new models — should cap at 10 total
 		o := &OpenCode{}
-		if err := o.Edit([]string{"new-0", "new-1", "new-2", "new-3", "new-4"}); err != nil {
+		if err := o.Edit(testLaunchModels("new-0", "new-1", "new-2", "new-3", "new-4")); err != nil {
 			t.Fatal(err)
 		}
 
@@ -763,7 +847,7 @@ func TestOpenCodeEdit_BaseURL(t *testing.T) {
 	setTestHome(t, tmpDir)
 
 	// Default OLLAMA_HOST
-	o.Edit([]string{"llama3.2"})
+	o.Edit(testLaunchModels("llama3.2"))
 
 	var cfg map[string]any
 	json.Unmarshal([]byte(o.configContent), &cfg)

cmd/launch/pi.go

@@ -4,7 +4,6 @@ import (
 	"context"
 	"encoding/json"
 	"fmt"
-	"net/http"
 	"os"
 	"os/exec"
 	"path/filepath"
@@ -14,7 +13,6 @@ import (
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/cmd/internal/fileutil"
 	"github.com/ollama/ollama/envconfig"
-	"github.com/ollama/ollama/types/model"
 )
 
 // Pi implements Runner and Editor for Pi (Pi Coding Agent) integration
@@ -28,7 +26,7 @@ const (
 
 func (p *Pi) String() string { return "Pi" }
 
-func (p *Pi) Run(model string, args []string) error {
+func (p *Pi) Run(_ string, _ []LaunchModel, args []string) error {
 	fmt.Fprintf(os.Stderr, "\n%sPreparing Pi...%s\n", ansiGray, ansiReset)
 	if err := ensureNpmInstalled(); err != nil {
 		return err
@@ -183,7 +181,7 @@ func (p *Pi) Paths() []string {
 	return paths
 }
 
-func (p *Pi) Edit(models []string) error {
+func (p *Pi) Edit(models []LaunchModel) error {
 	if len(models) == 0 {
 		return nil
 	}
@@ -225,7 +223,7 @@ func (p *Pi) Edit(models []string) error {
 	// Build set of selected models to track which need to be added
 	selectedSet := make(map[string]bool, len(models))
 	for _, m := range models {
-		selectedSet[m] = true
+		selectedSet[m.Name] = true
 	}
 
 	// Build new models list:
@@ -256,11 +254,9 @@ func (p *Pi) Edit(models []string) error {
 	}
 
 	// Add newly selected models that weren't already in the list
-	client := api.NewClient(envconfig.Host(), http.DefaultClient)
-	ctx := context.Background()
 	for _, model := range models {
-		if selectedSet[model] {
-			newModels = append(newModels, createConfig(ctx, client, model))
+		if selectedSet[model.Name] {
+			newModels = append(newModels, createConfig(model))
 		}
 	}
 
@@ -284,7 +280,7 @@ func (p *Pi) Edit(models []string) error {
 	}
 
 	settings["defaultProvider"] = "ollama"
-	settings["defaultModel"] = models[0]
+	settings["defaultModel"] = models[0].Name
 
 	settingsData, err := json.MarshalIndent(settings, "", "  ")
 	if err != nil {
@@ -342,54 +338,27 @@ func hasContextWindow(cfg map[string]any) bool {
 	}
 }
 
-// createConfig builds Pi model config with capability detection
-func createConfig(ctx context.Context, client *api.Client, modelID string) map[string]any {
+// createConfig builds Pi model config with capability detection.
+func createConfig(model LaunchModel) map[string]any {
 	cfg := map[string]any{
-		"id":      modelID,
+		"id":      model.Name,
 		"_launch": true,
 	}
-	if l, ok := lookupCloudModelLimit(modelID); ok {
-		cfg["contextWindow"] = l.Context
-	}
-
-	applyCloudContextFallback := func() {
-		if l, ok := lookupCloudModelLimit(modelID); ok {
-			cfg["contextWindow"] = l.Context
-		}
-	}
-
-	resp, err := client.Show(ctx, &api.ShowRequest{Model: modelID})
-	if err != nil {
-		applyCloudContextFallback()
-		return cfg
-	}
 
 	// Set input types based on vision capability
-	if slices.Contains(resp.Capabilities, model.CapabilityVision) {
+	if model.HasCapability("vision") {
 		cfg["input"] = []string{"text", "image"}
 	} else {
 		cfg["input"] = []string{"text"}
 	}
 
 	// Set reasoning based on thinking capability
-	if slices.Contains(resp.Capabilities, model.CapabilityThinking) {
+	if model.HasCapability("thinking") {
 		cfg["reasoning"] = true
 	}
 
-	// Extract context window from ModelInfo. For known cloud models, the
-	// pre-filled shared limit remains unless the server provides a positive value.
-	hasContextWindow := false
-	for key, val := range resp.ModelInfo {
-		if strings.HasSuffix(key, ".context_length") {
-			if ctxLen, ok := val.(float64); ok && ctxLen > 0 {
-				cfg["contextWindow"] = int(ctxLen)
-				hasContextWindow = true
-			}
-			break
-		}
-	}
-	if !hasContextWindow {
-		applyCloudContextFallback()
+	if model.ContextLength > 0 {
+		cfg["contextWindow"] = model.ContextLength
 	}
 
 	return cfg

cmd/launch/pi_test.go

@@ -1,19 +1,16 @@
 package launch
 
 import (
-	"context"
 	"encoding/json"
 	"fmt"
 	"net/http"
 	"net/http/httptest"
-	"net/url"
 	"os"
 	"path/filepath"
 	"runtime"
 	"strings"
 	"testing"
 
-	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/cmd/internal/fileutil"
 	"github.com/ollama/ollama/types/model"
 )
@@ -138,7 +135,7 @@ exit 0
 		})
 
 		p := &Pi{}
-		if err := p.Run("ignored", []string{"--version"}); err != nil {
+		if err := p.Run("ignored", nil, []string{"--version"}); err != nil {
 			t.Fatalf("Run() error = %v", err)
 		}
 
@@ -181,7 +178,7 @@ exit 0
 		})
 
 		p := &Pi{}
-		err := p.Run("ignored", nil)
+		err := p.Run("ignored", nil, nil)
 		if err == nil || !strings.Contains(err.Error(), "pi installation cancelled") {
 			t.Fatalf("expected install cancellation error, got %v", err)
 		}
@@ -203,7 +200,7 @@ exit 0
 		})
 
 		p := &Pi{}
-		if err := p.Run("ignored", []string{"session"}); err != nil {
+		if err := p.Run("ignored", nil, []string{"session"}); err != nil {
 			t.Fatalf("Run() error = %v", err)
 		}
 
@@ -238,7 +235,7 @@ exit 0
 		seedNpmNoop(t, tmpDir)
 
 		p := &Pi{}
-		if err := p.Run("ignored", []string{"doctor"}); err != nil {
+		if err := p.Run("ignored", nil, []string{"doctor"}); err != nil {
 			t.Fatalf("Run() error = %v", err)
 		}
 
@@ -266,7 +263,7 @@ exit 0
 
 		p := &Pi{}
 		stderr := captureStderr(t, func() {
-			if err := p.Run("ignored", []string{"session"}); err != nil {
+			if err := p.Run("ignored", nil, []string{"session"}); err != nil {
 				t.Fatalf("Run() should continue after web search update failure, got %v", err)
 			}
 		})
@@ -301,7 +298,7 @@ exit 0
 
 		p := &Pi{}
 		stderr := captureStderr(t, func() {
-			if err := p.Run("ignored", []string{"session"}); err != nil {
+			if err := p.Run("ignored", nil, []string{"session"}); err != nil {
 				t.Fatalf("Run() should continue after web search install failure, got %v", err)
 			}
 		})
@@ -331,7 +328,7 @@ exit 0
 
 		p := &Pi{}
 		stderr := captureStderr(t, func() {
-			if err := p.Run("ignored", []string{"session"}); err != nil {
+			if err := p.Run("ignored", nil, []string{"session"}); err != nil {
 				t.Fatalf("Run() error = %v", err)
 			}
 		})
@@ -360,7 +357,7 @@ exit 0
 		seedPiScript(t, tmpDir)
 
 		p := &Pi{}
-		err := p.Run("ignored", []string{"session"})
+		err := p.Run("ignored", nil, []string{"session"})
 		if err == nil || !strings.Contains(err.Error(), "npm (Node.js) is required to launch pi") {
 			t.Fatalf("expected missing npm error, got %v", err)
 		}
@@ -435,7 +432,7 @@ func TestPiEdit(t *testing.T) {
 	}
 
 	t.Run("returns nil for empty models", func(t *testing.T) {
-		if err := pi.Edit([]string{}); err != nil {
+		if err := pi.Edit(testLaunchModels()); err != nil {
 			t.Errorf("Edit([]) error = %v, want nil", err)
 		}
 	})
@@ -444,7 +441,7 @@ func TestPiEdit(t *testing.T) {
 		cleanup()
 
 		models := []string{"llama3.2", "qwen3:8b"}
-		if err := pi.Edit(models); err != nil {
+		if err := pi.Edit(launchModelsFromNames(models)); err != nil {
 			t.Fatalf("Edit() error = %v", err)
 		}
 
@@ -497,7 +494,7 @@ func TestPiEdit(t *testing.T) {
 		}
 
 		models := []string{"new-model"}
-		if err := pi.Edit(models); err != nil {
+		if err := pi.Edit(launchModelsFromNames(models)); err != nil {
 			t.Fatalf("Edit() error = %v", err)
 		}
 
@@ -550,7 +547,7 @@ func TestPiEdit(t *testing.T) {
 			t.Fatal(err)
 		}
 
-		if err := pi.Edit([]string{"glm-5:cloud"}); err != nil {
+		if err := pi.Edit(testLaunchModels("glm-5:cloud")); err != nil {
 			t.Fatalf("Edit() error = %v", err)
 		}
 
@@ -595,7 +592,7 @@ func TestPiEdit(t *testing.T) {
 		}
 
 		newModels := []string{"new-model-1", "new-model-2"}
-		if err := pi.Edit(newModels); err != nil {
+		if err := pi.Edit(launchModelsFromNames(newModels)); err != nil {
 			t.Fatalf("Edit() error = %v", err)
 		}
 
@@ -646,7 +643,7 @@ func TestPiEdit(t *testing.T) {
 		}
 
 		newModels := []string{"keep-model", "add-model"}
-		if err := pi.Edit(newModels); err != nil {
+		if err := pi.Edit(launchModelsFromNames(newModels)); err != nil {
 			t.Fatalf("Edit() error = %v", err)
 		}
 
@@ -683,7 +680,7 @@ func TestPiEdit(t *testing.T) {
 		}
 
 		models := []string{"test-model"}
-		if err := pi.Edit(models); err != nil {
+		if err := pi.Edit(launchModelsFromNames(models)); err != nil {
 			t.Fatalf("Edit() should not fail with corrupt config, got %v", err)
 		}
 
@@ -732,7 +729,7 @@ func TestPiEdit(t *testing.T) {
 
 		// Add a new ollama-managed model
 		newModels := []string{"new-ollama-model"}
-		if err := pi.Edit(newModels); err != nil {
+		if err := pi.Edit(launchModelsFromNames(newModels)); err != nil {
 			t.Fatalf("Edit() error = %v", err)
 		}
 
@@ -793,7 +790,7 @@ func TestPiEdit(t *testing.T) {
 		}
 
 		models := []string{"llama3.2"}
-		if err := pi.Edit(models); err != nil {
+		if err := pi.Edit(launchModelsFromNames(models)); err != nil {
 			t.Fatalf("Edit() error = %v", err)
 		}
 
@@ -831,7 +828,7 @@ func TestPiEdit(t *testing.T) {
 		os.MkdirAll(configDir, 0o755)
 
 		models := []string{"qwen3:8b"}
-		if err := pi.Edit(models); err != nil {
+		if err := pi.Edit(launchModelsFromNames(models)); err != nil {
 			t.Fatalf("Edit() error = %v", err)
 		}
 
@@ -865,7 +862,7 @@ func TestPiEdit(t *testing.T) {
 		}
 
 		models := []string{"test-model"}
-		if err := pi.Edit(models); err != nil {
+		if err := pi.Edit(launchModelsFromNames(models)); err != nil {
 			t.Fatalf("Edit() should not fail with corrupt settings, got %v", err)
 		}
 
@@ -921,7 +918,7 @@ func TestPiEdit_CreatesDistinctBackupsForEachManagedFile(t *testing.T) {
 		t.Fatal(err)
 	}
 
-	if err := pi.Edit([]string{"llama3.2"}); err != nil {
+	if err := pi.Edit(testLaunchModels("llama3.2")); err != nil {
 		t.Fatalf("Edit() error = %v", err)
 	}
 
@@ -1087,19 +1084,7 @@ func TestIsPiOllamaModel(t *testing.T) {
 
 func TestCreateConfig(t *testing.T) {
 	t.Run("sets vision input when model has vision capability", func(t *testing.T) {
-		srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-			if r.URL.Path == "/api/show" {
-				fmt.Fprintf(w, `{"capabilities":["vision"],"model_info":{}}`)
-				return
-			}
-			w.WriteHeader(http.StatusNotFound)
-		}))
-		defer srv.Close()
-
-		u, _ := url.Parse(srv.URL)
-		client := api.NewClient(u, srv.Client())
-
-		cfg := createConfig(context.Background(), client, "llava:7b")
+		cfg := createConfig(LaunchModel{Name: "llava:7b", Capabilities: []model.Capability{model.CapabilityVision}})
 
 		if cfg["id"] != "llava:7b" {
 			t.Errorf("id = %v, want llava:7b", cfg["id"])
@@ -1114,19 +1099,7 @@ func TestCreateConfig(t *testing.T) {
 	})
 
 	t.Run("sets text-only input when model lacks vision", func(t *testing.T) {
-		srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-			if r.URL.Path == "/api/show" {
-				fmt.Fprintf(w, `{"capabilities":["completion"],"model_info":{}}`)
-				return
-			}
-			w.WriteHeader(http.StatusNotFound)
-		}))
-		defer srv.Close()
-
-		u, _ := url.Parse(srv.URL)
-		client := api.NewClient(u, srv.Client())
-
-		cfg := createConfig(context.Background(), client, "llama3.2")
+		cfg := createConfig(LaunchModel{Name: "llama3.2", Capabilities: []model.Capability{model.CapabilityCompletion}})
 
 		input, ok := cfg["input"].([]string)
 		if !ok || len(input) != 1 || input[0] != "text" {
@@ -1138,39 +1111,15 @@ func TestCreateConfig(t *testing.T) {
 	})
 
 	t.Run("sets reasoning when model has thinking capability", func(t *testing.T) {
-		srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-			if r.URL.Path == "/api/show" {
-				fmt.Fprintf(w, `{"capabilities":["thinking"],"model_info":{}}`)
-				return
-			}
-			w.WriteHeader(http.StatusNotFound)
-		}))
-		defer srv.Close()
-
-		u, _ := url.Parse(srv.URL)
-		client := api.NewClient(u, srv.Client())
-
-		cfg := createConfig(context.Background(), client, "qwq")
+		cfg := createConfig(LaunchModel{Name: "qwq", Capabilities: []model.Capability{model.CapabilityThinking}})
 
 		if cfg["reasoning"] != true {
 			t.Error("expected reasoning = true for thinking model")
 		}
 	})
 
-	t.Run("extracts context window from model info", func(t *testing.T) {
-		srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-			if r.URL.Path == "/api/show" {
-				fmt.Fprintf(w, `{"capabilities":[],"model_info":{"llama.context_length":131072}}`)
-				return
-			}
-			w.WriteHeader(http.StatusNotFound)
-		}))
-		defer srv.Close()
-
-		u, _ := url.Parse(srv.URL)
-		client := api.NewClient(u, srv.Client())
-
-		cfg := createConfig(context.Background(), client, "llama3.2")
+	t.Run("sets context window from metadata", func(t *testing.T) {
+		cfg := createConfig(LaunchModel{Name: "llama3.2", ContextLength: 131072})
 
 		if cfg["contextWindow"] != 131072 {
 			t.Errorf("contextWindow = %v, want 131072", cfg["contextWindow"])
@@ -1178,19 +1127,11 @@ func TestCreateConfig(t *testing.T) {
 	})
 
 	t.Run("handles all capabilities together", func(t *testing.T) {
-		srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-			if r.URL.Path == "/api/show" {
-				fmt.Fprintf(w, `{"capabilities":["vision","thinking"],"model_info":{"qwen3.context_length":32768}}`)
-				return
-			}
-			w.WriteHeader(http.StatusNotFound)
-		}))
-		defer srv.Close()
-
-		u, _ := url.Parse(srv.URL)
-		client := api.NewClient(u, srv.Client())
-
-		cfg := createConfig(context.Background(), client, "qwen3-vision")
+		cfg := createConfig(LaunchModel{
+			Name:          "qwen3-vision",
+			Capabilities:  []model.Capability{model.CapabilityVision, model.CapabilityThinking},
+			ContextLength: 32768,
+		})
 
 		input := cfg["input"].([]string)
 		if len(input) != 2 || input[0] != "text" || input[1] != "image" {
@@ -1204,17 +1145,8 @@ func TestCreateConfig(t *testing.T) {
 		}
 	})
 
-	t.Run("returns minimal config when show fails", func(t *testing.T) {
-		srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-			w.WriteHeader(http.StatusNotFound)
-			fmt.Fprintf(w, `{"error":"model not found"}`)
-		}))
-		defer srv.Close()
-
-		u, _ := url.Parse(srv.URL)
-		client := api.NewClient(u, srv.Client())
-
-		cfg := createConfig(context.Background(), client, "missing-model")
+	t.Run("returns minimal config when metadata is unavailable", func(t *testing.T) {
+		cfg := createConfig(LaunchModel{Name: "missing-model"})
 
 		if cfg["id"] != "missing-model" {
 			t.Errorf("id = %v, want missing-model", cfg["id"])
@@ -1222,49 +1154,29 @@ func TestCreateConfig(t *testing.T) {
 		if cfg["_launch"] != true {
 			t.Error("expected _launch = true")
 		}
-		// Should not have capability fields
-		if _, ok := cfg["input"]; ok {
-			t.Error("input should not be set when show fails")
+		// Input defaults to text even when capabilities are unavailable.
+		input, ok := cfg["input"].([]string)
+		if !ok || len(input) != 1 || input[0] != "text" {
+			t.Errorf("input = %v, want [text]", cfg["input"])
 		}
 		if _, ok := cfg["reasoning"]; ok {
-			t.Error("reasoning should not be set when show fails")
+			t.Error("reasoning should not be set when metadata is unavailable")
 		}
 		if _, ok := cfg["contextWindow"]; ok {
-			t.Error("contextWindow should not be set when show fails")
+			t.Error("contextWindow should not be set when metadata is unavailable")
 		}
 	})
 
-	t.Run("cloud model falls back to hardcoded context when show fails", func(t *testing.T) {
-		srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-			w.WriteHeader(http.StatusNotFound)
-			fmt.Fprintf(w, `{"error":"model not found"}`)
-		}))
-		defer srv.Close()
-
-		u, _ := url.Parse(srv.URL)
-		client := api.NewClient(u, srv.Client())
-
-		cfg := createConfig(context.Background(), client, "kimi-k2.5:cloud")
+	t.Run("cloud model falls back to hardcoded context", func(t *testing.T) {
+		cfg := createConfig(fallbackLaunchModel("kimi-k2.5:cloud"))
 
 		if cfg["contextWindow"] != 262_144 {
 			t.Errorf("contextWindow = %v, want 262144", cfg["contextWindow"])
 		}
 	})
 
-	t.Run("cloud model falls back to hardcoded context when show omits model info", func(t *testing.T) {
-		srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-			if r.URL.Path == "/api/show" {
-				fmt.Fprintf(w, `{"capabilities":[],"model_info":{}}`)
-				return
-			}
-			w.WriteHeader(http.StatusNotFound)
-		}))
-		defer srv.Close()
-
-		u, _ := url.Parse(srv.URL)
-		client := api.NewClient(u, srv.Client())
-
-		cfg := createConfig(context.Background(), client, "glm-5:cloud")
+	t.Run("cloud model uses hardcoded context when tags omit context", func(t *testing.T) {
+		cfg := createConfig(fallbackLaunchModel("glm-5:cloud"))
 
 		if cfg["contextWindow"] != 202_752 {
 			t.Errorf("contextWindow = %v, want 202752", cfg["contextWindow"])
@@ -1272,35 +1184,14 @@ func TestCreateConfig(t *testing.T) {
 	})
 
 	t.Run("cloud model with dash suffix falls back to hardcoded context", func(t *testing.T) {
-		srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-			w.WriteHeader(http.StatusNotFound)
-			fmt.Fprintf(w, `{"error":"model not found"}`)
-		}))
-		defer srv.Close()
-
-		u, _ := url.Parse(srv.URL)
-		client := api.NewClient(u, srv.Client())
-
-		cfg := createConfig(context.Background(), client, "gpt-oss:120b-cloud")
+		cfg := createConfig(fallbackLaunchModel("gpt-oss:120b-cloud"))
 
 		if cfg["contextWindow"] != 131_072 {
 			t.Errorf("contextWindow = %v, want 131072", cfg["contextWindow"])
 		}
 	})
 	t.Run("skips zero context length", func(t *testing.T) {
-		srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-			if r.URL.Path == "/api/show" {
-				fmt.Fprintf(w, `{"capabilities":[],"model_info":{"llama.context_length":0}}`)
-				return
-			}
-			w.WriteHeader(http.StatusNotFound)
-		}))
-		defer srv.Close()
-
-		u, _ := url.Parse(srv.URL)
-		client := api.NewClient(u, srv.Client())
-
-		cfg := createConfig(context.Background(), client, "test-model")
+		cfg := createConfig(LaunchModel{Name: "test-model", ContextLength: 0})
 
 		if _, ok := cfg["contextWindow"]; ok {
 			t.Error("contextWindow should not be set for zero value")

cmd/launch/poolside.go

@@ -29,7 +29,7 @@ func (p *Poolside) args(model string, extra []string) []string {
 	return args
 }
 
-func (p *Poolside) Run(model string, args []string) error {
+func (p *Poolside) Run(model string, _ []LaunchModel, args []string) error {
 	if poolsideGOOS == "windows" {
 		return poolsideUnsupportedError()
 	}

cmd/launch/poolside_test.go

@@ -51,7 +51,7 @@ func TestPoolsideRunSetsOllamaEnv(t *testing.T) {
 	t.Setenv("OLLAMA_HOST", "http://127.0.0.1:11434")
 
 	p := &Poolside{}
-	if err := p.Run("qwen3.5", []string{"session"}); err != nil {
+	if err := p.Run("qwen3.5", nil, []string{"session"}); err != nil {
 		t.Fatalf("Run returned error: %v", err)
 	}
 
@@ -78,7 +78,7 @@ func TestPoolsideRunWindowsUnsupported(t *testing.T) {
 	t.Cleanup(func() { poolsideGOOS = prev })
 
 	p := &Poolside{}
-	err := p.Run("kimi-k2.6:cloud", nil)
+	err := p.Run("kimi-k2.6:cloud", nil, nil)
 	if err == nil {
 		t.Fatal("expected Windows unsupported error")
 	}

cmd/launch/registry.go

@@ -33,7 +33,7 @@ type IntegrationInfo struct {
 	Description string
 }
 
-var launcherIntegrationOrder = []string{"claude", "openclaw", "hermes", "opencode", "codex", "copilot", "droid", "pi", "pool"}
+var launcherIntegrationOrder = []string{"claude", "codex-app", "hermes", "openclaw", "opencode", "codex", "copilot", "droid", "pi", "pool"}
 
 var integrationSpecs = []*IntegrationSpec{
 	{
@@ -87,6 +87,18 @@ var integrationSpecs = []*IntegrationSpec{
 			Command: []string{"npm", "install", "-g", "@openai/codex"},
 		},
 	},
+	{
+		Name:        "codex-app",
+		Runner:      &CodexApp{},
+		Aliases:     []string{"codex-desktop", "codex-gui"},
+		Description: "An AI agent you can delegate real work to, by OpenAI",
+		Install: IntegrationInstallSpec{
+			CheckInstalled: func() bool {
+				return codexAppInstalled()
+			},
+			URL: "https://developers.openai.com/codex/quickstart",
+		},
+	},
 	{
 		Name:        "kimi",
 		Runner:      &Kimi{},

cmd/launch/runner_exec_only_test.go

@@ -84,7 +84,7 @@ func TestEditorRunsDoNotRewriteConfig(t *testing.T) {
 			t.Setenv("PATH", binDir)
 
 			configPath := tt.checkPath(home)
-			if err := tt.runner.Run("llama3.2", nil); err != nil {
+			if err := tt.runner.Run("llama3.2", nil, nil); err != nil {
 				t.Fatalf("Run returned error: %v", err)
 			}
 			if _, err := os.Stat(configPath); !os.IsNotExist(err) {

cmd/launch/test_config_helpers_test.go

@@ -41,6 +41,10 @@ func setTestHome(t *testing.T, dir string) {
 	setLaunchTestHome(t, dir)
 }
 
+func testLaunchModels(names ...string) []LaunchModel {
+	return launchModelsFromNames(names)
+}
+
 func SaveIntegration(appName string, models []string) error {
 	return config.SaveIntegration(appName, models)
 }

cmd/launch/vscode.go

@@ -126,7 +126,7 @@ const (
 	minVSCodeVersion      = "1.113"
 )
 
-func (v *VSCode) Run(model string, args []string) error {
+func (v *VSCode) Run(model string, _ []LaunchModel, args []string) error {
 	v.checkVSCodeVersion()
 	v.checkCopilotChatVersion()
 
@@ -238,7 +238,7 @@ func (v *VSCode) Paths() []string {
 	return nil
 }
 
-func (v *VSCode) Edit(models []string) error {
+func (v *VSCode) Edit(models []LaunchModel) error {
 	if len(models) == 0 {
 		return nil
 	}

cmd/launch/vscode_test.go

@@ -113,7 +113,7 @@ func TestVSCodeEdit(t *testing.T) {
 				os.WriteFile(clmPath, []byte(tt.setup), 0o644)
 			}
 
-			if err := v.Edit(tt.models); err != nil {
+			if err := v.Edit(launchModelsFromNames(tt.models)); err != nil {
 				t.Fatal(err)
 			}
 
@@ -134,7 +134,7 @@ func TestVSCodeEditCleansUpOldSettings(t *testing.T) {
 	os.MkdirAll(filepath.Dir(settingsPath), 0o755)
 	os.WriteFile(settingsPath, []byte(`{"github.copilot.chat.byok.ollamaEndpoint": "http://old:11434", "ollama.launch.configured": true, "editor.fontSize": 14}`), 0o644)
 
-	if err := v.Edit([]string{"llama3.2"}); err != nil {
+	if err := v.Edit(testLaunchModels("llama3.2")); err != nil {
 		t.Fatal(err)
 	}
 
@@ -180,7 +180,7 @@ func TestVSCodeEdit_CreatesDistinctBackupsForManagedFiles(t *testing.T) {
 		t.Fatal(err)
 	}
 
-	if err := v.Edit([]string{"llama3.2"}); err != nil {
+	if err := v.Edit(testLaunchModels("llama3.2")); err != nil {
 		t.Fatal(err)
 	}
 

cmd/tui/tui_test.go

@@ -29,6 +29,13 @@ func launcherTestState() *launch.LauncherState {
 				Selectable:  true,
 				Changeable:  true,
 			},
+			"codex-app": {
+				Name:        "codex-app",
+				DisplayName: "Codex App",
+				Description: "An AI agent you can delegate real work to, by OpenAI",
+				Selectable:  true,
+				Changeable:  true,
+			},
 			"openclaw": {
 				Name:            "openclaw",
 				DisplayName:     "OpenClaw",
@@ -122,12 +129,25 @@ func expectedExpandedSequence(state *launch.LauncherState) []string {
 func TestMenuRendersPinnedItemsAndMore(t *testing.T) {
 	state := launcherTestState()
 	menu := newModel(state)
+	wantPrefix := []string{"run", "claude", "codex-app", "hermes", "openclaw"}
+	if findMenuCursorByIntegration(menu.items, "codex-app") == -1 {
+		wantPrefix = []string{"run", "claude", "hermes", "openclaw", "opencode"}
+	}
+	if got := integrationSequence(menu.items); len(got) < len(wantPrefix) {
+		t.Fatalf("expected at least %d menu items, got %v", len(wantPrefix), got)
+	} else if diff := compareStrings(got[:len(wantPrefix)], wantPrefix); diff != "" {
+		t.Fatalf("unexpected primary TUI order: %s", diff)
+	}
+
 	view := menu.View()
-	for _, want := range []string{"Chat with a model", "Launch Claude Code", "Launch OpenClaw", "Launch Hermes Agent", "More..."} {
+	for _, want := range []string{"Chat with a model", "Launch Claude Code", "Launch Hermes Agent", "Launch OpenClaw", "More..."} {
 		if !strings.Contains(view, want) {
 			t.Fatalf("expected menu view to contain %q\n%s", want, view)
 		}
 	}
+	if findMenuCursorByIntegration(menu.items, "codex-app") != -1 && !strings.Contains(view, "Launch Codex App") {
+		t.Fatalf("expected menu view to contain Codex App\n%s", view)
+	}
 	if strings.Contains(view, "Launch Claude Desktop") {
 		t.Fatalf("expected hidden Claude Desktop to be absent\n%s", view)
 	}

convert/convert.go

@@ -147,7 +147,9 @@ func (ModelParameters) KV(t *Tokenizer) KV {
 	}
 
 	for _, sv := range t.SpecialVocabulary {
-		kv[fmt.Sprintf("tokenizer.ggml.add_%s_token", sv.Key())] = sv.AddToken
+		if sv.AddTokenSet {
+			kv[fmt.Sprintf("tokenizer.ggml.add_%s_token", sv.Key())] = sv.AddToken
+		}
 		kv[fmt.Sprintf("tokenizer.ggml.%s_token_id", sv.Key())] = uint32(sv.ID)
 		if len(sv.IDs) > 0 {
 			kv[fmt.Sprintf("tokenizer.ggml.%s_token_ids", sv.Key())] = sv.IDs
@@ -200,10 +202,32 @@ type ModelConverter interface {
 	specialTokenTypes() []string
 }
 
+// MultimodalConverter splits checkpoints with embedded vision/projector
+// weights into a text model GGUF and a separate projector GGUF.
+type MultimodalConverter interface {
+	ModelConverter
+	TextKV(*Tokenizer) KV
+	TextTensors([]Tensor, *Tokenizer) []*ggml.Tensor
+	ProjectorKV(*Tokenizer) KV
+	ProjectorTensors([]Tensor) []*ggml.Tensor
+}
+
 type moreParser interface {
 	parseMore(fs.FS) error
 }
 
+type extraTensorParser interface {
+	extraTensors(fs.FS) ([]Tensor, error)
+}
+
+type tokenizerAdjuster interface {
+	adjustTokenizer(*Tokenizer)
+}
+
+type tokenizerAwareTensorConverter interface {
+	TensorsWithTokenizer([]Tensor, *Tokenizer) []*ggml.Tensor
+}
+
 type AdapterConverter interface {
 	// KV maps parameters to LLM key-values
 	KV(ofs.Config) KV
@@ -288,6 +312,8 @@ func LoadModelMetadata(fsys fs.FS) (ModelKV, *Tokenizer, error) {
 		conv = &gemma2Model{}
 	case "Gemma3ForCausalLM", "Gemma3ForConditionalGeneration":
 		conv = &gemma3Model{Architecture: p.Architectures[0]}
+	case "Gemma3TextModel":
+		conv = &embeddingGemmaModel{}
 	case "Gemma3nForConditionalGeneration":
 		conv = &gemma3nModel{}
 	case "Gemma4ForCausalLM", "Gemma4ForConditionalGeneration":
@@ -348,6 +374,9 @@ func LoadModelMetadata(fsys fs.FS) (ModelKV, *Tokenizer, error) {
 	if err != nil {
 		return nil, nil, err
 	}
+	if ta, ok := conv.(tokenizerAdjuster); ok {
+		ta.adjustTokenizer(t)
+	}
 
 	vocabSize := int(cmp.Or(p.VocabSize, p.TextModel.VocabSize))
 
@@ -375,7 +404,7 @@ func LoadModelMetadata(fsys fs.FS) (ModelKV, *Tokenizer, error) {
 // and files it finds in the input path.
 // Supported input model formats include safetensors.
 // Supported input tokenizers files include tokenizer.json (preferred) and tokenizer.model.
-func ConvertModel(fsys fs.FS, f *os.File) error {
+func ConvertModel(fsys fs.FS, f *os.File, projectorFiles ...*os.File) error {
 	kv, t, err := LoadModelMetadata(fsys)
 	if err != nil {
 		return err
@@ -387,7 +416,47 @@ func ConvertModel(fsys fs.FS, f *os.File) error {
 		return err
 	}
 
-	return writeFile(f, conv.KV(t), conv.Tensors(ts))
+	if tp, ok := conv.(extraTensorParser); ok {
+		extra, err := tp.extraTensors(fsys)
+		if err != nil {
+			return err
+		}
+		ts = append(ts, extra...)
+	}
+
+	if err := ensureUniqueTensorNames(ts); err != nil {
+		return err
+	}
+
+	if mc, ok := conv.(MultimodalConverter); ok && len(projectorFiles) > 0 && projectorFiles[0] != nil {
+		projectorTensors := mc.ProjectorTensors(ts)
+		if len(projectorTensors) > 0 {
+			if err := writeFile(f, mc.TextKV(t), mc.TextTensors(ts, t)); err != nil {
+				return err
+			}
+			return writeFile(projectorFiles[0], mc.ProjectorKV(t), projectorTensors)
+		}
+	}
+
+	var tensors []*ggml.Tensor
+	if tc, ok := conv.(tokenizerAwareTensorConverter); ok {
+		tensors = tc.TensorsWithTokenizer(ts, t)
+	} else {
+		tensors = conv.Tensors(ts)
+	}
+
+	return writeFile(f, conv.KV(t), tensors)
+}
+
+func ensureUniqueTensorNames(ts []Tensor) error {
+	names := make(map[string]struct{}, len(ts))
+	for _, t := range ts {
+		if _, ok := names[t.Name()]; ok {
+			return fmt.Errorf("duplicate tensor name '%s' was found for this model", t.Name())
+		}
+		names[t.Name()] = struct{}{}
+	}
+	return nil
 }
 
 func writeFile(f *os.File, kv KV, ts []*ggml.Tensor) error {

convert/convert_embeddinggemma.go

@@ -0,0 +1,280 @@
+package convert
+
+import (
+	"cmp"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"io/fs"
+	"path"
+	"slices"
+	"strings"
+
+	"github.com/ollama/ollama/fs/ggml"
+)
+
+type embeddingGemmaModel struct {
+	gemmaModel
+	RopeLocalTheta float32 `json:"rope_local_base_freq"`
+	RopeTheta      float32 `json:"rope_theta"`
+	SlidingWindow  uint32  `json:"sliding_window"`
+
+	poolingType  uint32
+	denseModules []embeddingGemmaDenseModule
+}
+
+type embeddingGemmaDenseModule struct {
+	path       string
+	tensorName string
+	in, out    uint32
+}
+
+var (
+	_ ModelConverter    = (*embeddingGemmaModel)(nil)
+	_ moreParser        = (*embeddingGemmaModel)(nil)
+	_ extraTensorParser = (*embeddingGemmaModel)(nil)
+	_ tokenizerAdjuster = (*embeddingGemmaModel)(nil)
+)
+
+func (m *embeddingGemmaModel) KV(t *Tokenizer) KV {
+	kv := m.ModelParameters.KV(t)
+	kv["general.architecture"] = "gemma-embedding"
+	kv["gemma-embedding.context_length"] = cmp.Or(m.MaxPositionEmbeddings, uint32(2048))
+	kv["gemma-embedding.embedding_length"] = m.HiddenSize
+	kv["gemma-embedding.block_count"] = m.HiddenLayers
+	kv["gemma-embedding.feed_forward_length"] = m.IntermediateSize
+	kv["gemma-embedding.attention.head_count"] = m.NumAttentionHeads
+	kv["gemma-embedding.attention.head_count_kv"] = m.NumKeyValueHeads
+	kv["gemma-embedding.attention.layer_norm_rms_epsilon"] = cmp.Or(m.RMSNormEPS, float32(1e-6))
+	kv["gemma-embedding.attention.key_length"] = m.HeadDim
+	kv["gemma-embedding.attention.value_length"] = m.HeadDim
+	kv["gemma-embedding.attention.sliding_window"] = m.SlidingWindow
+	kv["gemma-embedding.rope.freq_base"] = cmp.Or(m.RopeTheta, float32(1000000.0))
+	kv["gemma-embedding.rope.freq_base_swa"] = cmp.Or(m.RopeLocalTheta, float32(10000.0))
+	kv["gemma-embedding.pooling_type"] = cmp.Or(m.poolingType, uint32(1))
+
+	for _, dense := range m.denseModules {
+		kv["gemma-embedding."+dense.tensorName+"_feat_in"] = dense.in
+		kv["gemma-embedding."+dense.tensorName+"_feat_out"] = dense.out
+	}
+
+	return kv
+}
+
+func (m *embeddingGemmaModel) parseMore(fsys fs.FS) error {
+	bts, err := fs.ReadFile(fsys, "modules.json")
+	if err != nil {
+		if errors.Is(err, fs.ErrNotExist) {
+			return errors.New("embeddinggemma requires sentence-transformers modules.json")
+		}
+		return err
+	}
+
+	var modules []struct {
+		Type string `json:"type"`
+		Path string `json:"path"`
+	}
+
+	if err := json.Unmarshal(bts, &modules); err != nil {
+		return err
+	}
+
+	m.poolingType = 1
+	m.denseModules = nil
+	for _, module := range modules {
+		switch module.Type {
+		case "sentence_transformers.models.Pooling":
+			poolingType, err := embeddingGemmaPoolingType(fsys, module.Path)
+			if err != nil {
+				return err
+			}
+			if poolingType != 0 {
+				m.poolingType = poolingType
+			}
+		case "sentence_transformers.models.Dense":
+			dense, ok, err := embeddingGemmaDenseModuleConfig(fsys, module.Path)
+			if err != nil {
+				return err
+			}
+			if ok {
+				m.denseModules = append(m.denseModules, dense)
+			}
+		}
+	}
+
+	slices.SortFunc(m.denseModules, func(a, b embeddingGemmaDenseModule) int {
+		return strings.Compare(a.tensorName, b.tensorName)
+	})
+
+	if len(m.denseModules) != 2 ||
+		m.denseModules[0].tensorName != "dense_2" ||
+		m.denseModules[1].tensorName != "dense_3" {
+		return errors.New("embeddinggemma requires sentence-transformers 2_Dense and 3_Dense modules")
+	}
+
+	return nil
+}
+
+func (m *embeddingGemmaModel) adjustTokenizer(t *Tokenizer) {
+	n := int(m.VocabSize)
+	if n == 0 || len(t.Vocabulary.Tokens) <= n {
+		return
+	}
+
+	t.Vocabulary.Tokens = t.Vocabulary.Tokens[:n]
+	if len(t.Vocabulary.Scores) > n {
+		t.Vocabulary.Scores = t.Vocabulary.Scores[:n]
+	}
+	if len(t.Vocabulary.Types) > n {
+		t.Vocabulary.Types = t.Vocabulary.Types[:n]
+	}
+}
+
+func embeddingGemmaPoolingType(fsys fs.FS, modulePath string) (uint32, error) {
+	if modulePath == "" {
+		return 0, nil
+	}
+
+	bts, err := fs.ReadFile(fsys, path.Join(modulePath, "config.json"))
+	if err != nil {
+		if errors.Is(err, fs.ErrNotExist) {
+			return 0, nil
+		}
+		return 0, err
+	}
+
+	var cfg struct {
+		PoolingModeMeanTokens bool `json:"pooling_mode_mean_tokens"`
+		PoolingModeCLSToken   bool `json:"pooling_mode_cls_token"`
+	}
+	if err := json.Unmarshal(bts, &cfg); err != nil {
+		return 0, err
+	}
+
+	switch {
+	case cfg.PoolingModeMeanTokens:
+		return 1, nil
+	case cfg.PoolingModeCLSToken:
+		return 2, nil
+	default:
+		return 0, nil
+	}
+}
+
+func embeddingGemmaDenseModuleConfig(fsys fs.FS, modulePath string) (embeddingGemmaDenseModule, bool, error) {
+	tensorName, ok := embeddingGemmaDenseTensorName(modulePath)
+	if !ok {
+		return embeddingGemmaDenseModule{}, false, nil
+	}
+
+	weightsPath := path.Join(modulePath, "model.safetensors")
+	if _, err := fs.Stat(fsys, weightsPath); err != nil {
+		if errors.Is(err, fs.ErrNotExist) {
+			return embeddingGemmaDenseModule{}, false, nil
+		}
+		return embeddingGemmaDenseModule{}, false, err
+	}
+
+	bts, err := fs.ReadFile(fsys, path.Join(modulePath, "config.json"))
+	if err != nil {
+		return embeddingGemmaDenseModule{}, false, err
+	}
+
+	var cfg struct {
+		InFeatures  uint32 `json:"in_features"`
+		OutFeatures uint32 `json:"out_features"`
+		Bias        bool   `json:"bias"`
+	}
+	if err := json.Unmarshal(bts, &cfg); err != nil {
+		return embeddingGemmaDenseModule{}, false, err
+	}
+	if cfg.InFeatures == 0 || cfg.OutFeatures == 0 {
+		return embeddingGemmaDenseModule{}, false, errors.New("embeddinggemma dense layer config missing in/out features")
+	}
+	if cfg.Bias {
+		return embeddingGemmaDenseModule{}, false, fmt.Errorf("embeddinggemma dense layer %s has unsupported bias", modulePath)
+	}
+
+	return embeddingGemmaDenseModule{
+		path:       weightsPath,
+		tensorName: tensorName,
+		in:         cfg.InFeatures,
+		out:        cfg.OutFeatures,
+	}, true, nil
+}
+
+func embeddingGemmaDenseTensorName(modulePath string) (string, bool) {
+	switch modulePath {
+	case "2_Dense":
+		return "dense_2", true
+	case "3_Dense":
+		return "dense_3", true
+	default:
+		return "", false
+	}
+}
+
+func (m *embeddingGemmaModel) extraTensors(fsys fs.FS) ([]Tensor, error) {
+	var extra []Tensor
+	for _, dense := range m.denseModules {
+		ts, err := parseSafetensors(fsys, strings.NewReplacer("linear.", dense.tensorName+"."), dense.path)
+		if err != nil {
+			return nil, err
+		}
+
+		foundWeight := false
+		for _, t := range ts {
+			if t.Name() == dense.tensorName+".weight" {
+				extra = append(extra, t)
+				foundWeight = true
+			}
+		}
+		if !foundWeight {
+			return nil, fmt.Errorf("embeddinggemma dense module %s missing linear.weight", dense.path)
+		}
+	}
+
+	return extra, nil
+}
+
+func (m *embeddingGemmaModel) Tensors(ts []Tensor) []*ggml.Tensor {
+	out := make([]*ggml.Tensor, 0, len(ts))
+	for _, t := range ts {
+		name := t.Name()
+		if name == "norm.weight" {
+			name = "output_norm.weight"
+		}
+		if strings.HasSuffix(name, "_norm.weight") {
+			t.SetRepacker(m.addOne)
+		}
+
+		out = append(out, &ggml.Tensor{
+			Name:     name,
+			Kind:     t.Kind(),
+			Shape:    t.Shape(),
+			WriterTo: t,
+		})
+	}
+
+	return out
+}
+
+func (m *embeddingGemmaModel) Replacements() []string {
+	return []string{
+		"embed_tokens.", "token_embd.",
+		"layers.", "blk.",
+		"input_layernorm", "attn_norm",
+		"self_attn.q_proj", "attn_q",
+		"self_attn.q_norm", "attn_q_norm",
+		"self_attn.k_proj", "attn_k",
+		"self_attn.k_norm", "attn_k_norm",
+		"self_attn.v_proj", "attn_v",
+		"self_attn.o_proj", "attn_output",
+		"mlp.gate_proj", "ffn_gate",
+		"mlp.down_proj", "ffn_down",
+		"mlp.up_proj", "ffn_up",
+		"post_attention_layernorm", "post_attention_norm",
+		"pre_feedforward_layernorm", "ffn_norm",
+		"post_feedforward_layernorm", "post_ffw_norm",
+	}
+}

convert/convert_embeddinggemma_test.go

@@ -0,0 +1,229 @@
+package convert
+
+import (
+	"bytes"
+	"encoding/binary"
+	"encoding/json"
+	"io"
+	"math"
+	"os"
+	"path/filepath"
+	"slices"
+	"testing"
+
+	"github.com/ollama/ollama/fs/ggml"
+)
+
+func TestConvertEmbeddingGemmaSentenceTransformers(t *testing.T) {
+	tempDir := t.TempDir()
+
+	writeJSONFile(t, filepath.Join(tempDir, "config.json"), map[string]any{
+		"architectures":               []string{"Gemma3TextModel"},
+		"vocab_size":                  uint32(4),
+		"max_position_embeddings":     uint32(2048),
+		"hidden_size":                 uint32(8),
+		"num_hidden_layers":           uint32(1),
+		"intermediate_size":           uint32(12),
+		"num_attention_heads":         uint32(1),
+		"num_key_value_heads":         uint32(1),
+		"head_dim":                    uint32(8),
+		"rms_norm_eps":                float32(1e-6),
+		"rope_theta":                  float32(1000000),
+		"rope_local_base_freq":        float32(10000),
+		"sliding_window":              uint32(512),
+		"use_bidirectional_attention": true,
+	})
+	writeJSONFile(t, filepath.Join(tempDir, "tokenizer.json"), map[string]any{
+		"model": map[string]any{
+			"vocab": map[string]int{
+				"<pad>": 0,
+				"<eos>": 1,
+				"<bos>": 2,
+				"<unk>": 3,
+			},
+		},
+		"added_tokens": []map[string]any{
+			{"id": 4, "content": "<image_soft_token>", "special": true},
+		},
+	})
+	writeJSONFile(t, filepath.Join(tempDir, "modules.json"), []map[string]string{
+		{"type": "sentence_transformers.models.Transformer", "path": ""},
+		{"type": "sentence_transformers.models.Pooling", "path": "1_Pooling"},
+		{"type": "sentence_transformers.models.Dense", "path": "2_Dense"},
+		{"type": "sentence_transformers.models.Dense", "path": "3_Dense"},
+		{"type": "sentence_transformers.models.Normalize", "path": "4_Normalize"},
+	})
+	writeJSONFile(t, filepath.Join(tempDir, "1_Pooling", "config.json"), map[string]any{
+		"pooling_mode_mean_tokens": true,
+	})
+	writeJSONFile(t, filepath.Join(tempDir, "2_Dense", "config.json"), map[string]any{
+		"in_features":  uint32(8),
+		"out_features": uint32(16),
+		"bias":         false,
+	})
+	writeJSONFile(t, filepath.Join(tempDir, "3_Dense", "config.json"), map[string]any{
+		"in_features":  uint32(16),
+		"out_features": uint32(8),
+		"bias":         false,
+	})
+
+	writeSafetensorsFile(t, filepath.Join(tempDir, "model.safetensors"), []safetensorFixtureTensor{
+		{name: "embed_tokens.weight", shape: []int{4, 8}},
+		{name: "norm.weight", shape: []int{8}},
+		{name: "layers.0.input_layernorm.weight", shape: []int{8}},
+		{name: "layers.0.self_attn.q_proj.weight", shape: []int{8, 8}},
+	})
+	writeSafetensorsFile(t, filepath.Join(tempDir, "2_Dense", "model.safetensors"), []safetensorFixtureTensor{
+		{name: "linear.weight", shape: []int{16, 8}},
+	})
+	writeSafetensorsFile(t, filepath.Join(tempDir, "3_Dense", "model.safetensors"), []safetensorFixtureTensor{
+		{name: "linear.weight", shape: []int{8, 16}},
+	})
+
+	f, kv, tensors := convertFull(t, os.DirFS(tempDir))
+	defer f.Close()
+
+	if got := kv.Architecture(); got != "gemma-embedding" {
+		t.Fatalf("architecture = %q, want gemma-embedding", got)
+	}
+
+	for key, want := range map[string]uint32{
+		"dense_2_feat_in":          8,
+		"dense_2_feat_out":         16,
+		"dense_3_feat_in":          16,
+		"dense_3_feat_out":         8,
+		"pooling_type":             1,
+		"attention.sliding_window": 512,
+	} {
+		if got := kv.Uint(key); got != want {
+			t.Errorf("%s = %d, want %d", key, got, want)
+		}
+	}
+
+	if got := kv.Float("rope.freq_base_swa"); got != 10000 {
+		t.Errorf("rope.freq_base_swa = %v, want 10000", got)
+	}
+	if got := kv.Strings("tokenizer.ggml.tokens"); len(got) != 4 {
+		t.Errorf("token count = %d, want 4", len(got))
+	}
+
+	names := tensorNames(tensors)
+	for _, name := range []string{
+		"token_embd.weight",
+		"output_norm.weight",
+		"blk.0.attn_norm.weight",
+		"blk.0.attn_q.weight",
+		"dense_2.weight",
+		"dense_3.weight",
+	} {
+		if !slices.Contains(names, name) {
+			t.Errorf("missing tensor %s", name)
+		}
+	}
+
+	assertF32TensorValues(t, f, tensors, "output_norm.weight", 1)
+	assertF32TensorValues(t, f, tensors, "blk.0.attn_norm.weight", 1)
+}
+
+type safetensorFixtureTensor struct {
+	name  string
+	shape []int
+}
+
+func writeJSONFile(t *testing.T, path string, value any) {
+	t.Helper()
+
+	if err := os.MkdirAll(filepath.Dir(path), 0o755); err != nil {
+		t.Fatal(err)
+	}
+
+	bts, err := json.Marshal(value)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	if err := os.WriteFile(path, bts, 0o644); err != nil {
+		t.Fatal(err)
+	}
+}
+
+func writeSafetensorsFile(t *testing.T, path string, tensors []safetensorFixtureTensor) {
+	t.Helper()
+
+	if err := os.MkdirAll(filepath.Dir(path), 0o755); err != nil {
+		t.Fatal(err)
+	}
+
+	offset := 0
+	metadata := map[string]*tensorData{}
+	for _, tensor := range tensors {
+		size := 4
+		for _, dim := range tensor.shape {
+			size *= dim
+		}
+
+		metadata[tensor.name] = &tensorData{
+			Offsets: []int{offset, offset + size},
+			Type:    "F32",
+			Shape:   tensor.shape,
+		}
+		offset += size
+	}
+
+	header, err := json.Marshal(metadata)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	var buf bytes.Buffer
+	if err := binary.Write(&buf, binary.LittleEndian, int64(len(header))); err != nil {
+		t.Fatal(err)
+	}
+	if _, err := buf.Write(header); err != nil {
+		t.Fatal(err)
+	}
+	if _, err := buf.Write(make([]byte, offset)); err != nil {
+		t.Fatal(err)
+	}
+
+	if err := os.WriteFile(path, buf.Bytes(), 0o644); err != nil {
+		t.Fatal(err)
+	}
+}
+
+func tensorNames(tensors ggml.Tensors) []string {
+	names := make([]string, 0, len(tensors.Items()))
+	for _, tensor := range tensors.Items() {
+		names = append(names, tensor.Name)
+	}
+	return names
+}
+
+func assertF32TensorValues(t *testing.T, f *os.File, tensors ggml.Tensors, name string, want float32) {
+	t.Helper()
+
+	var tensor *ggml.Tensor
+	for _, item := range tensors.Items() {
+		if item.Name == name {
+			tensor = item
+			break
+		}
+	}
+	if tensor == nil {
+		t.Fatalf("missing tensor %s", name)
+	}
+	if tensor.Kind != uint32(ggml.TensorTypeF32) {
+		t.Fatalf("%s kind = %d, want F32", name, tensor.Kind)
+	}
+
+	bts := make([]byte, tensor.Size())
+	reader := io.NewSectionReader(f, int64(tensors.Offset+tensor.Offset), int64(tensor.Size()))
+	if _, err := io.ReadFull(reader, bts); err != nil {
+		t.Fatal(err)
+	}
+	for i := 0; i < len(bts); i += 4 {
+		if got := math.Float32frombits(binary.LittleEndian.Uint32(bts[i:])); got != want {
+			t.Fatalf("%s[%d] = %v, want %v", name, i/4, got, want)
+		}
+	}
+}

convert/convert_gemma3.go

@@ -2,7 +2,11 @@ package convert
 
 import (
 	"cmp"
+	"fmt"
 	"slices"
+	"strings"
+
+	"github.com/ollama/ollama/fs/ggml"
 )
 
 type gemma3Model struct {
@@ -178,3 +182,42 @@ func (p *gemma3Model) Replacements() []string {
 		"multi_modal_projector", "mm",
 	}
 }
+
+func (p *gemma3Model) TensorsWithTokenizer(ts []Tensor, t *Tokenizer) []*ggml.Tensor {
+	vocabSize := uint64(0)
+	if t != nil && t.Vocabulary != nil {
+		vocabSize = uint64(len(t.Vocabulary.Tokens))
+	}
+
+	var out []*ggml.Tensor
+	for _, tensor := range ts {
+		name := tensor.Name()
+		gt := &ggml.Tensor{
+			Name:     name,
+			Kind:     tensor.Kind(),
+			Shape:    tensor.Shape(),
+			WriterTo: tensor,
+		}
+
+		if !strings.HasPrefix(name, "v.") && strings.HasSuffix(name, "_norm.weight") {
+			tensor.SetRepacker(p.addOne)
+		}
+
+		if vocabSize > 0 && name == "token_embd.weight" && len(gt.Shape) >= 2 && gt.Shape[0] > vocabSize {
+			gt.Shape = slices.Clone(gt.Shape)
+			embdDim := gt.Shape[1]
+			gt.Shape[0] = vocabSize
+			tensor.SetRepacker(func(_ string, data []float32, _ []uint64) ([]float32, error) {
+				n := vocabSize * embdDim
+				if uint64(len(data)) < n {
+					return nil, fmt.Errorf("gemma3 token_embd.weight has %d values, need %d", len(data), n)
+				}
+				return data[:n], nil
+			})
+		}
+
+		out = append(out, gt)
+	}
+
+	return out
+}

convert/convert_gemma3_test.go

@@ -0,0 +1,34 @@
+package convert
+
+import (
+	"slices"
+	"testing"
+)
+
+func TestGemma3TensorsWithTokenizerTruncatesPaddedEmbedding(t *testing.T) {
+	p := gemma3Model{}
+	embedding := &fakeTensor{
+		name:  "token_embd.weight",
+		shape: []uint64{5, 2},
+		data:  []float32{0, 1, 2, 3, 4, 5, 6, 7, 8, 9},
+	}
+
+	out := p.TensorsWithTokenizer([]Tensor{embedding}, &Tokenizer{
+		Vocabulary: &Vocabulary{Tokens: []string{"a", "b", "<image>"}},
+	})
+
+	if len(out) != 1 {
+		t.Fatalf("expected 1 tensor, got %d", len(out))
+	}
+	if got, want := out[0].Shape, []uint64{3, 2}; !slices.Equal(got, want) {
+		t.Fatalf("token_embd.weight shape = %v, want %v", got, want)
+	}
+
+	got, err := embedding.repacker(embedding.name, embedding.data, embedding.shape)
+	if err != nil {
+		t.Fatalf("unexpected repacker error: %v", err)
+	}
+	if want := embedding.data[:6]; !slices.Equal(got, want) {
+		t.Fatalf("truncated embedding = %v, want %v", got, want)
+	}
+}

convert/convert_gemma3n.go

@@ -1,6 +1,8 @@
 package convert
 
 import (
+	"encoding/json"
+	"fmt"
 	"slices"
 	"strings"
 
@@ -14,30 +16,58 @@ type gemma3nModel struct {
 	ModelParameters
 
 	TextModel struct {
-		ActivationSparsityPattern []float32 `json:"activation_sparsity_pattern"`
-		AltupActiveIdx            uint32    `json:"altup_active_idx"`
-		AltupCoefClip             float32   `json:"altup_coef_clip"`
-		AltupCorrectScale         bool      `json:"altup_correct_scale"`
-		AltupLRMultiplier         float32   `json:"altup_lr_multiplier"`
-		AltupNumInputs            uint32    `json:"altup_num_inputs"`
-		HeadDim                   uint32    `json:"head_dim"`
-		HiddenSize                uint32    `json:"hidden_size"`
-		HiddenSizePerLayerInput   uint32    `json:"hidden_size_per_layer_input"`
-		IntermediateSize          uint32    `json:"intermediate_size"`
-		MaxPositionEmbeddings     uint32    `json:"max_position_embeddings"`
-		NumAttentionHeads         uint32    `json:"num_attention_heads"`
-		NumHiddenLayers           uint32    `json:"num_hidden_layers"`
-		NumKeyValueHeads          uint32    `json:"num_key_value_heads"`
-		NumKVSharedLayers         uint32    `json:"num_kv_shared_layers"`
-		RMSNormEPS                float32   `json:"rms_norm_eps"`
-		RopeLocalBaseFreq         float32   `json:"rope_local_base_freq"`
-		RopeTheta                 float32   `json:"rope_theta"`
-		SlidingWindow             uint32    `json:"sliding_window"`
-		LayerTypes                []string  `json:"layer_types"`
+		ActivationSparsityPattern []float32               `json:"activation_sparsity_pattern"`
+		AltupActiveIdx            uint32                  `json:"altup_active_idx"`
+		AltupCoefClip             float32                 `json:"altup_coef_clip"`
+		AltupCorrectScale         bool                    `json:"altup_correct_scale"`
+		AltupLRMultiplier         float32                 `json:"altup_lr_multiplier"`
+		AltupNumInputs            uint32                  `json:"altup_num_inputs"`
+		HeadDim                   uint32                  `json:"head_dim"`
+		HiddenSize                uint32                  `json:"hidden_size"`
+		HiddenSizePerLayerInput   uint32                  `json:"hidden_size_per_layer_input"`
+		IntermediateSize          gemma3nIntermediateSize `json:"intermediate_size"`
+		MaxPositionEmbeddings     uint32                  `json:"max_position_embeddings"`
+		NumAttentionHeads         uint32                  `json:"num_attention_heads"`
+		NumHiddenLayers           uint32                  `json:"num_hidden_layers"`
+		NumKeyValueHeads          uint32                  `json:"num_key_value_heads"`
+		NumKVSharedLayers         uint32                  `json:"num_kv_shared_layers"`
+		RMSNormEPS                float32                 `json:"rms_norm_eps"`
+		RopeLocalBaseFreq         float32                 `json:"rope_local_base_freq"`
+		RopeTheta                 float32                 `json:"rope_theta"`
+		SlidingWindow             uint32                  `json:"sliding_window"`
+		LayerTypes                []string                `json:"layer_types"`
 	} `json:"text_config"`
 	VisionModel struct{} `json:"vision_config"`
 }
 
+type gemma3nIntermediateSize uint32
+
+func (s *gemma3nIntermediateSize) UnmarshalJSON(data []byte) error {
+	var scalar uint32
+	if err := json.Unmarshal(data, &scalar); err == nil {
+		*s = gemma3nIntermediateSize(scalar)
+		return nil
+	}
+
+	var values []uint32
+	if err := json.Unmarshal(data, &values); err != nil {
+		return err
+	}
+	if len(values) == 0 {
+		return fmt.Errorf("intermediate_size must not be empty")
+	}
+
+	first := values[0]
+	for _, v := range values[1:] {
+		if v != first {
+			return fmt.Errorf("intermediate_size values must match")
+		}
+	}
+
+	*s = gemma3nIntermediateSize(first)
+	return nil
+}
+
 func (m *gemma3nModel) KV(t *Tokenizer) KV {
 	kv := m.ModelParameters.KV(t)
 	kv["general.architecture"] = "gemma3n"
@@ -69,7 +99,7 @@ func (m *gemma3nModel) KV(t *Tokenizer) KV {
 	kv["gemma3n.context_length"] = m.TextModel.MaxPositionEmbeddings
 	kv["gemma3n.embedding_length_per_layer_input"] = m.TextModel.HiddenSizePerLayerInput
 	kv["gemma3n.embedding_length"] = m.TextModel.HiddenSize
-	kv["gemma3n.feed_forward_length"] = m.TextModel.IntermediateSize
+	kv["gemma3n.feed_forward_length"] = uint32(m.TextModel.IntermediateSize)
 	kv["gemma3n.head_dim"] = m.TextModel.HeadDim
 	kv["gemma3n.rope.freq_base_local"] = m.TextModel.RopeLocalBaseFreq
 	kv["gemma3n.rope.freq_base"] = m.TextModel.RopeTheta

convert/convert_gemma3n_test.go

@@ -0,0 +1,55 @@
+package convert
+
+import (
+	"encoding/json"
+	"testing"
+)
+
+func TestGemma3nIntermediateSize(t *testing.T) {
+	tests := []struct {
+		name    string
+		json    string
+		want    gemma3nIntermediateSize
+		wantErr bool
+	}{
+		{
+			name: "scalar",
+			json: `8192`,
+			want: 8192,
+		},
+		{
+			name: "uniform array",
+			json: `[8192,8192,8192]`,
+			want: 8192,
+		},
+		{
+			name:    "mixed array",
+			json:    `[8192,4096]`,
+			wantErr: true,
+		},
+		{
+			name:    "empty array",
+			json:    `[]`,
+			wantErr: true,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			var got gemma3nIntermediateSize
+			err := json.Unmarshal([]byte(tt.json), &got)
+			if tt.wantErr {
+				if err == nil {
+					t.Fatal("expected error")
+				}
+				return
+			}
+			if err != nil {
+				t.Fatal(err)
+			}
+			if got != tt.want {
+				t.Fatalf("got %d, want %d", got, tt.want)
+			}
+		})
+	}
+}

convert/convert_glm4moelite.go

@@ -39,48 +39,72 @@ type glm4MoeLiteModel struct {
 	ExpertWeightsScale     float32 `json:"routed_scaling_factor"`
 
 	LeadingDenseBlockCount uint32 `json:"first_k_dense_replace"`
+
+	ExpertGroupCount     uint32 `json:"n_group"`
+	ExpertGroupUsedCount uint32 `json:"topk_group"`
 }
 
 func (p *glm4MoeLiteModel) KV(t *Tokenizer) KV {
 	kv := p.ModelParameters.KV(t)
-	kv["general.architecture"] = "glm4moelite"
+	kv["general.architecture"] = "deepseek2"
 	kv["general.type"] = "model"
-	kv["glm4moelite.block_count"] = p.HiddenLayers
+	kv["deepseek2.block_count"] = p.HiddenLayers
 
 	numHeads := p.NumAttentionHeads
-	numKVHeads := p.NumKeyValueHeads
 
-	kv["glm4moelite.attention.head_count"] = numHeads
-	kv["glm4moelite.attention.head_count_kv"] = numKVHeads
-	kv["glm4moelite.attention.key_length"] = p.QKNopeHeadDim + p.QKRopeHeadDim
-	kv["glm4moelite.attention.kv_lora_rank"] = p.KVLoraRank
-	kv["glm4moelite.attention.layer_norm_rms_epsilon"] = p.RMSNormEPS
-	kv["glm4moelite.attention.q_lora_rank"] = p.QLoraRank
-	kv["glm4moelite.attention.value_length"] = p.VHeadDim
-	kv["glm4moelite.context_length"] = p.MaxPositionEmbeddings
-	kv["glm4moelite.embedding_length"] = p.HiddenSize
-	kv["glm4moelite.expert_count"] = p.ExpertCount
-	kv["glm4moelite.expert_feed_forward_length"] = p.ExpertIntermediateSize
-	kv["glm4moelite.expert_shared_count"] = p.ExpertSharedCount
+	kv["deepseek2.attention.head_count"] = numHeads
+	kv["deepseek2.attention.head_count_kv"] = uint32(1)
+	kv["deepseek2.attention.key_length"] = p.KVLoraRank + p.QKRopeHeadDim
+	kv["deepseek2.attention.kv_lora_rank"] = p.KVLoraRank
+	kv["deepseek2.attention.layer_norm_rms_epsilon"] = p.RMSNormEPS
+	kv["deepseek2.attention.q_lora_rank"] = p.QLoraRank
+	kv["deepseek2.attention.value_length"] = p.KVLoraRank
+	kv["deepseek2.context_length"] = p.MaxPositionEmbeddings
+	kv["deepseek2.embedding_length"] = p.HiddenSize
+	kv["deepseek2.expert_count"] = p.ExpertCount
+	kv["deepseek2.expert_feed_forward_length"] = p.ExpertIntermediateSize
+	kv["deepseek2.expert_shared_count"] = p.ExpertSharedCount
 
-	kv["glm4moelite.expert_gating_func"] = uint32(2)
-	kv["glm4moelite.expert_used_count"] = p.ExpertUsedCount
-	kv["glm4moelite.expert_weights_norm"] = p.ExpertWeightsNorm
-	kv["glm4moelite.expert_weights_scale"] = p.ExpertWeightsScale
-	kv["glm4moelite.feed_forward_length"] = p.IntermediateSize
-	kv["glm4moelite.leading_dense_block_count"] = p.LeadingDenseBlockCount
+	kv["deepseek2.expert_gating_func"] = uint32(2)
+	kv["deepseek2.expert_group_count"] = cmp.Or(p.ExpertGroupCount, uint32(1))
+	kv["deepseek2.expert_group_used_count"] = cmp.Or(p.ExpertGroupUsedCount, uint32(1))
+	kv["deepseek2.expert_used_count"] = p.ExpertUsedCount
+	kv["deepseek2.expert_weights_norm"] = p.ExpertWeightsNorm
+	kv["deepseek2.expert_weights_scale"] = p.ExpertWeightsScale
+	kv["deepseek2.feed_forward_length"] = p.IntermediateSize
+	kv["deepseek2.leading_dense_block_count"] = p.LeadingDenseBlockCount
 
-	kv["glm4moelite.rope.dimension_count"] = p.QKRopeHeadDim
-	kv["glm4moelite.rope.freq_base"] = cmp.Or(p.RopeTheta, float32(1000000.0))
+	kv["deepseek2.rope.dimension_count"] = p.QKRopeHeadDim
+	kv["deepseek2.rope.freq_base"] = cmp.Or(p.RopeTheta, float32(1000000.0))
 
-	kv["glm4moelite.attention.key_length_mla"] = p.KVLoraRank + p.QKRopeHeadDim
-	kv["glm4moelite.attention.value_length_mla"] = p.KVLoraRank
+	kv["deepseek2.attention.key_length_mla"] = p.QKNopeHeadDim + p.QKRopeHeadDim
+	kv["deepseek2.attention.value_length_mla"] = p.VHeadDim
 
 	kv["tokenizer.ggml.pre"] = "glm4"
+	setGLM4MoeLiteExtraEOGFromEOSIDs(kv)
 
 	return kv
 }
 
+func setGLM4MoeLiteExtraEOGFromEOSIDs(kv KV) {
+	switch ids := kv["tokenizer.ggml.eos_token_ids"].(type) {
+	case []int32:
+		if len(ids) >= 2 && ids[1] >= 0 {
+			kv["tokenizer.ggml.eot_token_id"] = uint32(ids[1])
+		}
+		if len(ids) >= 3 && ids[2] >= 0 {
+			kv["tokenizer.ggml.eom_token_id"] = uint32(ids[2])
+		}
+	case []uint32:
+		if len(ids) >= 2 {
+			kv["tokenizer.ggml.eot_token_id"] = ids[1]
+		}
+		if len(ids) >= 3 {
+			kv["tokenizer.ggml.eom_token_id"] = ids[2]
+		}
+	}
+}
+
 func (p *glm4MoeLiteModel) Replacements() []string {
 	return []string{
 		"lm_head", "output",

convert/convert_glm4moelite_test.go

@@ -0,0 +1,68 @@
+package convert
+
+import "testing"
+
+func TestGLM4MoeLiteKVUsesLlamaCppMetadata(t *testing.T) {
+	p := glm4MoeLiteModel{
+		ModelParameters:       ModelParameters{VocabSize: 151552},
+		MaxPositionEmbeddings: 202752,
+		HiddenSize:            2048,
+		HiddenLayers:          47,
+		IntermediateSize:      10240,
+		NumAttentionHeads:     20,
+		NumKeyValueHeads:      20,
+		RMSNormEPS:            1e-5,
+		RopeTheta:             1000000,
+		QKNopeHeadDim:         128,
+		QKRopeHeadDim:         64,
+		KVLoraRank:            512,
+		QLoraRank:             768,
+		VHeadDim:              128,
+		ExpertCount:           64,
+		ExpertSharedCount:     1,
+		ExpertUsedCount:       4,
+		ExpertWeightsNorm:     true,
+		ExpertWeightsScale:    1.8,
+	}
+
+	kv := p.KV(&Tokenizer{Vocabulary: &Vocabulary{Model: "gpt2", Tokens: []string{"a"}}})
+
+	if got := kv.Architecture(); got != "deepseek2" {
+		t.Fatalf("architecture = %q, want deepseek2", got)
+	}
+	for key, want := range map[string]uint32{
+		"attention.head_count":       20,
+		"attention.head_count_kv":    1,
+		"attention.key_length":       576,
+		"attention.value_length":     512,
+		"attention.key_length_mla":   192,
+		"attention.value_length_mla": 128,
+		"expert_group_count":         1,
+		"expert_group_used_count":    1,
+		"expert_gating_func":         2,
+		"rope.dimension_count":       64,
+	} {
+		if got := kv.Uint(key); got != want {
+			t.Errorf("%s = %d, want %d", key, got, want)
+		}
+	}
+	if got := kv.String("tokenizer.ggml.pre"); got != "glm4" {
+		t.Errorf("tokenizer.ggml.pre = %q, want glm4", got)
+	}
+}
+
+func TestGLM4MoeLiteKVPromotesExtraEOSIDs(t *testing.T) {
+	kv := KV{
+		"general.architecture":         "deepseek2",
+		"tokenizer.ggml.eos_token_ids": []int32{151329, 151330, 151336},
+	}
+
+	setGLM4MoeLiteExtraEOGFromEOSIDs(kv)
+
+	if got := kv.Uint("tokenizer.ggml.eot_token_id"); got != 151330 {
+		t.Errorf("eot token = %d, want 151330", got)
+	}
+	if got := kv.Uint("tokenizer.ggml.eom_token_id"); got != 151336 {
+		t.Errorf("eom token = %d, want 151336", got)
+	}
+}

convert/convert_glmocr.go

@@ -83,6 +83,7 @@ type glmOcrModel struct {
 		HiddenSize          uint32  `json:"hidden_size"`
 		IntermediateSize    uint32  `json:"intermediate_size"`
 		NumHiddenLayers     uint32  `json:"num_hidden_layers"`
+		NumNextNPredict     uint32  `json:"num_nextn_predict_layers"`
 		NumAttentionHeads   uint32  `json:"num_attention_heads"`
 		NumKeyValueHeads    uint32  `json:"num_key_value_heads"`
 		HeadDim             uint32  `json:"head_dim"`
@@ -131,7 +132,7 @@ type glmOcrModel struct {
 	} `json:"-"`
 }
 
-var _ ModelConverter = (*glmOcrModel)(nil)
+var _ MultimodalConverter = (*glmOcrModel)(nil)
 
 func (m *glmOcrModel) parseMore(fsys fs.FS) error {
 	bts, err := fs.ReadFile(fsys, "preprocessor_config.json")
@@ -145,9 +146,14 @@ func (m *glmOcrModel) parseMore(fsys fs.FS) error {
 func (m *glmOcrModel) KV(t *Tokenizer) KV {
 	kv := m.ModelParameters.KV(t)
 	kv["general.architecture"] = "glmocr"
+	applyGlmOcrTokenizerKV(kv, t)
 
 	// Text model parameters
-	kv["glmocr.block_count"] = cmp.Or(m.TextConfig.NumHiddenLayers, 16)
+	numHiddenLayers := cmp.Or(m.TextConfig.NumHiddenLayers, 16)
+	kv["glmocr.block_count"] = numHiddenLayers + m.TextConfig.NumNextNPredict
+	if m.TextConfig.NumNextNPredict > 0 {
+		kv["glmocr.nextn_predict_layers"] = m.TextConfig.NumNextNPredict
+	}
 	kv["glmocr.embedding_length"] = cmp.Or(m.TextConfig.HiddenSize, 1536)
 	kv["glmocr.attention.head_count"] = cmp.Or(m.TextConfig.NumAttentionHeads, 16)
 	kv["glmocr.attention.head_count_kv"] = cmp.Or(m.TextConfig.NumKeyValueHeads, 8)
@@ -175,8 +181,6 @@ func (m *glmOcrModel) KV(t *Tokenizer) KV {
 	kv["glmocr.vision.intermediate_size"] = cmp.Or(m.VisionConfig.IntermediateSize, 4096)
 	kv["glmocr.vision.attention.layer_norm_rms_epsilon"] = cmp.Or(m.VisionConfig.RMSNormEps, 1e-5)
 
-	// Preprocessor-derived image settings (min/max pixels and normalization)
-	// Note: fs.Config.keyValue() auto-prepends architecture prefix, so use full key
 	if m.Preprocessor.Size.ShortestEdge > 0 {
 		kv["glmocr.vision.min_pixels"] = m.Preprocessor.Size.ShortestEdge
 	}
@@ -190,7 +194,6 @@ func (m *glmOcrModel) KV(t *Tokenizer) KV {
 		kv["glmocr.vision.image_std"] = m.Preprocessor.ImageStd
 	}
 
-	// Special tokens
 	kv["glmocr.image_token_id"] = m.ImageTokenID
 	kv["glmocr.image_start_token_id"] = m.ImageStartTokenID
 	kv["glmocr.image_end_token_id"] = m.ImageEndTokenID
@@ -201,32 +204,249 @@ func (m *glmOcrModel) KV(t *Tokenizer) KV {
 	return kv
 }
 
+func applyGlmOcrTokenizerKV(kv KV, t *Tokenizer) {
+	kv["tokenizer.ggml.pre"] = "chatglm-bpe"
+	if id, ok := glmOcrTokenID(t, "<|endoftext|>"); ok {
+		kv["tokenizer.ggml.bos_token_id"] = uint32(id)
+		kv["tokenizer.ggml.unknown_token_id"] = uint32(id)
+	}
+	if id, ok := glmOcrTokenID(t, "<|user|>"); ok {
+		kv["tokenizer.ggml.eot_token_id"] = uint32(id)
+	}
+}
+
+func (m *glmOcrModel) TextKV(t *Tokenizer) KV {
+	kv := m.ModelParameters.KV(t)
+	kv["general.architecture"] = "glm4"
+	applyGlmOcrTokenizerKV(kv, t)
+
+	numHiddenLayers := cmp.Or(m.TextConfig.NumHiddenLayers, 16)
+	kv["block_count"] = numHiddenLayers + m.TextConfig.NumNextNPredict
+	if m.TextConfig.NumNextNPredict > 0 {
+		kv["nextn_predict_layers"] = m.TextConfig.NumNextNPredict
+	}
+	kv["embedding_length"] = cmp.Or(m.TextConfig.HiddenSize, 1536)
+	kv["attention.head_count"] = cmp.Or(m.TextConfig.NumAttentionHeads, 16)
+	kv["attention.head_count_kv"] = cmp.Or(m.TextConfig.NumKeyValueHeads, 8)
+	headDim := cmp.Or(m.TextConfig.HeadDim, m.TextConfig.HiddenSize/m.TextConfig.NumAttentionHeads)
+	kv["attention.key_length"] = headDim
+	kv["attention.value_length"] = headDim
+	kv["feed_forward_length"] = cmp.Or(m.TextConfig.IntermediateSize, 4608)
+	kv["attention.layer_norm_rms_epsilon"] = cmp.Or(m.TextConfig.RMSNormEps, 1e-5)
+	kv["context_length"] = cmp.Or(m.TextConfig.MaxPositionEmbed, 131072)
+	kv["rope.freq_base"] = cmp.Or(m.TextConfig.RopeParameters.RopeTheta, float32(10000))
+	partialRotaryFactor := cmp.Or(m.TextConfig.RopeParameters.PartialRotaryFactor, m.TextConfig.PartialRotaryFactor, float32(1.0))
+	kv["rope.dimension_count"] = uint32(float32(headDim) * partialRotaryFactor)
+	if len(m.TextConfig.RopeParameters.MRopeSection) > 0 {
+		sections := append([]int32(nil), m.TextConfig.RopeParameters.MRopeSection...)
+		for len(sections) < 4 {
+			sections = append(sections, 0)
+		}
+		kv["rope.dimension_sections"] = sections
+	}
+
+	return kv
+}
+
+func (m *glmOcrModel) ProjectorKV(*Tokenizer) KV {
+	kv := KV{
+		"general.architecture":                     "clip",
+		"general.type":                             "mmproj",
+		"general.file_type":                        uint32(1),
+		"general.quantization_version":             uint32(2),
+		"clip.has_vision_encoder":                  true,
+		"clip.projector_type":                      "glm4v",
+		"clip.use_silu":                            true,
+		"clip.vision.block_count":                  cmp.Or(m.VisionConfig.Depth, 24),
+		"clip.vision.embedding_length":             cmp.Or(m.VisionConfig.HiddenSize, 1024),
+		"clip.vision.attention.head_count":         cmp.Or(m.VisionConfig.NumHeads, 16),
+		"clip.vision.image_size":                   cmp.Or(m.VisionConfig.ImageSize, 336),
+		"clip.vision.patch_size":                   cmp.Or(m.VisionConfig.PatchSize, m.Preprocessor.PatchSize, 14),
+		"clip.vision.spatial_merge_size":           cmp.Or(m.VisionConfig.SpatialMergeSize, m.Preprocessor.MergeSize, 2),
+		"clip.vision.temporal_patch_size":          cmp.Or(m.VisionConfig.TemporalPatchSize, m.Preprocessor.TemporalPatchSize, 2),
+		"clip.vision.projection_dim":               cmp.Or(m.VisionConfig.OutHiddenSize, 1536),
+		"clip.vision.out_hidden_size":              cmp.Or(m.VisionConfig.OutHiddenSize, 1536),
+		"clip.vision.feed_forward_length":          cmp.Or(m.VisionConfig.IntermediateSize, 4096),
+		"clip.vision.intermediate_size":            cmp.Or(m.VisionConfig.IntermediateSize, 4096),
+		"clip.vision.attention.layer_norm_epsilon": cmp.Or(m.VisionConfig.RMSNormEps, 1e-5),
+		"clip.vision.image_token_id":               m.ImageTokenID,
+		"clip.vision.image_start_token_id":         m.ImageStartTokenID,
+		"clip.vision.image_end_token_id":           m.ImageEndTokenID,
+	}
+	if m.Preprocessor.Size.ShortestEdge > 0 {
+		kv["clip.vision.min_pixels"] = m.Preprocessor.Size.ShortestEdge
+	}
+	if m.Preprocessor.Size.LongestEdge > 0 {
+		kv["clip.vision.max_pixels"] = m.Preprocessor.Size.LongestEdge
+	}
+	if len(m.Preprocessor.ImageMean) == 3 {
+		kv["clip.vision.image_mean"] = m.Preprocessor.ImageMean
+	}
+	if len(m.Preprocessor.ImageStd) == 3 {
+		kv["clip.vision.image_std"] = m.Preprocessor.ImageStd
+	}
+
+	return kv
+}
+
+func glmOcrTokenID(t *Tokenizer, token string) (int, bool) {
+	if t == nil || t.Vocabulary == nil {
+		return 0, false
+	}
+	for i, candidate := range t.Vocabulary.Tokens {
+		if candidate == token {
+			return i, true
+		}
+	}
+	return 0, false
+}
+
+func isGlmOcrVisionTensor(name string) bool {
+	return strings.HasPrefix(name, "v.") || strings.HasPrefix(name, "mm.")
+}
+
+func (m *glmOcrModel) TextTensors(ts []Tensor, t *Tokenizer) []*ggml.Tensor {
+	textOnly := make([]Tensor, 0, len(ts))
+	for _, tensor := range ts {
+		if !isGlmOcrVisionTensor(tensor.Name()) {
+			textOnly = append(textOnly, tensor)
+		}
+	}
+	return m.Tensors(textOnly)
+}
+
+func (m *glmOcrModel) ProjectorTensors(ts []Tensor) []*ggml.Tensor {
+	var out []*ggml.Tensor
+	for _, t := range ts {
+		if !isGlmOcrVisionTensor(t.Name()) {
+			continue
+		}
+
+		name := t.Name()
+		switch {
+		case strings.HasSuffix(name, "patch_embd_0.weight"):
+			name = strings.Replace(name, "patch_embd_0.weight", "patch_embd.weight", 1)
+		case strings.HasSuffix(name, "patch_embd_1.weight"):
+			name = strings.Replace(name, "patch_embd_1.weight", "patch_embd.weight.1", 1)
+		case strings.HasSuffix(name, "patch_embd.weight.0"):
+			name = strings.Replace(name, "patch_embd.weight.0", "patch_embd.weight", 1)
+		}
+		if strings.HasSuffix(name, "patch_embd.weight") {
+			shape := t.Shape()
+			if len(shape) == 5 && shape[2] == 2 {
+				newShape := []uint64{shape[0], shape[1], shape[3], shape[4]}
+
+				t0 := t.Clone()
+				t0.SetRepacker(func(_ string, data []float32, shape []uint64) ([]float32, error) {
+					dims := make([]int, len(shape))
+					for i := range shape {
+						dims[i] = int(shape[i])
+					}
+					var tt tensor.Tensor = tensor.New(tensor.WithShape(dims...), tensor.WithBacking(data))
+					tt, err := tt.Slice(nil, nil, tensor.S(0, 1), nil, nil)
+					if err != nil {
+						return nil, err
+					}
+					tt = tensor.Materialize(tt)
+					newDims := []int{int(shape[0]), int(shape[1]), int(shape[3]), int(shape[4])}
+					if err := tt.Reshape(newDims...); err != nil {
+						return nil, err
+					}
+					if err := tt.Reshape(tt.Shape().TotalSize()); err != nil {
+						return nil, err
+					}
+					return native.VectorF32(tt.(*tensor.Dense))
+				})
+				out = append(out, &ggml.Tensor{
+					Name:     strings.Replace(name, "patch_embd.weight", "patch_embd.weight", 1),
+					Kind:     t.Kind(),
+					Shape:    newShape,
+					WriterTo: t0,
+				})
+
+				t1 := t.Clone()
+				t1.SetRepacker(func(_ string, data []float32, shape []uint64) ([]float32, error) {
+					dims := make([]int, len(shape))
+					for i := range shape {
+						dims[i] = int(shape[i])
+					}
+					var tt tensor.Tensor = tensor.New(tensor.WithShape(dims...), tensor.WithBacking(data))
+					tt, err := tt.Slice(nil, nil, tensor.S(1, 2), nil, nil)
+					if err != nil {
+						return nil, err
+					}
+					tt = tensor.Materialize(tt)
+					newDims := []int{int(shape[0]), int(shape[1]), int(shape[3]), int(shape[4])}
+					if err := tt.Reshape(newDims...); err != nil {
+						return nil, err
+					}
+					if err := tt.Reshape(tt.Shape().TotalSize()); err != nil {
+						return nil, err
+					}
+					return native.VectorF32(tt.(*tensor.Dense))
+				})
+				out = append(out, &ggml.Tensor{
+					Name:     strings.Replace(name, "patch_embd.weight", "patch_embd.weight.1", 1),
+					Kind:     t.Kind(),
+					Shape:    newShape,
+					WriterTo: t1,
+				})
+
+				continue
+			}
+		}
+
+		out = append(out, &ggml.Tensor{
+			Name:     name,
+			Kind:     t.Kind(),
+			Shape:    t.Shape(),
+			WriterTo: t,
+		})
+	}
+	return out
+}
+
 func (m *glmOcrModel) Tensors(ts []Tensor) []*ggml.Tensor {
 	var out []*ggml.Tensor
 
-	// Skip layers >= num_hidden_layers (Multi-Token Prediction layers not needed for basic inference)
 	numLayers := int(cmp.Or(m.TextConfig.NumHiddenLayers, 16))
-	skipLayer := func(name string) bool {
-		// Tensor names are already replaced to "blk.N.xxx" format
-		re := regexp.MustCompile(`^blk\.(\d+)`)
-		matches := re.FindStringSubmatch(name)
+	maxLayers := numLayers + int(m.TextConfig.NumNextNPredict)
+	layerRe := regexp.MustCompile(`^blk\.(\d+)`)
+	layerIndex := func(name string) (int, bool) {
+		matches := layerRe.FindStringSubmatch(name)
 		if matches == nil {
-			return false
+			return 0, false
 		}
 		blkNum, err := strconv.Atoi(matches[1])
 		if err != nil {
-			return false
+			return 0, false
 		}
-		return blkNum >= numLayers
+		return blkNum, true
 	}
 
 	for _, t := range ts {
 		name := t.Name()
 
-		// Skip next-n prediction layers (layers >= num_hidden_layers)
-		if skipLayer(name) {
+		blkNum, hasLayer := layerIndex(name)
+		if hasLayer && blkNum >= maxLayers {
 			continue
 		}
+		if hasLayer && blkNum >= numLayers {
+			switch {
+			case strings.HasSuffix(name, ".embed_tokens.weight"):
+				name = strings.Replace(name, ".embed_tokens.weight", ".nextn.embed_tokens.weight", 1)
+			case strings.HasSuffix(name, ".eh_proj.weight"):
+				name = strings.Replace(name, ".eh_proj.weight", ".nextn.eh_proj.weight", 1)
+			case strings.HasSuffix(name, ".enorm.weight"):
+				name = strings.Replace(name, ".enorm.weight", ".nextn.enorm.weight", 1)
+			case strings.HasSuffix(name, ".hnorm.weight"):
+				name = strings.Replace(name, ".hnorm.weight", ".nextn.hnorm.weight", 1)
+			case strings.HasSuffix(name, ".shared_head.head.weight"):
+				name = strings.Replace(name, ".shared_head.head.weight", ".nextn.shared_head_head.weight", 1)
+			case strings.HasSuffix(name, ".shared_head.norm.weight"):
+				name = strings.Replace(name, ".shared_head.norm.weight", ".nextn.shared_head_norm.weight", 1)
+			}
+		}
 
 		// Split ffn_gate_up into separate gate and up projections
 		if strings.Contains(name, "ffn_gate_up") {
@@ -440,16 +660,16 @@ func (m *glmOcrModel) Replacements() []string {
 		"self_attn.q_proj", "attn_q",
 		"self_attn.k_proj", "attn_k",
 		"self_attn.v_proj", "attn_v",
-		"self_attn.o_proj", "attn_out",
+		"self_attn.o_proj", "attn_output",
 
 		// Language model norms
 		"input_layernorm", "attn_norm",
 		"post_attention_layernorm", "ffn_norm",
-		"post_self_attn_layernorm", "post_attn_norm",
-		"post_mlp_layernorm", "post_ffn_norm",
+		"post_self_attn_layernorm", "post_attention_norm",
+		"post_mlp_layernorm", "post_ffw_norm",
 
-		// Language model MLP (remove mlp. prefix so ffn_* names work)
-		"mlp.gate_up_proj", "ffn_gate_up",
+		// Language model MLP
+		"mlp.gate_up_proj", "ffn_up",
 		"mlp.down_proj", "ffn_down",
 	}
 }

convert/convert_gptoss.go

@@ -30,7 +30,11 @@ type gptossModel struct {
 	RopeTheta             float32 `json:"rope_theta"`
 	RopeScalingFactor     float32 `json:"rope_scaling_factor"`
 	RopeScaling           struct {
-		Factor float32 `json:"factor"`
+		Type                          string  `json:"rope_type"`
+		Factor                        float32 `json:"factor"`
+		OriginalMaxPositionEmbeddings uint32  `json:"original_max_position_embeddings"`
+		BetaFast                      float32 `json:"beta_fast"`
+		BetaSlow                      float32 `json:"beta_slow"`
 	} `json:"rope_scaling"`
 	SlidingWindow uint32 `json:"sliding_window"`
 }
@@ -39,23 +43,32 @@ var _ ModelConverter = (*gptossModel)(nil)
 
 func (m *gptossModel) KV(t *Tokenizer) KV {
 	kv := m.ModelParameters.KV(t)
-	kv["general.architecture"] = "gptoss"
+	kv["general.architecture"] = "gpt-oss"
 	kv["general.file_type"] = uint32(4)
-	kv["gptoss.context_length"] = cmp.Or(m.MaxPositionEmbeddings, uint32(m.RopeScalingFactor*float32(m.InitialContextLength)))
-	kv["gptoss.block_count"] = m.HiddenLayers
-	kv["gptoss.embedding_length"] = m.HiddenSize
-	kv["gptoss.feed_forward_length"] = m.IntermediateSize
-	kv["gptoss.expert_count"] = cmp.Or(m.Experts, m.LocalExperts)
-	kv["gptoss.expert_used_count"] = m.ExpertsPerToken
-	kv["gptoss.attention.head_count"] = m.AttentionHeads
-	kv["gptoss.attention.head_count_kv"] = m.KeyValueHeads
-	kv["gptoss.attention.key_length"] = m.HeadDim
-	kv["gptoss.attention.value_length"] = m.HeadDim
-	kv["gptoss.attention.layer_norm_rms_epsilon"] = cmp.Or(m.RMSNormEpsilon, 1e-5)
-	kv["gptoss.attention.sliding_window"] = m.SlidingWindow
-	kv["gptoss.rope.freq_base"] = m.RopeTheta
-	kv["gptoss.rope.scaling.factor"] = cmp.Or(m.RopeScalingFactor, m.RopeScaling.Factor)
-	kv["gptoss.rope.scaling.original_context_length"] = m.InitialContextLength
+	kv["gpt-oss.context_length"] = cmp.Or(m.MaxPositionEmbeddings, uint32(m.RopeScalingFactor*float32(m.InitialContextLength)))
+	kv["gpt-oss.block_count"] = m.HiddenLayers
+	kv["gpt-oss.embedding_length"] = m.HiddenSize
+	kv["gpt-oss.feed_forward_length"] = m.IntermediateSize
+	kv["gpt-oss.expert_feed_forward_length"] = m.IntermediateSize
+	kv["gpt-oss.expert_count"] = cmp.Or(m.Experts, m.LocalExperts)
+	kv["gpt-oss.expert_used_count"] = m.ExpertsPerToken
+	kv["gpt-oss.attention.head_count"] = m.AttentionHeads
+	kv["gpt-oss.attention.head_count_kv"] = m.KeyValueHeads
+	kv["gpt-oss.attention.key_length"] = m.HeadDim
+	kv["gpt-oss.attention.value_length"] = m.HeadDim
+	kv["gpt-oss.attention.layer_norm_rms_epsilon"] = cmp.Or(m.RMSNormEpsilon, 1e-5)
+	kv["gpt-oss.attention.sliding_window"] = m.SlidingWindow
+	kv["gpt-oss.rope.freq_base"] = m.RopeTheta
+	kv["gpt-oss.rope.scaling.type"] = cmp.Or(m.RopeScaling.Type, "yarn")
+	kv["gpt-oss.rope.scaling.factor"] = cmp.Or(m.RopeScalingFactor, m.RopeScaling.Factor)
+	kv["gpt-oss.rope.scaling.original_context_length"] = cmp.Or(m.RopeScaling.OriginalMaxPositionEmbeddings, m.InitialContextLength)
+	if m.RopeScaling.BetaFast != 0 {
+		kv["gpt-oss.rope.scaling.yarn_beta_fast"] = m.RopeScaling.BetaFast
+	}
+	if m.RopeScaling.BetaSlow != 0 {
+		kv["gpt-oss.rope.scaling.yarn_beta_slow"] = m.RopeScaling.BetaSlow
+	}
+	kv["tokenizer.ggml.pre"] = "gpt-4o"
 	kv["tokenizer.ggml.bos_token_id"] = uint32(199998) // <|startoftext|>
 	kv["tokenizer.ggml.add_bos_token"] = false
 	kv["tokenizer.ggml.eos_token_id"] = uint32(199999) // <|endoftext|>
@@ -152,9 +165,9 @@ func (m *gptossModel) Replacements() []string {
 			"self_attn.q_proj", "attn_q",
 			"self_attn.k_proj", "attn_k",
 			"self_attn.v_proj", "attn_v",
-			"self_attn.o_proj", "attn_out",
-			"self_attn.sinks", "attn_sinks",
-			"post_attention_layernorm", "ffn_norm",
+			"self_attn.o_proj", "attn_output",
+			"self_attn.sinks", "attn_sinks.weight",
+			"post_attention_layernorm", "post_attention_norm",
 			"mlp.router", "ffn_gate_inp",
 			"mlp.experts.gate_up_proj_", "ffn_gate_up_exps.",
 			"mlp.experts.down_proj_", "ffn_down_exps.",
@@ -169,9 +182,9 @@ func (m *gptossModel) Replacements() []string {
 			"block", "blk",
 			"attn.norm", "attn_norm",
 			"attn.qkv", "attn_qkv",
-			"attn.sinks", "attn_sinks",
-			"attn.out", "attn_out",
-			"mlp.norm", "ffn_norm",
+			"attn.sinks", "attn_sinks.weight",
+			"attn.out", "attn_output",
+			"mlp.norm", "post_attention_norm",
 			"mlp.gate", "ffn_gate_inp",
 			"mlp.mlp1_", "ffn_gate_up_exps.",
 			"mlp.mlp2_", "ffn_down_exps.",

convert/convert_gptoss_test.go

@@ -0,0 +1,73 @@
+package convert
+
+import (
+	"strings"
+	"testing"
+)
+
+func TestGptOssCreatesLlamaCppMetadataAndNames(t *testing.T) {
+	m := &gptossModel{
+		HiddenLayers:          24,
+		MaxPositionEmbeddings: 131072,
+		HiddenSize:            2880,
+		IntermediateSize:      2880,
+		AttentionHeads:        64,
+		KeyValueHeads:         8,
+		HeadDim:               64,
+		LocalExperts:          32,
+		ExpertsPerToken:       4,
+		RopeTheta:             150000,
+		InitialContextLength:  4096,
+		SlidingWindow:         128,
+	}
+	m.RopeScaling.Type = "yarn"
+	m.RopeScaling.Factor = 32
+	m.RopeScaling.OriginalMaxPositionEmbeddings = 4096
+	m.RopeScaling.BetaFast = 32
+	m.RopeScaling.BetaSlow = 1
+
+	kv := m.KV(&Tokenizer{Vocabulary: &Vocabulary{Model: "gpt2"}, Pre: "default"})
+	for k, want := range map[string]any{
+		"general.architecture":                         "gpt-oss",
+		"tokenizer.ggml.pre":                           "gpt-4o",
+		"gpt-oss.context_length":                       uint32(131072),
+		"gpt-oss.expert_feed_forward_length":           uint32(2880),
+		"gpt-oss.rope.scaling.type":                    "yarn",
+		"gpt-oss.rope.scaling.factor":                  float32(32),
+		"gpt-oss.rope.scaling.original_context_length": uint32(4096),
+		"gpt-oss.rope.scaling.yarn_beta_fast":          float32(32),
+		"gpt-oss.rope.scaling.yarn_beta_slow":          float32(1),
+	} {
+		if got := kv[k]; got != want {
+			t.Fatalf("%s = %v (%T), want %v (%T)", k, got, got, want, want)
+		}
+	}
+	if _, ok := kv["gptoss.context_length"]; ok {
+		t.Fatal("unexpected Ollama-format gptoss metadata")
+	}
+
+	replacer := strings.NewReplacer(m.Replacements()...)
+	for name, want := range map[string]string{
+		"model.layers.0.self_attn.o_proj.weight":         "blk.0.attn_output.weight",
+		"model.layers.0.self_attn.sinks":                 "blk.0.attn_sinks.weight",
+		"model.layers.0.post_attention_layernorm.weight": "blk.0.post_attention_norm.weight",
+		"model.layers.0.mlp.experts.gate_up_proj_blocks": "blk.0.ffn_gate_up_exps.blocks",
+		"model.layers.0.mlp.experts.down_proj_scales":    "blk.0.ffn_down_exps.scales",
+	} {
+		if got := replacer.Replace(name); got != want {
+			t.Fatalf("Replace(%q) = %q, want %q", name, got, want)
+		}
+	}
+
+	m.MaxPositionEmbeddings = 0
+	replacer = strings.NewReplacer(m.Replacements()...)
+	for name, want := range map[string]string{
+		"block.0.attn.out.weight": "blk.0.attn_output.weight",
+		"block.0.attn.sinks":      "blk.0.attn_sinks.weight",
+		"block.0.mlp.norm.weight": "blk.0.post_attention_norm.weight",
+	} {
+		if got := replacer.Replace(name); got != want {
+			t.Fatalf("Replace(%q) = %q, want %q", name, got, want)
+		}
+	}
+}

convert/convert_llama.go

@@ -34,8 +34,6 @@ type llamaModel struct {
 		LowFrequencyFactor            float32 `json:"low_freq_factor"`
 		HighFrequencyFactor           float32 `json:"high_freq_factor"`
 		OriginalMaxPositionEmbeddings uint32  `json:"original_max_position_embeddings"`
-
-		factors ropeFactor
 	} `json:"rope_scaling"`
 	RMSNormEPS       float32 `json:"rms_norm_eps"`
 	LayerNormEPS     float32 `json:"layer_norm_eps"`
@@ -83,27 +81,6 @@ func (p *llamaModel) KV(t *Tokenizer) KV {
 	if p.RopeScaling.Type == "linear" {
 		kv["llama.rope.scaling.type"] = p.RopeScaling.Type
 		kv["llama.rope.scaling.factor"] = p.RopeScaling.Factor
-	} else if p.RopeScaling.RopeType == "llama3" {
-		dim := p.HiddenSize / p.NumAttentionHeads
-		for i := uint32(0); i < dim; i += 2 {
-			factor := cmp.Or(p.RopeScaling.Factor, 8.0)
-			factorLow := cmp.Or(p.RopeScaling.LowFrequencyFactor, 1.0)
-			factorHigh := cmp.Or(p.RopeScaling.HighFrequencyFactor, 4.0)
-
-			original := cmp.Or(p.RopeScaling.OriginalMaxPositionEmbeddings, 8192)
-			lambdaLow := float32(original) / factorLow
-			lambdaHigh := float32(original) / factorHigh
-
-			lambda := 2 * math.Pi * math.Pow(float64(p.RopeTheta), float64(i)/float64(dim))
-			if lambda < float64(lambdaHigh) {
-				p.RopeScaling.factors = append(p.RopeScaling.factors, 1.0)
-			} else if lambda > float64(lambdaLow) {
-				p.RopeScaling.factors = append(p.RopeScaling.factors, factor)
-			} else {
-				smooth := (float32(original)/float32(lambda) - factorLow) / (factorHigh - factorLow)
-				p.RopeScaling.factors = append(p.RopeScaling.factors, 1.0/((1-smooth)/factor+smooth))
-			}
-		}
 	}
 
 	if p.NumKeyValueHeads > 0 {
@@ -129,12 +106,12 @@ func (p *llamaModel) KV(t *Tokenizer) KV {
 func (p *llamaModel) Tensors(ts []Tensor) []*ggml.Tensor {
 	var out []*ggml.Tensor
 
-	if p.RopeScaling.factors != nil {
+	if factors := p.ropeFactors(); factors != nil {
 		out = append(out, &ggml.Tensor{
 			Name:     "rope_freqs.weight",
 			Kind:     0,
-			Shape:    []uint64{uint64(len(p.RopeScaling.factors))},
-			WriterTo: p.RopeScaling.factors,
+			Shape:    []uint64{uint64(len(factors))},
+			WriterTo: factors,
 		})
 	}
 
@@ -157,6 +134,40 @@ func (p *llamaModel) Tensors(ts []Tensor) []*ggml.Tensor {
 	return out
 }
 
+func (p *llamaModel) ropeFactors() ropeFactor {
+	if p.RopeScaling.RopeType != "llama3" || p.HiddenSize == 0 || p.NumAttentionHeads == 0 || p.RopeTheta == 0 {
+		return nil
+	}
+
+	dim := p.HiddenSize / p.NumAttentionHeads
+	if dim == 0 {
+		return nil
+	}
+
+	factors := make(ropeFactor, 0, dim/2)
+	for i := uint32(0); i < dim; i += 2 {
+		factor := cmp.Or(p.RopeScaling.Factor, float32(8))
+		factorLow := cmp.Or(p.RopeScaling.LowFrequencyFactor, float32(1))
+		factorHigh := cmp.Or(p.RopeScaling.HighFrequencyFactor, float32(4))
+
+		original := cmp.Or(p.RopeScaling.OriginalMaxPositionEmbeddings, uint32(8192))
+		lambdaLow := float32(original) / factorLow
+		lambdaHigh := float32(original) / factorHigh
+
+		lambda := 2 * math.Pi * math.Pow(float64(p.RopeTheta), float64(i)/float64(dim))
+		if lambda < float64(lambdaHigh) {
+			factors = append(factors, 1)
+		} else if lambda > float64(lambdaLow) {
+			factors = append(factors, factor)
+		} else {
+			smooth := (float32(original)/float32(lambda) - factorLow) / (factorHigh - factorLow)
+			factors = append(factors, 1/((1-smooth)/factor+smooth))
+		}
+	}
+
+	return factors
+}
+
 func (p *llamaModel) Replacements() []string {
 	return []string{
 		"lm_head", "output",

convert/convert_llama_test.go

@@ -0,0 +1,34 @@
+package convert
+
+import "testing"
+
+func TestLlama3RopeFactorsTensorDoesNotDependOnKVOrder(t *testing.T) {
+	m := &llamaModel{
+		HiddenSize:        2048,
+		NumAttentionHeads: 32,
+		RopeTheta:         500000,
+	}
+	m.RopeScaling.RopeType = "llama3"
+	m.RopeScaling.Factor = 32
+	m.RopeScaling.LowFrequencyFactor = 1
+	m.RopeScaling.HighFrequencyFactor = 4
+	m.RopeScaling.OriginalMaxPositionEmbeddings = 8192
+
+	tensors := m.Tensors(nil)
+	if len(tensors) != 1 {
+		t.Fatalf("expected rope tensor only, got %d tensors", len(tensors))
+	}
+	if tensors[0].Name != "rope_freqs.weight" {
+		t.Fatalf("expected rope_freqs.weight, got %q", tensors[0].Name)
+	}
+	if len(tensors[0].Shape) != 1 || tensors[0].Shape[0] != 32 {
+		t.Fatalf("expected rope tensor shape [32], got %v", tensors[0].Shape)
+	}
+
+	_ = m.KV(&Tokenizer{Vocabulary: &Vocabulary{}})
+
+	afterKV := m.Tensors(nil)
+	if len(afterKV) != 1 || afterKV[0].Name != "rope_freqs.weight" {
+		t.Fatalf("expected one rope tensor after KV call, got %#v", afterKV)
+	}
+}

convert/convert_mistral.go

@@ -79,20 +79,17 @@ func (p *mistral3Model) KV(t *Tokenizer) KV {
 	kv["mistral3.rope.freq_base"] = cmp.Or(p.TextModel.RopeTheta, p.TextModel.RopeParameters.RopeTheta)
 	kv["mistral3.rope.scaling.factor"] = p.TextModel.RopeParameters.Factor
 	kv["mistral3.rope.scaling.type"] = p.TextModel.RopeParameters.RopeType
-	kv["mistral3.rope.scaling.beta_fast"] = p.TextModel.RopeParameters.BetaFast
-	kv["mistral3.rope.scaling.beta_slow"] = p.TextModel.RopeParameters.BetaSlow
+	kv["mistral3.rope.scaling.yarn_beta_fast"] = p.TextModel.RopeParameters.BetaFast
+	kv["mistral3.rope.scaling.yarn_beta_slow"] = p.TextModel.RopeParameters.BetaSlow
 
-	if p.TextModel.RopeParameters.Mscale != nil {
-		kv["mistral3.rope.scaling.mscale"] = *p.TextModel.RopeParameters.Mscale
-	}
 	if p.TextModel.RopeParameters.MscaleAllDim != nil {
-		kv["mistral3.rope.scaling.mscale_all_dim"] = *p.TextModel.RopeParameters.MscaleAllDim
+		kv["mistral3.rope.scaling.yarn_log_multiplier"] = *p.TextModel.RopeParameters.MscaleAllDim
 	}
 	if p.TextModel.RopeParameters.OrigMaxPositionEmbeddings > 0 {
 		kv["mistral3.rope.scaling.original_context_length"] = p.TextModel.RopeParameters.OrigMaxPositionEmbeddings
 	}
 	if p.TextModel.RopeParameters.Llama4ScalingBeta != nil {
-		kv["mistral3.rope.scaling_beta"] = *p.TextModel.RopeParameters.Llama4ScalingBeta
+		kv["mistral3.attention.temperature_scale"] = *p.TextModel.RopeParameters.Llama4ScalingBeta
 	}
 
 	// Vision configuration

convert/convert_mistral_causal.go

@@ -58,24 +58,19 @@ func (p *mistral3CausalModel) KV(t *Tokenizer) KV {
 	kv["mistral3.rope.freq_base"] = cmp.Or(p.RopeTheta, p.RopeParameters.RopeTheta)
 	kv["mistral3.rope.scaling.factor"] = p.RopeParameters.Factor
 	kv["mistral3.rope.scaling.type"] = p.RopeParameters.RopeType
-	kv["mistral3.rope.scaling.beta_fast"] = p.RopeParameters.BetaFast
-	kv["mistral3.rope.scaling.beta_slow"] = p.RopeParameters.BetaSlow
-
-	if p.RopeParameters.Mscale != nil {
-		kv["mistral3.rope.scaling.mscale"] = *p.RopeParameters.Mscale
-	}
+	kv["mistral3.rope.scaling.yarn_beta_fast"] = p.RopeParameters.BetaFast
+	kv["mistral3.rope.scaling.yarn_beta_slow"] = p.RopeParameters.BetaSlow
 
 	if p.RopeParameters.MscaleAllDim != nil {
-		kv["mistral3.rope.scaling.mscale_all_dim"] = *p.RopeParameters.MscaleAllDim
+		kv["mistral3.rope.scaling.yarn_log_multiplier"] = *p.RopeParameters.MscaleAllDim
 	}
 
 	if p.RopeParameters.OrigMaxPositionEmbeddings > 0 {
 		kv["mistral3.rope.scaling.original_context_length"] = p.RopeParameters.OrigMaxPositionEmbeddings
-		kv["mistral3.rope.scaling_beta"] = *p.RopeParameters.Llama4ScalingBeta
 	}
 
 	if p.RopeParameters.Llama4ScalingBeta != nil {
-		kv["mistral3.rope.scaling_beta"] = *p.RopeParameters.Llama4ScalingBeta
+		kv["mistral3.attention.temperature_scale"] = *p.RopeParameters.Llama4ScalingBeta
 	}
 
 	return kv

convert/convert_mistral_test.go

@@ -0,0 +1,70 @@
+package convert
+
+import "testing"
+
+func TestMistral3KVUsesLlamaCppRopeScalingKeys(t *testing.T) {
+	mscale := float32(0.75)
+	mscaleAllDim := float32(0)
+	temperatureScale := float32(0.125)
+
+	multimodal := &mistral3Model{}
+	multimodal.TextModel.NumAttentionHeads = 1
+	multimodal.TextModel.HeadDim = 64
+	multimodal.TextModel.RopeParameters.BetaFast = 32
+	multimodal.TextModel.RopeParameters.BetaSlow = 1
+	multimodal.TextModel.RopeParameters.Mscale = &mscale
+	multimodal.TextModel.RopeParameters.MscaleAllDim = &mscaleAllDim
+	multimodal.TextModel.RopeParameters.Llama4ScalingBeta = &temperatureScale
+
+	causal := &mistral3CausalModel{NumAttentionHeads: 1, HeadDim: 64}
+	causal.RopeParameters.BetaFast = 32
+	causal.RopeParameters.BetaSlow = 1
+	causal.RopeParameters.Mscale = &mscale
+	causal.RopeParameters.MscaleAllDim = &mscaleAllDim
+	causal.RopeParameters.Llama4ScalingBeta = &temperatureScale
+
+	tests := []struct {
+		name string
+		kv   KV
+	}{
+		{name: "multimodal", kv: multimodal.KV(mistralTestTokenizer())},
+		{name: "causal", kv: causal.KV(mistralTestTokenizer())},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			assertKVEquals(t, tt.kv, "mistral3.rope.scaling.yarn_beta_fast", float32(32))
+			assertKVEquals(t, tt.kv, "mistral3.rope.scaling.yarn_beta_slow", float32(1))
+			assertKVEquals(t, tt.kv, "mistral3.rope.scaling.yarn_log_multiplier", mscaleAllDim)
+			assertKVEquals(t, tt.kv, "mistral3.attention.temperature_scale", temperatureScale)
+
+			for _, key := range []string{
+				"mistral3.rope.scaling.beta_fast",
+				"mistral3.rope.scaling.beta_slow",
+				"mistral3.rope.scaling.mscale",
+				"mistral3.rope.scaling.mscale_all_dim",
+				"mistral3.rope.scaling_beta",
+			} {
+				if _, ok := tt.kv[key]; ok {
+					t.Fatalf("unexpected legacy key %q", key)
+				}
+			}
+		})
+	}
+}
+
+func mistralTestTokenizer() *Tokenizer {
+	return &Tokenizer{Vocabulary: &Vocabulary{}}
+}
+
+func assertKVEquals[T comparable](t *testing.T, kv KV, key string, want T) {
+	t.Helper()
+
+	got, ok := kv[key]
+	if !ok {
+		t.Fatalf("missing key %q", key)
+	}
+	if got != want {
+		t.Fatalf("%s = %v, want %v", key, got, want)
+	}
+}

convert/convert_nemotron_h.go

@@ -131,8 +131,10 @@ type radioConfig struct {
 	} `json:"args"`
 }
 
-var _ ModelConverter = (*nemotronHModel)(nil)
-var _ ModelConverter = (*nemotronHNanoVLModel)(nil)
+var (
+	_ ModelConverter = (*nemotronHModel)(nil)
+	_ ModelConverter = (*nemotronHNanoVLModel)(nil)
+)
 
 func (n *nemotronHNanoVLModel) parseMore(fsys fs.FS) error {
 	if n.MaxSequenceLength > 0 {

convert/convert_olmo.go

@@ -36,39 +36,39 @@ var _ ModelConverter = (*olmoModel)(nil)
 
 func (p *olmoModel) KV(t *Tokenizer) KV {
 	kv := p.ModelParameters.KV(t)
-	kv["general.architecture"] = "olmo3"
-	kv["olmo3.block_count"] = p.NumHiddenLayers
-	kv["olmo3.context_length"] = p.MaxPositionEmbeddings
-	kv["olmo3.embedding_length"] = p.HiddenSize
-	kv["olmo3.feed_forward_length"] = p.IntermediateSize
-	kv["olmo3.attention.head_count"] = p.NumAttentionHeads
-	kv["olmo3.attention.head_count_kv"] = cmp.Or(p.NumKeyValueHeads, p.NumAttentionHeads)
+	kv["general.architecture"] = "olmo2"
+	kv["olmo2.block_count"] = p.NumHiddenLayers
+	kv["olmo2.context_length"] = p.MaxPositionEmbeddings
+	kv["olmo2.embedding_length"] = p.HiddenSize
+	kv["olmo2.feed_forward_length"] = p.IntermediateSize
+	kv["olmo2.attention.head_count"] = p.NumAttentionHeads
+	kv["olmo2.attention.head_count_kv"] = cmp.Or(p.NumKeyValueHeads, p.NumAttentionHeads)
 
 	if p.RopeTheta > 0 {
-		kv["olmo3.rope.freq_base"] = p.RopeTheta
+		kv["olmo2.rope.freq_base"] = p.RopeTheta
 	}
 
 	if p.RopeScaling != nil {
 		if p.RopeScaling.Factor > 0 {
-			kv["olmo3.rope.scaling.factor"] = p.RopeScaling.Factor
+			kv["olmo2.rope.scaling.factor"] = p.RopeScaling.Factor
 		}
 		if p.RopeScaling.OriginalMaxPositionEmbeds > 0 {
-			kv["olmo3.rope.scaling.original_context_length"] = p.RopeScaling.OriginalMaxPositionEmbeds
+			kv["olmo2.rope.scaling.original_context_length"] = p.RopeScaling.OriginalMaxPositionEmbeds
 		}
 		if p.RopeScaling.AttentionFactor > 0 {
-			kv["olmo3.rope.scaling.attn_factor"] = p.RopeScaling.AttentionFactor
+			kv["olmo2.rope.scaling.attn_factor"] = p.RopeScaling.AttentionFactor
 		}
 		if p.RopeScaling.RopeType != "" {
-			kv["olmo3.rope.scaling.type"] = p.RopeScaling.RopeType
+			kv["olmo2.rope.scaling.type"] = p.RopeScaling.RopeType
 		}
 	}
 
 	if p.RMSNormEPS > 0 {
-		kv["olmo3.attention.layer_norm_rms_epsilon"] = p.RMSNormEPS
+		kv["olmo2.attention.layer_norm_rms_epsilon"] = p.RMSNormEPS
 	}
 
 	if p.SlidingWindow > 0 {
-		kv["olmo3.attention.sliding_window"] = p.SlidingWindow
+		kv["olmo2.attention.sliding_window"] = p.SlidingWindow
 	}
 
 	if len(p.LayerTypes) > 0 {
@@ -76,7 +76,7 @@ func (p *olmoModel) KV(t *Tokenizer) KV {
 		for i, layerType := range p.LayerTypes {
 			slidingPattern[i] = (layerType == "sliding_attention")
 		}
-		kv["olmo3.attention.sliding_window_pattern"] = slidingPattern
+		kv["olmo2.attention.sliding_window_pattern"] = slidingPattern
 	}
 
 	return kv

convert/convert_qwen3next.go

@@ -1,15 +1,24 @@
 package convert
 
 import (
+	"bufio"
+	"bytes"
+	"encoding/binary"
 	"encoding/json"
 	"fmt"
+	"io"
 	"io/fs"
+	"maps"
 	"math"
+	"os"
 	"slices"
+	"strconv"
 	"strings"
 
+	"github.com/d4l3k/go-bfloat16"
 	"github.com/pdevine/tensor"
 	"github.com/pdevine/tensor/native"
+	"github.com/x448/float16"
 
 	"github.com/ollama/ollama/fs/ggml"
 )
@@ -32,6 +41,8 @@ type qwen3NextTextConfig struct {
 	MaxPositionEmbeddings uint32  `json:"max_position_embeddings"`
 	HiddenSize            uint32  `json:"hidden_size"`
 	NumHiddenLayers       uint32  `json:"num_hidden_layers"`
+	NumNextNPredictLayers uint32  `json:"num_nextn_predict_layers"`
+	MTPNumHiddenLayers    uint32  `json:"mtp_num_hidden_layers"`
 	IntermediateSize      uint32  `json:"intermediate_size"`
 	NumAttentionHeads     uint32  `json:"num_attention_heads"`
 	NumKeyValueHeads      uint32  `json:"num_key_value_heads"`
@@ -66,8 +77,11 @@ type qwen3NextTextConfig struct {
 type qwen3NextVisionConfig struct {
 	Depth                  uint32  `json:"depth"`
 	HiddenSize             uint32  `json:"hidden_size"`
+	IntermediateSize       uint32  `json:"intermediate_size"`
 	NumHeads               uint32  `json:"num_heads"`
+	NumPositionEmbeddings  uint32  `json:"num_position_embeddings"`
 	InChannels             uint32  `json:"in_channels"`
+	OutHiddenSize          uint32  `json:"out_hidden_size"`
 	PatchSize              uint32  `json:"patch_size"`
 	SpatialMergeSize       uint32  `json:"spatial_merge_size"`
 	RMSNormEps             float32 `json:"layer_norm_epsilon"`
@@ -96,12 +110,25 @@ type qwen3NextModel struct {
 	VisionEndTokenID   uint32 `json:"vision_end_token_id"`
 }
 
-var _ ModelConverter = (*qwen3NextModel)(nil)
+var (
+	_ ModelConverter      = (*qwen3NextModel)(nil)
+	_ MultimodalConverter = (*qwen3NextModel)(nil)
+)
 
 func (q *qwen3NextModel) parseMore(fsys fs.FS) error {
 	if q.TextConfig != nil {
 		q.qwen3NextTextConfig = *q.TextConfig
 	}
+	if q.NumNextNPredictLayers == 0 {
+		q.NumNextNPredictLayers = q.MTPNumHiddenLayers
+	}
+	if q.NumNextNPredictLayers == 0 {
+		nextn, err := qwen3NextInferNextNPredictLayers(fsys)
+		if err != nil {
+			return err
+		}
+		q.NumNextNPredictLayers = nextn
+	}
 
 	if q.RopeTheta == 0 {
 		q.RopeTheta = q.RopeParameters.RopeTheta
@@ -182,6 +209,150 @@ func (q *qwen3NextModel) parseMore(fsys fs.FS) error {
 	return nil
 }
 
+func qwen3NextInferNextNPredictLayers(fsys fs.FS) (uint32, error) {
+	paths, err := fs.Glob(fsys, "*.safetensors")
+	if err != nil {
+		return 0, err
+	}
+
+	maxLayer := -1
+	hasMTP := false
+	for _, p := range paths {
+		f, err := fsys.Open(p)
+		if err != nil {
+			return 0, err
+		}
+
+		var n int64
+		if err := binary.Read(f, binary.LittleEndian, &n); err != nil {
+			f.Close()
+			return 0, err
+		}
+
+		b := bytes.NewBuffer(make([]byte, 0, n))
+		if _, err = io.CopyN(b, f, n); err != nil {
+			f.Close()
+			return 0, err
+		}
+		f.Close()
+
+		var headers map[string]safetensorMetadata
+		if err := json.NewDecoder(b).Decode(&headers); err != nil {
+			return 0, err
+		}
+
+		for name, value := range headers {
+			if value.Type == "" || !strings.HasPrefix(name, "mtp.") {
+				continue
+			}
+			hasMTP = true
+			rest := strings.TrimPrefix(name, "mtp.layers.")
+			layer, suffix, ok := strings.Cut(rest, ".")
+			if !ok {
+				continue
+			}
+			n, err := strconv.Atoi(layer)
+			if err == nil && n > maxLayer && suffix != "" {
+				maxLayer = n
+			}
+		}
+	}
+
+	if maxLayer >= 0 {
+		return uint32(maxLayer + 1), nil
+	}
+	if hasMTP {
+		return 1, nil
+	}
+	return 0, nil
+}
+
+func ConvertQwen35MTPDraft(fsys fs.FS, f *os.File, baseKV ggml.KV, baseTensors []*ggml.Tensor) error {
+	arch := baseKV.Architecture()
+	if arch != "qwen35" && arch != "qwen35moe" {
+		return fmt.Errorf("MTP draft safetensors require a qwen3.5 base model, got %q", arch)
+	}
+
+	baseBlocks := baseKV.Uint("block_count")
+	if baseBlocks == 0 {
+		return fmt.Errorf("MTP draft safetensors require a base model with block_count")
+	}
+	if baseKV.Uint("nextn_predict_layers") > 0 {
+		return fmt.Errorf("MTP draft safetensors require a base model without embedded MTP layers")
+	}
+
+	nextn, err := qwen3NextInferNextNPredictLayers(fsys)
+	if err != nil {
+		return err
+	}
+	if nextn == 0 {
+		return fmt.Errorf("MTP draft safetensors did not contain mtp tensors")
+	}
+
+	q := &qwen3NextModel{
+		qwen3NextTextConfig: qwen3NextTextConfig{
+			NumHiddenLayers:       baseBlocks,
+			NumNextNPredictLayers: nextn,
+		},
+	}
+	ts, err := parseTensors(fsys, strings.NewReplacer(q.Replacements()...))
+	if err != nil {
+		return err
+	}
+	if err := ensureUniqueTensorNames(ts); err != nil {
+		return err
+	}
+
+	mtpTensors := q.Tensors(ts)
+	if len(mtpTensors) == 0 {
+		return fmt.Errorf("MTP draft safetensors did not produce GGUF tensors")
+	}
+	for _, tensor := range mtpTensors {
+		if !qwen35MTPDraftTensorName(tensor.Name, baseBlocks, nextn) {
+			return fmt.Errorf("MTP draft safetensors produced unexpected tensor %q", tensor.Name)
+		}
+		tensor.Shape = slices.Clone(tensor.Shape)
+		slices.Reverse(tensor.Shape)
+	}
+
+	kv := maps.Clone(baseKV)
+	qwen35RemoveSplitMetadata(kv, arch)
+	kv[arch+".block_count"] = baseBlocks + nextn
+	kv[arch+".nextn_predict_layers"] = nextn
+
+	tensors := make([]*ggml.Tensor, 0, len(baseTensors)+len(mtpTensors))
+	tensors = append(tensors, baseTensors...)
+	tensors = append(tensors, mtpTensors...)
+
+	var parameters uint64
+	for _, tensor := range tensors {
+		parameters += tensor.Elements()
+	}
+	kv["general.parameter_count"] = parameters
+
+	return ggml.WriteGGUF(f, kv, tensors)
+}
+
+func qwen35RemoveSplitMetadata(kv ggml.KV, arch string) {
+	for _, key := range []string{
+		"split.no",
+		"split.count",
+		"split.tensors.count",
+	} {
+		delete(kv, key)
+		delete(kv, arch+"."+key)
+	}
+}
+
+func qwen35MTPDraftTensorName(name string, base, nextn uint32) bool {
+	for i := range nextn {
+		if strings.HasPrefix(name, fmt.Sprintf("blk.%d.", base+i)) {
+			return true
+		}
+	}
+	return false
+}
+
 func (q *qwen3NextModel) kvHeadCounts() ([]uint32, error) {
 	if len(q.LayerTypes) > 0 {
 		kv := make([]uint32, q.NumHiddenLayers)
@@ -259,7 +430,10 @@ func (q *qwen3NextModel) KV(t *Tokenizer) KV {
 	}
 	kv["general.architecture"] = arch
 	kv["tokenizer.ggml.pre"] = "qwen35"
-	kv["block_count"] = q.NumHiddenLayers
+	kv["block_count"] = q.NumHiddenLayers + q.NumNextNPredictLayers
+	if q.NumNextNPredictLayers > 0 {
+		kv["nextn_predict_layers"] = q.NumNextNPredictLayers
+	}
 	kv["context_length"] = q.MaxPositionEmbeddings
 	kv["embedding_length"] = q.HiddenSize
 	kv["feed_forward_length"] = q.IntermediateSize
@@ -282,7 +456,11 @@ func (q *qwen3NextModel) KV(t *Tokenizer) KV {
 	if sections := q.ropeSections(); len(sections) > 0 {
 		kv["mrope_sections"] = sections
 		kv["rope.mrope_section"] = sections
-		kv["rope.dimension_sections"] = sections
+		dimensionSections := append([]int32(nil), sections...)
+		if len(dimensionSections) == 3 {
+			dimensionSections = append(dimensionSections, 0)
+		}
+		kv["rope.dimension_sections"] = dimensionSections
 	}
 	if q.RopeParameters.MRopeInterleaved {
 		kv["rope.mrope_interleaved"] = true
@@ -321,12 +499,21 @@ func (q *qwen3NextModel) KV(t *Tokenizer) KV {
 	}
 
 	if headCounts, err := q.kvHeadCounts(); err == nil {
-		kv["attention.head_count_kv"] = headCounts
+		var maxKV uint32
+		for _, count := range headCounts {
+			if count > maxKV {
+				maxKV = count
+			}
+		}
+		kv["attention.head_count_kv"] = maxKV
 	}
 
 	if q.VisionModel.Depth > 0 {
 		kv["vision.block_count"] = q.VisionModel.Depth
 		kv["vision.embedding_length"] = q.VisionModel.HiddenSize
+		if q.VisionModel.IntermediateSize > 0 {
+			kv["vision.feed_forward_length"] = q.VisionModel.IntermediateSize
+		}
 		kv["vision.attention.head_count"] = q.VisionModel.NumHeads
 		kv["vision.num_channels"] = q.VisionModel.InChannels
 		if q.VisionModel.PatchSize > 0 {
@@ -372,6 +559,378 @@ func (q *qwen3NextModel) KV(t *Tokenizer) KV {
 	return kv
 }
 
+func (q *qwen3NextModel) TextKV(t *Tokenizer) KV {
+	kv := q.KV(t)
+
+	for _, key := range []string{
+		"vision.block_count",
+		"vision.embedding_length",
+		"vision.feed_forward_length",
+		"vision.attention.head_count",
+		"vision.num_channels",
+		"vision.patch_size",
+		"vision.spatial_merge_size",
+		"vision.attention.layer_norm_epsilon",
+		"vision.rope.freq_base",
+		"vision.temporal_patch_size",
+		"vision.deepstack_visual_indexes",
+		"vision.shortest_edge",
+		"vision.longest_edge",
+		"vision.image_mean",
+		"vision.image_std",
+		"image_token_id",
+		"vision_start_token_id",
+		"vision_end_token_id",
+		"mrope_sections",
+		"rope.mrope_section",
+		"rope.mrope_interleaved",
+		"ssm.v_head_reordered",
+	} {
+		delete(kv, key)
+	}
+
+	return kv
+}
+
+func (q *qwen3NextModel) ProjectorKV(*Tokenizer) KV {
+	depth := q.VisionModel.Depth
+	deepstack := make([]bool, depth)
+	for _, idx := range q.VisionModel.DeepstackVisualIndexes {
+		if idx >= 0 && uint32(idx) < depth {
+			deepstack[idx] = true
+		}
+	}
+
+	imageSize := uint32(768)
+	if q.VisionModel.NumPositionEmbeddings > 0 && q.VisionModel.PatchSize > 0 {
+		root := uint32(math.Sqrt(float64(q.VisionModel.NumPositionEmbeddings)))
+		if root*root == q.VisionModel.NumPositionEmbeddings {
+			imageSize = root * q.VisionModel.PatchSize
+		}
+	}
+
+	projectionDim := q.VisionModel.OutHiddenSize
+	if projectionDim == 0 {
+		projectionDim = q.HiddenSize
+	}
+	layerNormEps := q.VisionModel.RMSNormEps
+	if layerNormEps == 0 {
+		layerNormEps = 1e-6
+	}
+
+	kv := KV{
+		"general.architecture":                     "clip",
+		"general.type":                             "mmproj",
+		"general.file_type":                        uint32(1),
+		"general.quantization_version":             uint32(2),
+		"clip.has_vision_encoder":                  true,
+		"clip.projector_type":                      "qwen3vl_merger",
+		"clip.use_gelu":                            true,
+		"clip.vision.block_count":                  depth,
+		"clip.vision.embedding_length":             q.VisionModel.HiddenSize,
+		"clip.vision.feed_forward_length":          q.VisionModel.IntermediateSize,
+		"clip.vision.attention.head_count":         q.VisionModel.NumHeads,
+		"clip.vision.image_size":                   imageSize,
+		"clip.vision.patch_size":                   q.VisionModel.PatchSize,
+		"clip.vision.projection_dim":               projectionDim,
+		"clip.vision.spatial_merge_size":           q.VisionModel.SpatialMergeSize,
+		"clip.vision.attention.layer_norm_epsilon": layerNormEps,
+		"clip.vision.is_deepstack_layers":          deepstack,
+	}
+	if len(q.VisionModel.ImageMean) > 0 {
+		kv["clip.vision.image_mean"] = q.VisionModel.ImageMean
+	}
+	if len(q.VisionModel.ImageStd) > 0 {
+		kv["clip.vision.image_std"] = q.VisionModel.ImageStd
+	}
+
+	return kv
+}
+
+func (q *qwen3NextModel) TextTensors(ts []Tensor, _ *Tokenizer) []*ggml.Tensor {
+	var text []Tensor
+	for _, t := range ts {
+		if qwen3NextVisionTensor(t.Name()) {
+			continue
+		}
+		text = append(text, t)
+	}
+
+	return q.Tensors(text)
+}
+
+func (q *qwen3NextModel) ProjectorTensors(ts []Tensor) []*ggml.Tensor {
+	if q.VisionModel.Depth == 0 {
+		return nil
+	}
+
+	rename := strings.NewReplacer(
+		"v.pos_embed", "v.position_embd",
+		"v.patch_embed", "v.patch_embd",
+		"v.merger.norm", "v.post_ln",
+		"v.merger.linear_fc1", "mm.0",
+		"v.merger.linear_fc2", "mm.2",
+		".mlp.linear_fc1", ".ffn_up",
+		".mlp.linear_fc2", ".ffn_down",
+		".norm1", ".ln1",
+		".norm2", ".ln2",
+	)
+
+	var out []*ggml.Tensor
+	for _, t := range ts {
+		name := t.Name()
+		if !qwen3NextVisionTensor(name) {
+			continue
+		}
+
+		if name == "v.patch_embed.weight" {
+			out = append(out, q.qwen35PatchEmbedTensors(t)...)
+			continue
+		}
+
+		outName := rename.Replace(name)
+		kind := t.Kind()
+		writer := io.WriterTo(t)
+		if outName == "v.position_embd.weight" {
+			kind = tensorKindFP32
+			writer = tensorFloat32Writer{tensor: t}
+		} else if sourceDType(t) == "BF16" && kind == tensorKindFP16 {
+			kind = tensorKindBF16
+			writer = tensorBF16Writer{tensor: t}
+		}
+		out = append(out, &ggml.Tensor{
+			Name:     outName,
+			Kind:     kind,
+			Shape:    slices.Clone(t.Shape()),
+			WriterTo: writer,
+		})
+	}
+
+	return out
+}
+
+func qwen3NextVisionTensor(name string) bool {
+	return strings.HasPrefix(name, "v.")
+}
+
+func (q *qwen3NextModel) qwen35PatchEmbedTensors(t Tensor) []*ggml.Tensor {
+	shape := t.Shape()
+	if len(shape) != 5 || shape[2] != 2 {
+		return nil
+	}
+
+	outShape := []uint64{shape[0], shape[1], shape[3], shape[4]}
+	return []*ggml.Tensor{
+		{
+			Name:     "v.patch_embd.weight",
+			Kind:     tensorKindFP32,
+			Shape:    slices.Clone(outShape),
+			WriterTo: tensorFloat32Writer{tensor: t, repacker: q.qwen35PatchEmbedSlice(0)},
+		},
+		{
+			Name:     "v.patch_embd.weight.1",
+			Kind:     tensorKindFP32,
+			Shape:    slices.Clone(outShape),
+			WriterTo: tensorFloat32Writer{tensor: t, repacker: q.qwen35PatchEmbedSlice(1)},
+		},
+	}
+}
+
+func (q *qwen3NextModel) qwen35PatchEmbedSlice(slice int) Repacker {
+	return func(_ string, data []float32, shape []uint64) ([]float32, error) {
+		if len(shape) != 5 || shape[2] != 2 {
+			return nil, fmt.Errorf("qwen3next: unexpected patch_embed shape %v", shape)
+		}
+
+		outChannels := int(shape[0])
+		inChannels := int(shape[1])
+		frames := int(shape[2])
+		height := int(shape[3])
+		width := int(shape[4])
+		if slice < 0 || slice >= frames {
+			return nil, fmt.Errorf("qwen3next: patch_embed slice %d out of range", slice)
+		}
+
+		expected := outChannels * inChannels * frames * height * width
+		if len(data) != expected {
+			return nil, fmt.Errorf("qwen3next: patch_embed data size %d, expected %d", len(data), expected)
+		}
+
+		out := make([]float32, outChannels*inChannels*height*width)
+		for oc := range outChannels {
+			for ic := range inChannels {
+				for y := range height {
+					for x := range width {
+						src := ((((oc*inChannels+ic)*frames+slice)*height + y) * width) + x
+						dst := (((oc*inChannels+ic)*height + y) * width) + x
+						out[dst] = data[src]
+					}
+				}
+			}
+		}
+
+		return out, nil
+	}
+}
+
+type tensorBF16Writer struct {
+	tensor   Tensor
+	repacker Repacker
+}
+
+func (w tensorBF16Writer) WriteTo(dst io.Writer) (int64, error) {
+	data, err := tensorFloat32Data(w.tensor)
+	if err != nil {
+		return 0, err
+	}
+	if w.repacker != nil {
+		data, err = w.repacker(w.tensor.Name(), data, w.tensor.Shape())
+		if err != nil {
+			return 0, err
+		}
+	}
+
+	u8s := bfloat16.EncodeFloat32(data)
+	if _, err := dst.Write(u8s); err != nil {
+		return 0, err
+	}
+	return int64(len(u8s)), nil
+}
+
+type tensorFloat32Writer struct {
+	tensor   Tensor
+	repacker Repacker
+}
+
+func (w tensorFloat32Writer) WriteTo(dst io.Writer) (int64, error) {
+	data, err := tensorFloat32Data(w.tensor)
+	if err != nil {
+		return 0, err
+	}
+	if w.repacker != nil {
+		data, err = w.repacker(w.tensor.Name(), data, w.tensor.Shape())
+		if err != nil {
+			return 0, err
+		}
+	}
+	if err := binary.Write(dst, binary.LittleEndian, data); err != nil {
+		return 0, err
+	}
+	return int64(len(data) * 4), nil
+}
+
+func tensorFloat32Data(t Tensor) ([]float32, error) {
+	if st, ok := tensorSafetensor(t); ok {
+		return safetensorFloat32Data(st)
+	}
+
+	var buf bytes.Buffer
+	if _, err := t.WriteTo(&buf); err != nil {
+		return nil, err
+	}
+
+	switch t.Kind() {
+	case tensorKindFP32:
+		out := make([]float32, buf.Len()/4)
+		if err := binary.Read(bytes.NewReader(buf.Bytes()), binary.LittleEndian, out); err != nil {
+			return nil, err
+		}
+		return out, nil
+	case tensorKindFP16:
+		raw := make([]uint16, buf.Len()/2)
+		if err := binary.Read(bytes.NewReader(buf.Bytes()), binary.LittleEndian, raw); err != nil {
+			return nil, err
+		}
+		out := make([]float32, len(raw))
+		for i, v := range raw {
+			out[i] = float16.Frombits(v).Float32()
+		}
+		return out, nil
+	case tensorKindBF16:
+		return bfloat16.DecodeFloat32(buf.Bytes()), nil
+	default:
+		return nil, fmt.Errorf("unsupported tensor kind %d for F32 writer", t.Kind())
+	}
+}
+
+func tensorSafetensor(t Tensor) (safetensor, bool) {
+	switch t := t.(type) {
+	case safetensor:
+		return t, true
+	case *safetensor:
+		return *t, true
+	default:
+		return safetensor{}, false
+	}
+}
+
+func safetensorFloat32Data(st safetensor) ([]float32, error) {
+	f, err := st.fs.Open(st.path)
+	if err != nil {
+		return nil, err
+	}
+	defer f.Close()
+
+	var r io.Reader
+	if readerAt, ok := f.(io.ReaderAt); ok {
+		r = io.NewSectionReader(readerAt, st.offset, st.size)
+	} else if seeker, ok := f.(io.Seeker); ok {
+		if _, err := seeker.Seek(st.offset, io.SeekStart); err != nil {
+			return nil, err
+		}
+		r = f
+	} else {
+		if _, err := io.CopyN(io.Discard, f, st.offset); err != nil {
+			return nil, err
+		}
+		r = f
+	}
+
+	br := bufio.NewReaderSize(r, min(32<<10, int(st.size)))
+	var out []float32
+	switch st.dtype {
+	case "F32":
+		out = make([]float32, st.size/4)
+		if err := binary.Read(br, binary.LittleEndian, out); err != nil {
+			return nil, err
+		}
+	case "F16":
+		raw := make([]uint16, st.size/2)
+		if err := binary.Read(br, binary.LittleEndian, raw); err != nil {
+			return nil, err
+		}
+		out = make([]float32, len(raw))
+		for i, v := range raw {
+			out[i] = float16.Frombits(v).Float32()
+		}
+	case "BF16":
+		raw := make([]uint8, st.size)
+		if err := binary.Read(br, binary.LittleEndian, raw); err != nil {
+			return nil, err
+		}
+		out = bfloat16.DecodeFloat32(raw)
+	case "F8_E4M3":
+		raw := make([]uint8, st.size)
+		if err := binary.Read(br, binary.LittleEndian, raw); err != nil {
+			return nil, err
+		}
+		out, err = st.decodeFP8E4M3(raw)
+		if err != nil {
+			return nil, err
+		}
+	default:
+		return nil, fmt.Errorf("unsupported safetensor dtype %q", st.dtype)
+	}
+
+	if st.repacker != nil {
+		out, err = st.repacker(st.Name(), out, st.Shape())
+		if err != nil {
+			return nil, err
+		}
+	}
+	return out, nil
+}
+
 func (q *qwen3NextModel) Tensors(ts []Tensor) []*ggml.Tensor {
 	var out []*ggml.Tensor
 
@@ -398,6 +957,13 @@ func (q *qwen3NextModel) Tensors(ts []Tensor) []*ggml.Tensor {
 		name := t.Name()
 		shape := t.Shape()
 
+		if names := q.mtpTensorNames(name); len(names) > 0 {
+			for _, name := range names {
+				out = q.appendDirectTensor(out, t, name)
+			}
+			continue
+		}
+
 		if strings.HasSuffix(name, ".ssm_in.weight") {
 			if qkv, gate, ok := q.splitQKVZTensor(t); ok {
 				out = append(out, qkv, gate)
@@ -464,7 +1030,7 @@ func (q *qwen3NextModel) Tensors(ts []Tensor) []*ggml.Tensor {
 			}
 			out = append(out, &ggml.Tensor{Name: name, Kind: t.Kind(), Shape: slices.Clone(shape), WriterTo: t})
 
-		case strings.HasSuffix(name, ".ssm_dt"):
+		case strings.HasSuffix(name, ".ssm_dt"), strings.HasSuffix(name, ".ssm_dt.bias"):
 			if q.shouldReorderVHeads() {
 				t.SetRepacker(q.repackReorderDim(0, 1))
 			}
@@ -499,6 +1065,73 @@ func (q *qwen3NextModel) Tensors(ts []Tensor) []*ggml.Tensor {
 	return out
 }
 
+func (q *qwen3NextModel) appendDirectTensor(out []*ggml.Tensor, t Tensor, name string) []*ggml.Tensor {
+	if qwen3NextShouldShiftNorm(name) {
+		t = t.Clone()
+		t.SetRepacker(q.addOne)
+	}
+	return append(out, &ggml.Tensor{Name: name, Kind: t.Kind(), Shape: slices.Clone(t.Shape()), WriterTo: t})
+}
+
+func qwen3NextShouldShiftNorm(name string) bool {
+	if strings.HasSuffix(name, ".ssm_norm.weight") {
+		return false
+	}
+	return strings.HasSuffix(name, "_norm.weight") ||
+		strings.HasSuffix(name, ".nextn.enorm.weight") ||
+		strings.HasSuffix(name, ".nextn.hnorm.weight")
+}
+
+func (q *qwen3NextModel) mtpTensorNames(name string) []string {
+	if !strings.HasPrefix(name, "mtp.") {
+		return nil
+	}
+
+	base := q.NumHiddenLayers
+	nextn := q.NumNextNPredictLayers
+	if nextn == 0 {
+		nextn = 1
+	}
+
+	if rest := strings.TrimPrefix(name, "mtp.layers."); rest != name {
+		layer, suffix, ok := strings.Cut(rest, ".")
+		if !ok {
+			return nil
+		}
+		idx, err := strconv.ParseUint(layer, 10, 32)
+		if err != nil {
+			return nil
+		}
+		return []string{fmt.Sprintf("blk.%d.%s", base+uint32(idx), suffix)}
+	}
+
+	var suffix string
+	switch name {
+	case "mtp.fc.weight":
+		suffix = "nextn.eh_proj.weight"
+	case "mtp.pre_fc_norm_embedding.weight":
+		suffix = "nextn.enorm.weight"
+	case "mtp.pre_fc_norm_hidden.weight":
+		suffix = "nextn.hnorm.weight"
+	case "mtp.norm.weight":
+		suffix = "nextn.shared_head_norm.weight"
+	case "mtp.embed_tokens.weight":
+		suffix = "nextn.embed_tokens.weight"
+	case "mtp.shared_head.head.weight":
+		suffix = "nextn.shared_head_head.weight"
+	case "mtp.shared_head.norm.weight":
+		suffix = "nextn.shared_head_norm.weight"
+	default:
+		return nil
+	}
+
+	names := make([]string, 0, nextn)
+	for i := range nextn {
+		names = append(names, fmt.Sprintf("blk.%d.%s", base+i, suffix))
+	}
+	return names
+}
+
 func (q *qwen3NextModel) repackReorderDim(dim, headDim int) Repacker {
 	return func(_ string, data []float32, shape []uint64) ([]float32, error) {
 		if !q.shouldReorderVHeads() {
@@ -925,7 +1558,7 @@ func (q *qwen3NextModel) Replacements() []string {
 		"linear_attn.in_proj_b", "ssm_beta",
 
 		"linear_attn.conv1d", "ssm_conv1d",
-		"linear_attn.dt_bias", "ssm_dt",
+		"linear_attn.dt_bias", "ssm_dt.bias",
 		"linear_attn.dt_proj", "ssm_dt",
 		"linear_attn.A_log", "ssm_a",
 		"linear_attn.norm", "ssm_norm",

convert/convert_qwen3next_test.go

@@ -4,10 +4,12 @@ import (
 	"bytes"
 	"encoding/binary"
 	"os"
+	"path/filepath"
 	"slices"
 	"strings"
 	"testing"
 
+	"github.com/d4l3k/go-bfloat16"
 	"github.com/ollama/ollama/fs/ggml"
 )
 
@@ -106,11 +108,7 @@ func TestQwen3NextKVLegacyConfig(t *testing.T) {
 		t.Fatalf("unexpected tokenizer pre: got %v want %v", got, want)
 	}
 
-	headCountKV, ok := kv["attention.head_count_kv"].([]uint32)
-	if !ok {
-		t.Fatalf("attention.head_count_kv has unexpected type: %T", kv["attention.head_count_kv"])
-	}
-	if got, want := headCountKV, []uint32{0, 2, 0, 2}; !slices.Equal(got, want) {
+	if got, want := kv["attention.head_count_kv"], uint32(2); got != want {
 		t.Fatalf("unexpected attention.head_count_kv: got %v want %v", got, want)
 	}
 
@@ -198,6 +196,7 @@ func TestQwen35KVFromTextConfig(t *testing.T) {
 		VisionModel: qwen3NextVisionConfig{
 			Depth:                  2,
 			HiddenSize:             128,
+			IntermediateSize:       512,
 			NumHeads:               4,
 			InChannels:             3,
 			PatchSize:              16,
@@ -225,11 +224,7 @@ func TestQwen35KVFromTextConfig(t *testing.T) {
 		t.Fatalf("unexpected architecture: got %v want %v", got, want)
 	}
 
-	headCountKV, ok := kv["attention.head_count_kv"].([]uint32)
-	if !ok {
-		t.Fatalf("attention.head_count_kv has unexpected type: %T", kv["attention.head_count_kv"])
-	}
-	if got, want := headCountKV, []uint32{0, 4, 0, 4}; !slices.Equal(got, want) {
+	if got, want := kv["attention.head_count_kv"], uint32(4); got != want {
 		t.Fatalf("unexpected attention.head_count_kv: got %v want %v", got, want)
 	}
 
@@ -248,7 +243,7 @@ func TestQwen35KVFromTextConfig(t *testing.T) {
 	if !ok {
 		t.Fatalf("rope.dimension_sections has unexpected type: %T", kv["rope.dimension_sections"])
 	}
-	if got, want := ropeSections, []int32{11, 11, 10}; !slices.Equal(got, want) {
+	if got, want := ropeSections, []int32{11, 11, 10, 0}; !slices.Equal(got, want) {
 		t.Fatalf("unexpected rope.dimension_sections: got %v want %v", got, want)
 	}
 
@@ -259,6 +254,254 @@ func TestQwen35KVFromTextConfig(t *testing.T) {
 	if got, want := kv["vision.block_count"], uint32(2); got != want {
 		t.Fatalf("unexpected vision.block_count: got %v want %v", got, want)
 	}
+	if got, want := kv["vision.feed_forward_length"], uint32(512); got != want {
+		t.Fatalf("unexpected vision.feed_forward_length: got %v want %v", got, want)
+	}
+}
+
+func TestQwen35MTPTensors(t *testing.T) {
+	m := &qwen3NextModel{
+		ModelParameters: ModelParameters{
+			ModelType: "qwen3_5",
+		},
+		qwen3NextTextConfig: qwen3NextTextConfig{
+			NumHiddenLayers:       32,
+			NumNextNPredictLayers: 1,
+		},
+	}
+
+	kv := m.KV(&Tokenizer{Vocabulary: &Vocabulary{}})
+	if got, want := kv["block_count"], uint32(33); got != want {
+		t.Fatalf("unexpected block_count: got %v want %v", got, want)
+	}
+	if got, want := kv["nextn_predict_layers"], uint32(1); got != want {
+		t.Fatalf("unexpected nextn_predict_layers: got %v want %v", got, want)
+	}
+
+	tensors := m.Tensors([]Tensor{
+		&fakeTensor{name: "mtp.fc.weight", shape: []uint64{2, 2}, data: make([]float32, 4)},
+		&fakeTensor{name: "mtp.pre_fc_norm_embedding.weight", shape: []uint64{2}, data: []float32{0, 1}},
+		&fakeTensor{name: "mtp.pre_fc_norm_hidden.weight", shape: []uint64{2}, data: []float32{0, 1}},
+		&fakeTensor{name: "mtp.norm.weight", shape: []uint64{2}, data: []float32{0, 1}},
+		&fakeTensor{name: "mtp.layers.0.attn_q.weight", shape: []uint64{2, 2}, data: make([]float32, 4)},
+		&fakeTensor{name: "mtp.layers.0.ffn_down.weight", shape: []uint64{2, 2}, data: make([]float32, 4)},
+	})
+
+	byName := map[string]*ggml.Tensor{}
+	for _, tensor := range tensors {
+		byName[tensor.Name] = tensor
+	}
+
+	for _, name := range []string{
+		"blk.32.nextn.eh_proj.weight",
+		"blk.32.nextn.enorm.weight",
+		"blk.32.nextn.hnorm.weight",
+		"blk.32.nextn.shared_head_norm.weight",
+		"blk.32.attn_q.weight",
+		"blk.32.ffn_down.weight",
+	} {
+		if _, ok := byName[name]; !ok {
+			t.Fatalf("missing MTP tensor %q", name)
+		}
+	}
+
+	for _, name := range []string{
+		"blk.32.nextn.enorm.weight",
+		"blk.32.nextn.hnorm.weight",
+		"blk.32.nextn.shared_head_norm.weight",
+	} {
+		if got, want := readTensorData(t, byName[name]), []float32{1, 2}; !slices.Equal(got, want) {
+			t.Fatalf("unexpected shifted norm values for %s: got %v want %v", name, got, want)
+		}
+	}
+}
+
+func TestQwen35NativeSplitKV(t *testing.T) {
+	m := &qwen3NextModel{
+		ModelParameters: ModelParameters{
+			ModelType: "qwen3_5",
+		},
+		TextConfig: &qwen3NextTextConfig{
+			MaxPositionEmbeddings: 16384,
+			HiddenSize:            2560,
+			NumHiddenLayers:       4,
+			IntermediateSize:      9216,
+			NumAttentionHeads:     16,
+			NumKeyValueHeads:      4,
+			HeadDim:               256,
+			RMSNormEPS:            1e-6,
+			FullAttentionInterval: 2,
+			LinearConvKernelDim:   4,
+			LinearKeyHeadDim:      128,
+			LinearNumKeyHeads:     16,
+			LinearNumValueHeads:   32,
+			LinearValueHeadDim:    128,
+			RopeParameters: qwen3NextRopeParams{
+				MRopeInterleaved:    true,
+				MropeSection:        []int32{11, 11, 10},
+				RopeTheta:           10_000_000,
+				PartialRotaryFactor: 0.25,
+			},
+		},
+		VisionModel: qwen3NextVisionConfig{
+			Depth:                 24,
+			HiddenSize:            1024,
+			IntermediateSize:      4096,
+			NumHeads:              16,
+			NumPositionEmbeddings: 2304,
+			InChannels:            3,
+			OutHiddenSize:         2560,
+			PatchSize:             16,
+			SpatialMergeSize:      2,
+		},
+		ImageTokenID:       248056,
+		VisionStartTokenID: 248053,
+		VisionEndTokenID:   248054,
+	}
+	m.VisionModel.ImageMean = []float32{0.5, 0.5, 0.5}
+	m.VisionModel.ImageStd = []float32{0.5, 0.5, 0.5}
+
+	if err := m.parseMore(os.DirFS(t.TempDir())); err != nil {
+		t.Fatal(err)
+	}
+
+	textKV := m.TextKV(&Tokenizer{Vocabulary: &Vocabulary{}})
+	for _, key := range []string{
+		"vision.block_count",
+		"image_token_id",
+		"vision_start_token_id",
+		"vision_end_token_id",
+		"mrope_sections",
+		"rope.mrope_section",
+		"rope.mrope_interleaved",
+		"ssm.v_head_reordered",
+	} {
+		if _, ok := textKV[key]; ok {
+			t.Fatalf("TextKV retained %q", key)
+		}
+	}
+	if got, want := textKV["rope.dimension_sections"], []int32{11, 11, 10, 0}; !slices.Equal(got.([]int32), want) {
+		t.Fatalf("unexpected rope.dimension_sections: got %v want %v", got, want)
+	}
+
+	projectorKV := m.ProjectorKV(&Tokenizer{Vocabulary: &Vocabulary{}})
+	if got, want := projectorKV["general.architecture"], "clip"; got != want {
+		t.Fatalf("unexpected projector architecture: got %v want %v", got, want)
+	}
+	if got, want := projectorKV["clip.projector_type"], "qwen3vl_merger"; got != want {
+		t.Fatalf("unexpected projector type: got %v want %v", got, want)
+	}
+	if got, want := projectorKV["clip.vision.feed_forward_length"], uint32(4096); got != want {
+		t.Fatalf("unexpected projector feed_forward_length: got %v want %v", got, want)
+	}
+	if got, want := projectorKV["clip.vision.image_size"], uint32(768); got != want {
+		t.Fatalf("unexpected projector image_size: got %v want %v", got, want)
+	}
+	if got, want := projectorKV["clip.vision.projection_dim"], uint32(2560); got != want {
+		t.Fatalf("unexpected projector projection_dim: got %v want %v", got, want)
+	}
+}
+
+func TestQwen35ProjectorTensors(t *testing.T) {
+	m := &qwen3NextModel{
+		VisionModel: qwen3NextVisionConfig{Depth: 1},
+	}
+
+	patch := &fakeTensor{
+		name:  "v.patch_embed.weight",
+		shape: []uint64{2, 2, 2, 1, 2},
+		data:  []float32{0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15},
+	}
+	tensors := m.ProjectorTensors([]Tensor{
+		patch,
+		&fakeTensor{name: "v.pos_embed.weight", shape: []uint64{4, 2}, data: []float32{0, 1, 2, 3, 4, 5, 6, 7}},
+		&fakeTensor{name: "v.blk.0.attn_qkv.weight", shape: []uint64{6, 2}, data: make([]float32, 12), sourceDType: "BF16", kind: tensorKindFP16},
+		&fakeTensor{name: "v.blk.0.mlp.linear_fc1.weight", shape: []uint64{8, 2}, data: make([]float32, 16), sourceDType: "BF16", kind: tensorKindFP16},
+		&fakeTensor{name: "token_embd.weight", shape: []uint64{2, 2}, data: make([]float32, 4)},
+		&fakeTensor{name: "mtp.fc.weight", shape: []uint64{2, 2}, data: make([]float32, 4)},
+	})
+
+	byName := map[string]*ggml.Tensor{}
+	for _, tensor := range tensors {
+		byName[tensor.Name] = tensor
+	}
+
+	if _, ok := byName["token_embd.weight"]; ok {
+		t.Fatalf("projector tensors included text tensor")
+	}
+	if _, ok := byName["mtp.fc.weight"]; ok {
+		t.Fatalf("projector tensors included MTP tensor")
+	}
+	if got := byName["v.position_embd.weight"]; got == nil || got.Kind != tensorKindFP32 {
+		t.Fatalf("position embedding was not promoted to F32: %#v", got)
+	}
+	if got := byName["v.blk.0.attn_qkv.weight"]; got == nil {
+		t.Fatalf("attn_qkv tensor missing")
+	} else if got.Kind != tensorKindBF16 {
+		t.Fatalf("attn_qkv tensor was not preserved as BF16: %#v", got)
+	}
+	if got := byName["v.blk.0.ffn_up.weight"]; got == nil {
+		t.Fatalf("ffn_up tensor missing")
+	} else if got.Kind != tensorKindBF16 {
+		t.Fatalf("ffn_up tensor was not preserved as BF16: %#v", got)
+	}
+
+	first := byName["v.patch_embd.weight"]
+	if first == nil {
+		t.Fatalf("first patch embedding slice missing")
+	}
+	if got, want := first.Shape, []uint64{2, 2, 1, 2}; !slices.Equal(got, want) {
+		t.Fatalf("unexpected first patch shape: got %v want %v", got, want)
+	}
+	if got, want := readTensorData(t, first), []float32{0, 1, 4, 5, 8, 9, 12, 13}; !slices.Equal(got, want) {
+		t.Fatalf("unexpected first patch data: got %v want %v", got, want)
+	}
+
+	second := byName["v.patch_embd.weight.1"]
+	if second == nil {
+		t.Fatalf("second patch embedding slice missing")
+	}
+	if got, want := readTensorData(t, second), []float32{2, 3, 6, 7, 10, 11, 14, 15}; !slices.Equal(got, want) {
+		t.Fatalf("unexpected second patch data: got %v want %v", got, want)
+	}
+}
+
+func TestQwen35BF16ProjectorWriterPreservesSource(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "tensor.bin")
+	values := []float32{1, -2, 3.5, 4.25}
+	raw := bfloat16.EncodeFloat32(values)
+	if err := os.WriteFile(path, raw, 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	st := safetensor{
+		fs:     os.DirFS(dir),
+		path:   "tensor.bin",
+		dtype:  "BF16",
+		offset: 0,
+		size:   int64(len(raw)),
+		tensorBase: &tensorBase{
+			name:  "v.blk.0.attn_qkv.weight",
+			shape: []uint64{2, 2},
+		},
+	}
+	tensor := &ggml.Tensor{
+		Name:     "v.blk.0.attn_qkv.weight",
+		Kind:     tensorKindBF16,
+		Shape:    []uint64{2, 2},
+		WriterTo: tensorBF16Writer{tensor: st},
+	}
+
+	var got bytes.Buffer
+	if n, err := tensor.WriteTo(&got); err != nil {
+		t.Fatal(err)
+	} else if n != int64(len(raw)) {
+		t.Fatalf("unexpected byte count: got %d want %d", n, len(raw))
+	}
+	if !bytes.Equal(got.Bytes(), raw) {
+		t.Fatalf("BF16 writer changed source bytes: got %x want %x", got.Bytes(), raw)
+	}
 }
 
 func TestQwen3NextReplacements(t *testing.T) {
@@ -273,6 +516,12 @@ func TestQwen3NextReplacements(t *testing.T) {
 	if got, want := r.Replace("model.layers.1.linear_attn.in_proj_qkvz.weight"), "blk.1.ssm_in.weight"; got != want {
 		t.Fatalf("unexpected legacy replacement: got %q want %q", got, want)
 	}
+	if got, want := r.Replace("model.layers.1.linear_attn.dt_bias"), "blk.1.ssm_dt.bias"; got != want {
+		t.Fatalf("unexpected dt bias replacement: got %q want %q", got, want)
+	}
+	if got, want := r.Replace("model.layers.1.linear_attn.dt_proj.weight"), "blk.1.ssm_dt.weight"; got != want {
+		t.Fatalf("unexpected dt projection replacement: got %q want %q", got, want)
+	}
 }
 
 func TestQwen35ReordersVHeads(t *testing.T) {
@@ -399,6 +648,33 @@ func TestQwen35ReordersSsmBetaRows(t *testing.T) {
 	}
 }
 
+func TestQwen35ReordersSsmDtBias(t *testing.T) {
+	m := &qwen3NextModel{
+		ModelParameters: ModelParameters{
+			ModelType: "qwen3_5",
+		},
+		qwen3NextTextConfig: qwen3NextTextConfig{
+			LinearNumKeyHeads:   2,
+			LinearNumValueHeads: 4,
+		},
+	}
+
+	out := m.Tensors([]Tensor{
+		&fakeTensor{
+			name:  "blk.0.ssm_dt.bias",
+			shape: []uint64{4},
+			data:  []float32{0, 1, 2, 3},
+		},
+	})
+	if len(out) != 1 {
+		t.Fatalf("unexpected output tensor count: got %d want 1", len(out))
+	}
+
+	if got, want := readTensorData(t, out[0]), []float32{0, 2, 1, 3}; !slices.Equal(got, want) {
+		t.Fatalf("unexpected ssm_dt.bias data: got %v want %v", got, want)
+	}
+}
+
 func TestQwen35ReordersConv1DChannelDim(t *testing.T) {
 	m := &qwen3NextModel{
 		ModelParameters: ModelParameters{

convert/convert_qwen3vl.go

@@ -3,8 +3,13 @@ package convert
 import (
 	"cmp"
 	"encoding/json"
+	"fmt"
+	"io"
 	"io/fs"
+	"math"
+	"regexp"
 	"slices"
+	"strconv"
 	"strings"
 
 	"github.com/ollama/ollama/fs/ggml"
@@ -25,6 +30,9 @@ type qwen3VLModel struct {
 		RopeTheta              float32 `json:"rope_theta"`
 		TemporalPatchSize      uint32  `json:"temporal_patch_size"`
 		DeepstackVisualIndexes []int32 `json:"deepstack_visual_indexes"`
+		IntermediateSize       uint32  `json:"intermediate_size"`
+		OutHiddenSize          uint32  `json:"out_hidden_size"`
+		NumPositionEmbeddings  uint32  `json:"num_position_embeddings"`
 
 		Size struct {
 			ShortestEdge uint32 `json:"shortest_edge"`
@@ -36,6 +44,8 @@ type qwen3VLModel struct {
 	} `json:"vision_config"`
 }
 
+var _ MultimodalConverter = (*qwen3VLModel)(nil)
+
 func (m *qwen3VLModel) parseMore(fsys fs.FS) error {
 	bts, err := fs.ReadFile(fsys, "preprocessor_config.json")
 	if err != nil {
@@ -55,8 +65,20 @@ func (m *qwen3VLModel) KV(t *Tokenizer) KV {
 	// override architecture
 	kv["general.architecture"] = arch
 
+	if sections := m.RopeScaling.MropeSection; len(sections) > 0 {
+		dimensionSections := append([]int32(nil), sections...)
+		if len(dimensionSections) == 3 {
+			dimensionSections = append(dimensionSections, 0)
+		}
+		kv["rope.dimension_sections"] = dimensionSections
+	}
+	kv["n_deepstack_layers"] = uint32(len(m.VisionModel.DeepstackVisualIndexes))
+
 	kv["vision.block_count"] = cmp.Or(m.VisionModel.Depth, 32)
 	kv["vision.embedding_length"] = m.VisionModel.HiddenSize
+	if m.VisionModel.IntermediateSize > 0 {
+		kv["vision.feed_forward_length"] = m.VisionModel.IntermediateSize
+	}
 	kv["vision.attention.head_count"] = cmp.Or(m.VisionModel.NumHeads, 16)
 	kv["vision.num_channels"] = m.VisionModel.InChannels
 	kv["vision.patch_size"] = cmp.Or(m.VisionModel.PatchSize, 14)
@@ -75,6 +97,234 @@ func (m *qwen3VLModel) KV(t *Tokenizer) KV {
 	return kv
 }
 
+func (m *qwen3VLModel) TextKV(t *Tokenizer) KV {
+	kv := m.KV(t)
+	for _, key := range []string{
+		"vision.block_count",
+		"vision.embedding_length",
+		"vision.feed_forward_length",
+		"vision.attention.head_count",
+		"vision.num_channels",
+		"vision.patch_size",
+		"vision.spatial_merge_size",
+		"vision.attention.layer_norm_epsilon",
+		"vision.rope.freq_base",
+		"vision.temporal_patch_size",
+		"vision.deepstack_visual_indexes",
+		"vision.shortest_edge",
+		"vision.longest_edge",
+		"vision.image_mean",
+		"vision.image_std",
+		"rope.mrope_section",
+	} {
+		delete(kv, key)
+	}
+
+	return kv
+}
+
+func (m *qwen3VLModel) ProjectorKV(*Tokenizer) KV {
+	depth := cmp.Or(m.VisionModel.Depth, uint32(32))
+	deepstack := make([]bool, depth)
+	for _, idx := range m.VisionModel.DeepstackVisualIndexes {
+		if idx >= 0 && uint32(idx) < depth {
+			deepstack[idx] = true
+		}
+	}
+
+	projectionDim := m.VisionModel.OutHiddenSize
+	if projectionDim == 0 {
+		projectionDim = m.HiddenSize
+	}
+	layerNormEps := m.VisionModel.RMSNormEps
+	if layerNormEps == 0 {
+		layerNormEps = 1e-6
+	}
+
+	kv := KV{
+		"general.architecture":                     "clip",
+		"general.type":                             "mmproj",
+		"general.file_type":                        uint32(1),
+		"general.quantization_version":             uint32(2),
+		"clip.has_vision_encoder":                  true,
+		"clip.projector_type":                      "qwen3vl_merger",
+		"clip.use_gelu":                            true,
+		"clip.vision.block_count":                  depth,
+		"clip.vision.embedding_length":             m.VisionModel.HiddenSize,
+		"clip.vision.feed_forward_length":          cmp.Or(m.VisionModel.IntermediateSize, m.VisionModel.HiddenSize*4),
+		"clip.vision.attention.head_count":         cmp.Or(m.VisionModel.NumHeads, uint32(16)),
+		"clip.vision.attention.layer_norm_epsilon": layerNormEps,
+		"clip.vision.num_channels":                 m.VisionModel.InChannels,
+		"clip.vision.patch_size":                   cmp.Or(m.VisionModel.PatchSize, uint32(14)),
+		"clip.vision.spatial_merge_size":           cmp.Or(m.VisionModel.SpatialMergeSize, uint32(2)),
+		"clip.vision.image_size":                   m.projectorImageSize(),
+		"clip.vision.projection_dim":               projectionDim,
+		"clip.vision.temporal_patch_size":          cmp.Or(m.VisionModel.TemporalPatchSize, uint32(2)),
+		"clip.vision.rope.freq_base":               cmp.Or(m.VisionModel.RopeTheta, float32(1e4)),
+		"clip.vision.is_deepstack_layers":          deepstack,
+	}
+	if m.VisionModel.Size.ShortestEdge > 0 {
+		kv["clip.vision.image_min_pixels"] = m.VisionModel.Size.ShortestEdge
+	}
+	if m.VisionModel.Size.LongestEdge > 0 {
+		kv["clip.vision.image_max_pixels"] = m.VisionModel.Size.LongestEdge
+	}
+	if len(m.VisionModel.ImageMean) == 3 {
+		kv["clip.vision.image_mean"] = m.VisionModel.ImageMean
+	}
+	if len(m.VisionModel.ImageStd) == 3 {
+		kv["clip.vision.image_std"] = m.VisionModel.ImageStd
+	}
+
+	return kv
+}
+
+func (m *qwen3VLModel) projectorImageSize() uint32 {
+	if m.VisionModel.NumPositionEmbeddings > 0 && m.VisionModel.PatchSize > 0 {
+		root := uint32(math.Sqrt(float64(m.VisionModel.NumPositionEmbeddings)))
+		if root*root == m.VisionModel.NumPositionEmbeddings {
+			return root * m.VisionModel.PatchSize
+		}
+	}
+	return uint32(768)
+}
+
+func qwen3VLVisionTensor(name string) bool {
+	return strings.HasPrefix(name, "v.") || strings.HasPrefix(name, "mm.")
+}
+
+func (m *qwen3VLModel) TextTensors(ts []Tensor, _ *Tokenizer) []*ggml.Tensor {
+	var textOnly []Tensor
+	for _, t := range ts {
+		if qwen3VLVisionTensor(t.Name()) {
+			continue
+		}
+		textOnly = append(textOnly, t)
+	}
+
+	return m.qwen3Model.Tensors(textOnly)
+}
+
+func (m *qwen3VLModel) qwen3VLProjectorRename(name string) string {
+	if strings.HasPrefix(name, "v.merger.") {
+		name = strings.Replace(name, "v.merger.linear_fc1", "mm.0", 1)
+		name = strings.Replace(name, "v.merger.linear_fc2", "mm.2", 1)
+		name = strings.Replace(name, "v.merger.norm", "v.post_ln", 1)
+		return name
+	}
+
+	if strings.HasPrefix(name, "v.deepstack.") {
+		re := regexp.MustCompile(`^v\.deepstack\.(\d+)\.(.+)$`)
+		if matches := re.FindStringSubmatch(name); matches != nil {
+			seqIdx, err := strconv.Atoi(matches[1])
+			if err == nil && seqIdx < len(m.VisionModel.DeepstackVisualIndexes) {
+				blockIdx := m.VisionModel.DeepstackVisualIndexes[seqIdx]
+				suffix := matches[2]
+				suffix = strings.Replace(suffix, "linear_fc1", "fc1", 1)
+				suffix = strings.Replace(suffix, "linear_fc2", "fc2", 1)
+				return fmt.Sprintf("v.deepstack.%d.%s", blockIdx, suffix)
+			}
+		}
+	}
+
+	return name
+}
+
+func (m *qwen3VLModel) ProjectorTensors(ts []Tensor) []*ggml.Tensor {
+	var out []*ggml.Tensor
+
+	for _, t := range ts {
+		if !qwen3VLVisionTensor(t.Name()) {
+			continue
+		}
+
+		name := m.qwen3VLProjectorRename(t.Name())
+		if name == "v.patch_embd.weight" {
+			out = append(out, m.qwen3VLPatchEmbedTensors(t)...)
+			continue
+		}
+
+		kind := t.Kind()
+		var writer io.WriterTo = t
+		if name == "v.position_embd.weight" {
+			kind = tensorKindFP32
+			writer = tensorFloat32Writer{tensor: t}
+		} else if sourceDType(t) == "BF16" && kind == tensorKindFP16 {
+			kind = tensorKindBF16
+			writer = tensorBF16Writer{tensor: t}
+		}
+
+		out = append(out, &ggml.Tensor{
+			Name:     name,
+			Kind:     kind,
+			Shape:    slices.Clone(t.Shape()),
+			WriterTo: writer,
+		})
+	}
+
+	return out
+}
+
+func (m *qwen3VLModel) qwen3VLPatchEmbedTensors(t Tensor) []*ggml.Tensor {
+	shape := t.Shape()
+	if len(shape) != 5 || shape[2] != 2 {
+		return nil
+	}
+
+	outShape := []uint64{shape[0], shape[1], shape[3], shape[4]}
+	return []*ggml.Tensor{
+		{
+			Name:     "v.patch_embd.weight",
+			Kind:     tensorKindFP32,
+			Shape:    slices.Clone(outShape),
+			WriterTo: tensorFloat32Writer{tensor: t, repacker: qwenTemporalPatchEmbedSlice(0)},
+		},
+		{
+			Name:     "v.patch_embd.weight.1",
+			Kind:     tensorKindFP32,
+			Shape:    slices.Clone(outShape),
+			WriterTo: tensorFloat32Writer{tensor: t, repacker: qwenTemporalPatchEmbedSlice(1)},
+		},
+	}
+}
+
+func qwenTemporalPatchEmbedSlice(slice int) Repacker {
+	return func(_ string, data []float32, shape []uint64) ([]float32, error) {
+		if len(shape) != 5 || shape[2] != 2 {
+			return nil, fmt.Errorf("qwen temporal patch embedding shape %v", shape)
+		}
+
+		outChannels := int(shape[0])
+		inChannels := int(shape[1])
+		frames := int(shape[2])
+		height := int(shape[3])
+		width := int(shape[4])
+		if slice < 0 || slice >= frames {
+			return nil, fmt.Errorf("qwen temporal patch embedding slice %d out of range", slice)
+		}
+
+		expected := outChannels * inChannels * frames * height * width
+		if len(data) != expected {
+			return nil, fmt.Errorf("qwen temporal patch embedding data size %d, expected %d", len(data), expected)
+		}
+
+		out := make([]float32, outChannels*inChannels*height*width)
+		for oc := range outChannels {
+			for ic := range inChannels {
+				for y := range height {
+					for x := range width {
+						src := ((((oc*inChannels+ic)*frames+slice)*height + y) * width) + x
+						dst := (((oc*inChannels+ic)*height + y) * width) + x
+						out[dst] = data[src]
+					}
+				}
+			}
+		}
+
+		return out, nil
+	}
+}
+
 func (m *qwen3VLModel) Tensors(ts []Tensor) []*ggml.Tensor {
 	var rest []Tensor
 	var out []*ggml.Tensor
@@ -107,10 +357,15 @@ func (m *qwen3VLModel) Replacements() []string {
 		m.qwen3Model.Replacements(),
 		"model.language_", "",
 		"model.visual", "v",
-		"patch_embed.proj", "patch_embed",
+		"patch_embed.proj", "patch_embd",
+		"pos_embed", "position_embd",
 		"blocks", "blk",
 		"attn.qkv", "attn_qkv",
 		"attn.proj", "attn_out",
-		"deepstack_merger_list", "deepstack_merger",
+		"norm1", "ln1",
+		"norm2", "ln2",
+		"mlp.linear_fc1", "ffn_up",
+		"mlp.linear_fc2", "ffn_down",
+		"deepstack_merger_list", "deepstack",
 	)
 }

convert/convert_qwen3vl_test.go

@@ -0,0 +1,147 @@
+package convert
+
+import (
+	"slices"
+	"testing"
+
+	"github.com/ollama/ollama/fs/ggml"
+)
+
+func TestQwen3VLTextAndProjectorKV(t *testing.T) {
+	m := &qwen3VLModel{
+		qwen3Model: qwen3Model{
+			HiddenSize: 2048,
+		},
+	}
+	m.RopeScaling.Type = "mrope"
+	m.RopeScaling.MropeSection = []int32{24, 20, 20}
+	m.VisionModel.Depth = 24
+	m.VisionModel.HiddenSize = 1024
+	m.VisionModel.IntermediateSize = 4096
+	m.VisionModel.OutHiddenSize = 2048
+	m.VisionModel.NumHeads = 16
+	m.VisionModel.InChannels = 3
+	m.VisionModel.PatchSize = 16
+	m.VisionModel.SpatialMergeSize = 2
+	m.VisionModel.NumPositionEmbeddings = 2304
+	m.VisionModel.TemporalPatchSize = 2
+	m.VisionModel.RMSNormEps = 1e-6
+	m.VisionModel.RopeTheta = 10000
+	m.VisionModel.DeepstackVisualIndexes = []int32{5, 11, 17}
+	m.VisionModel.ImageMean = []float32{0.5, 0.5, 0.5}
+	m.VisionModel.ImageStd = []float32{0.5, 0.5, 0.5}
+
+	textKV := m.TextKV(&Tokenizer{Vocabulary: &Vocabulary{}})
+	if got, want := textKV["general.architecture"], "qwen3vl"; got != want {
+		t.Fatalf("unexpected text architecture: got %v want %v", got, want)
+	}
+	if got, want := textKV["rope.dimension_sections"], []int32{24, 20, 20, 0}; !slices.Equal(got.([]int32), want) {
+		t.Fatalf("unexpected rope.dimension_sections: got %v want %v", got, want)
+	}
+	if got, want := textKV["n_deepstack_layers"], uint32(3); got != want {
+		t.Fatalf("unexpected n_deepstack_layers: got %v want %v", got, want)
+	}
+	for _, key := range []string{"vision.block_count", "vision.deepstack_visual_indexes", "rope.mrope_section"} {
+		if _, ok := textKV[key]; ok {
+			t.Fatalf("TextKV retained %q", key)
+		}
+	}
+
+	projectorKV := m.ProjectorKV(&Tokenizer{Vocabulary: &Vocabulary{}})
+	if got, want := projectorKV["general.architecture"], "clip"; got != want {
+		t.Fatalf("unexpected projector architecture: got %v want %v", got, want)
+	}
+	if got, want := projectorKV["general.type"], "mmproj"; got != want {
+		t.Fatalf("unexpected projector type: got %v want %v", got, want)
+	}
+	if got, want := projectorKV["clip.projector_type"], "qwen3vl_merger"; got != want {
+		t.Fatalf("unexpected projector type: got %v want %v", got, want)
+	}
+	if got, want := projectorKV["clip.vision.feed_forward_length"], uint32(4096); got != want {
+		t.Fatalf("unexpected feed_forward_length: got %v want %v", got, want)
+	}
+	if got, want := projectorKV["clip.vision.image_size"], uint32(768); got != want {
+		t.Fatalf("unexpected image_size: got %v want %v", got, want)
+	}
+	mask, ok := projectorKV["clip.vision.is_deepstack_layers"].([]bool)
+	if !ok {
+		t.Fatalf("deepstack mask has unexpected type: %T", projectorKV["clip.vision.is_deepstack_layers"])
+	}
+	if len(mask) != 24 || !mask[5] || !mask[11] || !mask[17] {
+		t.Fatalf("unexpected deepstack mask: %v", mask)
+	}
+}
+
+func TestQwen3VLProjectorTensors(t *testing.T) {
+	m := &qwen3VLModel{}
+	m.VisionModel.DeepstackVisualIndexes = []int32{5, 11, 17}
+
+	tensors := m.ProjectorTensors([]Tensor{
+		&fakeTensor{
+			name:  "v.patch_embd.weight",
+			shape: []uint64{2, 2, 2, 1, 2},
+			data:  []float32{0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15},
+		},
+		&fakeTensor{name: "v.position_embd.weight", shape: []uint64{4, 2}, data: []float32{0, 1, 2, 3, 4, 5, 6, 7}},
+		&fakeTensor{name: "v.merger.linear_fc1.weight", shape: []uint64{4, 2}, data: make([]float32, 8)},
+		&fakeTensor{name: "v.merger.linear_fc2.bias", shape: []uint64{4}, data: make([]float32, 4)},
+		&fakeTensor{name: "v.merger.norm.weight", shape: []uint64{2}, data: make([]float32, 2)},
+		&fakeTensor{name: "v.deepstack.0.linear_fc1.weight", shape: []uint64{4, 2}, data: make([]float32, 8)},
+		&fakeTensor{name: "v.deepstack.1.norm.bias", shape: []uint64{2}, data: make([]float32, 2)},
+		&fakeTensor{name: "v.blk.0.attn_qkv.weight", shape: []uint64{6, 2}, data: make([]float32, 12), sourceDType: "BF16", kind: tensorKindFP16},
+		&fakeTensor{name: "token_embd.weight", shape: []uint64{2, 2}, data: make([]float32, 4)},
+	})
+
+	byName := map[string]uint32{}
+	for _, tensor := range tensors {
+		byName[tensor.Name] = tensor.Kind
+	}
+
+	if _, ok := byName["token_embd.weight"]; ok {
+		t.Fatalf("projector tensors included text tensor")
+	}
+	if got := byName["v.position_embd.weight"]; got != tensorKindFP32 {
+		t.Fatalf("position embedding was not promoted to F32: %d", got)
+	}
+	if got := byName["v.blk.0.attn_qkv.weight"]; got != tensorKindBF16 {
+		t.Fatalf("BF16 projector tensor was not preserved: %d", got)
+	}
+	for _, name := range []string{
+		"mm.0.weight",
+		"mm.2.bias",
+		"v.post_ln.weight",
+		"v.deepstack.5.fc1.weight",
+		"v.deepstack.11.norm.bias",
+	} {
+		if _, ok := byName[name]; !ok {
+			t.Fatalf("missing projector tensor %q", name)
+		}
+	}
+
+	firstTensor := tensorsByName(tensors)["v.patch_embd.weight"]
+	if firstTensor == nil {
+		t.Fatalf("first patch embedding slice missing")
+	}
+	if got, want := firstTensor.Shape, []uint64{2, 2, 1, 2}; !slices.Equal(got, want) {
+		t.Fatalf("unexpected first patch shape: got %v want %v", got, want)
+	}
+	if got, want := readTensorData(t, firstTensor), []float32{0, 1, 4, 5, 8, 9, 12, 13}; !slices.Equal(got, want) {
+		t.Fatalf("unexpected first patch data: got %v want %v", got, want)
+	}
+
+	secondTensor := tensorsByName(tensors)["v.patch_embd.weight.1"]
+	if secondTensor == nil {
+		t.Fatalf("second patch embedding slice missing")
+	}
+	if got, want := readTensorData(t, secondTensor), []float32{2, 3, 6, 7, 10, 11, 14, 15}; !slices.Equal(got, want) {
+		t.Fatalf("unexpected second patch data: got %v want %v", got, want)
+	}
+}
+
+func tensorsByName(tensors []*ggml.Tensor) map[string]*ggml.Tensor {
+	byName := map[string]*ggml.Tensor{}
+	for _, tensor := range tensors {
+		byName[tensor.Name] = tensor
+	}
+	return byName
+}

convert/tensor_test.go

@@ -22,6 +22,7 @@ type fakeTensor struct {
 	data  []float32
 
 	sourceDType string
+	kind        uint32
 	repacker    Repacker
 }
 
@@ -34,6 +35,9 @@ func (f fakeTensor) Shape() []uint64 {
 }
 
 func (f fakeTensor) Kind() uint32 {
+	if f.kind != 0 {
+		return f.kind
+	}
 	return 0
 }
 
@@ -51,6 +55,7 @@ func (f fakeTensor) Clone() Tensor {
 		shape:       slices.Clone(f.shape),
 		data:        slices.Clone(f.data),
 		sourceDType: f.sourceDType,
+		kind:        f.kind,
 		repacker:    f.repacker,
 	}
 }

convert/tokenizer.go

@@ -149,6 +149,7 @@ func parseTokenizer(fsys fs.FS, specialTokenTypes []string) (*Tokenizer, error)
 				if err := json.Unmarshal(bts, &sv.AddToken); err != nil {
 					return nil, err
 				}
+				sv.AddTokenSet = true
 			}
 
 			if bts, ok := p[fmt.Sprintf("%s_token", st)]; ok {
@@ -314,6 +315,10 @@ type SpecialVocabulary struct {
 	ID       int
 	Content  string
 	AddToken bool
+	// AddTokenSet tracks whether tokenizer_config.json explicitly defined the
+	// add_*_token setting. Missing and explicit false have different GGUF
+	// semantics for some tokenizers.
+	AddTokenSet bool
 
 	// IDs is populated by generation_config.json
 	IDs []int32

convert/tokenizer_test.go

@@ -184,8 +184,8 @@ func TestParseTokenizer(t *testing.T) {
 				},
 				SpecialVocabulary: []*SpecialVocabulary{
 					{Type: "pad", Content: "<pad>", ID: 0, AddToken: false},
-					{Type: "eos", Content: "<eos>", ID: 1, AddToken: false},
-					{Type: "bos", Content: "<bos>", ID: 2, AddToken: true},
+					{Type: "eos", Content: "<eos>", ID: 1, AddToken: false, AddTokenSet: true},
+					{Type: "bos", Content: "<bos>", ID: 2, AddToken: true, AddTokenSet: true},
 					{Type: "unk", Content: "<unk>", ID: 3, AddToken: false},
 				},
 				Pre: "default",
@@ -380,8 +380,8 @@ func TestParseTokenizer(t *testing.T) {
 					Types:  []int32{3, 3, 3, 3},
 				},
 				SpecialVocabulary: []*SpecialVocabulary{
-					{Type: "eos", Content: "<eos>", ID: 1, IDs: []int32{1, 2, 3}, AddToken: false},
-					{Type: "bos", Content: "<bos>", ID: 0, AddToken: true},
+					{Type: "eos", Content: "<eos>", ID: 1, IDs: []int32{1, 2, 3}, AddToken: false, AddTokenSet: true},
+					{Type: "bos", Content: "<bos>", ID: 0, AddToken: true, AddTokenSet: true},
 				},
 				Pre: "default",
 			},
@@ -423,3 +423,23 @@ func TestParseTokenizer(t *testing.T) {
 		})
 	}
 }
+
+func TestModelParametersKVOmitsMissingAddToken(t *testing.T) {
+	kv := ModelParameters{}.KV(&Tokenizer{
+		Vocabulary: &Vocabulary{Model: "gpt2"},
+		SpecialVocabulary: []*SpecialVocabulary{
+			{Type: "bos", Content: "<bos>", ID: 1},
+			{Type: "eos", Content: "<eos>", ID: 2, AddToken: false, AddTokenSet: true},
+		},
+	})
+
+	if _, ok := kv["tokenizer.ggml.add_bos_token"]; ok {
+		t.Errorf("tokenizer.ggml.add_bos_token should be omitted when add_bos_token is absent")
+	}
+	if got := kv["tokenizer.ggml.bos_token_id"]; got != uint32(1) {
+		t.Errorf("tokenizer.ggml.bos_token_id = %v, want 1", got)
+	}
+	if got, ok := kv["tokenizer.ggml.add_eos_token"]; !ok || got != false {
+		t.Errorf("tokenizer.ggml.add_eos_token = %v, %v; want explicit false", got, ok)
+	}
+}

discover/amd.go

@@ -0,0 +1,487 @@
+// AMD discovery needs a small amount of backend-specific handling beyond the
+// generic llama-server device list. ROCm devices expose their real capability
+// as gfx targets, and the shipped rocBLAS kernels define which of those
+// targets are actually usable. On Linux, KFD topology and DRM sysfs attributes
+// provide the integrated-vs-discrete signal needed for scheduler decisions. On
+// Windows, older HIP driver installs can also leave ROCm libraries present but
+// too old to support GPU inference. These helpers keep that extra validation
+// and warning logic in one place.
+package discover
+
+import (
+	"bufio"
+	"log/slog"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"regexp"
+	"runtime"
+	"sort"
+	"strconv"
+	"strings"
+
+	"github.com/ollama/ollama/ml"
+)
+
+// gfxTargetRegex matches ROCm stderr lines like:
+//
+//	Device 0: AMD Radeon RX 6700 XT, gfx1031 (0x1031), VMM: no, Wave Size: 32, VRAM: 12272 MiB
+//	Device 1: AMD Radeon Pro VII, gfx906:sramecc+:xnack- (0x906), VMM: no, Wave Size: 64, VRAM: 16368 MiB
+var gfxTargetRegex = regexp.MustCompile(
+	`Device\s+(\d+):.*,\s+(gfx[0-9a-f]+)[\s:(]`,
+)
+
+var pciIDRegex = regexp.MustCompile(`^[0-9a-fA-F]{4}:[0-9a-fA-F]{2}:[0-9a-fA-F]{2}\.[0-7]$`)
+
+func parseROCmGFXTargets(output string) map[int]string {
+	gfxByIndex := make(map[int]string)
+
+	scanner := bufio.NewScanner(strings.NewReader(output))
+	for scanner.Scan() {
+		if matches := gfxTargetRegex.FindStringSubmatch(scanner.Text()); matches != nil {
+			idx, _ := strconv.Atoi(matches[1])
+			gfxByIndex[idx] = matches[2]
+		}
+	}
+
+	return gfxByIndex
+}
+
+func parseGFXTarget(gfx string) (int, int) {
+	gfx, ok := strings.CutPrefix(gfx, "gfx")
+	if !ok || len(gfx) < 3 {
+		return 0, 0
+	}
+
+	major, err := strconv.ParseInt(gfx[:len(gfx)-2], 16, 32)
+	if err != nil {
+		return 0, 0
+	}
+	minor, err := strconv.ParseInt(gfx[len(gfx)-2:], 16, 32)
+	if err != nil {
+		return 0, 0
+	}
+
+	return int(major), int(minor)
+}
+
+// HSA_OVERRIDE_GFX_VERSION changes the effective HIP/rocBLAS target even
+// though KFD/sysfs still reports the physical ASIC.
+func hsaOverrideGFXTarget() string {
+	return rocmGFXTargetOverride(os.Getenv("HSA_OVERRIDE_GFX_VERSION"))
+}
+
+func rocmGFXTargetOverride(value string) string {
+	value = strings.TrimSpace(value)
+	if value == "" {
+		return ""
+	}
+	if strings.HasPrefix(value, "gfx") {
+		if major, minor := parseGFXTarget(value); major != 0 || minor != 0 {
+			return value
+		}
+		return ""
+	}
+
+	parts := strings.Split(value, ".")
+	if len(parts) != 3 {
+		return ""
+	}
+
+	var digits [3]uint64
+	for i, part := range parts {
+		digit, err := strconv.ParseUint(part, 10, 8)
+		if err != nil || digit > 0xf {
+			return ""
+		}
+		digits[i] = digit
+	}
+
+	return "gfx" +
+		strconv.FormatUint(digits[0], 10) +
+		strconv.FormatUint(digits[1], 16) +
+		strconv.FormatUint(digits[2], 16)
+}
+
+func setROCmGFXTarget(device *ml.DeviceInfo, gfx string) {
+	if gfx == "" || device.Library != "ROCm" {
+		return
+	}
+	device.GFXTarget = gfx
+	device.ComputeMajor, device.ComputeMinor = parseGFXTarget(gfx)
+}
+
+// rocblasGFXTargets scans the rocblas library directory for supported gfx targets
+// by looking for TensileLibrary_lazy_gfxNNNN.dat files.
+func rocblasGFXTargets(libDirs []string) map[string]bool {
+	targets := make(map[string]bool)
+	for _, dir := range libDirs {
+		files, _ := filepath.Glob(filepath.Join(dir, "rocblas", "library", "TensileLibrary_lazy_gfx*.dat"))
+		for _, f := range files {
+			base := filepath.Base(f)
+			if t, ok := strings.CutPrefix(base, "TensileLibrary_lazy_"); ok {
+				if t, ok = strings.CutSuffix(t, ".dat"); ok {
+					targets[t] = true
+				}
+			}
+		}
+	}
+	return targets
+}
+
+type rocmLinuxSysfsDevice struct {
+	pciID      string
+	gfxTarget  string
+	integrated bool
+	known      bool
+}
+
+func refineLinuxROCmDevices(devices []ml.DeviceInfo) []ml.DeviceInfo {
+	if runtime.GOOS != "linux" {
+		return devices
+	}
+	applyLinuxROCmRefinement(devices, "/sys")
+	return devices
+}
+
+func applyLinuxROCmRefinement(devices []ml.DeviceInfo, sysfsRoot string) bool {
+	var rocmIndexes []int
+	for i, device := range devices {
+		if device.Library == "ROCm" {
+			rocmIndexes = append(rocmIndexes, i)
+		}
+	}
+	if len(rocmIndexes) == 0 {
+		return false
+	}
+
+	sysfsDevices, err := readROCmLinuxSysfsDevices(sysfsRoot)
+	if err != nil {
+		slog.Debug("linux rocm device refinement unavailable", "error", err)
+		return false
+	}
+	if len(sysfsDevices) != len(rocmIndexes) {
+		slog.Debug("linux rocm device refinement skipped: device count mismatch",
+			"llama_server_count", len(rocmIndexes), "kfd_count", len(sysfsDevices))
+		return false
+	}
+
+	byPCI := map[string]rocmLinuxSysfsDevice{}
+	byGFX := uniqueROCmSysfsDevicesByGFX(sysfsDevices)
+	for _, sysfsDevice := range sysfsDevices {
+		if sysfsDevice.pciID != "" {
+			byPCI[strings.ToLower(sysfsDevice.pciID)] = sysfsDevice
+		}
+	}
+
+	refined := 0
+	for i, rocmIndex := range rocmIndexes {
+		device := &devices[rocmIndex]
+		sysfsDevice, ok := matchROCmLinuxSysfsDevice(*device, i, sysfsDevices, byPCI, byGFX)
+		if !ok {
+			slog.Debug("linux rocm device refinement skipped: no stable match",
+				"device", device.Name, "pci_id", device.PCIID, "gfx", device.GFXTarget)
+			continue
+		}
+		applyROCmLinuxSysfsDevice(device, sysfsDevice)
+		refined++
+	}
+
+	if refined == 0 {
+		return false
+	}
+
+	slog.Debug("linux rocm device refinement applied", "devices", refined)
+	return true
+}
+
+func uniqueROCmSysfsDevicesByGFX(sysfsDevices []rocmLinuxSysfsDevice) map[string]rocmLinuxSysfsDevice {
+	byGFX := map[string]rocmLinuxSysfsDevice{}
+	duplicates := map[string]bool{}
+	for _, sysfsDevice := range sysfsDevices {
+		if sysfsDevice.gfxTarget == "" {
+			continue
+		}
+		if _, ok := byGFX[sysfsDevice.gfxTarget]; ok {
+			duplicates[sysfsDevice.gfxTarget] = true
+			continue
+		}
+		byGFX[sysfsDevice.gfxTarget] = sysfsDevice
+	}
+	for gfx := range duplicates {
+		delete(byGFX, gfx)
+	}
+	return byGFX
+}
+
+func matchROCmLinuxSysfsDevice(device ml.DeviceInfo, index int, sysfsDevices []rocmLinuxSysfsDevice, byPCI, byGFX map[string]rocmLinuxSysfsDevice) (rocmLinuxSysfsDevice, bool) {
+	// ROCm visibility envs can remap backend ordinals while sysfs stays in
+	// physical KFD order, so prefer stable identity before index fallback.
+	if device.PCIID != "" {
+		if sysfsDevice, ok := byPCI[strings.ToLower(device.PCIID)]; ok {
+			return sysfsDevice, true
+		}
+	}
+
+	if device.GFXTarget != "" {
+		if sysfsDevice, ok := byGFX[device.GFXTarget]; ok {
+			return sysfsDevice, true
+		}
+	}
+
+	if index >= len(sysfsDevices) {
+		return rocmLinuxSysfsDevice{}, false
+	}
+	sysfsDevice := sysfsDevices[index]
+	if sysfsDevice.gfxTarget != "" && device.GFXTarget != "" && sysfsDevice.gfxTarget != device.GFXTarget {
+		slog.Debug("linux rocm device refinement index mismatch",
+			"device", device.Name, "llama_server_gfx", device.GFXTarget, "kfd_gfx", sysfsDevice.gfxTarget)
+		return rocmLinuxSysfsDevice{}, false
+	}
+	return sysfsDevice, true
+}
+
+func applyROCmLinuxSysfsDevice(device *ml.DeviceInfo, sysfsDevice rocmLinuxSysfsDevice) {
+	if sysfsDevice.pciID != "" {
+		device.PCIID = sysfsDevice.pciID
+	}
+	if sysfsDevice.known {
+		device.Integrated = sysfsDevice.integrated
+	}
+}
+
+func readROCmLinuxSysfsDevices(sysfsRoot string) ([]rocmLinuxSysfsDevice, error) {
+	nodeRoot := filepath.Join(sysfsRoot, "class", "kfd", "kfd", "topology", "nodes")
+	entries, err := os.ReadDir(nodeRoot)
+	if err != nil {
+		return nil, err
+	}
+
+	sort.Slice(entries, func(i, j int) bool {
+		left, _ := strconv.Atoi(entries[i].Name())
+		right, _ := strconv.Atoi(entries[j].Name())
+		return left < right
+	})
+
+	var devices []rocmLinuxSysfsDevice
+	for _, entry := range entries {
+		if !entry.IsDir() {
+			continue
+		}
+		properties, err := readKFDNodeProperties(filepath.Join(nodeRoot, entry.Name(), "properties"))
+		if err != nil || !properties.isGPU() {
+			continue
+		}
+
+		device, err := readROCmDRMDevice(sysfsRoot, properties.drmRenderMinor)
+		if err != nil {
+			slog.Debug("linux rocm sysfs device skipped", "node", entry.Name(), "error", err)
+			continue
+		}
+		device.gfxTarget = gfxTargetFromKFDVersion(properties.gfxTargetVersion)
+		devices = append(devices, device)
+	}
+
+	return devices, nil
+}
+
+type kfdNodeProperties struct {
+	vendorID         uint64
+	deviceID         uint64
+	drmRenderMinor   int
+	gfxTargetVersion uint64
+}
+
+func (p kfdNodeProperties) isGPU() bool {
+	return p.vendorID != 0 && p.deviceID != 0 && p.drmRenderMinor != 0
+}
+
+func readKFDNodeProperties(path string) (kfdNodeProperties, error) {
+	file, err := os.Open(path)
+	if err != nil {
+		return kfdNodeProperties{}, err
+	}
+	defer file.Close()
+
+	values := make(map[string]string)
+	scanner := bufio.NewScanner(file)
+	for scanner.Scan() {
+		fields := strings.Fields(scanner.Text())
+		if len(fields) >= 2 {
+			values[fields[0]] = fields[1]
+		}
+	}
+	if err := scanner.Err(); err != nil {
+		return kfdNodeProperties{}, err
+	}
+
+	vendorID, _ := parseSysfsUint(values["vendor_id"])
+	deviceID, _ := parseSysfsUint(values["device_id"])
+	renderMinor, _ := parseSysfsUint(values["drm_render_minor"])
+	gfxVersion, _ := parseSysfsUint(values["gfx_target_version"])
+
+	return kfdNodeProperties{
+		vendorID:         vendorID,
+		deviceID:         deviceID,
+		drmRenderMinor:   int(renderMinor),
+		gfxTargetVersion: gfxVersion,
+	}, nil
+}
+
+func readROCmDRMDevice(sysfsRoot string, renderMinor int) (rocmLinuxSysfsDevice, error) {
+	devicePath := filepath.Join(sysfsRoot, "class", "drm", "renderD"+strconv.Itoa(renderMinor), "device")
+	resolvedDevicePath, err := filepath.EvalSymlinks(devicePath)
+	if err != nil {
+		return rocmLinuxSysfsDevice{}, err
+	}
+
+	vendor, err := readSysfsString(filepath.Join(resolvedDevicePath, "vendor"))
+	if err != nil {
+		return rocmLinuxSysfsDevice{}, err
+	}
+	if !strings.EqualFold(vendor, "0x1002") {
+		return rocmLinuxSysfsDevice{}, nil
+	}
+
+	driver, err := readSysfsDriverName(filepath.Join(resolvedDevicePath, "driver"))
+	if err != nil {
+		return rocmLinuxSysfsDevice{}, err
+	}
+	if driver != "amdgpu" {
+		return rocmLinuxSysfsDevice{}, nil
+	}
+
+	device := rocmLinuxSysfsDevice{pciID: pciIDFromPath(resolvedDevicePath)}
+	if sysfsFileExists(filepath.Join(resolvedDevicePath, "mem_info_vram_vendor")) ||
+		sysfsFileExists(filepath.Join(resolvedDevicePath, "board_info")) {
+		device.known = true
+		return device, nil
+	}
+
+	vramTotal, ok := readROCmLinuxMemoryInfo(resolvedDevicePath, "mem_info_vram_total")
+	if !ok {
+		return device, nil
+	}
+	gttTotal, ok := readROCmLinuxMemoryInfo(resolvedDevicePath, "mem_info_gtt_total")
+	if !ok {
+		return device, nil
+	}
+
+	const (
+		maxIntegratedVRAM = 4 << 30
+		minSharedGTT      = 8 << 30
+	)
+	if vramTotal > 0 && vramTotal <= maxIntegratedVRAM && gttTotal >= minSharedGTT && gttTotal >= 4*vramTotal {
+		device.integrated = true
+		device.known = true
+	}
+
+	return device, nil
+}
+
+func readROCmLinuxMemoryInfo(devicePath, name string) (uint64, bool) {
+	value, err := readSysfsUint(filepath.Join(devicePath, name))
+	return value, err == nil
+}
+
+func readSysfsString(path string) (string, error) {
+	data, err := os.ReadFile(path)
+	if err != nil {
+		return "", err
+	}
+	return strings.TrimSpace(string(data)), nil
+}
+
+func readSysfsDriverName(path string) (string, error) {
+	driver, readErr := readSysfsString(path)
+	if readErr == nil {
+		return driver, nil
+	}
+	driverPath, err := filepath.EvalSymlinks(path)
+	if err == nil {
+		return filepath.Base(driverPath), nil
+	}
+	return "", readErr
+}
+
+func readSysfsUint(path string) (uint64, error) {
+	value, err := readSysfsString(path)
+	if err != nil {
+		return 0, err
+	}
+	return parseSysfsUint(value)
+}
+
+func parseSysfsUint(value string) (uint64, error) {
+	return strconv.ParseUint(strings.TrimSpace(value), 0, 64)
+}
+
+func sysfsFileExists(path string) bool {
+	_, err := os.Stat(path)
+	return err == nil
+}
+
+func pciIDFromPath(path string) string {
+	base := filepath.Base(path)
+	if pciIDRegex.MatchString(base) {
+		return base
+	}
+	return ""
+}
+
+func gfxTargetFromKFDVersion(version uint64) string {
+	if version == 0 {
+		return ""
+	}
+	major := version / 10000
+	minor := (version / 100) % 100
+	stepping := version % 100
+	if minor > 0xf || stepping > 0xf {
+		return ""
+	}
+	return "gfx" + strconv.FormatUint(major, 10) + strconv.FormatUint(minor, 16) + strconv.FormatUint(stepping, 16)
+}
+
+// filterUnsupportedROCmDevices removes ROCm devices whose gfx target doesn't have
+// matching rocblas kernels bundled.
+func filterUnsupportedROCmDevices(devices []ml.DeviceInfo, libDirs []string) []ml.DeviceInfo {
+	supported := rocblasGFXTargets(libDirs)
+	if len(supported) == 0 {
+		return devices
+	}
+
+	override := hsaOverrideGFXTarget()
+	var filtered []ml.DeviceInfo
+	for _, dev := range devices {
+		if dev.Library != "ROCm" {
+			filtered = append(filtered, dev)
+			continue
+		}
+
+		setROCmGFXTarget(&dev, override)
+		gfx := dev.GFXTarget
+		if gfx == "" {
+			filtered = append(filtered, dev)
+			continue
+		}
+		if supported[gfx] {
+			filtered = append(filtered, dev)
+		} else {
+			slog.Warn("dropping ROCm device — no rocblas support for gfx target",
+				"device", dev.Name, "gfx_target", gfx, "supported", supported,
+				"hint", "set HSA_OVERRIDE_GFX_VERSION to map to a supported target")
+		}
+	}
+	return filtered
+}
+
+func detectOldAMDDriverWindows() {
+	if runtime.GOOS != "windows" {
+		return
+	}
+	_, errV6 := exec.LookPath("amdhip64_6.dll")
+	_, errV7 := exec.LookPath("amdhip64_7.dll")
+	if errV6 == nil && errV7 != nil {
+		slog.Warn("AMD driver is too old. Update your AMD driver to enable GPU inference.")
+	}
+}

discover/amd_test.go

@@ -0,0 +1,289 @@
+package discover
+
+import (
+	"os"
+	"path/filepath"
+	"runtime"
+	"strconv"
+	"testing"
+
+	"github.com/ollama/ollama/ml"
+)
+
+func TestApplyLinuxROCmRefinement(t *testing.T) {
+	if runtime.GOOS == "windows" {
+		t.Skip("fake Linux PCI sysfs paths use ':' which is not valid in Windows filenames")
+	}
+
+	tests := []struct {
+		name           string
+		nodes          []fakeROCmNode
+		devices        []ml.DeviceInfo
+		applied        bool
+		wantIntegrated []bool
+		wantPCIIDs     []string
+	}{
+		{
+			name: "apu is integrated",
+			nodes: []fakeROCmNode{{
+				node:        1,
+				renderMinor: 128,
+				gfxVersion:  "90012",
+				vramTotal:   2 << 30,
+				gttTotal:    32 << 30,
+			}},
+			devices: []ml.DeviceInfo{{
+				DeviceID:  ml.DeviceID{ID: "0", Library: "ROCm"},
+				Name:      "ROCm0",
+				GFXTarget: "gfx90c",
+			}},
+			applied:        true,
+			wantIntegrated: []bool{true},
+		},
+		{
+			name: "low vram dgpu is not integrated",
+			nodes: []fakeROCmNode{{
+				node:        1,
+				renderMinor: 128,
+				gfxVersion:  "100601",
+				vramTotal:   4 << 30,
+				gttTotal:    32 << 30,
+				vramVendor:  true,
+				boardInfo:   true,
+			}},
+			devices: []ml.DeviceInfo{{
+				DeviceID:  ml.DeviceID{ID: "0", Library: "ROCm"},
+				Name:      "ROCm0",
+				GFXTarget: "gfx1061",
+			}},
+			applied:        true,
+			wantIntegrated: []bool{false},
+		},
+		{
+			name: "mixed system follows kfd order not drm order",
+			nodes: []fakeROCmNode{
+				{
+					node:        1,
+					renderMinor: 129,
+					gfxVersion:  "110000",
+					vramTotal:   48 << 30,
+					gttTotal:    64 << 30,
+					vramVendor:  true,
+					boardInfo:   true,
+				},
+				{
+					node:        2,
+					renderMinor: 128,
+					gfxVersion:  "110003",
+					vramTotal:   512 << 20,
+					gttTotal:    32 << 30,
+				},
+			},
+			devices: []ml.DeviceInfo{
+				{DeviceID: ml.DeviceID{ID: "0", Library: "ROCm"}, Name: "ROCm0", GFXTarget: "gfx1100"},
+				{DeviceID: ml.DeviceID{ID: "1", Library: "ROCm"}, Name: "ROCm1", GFXTarget: "gfx1103"},
+			},
+			applied:        true,
+			wantIntegrated: []bool{false, true},
+		},
+		{
+			name: "remapped visible order matches existing pci identity",
+			nodes: []fakeROCmNode{
+				{
+					node:        1,
+					renderMinor: 128,
+					pciID:       "0000:e3:00.0",
+					gfxVersion:  "110000",
+					vramTotal:   48 << 30,
+					gttTotal:    64 << 30,
+					vramVendor:  true,
+					boardInfo:   true,
+				},
+				{
+					node:        2,
+					renderMinor: 129,
+					pciID:       "0000:c3:00.0",
+					gfxVersion:  "120000",
+					vramTotal:   2 << 30,
+					gttTotal:    32 << 30,
+				},
+			},
+			devices: []ml.DeviceInfo{
+				{DeviceID: ml.DeviceID{ID: "0", Library: "ROCm"}, Name: "ROCm0", GFXTarget: "gfx1200", PCIID: "0000:c3:00.0"},
+				{DeviceID: ml.DeviceID{ID: "1", Library: "ROCm"}, Name: "ROCm1", GFXTarget: "gfx1100", PCIID: "0000:e3:00.0"},
+			},
+			applied:        true,
+			wantIntegrated: []bool{true, false},
+			wantPCIIDs:     []string{"0000:c3:00.0", "0000:e3:00.0"},
+		},
+		{
+			name: "remapped visible order matches unique gfx when pci is absent",
+			nodes: []fakeROCmNode{
+				{
+					node:        1,
+					renderMinor: 128,
+					pciID:       "0000:e3:00.0",
+					gfxVersion:  "110000",
+					vramTotal:   48 << 30,
+					gttTotal:    64 << 30,
+					vramVendor:  true,
+					boardInfo:   true,
+				},
+				{
+					node:        2,
+					renderMinor: 129,
+					pciID:       "0000:c3:00.0",
+					gfxVersion:  "120000",
+					vramTotal:   2 << 30,
+					gttTotal:    32 << 30,
+				},
+			},
+			devices: []ml.DeviceInfo{
+				{DeviceID: ml.DeviceID{ID: "0", Library: "ROCm"}, Name: "ROCm0", GFXTarget: "gfx1200"},
+				{DeviceID: ml.DeviceID{ID: "1", Library: "ROCm"}, Name: "ROCm1", GFXTarget: "gfx1100"},
+			},
+			applied:        true,
+			wantIntegrated: []bool{true, false},
+			wantPCIIDs:     []string{"0000:c3:00.0", "0000:e3:00.0"},
+		},
+		{
+			name: "missing kfd data leaves devices unchanged",
+			devices: []ml.DeviceInfo{{
+				DeviceID:   ml.DeviceID{ID: "0", Library: "ROCm"},
+				Name:       "ROCm0",
+				Integrated: true,
+			}},
+			wantIntegrated: []bool{true},
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			sysfsRoot := t.TempDir()
+			for _, node := range tt.nodes {
+				writeFakeROCmNode(t, sysfsRoot, node)
+			}
+
+			devices := append([]ml.DeviceInfo(nil), tt.devices...)
+			applied := applyLinuxROCmRefinement(devices, sysfsRoot)
+			if applied != tt.applied {
+				t.Fatalf("applied = %v, want %v", applied, tt.applied)
+			}
+			for i, want := range tt.wantIntegrated {
+				if devices[i].Integrated != want {
+					t.Fatalf("device %d integrated = %v, want %v", i, devices[i].Integrated, want)
+				}
+			}
+			for i, want := range tt.wantPCIIDs {
+				if devices[i].PCIID != want {
+					t.Fatalf("device %d PCIID = %q, want %q", i, devices[i].PCIID, want)
+				}
+			}
+		})
+	}
+}
+
+func TestSameRefreshDeviceMatchesROCmByPCI(t *testing.T) {
+	updated := ml.DeviceInfo{
+		DeviceID: ml.DeviceID{ID: "0", Library: "ROCm"},
+		PCIID:    "0000:c3:00.0",
+	}
+	existing := ml.DeviceInfo{
+		DeviceID: ml.DeviceID{ID: "1", Library: "ROCm"},
+		PCIID:    "0000:C3:00.0",
+	}
+	if !sameRefreshDevice(updated, existing) {
+		t.Fatal("sameRefreshDevice did not match remapped ROCm device by PCI ID")
+	}
+}
+
+func TestFilterUnsupportedROCmDevicesRespectsHSAOverride(t *testing.T) {
+	t.Setenv("HSA_OVERRIDE_GFX_VERSION", "10.3.0")
+
+	libDir := t.TempDir()
+	rocblasDir := filepath.Join(libDir, "rocblas", "library")
+	if err := os.MkdirAll(rocblasDir, 0o755); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(rocblasDir, "TensileLibrary_lazy_gfx1030.dat"), nil, 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	devices := filterUnsupportedROCmDevices([]ml.DeviceInfo{{
+		DeviceID:     ml.DeviceID{ID: "0", Library: "ROCm"},
+		Name:         "ROCm0",
+		GFXTarget:    "gfx1031",
+		ComputeMajor: 0x10,
+		ComputeMinor: 0x31,
+	}}, []string{libDir})
+	if len(devices) != 1 {
+		t.Fatalf("got %d devices, want 1", len(devices))
+	}
+	if got := devices[0].GFXTarget; got != "gfx1030" {
+		t.Fatalf("GFXTarget = %q, want gfx1030", got)
+	}
+	if got := devices[0].Compute(); got != "gfx1030" {
+		t.Fatalf("Compute() = %q, want gfx1030", got)
+	}
+}
+
+type fakeROCmNode struct {
+	node        int
+	renderMinor int
+	pciID       string
+	gfxVersion  string
+	vramTotal   uint64
+	gttTotal    uint64
+	vramVendor  bool
+	boardInfo   bool
+}
+
+func writeFakeROCmNode(t *testing.T, sysfsRoot string, node fakeROCmNode) {
+	t.Helper()
+
+	nodeDir := filepath.Join(sysfsRoot, "class", "kfd", "kfd", "topology", "nodes", strconv.Itoa(node.node))
+	if err := os.MkdirAll(nodeDir, 0o755); err != nil {
+		t.Fatal(err)
+	}
+	properties := "vendor_id 4098\n" +
+		"device_id 1234\n" +
+		"drm_render_minor " + strconv.Itoa(node.renderMinor) + "\n" +
+		"gfx_target_version " + node.gfxVersion + "\n"
+	if err := os.WriteFile(filepath.Join(nodeDir, "properties"), []byte(properties), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	deviceDir := filepath.Join(sysfsRoot, "class", "drm", "renderD"+strconv.Itoa(node.renderMinor), "device")
+	if node.pciID != "" {
+		targetDir := filepath.Join(sysfsRoot, "devices", node.pciID)
+		if err := os.MkdirAll(targetDir, 0o755); err != nil {
+			t.Fatal(err)
+		}
+		if err := os.MkdirAll(filepath.Dir(deviceDir), 0o755); err != nil {
+			t.Fatal(err)
+		}
+		if err := os.Symlink(targetDir, deviceDir); err != nil {
+			t.Skipf("symlink unavailable for fake sysfs PCI path: %v", err)
+		}
+		deviceDir = targetDir
+	} else if err := os.MkdirAll(deviceDir, 0o755); err != nil {
+		t.Fatal(err)
+	}
+	writeFakeSysfsFile(t, deviceDir, "vendor", "0x1002\n")
+	writeFakeSysfsFile(t, deviceDir, "driver", "amdgpu\n")
+	writeFakeSysfsFile(t, deviceDir, "mem_info_vram_total", strconv.FormatUint(node.vramTotal, 10)+"\n")
+	writeFakeSysfsFile(t, deviceDir, "mem_info_gtt_total", strconv.FormatUint(node.gttTotal, 10)+"\n")
+	if node.vramVendor {
+		writeFakeSysfsFile(t, deviceDir, "mem_info_vram_vendor", "samsung\n")
+	}
+	if node.boardInfo {
+		writeFakeSysfsFile(t, deviceDir, "board_info", "type : cem\n")
+	}
+}
+
+func writeFakeSysfsFile(t *testing.T, dir, name, content string) {
+	t.Helper()
+	if err := os.WriteFile(filepath.Join(dir, name), []byte(content), 0o644); err != nil {
+		t.Fatal(err)
+	}
+}

discover/cpu_linux.go

@@ -4,13 +4,8 @@ import (
 	"bufio"
 	"errors"
 	"fmt"
-	"io"
-	"log/slog"
 	"os"
 	"path/filepath"
-	"reflect"
-	"regexp"
-	"sort"
 	"strconv"
 	"strings"
 
@@ -92,143 +87,6 @@ func getUint64ValueFromFile(path string) (uint64, error) {
 	return 0, errors.New("empty file content")
 }
 
-const CpuInfoFilename = "/proc/cpuinfo"
-
-type linuxCpuInfo struct {
-	ID         string `cpuinfo:"processor"`
-	VendorID   string `cpuinfo:"vendor_id"`
-	ModelName  string `cpuinfo:"model name"`
-	PhysicalID string `cpuinfo:"physical id"`
-	Siblings   string `cpuinfo:"siblings"`
-	CoreID     string `cpuinfo:"core id"`
-}
-
-func GetCPUDetails() []CPU {
-	file, err := os.Open(CpuInfoFilename)
-	if err != nil {
-		slog.Warn("failed to get CPU details", "error", err)
-		return nil
-	}
-	defer file.Close()
-	cpus := linuxCPUDetails(file)
-	return overwriteThreadCountByLinuxCgroups(cpus)
-}
-
-func overwriteThreadCountByLinuxCgroups(cpus []CPU) []CPU {
-	file, err := os.Open("/sys/fs/cgroup/cpu.max")
-	if err != nil {
-		return cpus
-	}
-	defer file.Close()
-
-	scanner := bufio.NewScanner(file)
-	for scanner.Scan() {
-		line := scanner.Text()
-		if sl := strings.Split(line, " "); len(sl) == 2 {
-			allowdUs, err := strconv.ParseInt(sl[0], 10, 64)
-			if err != nil {
-				slog.Warn("failed to parse CPU allowed micro secs", "error", err)
-				return cpus
-			}
-			unitUs, err := strconv.ParseInt(sl[1], 10, 64)
-			if err != nil {
-				slog.Warn("failed to parse CPU unit micro secs", "error", err)
-				return cpus
-			}
-
-			threads := int(max(allowdUs/unitUs, 1))
-
-			cpu := cpus[0]
-			cpu.CoreCount = threads
-			cpu.ThreadCount = threads
-			return []CPU{cpu}
-		}
-	}
-	return cpus
-}
-
-func linuxCPUDetails(file io.Reader) []CPU {
-	reColumns := regexp.MustCompile("\t+: ")
-	scanner := bufio.NewScanner(file)
-	cpuInfos := []linuxCpuInfo{}
-	cpu := &linuxCpuInfo{}
-	for scanner.Scan() {
-		line := scanner.Text()
-		if sl := reColumns.Split(line, 2); len(sl) > 1 {
-			t := reflect.TypeOf(cpu).Elem()
-			s := reflect.ValueOf(cpu).Elem()
-			for i := range t.NumField() {
-				field := t.Field(i)
-				tag := field.Tag.Get("cpuinfo")
-				if tag == sl[0] {
-					s.FieldByName(field.Name).SetString(sl[1])
-					break
-				}
-			}
-		} else if strings.TrimSpace(line) == "" && cpu.ID != "" {
-			cpuInfos = append(cpuInfos, *cpu)
-			cpu = &linuxCpuInfo{}
-		}
-	}
-	if cpu.ID != "" {
-		cpuInfos = append(cpuInfos, *cpu)
-	}
-
-	// Process the sockets/cores/threads
-	socketByID := map[string]*CPU{}
-	coreBySocket := map[string]map[string]struct{}{}
-	threadsByCoreBySocket := map[string]map[string]int{}
-	for _, c := range cpuInfos {
-		if _, found := socketByID[c.PhysicalID]; !found {
-			socketByID[c.PhysicalID] = &CPU{
-				ID:        c.PhysicalID,
-				VendorID:  c.VendorID,
-				ModelName: c.ModelName,
-			}
-			coreBySocket[c.PhysicalID] = map[string]struct{}{}
-			threadsByCoreBySocket[c.PhysicalID] = map[string]int{}
-		}
-		if c.CoreID != "" {
-			coreBySocket[c.PhysicalID][c.PhysicalID+":"+c.CoreID] = struct{}{}
-			threadsByCoreBySocket[c.PhysicalID][c.PhysicalID+":"+c.CoreID]++
-		} else {
-			coreBySocket[c.PhysicalID][c.PhysicalID+":"+c.ID] = struct{}{}
-			threadsByCoreBySocket[c.PhysicalID][c.PhysicalID+":"+c.ID]++
-		}
-	}
-
-	// Tally up the values from the tracking maps
-	for id, s := range socketByID {
-		s.CoreCount = len(coreBySocket[id])
-		s.ThreadCount = 0
-
-		// This only works if HT is enabled, consider a more reliable model, maybe cache size comparisons?
-		efficiencyCoreCount := 0
-		for _, threads := range threadsByCoreBySocket[id] {
-			s.ThreadCount += threads
-			if threads == 1 {
-				efficiencyCoreCount++
-			}
-		}
-		if efficiencyCoreCount == s.CoreCount {
-			// 1:1 mapping means they're not actually efficiency cores, but regular cores
-			s.EfficiencyCoreCount = 0
-		} else {
-			s.EfficiencyCoreCount = efficiencyCoreCount
-		}
-	}
-	keys := make([]string, 0, len(socketByID))
-	result := make([]CPU, 0, len(socketByID))
-	for k := range socketByID {
-		keys = append(keys, k)
-	}
-	sort.Strings(keys)
-	for _, k := range keys {
-		result = append(result, *socketByID[k])
-	}
-	return result
-}
-
 func IsNUMA() bool {
 	ids := map[string]any{}
 	packageIds, _ := filepath.Glob("/sys/devices/system/cpu/cpu*/topology/physical_package_id")