simplify

Remove the --experimental flag requirement for creating image generation
models. Now ollama create automatically detects imagegen models by checking
for model_index.json in the model directory and routes them to the direct
creation path.

This simplifies the workflow for creating FLUX and other diffusion models -
users no longer need to pass --experimental.

.github/ISSUE_TEMPLATE/10_bug_report.yml

@@ -13,7 +13,7 @@ body:
     id: logs
     attributes:
       label: Relevant log output
-      description: Please copy and paste any relevant log output. See [Troubleshooting Guide](https://github.com/ollama/ollama/blob/main/docs/troubleshooting.md#how-to-troubleshoot-issues) for details.
+      description: Please copy and paste any relevant log output. See [Troubleshooting Guide](https://github.com/ollama/ollama/blob/main/docs/troubleshooting.mdx#how-to-troubleshoot-issues) for details.
       render: shell
     validations:
       required: false

.github/workflows/release.yaml

@@ -68,6 +68,7 @@ jobs:
           name: bundles-darwin
           path: |
             dist/*.tgz
+            dist/*.tar.zst
             dist/*.zip
             dist/*.dmg
 
@@ -371,13 +372,17 @@ jobs:
           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 ;;
+              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 }}.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 }}-${{ matrix.arch }}-rocm.tar.in ;;
@@ -392,13 +397,13 @@ jobs:
           done
       - run: |
           for ARCHIVE in dist/${{ matrix.os }}-${{ matrix.arch }}/*.tar.in; do
-            tar c -C dist/${{ matrix.os }}-${{ matrix.arch }} -T $ARCHIVE --owner 0 --group 0 | pigz -9vc >$(basename ${ARCHIVE//.*/}.tgz);
+            tar c -C dist/${{ 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: |
-            *.tgz
+            *.tar.zst
 
   # Build each Docker variant (OS, arch, and flavor) separately. Using QEMU is unreliable and slower.
   docker-build-push:
@@ -531,7 +536,7 @@ jobs:
       - name: Upload release artifacts
         run: |
           pids=()
-          for payload in dist/*.txt dist/*.zip dist/*.tgz dist/*.exe dist/*.dmg ; do
+          for payload in dist/*.txt dist/*.zip dist/*.tgz dist/*.tar.zst dist/*.exe dist/*.dmg ; do
             echo "Uploading $payload"
             gh release upload ${GITHUB_REF_NAME} $payload --clobber &
             pids[$!]=$!

CMakeLists.txt

@@ -2,6 +2,22 @@ cmake_minimum_required(VERSION 3.21)
 
 project(Ollama C CXX)
 
+# Handle cross-compilation on macOS: when CMAKE_OSX_ARCHITECTURES is set to a
+# single architecture different from the host, override CMAKE_SYSTEM_PROCESSOR
+# to match. This is necessary because CMAKE_SYSTEM_PROCESSOR defaults to the
+# host architecture, but downstream projects (like MLX) use it to detect the
+# target architecture.
+if(CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_OSX_ARCHITECTURES MATCHES ";")
+    # Single architecture specified
+    if(CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" AND NOT CMAKE_SYSTEM_PROCESSOR STREQUAL "x86_64")
+        message(STATUS "Cross-compiling for x86_64: overriding CMAKE_SYSTEM_PROCESSOR from ${CMAKE_SYSTEM_PROCESSOR} to x86_64")
+        set(CMAKE_SYSTEM_PROCESSOR "x86_64")
+    elseif(CMAKE_OSX_ARCHITECTURES STREQUAL "arm64" AND NOT CMAKE_SYSTEM_PROCESSOR STREQUAL "arm64")
+        message(STATUS "Cross-compiling for arm64: overriding CMAKE_SYSTEM_PROCESSOR from ${CMAKE_SYSTEM_PROCESSOR} to arm64")
+        set(CMAKE_SYSTEM_PROCESSOR "arm64")
+    endif()
+endif()
+
 include(CheckLanguage)
 include(GNUInstallDirs)
 
@@ -12,7 +28,7 @@ set(BUILD_SHARED_LIBS ON)
 
 set(CMAKE_CXX_STANDARD 17)
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
-set(CMAKE_CXX_EXTENSIONS OFF)
+set(CMAKE_CXX_EXTENSIONS ON) # Recent versions of MLX Requires gnu++17 extensions to compile properly
 
 set(GGML_BUILD ON)
 set(GGML_SHARED ON)
@@ -32,9 +48,10 @@ if((CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_OSX_ARCHITECTURES MATCHES "arm64")
     set(GGML_CPU_ALL_VARIANTS ON)
 endif()
 
-if (CMAKE_OSX_ARCHITECTURES MATCHES "x86_64")
+if(APPLE)
     set(CMAKE_BUILD_RPATH "@loader_path")
     set(CMAKE_INSTALL_RPATH "@loader_path")
+    set(CMAKE_BUILD_WITH_INSTALL_RPATH ON)
 endif()
 
 set(OLLAMA_BUILD_DIR ${CMAKE_BINARY_DIR}/lib/ollama)
@@ -147,14 +164,56 @@ if(CMAKE_HIP_COMPILER)
     endif()
 endif()
 
-find_package(Vulkan)
-if(Vulkan_FOUND)
-    add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-vulkan)
-    install(TARGETS ggml-vulkan
-        RUNTIME_DEPENDENCIES
-            PRE_INCLUDE_REGEXES vulkan
-            PRE_EXCLUDE_REGEXES ".*"
-        RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT Vulkan
-        LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT Vulkan
-    )
+if(NOT APPLE)
+    find_package(Vulkan)
+    if(Vulkan_FOUND)
+        add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-vulkan)
+        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/ml/backend/mlx)
+
+    # Find CUDA toolkit if MLX is built with CUDA support
+    find_package(CUDAToolkit)
+
+    install(TARGETS mlx mlxc
+        RUNTIME_DEPENDENCIES
+            DIRECTORIES ${CUDAToolkit_BIN_DIR} ${CUDAToolkit_BIN_DIR}/x64 ${CUDAToolkit_LIBRARY_DIR}
+            PRE_INCLUDE_REGEXES cublas cublasLt cudart nvrtc nvrtc-builtins cudnn nccl openblas gfortran
+            PRE_EXCLUDE_REGEXES ".*"
+        RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT MLX
+        LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT MLX
+        FRAMEWORK DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT MLX
+    )
+
+    # 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 ${OLLAMA_INSTALL_DIR}
+            COMPONENT MLX)
+    endif()
+
+    # Manually install cudart and cublas since they might not be picked up as direct dependencies
+    if(CUDAToolkit_FOUND)
+        file(GLOB CUDART_LIBS
+            "${CUDAToolkit_LIBRARY_DIR}/libcudart.so*"
+            "${CUDAToolkit_LIBRARY_DIR}/libcublas.so*")
+        if(CUDART_LIBS)
+            install(FILES ${CUDART_LIBS}
+                DESTINATION ${OLLAMA_INSTALL_DIR}
+                COMPONENT MLX)
+        endif()
+    endif()
+endif()

CMakePresets.json

@@ -41,7 +41,7 @@
       "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",
+        "CMAKE_CUDA_FLAGS": "-t 4",
         "OLLAMA_RUNNER_DIR": "cuda_v13"
       }
     },
@@ -83,6 +83,28 @@
       "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": {
+        "OLLAMA_RUNNER_DIR": "mlx_cuda_v13"
+      }
     }
   ],
   "buildPresets": [
@@ -140,6 +162,21 @@
       "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"
     }
   ]
 }

Dockerfile

@@ -32,7 +32,7 @@ ENV PATH=/${VULKANVERSION}/x86_64/bin:$PATH
 FROM --platform=linux/arm64 almalinux:8 AS base-arm64
 # install epel-release for ccache
 RUN yum install -y yum-utils epel-release \
-    && dnf install -y clang ccache \
+    && dnf install -y clang ccache git \
     && yum-config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/rhel8/sbsa/cuda-rhel8.repo
 ENV CC=clang CXX=clang++
 
@@ -131,8 +131,32 @@ COPY ml/backend/ggml/ggml ml/backend/ggml/ggml
 RUN --mount=type=cache,target=/root/.ccache \
     cmake --preset 'Vulkan' \
         && cmake --build --parallel --preset 'Vulkan' \
-        && cmake --install build --component Vulkan --strip --parallel 8 
+        && cmake --install build --component Vulkan --strip --parallel 8
 
+FROM base AS mlx
+ARG CUDA13VERSION=13.0
+RUN dnf install -y cuda-toolkit-${CUDA13VERSION//./-} \
+    && dnf install -y openblas-devel lapack-devel \
+    && dnf install -y libcudnn9-cuda-13 libcudnn9-devel-cuda-13 \
+    && dnf install -y libnccl libnccl-devel
+ENV PATH=/usr/local/cuda-13/bin:$PATH
+ENV BLAS_INCLUDE_DIRS=/usr/include/openblas
+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"
+ARG PARALLEL
+WORKDIR /go/src/github.com/ollama/ollama
+COPY CMakeLists.txt CMakePresets.json .
+COPY ml/backend/ggml/ggml ml/backend/ggml/ggml
+COPY x/ml/backend/mlx x/ml/backend/mlx
+COPY go.mod go.sum .
+COPY MLX_VERSION .
+RUN curl -fsSL https://golang.org/dl/go$(awk '/^go/ { print $2 }' go.mod).linux-$(case $(uname -m) in x86_64) echo amd64 ;; aarch64) echo arm64 ;; esac).tar.gz | tar xz -C /usr/local
+ENV PATH=/usr/local/go/bin:$PATH
+RUN go mod download
+RUN --mount=type=cache,target=/root/.ccache \
+    cmake --preset 'MLX CUDA 13' -DBLAS_INCLUDE_DIRS=/usr/include/openblas -DLAPACK_INCLUDE_DIRS=/usr/include/openblas \
+        && cmake --build --parallel ${PARALLEL} --preset 'MLX CUDA 13' \
+        && cmake --install build --component MLX --strip --parallel ${PARALLEL}
 
 FROM base AS build
 WORKDIR /go/src/github.com/ollama/ollama
@@ -141,18 +165,21 @@ RUN curl -fsSL https://golang.org/dl/go$(awk '/^go/ { print $2 }' go.mod).linux-
 ENV PATH=/usr/local/go/bin:$PATH
 RUN go mod download
 COPY . .
+# Clone mlx-c headers for CGO (version from MLX_VERSION file)
+RUN git clone --depth 1 --branch "$(cat MLX_VERSION)" https://github.com/ml-explore/mlx-c.git build/_deps/mlx-c-src
 ARG GOFLAGS="'-ldflags=-w -s'"
 ENV CGO_ENABLED=1
-ARG CGO_CFLAGS
+ENV CGO_CFLAGS="-I/go/src/github.com/ollama/ollama/build/_deps/mlx-c-src"
 ARG CGO_CXXFLAGS
 RUN --mount=type=cache,target=/root/.cache/go-build \
-    go build -trimpath -buildmode=pie -o /bin/ollama .
+    go build -tags mlx -trimpath -buildmode=pie -o /bin/ollama .
 
 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=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/
@@ -171,7 +198,7 @@ COPY --from=build /bin/ollama /bin/ollama
 
 FROM ubuntu:24.04
 RUN apt-get update \
-    && apt-get install -y ca-certificates libvulkan1 \
+    && apt-get install -y ca-certificates libvulkan1 libopenblas0 \
     && apt-get clean \
     && rm -rf /var/lib/apt/lists/*
 COPY --from=archive /bin /usr/bin

MLX_VERSION

@@ -0,0 +1 @@
+v0.4.1

README.md

@@ -48,7 +48,7 @@ ollama run gemma3
 
 ## Model library
 
-Ollama supports a list of models available on [ollama.com/library](https://ollama.com/library 'ollama model library')
+Ollama supports a list of models available on [ollama.com/library](https://ollama.com/library "ollama model library")
 
 Here are some example models that can be downloaded:
 
@@ -79,7 +79,7 @@ Here are some example models that can be downloaded:
 | Code Llama         | 7B         | 3.8GB | `ollama run codellama`           |
 | Llama 2 Uncensored | 7B         | 3.8GB | `ollama run llama2-uncensored`   |
 | LLaVA              | 7B         | 4.5GB | `ollama run llava`               |
-| Granite-3.3         | 8B         | 4.9GB | `ollama run granite3.3`          |
+| Granite-3.3        | 8B         | 4.9GB | `ollama run granite3.3`          |
 
 > [!NOTE]
 > You should have at least 8 GB of RAM available to run the 7B models, 16 GB to run the 13B models, and 32 GB to run the 33B models.
@@ -260,6 +260,38 @@ Finally, in a separate shell, run a model:
 ./ollama run llama3.2
 ```
 
+## Building with MLX (experimental)
+
+First build the MLX libraries:
+
+```shell
+cmake --preset MLX
+cmake --build --preset MLX --parallel
+cmake --install build --component MLX
+```
+
+When building with the `-tags mlx` flag, the main `ollama` binary includes MLX support for experimental features like image generation:
+
+```shell
+go build -tags mlx .
+```
+
+Finally, start the server:
+
+```
+./ollama serve
+```
+
+### Building MLX with CUDA
+
+When building with CUDA, use the preset "MLX CUDA 13" or "MLX CUDA 12" to enable CUDA with default architectures:
+
+```shell
+cmake --preset 'MLX CUDA 13'
+cmake --build --preset 'MLX CUDA 13' --parallel
+cmake --install build --component MLX
+```
+
 ## REST API
 
 Ollama has a REST API for running and managing models.
@@ -290,6 +322,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 
 ### Web & Desktop
 
+- [Onyx](https://github.com/onyx-dot-app/onyx)
 - [Open WebUI](https://github.com/open-webui/open-webui)
 - [SwiftChat (macOS with ReactNative)](https://github.com/aws-samples/swift-chat)
 - [Enchanted (macOS native)](https://github.com/AugustDev/enchanted)
@@ -421,7 +454,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [AppFlowy](https://github.com/AppFlowy-IO/AppFlowy) (AI collaborative workspace with Ollama, cross-platform and self-hostable)
 - [Lumina](https://github.com/cushydigit/lumina.git) (A lightweight, minimal React.js frontend for interacting with Ollama servers)
 - [Tiny Notepad](https://pypi.org/project/tiny-notepad) (A lightweight, notepad-like interface to chat with ollama available on PyPI)
-- [macLlama (macOS native)](https://github.com/hellotunamayo/macLlama) (A native macOS GUI application for interacting with Ollama models, featuring a chat interface.) 
+- [macLlama (macOS native)](https://github.com/hellotunamayo/macLlama) (A native macOS GUI application for interacting with Ollama models, featuring a chat interface.)
 - [GPTranslate](https://github.com/philberndt/GPTranslate) (A fast and lightweight, AI powered desktop translation application written with Rust and Tauri. Features real-time translation with OpenAI/Azure/Ollama.)
 - [ollama launcher](https://github.com/NGC13009/ollama-launcher) (A launcher for Ollama, aiming to provide users with convenient functions such as ollama server launching, management, or configuration.)
 - [ai-hub](https://github.com/Aj-Seven/ai-hub) (AI Hub supports multiple models via API keys and Chat support via Ollama API.)
@@ -493,7 +526,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 ### Database
 
 - [pgai](https://github.com/timescale/pgai) - PostgreSQL as a vector database (Create and search embeddings from Ollama models using pgvector)
-   - [Get started guide](https://github.com/timescale/pgai/blob/main/docs/vectorizer-quick-start.md)
+  - [Get started guide](https://github.com/timescale/pgai/blob/main/docs/vectorizer-quick-start.md)
 - [MindsDB](https://github.com/mindsdb/mindsdb/blob/staging/mindsdb/integrations/handlers/ollama_handler/README.md) (Connects Ollama models with nearly 200 data platforms and apps)
 - [chromem-go](https://github.com/philippgille/chromem-go/blob/v0.5.0/embed_ollama.go) with [example](https://github.com/philippgille/chromem-go/tree/v0.5.0/examples/rag-wikipedia-ollama)
 - [Kangaroo](https://github.com/dbkangaroo/kangaroo) (AI-powered SQL client and admin tool for popular databases)
@@ -636,6 +669,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [llama.cpp](https://github.com/ggml-org/llama.cpp) project founded by Georgi Gerganov.
 
 ### Observability
+
 - [Opik](https://www.comet.com/docs/opik/cookbook/ollama) is an open-source platform to debug, evaluate, and monitor your LLM applications, RAG systems, and agentic workflows with comprehensive tracing, automated evaluations, and production-ready dashboards. Opik supports native integration to Ollama.
 - [Lunary](https://lunary.ai/docs/integrations/ollama) is the leading open-source LLM observability platform. It provides a variety of enterprise-grade features such as real-time analytics, prompt templates management, PII masking, and comprehensive agent tracing.
 - [OpenLIT](https://github.com/openlit/openlit) is an OpenTelemetry-native tool for monitoring Ollama Applications & GPUs using traces and metrics.
@@ -644,4 +678,5 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [MLflow Tracing](https://mlflow.org/docs/latest/llms/tracing/index.html#automatic-tracing) is an open source LLM observability tool with a convenient API to log and visualize traces, making it easy to debug and evaluate GenAI applications.
 
 ### Security
+
 - [Ollama Fortress](https://github.com/ParisNeo/ollama_proxy_server)

anthropic/anthropic.go

@@ -0,0 +1,778 @@
+package anthropic
+
+import (
+	"crypto/rand"
+	"encoding/base64"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"log/slog"
+	"net/http"
+	"strings"
+	"time"
+
+	"github.com/ollama/ollama/api"
+)
+
+// Error types matching Anthropic API
+type Error struct {
+	Type    string `json:"type"`
+	Message string `json:"message"`
+}
+
+type ErrorResponse struct {
+	Type      string `json:"type"` // always "error"
+	Error     Error  `json:"error"`
+	RequestID string `json:"request_id,omitempty"`
+}
+
+// NewError creates a new ErrorResponse with the appropriate error type based on HTTP status code
+func NewError(code int, message string) ErrorResponse {
+	var etype string
+	switch code {
+	case http.StatusBadRequest:
+		etype = "invalid_request_error"
+	case http.StatusUnauthorized:
+		etype = "authentication_error"
+	case http.StatusForbidden:
+		etype = "permission_error"
+	case http.StatusNotFound:
+		etype = "not_found_error"
+	case http.StatusTooManyRequests:
+		etype = "rate_limit_error"
+	case http.StatusServiceUnavailable, 529:
+		etype = "overloaded_error"
+	default:
+		etype = "api_error"
+	}
+
+	return ErrorResponse{
+		Type:      "error",
+		Error:     Error{Type: etype, Message: message},
+		RequestID: generateID("req"),
+	}
+}
+
+// Request types
+
+// MessagesRequest represents an Anthropic Messages API request
+type MessagesRequest struct {
+	Model         string          `json:"model"`
+	MaxTokens     int             `json:"max_tokens"`
+	Messages      []MessageParam  `json:"messages"`
+	System        any             `json:"system,omitempty"` // string or []ContentBlock
+	Stream        bool            `json:"stream,omitempty"`
+	Temperature   *float64        `json:"temperature,omitempty"`
+	TopP          *float64        `json:"top_p,omitempty"`
+	TopK          *int            `json:"top_k,omitempty"`
+	StopSequences []string        `json:"stop_sequences,omitempty"`
+	Tools         []Tool          `json:"tools,omitempty"`
+	ToolChoice    *ToolChoice     `json:"tool_choice,omitempty"`
+	Thinking      *ThinkingConfig `json:"thinking,omitempty"`
+	Metadata      *Metadata       `json:"metadata,omitempty"`
+}
+
+// MessageParam represents a message in the request
+type MessageParam struct {
+	Role    string `json:"role"`    // "user" or "assistant"
+	Content any    `json:"content"` // string or []ContentBlock
+}
+
+// ContentBlock represents a content block in a message.
+// Text and Thinking use pointers so they serialize as the field being present (even if empty)
+// only when set, which is required for SDK streaming accumulation.
+type ContentBlock struct {
+	Type string `json:"type"` // text, image, tool_use, tool_result, thinking
+
+	// For text blocks - pointer so field only appears when set (SDK requires it for accumulation)
+	Text *string `json:"text,omitempty"`
+
+	// For image blocks
+	Source *ImageSource `json:"source,omitempty"`
+
+	// For tool_use blocks
+	ID    string `json:"id,omitempty"`
+	Name  string `json:"name,omitempty"`
+	Input any    `json:"input,omitempty"`
+
+	// For tool_result blocks
+	ToolUseID string `json:"tool_use_id,omitempty"`
+	Content   any    `json:"content,omitempty"` // string or []ContentBlock
+	IsError   bool   `json:"is_error,omitempty"`
+
+	// For thinking blocks - pointer so field only appears when set (SDK requires it for accumulation)
+	Thinking  *string `json:"thinking,omitempty"`
+	Signature string  `json:"signature,omitempty"`
+}
+
+// ImageSource represents the source of an image
+type ImageSource struct {
+	Type      string `json:"type"` // "base64" or "url"
+	MediaType string `json:"media_type,omitempty"`
+	Data      string `json:"data,omitempty"`
+	URL       string `json:"url,omitempty"`
+}
+
+// Tool represents a tool definition
+type Tool struct {
+	Type        string          `json:"type,omitempty"` // "custom" for user-defined tools
+	Name        string          `json:"name"`
+	Description string          `json:"description,omitempty"`
+	InputSchema json.RawMessage `json:"input_schema,omitempty"`
+}
+
+// ToolChoice controls how the model uses tools
+type ToolChoice struct {
+	Type                   string `json:"type"` // "auto", "any", "tool", "none"
+	Name                   string `json:"name,omitempty"`
+	DisableParallelToolUse bool   `json:"disable_parallel_tool_use,omitempty"`
+}
+
+// ThinkingConfig controls extended thinking
+type ThinkingConfig struct {
+	Type         string `json:"type"` // "enabled" or "disabled"
+	BudgetTokens int    `json:"budget_tokens,omitempty"`
+}
+
+// Metadata for the request
+type Metadata struct {
+	UserID string `json:"user_id,omitempty"`
+}
+
+// Response types
+
+// MessagesResponse represents an Anthropic Messages API response
+type MessagesResponse struct {
+	ID           string         `json:"id"`
+	Type         string         `json:"type"` // "message"
+	Role         string         `json:"role"` // "assistant"
+	Model        string         `json:"model"`
+	Content      []ContentBlock `json:"content"`
+	StopReason   string         `json:"stop_reason,omitempty"`
+	StopSequence string         `json:"stop_sequence,omitempty"`
+	Usage        Usage          `json:"usage"`
+}
+
+// Usage contains token usage information
+type Usage struct {
+	InputTokens  int `json:"input_tokens"`
+	OutputTokens int `json:"output_tokens"`
+}
+
+// Streaming event types
+
+// MessageStartEvent is sent at the start of streaming
+type MessageStartEvent struct {
+	Type    string           `json:"type"` // "message_start"
+	Message MessagesResponse `json:"message"`
+}
+
+// ContentBlockStartEvent signals the start of a content block
+type ContentBlockStartEvent struct {
+	Type         string       `json:"type"` // "content_block_start"
+	Index        int          `json:"index"`
+	ContentBlock ContentBlock `json:"content_block"`
+}
+
+// ContentBlockDeltaEvent contains incremental content updates
+type ContentBlockDeltaEvent struct {
+	Type  string `json:"type"` // "content_block_delta"
+	Index int    `json:"index"`
+	Delta Delta  `json:"delta"`
+}
+
+// Delta represents an incremental update
+type Delta struct {
+	Type        string `json:"type"` // "text_delta", "input_json_delta", "thinking_delta", "signature_delta"
+	Text        string `json:"text,omitempty"`
+	PartialJSON string `json:"partial_json,omitempty"`
+	Thinking    string `json:"thinking,omitempty"`
+	Signature   string `json:"signature,omitempty"`
+}
+
+// ContentBlockStopEvent signals the end of a content block
+type ContentBlockStopEvent struct {
+	Type  string `json:"type"` // "content_block_stop"
+	Index int    `json:"index"`
+}
+
+// MessageDeltaEvent contains updates to the message
+type MessageDeltaEvent struct {
+	Type  string       `json:"type"` // "message_delta"
+	Delta MessageDelta `json:"delta"`
+	Usage DeltaUsage   `json:"usage"`
+}
+
+// MessageDelta contains stop information
+type MessageDelta struct {
+	StopReason   string `json:"stop_reason,omitempty"`
+	StopSequence string `json:"stop_sequence,omitempty"`
+}
+
+// DeltaUsage contains cumulative token usage
+type DeltaUsage struct {
+	OutputTokens int `json:"output_tokens"`
+}
+
+// MessageStopEvent signals the end of the message
+type MessageStopEvent struct {
+	Type string `json:"type"` // "message_stop"
+}
+
+// PingEvent is a keepalive event
+type PingEvent struct {
+	Type string `json:"type"` // "ping"
+}
+
+// StreamErrorEvent is an error during streaming
+type StreamErrorEvent struct {
+	Type  string `json:"type"` // "error"
+	Error Error  `json:"error"`
+}
+
+// FromMessagesRequest converts an Anthropic MessagesRequest to an Ollama api.ChatRequest
+func FromMessagesRequest(r MessagesRequest) (*api.ChatRequest, error) {
+	var messages []api.Message
+
+	if r.System != nil {
+		switch sys := r.System.(type) {
+		case string:
+			if sys != "" {
+				messages = append(messages, api.Message{Role: "system", Content: sys})
+			}
+		case []any:
+			// System can be an array of content blocks
+			var content strings.Builder
+			for _, block := range sys {
+				if blockMap, ok := block.(map[string]any); ok {
+					if blockMap["type"] == "text" {
+						if text, ok := blockMap["text"].(string); ok {
+							content.WriteString(text)
+						}
+					}
+				}
+			}
+			if content.Len() > 0 {
+				messages = append(messages, api.Message{Role: "system", Content: content.String()})
+			}
+		}
+	}
+
+	for _, msg := range r.Messages {
+		converted, err := convertMessage(msg)
+		if err != nil {
+			return nil, err
+		}
+		messages = append(messages, converted...)
+	}
+
+	options := make(map[string]any)
+
+	options["num_predict"] = r.MaxTokens
+
+	if r.Temperature != nil {
+		options["temperature"] = *r.Temperature
+	}
+
+	if r.TopP != nil {
+		options["top_p"] = *r.TopP
+	}
+
+	if r.TopK != nil {
+		options["top_k"] = *r.TopK
+	}
+
+	if len(r.StopSequences) > 0 {
+		options["stop"] = r.StopSequences
+	}
+
+	var tools api.Tools
+	for _, t := range r.Tools {
+		tool, err := convertTool(t)
+		if err != nil {
+			return nil, err
+		}
+		tools = append(tools, tool)
+	}
+
+	var think *api.ThinkValue
+	if r.Thinking != nil && r.Thinking.Type == "enabled" {
+		think = &api.ThinkValue{Value: true}
+	}
+
+	stream := r.Stream
+
+	return &api.ChatRequest{
+		Model:    r.Model,
+		Messages: messages,
+		Options:  options,
+		Stream:   &stream,
+		Tools:    tools,
+		Think:    think,
+	}, nil
+}
+
+// convertMessage converts an Anthropic MessageParam to Ollama api.Message(s)
+func convertMessage(msg MessageParam) ([]api.Message, error) {
+	var messages []api.Message
+	role := strings.ToLower(msg.Role)
+
+	switch content := msg.Content.(type) {
+	case string:
+		messages = append(messages, api.Message{Role: role, Content: content})
+
+	case []any:
+		var textContent strings.Builder
+		var images []api.ImageData
+		var toolCalls []api.ToolCall
+		var thinking string
+		var toolResults []api.Message
+
+		for _, block := range content {
+			blockMap, ok := block.(map[string]any)
+			if !ok {
+				return nil, errors.New("invalid content block format")
+			}
+
+			blockType, _ := blockMap["type"].(string)
+
+			switch blockType {
+			case "text":
+				if text, ok := blockMap["text"].(string); ok {
+					textContent.WriteString(text)
+				}
+
+			case "image":
+				source, ok := blockMap["source"].(map[string]any)
+				if !ok {
+					return nil, errors.New("invalid image source")
+				}
+
+				sourceType, _ := source["type"].(string)
+				if sourceType == "base64" {
+					data, _ := source["data"].(string)
+					decoded, err := base64.StdEncoding.DecodeString(data)
+					if err != nil {
+						return nil, fmt.Errorf("invalid base64 image data: %w", err)
+					}
+					images = append(images, decoded)
+				} else {
+					return nil, fmt.Errorf("invalid image source type: %s. Only base64 images are supported.", sourceType)
+				}
+				// URL images would need to be fetched - skip for now
+
+			case "tool_use":
+				id, ok := blockMap["id"].(string)
+				if !ok {
+					return nil, errors.New("tool_use block missing required 'id' field")
+				}
+				name, ok := blockMap["name"].(string)
+				if !ok {
+					return nil, errors.New("tool_use block missing required 'name' field")
+				}
+				tc := api.ToolCall{
+					ID: id,
+					Function: api.ToolCallFunction{
+						Name: name,
+					},
+				}
+				if input, ok := blockMap["input"].(map[string]any); ok {
+					tc.Function.Arguments = mapToArgs(input)
+				}
+				toolCalls = append(toolCalls, tc)
+
+			case "tool_result":
+				toolUseID, _ := blockMap["tool_use_id"].(string)
+				var resultContent string
+
+				switch c := blockMap["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
+								}
+							}
+						}
+					}
+				}
+
+				toolResults = append(toolResults, api.Message{
+					Role:       "tool",
+					Content:    resultContent,
+					ToolCallID: toolUseID,
+				})
+
+			case "thinking":
+				if t, ok := blockMap["thinking"].(string); ok {
+					thinking = t
+				}
+			}
+		}
+
+		if textContent.Len() > 0 || len(images) > 0 || len(toolCalls) > 0 || thinking != "" {
+			m := api.Message{
+				Role:      role,
+				Content:   textContent.String(),
+				Images:    images,
+				ToolCalls: toolCalls,
+				Thinking:  thinking,
+			}
+			messages = append(messages, m)
+		}
+
+		// Add tool results as separate messages
+		messages = append(messages, toolResults...)
+
+	default:
+		return nil, fmt.Errorf("invalid message content type: %T", content)
+	}
+
+	return messages, nil
+}
+
+// convertTool converts an Anthropic Tool to an Ollama api.Tool
+func convertTool(t Tool) (api.Tool, error) {
+	var params api.ToolFunctionParameters
+	if len(t.InputSchema) > 0 {
+		if err := json.Unmarshal(t.InputSchema, &params); err != nil {
+			return api.Tool{}, fmt.Errorf("invalid input_schema for tool %q: %w", t.Name, err)
+		}
+	}
+
+	return api.Tool{
+		Type: "function",
+		Function: api.ToolFunction{
+			Name:        t.Name,
+			Description: t.Description,
+			Parameters:  params,
+		},
+	}, nil
+}
+
+// ToMessagesResponse converts an Ollama api.ChatResponse to an Anthropic MessagesResponse
+func ToMessagesResponse(id string, r api.ChatResponse) MessagesResponse {
+	var content []ContentBlock
+
+	if r.Message.Thinking != "" {
+		content = append(content, ContentBlock{
+			Type:     "thinking",
+			Thinking: ptr(r.Message.Thinking),
+		})
+	}
+
+	if r.Message.Content != "" {
+		content = append(content, ContentBlock{
+			Type: "text",
+			Text: ptr(r.Message.Content),
+		})
+	}
+
+	for _, tc := range r.Message.ToolCalls {
+		content = append(content, ContentBlock{
+			Type:  "tool_use",
+			ID:    tc.ID,
+			Name:  tc.Function.Name,
+			Input: tc.Function.Arguments,
+		})
+	}
+
+	stopReason := mapStopReason(r.DoneReason, len(r.Message.ToolCalls) > 0)
+
+	return MessagesResponse{
+		ID:         id,
+		Type:       "message",
+		Role:       "assistant",
+		Model:      r.Model,
+		Content:    content,
+		StopReason: stopReason,
+		Usage: Usage{
+			InputTokens:  r.Metrics.PromptEvalCount,
+			OutputTokens: r.Metrics.EvalCount,
+		},
+	}
+}
+
+// mapStopReason converts Ollama done_reason to Anthropic stop_reason
+func mapStopReason(reason string, hasToolCalls bool) string {
+	if hasToolCalls {
+		return "tool_use"
+	}
+
+	switch reason {
+	case "stop":
+		return "end_turn"
+	case "length":
+		return "max_tokens"
+	default:
+		if reason != "" {
+			return "stop_sequence"
+		}
+		return ""
+	}
+}
+
+// StreamConverter manages state for converting Ollama streaming responses to Anthropic format
+type StreamConverter struct {
+	ID              string
+	Model           string
+	firstWrite      bool
+	contentIndex    int
+	inputTokens     int
+	outputTokens    int
+	thinkingStarted bool
+	thinkingDone    bool
+	textStarted     bool
+	toolCallsSent   map[string]bool
+}
+
+func NewStreamConverter(id, model string) *StreamConverter {
+	return &StreamConverter{
+		ID:            id,
+		Model:         model,
+		firstWrite:    true,
+		toolCallsSent: make(map[string]bool),
+	}
+}
+
+// StreamEvent represents a streaming event to be sent to the client
+type StreamEvent struct {
+	Event string
+	Data  any
+}
+
+// Process converts an Ollama ChatResponse to Anthropic streaming events
+func (c *StreamConverter) Process(r api.ChatResponse) []StreamEvent {
+	var events []StreamEvent
+
+	if c.firstWrite {
+		c.firstWrite = false
+		c.inputTokens = r.Metrics.PromptEvalCount
+
+		events = append(events, StreamEvent{
+			Event: "message_start",
+			Data: MessageStartEvent{
+				Type: "message_start",
+				Message: MessagesResponse{
+					ID:      c.ID,
+					Type:    "message",
+					Role:    "assistant",
+					Model:   c.Model,
+					Content: []ContentBlock{},
+					Usage: Usage{
+						InputTokens:  c.inputTokens,
+						OutputTokens: 0,
+					},
+				},
+			},
+		})
+	}
+
+	if r.Message.Thinking != "" && !c.thinkingDone {
+		if !c.thinkingStarted {
+			c.thinkingStarted = true
+			events = append(events, StreamEvent{
+				Event: "content_block_start",
+				Data: ContentBlockStartEvent{
+					Type:  "content_block_start",
+					Index: c.contentIndex,
+					ContentBlock: ContentBlock{
+						Type:     "thinking",
+						Thinking: ptr(""),
+					},
+				},
+			})
+		}
+
+		events = append(events, StreamEvent{
+			Event: "content_block_delta",
+			Data: ContentBlockDeltaEvent{
+				Type:  "content_block_delta",
+				Index: c.contentIndex,
+				Delta: Delta{
+					Type:     "thinking_delta",
+					Thinking: r.Message.Thinking,
+				},
+			},
+		})
+	}
+
+	if r.Message.Content != "" {
+		if c.thinkingStarted && !c.thinkingDone {
+			c.thinkingDone = true
+			events = append(events, StreamEvent{
+				Event: "content_block_stop",
+				Data: ContentBlockStopEvent{
+					Type:  "content_block_stop",
+					Index: c.contentIndex,
+				},
+			})
+			c.contentIndex++
+		}
+
+		if !c.textStarted {
+			c.textStarted = true
+			events = append(events, StreamEvent{
+				Event: "content_block_start",
+				Data: ContentBlockStartEvent{
+					Type:  "content_block_start",
+					Index: c.contentIndex,
+					ContentBlock: ContentBlock{
+						Type: "text",
+						Text: ptr(""),
+					},
+				},
+			})
+		}
+
+		events = append(events, StreamEvent{
+			Event: "content_block_delta",
+			Data: ContentBlockDeltaEvent{
+				Type:  "content_block_delta",
+				Index: c.contentIndex,
+				Delta: Delta{
+					Type: "text_delta",
+					Text: r.Message.Content,
+				},
+			},
+		})
+	}
+
+	for _, tc := range r.Message.ToolCalls {
+		if c.toolCallsSent[tc.ID] {
+			continue
+		}
+
+		if c.textStarted {
+			events = append(events, StreamEvent{
+				Event: "content_block_stop",
+				Data: ContentBlockStopEvent{
+					Type:  "content_block_stop",
+					Index: c.contentIndex,
+				},
+			})
+			c.contentIndex++
+			c.textStarted = false
+		}
+
+		argsJSON, err := json.Marshal(tc.Function.Arguments)
+		if err != nil {
+			slog.Error("failed to marshal tool arguments", "error", err, "tool_id", tc.ID)
+			continue
+		}
+
+		events = append(events, StreamEvent{
+			Event: "content_block_start",
+			Data: ContentBlockStartEvent{
+				Type:  "content_block_start",
+				Index: c.contentIndex,
+				ContentBlock: ContentBlock{
+					Type:  "tool_use",
+					ID:    tc.ID,
+					Name:  tc.Function.Name,
+					Input: map[string]any{},
+				},
+			},
+		})
+
+		events = append(events, StreamEvent{
+			Event: "content_block_delta",
+			Data: ContentBlockDeltaEvent{
+				Type:  "content_block_delta",
+				Index: c.contentIndex,
+				Delta: Delta{
+					Type:        "input_json_delta",
+					PartialJSON: string(argsJSON),
+				},
+			},
+		})
+
+		events = append(events, StreamEvent{
+			Event: "content_block_stop",
+			Data: ContentBlockStopEvent{
+				Type:  "content_block_stop",
+				Index: c.contentIndex,
+			},
+		})
+
+		c.toolCallsSent[tc.ID] = true
+		c.contentIndex++
+	}
+
+	if r.Done {
+		if c.textStarted {
+			events = append(events, StreamEvent{
+				Event: "content_block_stop",
+				Data: ContentBlockStopEvent{
+					Type:  "content_block_stop",
+					Index: c.contentIndex,
+				},
+			})
+		} else if c.thinkingStarted && !c.thinkingDone {
+			events = append(events, StreamEvent{
+				Event: "content_block_stop",
+				Data: ContentBlockStopEvent{
+					Type:  "content_block_stop",
+					Index: c.contentIndex,
+				},
+			})
+		}
+
+		c.outputTokens = r.Metrics.EvalCount
+		stopReason := mapStopReason(r.DoneReason, len(c.toolCallsSent) > 0)
+
+		events = append(events, StreamEvent{
+			Event: "message_delta",
+			Data: MessageDeltaEvent{
+				Type: "message_delta",
+				Delta: MessageDelta{
+					StopReason: stopReason,
+				},
+				Usage: DeltaUsage{
+					OutputTokens: c.outputTokens,
+				},
+			},
+		})
+
+		events = append(events, StreamEvent{
+			Event: "message_stop",
+			Data: MessageStopEvent{
+				Type: "message_stop",
+			},
+		})
+	}
+
+	return events
+}
+
+// generateID generates a unique ID with the given prefix using crypto/rand
+func generateID(prefix string) string {
+	b := make([]byte, 12)
+	if _, err := rand.Read(b); err != nil {
+		// Fallback to time-based ID if crypto/rand fails
+		return fmt.Sprintf("%s_%d", prefix, time.Now().UnixNano())
+	}
+	return fmt.Sprintf("%s_%x", prefix, b)
+}
+
+// GenerateMessageID generates a unique message ID
+func GenerateMessageID() string {
+	return generateID("msg")
+}
+
+// ptr returns a pointer to the given string value
+func ptr(s string) *string {
+	return &s
+}
+
+// mapToArgs converts a map to ToolCallFunctionArguments
+func mapToArgs(m map[string]any) api.ToolCallFunctionArguments {
+	args := api.NewToolCallFunctionArguments()
+	for k, v := range m {
+		args.Set(k, v)
+	}
+	return args
+}

anthropic/anthropic_test.go

@@ -0,0 +1,953 @@
+package anthropic
+
+import (
+	"encoding/base64"
+	"encoding/json"
+	"testing"
+
+	"github.com/google/go-cmp/cmp"
+
+	"github.com/ollama/ollama/api"
+)
+
+const (
+	testImage = `iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAQAAAC1HAwCAAAAC0lEQVR42mNk+A8AAQUBAScY42YAAAAASUVORK5CYII=`
+)
+
+// testArgs creates ToolCallFunctionArguments from a map (convenience function for tests)
+func testArgs(m map[string]any) api.ToolCallFunctionArguments {
+	args := api.NewToolCallFunctionArguments()
+	for k, v := range m {
+		args.Set(k, v)
+	}
+	return args
+}
+
+func TestFromMessagesRequest_Basic(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		Messages: []MessageParam{
+			{Role: "user", Content: "Hello"},
+		},
+	}
+
+	result, err := FromMessagesRequest(req)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if result.Model != "test-model" {
+		t.Errorf("expected model 'test-model', got %q", result.Model)
+	}
+
+	if len(result.Messages) != 1 {
+		t.Fatalf("expected 1 message, got %d", len(result.Messages))
+	}
+
+	if result.Messages[0].Role != "user" || result.Messages[0].Content != "Hello" {
+		t.Errorf("unexpected message: %+v", result.Messages[0])
+	}
+
+	if numPredict, ok := result.Options["num_predict"].(int); !ok || numPredict != 1024 {
+		t.Errorf("expected num_predict 1024, got %v", result.Options["num_predict"])
+	}
+}
+
+func TestFromMessagesRequest_WithSystemPrompt(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		System:    "You are a helpful assistant.",
+		Messages: []MessageParam{
+			{Role: "user", Content: "Hello"},
+		},
+	}
+
+	result, err := FromMessagesRequest(req)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if len(result.Messages) != 2 {
+		t.Fatalf("expected 2 messages, got %d", len(result.Messages))
+	}
+
+	if result.Messages[0].Role != "system" || result.Messages[0].Content != "You are a helpful assistant." {
+		t.Errorf("unexpected system message: %+v", result.Messages[0])
+	}
+}
+
+func TestFromMessagesRequest_WithSystemPromptArray(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		System: []any{
+			map[string]any{"type": "text", "text": "You are helpful."},
+			map[string]any{"type": "text", "text": " Be concise."},
+		},
+		Messages: []MessageParam{
+			{Role: "user", Content: "Hello"},
+		},
+	}
+
+	result, err := FromMessagesRequest(req)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if len(result.Messages) != 2 {
+		t.Fatalf("expected 2 messages, got %d", len(result.Messages))
+	}
+
+	if result.Messages[0].Content != "You are helpful. Be concise." {
+		t.Errorf("unexpected system message content: %q", result.Messages[0].Content)
+	}
+}
+
+func TestFromMessagesRequest_WithOptions(t *testing.T) {
+	temp := 0.7
+	topP := 0.9
+	topK := 40
+	req := MessagesRequest{
+		Model:         "test-model",
+		MaxTokens:     2048,
+		Messages:      []MessageParam{{Role: "user", Content: "Hello"}},
+		Temperature:   &temp,
+		TopP:          &topP,
+		TopK:          &topK,
+		StopSequences: []string{"\n", "END"},
+	}
+
+	result, err := FromMessagesRequest(req)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if result.Options["temperature"] != 0.7 {
+		t.Errorf("expected temperature 0.7, got %v", result.Options["temperature"])
+	}
+	if result.Options["top_p"] != 0.9 {
+		t.Errorf("expected top_p 0.9, got %v", result.Options["top_p"])
+	}
+	if result.Options["top_k"] != 40 {
+		t.Errorf("expected top_k 40, got %v", result.Options["top_k"])
+	}
+	if diff := cmp.Diff([]string{"\n", "END"}, result.Options["stop"]); diff != "" {
+		t.Errorf("stop sequences mismatch: %s", diff)
+	}
+}
+
+func TestFromMessagesRequest_WithImage(t *testing.T) {
+	imgData, _ := base64.StdEncoding.DecodeString(testImage)
+
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		Messages: []MessageParam{
+			{
+				Role: "user",
+				Content: []any{
+					map[string]any{"type": "text", "text": "What's in this 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))
+	}
+
+	if result.Messages[0].Content != "What's in this image?" {
+		t.Errorf("expected content 'What's in this image?', got %q", result.Messages[0].Content)
+	}
+
+	if len(result.Messages[0].Images) != 1 {
+		t.Fatalf("expected 1 image, got %d", len(result.Messages[0].Images))
+	}
+
+	if string(result.Messages[0].Images[0]) != string(imgData) {
+		t.Error("image data mismatch")
+	}
+}
+
+func TestFromMessagesRequest_WithToolUse(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		Messages: []MessageParam{
+			{Role: "user", Content: "What's the weather in Paris?"},
+			{
+				Role: "assistant",
+				Content: []any{
+					map[string]any{
+						"type":  "tool_use",
+						"id":    "call_123",
+						"name":  "get_weather",
+						"input": map[string]any{"location": "Paris"},
+					},
+				},
+			},
+		},
+	}
+
+	result, err := FromMessagesRequest(req)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if len(result.Messages) != 2 {
+		t.Fatalf("expected 2 messages, got %d", len(result.Messages))
+	}
+
+	if len(result.Messages[1].ToolCalls) != 1 {
+		t.Fatalf("expected 1 tool call, got %d", len(result.Messages[1].ToolCalls))
+	}
+
+	tc := result.Messages[1].ToolCalls[0]
+	if tc.ID != "call_123" {
+		t.Errorf("expected tool call ID 'call_123', got %q", tc.ID)
+	}
+	if tc.Function.Name != "get_weather" {
+		t.Errorf("expected tool name 'get_weather', got %q", tc.Function.Name)
+	}
+}
+
+func TestFromMessagesRequest_WithToolResult(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		Messages: []MessageParam{
+			{
+				Role: "user",
+				Content: []any{
+					map[string]any{
+						"type":        "tool_result",
+						"tool_use_id": "call_123",
+						"content":     "The weather in Paris is sunny, 22°C",
+					},
+				},
+			},
+		},
+	}
+
+	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_123" {
+		t.Errorf("expected tool_call_id 'call_123', got %q", msg.ToolCallID)
+	}
+	if msg.Content != "The weather in Paris is sunny, 22°C" {
+		t.Errorf("unexpected content: %q", msg.Content)
+	}
+}
+
+func TestFromMessagesRequest_WithTools(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		Messages:  []MessageParam{{Role: "user", Content: "Hello"}},
+		Tools: []Tool{
+			{
+				Name:        "get_weather",
+				Description: "Get current weather",
+				InputSchema: json.RawMessage(`{"type":"object","properties":{"location":{"type":"string"}},"required":["location"]}`),
+			},
+		},
+	}
+
+	result, err := FromMessagesRequest(req)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if len(result.Tools) != 1 {
+		t.Fatalf("expected 1 tool, got %d", len(result.Tools))
+	}
+
+	tool := result.Tools[0]
+	if tool.Type != "function" {
+		t.Errorf("expected type 'function', got %q", tool.Type)
+	}
+	if tool.Function.Name != "get_weather" {
+		t.Errorf("expected name 'get_weather', got %q", tool.Function.Name)
+	}
+	if tool.Function.Description != "Get current weather" {
+		t.Errorf("expected description 'Get current weather', got %q", tool.Function.Description)
+	}
+}
+
+func TestFromMessagesRequest_WithThinking(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		Messages:  []MessageParam{{Role: "user", Content: "Hello"}},
+		Thinking:  &ThinkingConfig{Type: "enabled", BudgetTokens: 1000},
+	}
+
+	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 v, ok := result.Think.Value.(bool); !ok || !v {
+		t.Errorf("expected Think.Value to be true, got %v", result.Think.Value)
+	}
+}
+
+// TestFromMessagesRequest_ThinkingOnlyBlock verifies that messages containing only
+// a thinking block (no text, images, or tool calls) are preserved and not dropped.
+func TestFromMessagesRequest_ThinkingOnlyBlock(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		Messages: []MessageParam{
+			{Role: "user", Content: "Hello"},
+			{
+				Role: "assistant",
+				Content: []any{
+					map[string]any{
+						"type":     "thinking",
+						"thinking": "Let me think about this...",
+					},
+				},
+			},
+		},
+	}
+
+	result, err := FromMessagesRequest(req)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if len(result.Messages) != 2 {
+		t.Fatalf("expected 2 messages, got %d", len(result.Messages))
+	}
+
+	assistantMsg := result.Messages[1]
+	if assistantMsg.Thinking != "Let me think about this..." {
+		t.Errorf("expected thinking content, got %q", assistantMsg.Thinking)
+	}
+}
+
+func TestFromMessagesRequest_ToolUseMissingID(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		Messages: []MessageParam{
+			{
+				Role: "assistant",
+				Content: []any{
+					map[string]any{
+						"type": "tool_use",
+						"name": "get_weather",
+					},
+				},
+			},
+		},
+	}
+
+	_, err := FromMessagesRequest(req)
+	if err == nil {
+		t.Fatal("expected error for missing tool_use id")
+	}
+	if err.Error() != "tool_use block missing required 'id' field" {
+		t.Errorf("unexpected error message: %v", err)
+	}
+}
+
+func TestFromMessagesRequest_ToolUseMissingName(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		Messages: []MessageParam{
+			{
+				Role: "assistant",
+				Content: []any{
+					map[string]any{
+						"type": "tool_use",
+						"id":   "call_123",
+					},
+				},
+			},
+		},
+	}
+
+	_, err := FromMessagesRequest(req)
+	if err == nil {
+		t.Fatal("expected error for missing tool_use name")
+	}
+	if err.Error() != "tool_use block missing required 'name' field" {
+		t.Errorf("unexpected error message: %v", err)
+	}
+}
+
+func TestFromMessagesRequest_InvalidToolSchema(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		Messages:  []MessageParam{{Role: "user", Content: "Hello"}},
+		Tools: []Tool{
+			{
+				Name:        "bad_tool",
+				InputSchema: json.RawMessage(`{invalid json`),
+			},
+		},
+	}
+
+	_, err := FromMessagesRequest(req)
+	if err == nil {
+		t.Fatal("expected error for invalid tool schema")
+	}
+}
+
+func TestToMessagesResponse_Basic(t *testing.T) {
+	resp := api.ChatResponse{
+		Model: "test-model",
+		Message: api.Message{
+			Role:    "assistant",
+			Content: "Hello there!",
+		},
+		Done:       true,
+		DoneReason: "stop",
+		Metrics: api.Metrics{
+			PromptEvalCount: 10,
+			EvalCount:       5,
+		},
+	}
+
+	result := ToMessagesResponse("msg_123", resp)
+
+	if result.ID != "msg_123" {
+		t.Errorf("expected ID 'msg_123', got %q", result.ID)
+	}
+	if result.Type != "message" {
+		t.Errorf("expected type 'message', got %q", result.Type)
+	}
+	if result.Role != "assistant" {
+		t.Errorf("expected role 'assistant', got %q", result.Role)
+	}
+	if len(result.Content) != 1 {
+		t.Fatalf("expected 1 content block, got %d", len(result.Content))
+	}
+	if result.Content[0].Type != "text" || result.Content[0].Text == nil || *result.Content[0].Text != "Hello there!" {
+		t.Errorf("unexpected content: %+v", result.Content[0])
+	}
+	if result.StopReason != "end_turn" {
+		t.Errorf("expected stop_reason 'end_turn', got %q", result.StopReason)
+	}
+	if result.Usage.InputTokens != 10 || result.Usage.OutputTokens != 5 {
+		t.Errorf("unexpected usage: %+v", result.Usage)
+	}
+}
+
+func TestToMessagesResponse_WithToolCalls(t *testing.T) {
+	resp := api.ChatResponse{
+		Model: "test-model",
+		Message: api.Message{
+			Role: "assistant",
+			ToolCalls: []api.ToolCall{
+				{
+					ID: "call_123",
+					Function: api.ToolCallFunction{
+						Name:      "get_weather",
+						Arguments: testArgs(map[string]any{"location": "Paris"}),
+					},
+				},
+			},
+		},
+		Done:       true,
+		DoneReason: "stop",
+	}
+
+	result := ToMessagesResponse("msg_123", resp)
+
+	if len(result.Content) != 1 {
+		t.Fatalf("expected 1 content block, got %d", len(result.Content))
+	}
+	if result.Content[0].Type != "tool_use" {
+		t.Errorf("expected type 'tool_use', got %q", result.Content[0].Type)
+	}
+	if result.Content[0].ID != "call_123" {
+		t.Errorf("expected ID 'call_123', got %q", result.Content[0].ID)
+	}
+	if result.Content[0].Name != "get_weather" {
+		t.Errorf("expected name 'get_weather', got %q", result.Content[0].Name)
+	}
+	if result.StopReason != "tool_use" {
+		t.Errorf("expected stop_reason 'tool_use', got %q", result.StopReason)
+	}
+}
+
+func TestToMessagesResponse_WithThinking(t *testing.T) {
+	resp := api.ChatResponse{
+		Model: "test-model",
+		Message: api.Message{
+			Role:     "assistant",
+			Content:  "The answer is 42.",
+			Thinking: "Let me think about this...",
+		},
+		Done:       true,
+		DoneReason: "stop",
+	}
+
+	result := ToMessagesResponse("msg_123", resp)
+
+	if len(result.Content) != 2 {
+		t.Fatalf("expected 2 content blocks, got %d", len(result.Content))
+	}
+	if result.Content[0].Type != "thinking" {
+		t.Errorf("expected first block type 'thinking', got %q", result.Content[0].Type)
+	}
+	if result.Content[0].Thinking == nil || *result.Content[0].Thinking != "Let me think about this..." {
+		t.Errorf("unexpected thinking content: %v", result.Content[0].Thinking)
+	}
+	if result.Content[1].Type != "text" {
+		t.Errorf("expected second block type 'text', got %q", result.Content[1].Type)
+	}
+}
+
+func TestMapStopReason(t *testing.T) {
+	tests := []struct {
+		reason       string
+		hasToolCalls bool
+		want         string
+	}{
+		{"stop", false, "end_turn"},
+		{"length", false, "max_tokens"},
+		{"stop", true, "tool_use"},
+		{"other", false, "stop_sequence"},
+		{"", false, ""},
+	}
+
+	for _, tt := range tests {
+		got := mapStopReason(tt.reason, tt.hasToolCalls)
+		if got != tt.want {
+			t.Errorf("mapStopReason(%q, %v) = %q, want %q", tt.reason, tt.hasToolCalls, got, tt.want)
+		}
+	}
+}
+
+func TestNewError(t *testing.T) {
+	tests := []struct {
+		code int
+		want string
+	}{
+		{400, "invalid_request_error"},
+		{401, "authentication_error"},
+		{403, "permission_error"},
+		{404, "not_found_error"},
+		{429, "rate_limit_error"},
+		{500, "api_error"},
+		{503, "overloaded_error"},
+		{529, "overloaded_error"},
+	}
+
+	for _, tt := range tests {
+		result := NewError(tt.code, "test message")
+		if result.Type != "error" {
+			t.Errorf("NewError(%d) type = %q, want 'error'", tt.code, result.Type)
+		}
+		if result.Error.Type != tt.want {
+			t.Errorf("NewError(%d) error.type = %q, want %q", tt.code, result.Error.Type, tt.want)
+		}
+		if result.Error.Message != "test message" {
+			t.Errorf("NewError(%d) message = %q, want 'test message'", tt.code, result.Error.Message)
+		}
+		if result.RequestID == "" {
+			t.Errorf("NewError(%d) request_id should not be empty", tt.code)
+		}
+	}
+}
+
+func TestGenerateMessageID(t *testing.T) {
+	id1 := GenerateMessageID()
+	id2 := GenerateMessageID()
+
+	if id1 == "" {
+		t.Error("GenerateMessageID returned empty string")
+	}
+	if id1 == id2 {
+		t.Error("GenerateMessageID returned duplicate IDs")
+	}
+	if len(id1) < 10 {
+		t.Errorf("GenerateMessageID returned short ID: %q", id1)
+	}
+	if id1[:4] != "msg_" {
+		t.Errorf("GenerateMessageID should start with 'msg_', got %q", id1[:4])
+	}
+}
+
+func TestStreamConverter_Basic(t *testing.T) {
+	conv := NewStreamConverter("msg_123", "test-model")
+
+	// First chunk
+	resp1 := api.ChatResponse{
+		Model: "test-model",
+		Message: api.Message{
+			Role:    "assistant",
+			Content: "Hello",
+		},
+		Metrics: api.Metrics{PromptEvalCount: 10},
+	}
+
+	events1 := conv.Process(resp1)
+	if len(events1) < 3 {
+		t.Fatalf("expected at least 3 events for first chunk, got %d", len(events1))
+	}
+
+	// Should have message_start, content_block_start, content_block_delta
+	if events1[0].Event != "message_start" {
+		t.Errorf("expected first event 'message_start', got %q", events1[0].Event)
+	}
+	if events1[1].Event != "content_block_start" {
+		t.Errorf("expected second event 'content_block_start', got %q", events1[1].Event)
+	}
+	if events1[2].Event != "content_block_delta" {
+		t.Errorf("expected third event 'content_block_delta', got %q", events1[2].Event)
+	}
+
+	// Final chunk
+	resp2 := api.ChatResponse{
+		Model: "test-model",
+		Message: api.Message{
+			Role:    "assistant",
+			Content: " world!",
+		},
+		Done:       true,
+		DoneReason: "stop",
+		Metrics:    api.Metrics{EvalCount: 5},
+	}
+
+	events2 := conv.Process(resp2)
+
+	// Should have content_block_delta, content_block_stop, message_delta, message_stop
+	hasStop := false
+	for _, e := range events2 {
+		if e.Event == "message_stop" {
+			hasStop = true
+		}
+	}
+	if !hasStop {
+		t.Error("expected message_stop event in final chunk")
+	}
+}
+
+func TestStreamConverter_WithToolCalls(t *testing.T) {
+	conv := NewStreamConverter("msg_123", "test-model")
+
+	resp := api.ChatResponse{
+		Model: "test-model",
+		Message: api.Message{
+			Role: "assistant",
+			ToolCalls: []api.ToolCall{
+				{
+					ID: "call_123",
+					Function: api.ToolCallFunction{
+						Name:      "get_weather",
+						Arguments: testArgs(map[string]any{"location": "Paris"}),
+					},
+				},
+			},
+		},
+		Done:       true,
+		DoneReason: "stop",
+		Metrics:    api.Metrics{PromptEvalCount: 10, EvalCount: 5},
+	}
+
+	events := conv.Process(resp)
+
+	hasToolStart := false
+	hasToolDelta := false
+	for _, e := range events {
+		if e.Event == "content_block_start" {
+			if start, ok := e.Data.(ContentBlockStartEvent); ok {
+				if start.ContentBlock.Type == "tool_use" {
+					hasToolStart = true
+				}
+			}
+		}
+		if e.Event == "content_block_delta" {
+			if delta, ok := e.Data.(ContentBlockDeltaEvent); ok {
+				if delta.Delta.Type == "input_json_delta" {
+					hasToolDelta = true
+				}
+			}
+		}
+	}
+
+	if !hasToolStart {
+		t.Error("expected tool_use content_block_start event")
+	}
+	if !hasToolDelta {
+		t.Error("expected input_json_delta event")
+	}
+}
+
+func TestStreamConverter_ToolCallWithUnmarshalableArgs(t *testing.T) {
+	// Test that unmarshalable arguments (like channels) are handled gracefully
+	// and don't cause a panic or corrupt stream
+	conv := NewStreamConverter("msg_123", "test-model")
+
+	// Create a channel which cannot be JSON marshaled
+	unmarshalable := make(chan int)
+	badArgs := api.NewToolCallFunctionArguments()
+	badArgs.Set("channel", unmarshalable)
+
+	resp := api.ChatResponse{
+		Model: "test-model",
+		Message: api.Message{
+			Role: "assistant",
+			ToolCalls: []api.ToolCall{
+				{
+					ID: "call_bad",
+					Function: api.ToolCallFunction{
+						Name:      "bad_function",
+						Arguments: badArgs,
+					},
+				},
+			},
+		},
+		Done:       true,
+		DoneReason: "stop",
+	}
+
+	// Should not panic and should skip the unmarshalable tool call
+	events := conv.Process(resp)
+
+	// Verify no tool_use block was started (since marshal failed before block start)
+	hasToolStart := false
+	for _, e := range events {
+		if e.Event == "content_block_start" {
+			if start, ok := e.Data.(ContentBlockStartEvent); ok {
+				if start.ContentBlock.Type == "tool_use" {
+					hasToolStart = true
+				}
+			}
+		}
+	}
+
+	if hasToolStart {
+		t.Error("expected no tool_use block when arguments cannot be marshaled")
+	}
+}
+
+func TestStreamConverter_MultipleToolCallsWithMixedValidity(t *testing.T) {
+	// Test that valid tool calls still work when mixed with invalid ones
+	conv := NewStreamConverter("msg_123", "test-model")
+
+	unmarshalable := make(chan int)
+	badArgs := api.NewToolCallFunctionArguments()
+	badArgs.Set("channel", unmarshalable)
+
+	resp := api.ChatResponse{
+		Model: "test-model",
+		Message: api.Message{
+			Role: "assistant",
+			ToolCalls: []api.ToolCall{
+				{
+					ID: "call_good",
+					Function: api.ToolCallFunction{
+						Name:      "good_function",
+						Arguments: testArgs(map[string]any{"location": "Paris"}),
+					},
+				},
+				{
+					ID: "call_bad",
+					Function: api.ToolCallFunction{
+						Name:      "bad_function",
+						Arguments: badArgs,
+					},
+				},
+			},
+		},
+		Done:       true,
+		DoneReason: "stop",
+	}
+
+	events := conv.Process(resp)
+
+	// Count tool_use blocks - should only have 1 (the valid one)
+	toolStartCount := 0
+	toolDeltaCount := 0
+	for _, e := range events {
+		if e.Event == "content_block_start" {
+			if start, ok := e.Data.(ContentBlockStartEvent); ok {
+				if start.ContentBlock.Type == "tool_use" {
+					toolStartCount++
+					if start.ContentBlock.Name != "good_function" {
+						t.Errorf("expected tool name 'good_function', got %q", start.ContentBlock.Name)
+					}
+				}
+			}
+		}
+		if e.Event == "content_block_delta" {
+			if delta, ok := e.Data.(ContentBlockDeltaEvent); ok {
+				if delta.Delta.Type == "input_json_delta" {
+					toolDeltaCount++
+				}
+			}
+		}
+	}
+
+	if toolStartCount != 1 {
+		t.Errorf("expected 1 tool_use block, got %d", toolStartCount)
+	}
+	if toolDeltaCount != 1 {
+		t.Errorf("expected 1 input_json_delta, got %d", toolDeltaCount)
+	}
+}
+
+// TestContentBlockJSON_EmptyFieldsPresent verifies that empty text and thinking fields
+// are serialized in JSON output. The Anthropic SDK requires these fields to be present
+// (even when empty) in content_block_start events to properly accumulate streaming deltas.
+// Without these fields, the SDK throws: "TypeError: unsupported operand type(s) for +=: 'NoneType' and 'str'"
+func TestContentBlockJSON_EmptyFieldsPresent(t *testing.T) {
+	tests := []struct {
+		name     string
+		block    ContentBlock
+		wantKeys []string
+	}{
+		{
+			name: "text block includes empty text field",
+			block: ContentBlock{
+				Type: "text",
+				Text: ptr(""),
+			},
+			wantKeys: []string{"type", "text"},
+		},
+		{
+			name: "thinking block includes empty thinking field",
+			block: ContentBlock{
+				Type:     "thinking",
+				Thinking: ptr(""),
+			},
+			wantKeys: []string{"type", "thinking"},
+		},
+		{
+			name: "text block with content",
+			block: ContentBlock{
+				Type: "text",
+				Text: ptr("hello"),
+			},
+			wantKeys: []string{"type", "text"},
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			data, err := json.Marshal(tt.block)
+			if err != nil {
+				t.Fatalf("failed to marshal: %v", err)
+			}
+
+			var result map[string]any
+			if err := json.Unmarshal(data, &result); err != nil {
+				t.Fatalf("failed to unmarshal: %v", err)
+			}
+
+			for _, key := range tt.wantKeys {
+				if _, ok := result[key]; !ok {
+					t.Errorf("expected key %q to be present in JSON output, got: %s", key, string(data))
+				}
+			}
+		})
+	}
+}
+
+// TestStreamConverter_ContentBlockStartIncludesEmptyFields verifies that content_block_start
+// events include the required empty fields for SDK compatibility.
+func TestStreamConverter_ContentBlockStartIncludesEmptyFields(t *testing.T) {
+	t.Run("text block start includes empty text", func(t *testing.T) {
+		conv := NewStreamConverter("msg_123", "test-model")
+
+		resp := api.ChatResponse{
+			Model:   "test-model",
+			Message: api.Message{Role: "assistant", Content: "hello"},
+		}
+
+		events := conv.Process(resp)
+
+		var foundTextStart bool
+		for _, e := range events {
+			if e.Event == "content_block_start" {
+				if start, ok := e.Data.(ContentBlockStartEvent); ok {
+					if start.ContentBlock.Type == "text" {
+						foundTextStart = true
+						// Marshal and verify the text field is present
+						data, _ := json.Marshal(start)
+						var result map[string]any
+						json.Unmarshal(data, &result)
+						cb := result["content_block"].(map[string]any)
+						if _, ok := cb["text"]; !ok {
+							t.Error("content_block_start for text should include 'text' field")
+						}
+					}
+				}
+			}
+		}
+
+		if !foundTextStart {
+			t.Error("expected text content_block_start event")
+		}
+	})
+
+	t.Run("thinking block start includes empty thinking", func(t *testing.T) {
+		conv := NewStreamConverter("msg_123", "test-model")
+
+		resp := api.ChatResponse{
+			Model:   "test-model",
+			Message: api.Message{Role: "assistant", Thinking: "let me think..."},
+		}
+
+		events := conv.Process(resp)
+
+		var foundThinkingStart bool
+		for _, e := range events {
+			if e.Event == "content_block_start" {
+				if start, ok := e.Data.(ContentBlockStartEvent); ok {
+					if start.ContentBlock.Type == "thinking" {
+						foundThinkingStart = true
+						data, _ := json.Marshal(start)
+						var result map[string]any
+						json.Unmarshal(data, &result)
+						cb := result["content_block"].(map[string]any)
+						if _, ok := cb["thinking"]; !ok {
+							t.Error("content_block_start for thinking should include 'thinking' field")
+						}
+					}
+				}
+			}
+		}
+
+		if !foundThinkingStart {
+			t.Error("expected thinking content_block_start event")
+		}
+	})
+}

api/client.go

@@ -165,7 +165,7 @@ func (c *Client) do(ctx context.Context, method, path string, reqData, respData
 	return nil
 }
 
-const maxBufferSize = 512 * format.KiloByte
+const maxBufferSize = 8 * format.MegaByte
 
 func (c *Client) stream(ctx context.Context, method, path string, data any, fn func([]byte) error) error {
 	var buf io.Reader

api/types.go

@@ -127,6 +127,20 @@ type GenerateRequest struct {
 	// each with an associated log probability. Only applies when Logprobs is true.
 	// Valid values are 0-20. Default is 0 (only return the selected token's logprob).
 	TopLogprobs int `json:"top_logprobs,omitempty"`
+
+	// Experimental: Image generation fields (may change or be removed)
+
+	// Width is the width of the generated image in pixels.
+	// Only used for image generation models.
+	Width int32 `json:"width,omitempty"`
+
+	// Height is the height of the generated image in pixels.
+	// Only used for image generation models.
+	Height int32 `json:"height,omitempty"`
+
+	// Steps is the number of diffusion steps for image generation.
+	// Only used for image generation models.
+	Steps int32 `json:"steps,omitempty"`
 }
 
 // ChatRequest describes a request sent by [Client.Chat].
@@ -860,6 +874,20 @@ type GenerateResponse struct {
 	// Logprobs contains log probability information for the generated tokens,
 	// if requested via the Logprobs parameter.
 	Logprobs []Logprob `json:"logprobs,omitempty"`
+
+	// Experimental: Image generation fields (may change or be removed)
+
+	// Image contains a base64-encoded generated image.
+	// Only present for image generation models.
+	Image string `json:"image,omitempty"`
+
+	// Completed is the number of completed steps in image generation.
+	// Only present for image generation models during streaming.
+	Completed int64 `json:"completed,omitempty"`
+
+	// Total is the total number of steps for image generation.
+	// Only present for image generation models during streaming.
+	Total int64 `json:"total,omitempty"`
 }
 
 // ModelDetails provides details about a model.

app/cmd/app/app_darwin.m

@@ -14,6 +14,7 @@ extern NSString *SystemWidePath;
 @interface AppDelegate () <NSWindowDelegate, WKNavigationDelegate, WKUIDelegate>
 @property(strong, nonatomic) NSStatusItem *statusItem;
 @property(assign, nonatomic) BOOL updateAvailable;
+@property(assign, nonatomic) BOOL systemShutdownInProgress;
 @end
 
 @implementation AppDelegate
@@ -40,6 +41,13 @@ bool firstTimeRun,startHidden; // Set in run before initialization
 }
 
 - (void)applicationDidFinishLaunching:(NSNotification *)aNotification {
+    // Register for system shutdown/restart notification so we can allow termination
+    [[[NSWorkspace sharedWorkspace] notificationCenter]
+        addObserver:self
+           selector:@selector(systemWillPowerOff:)
+               name:NSWorkspaceWillPowerOffNotification
+             object:nil];
+
     // if we're in development mode, set the app icon
     NSString *bundlePath = [[NSBundle mainBundle] bundlePath];
     if (![bundlePath hasSuffix:@".app"]) {
@@ -278,7 +286,18 @@ bool firstTimeRun,startHidden; // Set in run before initialization
     [NSApp activateIgnoringOtherApps:YES];
 }
 
+- (void)systemWillPowerOff:(NSNotification *)notification {
+    // Set flag so applicationShouldTerminate: knows to allow termination.
+    // The system will call applicationShouldTerminate: after posting this notification.
+    self.systemShutdownInProgress = YES;
+}
+
 - (NSApplicationTerminateReply)applicationShouldTerminate:(NSApplication *)sender {
+    // Allow termination if the system is shutting down or restarting
+    if (self.systemShutdownInProgress) {
+        return NSTerminateNow;
+    }
+    // Otherwise just hide the app (for Cmd+Q, close button, etc.)
     [NSApp hide:nil];
     [NSApp setActivationPolicy:NSApplicationActivationPolicyAccessory];
     return NSTerminateCancel;

cmd/cmd.go

@@ -46,6 +46,9 @@ import (
 	"github.com/ollama/ollama/types/syncmap"
 	"github.com/ollama/ollama/version"
 	xcmd "github.com/ollama/ollama/x/cmd"
+	"github.com/ollama/ollama/x/create"
+	xcreateclient "github.com/ollama/ollama/x/create/client"
+	"github.com/ollama/ollama/x/imagegen"
 )
 
 const ConnectInstructions = "To sign in, navigate to:\n    %s\n\n"
@@ -91,11 +94,27 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 	p := progress.NewProgress(os.Stderr)
 	defer p.Stop()
 
+	// Validate model name early to fail fast
+	modelName := args[0]
+	name := model.ParseName(modelName)
+	if !name.IsValid() {
+		return fmt.Errorf("invalid model name: %s", modelName)
+	}
+
 	var reader io.Reader
 
 	filename, err := getModelfileName(cmd)
 	if os.IsNotExist(err) {
 		if filename == "" {
+			// No Modelfile found - check if current directory is an image gen model
+			if create.IsTensorModelDir(".") {
+				quantize, _ := cmd.Flags().GetString("quantize")
+				return xcreateclient.CreateModel(xcreateclient.CreateOptions{
+					ModelName: modelName,
+					ModelDir:  ".",
+					Quantize:  quantize,
+				}, p)
+			}
 			reader = strings.NewReader("FROM .\n")
 		} else {
 			return errModelfileNotFound
@@ -117,6 +136,28 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 		return err
 	}
 
+	// Check if this is a tensor model (image generation) and handle it directly
+	quantize, _ := cmd.Flags().GetString("quantize")
+	modelDir := filepath.Dir(filename)
+	for _, cmd := range modelfile.Commands {
+		if cmd.Name == "model" {
+			if filepath.IsAbs(cmd.Args) {
+				modelDir = cmd.Args
+			} else {
+				modelDir = filepath.Join(filepath.Dir(filename), cmd.Args)
+			}
+			break
+		}
+	}
+	if create.IsTensorModelDir(modelDir) {
+		return xcreateclient.CreateModel(xcreateclient.CreateOptions{
+			ModelName: modelName,
+			ModelDir:  modelDir,
+			Quantize:  quantize,
+			Modelfile: xcreateclient.ExtractModelfileConfig(modelfile),
+		}, p)
+	}
+
 	status := "gathering model components"
 	spinner := progress.NewSpinner(status)
 	p.Add(status, spinner)
@@ -127,8 +168,7 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 	}
 	spinner.Stop()
 
-	req.Model = args[0]
-	quantize, _ := cmd.Flags().GetString("quantize")
+	req.Model = modelName
 	if quantize != "" {
 		req.Quantize = quantize
 	}
@@ -457,6 +497,7 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 	}
 
 	name := args[0]
+
 	info, err := func() (*api.ShowResponse, error) {
 		showReq := &api.ShowRequest{Name: name}
 		info, err := client.Show(cmd.Context(), showReq)
@@ -518,9 +559,18 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 		return generateEmbedding(cmd, name, opts.Prompt, opts.KeepAlive, truncate, dimensions)
 	}
 
+	// Check if this is an image generation model
+	if slices.Contains(info.Capabilities, model.CapabilityImage) {
+		if opts.Prompt == "" && !interactive {
+			return errors.New("image generation models require a prompt. Usage: ollama run " + name + " \"your prompt here\"")
+		}
+		return imagegen.RunCLI(cmd, name, opts.Prompt, interactive, opts.KeepAlive)
+	}
+
 	// Check for experimental flag
 	isExperimental, _ := cmd.Flags().GetBool("experimental")
-	yoloMode, _ := cmd.Flags().GetBool("yolo")
+	yoloMode, _ := cmd.Flags().GetBool("experimental-yolo")
+	enableWebsearch, _ := cmd.Flags().GetBool("experimental-websearch")
 
 	if interactive {
 		if err := loadOrUnloadModel(cmd, &opts); err != nil {
@@ -550,7 +600,7 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 
 		// Use experimental agent loop with tools
 		if isExperimental {
-			return xcmd.GenerateInteractive(cmd, opts.Model, opts.WordWrap, opts.Options, opts.Think, opts.HideThinking, opts.KeepAlive, yoloMode)
+			return xcmd.GenerateInteractive(cmd, opts.Model, opts.WordWrap, opts.Options, opts.Think, opts.HideThinking, opts.KeepAlive, yoloMode, enableWebsearch)
 		}
 
 		return generateInteractive(cmd, opts)
@@ -656,7 +706,11 @@ func PushHandler(cmd *cobra.Command, args []string) error {
 
 			bar, ok := bars[resp.Digest]
 			if !ok {
-				bar = progress.NewBar(fmt.Sprintf("pushing %s...", resp.Digest[7:19]), resp.Total, resp.Completed)
+				msg := resp.Status
+				if msg == "" {
+					msg = fmt.Sprintf("pushing %s...", resp.Digest[7:19])
+				}
+				bar = progress.NewBar(msg, resp.Total, resp.Completed)
 				bars[resp.Digest] = bar
 				p.Add(resp.Digest, bar)
 			}
@@ -805,11 +859,11 @@ func DeleteHandler(cmd *cobra.Command, args []string) error {
 	for _, arg := range args {
 		// Unload the model if it's running before deletion
 		if err := loadOrUnloadModel(cmd, &runOptions{
-			Model:     args[0],
+			Model:     arg,
 			KeepAlive: &api.Duration{Duration: 0},
 		}); err != nil {
 			if !strings.Contains(strings.ToLower(err.Error()), "not found") {
-				fmt.Fprintf(os.Stderr, "Warning: unable to stop model '%s'\n", args[0])
+				fmt.Fprintf(os.Stderr, "Warning: unable to stop model '%s'\n", arg)
 			}
 		}
 
@@ -1765,7 +1819,11 @@ func NewCLI() *cobra.Command {
 	runCmd.Flags().Bool("truncate", false, "For embedding models: truncate inputs exceeding context length (default: true). Set --truncate=false to error instead")
 	runCmd.Flags().Int("dimensions", 0, "Truncate output embeddings to specified dimension (embedding models only)")
 	runCmd.Flags().Bool("experimental", false, "Enable experimental agent loop with tools")
-	runCmd.Flags().BoolP("yolo", "y", false, "Skip all tool approval prompts (use with caution)")
+	runCmd.Flags().Bool("experimental-yolo", false, "Skip all tool approval prompts (use with caution)")
+	runCmd.Flags().Bool("experimental-websearch", false, "Enable web search tool in experimental mode")
+
+	// Image generation flags (width, height, steps, seed, etc.)
+	imagegen.RegisterFlags(runCmd)
 
 	stopCmd := &cobra.Command{
 		Use:     "stop MODEL",
@@ -1880,6 +1938,7 @@ func NewCLI() *cobra.Command {
 	} {
 		switch cmd {
 		case runCmd:
+			imagegen.AppendFlagsDocs(cmd)
 			appendEnvDocs(cmd, []envconfig.EnvVar{envVars["OLLAMA_HOST"], envVars["OLLAMA_NOHISTORY"]})
 		case serveCmd:
 			appendEnvDocs(cmd, []envconfig.EnvVar{

cmd/cmd_test.go

@@ -1547,6 +1547,79 @@ func TestRunOptions_Copy_ThinkValueVariants(t *testing.T) {
 	}
 }
 
+func TestShowInfoImageGen(t *testing.T) {
+	var b bytes.Buffer
+	err := showInfo(&api.ShowResponse{
+		Details: api.ModelDetails{
+			Family:            "ZImagePipeline",
+			ParameterSize:     "10.3B",
+			QuantizationLevel: "FP8",
+		},
+		Capabilities: []model.Capability{model.CapabilityImage},
+		Requires:     "0.14.0",
+	}, false, &b)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	expect := "  Model\n" +
+		"    architecture    ZImagePipeline    \n" +
+		"    parameters      10.3B             \n" +
+		"    quantization    FP8               \n" +
+		"    requires        0.14.0            \n" +
+		"\n" +
+		"  Capabilities\n" +
+		"    image    \n" +
+		"\n"
+	if diff := cmp.Diff(expect, b.String()); diff != "" {
+		t.Errorf("unexpected output (-want +got):\n%s", diff)
+	}
+}
+
+func TestPushProgressMessage(t *testing.T) {
+	tests := []struct {
+		name    string
+		status  string
+		digest  string
+		wantMsg string
+	}{
+		{
+			name:    "uses status when provided",
+			status:  "uploading model",
+			digest:  "sha256:abc123456789def",
+			wantMsg: "uploading model",
+		},
+		{
+			name:    "falls back to digest when status empty",
+			status:  "",
+			digest:  "sha256:abc123456789def",
+			wantMsg: "pushing abc123456789...",
+		},
+		{
+			name:    "handles short digest gracefully",
+			status:  "",
+			digest:  "sha256:abc",
+			wantMsg: "pushing sha256:abc...",
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			msg := tt.status
+			if msg == "" {
+				if len(tt.digest) >= 19 {
+					msg = fmt.Sprintf("pushing %s...", tt.digest[7:19])
+				} else {
+					msg = fmt.Sprintf("pushing %s...", tt.digest)
+				}
+			}
+			if msg != tt.wantMsg {
+				t.Errorf("got %q, want %q", msg, tt.wantMsg)
+			}
+		})
+	}
+}
+
 func TestRunOptions_Copy_Independence(t *testing.T) {
 	// Test that modifications to original don't affect copy
 	originalThink := &api.ThinkValue{Value: "original"}

cmd/interactive.go

@@ -116,7 +116,7 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 		Prompt:         ">>> ",
 		AltPrompt:      "... ",
 		Placeholder:    "Send a message (/? for help)",
-		AltPlaceholder: `Use """ to end multi-line input`,
+		AltPlaceholder: "Press Enter to send",
 	})
 	if err != nil {
 		return err

convert/convert.go

@@ -6,11 +6,14 @@ import (
 	"errors"
 	"fmt"
 	"io/fs"
+	"iter"
 	"log/slog"
+	"maps"
 	"os"
 	"slices"
 	"strings"
 
+	ofs "github.com/ollama/ollama/fs"
 	"github.com/ollama/ollama/fs/ggml"
 )
 
@@ -18,8 +21,13 @@ type ModelParameters struct {
 	Architectures []string `json:"architectures"`
 	VocabSize     uint32   `json:"vocab_size"`
 
+	// TODO is this needed?
+	ModelType string `json:"model_type"`
+
 	TextModel struct {
-		VocabSize uint32 `json:"vocab_size"`
+		VocabSize  uint32 `json:"vocab_size"`
+		HiddenSize uint32 `json:"hidden_size"`
+		ModelType  string `json:"model_type"`
 	} `json:"text_config"`
 }
 
@@ -33,8 +41,94 @@ type AdapterParameters struct {
 	} `json:"lora_parameters"`
 }
 
-func (ModelParameters) KV(t *Tokenizer) ggml.KV {
-	kv := ggml.KV{
+type KV map[string]any
+
+func (kv KV) Architecture() string {
+	return kv.String("general.architecture", "unknown")
+}
+
+type valueTypes interface {
+	uint8 | int8 | uint16 | int16 |
+		uint32 | int32 | uint64 | int64 |
+		string | float32 | float64 | bool
+}
+
+type arrayValueTypes interface {
+	[]uint8 | []int8 | []uint16 | []int16 |
+		[]uint32 | []int32 | []uint64 | []int64 |
+		[]string | []float32 | []float64 | []bool
+}
+
+func keyValue[T valueTypes | arrayValueTypes](kv KV, key string, defaultValue ...T) (T, bool) {
+	if !strings.HasPrefix(key, "tokenizer.") && !strings.HasPrefix(key, "general.") {
+		key = kv.Architecture() + "." + key
+	}
+
+	if val, ok := kv[key].(T); ok {
+		return val, true
+	}
+	return defaultValue[0], false
+}
+
+func (kv KV) String(key string, defaultValue ...string) string {
+	val, _ := keyValue(kv, key, append(defaultValue, "")...)
+	return val
+}
+
+func (kv KV) Uint(key string, defaultValue ...uint32) uint32 {
+	val, _ := keyValue(kv, key, append(defaultValue, 0)...)
+	return val
+}
+
+func (kv KV) Float(key string, defaultValue ...float32) float32 {
+	val, _ := keyValue(kv, key, append(defaultValue, 0)...)
+	return val
+}
+
+func (kv KV) Bool(key string, defaultValue ...bool) bool {
+	val, _ := keyValue(kv, key, append(defaultValue, false)...)
+	return val
+}
+
+func (kv KV) Strings(key string, defaultValue ...[]string) []string {
+	val, _ := keyValue(kv, key, append(defaultValue, []string{""})...)
+	return val
+}
+
+func (kv KV) Ints(key string, defaultValue ...[]int32) []int32 {
+	val, _ := keyValue(kv, key, append(defaultValue, []int32{0})...)
+	return val
+}
+
+func (kv KV) Uints(key string, defaultValue ...[]uint32) []uint32 {
+	val, _ := keyValue(kv, key, append(defaultValue, []uint32{0})...)
+	return val
+}
+
+func (kv KV) Floats(key string, defaultValue ...[]float32) []float32 {
+	val, _ := keyValue(kv, key, append(defaultValue, []float32{0})...)
+	return val
+}
+
+func (kv KV) Bools(key string, defaultValue ...[]bool) []bool {
+	val, _ := keyValue(kv, key, append(defaultValue, []bool{false})...)
+	return val
+}
+
+func (kv KV) Len() int {
+	return len(kv)
+}
+
+func (kv KV) Keys() iter.Seq[string] {
+	return maps.Keys(kv)
+}
+
+func (kv KV) Value(key string) any {
+	return kv[key]
+}
+
+func (ModelParameters) KV(t *Tokenizer) KV {
+	kv := KV{
 		"general.file_type":            uint32(1),
 		"general.quantization_version": uint32(2),
 		"tokenizer.ggml.pre":           t.Pre,
@@ -63,7 +157,7 @@ func (ModelParameters) KV(t *Tokenizer) ggml.KV {
 	return kv
 }
 
-func (p AdapterParameters) KV() ggml.KV {
+func (p AdapterParameters) KV() KV {
 	var alpha float32
 	if p.LoraParameters.Alpha == 0 {
 		alpha = float32(p.Alpha)
@@ -71,7 +165,7 @@ func (p AdapterParameters) KV() ggml.KV {
 		alpha = p.LoraParameters.Alpha
 	}
 
-	kv := ggml.KV{
+	kv := KV{
 		"adapter.lora.alpha": alpha,
 		"adapter.type":       "lora",
 		"general.file_type":  uint32(1),
@@ -88,9 +182,14 @@ func (ModelParameters) specialTokenTypes() []string {
 	}
 }
 
-type ModelConverter interface {
+type ModelKV interface {
 	// KV maps parameters to LLM key-values
-	KV(*Tokenizer) ggml.KV
+	KV(*Tokenizer) KV
+}
+
+type ModelConverter interface {
+	ModelKV
+
 	// Tensors maps input tensors to LLM tensors. Model specific modifications can be done here.
 	Tensors([]Tensor) []*ggml.Tensor
 	// Replacements returns a list of string pairs to replace in tensor names.
@@ -107,7 +206,7 @@ type moreParser interface {
 
 type AdapterConverter interface {
 	// KV maps parameters to LLM key-values
-	KV(ggml.KV) ggml.KV
+	KV(ofs.Config) KV
 	// Tensors maps input tensors to LLM tensors. Adapter specific modifications can be done here.
 	Tensors([]Tensor) []*ggml.Tensor
 	// Replacements returns a list of string pairs to replace in tensor names.
@@ -115,7 +214,7 @@ type AdapterConverter interface {
 	Replacements() []string
 }
 
-func ConvertAdapter(fsys fs.FS, f *os.File, baseKV ggml.KV) error {
+func ConvertAdapter(fsys fs.FS, f *os.File, baseKV ofs.Config) error {
 	bts, err := fs.ReadFile(fsys, "adapter_config.json")
 	if err != nil {
 		return err
@@ -126,8 +225,8 @@ func ConvertAdapter(fsys fs.FS, f *os.File, baseKV ggml.KV) error {
 		return err
 	}
 
-	arch, ok := baseKV["general.architecture"]
-	if !ok {
+	arch := baseKV.Architecture()
+	if arch == "" {
 		return errors.New("architecture not set for the base model")
 	}
 
@@ -153,23 +252,19 @@ func ConvertAdapter(fsys fs.FS, f *os.File, baseKV ggml.KV) error {
 	return writeFile(f, conv.KV(baseKV), conv.Tensors(ts))
 }
 
-// Convert writes an Ollama compatible model to the provided io.WriteSeeker based on configurations
-// 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 LoadModelMetadata(fsys fs.FS) (ModelKV, *Tokenizer, error) {
 	bts, err := fs.ReadFile(fsys, "config.json")
 	if err != nil {
-		return err
+		return nil, nil, err
 	}
 
 	var p ModelParameters
 	if err := json.Unmarshal(bts, &p); err != nil {
-		return err
+		return nil, nil, err
 	}
 
 	if len(p.Architectures) < 1 {
-		return errors.New("unknown architecture")
+		return nil, nil, errors.New("unknown architecture")
 	}
 
 	var conv ModelConverter
@@ -216,23 +311,25 @@ func ConvertModel(fsys fs.FS, f *os.File) error {
 		conv = &deepseekocr{}
 	case "DeepseekV3ForCausalLM":
 		conv = &deepseek2Model{}
+	case "Glm4MoeLiteForCausalLM":
+		conv = &glm4MoeLiteModel{}
 	default:
-		return fmt.Errorf("unsupported architecture %q", p.Architectures[0])
+		return nil, nil, fmt.Errorf("unsupported architecture %q", p.Architectures[0])
 	}
 
 	if err := json.Unmarshal(bts, conv); err != nil {
-		return err
+		return nil, nil, err
 	}
 
 	if t, ok := conv.(moreParser); ok {
 		if err := t.parseMore(fsys); err != nil {
-			return err
+			return nil, nil, err
 		}
 	}
 
 	t, err := parseTokenizer(fsys, conv.specialTokenTypes())
 	if err != nil {
-		return err
+		return nil, nil, err
 	}
 
 	vocabSize := int(cmp.Or(p.VocabSize, p.TextModel.VocabSize))
@@ -254,6 +351,19 @@ func ConvertModel(fsys fs.FS, f *os.File) error {
 	default:
 		slog.Debug("vocabulary", "size", len(t.Vocabulary.Tokens))
 	}
+	return conv, t, nil
+}
+
+// Convert writes an Ollama compatible model to the provided io.WriteSeeker based on configurations
+// 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 {
+	kv, t, err := LoadModelMetadata(fsys)
+	if err != nil {
+		return err
+	}
+	conv := kv.(ModelConverter)
 
 	ts, err := parseTensors(fsys, strings.NewReplacer(conv.Replacements()...))
 	if err != nil {
@@ -263,7 +373,7 @@ func ConvertModel(fsys fs.FS, f *os.File) error {
 	return writeFile(f, conv.KV(t), conv.Tensors(ts))
 }
 
-func writeFile(f *os.File, kv ggml.KV, ts []*ggml.Tensor) error {
+func writeFile(f *os.File, kv KV, ts []*ggml.Tensor) error {
 	for i := range ts {
 		ts[i].Shape = slices.Clone(ts[i].Shape)
 		slices.Reverse(ts[i].Shape)

convert/convert_bert.go

@@ -88,7 +88,7 @@ func (p *bertModel) parseMore(fsys fs.FS) error {
 	return nil
 }
 
-func (p *bertModel) KV(t *Tokenizer) ggml.KV {
+func (p *bertModel) KV(t *Tokenizer) KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "bert"
 	kv["bert.attention.causal"] = false

convert/convert_commandr.go

@@ -24,7 +24,7 @@ type commandrModel struct {
 
 var _ ModelConverter = (*commandrModel)(nil)
 
-func (p *commandrModel) KV(t *Tokenizer) ggml.KV {
+func (p *commandrModel) KV(t *Tokenizer) KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "command-r"
 	kv["general.name"] = "command-r"

convert/convert_deepseek2.go

@@ -47,7 +47,7 @@ type deepseek2Model struct {
 	Architecture string
 }
 
-func (p *deepseek2Model) KV(t *Tokenizer) ggml.KV {
+func (p *deepseek2Model) KV(t *Tokenizer) KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "deepseek2"
 	kv["general.type"] = "model"

convert/convert_deepseekocr.go

@@ -41,7 +41,7 @@ type deepseekocr struct {
 	} `json:"vision_config"`
 }
 
-func (m *deepseekocr) KV(t *Tokenizer) ggml.KV {
+func (m *deepseekocr) KV(t *Tokenizer) KV {
 	kv := m.ModelParameters.KV(t)
 	kv["general.architecture"] = "deepseekocr"
 	kv["block_count"] = m.LanguageConfig.HiddenLayers

convert/convert_gemma.go

@@ -23,7 +23,7 @@ type gemmaModel struct {
 
 var _ ModelConverter = (*gemmaModel)(nil)
 
-func (p *gemmaModel) KV(t *Tokenizer) ggml.KV {
+func (p *gemmaModel) KV(t *Tokenizer) KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "gemma"
 	kv["gemma.context_length"] = p.MaxPositionEmbeddings

convert/convert_gemma2.go

@@ -1,7 +1,5 @@
 package convert
 
-import "github.com/ollama/ollama/fs/ggml"
-
 type gemma2Model struct {
 	gemmaModel
 	SlidingWindow         uint32  `json:"sliding_window"`
@@ -9,7 +7,7 @@ type gemma2Model struct {
 	FinalLogitSoftcap     float32 `json:"final_logit_softcapping"`
 }
 
-func (p *gemma2Model) KV(t *Tokenizer) ggml.KV {
+func (p *gemma2Model) KV(t *Tokenizer) KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "gemma2"
 	kv["gemma2.context_length"] = p.MaxPositionEmbeddings

convert/convert_gemma2_adapter.go

@@ -6,6 +6,7 @@ import (
 	"github.com/pdevine/tensor"
 	"github.com/pdevine/tensor/native"
 
+	"github.com/ollama/ollama/fs"
 	"github.com/ollama/ollama/fs/ggml"
 )
 
@@ -15,7 +16,7 @@ type gemma2Adapter struct {
 
 var _ AdapterConverter = (*gemma2Adapter)(nil)
 
-func (p *gemma2Adapter) KV(baseKV ggml.KV) ggml.KV {
+func (p *gemma2Adapter) KV(baseKV fs.Config) KV {
 	kv := p.AdapterParameters.KV()
 	kv["general.architecture"] = "gemma2"
 	return kv

convert/convert_gemma3.go

@@ -3,8 +3,6 @@ package convert
 import (
 	"cmp"
 	"slices"
-
-	"github.com/ollama/ollama/fs/ggml"
 )
 
 type gemma3Model struct {
@@ -55,7 +53,7 @@ const (
 	gemma27BLayerCount = 62
 )
 
-func (p *gemma3Model) KV(t *Tokenizer) ggml.KV {
+func (p *gemma3Model) KV(t *Tokenizer) KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "gemma3"
 

convert/convert_gemma3n.go

@@ -38,7 +38,7 @@ type gemma3nModel struct {
 	VisionModel struct{} `json:"vision_config"`
 }
 
-func (m *gemma3nModel) KV(t *Tokenizer) ggml.KV {
+func (m *gemma3nModel) KV(t *Tokenizer) KV {
 	kv := m.ModelParameters.KV(t)
 	kv["general.architecture"] = "gemma3n"
 	kv["gemma3n.activation_sparsity_scale"] = slices.Collect(func(yield func(float32) bool) {

convert/convert_glm4moelite.go

@@ -0,0 +1,150 @@
+package convert
+
+import (
+	"cmp"
+	"fmt"
+	"log/slog"
+	"regexp"
+	"strconv"
+
+	"github.com/ollama/ollama/fs/ggml"
+)
+
+type glm4MoeLiteModel struct {
+	ModelParameters
+	MaxPositionEmbeddings uint32  `json:"max_position_embeddings"`
+	HiddenSize            uint32  `json:"hidden_size"`
+	HiddenLayers          uint32  `json:"num_hidden_layers"`
+	IntermediateSize      uint32  `json:"intermediate_size"`
+	NumAttentionHeads     uint32  `json:"num_attention_heads"`
+	NumKeyValueHeads      uint32  `json:"num_key_value_heads"`
+	RMSNormEPS            float32 `json:"rms_norm_eps"`
+
+	RopeTheta     float32 `json:"rope_theta"`
+	QKNopeHeadDim uint32  `json:"qk_nope_head_dim"`
+	QKRopeHeadDim uint32  `json:"qk_rope_head_dim"`
+	KVLoraRank    uint32  `json:"kv_lora_rank"`
+	QLoraRank     uint32  `json:"q_lora_rank"`
+	VHeadDim      uint32  `json:"v_head_dim"`
+
+	ExpertCount            uint32  `json:"n_routed_experts"`
+	ExpertSharedCount      uint32  `json:"n_shared_experts"`
+	ExpertIntermediateSize uint32  `json:"moe_intermediate_size"`
+	ExpertUsedCount        uint32  `json:"num_experts_per_tok"`
+	ExpertWeightsNorm      bool    `json:"norm_topk_prob"`
+	ExpertWeightsScale     float32 `json:"routed_scaling_factor"`
+
+	LeadingDenseBlockCount uint32 `json:"first_k_dense_replace"`
+}
+
+func (p *glm4MoeLiteModel) KV(t *Tokenizer) KV {
+	kv := p.ModelParameters.KV(t)
+	kv["general.architecture"] = "glm4moelite"
+	kv["general.type"] = "model"
+	kv["glm4moelite.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["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["glm4moelite.rope.dimension_count"] = p.QKRopeHeadDim
+	kv["glm4moelite.rope.freq_base"] = cmp.Or(p.RopeTheta, float32(1000000.0))
+
+	kv["tokenizer.ggml.pre"] = "glm4"
+
+	return kv
+}
+
+func (p *glm4MoeLiteModel) Replacements() []string {
+	return []string{
+		"lm_head", "output",
+		"model.embed_tokens", "token_embd",
+		"model.norm", "output_norm",
+		"model.layers", "blk",
+		"input_layernorm", "attn_norm",
+		"self_attn.kv_a_proj_with_mqa", "attn_kv_a_mqa",
+		"self_attn.kv_a_layernorm", "attn_kv_a_norm",
+		"self_attn.kv_b_proj", "attn_kv_b",
+		"self_attn.q_a_proj", "attn_q_a",
+		"self_attn.q_a_layernorm", "attn_q_a_norm",
+		"self_attn.q_b_proj", "attn_q_b",
+		"self_attn.o_proj", "attn_output",
+		"post_attention_layernorm", "ffn_norm",
+		"mlp.shared_experts.down_proj", "ffn_down_shexp",
+		"mlp.shared_experts.gate_proj", "ffn_gate_shexp",
+		"mlp.shared_experts.up_proj", "ffn_up_shexp",
+		"mlp.gate_proj", "ffn_gate",
+		"mlp.down_proj", "ffn_down",
+		"mlp.up_proj", "ffn_up",
+		"mlp.gate.e_score_correction_bias", "exp_probs_b.bias",
+		"mlp.gate", "ffn_gate_inp",
+	}
+}
+
+func (p *glm4MoeLiteModel) Tensors(s []Tensor) (out []*ggml.Tensor) {
+	merges := make([]merge, p.HiddenLayers*3)
+	for i := range p.HiddenLayers {
+		merges[i*3+0] = merge{
+			fmt.Sprintf("blk.%d.mlp.experts.*.gate_proj.weight", i),
+			fmt.Sprintf("blk.%d.ffn_gate_exps.weight", i),
+		}
+		merges[i*3+1] = merge{
+			fmt.Sprintf("blk.%d.mlp.experts.*.up_proj.weight", i),
+			fmt.Sprintf("blk.%d.ffn_up_exps.weight", i),
+		}
+		merges[i*3+2] = merge{
+			fmt.Sprintf("blk.%d.mlp.experts.*.down_proj.weight", i),
+			fmt.Sprintf("blk.%d.ffn_down_exps.weight", i),
+		}
+	}
+
+	skipLayer := func(n string, minValue uint32) bool {
+		re := regexp.MustCompile(`^blk\.(\d+)`)
+		matches := re.FindStringSubmatch(n)
+		if matches == nil {
+			return false
+		}
+
+		blkNum, err := strconv.Atoi(matches[1])
+		if err != nil {
+			return false
+		}
+
+		return uint32(blkNum) >= minValue
+	}
+
+	out, s = mergeTensors(s, merges...)
+	for _, t := range s {
+		// skip any additional layers (such as the Multi-Token Prediction layer)
+		if skipLayer(t.Name(), p.HiddenLayers) {
+			slog.Debug("skipping layer", "name", t.Name())
+			continue
+		}
+		out = append(out, &ggml.Tensor{
+			Name:     t.Name(),
+			Kind:     t.Kind(),
+			Shape:    t.Shape(),
+			WriterTo: t,
+		})
+	}
+	return out
+}

convert/convert_gptoss.go

@@ -37,7 +37,7 @@ type gptossModel struct {
 
 var _ ModelConverter = (*gptossModel)(nil)
 
-func (m *gptossModel) KV(t *Tokenizer) ggml.KV {
+func (m *gptossModel) KV(t *Tokenizer) KV {
 	kv := m.ModelParameters.KV(t)
 	kv["general.architecture"] = "gptoss"
 	kv["general.file_type"] = uint32(4)

convert/convert_llama.go

@@ -48,7 +48,7 @@ type llamaModel struct {
 
 var _ ModelConverter = (*llamaModel)(nil)
 
-func (p *llamaModel) KV(t *Tokenizer) ggml.KV {
+func (p *llamaModel) KV(t *Tokenizer) KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "llama"
 	kv["llama.vocab_size"] = p.VocabSize

convert/convert_llama4.go

@@ -35,7 +35,7 @@ type llama4Model struct {
 }
 
 // KV implements ModelConverter.
-func (p *llama4Model) KV(t *Tokenizer) ggml.KV {
+func (p *llama4Model) KV(t *Tokenizer) KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "llama4"
 

convert/convert_llama_adapter.go

@@ -7,6 +7,7 @@ import (
 	"github.com/pdevine/tensor"
 	"github.com/pdevine/tensor/native"
 
+	"github.com/ollama/ollama/fs"
 	"github.com/ollama/ollama/fs/ggml"
 )
 
@@ -18,13 +19,13 @@ type llamaAdapter struct {
 
 var _ AdapterConverter = (*llamaAdapter)(nil)
 
-func (p *llamaAdapter) KV(baseKV ggml.KV) ggml.KV {
+func (p *llamaAdapter) KV(baseKV fs.Config) KV {
 	kv := p.AdapterParameters.KV()
 	kv["general.architecture"] = "llama"
-	kv["llama.attention.head_count"] = baseKV["llama.attention.head_count"]
-	kv["llama.attention.head_count_kv"] = baseKV["llama.attention.head_count_kv"]
+	kv["llama.attention.head_count"] = baseKV.Value("llama.attention.head_count")
+	kv["llama.attention.head_count_kv"] = baseKV.Value("llama.attention.head_count_kv")
 
-	p.NumAttentionHeads = baseKV["llama.attention.head_count"].(uint32)
+	p.NumAttentionHeads = baseKV.Value("llama.attention.head_count").(uint32)
 
 	return kv
 }

convert/convert_mistral.go

@@ -60,7 +60,7 @@ type mistral3Model struct {
 	ProjectorHiddenAct      string `json:"projector_hidden_act"`
 }
 
-func (p *mistral3Model) KV(t *Tokenizer) ggml.KV {
+func (p *mistral3Model) KV(t *Tokenizer) KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "mistral3"
 	kv["mistral3.vocab_size"] = p.TextModel.VocabSize

convert/convert_mistral_causal.go

@@ -39,7 +39,7 @@ type mistral3CausalModel struct {
 	} `json:"rope_parameters"`
 }
 
-func (p *mistral3CausalModel) KV(t *Tokenizer) ggml.KV {
+func (p *mistral3CausalModel) KV(t *Tokenizer) KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "mistral3"
 	kv["mistral3.vocab_size"] = p.VocabSize

convert/convert_mixtral.go

@@ -12,7 +12,7 @@ type mixtralModel struct {
 	NumExpertsPerToken uint32 `json:"num_experts_per_tok"`
 }
 
-func (p *mixtralModel) KV(t *Tokenizer) ggml.KV {
+func (p *mixtralModel) KV(t *Tokenizer) KV {
 	kv := p.llamaModel.KV(t)
 
 	if p.NumLocalExperts > 0 {

convert/convert_mllama.go

@@ -34,7 +34,7 @@ type mllamaModel struct {
 	} `json:"vision_config"`
 }
 
-func (m *mllamaModel) KV(t *Tokenizer) ggml.KV {
+func (m *mllamaModel) KV(t *Tokenizer) KV {
 	kv := m.ModelParameters.KV(t)
 	kv["general.architecture"] = "mllama"
 

convert/convert_nomicbert.go

@@ -87,7 +87,7 @@ func (p *nomicbertModel) parseMore(fsys fs.FS) error {
 	return nil
 }
 
-func (p *nomicbertModel) KV(t *Tokenizer) ggml.KV {
+func (p *nomicbertModel) KV(t *Tokenizer) KV {
 	kv := p.ModelParameters.KV(t)
 
 	// Determine architecture based on MoE parameters (following qwen3 pattern)

convert/convert_olmo.go

@@ -34,7 +34,7 @@ type olmoModel struct {
 
 var _ ModelConverter = (*olmoModel)(nil)
 
-func (p *olmoModel) KV(t *Tokenizer) ggml.KV {
+func (p *olmoModel) KV(t *Tokenizer) KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "olmo3"
 	kv["olmo3.block_count"] = p.NumHiddenLayers

convert/convert_phi3.go

@@ -37,7 +37,7 @@ type phi3Model struct {
 
 var _ ModelConverter = (*phi3Model)(nil)
 
-func (p *phi3Model) KV(t *Tokenizer) ggml.KV {
+func (p *phi3Model) KV(t *Tokenizer) KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "phi3"
 	kv["phi3.context_length"] = p.MaxPositionEmbeddings

convert/convert_qwen2.go

@@ -22,7 +22,7 @@ type qwen2Model struct {
 
 var _ ModelConverter = (*qwen2Model)(nil)
 
-func (q *qwen2Model) KV(t *Tokenizer) ggml.KV {
+func (q *qwen2Model) KV(t *Tokenizer) KV {
 	kv := q.ModelParameters.KV(t)
 	kv["general.architecture"] = "qwen2"
 	kv["qwen2.block_count"] = q.HiddenLayers

convert/convert_qwen25vl.go

@@ -29,7 +29,7 @@ type qwen25VLModel struct {
 
 var _ ModelConverter = (*qwen25VLModel)(nil)
 
-func (q *qwen25VLModel) KV(t *Tokenizer) ggml.KV {
+func (q *qwen25VLModel) KV(t *Tokenizer) KV {
 	kv := q.ModelParameters.KV(t)
 	kv["general.architecture"] = "qwen25vl"
 

convert/convert_qwen3.go

@@ -32,7 +32,7 @@ type qwen3Model struct {
 }
 
 // KV implements ModelConverter.
-func (q *qwen3Model) KV(t *Tokenizer) ggml.KV {
+func (q *qwen3Model) KV(t *Tokenizer) KV {
 	arch := "qwen3"
 	if q.NumExperts > 0 {
 		arch += "moe"

convert/convert_qwen3vl.go

@@ -45,7 +45,7 @@ func (m *qwen3VLModel) parseMore(fsys fs.FS) error {
 	return json.Unmarshal(bts, &m.VisionModel)
 }
 
-func (m *qwen3VLModel) KV(t *Tokenizer) ggml.KV {
+func (m *qwen3VLModel) KV(t *Tokenizer) KV {
 	kv := m.qwen3Model.KV(t)
 
 	arch := "qwen3vl"

convert/convert_test.go

@@ -19,6 +19,7 @@ import (
 	"testing"
 
 	"github.com/google/go-cmp/cmp"
+	fsc "github.com/ollama/ollama/fs"
 	"github.com/ollama/ollama/fs/ggml"
 )
 
@@ -28,7 +29,7 @@ type tensorData struct {
 	Shape   []int  `json:"shape"`
 }
 
-func convertFull(t *testing.T, fsys fs.FS) (*os.File, ggml.KV, ggml.Tensors) {
+func convertFull(t *testing.T, fsys fs.FS) (*os.File, fsc.Config, ggml.Tensors) {
 	t.Helper()
 
 	f, err := os.CreateTemp(t.TempDir(), "f16")
@@ -59,9 +60,10 @@ func convertFull(t *testing.T, fsys fs.FS) (*os.File, ggml.KV, ggml.Tensors) {
 	return r, m.KV(), m.Tensors()
 }
 
-func generateResultsJSON(t *testing.T, f *os.File, kv ggml.KV, tensors ggml.Tensors) map[string]string {
+func generateResultsJSON(t *testing.T, f *os.File, kv fsc.Config, tensors ggml.Tensors) map[string]string {
 	actual := make(map[string]string)
-	for k, v := range kv {
+	for k := range kv.Keys() {
+		v := kv.Value(k)
 		if s, ok := v.(json.Marshaler); !ok {
 			actual[k] = fmt.Sprintf("%v", v)
 		} else {
@@ -277,7 +279,7 @@ func generateSafetensorTestData(t *testing.T, tempDir string, tensorData map[str
 func TestConvertAdapter(t *testing.T) {
 	type AdapterCase struct {
 		Name     string
-		BaseKV   map[string]any
+		BaseKV   KV
 		Expected map[string]string
 	}
 

docs/README.md

@@ -14,6 +14,7 @@
 * [API Reference](https://docs.ollama.com/api)
 * [Modelfile Reference](https://docs.ollama.com/modelfile)
 * [OpenAI Compatibility](https://docs.ollama.com/api/openai-compatibility)
+* [Anthropic Compatibility](./api/anthropic-compatibility.mdx)
 
 ### Resources
 

docs/api.md

@@ -16,6 +16,7 @@
 - [Generate Embeddings](#generate-embeddings)
 - [List Running Models](#list-running-models)
 - [Version](#version)
+- [Experimental: Image Generation](#image-generation-experimental)
 
 ## Conventions
 
@@ -58,6 +59,15 @@ Advanced parameters (optional):
 - `keep_alive`: controls how long the model will stay loaded into memory following the request (default: `5m`)
 - `context` (deprecated): the context parameter returned from a previous request to `/generate`, this can be used to keep a short conversational memory
 
+Experimental image generation parameters (for image generation models only):
+
+> [!WARNING]
+> These parameters are experimental and may change in future versions.
+
+- `width`: width of the generated image in pixels
+- `height`: height of the generated image in pixels
+- `steps`: number of diffusion steps
+
 #### Structured outputs
 
 Structured outputs are supported by providing a JSON schema in the `format` parameter. The model will generate a response that matches the schema. See the [structured outputs](#request-structured-outputs) example below.
@@ -1867,3 +1877,55 @@ curl http://localhost:11434/api/version
   "version": "0.5.1"
 }
 ```
+
+## Experimental Features
+
+### Image Generation (Experimental)
+
+> [!WARNING]
+> Image generation is experimental and may change in future versions.
+
+Image generation is now supported through the standard `/api/generate` endpoint when using image generation models. The API automatically detects when an image generation model is being used.
+
+See the [Generate a completion](#generate-a-completion) section for the full API documentation. The experimental image generation parameters (`width`, `height`, `steps`) are documented there.
+
+#### Example
+
+##### Request
+
+```shell
+curl http://localhost:11434/api/generate -d '{
+  "model": "x/z-image-turbo",
+  "prompt": "a sunset over mountains",
+  "width": 1024,
+  "height": 768
+}'
+```
+
+##### Response (streaming)
+
+Progress updates during generation:
+
+```json
+{
+  "model": "x/z-image-turbo",
+  "created_at": "2024-01-15T10:30:00.000000Z",
+  "completed": 5,
+  "total": 20,
+  "done": false
+}
+```
+
+##### Final Response
+
+```json
+{
+  "model": "x/z-image-turbo",
+  "created_at": "2024-01-15T10:30:15.000000Z",
+  "image": "iVBORw0KGgoAAAANSUhEUg...",
+  "done": true,
+  "done_reason": "stop",
+  "total_duration": 15000000000,
+  "load_duration": 2000000000
+}
+```

docs/api/anthropic-compatibility.mdx

@@ -0,0 +1,408 @@
+---
+title: Anthropic compatibility
+---
+
+Ollama provides compatibility with the [Anthropic Messages API](https://docs.anthropic.com/en/api/messages) to help connect existing applications to Ollama, including tools like Claude Code.
+
+## Recommended models
+
+For coding use cases, models like `glm-4.7:cloud`, `minimax-m2.1:cloud`, and `qwen3-coder` are recommended.
+
+Pull a model before use:
+```shell
+ollama pull qwen3-coder
+ollama pull glm-4.7:cloud
+```
+
+## Usage
+
+### Environment variables
+
+To use Ollama with tools that expect the Anthropic API (like Claude Code), set these environment variables:
+
+```shell
+export ANTHROPIC_AUTH_TOKEN=ollama  # required but ignored
+export ANTHROPIC_BASE_URL=http://localhost:11434
+export ANTHROPIC_API_KEY=ollama  # required but ignored
+```
+
+### Simple `/v1/messages` example
+
+<CodeGroup dropdown>
+
+```python basic.py
+import anthropic
+
+client = anthropic.Anthropic(
+    base_url='http://localhost:11434',
+    api_key='ollama',  # required but ignored
+)
+
+message = client.messages.create(
+    model='qwen3-coder',
+    max_tokens=1024,
+    messages=[
+        {'role': 'user', 'content': 'Hello, how are you?'}
+    ]
+)
+print(message.content[0].text)
+```
+
+```javascript basic.js
+import Anthropic from "@anthropic-ai/sdk";
+
+const anthropic = new Anthropic({
+  baseURL: "http://localhost:11434",
+  apiKey: "ollama", // required but ignored
+});
+
+const message = await anthropic.messages.create({
+  model: "qwen3-coder",
+  max_tokens: 1024,
+  messages: [{ role: "user", content: "Hello, how are you?" }],
+});
+
+console.log(message.content[0].text);
+```
+
+```shell basic.sh
+curl -X POST http://localhost:11434/v1/messages \
+-H "Content-Type: application/json" \
+-H "x-api-key: ollama" \
+-H "anthropic-version: 2023-06-01" \
+-d '{
+  "model": "qwen3-coder",
+  "max_tokens": 1024,
+  "messages": [{ "role": "user", "content": "Hello, how are you?" }]
+}'
+```
+
+</CodeGroup>
+
+### Streaming example
+
+<CodeGroup dropdown>
+
+```python streaming.py
+import anthropic
+
+client = anthropic.Anthropic(
+    base_url='http://localhost:11434',
+    api_key='ollama',
+)
+
+with client.messages.stream(
+    model='qwen3-coder',
+    max_tokens=1024,
+    messages=[{'role': 'user', 'content': 'Count from 1 to 10'}]
+) as stream:
+    for text in stream.text_stream:
+        print(text, end='', flush=True)
+```
+
+```javascript streaming.js
+import Anthropic from "@anthropic-ai/sdk";
+
+const anthropic = new Anthropic({
+  baseURL: "http://localhost:11434",
+  apiKey: "ollama",
+});
+
+const stream = await anthropic.messages.stream({
+  model: "qwen3-coder",
+  max_tokens: 1024,
+  messages: [{ role: "user", content: "Count from 1 to 10" }],
+});
+
+for await (const event of stream) {
+  if (
+    event.type === "content_block_delta" &&
+    event.delta.type === "text_delta"
+  ) {
+    process.stdout.write(event.delta.text);
+  }
+}
+```
+
+```shell streaming.sh
+curl -X POST http://localhost:11434/v1/messages \
+-H "Content-Type: application/json" \
+-d '{
+  "model": "qwen3-coder",
+  "max_tokens": 1024,
+  "stream": true,
+  "messages": [{ "role": "user", "content": "Count from 1 to 10" }]
+}'
+```
+
+</CodeGroup>
+
+### Tool calling example
+
+<CodeGroup dropdown>
+
+```python tools.py
+import anthropic
+
+client = anthropic.Anthropic(
+    base_url='http://localhost:11434',
+    api_key='ollama',
+)
+
+message = client.messages.create(
+    model='qwen3-coder',
+    max_tokens=1024,
+    tools=[
+        {
+            'name': 'get_weather',
+            'description': 'Get the current weather in a location',
+            'input_schema': {
+                'type': 'object',
+                'properties': {
+                    'location': {
+                        'type': 'string',
+                        'description': 'The city and state, e.g. San Francisco, CA'
+                    }
+                },
+                'required': ['location']
+            }
+        }
+    ],
+    messages=[{'role': 'user', 'content': "What's the weather in San Francisco?"}]
+)
+
+for block in message.content:
+    if block.type == 'tool_use':
+        print(f'Tool: {block.name}')
+        print(f'Input: {block.input}')
+```
+
+```javascript tools.js
+import Anthropic from "@anthropic-ai/sdk";
+
+const anthropic = new Anthropic({
+  baseURL: "http://localhost:11434",
+  apiKey: "ollama",
+});
+
+const message = await anthropic.messages.create({
+  model: "qwen3-coder",
+  max_tokens: 1024,
+  tools: [
+    {
+      name: "get_weather",
+      description: "Get the current weather in a location",
+      input_schema: {
+        type: "object",
+        properties: {
+          location: {
+            type: "string",
+            description: "The city and state, e.g. San Francisco, CA",
+          },
+        },
+        required: ["location"],
+      },
+    },
+  ],
+  messages: [{ role: "user", content: "What's the weather in San Francisco?" }],
+});
+
+for (const block of message.content) {
+  if (block.type === "tool_use") {
+    console.log("Tool:", block.name);
+    console.log("Input:", block.input);
+  }
+}
+```
+
+```shell tools.sh
+curl -X POST http://localhost:11434/v1/messages \
+-H "Content-Type: application/json" \
+-d '{
+  "model": "qwen3-coder",
+  "max_tokens": 1024,
+  "tools": [
+    {
+      "name": "get_weather",
+      "description": "Get the current weather in a location",
+      "input_schema": {
+        "type": "object",
+        "properties": {
+          "location": {
+            "type": "string",
+            "description": "The city and state"
+          }
+        },
+        "required": ["location"]
+      }
+    }
+  ],
+  "messages": [{ "role": "user", "content": "What is the weather in San Francisco?" }]
+}'
+```
+
+</CodeGroup>
+
+## Using with Claude Code
+
+[Claude Code](https://code.claude.com/docs/en/overview) can be configured to use Ollama as its backend:
+
+```shell
+ANTHROPIC_AUTH_TOKEN=ollama ANTHROPIC_BASE_URL=http://localhost:11434 ANTHROPIC_API_KEY=ollama claude --model qwen3-coder
+```
+
+Or set the environment variables in your shell profile:
+
+```shell
+export ANTHROPIC_AUTH_TOKEN=ollama
+export ANTHROPIC_BASE_URL=http://localhost:11434
+export ANTHROPIC_API_KEY=ollama
+```
+
+Then run Claude Code with any Ollama model:
+
+```shell
+# Local models
+claude --model qwen3-coder
+claude --model gpt-oss:20b
+
+# Cloud models
+claude --model glm-4.7:cloud
+claude --model minimax-m2.1:cloud
+```
+
+## Endpoints
+
+### `/v1/messages`
+
+#### Supported features
+
+- [x] Messages
+- [x] Streaming
+- [x] System prompts
+- [x] Multi-turn conversations
+- [x] Vision (images)
+- [x] Tools (function calling)
+- [x] Tool results
+- [x] Thinking/extended thinking
+
+#### Supported request fields
+
+- [x] `model`
+- [x] `max_tokens`
+- [x] `messages`
+  - [x] Text `content`
+  - [x] Image `content` (base64)
+  - [x] Array of content blocks
+  - [x] `tool_use` blocks
+  - [x] `tool_result` blocks
+  - [x] `thinking` blocks
+- [x] `system` (string or array)
+- [x] `stream`
+- [x] `temperature`
+- [x] `top_p`
+- [x] `top_k`
+- [x] `stop_sequences`
+- [x] `tools`
+- [x] `thinking`
+- [ ] `tool_choice`
+- [ ] `metadata`
+
+#### Supported response fields
+
+- [x] `id`
+- [x] `type`
+- [x] `role`
+- [x] `model`
+- [x] `content` (text, tool_use, thinking blocks)
+- [x] `stop_reason` (end_turn, max_tokens, tool_use)
+- [x] `usage` (input_tokens, output_tokens)
+
+#### Streaming events
+
+- [x] `message_start`
+- [x] `content_block_start`
+- [x] `content_block_delta` (text_delta, input_json_delta, thinking_delta)
+- [x] `content_block_stop`
+- [x] `message_delta`
+- [x] `message_stop`
+- [x] `ping`
+- [x] `error`
+
+## Models
+
+Ollama supports both local and cloud models.
+
+### Local models
+
+Pull a local model before use:
+
+```shell
+ollama pull qwen3-coder
+```
+
+Recommended local models:
+- `qwen3-coder` - Excellent for coding tasks
+- `gpt-oss:20b` - Strong general-purpose model
+
+### Cloud models
+
+Cloud models are available immediately without pulling:
+
+- `glm-4.7:cloud` - High-performance cloud model
+- `minimax-m2.1:cloud` - Fast cloud model
+
+### Default model names
+
+For tooling that relies on default Anthropic model names such as `claude-3-5-sonnet`, use `ollama cp` to copy an existing model name:
+
+```shell
+ollama cp qwen3-coder claude-3-5-sonnet
+```
+
+Afterwards, this new model name can be specified in the `model` field:
+
+```shell
+curl http://localhost:11434/v1/messages \
+    -H "Content-Type: application/json" \
+    -d '{
+        "model": "claude-3-5-sonnet",
+        "max_tokens": 1024,
+        "messages": [
+            {
+                "role": "user",
+                "content": "Hello!"
+            }
+        ]
+    }'
+```
+
+## Differences from the Anthropic API
+
+### Behavior differences
+
+- API key is accepted but not validated
+- `anthropic-version` header is accepted but not used
+- Token counts are approximations based on the underlying model's tokenizer
+
+### Not supported
+
+The following Anthropic API features are not currently supported:
+
+| Feature | Description |
+|---------|-------------|
+| `/v1/messages/count_tokens` | Token counting endpoint |
+| `tool_choice` | Forcing specific tool use or disabling tools |
+| `metadata` | Request metadata (user_id) |
+| Prompt caching | `cache_control` blocks for caching prefixes |
+| Batches API | `/v1/messages/batches` for async batch processing |
+| Citations | `citations` content blocks |
+| PDF support | `document` content blocks with PDF files |
+| Server-sent errors | `error` events during streaming (errors return HTTP status) |
+
+### Partial support
+
+| Feature | Status |
+|---------|--------|
+| Image content | Base64 images supported; URL images not supported |
+| Extended thinking | Basic support; `budget_tokens` accepted but not enforced |

docs/api/openai-compatibility.mdx

@@ -275,6 +275,73 @@ curl -X POST http://localhost:11434/v1/chat/completions \
 - [x] `dimensions`
 - [ ] `user`
 
+### `/v1/images/generations` (experimental)
+
+> Note: This endpoint is experimental and may change or be removed in future versions.
+
+Generate images using image generation models.
+
+<CodeGroup dropdown>
+
+```python images.py
+from openai import OpenAI
+
+client = OpenAI(
+    base_url='http://localhost:11434/v1/',
+    api_key='ollama',  # required but ignored
+)
+
+response = client.images.generate(
+    model='x/z-image-turbo',
+    prompt='A cute robot learning to paint',
+    size='1024x1024',
+    response_format='b64_json',
+)
+print(response.data[0].b64_json[:50] + '...')
+```
+
+```javascript images.js
+import OpenAI from "openai";
+
+const openai = new OpenAI({
+  baseURL: "http://localhost:11434/v1/",
+  apiKey: "ollama", // required but ignored
+});
+
+const response = await openai.images.generate({
+  model: "x/z-image-turbo",
+  prompt: "A cute robot learning to paint",
+  size: "1024x1024",
+  response_format: "b64_json",
+});
+
+console.log(response.data[0].b64_json.slice(0, 50) + "...");
+```
+
+```shell images.sh
+curl -X POST http://localhost:11434/v1/images/generations \
+-H "Content-Type: application/json" \
+-d '{
+  "model": "x/z-image-turbo",
+  "prompt": "A cute robot learning to paint",
+  "size": "1024x1024",
+  "response_format": "b64_json"
+}'
+```
+
+</CodeGroup>
+
+#### Supported request fields
+
+- [x] `model`
+- [x] `prompt`
+- [x] `size` (e.g. "1024x1024")
+- [x] `response_format` (only `b64_json` supported)
+- [ ] `n`
+- [ ] `quality`
+- [ ] `style`
+- [ ] `user`
+
 ### `/v1/responses`
 
 > Note: Added in Ollama v0.13.3

docs/capabilities/web-search.mdx

@@ -110,7 +110,7 @@ More Ollama [Python example](https://github.com/ollama/ollama-python/blob/main/e
 import { Ollama } from "ollama";
 
 const client = new Ollama();
-const results = await client.webSearch({ query: "what is ollama?" });
+const results = await client.webSearch("what is ollama?");
 console.log(JSON.stringify(results, null, 2));
 ```
 
@@ -213,7 +213,7 @@ models](https://ollama.com/models)\n\nAvailable for macOS, Windows, and Linux',
 import { Ollama } from "ollama";
 
 const client = new Ollama();
-const fetchResult = await client.webFetch({ url: "https://ollama.com" });
+const fetchResult = await client.webFetch("https://ollama.com");
 console.log(JSON.stringify(fetchResult, null, 2));
 ```
 

docs/docs.json

@@ -32,7 +32,9 @@
     "codeblocks": "system"
   },
   "contextual": {
-    "options": ["copy"]
+    "options": [
+      "copy"
+    ]
   },
   "navbar": {
     "links": [
@@ -52,7 +54,9 @@
       "display": "simple"
     },
     "examples": {
-      "languages": ["curl"]
+      "languages": [
+        "curl"
+      ]
     }
   },
   "redirects": [
@@ -97,6 +101,7 @@
           {
             "group": "Integrations",
             "pages": [
+              "/integrations/claude-code",
               "/integrations/vscode",
               "/integrations/jetbrains",
               "/integrations/codex",
@@ -106,7 +111,9 @@
               "/integrations/zed",
               "/integrations/roo-code",
               "/integrations/n8n",
-              "/integrations/xcode"
+              "/integrations/xcode",
+              "/integrations/onyx",
+              "/integrations/marimo"
             ]
           },
           {
@@ -139,7 +146,8 @@
               "/api/streaming",
               "/api/usage",
               "/api/errors",
-              "/api/openai-compatibility"
+              "/api/openai-compatibility",
+              "/api/anthropic-compatibility"
             ]
           },
           {

docs/faq.mdx

@@ -22,7 +22,7 @@ Please refer to the [GPU docs](./gpu).
 
 ## How can I specify the context window size?
 
-By default, Ollama uses a context window size of 2048 tokens.
+By default, Ollama uses a context window size of 4096 tokens.
 
 This can be overridden with the `OLLAMA_CONTEXT_LENGTH` environment variable. For example, to set the default context window to 8K, use:
 

docs/integrations/claude-code.mdx

@@ -0,0 +1,78 @@
+---
+title: Claude Code
+---
+
+Claude Code is Anthropic's agentic coding tool that can read, modify, and execute code in your working directory. 
+
+Open models can be used with Claude Code through Ollama's Anthropic-compatible API, enabling you to use models such as `qwen3-coder`, `gpt-oss:20b`, or other models.
+
+![Claude Code with Ollama](https://files.ollama.com/claude-code.png)
+
+## Install
+
+Install [Claude Code](https://code.claude.com/docs/en/overview):
+
+<CodeGroup>
+
+```shell macOS / Linux
+curl -fsSL https://claude.ai/install.sh | bash
+```
+
+```powershell Windows
+irm https://claude.ai/install.ps1 | iex
+```
+
+</CodeGroup>
+
+## Usage with Ollama
+
+Claude Code connects to Ollama using the Anthropic-compatible API.
+
+1. Set the environment variables:
+
+```shell
+export ANTHROPIC_AUTH_TOKEN=ollama
+export ANTHROPIC_BASE_URL=http://localhost:11434
+```
+
+2. Run Claude Code with an Ollama model:
+
+```shell
+claude --model gpt-oss:20b
+```
+
+Or run with environment variables inline:
+
+```shell
+ANTHROPIC_AUTH_TOKEN=ollama ANTHROPIC_BASE_URL=http://localhost:11434 claude --model gpt-oss:20b
+```
+
+**Note:** Claude Code requires a large context window. We recommend at least 32K tokens. See the [context length documentation](/context-length) for how to adjust context length in Ollama.
+
+## Connecting to ollama.com
+
+1. Create an [API key](https://ollama.com/settings/keys) on ollama.com
+2. Set the environment variables:
+
+```shell
+export ANTHROPIC_BASE_URL=https://ollama.com
+export ANTHROPIC_API_KEY=<your-api-key>
+```
+
+3. Run Claude Code with a cloud model:
+
+```shell
+claude --model glm-4.7:cloud
+```
+
+## Recommended Models
+
+### Cloud models
+- `glm-4.7:cloud` - High-performance cloud model
+- `minimax-m2.1:cloud` - Fast cloud model
+- `qwen3-coder:480b` - Large coding model
+
+### Local models
+- `qwen3-coder` - Excellent for coding tasks
+- `gpt-oss:20b` - Strong general-purpose model
+- `gpt-oss:120b` - Larger general-purpose model for more complex tasks

docs/integrations/marimo.mdx

@@ -0,0 +1,73 @@
+---
+title: marimo
+---
+
+## Install
+
+Install [marimo](https://marimo.io). You can use `pip` or `uv` for this. You 
+can also use `uv` to create a sandboxed environment for marimo by running:
+
+```
+uvx marimo edit --sandbox notebook.py
+```
+
+## Usage with Ollama
+
+1. In marimo, go to the user settings and go to the AI tab. From here
+you can find and configure Ollama as an AI provider. For local use you
+would typically point the base url to `http://localhost:11434/v1`.
+
+<div style={{ display: 'flex', justifyContent: 'center' }}>
+  <img 
+    src="/images/marimo-settings.png" 
+    alt="Ollama settings in marimo"
+    width="50%"
+  />
+</div>
+
+2. Once the AI provider is set up, you can turn on/off specific AI models you'd like to access. 
+
+<div style={{ display: 'flex', justifyContent: 'center' }}>
+  <img 
+    src="/images/marimo-models.png" 
+    alt="Selecting an Ollama model"
+    width="50%"
+  />
+</div>
+
+3. You can also add a model to the list of available models by scrolling to the bottom and using the UI there. 
+
+<div style={{ display: 'flex', justifyContent: 'center' }}>
+  <img 
+    src="/images/marimo-add-model.png" 
+    alt="Adding a new Ollama model"
+    width="50%"
+  />
+</div>
+
+4. Once configured, you can now use Ollama for AI chats in marimo.
+
+<div style={{ display: 'flex', justifyContent: 'center' }}>
+  <img 
+    src="/images/marimo-chat.png" 
+    alt="Configure code completion"
+    width="50%"
+  />
+</div>
+
+4. Alternatively, you can now use Ollama for **inline code completion** in marimo. This can be configured in the "AI Features" tab. 
+
+<div style={{ display: 'flex', justifyContent: 'center' }}>
+  <img 
+    src="/images/marimo-code-completion.png" 
+    alt="Configure code completion"
+    width="50%"
+  />
+</div>
+
+
+## Connecting to ollama.com
+
+1. Sign in to ollama cloud via `ollama signin` 
+2. In the ollama model settings add a model that ollama hosts, like `gpt-oss:120b`.
+3. You can now refer to this model in marimo!

docs/integrations/onyx.mdx

@@ -0,0 +1,63 @@
+---
+title: Onyx
+---
+
+## Overview
+[Onyx](http://onyx.app/) is a self-hostable Chat UI that integrates with all Ollama models. Features include:
+- Creating custom Agents
+- Web search
+- Deep Research
+- RAG over uploaded documents and connected apps
+- Connectors to applications like Google Drive, Email, Slack, etc.
+- MCP and OpenAPI Actions support
+- Image generation
+- User/Groups management, RBAC, SSO, etc.
+
+Onyx can be deployed for single users or large organizations.
+
+## Install Onyx
+
+Deploy Onyx with the [quickstart guide](https://docs.onyx.app/deployment/getting_started/quickstart).
+
+<Info>
+Resourcing/scaling docs [here](https://docs.onyx.app/deployment/getting_started/resourcing).
+</Info>
+
+## Usage with Ollama 
+
+1. Login to your Onyx deployment (create an account first).
+<div style={{ display: 'flex', justifyContent: 'center' }}>
+  <img 
+    src="/images/onyx-login.png" 
+    alt="Onyx Login Page"
+    width="75%"
+  />
+</div>
+2. In the set-up process select `Ollama` as the LLM provider.
+<div style={{ display: 'flex', justifyContent: 'center' }}>
+  <img 
+    src="/images/onyx-ollama-llm.png" 
+    alt="Onyx Set Up Form"
+    width="75%"
+  />
+</div>
+3. Provide your **Ollama API URL** and select your models.
+<Note>If you're running Onyx in Docker, to access your computer's local network use `http://host.docker.internal` instead of `http://127.0.0.1`.</Note>
+<div style={{ display: 'flex', justifyContent: 'center' }}>
+  <img 
+    src="/images/onyx-ollama-form.png" 
+    alt="Selecting Ollama Models"
+    width="75%"
+  />
+</div>
+
+You can also easily connect up Onyx Cloud with the `Ollama Cloud` tab of the setup.
+
+## Send your first query
+<div style={{ display: 'flex', justifyContent: 'center' }}>
+  <img 
+    src="/images/onyx-query.png" 
+    alt="Onyx Query Example"
+    width="75%"
+  />
+</div>

docs/linux.mdx

@@ -20,8 +20,8 @@ curl -fsSL https://ollama.com/install.sh | sh
 Download and extract the package:
 
 ```shell
-curl -fsSL https://ollama.com/download/ollama-linux-amd64.tgz \
-    | sudo tar zx -C /usr
+curl -fsSL https://ollama.com/download/ollama-linux-amd64.tar.zst \
+    | sudo tar x -C /usr
 ```
 
 Start Ollama:
@@ -41,8 +41,8 @@ ollama -v
 If you have an AMD GPU, also download and extract the additional ROCm package:
 
 ```shell
-curl -fsSL https://ollama.com/download/ollama-linux-amd64-rocm.tgz \
-    | sudo tar zx -C /usr
+curl -fsSL https://ollama.com/download/ollama-linux-amd64-rocm.tar.zst \
+    | sudo tar x -C /usr
 ```
 
 ### ARM64 install
@@ -50,8 +50,8 @@ curl -fsSL https://ollama.com/download/ollama-linux-amd64-rocm.tgz \
 Download and extract the ARM64-specific package:
 
 ```shell
-curl -fsSL https://ollama.com/download/ollama-linux-arm64.tgz \
-    | sudo tar zx -C /usr
+curl -fsSL https://ollama.com/download/ollama-linux-arm64.tar.zst \
+    | sudo tar x -C /usr
 ```
 
 ### Adding Ollama as a startup service (recommended)
@@ -146,8 +146,8 @@ curl -fsSL https://ollama.com/install.sh | sh
 Or by re-downloading Ollama:
 
 ```shell
-curl -fsSL https://ollama.com/download/ollama-linux-amd64.tgz \
-    | sudo tar zx -C /usr
+curl -fsSL https://ollama.com/download/ollama-linux-amd64.tar.zst \
+    | sudo tar x -C /usr
 ```
 
 ## Installing specific versions

docs/troubleshooting.md

@@ -1,3 +0,0 @@
-# Troubleshooting
-
-For troubleshooting, see [https://docs.ollama.com/troubleshooting](https://docs.ollama.com/troubleshooting)

fs/config.go

@@ -1,5 +1,7 @@
 package fs
 
+import "iter"
+
 type Config interface {
 	Architecture() string
 	String(string, ...string) string
@@ -11,4 +13,8 @@ type Config interface {
 	Ints(string, ...[]int32) []int32
 	Floats(string, ...[]float32) []float32
 	Bools(string, ...[]bool) []bool
+
+	Len() int
+	Keys() iter.Seq[string]
+	Value(key string) any
 }

fs/ggml/ggml.go

@@ -6,7 +6,9 @@ import (
 	"errors"
 	"fmt"
 	"io"
+	"iter"
 	"log/slog"
+	"maps"
 	"math"
 	"slices"
 	"strings"
@@ -239,6 +241,18 @@ func (kv KV) Bools(key string, defaultValue ...[]bool) []bool {
 	return val.values
 }
 
+func (kv KV) Len() int {
+	return len(kv)
+}
+
+func (kv KV) Keys() iter.Seq[string] {
+	return maps.Keys(kv)
+}
+
+func (kv KV) Value(key string) any {
+	return kv[key]
+}
+
 func (kv KV) OllamaEngineRequired() bool {
 	return slices.Contains([]string{
 		"bert",
@@ -255,6 +269,7 @@ func (kv KV) OllamaEngineRequired() bool {
 		"qwen25vl",
 		"qwen3", "qwen3moe",
 		"qwen3vl", "qwen3vlmoe",
+		"glm4moelite",
 	}, kv.Architecture())
 }
 
@@ -842,6 +857,7 @@ func (f GGML) FlashAttention() bool {
 	return slices.Contains([]string{
 		"bert",
 		"gemma3",
+		"glm4moelite",
 		"gptoss", "gpt-oss",
 		"mistral3",
 		"olmo3",

fs/ggml/gguf.go

@@ -8,12 +8,12 @@ import (
 	"fmt"
 	"io"
 	"log/slog"
-	"maps"
 	"os"
 	"runtime"
 	"slices"
 	"strings"
 
+	"github.com/ollama/ollama/fs"
 	"golang.org/x/sync/errgroup"
 )
 
@@ -508,7 +508,7 @@ func writeGGUFArray[S ~[]E, E any](w io.Writer, t uint32, s S) error {
 	return binary.Write(w, binary.LittleEndian, s)
 }
 
-func WriteGGUF(f *os.File, kv KV, ts []*Tensor) error {
+func WriteGGUF(f *os.File, kv fs.Config, ts []*Tensor) error {
 	arch := kv.String("general.architecture")
 	if arch == "" {
 		return fmt.Errorf("architecture not set")
@@ -526,12 +526,12 @@ func WriteGGUF(f *os.File, kv KV, ts []*Tensor) error {
 		return err
 	}
 
-	if err := binary.Write(f, binary.LittleEndian, uint64(len(kv))); err != nil {
+	if err := binary.Write(f, binary.LittleEndian, uint64(kv.Len())); err != nil {
 		return err
 	}
 
-	for _, key := range slices.Sorted(maps.Keys(kv)) {
-		if err := ggufWriteKV(f, arch, key, kv[key]); err != nil {
+	for _, key := range slices.Sorted(kv.Keys()) {
+		if err := ggufWriteKV(f, arch, key, kv.Value(key)); err != nil {
 			return err
 		}
 	}

integration/tools_test.go

@@ -131,7 +131,7 @@ func TestAPIToolCalling(t *testing.T) {
 					t.Errorf("unexpected tool called: got %q want %q", lastToolCall.Function.Name, "get_weather")
 				}
 
-				if _, ok := lastToolCall.Function.Arguments["location"]; !ok {
+				if _, ok := lastToolCall.Function.Arguments.Get("location"); !ok {
 					t.Errorf("expected tool arguments to include 'location', got: %s", lastToolCall.Function.Arguments.String())
 				}
 			case <-ctx.Done():

llm/server.go

@@ -1464,6 +1464,12 @@ type CompletionRequest struct {
 
 	// TopLogprobs specifies the number of most likely alternative tokens to return (0-20)
 	TopLogprobs int
+
+	// Image generation fields
+	Width  int32 `json:"width,omitempty"`
+	Height int32 `json:"height,omitempty"`
+	Steps  int32 `json:"steps,omitempty"`
+	Seed   int64 `json:"seed,omitempty"`
 }
 
 // DoneReason represents the reason why a completion response is done
@@ -1512,6 +1518,15 @@ type CompletionResponse struct {
 
 	// Logprobs contains log probability information if requested
 	Logprobs []Logprob `json:"logprobs,omitempty"`
+
+	// Image contains base64-encoded image data for image generation
+	Image string `json:"image,omitempty"`
+
+	// Step is the current step in image generation
+	Step int `json:"step,omitempty"`
+
+	// TotalSteps is the total number of steps for image generation
+	TotalSteps int `json:"total_steps,omitempty"`
 }
 
 func (s *llmServer) Completion(ctx context.Context, req CompletionRequest, fn func(CompletionResponse)) error {

middleware/anthropic.go

@@ -0,0 +1,152 @@
+package middleware
+
+import (
+	"bytes"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+
+	"github.com/gin-gonic/gin"
+
+	"github.com/ollama/ollama/anthropic"
+	"github.com/ollama/ollama/api"
+)
+
+// AnthropicWriter wraps the response writer to transform Ollama responses to Anthropic format
+type AnthropicWriter struct {
+	BaseWriter
+	stream    bool
+	id        string
+	model     string
+	converter *anthropic.StreamConverter
+}
+
+func (w *AnthropicWriter) writeError(data []byte) (int, error) {
+	var errData struct {
+		Error string `json:"error"`
+	}
+	if err := json.Unmarshal(data, &errData); err != nil {
+		return 0, err
+	}
+
+	w.ResponseWriter.Header().Set("Content-Type", "application/json")
+	err := json.NewEncoder(w.ResponseWriter).Encode(anthropic.NewError(w.ResponseWriter.Status(), errData.Error))
+	if err != nil {
+		return 0, err
+	}
+
+	return len(data), nil
+}
+
+func (w *AnthropicWriter) writeEvent(eventType string, data any) error {
+	d, err := json.Marshal(data)
+	if err != nil {
+		return err
+	}
+	_, err = w.ResponseWriter.Write([]byte(fmt.Sprintf("event: %s\ndata: %s\n\n", eventType, d)))
+	if err != nil {
+		return err
+	}
+	if f, ok := w.ResponseWriter.(http.Flusher); ok {
+		f.Flush()
+	}
+	return nil
+}
+
+func (w *AnthropicWriter) writeResponse(data []byte) (int, error) {
+	var chatResponse api.ChatResponse
+	err := json.Unmarshal(data, &chatResponse)
+	if err != nil {
+		return 0, err
+	}
+
+	if w.stream {
+		w.ResponseWriter.Header().Set("Content-Type", "text/event-stream")
+
+		events := w.converter.Process(chatResponse)
+		for _, event := range events {
+			if err := w.writeEvent(event.Event, event.Data); err != nil {
+				return 0, err
+			}
+		}
+		return len(data), nil
+	}
+
+	w.ResponseWriter.Header().Set("Content-Type", "application/json")
+	response := anthropic.ToMessagesResponse(w.id, chatResponse)
+	return len(data), json.NewEncoder(w.ResponseWriter).Encode(response)
+}
+
+func (w *AnthropicWriter) Write(data []byte) (int, error) {
+	code := w.ResponseWriter.Status()
+	if code != http.StatusOK {
+		return w.writeError(data)
+	}
+
+	return w.writeResponse(data)
+}
+
+// AnthropicMessagesMiddleware handles Anthropic Messages API requests
+func AnthropicMessagesMiddleware() gin.HandlerFunc {
+	return func(c *gin.Context) {
+		var req anthropic.MessagesRequest
+		err := c.ShouldBindJSON(&req)
+		if err != nil {
+			c.AbortWithStatusJSON(http.StatusBadRequest, anthropic.NewError(http.StatusBadRequest, err.Error()))
+			return
+		}
+
+		if req.Model == "" {
+			c.AbortWithStatusJSON(http.StatusBadRequest, anthropic.NewError(http.StatusBadRequest, "model is required"))
+			return
+		}
+
+		if req.MaxTokens <= 0 {
+			c.AbortWithStatusJSON(http.StatusBadRequest, anthropic.NewError(http.StatusBadRequest, "max_tokens is required and must be positive"))
+			return
+		}
+
+		if len(req.Messages) == 0 {
+			c.AbortWithStatusJSON(http.StatusBadRequest, anthropic.NewError(http.StatusBadRequest, "messages is required"))
+			return
+		}
+
+		chatReq, err := anthropic.FromMessagesRequest(req)
+		if err != nil {
+			c.AbortWithStatusJSON(http.StatusBadRequest, anthropic.NewError(http.StatusBadRequest, err.Error()))
+			return
+		}
+
+		// Set think to nil when being used with Anthropic API to connect to tools like claude code
+		c.Set("relax_thinking", true)
+
+		var b bytes.Buffer
+		if err := json.NewEncoder(&b).Encode(chatReq); err != nil {
+			c.AbortWithStatusJSON(http.StatusInternalServerError, anthropic.NewError(http.StatusInternalServerError, err.Error()))
+			return
+		}
+
+		c.Request.Body = io.NopCloser(&b)
+
+		messageID := anthropic.GenerateMessageID()
+
+		w := &AnthropicWriter{
+			BaseWriter: BaseWriter{ResponseWriter: c.Writer},
+			stream:     req.Stream,
+			id:         messageID,
+			model:      req.Model,
+			converter:  anthropic.NewStreamConverter(messageID, req.Model),
+		}
+
+		if req.Stream {
+			c.Writer.Header().Set("Content-Type", "text/event-stream")
+			c.Writer.Header().Set("Cache-Control", "no-cache")
+			c.Writer.Header().Set("Connection", "keep-alive")
+		}
+
+		c.Writer = w
+
+		c.Next()
+	}
+}

middleware/anthropic_test.go

@@ -0,0 +1,607 @@
+package middleware
+
+import (
+	"bytes"
+	"encoding/json"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	"github.com/gin-gonic/gin"
+	"github.com/google/go-cmp/cmp"
+	"github.com/google/go-cmp/cmp/cmpopts"
+
+	"github.com/ollama/ollama/anthropic"
+	"github.com/ollama/ollama/api"
+)
+
+func captureAnthropicRequest(capturedRequest any) gin.HandlerFunc {
+	return func(c *gin.Context) {
+		bodyBytes, _ := io.ReadAll(c.Request.Body)
+		c.Request.Body = io.NopCloser(bytes.NewReader(bodyBytes))
+		_ = json.Unmarshal(bodyBytes, capturedRequest)
+		c.Next()
+	}
+}
+
+// testProps creates ToolPropertiesMap from a map (convenience function for tests)
+func testProps(m map[string]api.ToolProperty) *api.ToolPropertiesMap {
+	props := api.NewToolPropertiesMap()
+	for k, v := range m {
+		props.Set(k, v)
+	}
+	return props
+}
+
+func TestAnthropicMessagesMiddleware(t *testing.T) {
+	type testCase struct {
+		name string
+		body string
+		req  api.ChatRequest
+		err  anthropic.ErrorResponse
+	}
+
+	var capturedRequest *api.ChatRequest
+	stream := true
+
+	testCases := []testCase{
+		{
+			name: "basic message",
+			body: `{
+				"model": "test-model",
+				"max_tokens": 1024,
+				"messages": [
+					{"role": "user", "content": "Hello"}
+				]
+			}`,
+			req: api.ChatRequest{
+				Model: "test-model",
+				Messages: []api.Message{
+					{Role: "user", Content: "Hello"},
+				},
+				Options: map[string]any{"num_predict": 1024},
+				Stream:  &False,
+			},
+		},
+		{
+			name: "with system prompt",
+			body: `{
+				"model": "test-model",
+				"max_tokens": 1024,
+				"system": "You are helpful.",
+				"messages": [
+					{"role": "user", "content": "Hello"}
+				]
+			}`,
+			req: api.ChatRequest{
+				Model: "test-model",
+				Messages: []api.Message{
+					{Role: "system", Content: "You are helpful."},
+					{Role: "user", Content: "Hello"},
+				},
+				Options: map[string]any{"num_predict": 1024},
+				Stream:  &False,
+			},
+		},
+		{
+			name: "with options",
+			body: `{
+				"model": "test-model",
+				"max_tokens": 2048,
+				"temperature": 0.7,
+				"top_p": 0.9,
+				"top_k": 40,
+				"stop_sequences": ["\n", "END"],
+				"messages": [
+					{"role": "user", "content": "Hello"}
+				]
+			}`,
+			req: api.ChatRequest{
+				Model: "test-model",
+				Messages: []api.Message{
+					{Role: "user", Content: "Hello"},
+				},
+				Options: map[string]any{
+					"num_predict": 2048,
+					"temperature": 0.7,
+					"top_p":       0.9,
+					"top_k":       40,
+					"stop":        []string{"\n", "END"},
+				},
+				Stream: &False,
+			},
+		},
+		{
+			name: "streaming",
+			body: `{
+				"model": "test-model",
+				"max_tokens": 1024,
+				"stream": true,
+				"messages": [
+					{"role": "user", "content": "Hello"}
+				]
+			}`,
+			req: api.ChatRequest{
+				Model: "test-model",
+				Messages: []api.Message{
+					{Role: "user", Content: "Hello"},
+				},
+				Options: map[string]any{"num_predict": 1024},
+				Stream:  &stream,
+			},
+		},
+		{
+			name: "with tools",
+			body: `{
+				"model": "test-model",
+				"max_tokens": 1024,
+				"messages": [
+					{"role": "user", "content": "What's the weather?"}
+				],
+				"tools": [{
+					"name": "get_weather",
+					"description": "Get current weather",
+					"input_schema": {
+						"type": "object",
+						"properties": {
+							"location": {"type": "string"}
+						},
+						"required": ["location"]
+					}
+				}]
+			}`,
+			req: api.ChatRequest{
+				Model: "test-model",
+				Messages: []api.Message{
+					{Role: "user", Content: "What's the weather?"},
+				},
+				Tools: []api.Tool{
+					{
+						Type: "function",
+						Function: api.ToolFunction{
+							Name:        "get_weather",
+							Description: "Get current weather",
+							Parameters: api.ToolFunctionParameters{
+								Type:     "object",
+								Required: []string{"location"},
+								Properties: testProps(map[string]api.ToolProperty{
+									"location": {Type: api.PropertyType{"string"}},
+								}),
+							},
+						},
+					},
+				},
+				Options: map[string]any{"num_predict": 1024},
+				Stream:  &False,
+			},
+		},
+		{
+			name: "with tool result",
+			body: `{
+				"model": "test-model",
+				"max_tokens": 1024,
+				"messages": [
+					{"role": "user", "content": "What's the weather?"},
+					{"role": "assistant", "content": [
+						{"type": "tool_use", "id": "call_123", "name": "get_weather", "input": {"location": "Paris"}}
+					]},
+					{"role": "user", "content": [
+						{"type": "tool_result", "tool_use_id": "call_123", "content": "Sunny, 22°C"}
+					]}
+				]
+			}`,
+			req: api.ChatRequest{
+				Model: "test-model",
+				Messages: []api.Message{
+					{Role: "user", Content: "What's the weather?"},
+					{
+						Role: "assistant",
+						ToolCalls: []api.ToolCall{
+							{
+								ID: "call_123",
+								Function: api.ToolCallFunction{
+									Name:      "get_weather",
+									Arguments: testArgs(map[string]any{"location": "Paris"}),
+								},
+							},
+						},
+					},
+					{Role: "tool", Content: "Sunny, 22°C", ToolCallID: "call_123"},
+				},
+				Options: map[string]any{"num_predict": 1024},
+				Stream:  &False,
+			},
+		},
+		{
+			name: "with thinking enabled",
+			body: `{
+				"model": "test-model",
+				"max_tokens": 1024,
+				"thinking": {"type": "enabled", "budget_tokens": 1000},
+				"messages": [
+					{"role": "user", "content": "Hello"}
+				]
+			}`,
+			req: api.ChatRequest{
+				Model: "test-model",
+				Messages: []api.Message{
+					{Role: "user", Content: "Hello"},
+				},
+				Options: map[string]any{"num_predict": 1024},
+				Stream:  &False,
+				Think:   &api.ThinkValue{Value: true},
+			},
+		},
+		{
+			name: "missing model error",
+			body: `{
+				"max_tokens": 1024,
+				"messages": [
+					{"role": "user", "content": "Hello"}
+				]
+			}`,
+			err: anthropic.ErrorResponse{
+				Type: "error",
+				Error: anthropic.Error{
+					Type:    "invalid_request_error",
+					Message: "model is required",
+				},
+			},
+		},
+		{
+			name: "missing max_tokens error",
+			body: `{
+				"model": "test-model",
+				"messages": [
+					{"role": "user", "content": "Hello"}
+				]
+			}`,
+			err: anthropic.ErrorResponse{
+				Type: "error",
+				Error: anthropic.Error{
+					Type:    "invalid_request_error",
+					Message: "max_tokens is required and must be positive",
+				},
+			},
+		},
+		{
+			name: "missing messages error",
+			body: `{
+				"model": "test-model",
+				"max_tokens": 1024
+			}`,
+			err: anthropic.ErrorResponse{
+				Type: "error",
+				Error: anthropic.Error{
+					Type:    "invalid_request_error",
+					Message: "messages is required",
+				},
+			},
+		},
+		{
+			name: "tool_use missing id error",
+			body: `{
+				"model": "test-model",
+				"max_tokens": 1024,
+				"messages": [
+					{"role": "assistant", "content": [
+						{"type": "tool_use", "name": "test"}
+					]}
+				]
+			}`,
+			err: anthropic.ErrorResponse{
+				Type: "error",
+				Error: anthropic.Error{
+					Type:    "invalid_request_error",
+					Message: "tool_use block missing required 'id' field",
+				},
+			},
+		},
+	}
+
+	endpoint := func(c *gin.Context) {
+		c.Status(http.StatusOK)
+	}
+
+	gin.SetMode(gin.TestMode)
+	router := gin.New()
+	router.Use(AnthropicMessagesMiddleware(), captureAnthropicRequest(&capturedRequest))
+	router.Handle(http.MethodPost, "/v1/messages", endpoint)
+
+	for _, tc := range testCases {
+		t.Run(tc.name, func(t *testing.T) {
+			req, _ := http.NewRequest(http.MethodPost, "/v1/messages", strings.NewReader(tc.body))
+			req.Header.Set("Content-Type", "application/json")
+
+			defer func() { capturedRequest = nil }()
+
+			resp := httptest.NewRecorder()
+			router.ServeHTTP(resp, req)
+
+			if tc.err.Type != "" {
+				// Expect error
+				if resp.Code == http.StatusOK {
+					t.Fatalf("expected error response, got 200 OK")
+				}
+				var errResp anthropic.ErrorResponse
+				if err := json.Unmarshal(resp.Body.Bytes(), &errResp); err != nil {
+					t.Fatalf("failed to unmarshal error: %v", err)
+				}
+				if errResp.Type != tc.err.Type {
+					t.Errorf("expected error type %q, got %q", tc.err.Type, errResp.Type)
+				}
+				if errResp.Error.Type != tc.err.Error.Type {
+					t.Errorf("expected error.type %q, got %q", tc.err.Error.Type, errResp.Error.Type)
+				}
+				if errResp.Error.Message != tc.err.Error.Message {
+					t.Errorf("expected error.message %q, got %q", tc.err.Error.Message, errResp.Error.Message)
+				}
+				return
+			}
+
+			if resp.Code != http.StatusOK {
+				t.Fatalf("unexpected status code: %d, body: %s", resp.Code, resp.Body.String())
+			}
+
+			if capturedRequest == nil {
+				t.Fatal("request was not captured")
+			}
+
+			// Compare relevant fields
+			if capturedRequest.Model != tc.req.Model {
+				t.Errorf("model mismatch: got %q, want %q", capturedRequest.Model, tc.req.Model)
+			}
+
+			if diff := cmp.Diff(tc.req.Messages, capturedRequest.Messages,
+				cmpopts.IgnoreUnexported(api.ToolCallFunctionArguments{}, api.ToolPropertiesMap{})); diff != "" {
+				t.Errorf("messages mismatch (-want +got):\n%s", diff)
+			}
+
+			if tc.req.Stream != nil && capturedRequest.Stream != nil {
+				if *tc.req.Stream != *capturedRequest.Stream {
+					t.Errorf("stream mismatch: got %v, want %v", *capturedRequest.Stream, *tc.req.Stream)
+				}
+			}
+
+			if tc.req.Think != nil {
+				if capturedRequest.Think == nil {
+					t.Error("expected Think to be set")
+				} else if capturedRequest.Think.Value != tc.req.Think.Value {
+					t.Errorf("Think mismatch: got %v, want %v", capturedRequest.Think.Value, tc.req.Think.Value)
+				}
+			}
+		})
+	}
+}
+
+func TestAnthropicMessagesMiddleware_Headers(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
+	t.Run("streaming sets correct headers", func(t *testing.T) {
+		router := gin.New()
+		router.Use(AnthropicMessagesMiddleware())
+		router.POST("/v1/messages", func(c *gin.Context) {
+			// Check headers were set
+			if c.Writer.Header().Get("Content-Type") != "text/event-stream" {
+				t.Errorf("expected Content-Type text/event-stream, got %q", c.Writer.Header().Get("Content-Type"))
+			}
+			if c.Writer.Header().Get("Cache-Control") != "no-cache" {
+				t.Errorf("expected Cache-Control no-cache, got %q", c.Writer.Header().Get("Cache-Control"))
+			}
+			c.Status(http.StatusOK)
+		})
+
+		body := `{"model": "test", "max_tokens": 100, "stream": true, "messages": [{"role": "user", "content": "Hi"}]}`
+		req, _ := http.NewRequest(http.MethodPost, "/v1/messages", strings.NewReader(body))
+		req.Header.Set("Content-Type", "application/json")
+
+		resp := httptest.NewRecorder()
+		router.ServeHTTP(resp, req)
+	})
+}
+
+func TestAnthropicMessagesMiddleware_InvalidJSON(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	router := gin.New()
+	router.Use(AnthropicMessagesMiddleware())
+	router.POST("/v1/messages", func(c *gin.Context) {
+		c.Status(http.StatusOK)
+	})
+
+	req, _ := http.NewRequest(http.MethodPost, "/v1/messages", strings.NewReader(`{invalid json`))
+	req.Header.Set("Content-Type", "application/json")
+
+	resp := httptest.NewRecorder()
+	router.ServeHTTP(resp, req)
+
+	if resp.Code != http.StatusBadRequest {
+		t.Errorf("expected status 400, got %d", resp.Code)
+	}
+
+	var errResp anthropic.ErrorResponse
+	if err := json.Unmarshal(resp.Body.Bytes(), &errResp); err != nil {
+		t.Fatalf("failed to unmarshal error: %v", err)
+	}
+
+	if errResp.Type != "error" {
+		t.Errorf("expected type 'error', got %q", errResp.Type)
+	}
+	if errResp.Error.Type != "invalid_request_error" {
+		t.Errorf("expected error type 'invalid_request_error', got %q", errResp.Error.Type)
+	}
+}
+
+func TestAnthropicWriter_NonStreaming(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
+	router := gin.New()
+	router.Use(AnthropicMessagesMiddleware())
+	router.POST("/v1/messages", func(c *gin.Context) {
+		// Simulate Ollama response
+		resp := api.ChatResponse{
+			Model: "test-model",
+			Message: api.Message{
+				Role:    "assistant",
+				Content: "Hello there!",
+			},
+			Done:       true,
+			DoneReason: "stop",
+			Metrics: api.Metrics{
+				PromptEvalCount: 10,
+				EvalCount:       5,
+			},
+		}
+		data, _ := json.Marshal(resp)
+		c.Writer.WriteHeader(http.StatusOK)
+		_, _ = c.Writer.Write(data)
+	})
+
+	body := `{"model": "test-model", "max_tokens": 100, "messages": [{"role": "user", "content": "Hi"}]}`
+	req, _ := http.NewRequest(http.MethodPost, "/v1/messages", strings.NewReader(body))
+	req.Header.Set("Content-Type", "application/json")
+
+	resp := httptest.NewRecorder()
+	router.ServeHTTP(resp, req)
+
+	if resp.Code != http.StatusOK {
+		t.Fatalf("expected status 200, got %d", resp.Code)
+	}
+
+	var result anthropic.MessagesResponse
+	if err := json.Unmarshal(resp.Body.Bytes(), &result); err != nil {
+		t.Fatalf("failed to unmarshal response: %v", err)
+	}
+
+	if result.Type != "message" {
+		t.Errorf("expected type 'message', got %q", result.Type)
+	}
+	if result.Role != "assistant" {
+		t.Errorf("expected role 'assistant', got %q", result.Role)
+	}
+	if len(result.Content) != 1 {
+		t.Fatalf("expected 1 content block, got %d", len(result.Content))
+	}
+	if result.Content[0].Text == nil || *result.Content[0].Text != "Hello there!" {
+		t.Errorf("expected text 'Hello there!', got %v", result.Content[0].Text)
+	}
+	if result.StopReason != "end_turn" {
+		t.Errorf("expected stop_reason 'end_turn', got %q", result.StopReason)
+	}
+	if result.Usage.InputTokens != 10 {
+		t.Errorf("expected input_tokens 10, got %d", result.Usage.InputTokens)
+	}
+	if result.Usage.OutputTokens != 5 {
+		t.Errorf("expected output_tokens 5, got %d", result.Usage.OutputTokens)
+	}
+}
+
+// TestAnthropicWriter_ErrorFromRoutes tests error handling when routes.go sends
+// gin.H{"error": "message"} without a StatusCode field (which is the common case)
+func TestAnthropicWriter_ErrorFromRoutes(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
+	tests := []struct {
+		name          string
+		statusCode    int
+		errorPayload  any
+		wantErrorType string
+		wantMessage   string
+	}{
+		// routes.go sends errors without StatusCode in JSON, so we must use HTTP status
+		{
+			name:          "404 with gin.H error (model not found)",
+			statusCode:    http.StatusNotFound,
+			errorPayload:  gin.H{"error": "model 'nonexistent' not found"},
+			wantErrorType: "not_found_error",
+			wantMessage:   "model 'nonexistent' not found",
+		},
+		{
+			name:          "400 with gin.H error (bad request)",
+			statusCode:    http.StatusBadRequest,
+			errorPayload:  gin.H{"error": "model is required"},
+			wantErrorType: "invalid_request_error",
+			wantMessage:   "model is required",
+		},
+		{
+			name:          "500 with gin.H error (internal error)",
+			statusCode:    http.StatusInternalServerError,
+			errorPayload:  gin.H{"error": "something went wrong"},
+			wantErrorType: "api_error",
+			wantMessage:   "something went wrong",
+		},
+		{
+			name:       "404 with api.StatusError",
+			statusCode: http.StatusNotFound,
+			errorPayload: api.StatusError{
+				StatusCode:   http.StatusNotFound,
+				ErrorMessage: "model not found via StatusError",
+			},
+			wantErrorType: "not_found_error",
+			wantMessage:   "model not found via StatusError",
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			router := gin.New()
+			router.Use(AnthropicMessagesMiddleware())
+			router.POST("/v1/messages", func(c *gin.Context) {
+				// Simulate what routes.go does - set status and write error JSON
+				data, _ := json.Marshal(tt.errorPayload)
+				c.Writer.WriteHeader(tt.statusCode)
+				_, _ = c.Writer.Write(data)
+			})
+
+			body := `{"model": "test-model", "max_tokens": 100, "messages": [{"role": "user", "content": "Hi"}]}`
+			req, _ := http.NewRequest(http.MethodPost, "/v1/messages", strings.NewReader(body))
+			req.Header.Set("Content-Type", "application/json")
+
+			resp := httptest.NewRecorder()
+			router.ServeHTTP(resp, req)
+
+			if resp.Code != tt.statusCode {
+				t.Errorf("expected status %d, got %d", tt.statusCode, resp.Code)
+			}
+
+			var errResp anthropic.ErrorResponse
+			if err := json.Unmarshal(resp.Body.Bytes(), &errResp); err != nil {
+				t.Fatalf("failed to unmarshal error response: %v\nbody: %s", err, resp.Body.String())
+			}
+
+			if errResp.Type != "error" {
+				t.Errorf("expected type 'error', got %q", errResp.Type)
+			}
+			if errResp.Error.Type != tt.wantErrorType {
+				t.Errorf("expected error type %q, got %q", tt.wantErrorType, errResp.Error.Type)
+			}
+			if errResp.Error.Message != tt.wantMessage {
+				t.Errorf("expected message %q, got %q", tt.wantMessage, errResp.Error.Message)
+			}
+		})
+	}
+}
+
+func TestAnthropicMessagesMiddleware_SetsRelaxThinkingFlag(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
+	var flagSet bool
+	router := gin.New()
+	router.Use(AnthropicMessagesMiddleware())
+	router.POST("/v1/messages", func(c *gin.Context) {
+		_, flagSet = c.Get("relax_thinking")
+		c.Status(http.StatusOK)
+	})
+
+	body := `{"model": "test-model", "max_tokens": 100, "messages": [{"role": "user", "content": "Hi"}]}`
+	req, _ := http.NewRequest(http.MethodPost, "/v1/messages", strings.NewReader(body))
+	req.Header.Set("Content-Type", "application/json")
+
+	resp := httptest.NewRecorder()
+	router.ServeHTTP(resp, req)
+
+	if !flagSet {
+		t.Error("expected relax_thinking flag to be set in context")
+	}
+}

middleware/openai.go

@@ -8,6 +8,7 @@ import (
 	"math/rand"
 	"net/http"
 	"strings"
+	"time"
 
 	"github.com/gin-gonic/gin"
 
@@ -441,6 +442,7 @@ type ResponsesWriter struct {
 	stream     bool
 	responseID string
 	itemID     string
+	request    openai.ResponsesRequest
 }
 
 func (w *ResponsesWriter) writeEvent(eventType string, data any) error {
@@ -478,7 +480,9 @@ func (w *ResponsesWriter) writeResponse(data []byte) (int, error) {
 
 	// Non-streaming response
 	w.ResponseWriter.Header().Set("Content-Type", "application/json")
-	response := openai.ToResponse(w.model, w.responseID, w.itemID, chatResponse)
+	response := openai.ToResponse(w.model, w.responseID, w.itemID, chatResponse, w.request)
+	completedAt := time.Now().Unix()
+	response.CompletedAt = &completedAt
 	return len(data), json.NewEncoder(w.ResponseWriter).Encode(response)
 }
 
@@ -523,11 +527,12 @@ func ResponsesMiddleware() gin.HandlerFunc {
 
 		w := &ResponsesWriter{
 			BaseWriter: BaseWriter{ResponseWriter: c.Writer},
-			converter:  openai.NewResponsesStreamConverter(responseID, itemID, req.Model),
+			converter:  openai.NewResponsesStreamConverter(responseID, itemID, req.Model, req),
 			model:      req.Model,
 			stream:     streamRequested,
 			responseID: responseID,
 			itemID:     itemID,
+			request:    req,
 		}
 
 		// Set headers based on streaming mode
@@ -541,3 +546,66 @@ func ResponsesMiddleware() gin.HandlerFunc {
 		c.Next()
 	}
 }
+
+type ImageWriter struct {
+	BaseWriter
+}
+
+func (w *ImageWriter) writeResponse(data []byte) (int, error) {
+	var generateResponse api.GenerateResponse
+	if err := json.Unmarshal(data, &generateResponse); err != nil {
+		return 0, err
+	}
+
+	// Only write response when done with image
+	if generateResponse.Done && generateResponse.Image != "" {
+		w.ResponseWriter.Header().Set("Content-Type", "application/json")
+		return len(data), json.NewEncoder(w.ResponseWriter).Encode(openai.ToImageGenerationResponse(generateResponse))
+	}
+
+	return len(data), nil
+}
+
+func (w *ImageWriter) Write(data []byte) (int, error) {
+	code := w.ResponseWriter.Status()
+	if code != http.StatusOK {
+		return w.writeError(data)
+	}
+
+	return w.writeResponse(data)
+}
+
+func ImageGenerationsMiddleware() gin.HandlerFunc {
+	return func(c *gin.Context) {
+		var req openai.ImageGenerationRequest
+		if err := c.ShouldBindJSON(&req); err != nil {
+			c.AbortWithStatusJSON(http.StatusBadRequest, openai.NewError(http.StatusBadRequest, err.Error()))
+			return
+		}
+
+		if req.Prompt == "" {
+			c.AbortWithStatusJSON(http.StatusBadRequest, openai.NewError(http.StatusBadRequest, "prompt is required"))
+			return
+		}
+
+		if req.Model == "" {
+			c.AbortWithStatusJSON(http.StatusBadRequest, openai.NewError(http.StatusBadRequest, "model is required"))
+			return
+		}
+
+		var b bytes.Buffer
+		if err := json.NewEncoder(&b).Encode(openai.FromImageGenerationRequest(req)); err != nil {
+			c.AbortWithStatusJSON(http.StatusInternalServerError, openai.NewError(http.StatusInternalServerError, err.Error()))
+			return
+		}
+
+		c.Request.Body = io.NopCloser(&b)
+
+		w := &ImageWriter{
+			BaseWriter: BaseWriter{ResponseWriter: c.Writer},
+		}
+
+		c.Writer = w
+		c.Next()
+	}
+}

middleware/openai_test.go

@@ -961,3 +961,154 @@ func TestRetrieveMiddleware(t *testing.T) {
 		}
 	}
 }
+
+func TestImageGenerationsMiddleware(t *testing.T) {
+	type testCase struct {
+		name string
+		body string
+		req  api.GenerateRequest
+		err  openai.ErrorResponse
+	}
+
+	var capturedRequest *api.GenerateRequest
+
+	testCases := []testCase{
+		{
+			name: "image generation basic",
+			body: `{
+				"model": "test-model",
+				"prompt": "a beautiful sunset"
+			}`,
+			req: api.GenerateRequest{
+				Model:  "test-model",
+				Prompt: "a beautiful sunset",
+			},
+		},
+		{
+			name: "image generation with size",
+			body: `{
+				"model": "test-model",
+				"prompt": "a beautiful sunset",
+				"size": "512x768"
+			}`,
+			req: api.GenerateRequest{
+				Model:  "test-model",
+				Prompt: "a beautiful sunset",
+				Width:  512,
+				Height: 768,
+			},
+		},
+		{
+			name: "image generation missing prompt",
+			body: `{
+				"model": "test-model"
+			}`,
+			err: openai.ErrorResponse{
+				Error: openai.Error{
+					Message: "prompt is required",
+					Type:    "invalid_request_error",
+				},
+			},
+		},
+		{
+			name: "image generation missing model",
+			body: `{
+				"prompt": "a beautiful sunset"
+			}`,
+			err: openai.ErrorResponse{
+				Error: openai.Error{
+					Message: "model is required",
+					Type:    "invalid_request_error",
+				},
+			},
+		},
+	}
+
+	endpoint := func(c *gin.Context) {
+		c.Status(http.StatusOK)
+	}
+
+	gin.SetMode(gin.TestMode)
+	router := gin.New()
+	router.Use(ImageGenerationsMiddleware(), captureRequestMiddleware(&capturedRequest))
+	router.Handle(http.MethodPost, "/api/generate", endpoint)
+
+	for _, tc := range testCases {
+		t.Run(tc.name, func(t *testing.T) {
+			req, _ := http.NewRequest(http.MethodPost, "/api/generate", strings.NewReader(tc.body))
+			req.Header.Set("Content-Type", "application/json")
+
+			defer func() { capturedRequest = nil }()
+
+			resp := httptest.NewRecorder()
+			router.ServeHTTP(resp, req)
+
+			if tc.err.Error.Message != "" {
+				var errResp openai.ErrorResponse
+				if err := json.Unmarshal(resp.Body.Bytes(), &errResp); err != nil {
+					t.Fatal(err)
+				}
+				if diff := cmp.Diff(tc.err, errResp); diff != "" {
+					t.Fatalf("errors did not match:\n%s", diff)
+				}
+				return
+			}
+
+			if resp.Code != http.StatusOK {
+				t.Fatalf("expected status 200, got %d: %s", resp.Code, resp.Body.String())
+			}
+
+			if diff := cmp.Diff(&tc.req, capturedRequest); diff != "" {
+				t.Fatalf("requests did not match:\n%s", diff)
+			}
+		})
+	}
+}
+
+func TestImageWriterResponse(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
+	// Test that ImageWriter transforms GenerateResponse to OpenAI format
+	endpoint := func(c *gin.Context) {
+		resp := api.GenerateResponse{
+			Model:     "test-model",
+			CreatedAt: time.Unix(1234567890, 0).UTC(),
+			Done:      true,
+			Image:     "dGVzdC1pbWFnZS1kYXRh", // base64 of "test-image-data"
+		}
+		data, _ := json.Marshal(resp)
+		c.Writer.Write(append(data, '\n'))
+	}
+
+	router := gin.New()
+	router.Use(ImageGenerationsMiddleware())
+	router.Handle(http.MethodPost, "/api/generate", endpoint)
+
+	body := `{"model": "test-model", "prompt": "test"}`
+	req, _ := http.NewRequest(http.MethodPost, "/api/generate", strings.NewReader(body))
+	req.Header.Set("Content-Type", "application/json")
+
+	resp := httptest.NewRecorder()
+	router.ServeHTTP(resp, req)
+
+	if resp.Code != http.StatusOK {
+		t.Fatalf("expected status 200, got %d: %s", resp.Code, resp.Body.String())
+	}
+
+	var imageResp openai.ImageGenerationResponse
+	if err := json.Unmarshal(resp.Body.Bytes(), &imageResp); err != nil {
+		t.Fatalf("failed to unmarshal response: %v", err)
+	}
+
+	if imageResp.Created != 1234567890 {
+		t.Errorf("expected created 1234567890, got %d", imageResp.Created)
+	}
+
+	if len(imageResp.Data) != 1 {
+		t.Fatalf("expected 1 image, got %d", len(imageResp.Data))
+	}
+
+	if imageResp.Data[0].B64JSON != "dGVzdC1pbWFnZS1kYXRh" {
+		t.Errorf("expected image data 'dGVzdC1pbWFnZS1kYXRh', got %s", imageResp.Data[0].B64JSON)
+	}
+}

model/models/glm4moelite/model.go

@@ -0,0 +1,304 @@
+package glm4moelite
+
+import (
+	"math"
+
+	"github.com/ollama/ollama/fs"
+	"github.com/ollama/ollama/kvcache"
+	"github.com/ollama/ollama/ml"
+	"github.com/ollama/ollama/ml/nn"
+	"github.com/ollama/ollama/model"
+	"github.com/ollama/ollama/model/input"
+)
+
+type Options struct {
+	numExpertsUsed      int
+	numExperts          int
+	normTopKProb        bool
+	routedScalingFactor float32
+
+	kvLoraRank,
+	qkNopeHeadDim,
+	qkRopeHeadDim,
+	kqNopeHeadDim,
+	qkHeadDim int
+	qLoraRank int
+	vHeadDim  int
+
+	hiddenSize,
+	numHeads,
+	numKVHeads int
+
+	eps,
+	ropeBase float32
+	kqScale float64
+}
+
+func (o Options) applyRotaryPositionEmbeddings(ctx ml.Context, t, p ml.Tensor) ml.Tensor {
+	return nn.RoPE(ctx, t, p, o.qkRopeHeadDim, o.ropeBase, 1.0)
+}
+
+type Attention struct {
+	Q *nn.Linear `gguf:"attn_q"`
+
+	QA     *nn.Linear  `gguf:"attn_q_a"`
+	QANorm *nn.RMSNorm `gguf:"attn_q_a_norm"`
+	QB     *nn.Linear  `gguf:"attn_q_b"`
+
+	KVA     *nn.Linear  `gguf:"attn_kv_a_mqa"`
+	KVANorm *nn.RMSNorm `gguf:"attn_kv_a_norm"`
+	KVB     *nn.Linear  `gguf:"attn_kv_b"`
+
+	Output *nn.Linear `gguf:"attn_out,alt:attn_output"`
+}
+
+func (attn *Attention) Forward(ctx ml.Context, hiddenStates, positions ml.Tensor, cache kvcache.Cache, opts *Options) ml.Tensor {
+	seqLength := hiddenStates.Dim(1)
+
+	var query ml.Tensor
+	if opts.qLoraRank == 0 {
+		query = attn.Q.Forward(ctx, hiddenStates)
+	} else {
+		query = attn.QA.Forward(ctx, hiddenStates)
+		query = attn.QANorm.Forward(ctx, query, opts.eps)
+		query = attn.QB.Forward(ctx, query)
+	}
+
+	query = query.Reshape(ctx, query.Dim(0)/opts.numHeads, opts.numHeads, seqLength)
+	queryChunks := query.ChunkSections(ctx, 0, opts.qkNopeHeadDim, opts.qkRopeHeadDim)
+
+	compressedKV := attn.KVA.Forward(ctx, hiddenStates)
+	kPass := compressedKV.Slice(ctx, 0, 0, opts.kvLoraRank, 1)
+	kRot := compressedKV.View(ctx,
+		opts.kvLoraRank*compressedKV.Stride(0), opts.qkRopeHeadDim,
+		compressedKV.Stride(1), 1,
+		compressedKV.Stride(1), compressedKV.Dim(1),
+	)
+
+	qRot := opts.applyRotaryPositionEmbeddings(ctx, queryChunks[1], positions)
+	kRot = opts.applyRotaryPositionEmbeddings(ctx, kRot, positions)
+	kPass = attn.KVANorm.Forward(ctx, kPass, opts.eps)
+	kPass = attn.KVB.Forward(ctx, kPass)
+
+	kv := kPass.Reshape(ctx, kPass.Dim(0)/opts.numKVHeads, opts.numKVHeads, seqLength)
+	kvChunks := kv.ChunkSections(ctx, 0, opts.kqNopeHeadDim, opts.vHeadDim)
+
+	kRot = kRot.Repeat(ctx, 1, queryChunks[0].Dim(1))
+	query = qRot.Concat(ctx, queryChunks[0], 0)
+	key := kRot.Concat(ctx, kvChunks[0], 0)
+	attention := nn.Attention(ctx, query, key, kvChunks[1], opts.kqScale, cache)
+
+	attention = attention.Reshape(ctx, attention.Dim(0)*attention.Dim(1), seqLength)
+	return attn.Output.Forward(ctx, attention)
+}
+
+type MLP interface {
+	Forward(ml.Context, ml.Tensor, *Options) ml.Tensor
+}
+
+type sparse struct {
+	Router       *nn.Linear `gguf:"ffn_gate_inp"`
+	Gate         *nn.Linear `gguf:"ffn_gate_exps"`
+	Up           *nn.Linear `gguf:"ffn_up_exps"`
+	Down         *nn.Linear `gguf:"ffn_down_exps"`
+	SharedExpert *dense     `gguf:",suf:_shexp"`
+	ExpProbsBias ml.Tensor  `gguf:"exp_probs_b.bias,alt:exp_probs_b"`
+}
+
+func (moe *sparse) Moe(ctx ml.Context, hiddenStates, topKIndices, topKWeights ml.Tensor, opts *Options) ml.Tensor {
+	hiddenStates = hiddenStates.Reshape(ctx, hiddenStates.Dim(0), 1, hiddenStates.Dim(1))
+
+	upStates := moe.Up.Weight.MulmatID(ctx, hiddenStates, topKIndices)
+	hiddenStates = moe.Gate.Weight.MulmatID(ctx, hiddenStates, topKIndices)
+	hiddenStates = hiddenStates.SILU(ctx, upStates)
+
+	experts := moe.Down.Weight.MulmatID(ctx, hiddenStates, topKIndices)
+	experts = experts.Mul(ctx, topKWeights)
+
+	nextStates := experts.View(ctx, 0, experts.Dim(0), experts.Stride(2), experts.Dim(2))
+	for i := 1; i < opts.numExpertsUsed; i++ {
+		nextStates = nextStates.Add(ctx, experts.View(ctx, i*experts.Stride(1), experts.Dim(0), experts.Stride(2), experts.Dim(2)))
+	}
+	return nextStates
+}
+
+func (moe *sparse) topKIndices(ctx ml.Context, scores ml.Tensor, opts *Options) ml.Tensor {
+	if moe.ExpProbsBias != nil {
+		scores = scores.Add(ctx, moe.ExpProbsBias)
+	}
+	topKIndices := scores.TopK(ctx, opts.numExpertsUsed)
+	return topKIndices
+}
+
+func (moe *sparse) Forward(ctx ml.Context, hiddenStates ml.Tensor, opts *Options) ml.Tensor {
+	residuals := hiddenStates
+
+	routerLogits := moe.Router.Forward(ctx, hiddenStates)
+	scores := routerLogits.Sigmoid(ctx)
+	topKIndices := moe.topKIndices(ctx, scores, opts)
+	topKWeights := scores.Reshape(ctx, 1, opts.numExperts, hiddenStates.Dim(1)).Rows(ctx, topKIndices)
+
+	if opts.normTopKProb {
+		topKWeights = topKWeights.Reshape(ctx, opts.numExpertsUsed, hiddenStates.Dim(1))
+		topKWeights = topKWeights.Div(ctx, topKWeights.SumRows(ctx))
+		topKWeights = topKWeights.Reshape(ctx, 1, opts.numExpertsUsed, hiddenStates.Dim(1))
+	}
+
+	topKWeights = topKWeights.Scale(ctx, float64(opts.routedScalingFactor))
+	hiddenStates = moe.Moe(ctx, hiddenStates, topKIndices, topKWeights, opts)
+	sharedExpertResult := moe.SharedExpert.Forward(ctx, residuals, opts)
+
+	hiddenStates = hiddenStates.Add(ctx, sharedExpertResult)
+	return hiddenStates
+}
+
+type dense struct {
+	Gate *nn.Linear `gguf:"ffn_gate"`
+	Up   *nn.Linear `gguf:"ffn_up"`
+	Down *nn.Linear `gguf:"ffn_down"`
+}
+
+func (mlp *dense) Forward(ctx ml.Context, hiddenStates ml.Tensor, opts *Options) ml.Tensor {
+	hiddenStates = mlp.Gate.Forward(ctx, hiddenStates).SILU(ctx, mlp.Up.Forward(ctx, hiddenStates))
+	return mlp.Down.Forward(ctx, hiddenStates)
+}
+
+type Layer struct {
+	AttentionNorm *nn.RMSNorm `gguf:"attn_norm"`
+	Attention     *Attention
+
+	MLPNorm *nn.RMSNorm `gguf:"ffn_norm"`
+	MLP     MLP
+}
+
+func (t *Layer) Forward(ctx ml.Context, hiddenStates, positions, outputs ml.Tensor, cache kvcache.Cache, opts *Options) ml.Tensor {
+	residual := hiddenStates
+	hiddenStates = t.AttentionNorm.Forward(ctx, hiddenStates, opts.eps)
+	hiddenStates = t.Attention.Forward(ctx, hiddenStates, positions, cache, opts)
+
+	if outputs != nil {
+		hiddenStates = hiddenStates.Rows(ctx, outputs)
+		residual = residual.Rows(ctx, outputs)
+	}
+
+	hiddenStates = hiddenStates.Add(ctx, residual)
+	residual = hiddenStates
+
+	hiddenStates = t.MLPNorm.Forward(ctx, hiddenStates, opts.eps)
+	hiddenStates = t.MLP.Forward(ctx, hiddenStates, opts)
+	hiddenStates = hiddenStates.Add(ctx, residual)
+	return hiddenStates
+}
+
+type Model struct {
+	model.Base
+	model.BytePairEncoding
+
+	TokenEmbedding *nn.Embedding `gguf:"token_embd"`
+	Layers         []Layer       `gguf:"blk"`
+
+	OutputNorm *nn.RMSNorm `gguf:"output_norm"`
+	Output     *nn.Linear  `gguf:"output,alt:token_embd"`
+
+	*Options
+}
+
+func New(c fs.Config) (model.Model, error) {
+	layers := make([]Layer, c.Uint("block_count"))
+
+	firstDenseLayerIndex := int(c.Uint("leading_dense_block_count"))
+	for i := range layers {
+		if i < firstDenseLayerIndex {
+			layers[i].MLP = &dense{}
+		} else {
+			layers[i].MLP = &sparse{}
+		}
+	}
+
+	keyLength := int(c.Uint("attention.key_length"))
+	valueLength := int(c.Uint("attention.value_length"))
+
+	kqScale := 1.0 / math.Sqrt(float64(keyLength))
+
+	var pre []string
+	switch c.String("tokenizer.ggml.pre") {
+	case "glm4":
+		pre = []string{
+			`(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\r\n\p{L}\p{N}]?\p{L}+|\p{N}{1,3}| ?[^\s\p{L}\p{N}]+[\r\n]*|\s*[\r\n]+|\s+(?!\S)|\s+`,
+		}
+	default:
+		return nil, model.ErrUnsupportedTokenizer
+	}
+
+	m := Model{
+		BytePairEncoding: model.NewBytePairEncoding(
+			&model.Vocabulary{
+				Values: c.Strings("tokenizer.ggml.tokens"),
+				Types:  c.Ints("tokenizer.ggml.token_type"),
+				Merges: c.Strings("tokenizer.ggml.merges"),
+				AddBOS: c.Bool("tokenizer.ggml.add_bos_token", true),
+				BOS:    []int32{int32(c.Uint("tokenizer.ggml.bos_token_id"))},
+				AddEOS: c.Bool("tokenizer.ggml.add_eos_token", false),
+				EOS: append(
+					[]int32{int32(c.Uint("tokenizer.ggml.eos_token_id"))},
+					c.Ints("tokenizer.ggml.eos_token_ids")...,
+				),
+			},
+			pre...,
+		),
+		Layers: layers,
+		Options: &Options{
+			hiddenSize:     int(c.Uint("embedding_length")),
+			numHeads:       int(c.Uint("attention.head_count")),
+			numKVHeads:     int(c.Uint("attention.head_count_kv")),
+			eps:            c.Float("attention.layer_norm_rms_epsilon"),
+			ropeBase:       c.Float("rope.freq_base"),
+			numExperts:     int(c.Uint("expert_count")),
+			numExpertsUsed: int(c.Uint("expert_used_count")),
+			normTopKProb:   c.Bool("expert_weights_norm", true),
+
+			qLoraRank:     int(c.Uint("attention.q_lora_rank")),
+			kvLoraRank:    int(c.Uint("attention.kv_lora_rank")),
+			qkHeadDim:     keyLength,
+			vHeadDim:      valueLength,
+			qkRopeHeadDim: int(c.Uint("rope.dimension_count")),
+			qkNopeHeadDim: keyLength - int(c.Uint("rope.dimension_count")),
+			kqNopeHeadDim: keyLength - int(c.Uint("rope.dimension_count")),
+
+			routedScalingFactor: c.Float("expert_weights_scale"),
+
+			kqScale: kqScale,
+		},
+	}
+
+	m.Cache = kvcache.NewCausalCache(m.Shift)
+	return &m, nil
+}
+
+func (m Model) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) {
+	return m.applyRotaryPositionEmbeddings(ctx, key, shift), nil
+}
+
+func (m *Model) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
+	positions := ctx.Input().FromInts(batch.Positions, len(batch.Positions))
+
+	hiddenStates := m.TokenEmbedding.Forward(ctx, batch.Inputs)
+
+	for i, layer := range m.Layers {
+		m.Cache.SetLayer(i)
+
+		var outputs ml.Tensor
+		if i == len(m.Layers)-1 {
+			outputs = batch.Outputs
+		}
+
+		hiddenStates = layer.Forward(ctx, hiddenStates, positions, outputs, m.Cache, m.Options)
+	}
+
+	hiddenStates = m.OutputNorm.Forward(ctx, hiddenStates, m.eps)
+	return m.Output.Forward(ctx, hiddenStates), nil
+}
+
+func init() {
+	model.Register("glm4moelite", New)
+}

model/models/models.go

@@ -7,6 +7,7 @@ import (
 	_ "github.com/ollama/ollama/model/models/gemma2"
 	_ "github.com/ollama/ollama/model/models/gemma3"
 	_ "github.com/ollama/ollama/model/models/gemma3n"
+	_ "github.com/ollama/ollama/model/models/glm4moelite"
 	_ "github.com/ollama/ollama/model/models/gptoss"
 	_ "github.com/ollama/ollama/model/models/llama"
 	_ "github.com/ollama/ollama/model/models/llama4"

model/parsers/glm46.go

@@ -0,0 +1,410 @@
+package parsers
+
+import (
+	"context"
+	"encoding/xml"
+	"fmt"
+	"log/slog"
+	"strings"
+	"unicode"
+
+	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/logutil"
+)
+
+type glm46ParserState int
+
+const (
+	glm46ParserState_LookingForThinkingOpen glm46ParserState = iota
+	glm46ParserState_ThinkingStartedEatingWhitespace
+	glm46ParserState_CollectingThinking
+	glm46ParserState_ThinkingDoneEatingWhitespace
+	glm46ParserState_CollectingContent
+	glm46ParserState_ToolStartedEatingWhitespace
+	glm46ParserState_CollectingToolContent
+)
+
+const (
+	glm46ThinkingOpenTag  = "<think>"
+	glm46ThinkingCloseTag = "</think>"
+	glm46ToolOpenTag      = "<tool_call>"
+	glm46ToolCloseTag     = "</tool_call>"
+)
+
+type GLM46Parser struct {
+	state  glm46ParserState
+	buffer strings.Builder
+	tools  []api.Tool
+}
+
+func (p *GLM46Parser) HasToolSupport() bool {
+	return true
+}
+
+func (p *GLM46Parser) HasThinkingSupport() bool {
+	return true
+}
+
+// func (p *GLM46Parser) Init(tools []api.Tool, lastMessage *api.Message) []api.Tool {
+func (p *GLM46Parser) Init(tools []api.Tool, lastMessage *api.Message, thinkValue *api.ThinkValue) []api.Tool {
+	p.tools = tools
+	return tools
+}
+
+type glm46Event interface {
+	isGLM46Event()
+}
+
+type glm46EventContent struct {
+	content string
+}
+
+func (glm46EventContent) isGLM46Event() {}
+
+type glm46EventRawToolCall struct {
+	raw string
+}
+
+func (glm46EventRawToolCall) isGLM46Event() {}
+
+type glm46EventThinkingContent struct {
+	content string
+}
+
+func (glm46EventThinkingContent) isGLM46Event() {}
+
+func (p *GLM46Parser) Add(s string, done bool) (content string, thinking string, calls []api.ToolCall, err error) {
+	p.buffer.WriteString(s)
+	events := p.parseEvents()
+
+	var toolCalls []api.ToolCall
+	var contentSb strings.Builder
+	var thinkingSb strings.Builder
+
+	for _, event := range events {
+		switch event := event.(type) {
+		case glm46EventRawToolCall:
+			toolCall, err := parseGLM46ToolCall(event, p.tools)
+			if err != nil {
+				slog.Warn("glm-4.6 tool call parsing failed", "error", err)
+				return "", "", nil, err
+			}
+			toolCalls = append(toolCalls, toolCall)
+		case glm46EventThinkingContent:
+			thinkingSb.WriteString(event.content)
+		case glm46EventContent:
+			// TODO(drifkin): if the same turn contains multiple interleaved content
+			// events, we naively append them together here.
+			contentSb.WriteString(event.content)
+		}
+	}
+
+	return contentSb.String(), thinkingSb.String(), toolCalls, nil
+}
+
+func (p *GLM46Parser) parseEvents() []glm46Event {
+	var all []glm46Event
+
+	keepLooping := true
+	for keepLooping {
+		var events []glm46Event
+		events, keepLooping = p.eat()
+		if len(events) > 0 {
+			all = append(all, events...)
+		}
+	}
+
+	if len(all) > 0 {
+		slog.Log(context.TODO(), logutil.LevelTrace, "glm-4.6 events parsed", "events", all, "state", p.state, "buffer", p.buffer.String())
+	}
+
+	return all
+}
+
+// eatLeadingWhitespaceAndTransitionTo consumes leading whitespace from the buffer
+// and transitions to the next state. Returns (nil, false) if only whitespace remains
+// in the buffer (needs more input), or (nil, true) if we successfully transitioned.
+func (p *GLM46Parser) eatLeadingWhitespaceAndTransitionTo(nextState glm46ParserState) ([]glm46Event, bool) {
+	trimmed := strings.TrimLeftFunc(p.buffer.String(), unicode.IsSpace)
+	p.buffer.Reset()
+	if trimmed == "" {
+		return nil, false // Still only whitespace, keep waiting for more input
+	}
+	p.state = nextState
+	p.buffer.WriteString(trimmed)
+	return nil, true // Successfully transitioned
+}
+
+// glm46SplitAtTag splits the buffer at the given tag, returns the content before (trimmed of trailing whitespace),
+// the content after (optionally trimmed of leading whitespace), and updates the buffer
+func glm46SplitAtTag(p *GLM46Parser, tag string, trimAfter bool) (string, string) {
+	split := strings.SplitN(p.buffer.String(), tag, 2)
+	before := split[0]
+	before = strings.TrimRightFunc(before, unicode.IsSpace)
+	after := split[1]
+	if trimAfter {
+		after = strings.TrimLeftFunc(after, unicode.IsSpace)
+	}
+	p.buffer.Reset()
+	p.buffer.WriteString(after)
+	return before, after
+}
+
+func (p *GLM46Parser) eat() ([]glm46Event, bool) {
+	var events []glm46Event
+
+	switch p.state {
+	case glm46ParserState_LookingForThinkingOpen:
+		trimmed := strings.TrimLeftFunc(p.buffer.String(), unicode.IsSpace)
+		if strings.HasPrefix(trimmed, glm46ThinkingOpenTag) {
+			// Found <think> opening tag
+			after := strings.TrimPrefix(trimmed, glm46ThinkingOpenTag)
+			after = strings.TrimLeftFunc(after, unicode.IsSpace)
+			p.buffer.Reset()
+			p.buffer.WriteString(after)
+			if after == "" {
+				p.state = glm46ParserState_ThinkingStartedEatingWhitespace
+			} else {
+				p.state = glm46ParserState_CollectingThinking
+			}
+			return events, true
+		} else if strings.HasPrefix(glm46ThinkingOpenTag, trimmed) {
+			// Partial opening tag seen, keep accumulating
+			return events, false
+		} else if trimmed == "" {
+			// Only whitespace, keep accumulating
+			return events, false
+		} else {
+			// No thinking tag found, skip to content collection
+			p.state = glm46ParserState_CollectingContent
+			// Don't trim - we want to keep the original content
+			return events, true
+		}
+
+	case glm46ParserState_ThinkingStartedEatingWhitespace:
+		return p.eatLeadingWhitespaceAndTransitionTo(glm46ParserState_CollectingThinking)
+
+	case glm46ParserState_CollectingThinking:
+		acc := p.buffer.String()
+		if strings.Contains(acc, glm46ThinkingCloseTag) {
+			thinking, remaining := glm46SplitAtTag(p, glm46ThinkingCloseTag, true)
+			if len(thinking) > 0 {
+				events = append(events, glm46EventThinkingContent{content: thinking})
+			}
+			if remaining == "" {
+				p.state = glm46ParserState_ThinkingDoneEatingWhitespace
+			} else {
+				p.state = glm46ParserState_CollectingContent
+			}
+			return events, true
+		} else if overlapLen := overlap(acc, glm46ThinkingCloseTag); overlapLen > 0 {
+			// Partial closing tag - withhold it along with any trailing whitespace before it
+			beforePartialTag := acc[:len(acc)-overlapLen]
+			trailingWhitespaceLen := trailingWhitespaceLen(beforePartialTag)
+			ambiguousStart := len(beforePartialTag) - trailingWhitespaceLen
+
+			unambiguous := acc[:ambiguousStart]
+			ambiguous := acc[ambiguousStart:]
+			p.buffer.Reset()
+			p.buffer.WriteString(ambiguous)
+			if len(unambiguous) > 0 {
+				events = append(events, glm46EventThinkingContent{content: unambiguous})
+			}
+			return events, false
+		} else {
+			// Pure thinking content - withhold trailing whitespace (might precede closing tag)
+			whitespaceLen := trailingWhitespaceLen(acc)
+			ambiguousStart := len(acc) - whitespaceLen
+
+			unambiguous := acc[:ambiguousStart]
+			ambiguous := acc[ambiguousStart:]
+			p.buffer.Reset()
+			p.buffer.WriteString(ambiguous)
+			if len(unambiguous) > 0 {
+				events = append(events, glm46EventThinkingContent{content: unambiguous})
+			}
+			return events, false
+		}
+
+	case glm46ParserState_ThinkingDoneEatingWhitespace:
+		return p.eatLeadingWhitespaceAndTransitionTo(glm46ParserState_CollectingContent)
+
+	case glm46ParserState_CollectingContent:
+		if strings.Contains(p.buffer.String(), glm46ToolOpenTag) {
+			before, after := glm46SplitAtTag(p, glm46ToolOpenTag, true)
+			if len(before) > 0 {
+				events = append(events, glm46EventContent{content: before})
+			}
+			if after == "" {
+				p.state = glm46ParserState_ToolStartedEatingWhitespace
+			} else {
+				p.state = glm46ParserState_CollectingToolContent
+			}
+			return events, true
+		} else if overlapLen := overlap(p.buffer.String(), glm46ToolOpenTag); overlapLen > 0 {
+			beforePartialTag := p.buffer.String()[:len(p.buffer.String())-overlapLen]
+			trailingWhitespaceLen := trailingWhitespaceLen(beforePartialTag)
+			ambiguousStart := len(beforePartialTag) - trailingWhitespaceLen
+
+			unambiguous := p.buffer.String()[:ambiguousStart]
+			ambiguous := p.buffer.String()[ambiguousStart:]
+			p.buffer.Reset()
+			p.buffer.WriteString(ambiguous)
+			if len(unambiguous) > 0 {
+				events = append(events, glm46EventContent{content: unambiguous})
+			}
+			return events, false
+		} else {
+			whitespaceLen := trailingWhitespaceLen(p.buffer.String())
+			ambiguousStart := len(p.buffer.String()) - whitespaceLen
+
+			unambiguous := p.buffer.String()[:ambiguousStart]
+			ambiguous := p.buffer.String()[ambiguousStart:]
+			p.buffer.Reset()
+			p.buffer.WriteString(ambiguous)
+			if len(unambiguous) > 0 {
+				events = append(events, glm46EventContent{content: unambiguous})
+			}
+			return events, false
+		}
+
+	case glm46ParserState_ToolStartedEatingWhitespace:
+		return p.eatLeadingWhitespaceAndTransitionTo(glm46ParserState_CollectingToolContent)
+
+	case glm46ParserState_CollectingToolContent:
+		acc := p.buffer.String()
+		if strings.Contains(acc, glm46ToolCloseTag) {
+			toolContent, _ := glm46SplitAtTag(p, glm46ToolCloseTag, true)
+			if len(toolContent) == 0 {
+				slog.Warn("glm46 tool call closing tag found but no content before it")
+			}
+			events = append(events, glm46EventRawToolCall{raw: toolContent})
+			p.state = glm46ParserState_CollectingContent
+			return events, true
+		} else {
+			// Keep accumulating - tool calls are not streamed
+			// We just wait for the closing tag
+			return events, false
+		}
+
+	default:
+		panic("unreachable")
+	}
+}
+
+// GLMToolCallXML represents the structure of a GLM-4.6 tool call for XML parsing
+type GLMToolCallXML struct {
+	XMLName xml.Name `xml:"tool_call"`
+	Content string   `xml:",chardata"` // Function name (text nodes between tags)
+	Keys    []string `xml:"arg_key"`   // All arg_key elements in document order
+	Values  []string `xml:"arg_value"` // All arg_value elements in document order
+}
+
+// escapeGLM46Content escapes XML entities in text content while preserving arg_key/arg_value tags
+func escapeGLM46Content(s string) string {
+	var result strings.Builder
+	inTag := false
+
+	for i := range len(s) {
+		ch := s[i]
+
+		if ch == '<' {
+			// Check if this is a known tag
+			if strings.HasPrefix(s[i:], "<arg_key>") ||
+				strings.HasPrefix(s[i:], "</arg_key>") ||
+				strings.HasPrefix(s[i:], "<arg_value>") ||
+				strings.HasPrefix(s[i:], "</arg_value>") {
+				inTag = true
+			}
+		}
+
+		if inTag {
+			result.WriteByte(ch)
+			if ch == '>' {
+				inTag = false
+			}
+		} else {
+			// Escape special characters in text content
+			switch ch {
+			case '&':
+				result.WriteString("&amp;")
+			case '<':
+				result.WriteString("&lt;")
+			case '>':
+				result.WriteString("&gt;")
+			default:
+				result.WriteByte(ch)
+			}
+		}
+	}
+
+	return result.String()
+}
+
+func parseGLM46ToolCall(raw glm46EventRawToolCall, tools []api.Tool) (api.ToolCall, error) {
+	// Escape any unescaped entities in text content
+	// We need to escape text between tags, but not the tags themselves
+	escaped := escapeGLM46Content(raw.raw)
+
+	// Wrap the content in a root element to make it valid XML
+	xmlString := "<tool_call>" + escaped + "</tool_call>"
+
+	// Parse XML into struct
+	var parsed GLMToolCallXML
+	if err := xml.Unmarshal([]byte(xmlString), &parsed); err != nil {
+		return api.ToolCall{}, fmt.Errorf("failed to parse XML: %w", err)
+	}
+
+	// Extract and trim function name
+	functionName := strings.TrimSpace(parsed.Content)
+	if functionName == "" {
+		return api.ToolCall{}, fmt.Errorf("empty function name")
+	}
+
+	// Verify keys and values are paired correctly
+	if len(parsed.Keys) != len(parsed.Values) {
+		return api.ToolCall{}, fmt.Errorf("mismatched arg_key and arg_value counts: %d keys, %d values", len(parsed.Keys), len(parsed.Values))
+	}
+
+	// Find the matching tool to get parameter types
+	var matchedTool *api.Tool
+	for i := range tools {
+		if tools[i].Function.Name == functionName {
+			matchedTool = &tools[i]
+			break
+		}
+	}
+
+	// Build arguments map by pairing keys and values
+	toolCall := api.ToolCall{
+		Function: api.ToolCallFunction{
+			Name:      functionName,
+			Arguments: api.NewToolCallFunctionArguments(),
+		},
+	}
+
+	for i := range parsed.Keys {
+		key := strings.TrimSpace(parsed.Keys[i])
+		value := parsed.Values[i] // Don't trim here - parseValue handles it
+
+		// Look up parameter type
+		var paramType api.PropertyType
+		if matchedTool != nil && matchedTool.Function.Parameters.Properties != nil {
+			if prop, ok := matchedTool.Function.Parameters.Properties.Get(key); ok {
+				// Handle anyOf by collecting all types from the union
+				if len(prop.AnyOf) > 0 {
+					for _, anyOfProp := range prop.AnyOf {
+						paramType = append(paramType, anyOfProp.Type...)
+					}
+				} else {
+					paramType = prop.Type
+				}
+			}
+		}
+
+		// Parse value with type coercion
+		toolCall.Function.Arguments.Set(key, parseValue(value, paramType))
+	}
+
+	return toolCall, nil
+}

model/parsers/glm46_test.go

@@ -0,0 +1,862 @@
+package parsers
+
+import (
+	"encoding/xml"
+	"reflect"
+	"testing"
+
+	"github.com/ollama/ollama/api"
+)
+
+func TestGLM46ParserStreaming(t *testing.T) {
+	type step struct {
+		input      string
+		wantEvents []glm46Event
+	}
+
+	cases := []struct {
+		desc  string
+		steps []step
+		only  bool
+	}{
+		{
+			desc: "leading whitespace before think tag",
+			steps: []step{
+				{
+					input:      "   \n\t  ",
+					wantEvents: []glm46Event{},
+				},
+				{
+					input:      "<think>thinking</think>",
+					wantEvents: []glm46Event{glm46EventThinkingContent{content: "thinking"}},
+				},
+			},
+		},
+		{
+			desc: "think tag with whitespace inside",
+			steps: []step{
+				{
+					input: "<think>  \n  thinking content  \n  </think>regular content",
+					wantEvents: []glm46Event{
+						glm46EventThinkingContent{content: "thinking content"},
+						glm46EventContent{content: "regular content"},
+					},
+				},
+			},
+		},
+		{
+			desc: "tool call with leading whitespace after opening tag",
+			steps: []step{
+				{
+					input: "<think></think><tool_call>  \n  test  \n  </tool_call>",
+					wantEvents: []glm46Event{
+						glm46EventRawToolCall{raw: "test"},
+					},
+				},
+			},
+		},
+		{
+			desc: "simple thinking then content",
+			steps: []step{
+				{
+					input: "<think>I am thinking</think>Now I respond",
+					wantEvents: []glm46Event{
+						glm46EventThinkingContent{content: "I am thinking"},
+						glm46EventContent{content: "Now I respond"},
+					},
+				},
+			},
+		},
+		{
+			desc: "streamed thinking content",
+			steps: []step{
+				{
+					input:      "<think>hello",
+					wantEvents: []glm46Event{glm46EventThinkingContent{content: "hello"}},
+				},
+				{
+					input:      " world",
+					wantEvents: []glm46Event{glm46EventThinkingContent{content: " world"}},
+				},
+				{
+					input: "</think>content",
+					wantEvents: []glm46Event{
+						glm46EventContent{content: "content"},
+					},
+				},
+			},
+		},
+		{
+			desc: "content before tool call",
+			steps: []step{
+				{
+					input: "<think>Let me call a tool</think>here is text<tool_call>",
+					wantEvents: []glm46Event{
+						glm46EventThinkingContent{content: "Let me call a tool"},
+						glm46EventContent{content: "here is text"},
+					},
+				},
+				{
+					input: "function_name\n<arg_key>param</arg_key>\n<arg_value>value</arg_value>\n</tool_call>",
+					wantEvents: []glm46Event{
+						glm46EventRawToolCall{raw: "function_name\n<arg_key>param</arg_key>\n<arg_value>value</arg_value>"},
+					},
+				},
+			},
+		},
+		{
+			desc: "tool call with content after",
+			steps: []step{
+				{
+					input: "<think>thinking</think><tool_call>test</tool_call>after tool",
+					wantEvents: []glm46Event{
+						glm46EventThinkingContent{content: "thinking"},
+						glm46EventRawToolCall{raw: "test"},
+						glm46EventContent{content: "after tool"},
+					},
+				},
+			},
+		},
+		{
+			desc: "trailing whitespace between content and tool call is trimmed",
+			steps: []step{
+				{
+					input: "<think>thinking</think>content\n  \t  <tool_call>test</tool_call>",
+					wantEvents: []glm46Event{
+						glm46EventThinkingContent{content: "thinking"},
+						glm46EventContent{content: "content"},
+						glm46EventRawToolCall{raw: "test"},
+					},
+				},
+			},
+		},
+		{
+			desc: "trailing whitespace between tool call and content is trimmed",
+			steps: []step{
+				{
+					input: "<think>think</think><tool_call>test</tool_call>\n\t  after",
+					wantEvents: []glm46Event{
+						glm46EventThinkingContent{content: "think"},
+						glm46EventRawToolCall{raw: "test"},
+						glm46EventContent{content: "after"},
+					},
+				},
+			},
+		},
+		{
+			desc: "split thinking close tag",
+			steps: []step{
+				{
+					input:      "<think>thinking content</th",
+					wantEvents: []glm46Event{glm46EventThinkingContent{content: "thinking content"}},
+				},
+				{
+					input: "ink>after",
+					wantEvents: []glm46Event{
+						glm46EventContent{content: "after"},
+					},
+				},
+			},
+		},
+		{
+			desc: "split thinking open tag",
+			steps: []step{
+				{
+					input:      "  <thi",
+					wantEvents: []glm46Event{},
+				},
+				{
+					input:      "nk>content</think>",
+					wantEvents: []glm46Event{glm46EventThinkingContent{content: "content"}},
+				},
+			},
+		},
+		{
+			desc: "split tool open tag",
+			steps: []step{
+				{
+					input:      "<think>think</think>content<tool",
+					wantEvents: []glm46Event{glm46EventThinkingContent{content: "think"}, glm46EventContent{content: "content"}},
+				},
+				{
+					input:      "_call>inside",
+					wantEvents: []glm46Event{},
+				},
+				{
+					input: "</tool_call>",
+					wantEvents: []glm46Event{
+						glm46EventRawToolCall{raw: "inside"},
+					},
+				},
+			},
+		},
+		{
+			desc: "partial thinking close tag fakeout",
+			steps: []step{
+				{
+					input:      "<think>content</th",
+					wantEvents: []glm46Event{glm46EventThinkingContent{content: "content"}},
+				},
+				{
+					input:      "ought more",
+					wantEvents: []glm46Event{glm46EventThinkingContent{content: "</thought more"}},
+				},
+			},
+		},
+		{
+			desc: "partial thinking open tag fakeout",
+			steps: []step{
+				{
+					input:      "  <thi",
+					wantEvents: []glm46Event{},
+				},
+				{
+					input: "nking is fun",
+					wantEvents: []glm46Event{
+						glm46EventContent{content: "  <thinking is fun"},
+					},
+				},
+			},
+		},
+		{
+			desc: "partial tool open tag fakeout",
+			steps: []step{
+				{
+					input: "<think></think>content\n<tool",
+					wantEvents: []glm46Event{
+						glm46EventContent{content: "content"},
+					},
+				},
+				{
+					input: " fakeout",
+					wantEvents: []glm46Event{
+						glm46EventContent{content: "\n<tool fakeout"},
+					},
+				},
+			},
+		},
+		{
+			desc: "partial tool close tag fakeout",
+			steps: []step{
+				{
+					input:      "<think></think><tool_call>content</tool",
+					wantEvents: []glm46Event{},
+				},
+				{
+					input:      " fakeout",
+					wantEvents: []glm46Event{},
+				},
+				{
+					input: "</tool_call>",
+					wantEvents: []glm46Event{
+						glm46EventRawToolCall{raw: "content</tool fakeout"},
+					},
+				},
+			},
+		},
+		{
+			desc: "empty thinking tag",
+			steps: []step{
+				{
+					input: "<think></think>content here",
+					wantEvents: []glm46Event{
+						glm46EventContent{content: "content here"},
+					},
+				},
+			},
+		},
+		{
+			desc: "multiple tool calls in sequence",
+			steps: []step{
+				{
+					input: "<think>think</think><tool_call>first</tool_call>between<tool_call>second</tool_call>end",
+					wantEvents: []glm46Event{
+						glm46EventThinkingContent{content: "think"},
+						glm46EventRawToolCall{raw: "first"},
+						glm46EventContent{content: "between"},
+						glm46EventRawToolCall{raw: "second"},
+						glm46EventContent{content: "end"},
+					},
+				},
+			},
+		},
+		{
+			desc: "no thinking tag - direct to content",
+			steps: []step{
+				{
+					input: "just content here",
+					wantEvents: []glm46Event{
+						glm46EventContent{content: "just content here"},
+					},
+				},
+			},
+		},
+		{
+			desc: "no thinking tag - skip to content then tool call",
+			steps: []step{
+				{
+					input: "Here's the answer:<tool_call>test</tool_call>done",
+					wantEvents: []glm46Event{
+						glm46EventContent{content: "Here's the answer:"},
+						glm46EventRawToolCall{raw: "test"},
+						glm46EventContent{content: "done"},
+					},
+				},
+			},
+		},
+		{
+			desc: "no thinking tag - whitespace preserved when no tags",
+			steps: []step{
+				{
+					input: "  \n  content with leading whitespace",
+					wantEvents: []glm46Event{
+						glm46EventContent{content: "  \n  content with leading whitespace"},
+					},
+				},
+			},
+		},
+		{
+			desc: "whitespace after think close tag gets eaten",
+			steps: []step{
+				{
+					input: "<think>thinking</think>  \n\t  content",
+					wantEvents: []glm46Event{
+						glm46EventThinkingContent{content: "thinking"},
+						glm46EventContent{content: "content"},
+					},
+				},
+			},
+		},
+		{
+			desc: "whitespace after tool_call close tag gets eaten",
+			steps: []step{
+				{
+					input: "<think></think><tool_call>test</tool_call>  \n\t  content",
+					wantEvents: []glm46Event{
+						glm46EventRawToolCall{raw: "test"},
+						glm46EventContent{content: "content"},
+					},
+				},
+			},
+		},
+		{
+			desc: "thinking content withholds trailing whitespace (single chunk)",
+			steps: []step{
+				{
+					input: "<think>thinking content   ",
+					wantEvents: []glm46Event{
+						glm46EventThinkingContent{content: "thinking content"},
+					},
+				},
+				{
+					input: "</think>after",
+					wantEvents: []glm46Event{
+						glm46EventContent{content: "after"},
+					},
+				},
+			},
+		},
+		{
+			desc: "thinking content withholds trailing whitespace with newlines",
+			steps: []step{
+				{
+					input: "<think>thinking\n\n  ",
+					wantEvents: []glm46Event{
+						glm46EventThinkingContent{content: "thinking"},
+					},
+				},
+				{
+					input: "</think>content",
+					wantEvents: []glm46Event{
+						glm46EventContent{content: "content"},
+					},
+				},
+			},
+		},
+		{
+			desc: "thinking content trailing whitespace emitted when more content arrives",
+			steps: []step{
+				{
+					input: "<think>thinking   ",
+					wantEvents: []glm46Event{
+						glm46EventThinkingContent{content: "thinking"},
+					},
+				},
+				{
+					input: "more thinking",
+					wantEvents: []glm46Event{
+						glm46EventThinkingContent{content: "   more thinking"},
+					},
+				},
+				{
+					input:      "</think>",
+					wantEvents: []glm46Event{},
+				},
+			},
+		},
+		{
+			desc: "thinking content withholds trailing whitespace before partial close tag",
+			steps: []step{
+				{
+					input: "<think>thinking   </th",
+					wantEvents: []glm46Event{
+						glm46EventThinkingContent{content: "thinking"},
+					},
+				},
+				{
+					input: "ink>content",
+					wantEvents: []glm46Event{
+						glm46EventContent{content: "content"},
+					},
+				},
+			},
+		},
+	}
+
+	anyOnlies := false
+	for _, tc := range cases {
+		if tc.only {
+			anyOnlies = true
+		}
+	}
+
+	for _, tc := range cases {
+		if anyOnlies && !tc.only {
+			continue
+		}
+
+		t.Run(tc.desc, func(t *testing.T) {
+			parser := GLM46Parser{}
+
+			for i, step := range tc.steps {
+				parser.buffer.WriteString(step.input)
+				gotEvents := parser.parseEvents()
+
+				if len(gotEvents) == 0 && len(step.wantEvents) == 0 {
+					// avoid deep equal on empty vs. nil slices
+					continue
+				}
+
+				if !reflect.DeepEqual(gotEvents, step.wantEvents) {
+					t.Errorf("step %d: input %q: got events %#v, want %#v", i, step.input, gotEvents, step.wantEvents)
+				}
+			}
+		})
+	}
+}
+
+// TestGLMToolCallXMLOrderPreservation verifies that xml.Unmarshal preserves
+// document order when collecting multiple elements with the same tag name into slices.
+// This is a critical assumption for the GLM-4.6 parser's struct-based approach.
+func TestGLMToolCallXMLOrderPreservation(t *testing.T) {
+	testCases := []struct {
+		name       string
+		xml        string
+		wantKeys   []string
+		wantValues []string
+	}{
+		{
+			name: "alternating keys and values",
+			xml: `<tool_call>
+function_name
+<arg_key>first</arg_key>
+<arg_value>A</arg_value>
+<arg_key>second</arg_key>
+<arg_value>B</arg_value>
+<arg_key>third</arg_key>
+<arg_value>C</arg_value>
+</tool_call>`,
+			wantKeys:   []string{"first", "second", "third"},
+			wantValues: []string{"A", "B", "C"},
+		},
+		{
+			name: "all keys then all values",
+			xml: `<tool_call>
+function_name
+<arg_key>key1</arg_key>
+<arg_key>key2</arg_key>
+<arg_key>key3</arg_key>
+<arg_value>val1</arg_value>
+<arg_value>val2</arg_value>
+<arg_value>val3</arg_value>
+</tool_call>`,
+			wantKeys:   []string{"key1", "key2", "key3"},
+			wantValues: []string{"val1", "val2", "val3"},
+		},
+		{
+			name: "mixed grouping",
+			xml: `<tool_call>
+function_name
+<arg_key>a</arg_key>
+<arg_value>1</arg_value>
+<arg_key>b</arg_key>
+<arg_key>c</arg_key>
+<arg_value>2</arg_value>
+<arg_value>3</arg_value>
+</tool_call>`,
+			wantKeys:   []string{"a", "b", "c"},
+			wantValues: []string{"1", "2", "3"},
+		},
+		{
+			name: "reverse order - all values then all keys",
+			xml: `<tool_call>
+function_name
+<arg_value>X</arg_value>
+<arg_value>Y</arg_value>
+<arg_value>Z</arg_value>
+<arg_key>x</arg_key>
+<arg_key>y</arg_key>
+<arg_key>z</arg_key>
+</tool_call>`,
+			wantKeys:   []string{"x", "y", "z"},
+			wantValues: []string{"X", "Y", "Z"},
+		},
+	}
+
+	for _, tc := range testCases {
+		t.Run(tc.name, func(t *testing.T) {
+			var parsed GLMToolCallXML
+			err := xml.Unmarshal([]byte(tc.xml), &parsed)
+			if err != nil {
+				t.Fatalf("failed to unmarshal XML: %v", err)
+			}
+
+			if !reflect.DeepEqual(parsed.Keys, tc.wantKeys) {
+				t.Errorf("Keys order mismatch:\ngot:  %v\nwant: %v", parsed.Keys, tc.wantKeys)
+			}
+
+			if !reflect.DeepEqual(parsed.Values, tc.wantValues) {
+				t.Errorf("Values order mismatch:\ngot:  %v\nwant: %v", parsed.Values, tc.wantValues)
+			}
+		})
+	}
+}
+
+func TestGLM46ToolCallParsing(t *testing.T) {
+	type testCase struct {
+		name         string
+		rawToolCall  string
+		tools        []api.Tool
+		wantToolCall api.ToolCall
+	}
+
+	cases := []testCase{
+		{
+			name:  "simple tool call",
+			tools: []api.Tool{},
+			rawToolCall: `get-current-weather
+<arg_key>location</arg_key>
+<arg_value>New York, NY</arg_value>
+<arg_key>unit</arg_key>
+<arg_value>celsius</arg_value>`,
+			wantToolCall: api.ToolCall{
+				Function: api.ToolCallFunction{
+					Name:      "get-current-weather",
+					Arguments: args(`{"location": "New York, NY", "unit": "celsius"}`),
+				},
+			},
+		},
+		{
+			name: "tool call with typed parameters",
+			tools: []api.Tool{
+				tool("calculate", map[string]api.ToolProperty{
+					"x":       {Type: api.PropertyType{"number"}},
+					"y":       {Type: api.PropertyType{"integer"}},
+					"enabled": {Type: api.PropertyType{"boolean"}},
+					"items":   {Type: api.PropertyType{"array"}},
+				}),
+			},
+			rawToolCall: `calculate
+<arg_key>x</arg_key>
+<arg_value>3.14</arg_value>
+<arg_key>y</arg_key>
+<arg_value>42</arg_value>
+<arg_key>enabled</arg_key>
+<arg_value>true</arg_value>
+<arg_key>items</arg_key>
+<arg_value>["a", "b", "c"]</arg_value>`,
+			wantToolCall: api.ToolCall{
+				Function: api.ToolCallFunction{
+					Name:      "calculate",
+					Arguments: args(`{"enabled": true, "items": ["a", "b", "c"], "x": 3.14, "y": 42}`),
+				},
+			},
+		},
+		{
+			name:  "function name with whitespace",
+			tools: []api.Tool{},
+			rawToolCall: `  get-weather
+<arg_key>city</arg_key>
+<arg_value>Paris</arg_value>`,
+			wantToolCall: api.ToolCall{
+				Function: api.ToolCallFunction{
+					Name:      "get-weather",
+					Arguments: args(`{"city": "Paris"}`),
+				},
+			},
+		},
+		{
+			name:  "values with special characters",
+			tools: []api.Tool{},
+			rawToolCall: `execute-command
+<arg_key>command</arg_key>
+<arg_value>ls && echo "done"</arg_value>
+<arg_key>message</arg_key>
+<arg_value>a < b and c > d</arg_value>`,
+			wantToolCall: api.ToolCall{
+				Function: api.ToolCallFunction{
+					Name:      "execute-command",
+					Arguments: args(`{"command": "ls && echo \"done\"", "message": "a < b and c > d"}`),
+				},
+			},
+		},
+		{
+			name:  "unicode in function names and values",
+			tools: []api.Tool{},
+			rawToolCall: `获取天气
+<arg_key>城市</arg_key>
+<arg_value>北京</arg_value>
+<arg_key>message</arg_key>
+<arg_value>Hello! 你好! 🌟</arg_value>`,
+			wantToolCall: api.ToolCall{
+				Function: api.ToolCallFunction{
+					Name:      "获取天气",
+					Arguments: args(`{"message": "Hello! 你好! 🌟", "城市": "北京"}`),
+				},
+			},
+		},
+		{
+			name:  "empty value",
+			tools: []api.Tool{},
+			rawToolCall: `test-function
+<arg_key>param1</arg_key>
+<arg_value></arg_value>`,
+			wantToolCall: api.ToolCall{
+				Function: api.ToolCallFunction{
+					Name:      "test-function",
+					Arguments: args(`{"param1": ""}`),
+				},
+			},
+		},
+		{
+			name:  "special chars in arg_key names",
+			tools: []api.Tool{},
+			rawToolCall: `test-function
+<arg_key>param<1></arg_key>
+<arg_value>value1</arg_value>
+<arg_key>a&b</arg_key>
+<arg_value>value2</arg_value>
+<arg_key>x>y</arg_key>
+<arg_value>value3</arg_value>`,
+			wantToolCall: api.ToolCall{
+				Function: api.ToolCallFunction{
+					Name:      "test-function",
+					Arguments: args(`{"a&b": "value2", "param<1>": "value1", "x>y": "value3"}`),
+				},
+			},
+		},
+		{
+			name:  "multiple consecutive ampersands",
+			tools: []api.Tool{},
+			rawToolCall: `test-function
+<arg_key>param</arg_key>
+<arg_value>test &&&& more</arg_value>`,
+			wantToolCall: api.ToolCall{
+				Function: api.ToolCallFunction{
+					Name:      "test-function",
+					Arguments: args(`{"param": "test &&&& more"}`),
+				},
+			},
+		},
+		{
+			name:  "mixed special chars together",
+			tools: []api.Tool{},
+			rawToolCall: `test-function
+<arg_key>param</arg_key>
+<arg_value><>&<>&</arg_value>`,
+			wantToolCall: api.ToolCall{
+				Function: api.ToolCallFunction{
+					Name:      "test-function",
+					Arguments: args(`{"param": "<>&<>&"}`),
+				},
+			},
+		},
+		{
+			name:  "newlines and tabs in parameter values",
+			tools: []api.Tool{},
+			rawToolCall: `test-function
+<arg_key>multiline</arg_key>
+<arg_value>line1
+	indented line2
+line3</arg_value>`,
+			wantToolCall: api.ToolCall{
+				Function: api.ToolCallFunction{
+					Name:      "test-function",
+					Arguments: args(`{"multiline": "line1\n\tindented line2\nline3"}`),
+				},
+			},
+		},
+		{
+			name:  "single and double quotes in values",
+			tools: []api.Tool{},
+			rawToolCall: `test-function
+<arg_key>quotes</arg_key>
+<arg_value>She said "Hello's there!"</arg_value>`,
+			wantToolCall: api.ToolCall{
+				Function: api.ToolCallFunction{
+					Name:      "test-function",
+					Arguments: args(`{"quotes": "She said \"Hello's there!\""}`),
+				},
+			},
+		},
+		{
+			name:  "CDATA-like content that should be treated as text",
+			tools: []api.Tool{},
+			rawToolCall: `test-function
+<arg_key>cdata</arg_key>
+<arg_value><![CDATA[not actual cdata]]></arg_value>`,
+			wantToolCall: api.ToolCall{
+				Function: api.ToolCallFunction{
+					Name:      "test-function",
+					Arguments: args(`{"cdata": "<![CDATA[not actual cdata]]>"}`),
+				},
+			},
+		},
+		{
+			name:  "all special XML entities",
+			tools: []api.Tool{},
+			rawToolCall: `test-function
+<arg_key>entities</arg_key>
+<arg_value>&lt;&gt;&amp;&apos;&quot;</arg_value>`,
+			wantToolCall: api.ToolCall{
+				Function: api.ToolCallFunction{
+					Name:      "test-function",
+					Arguments: args(`{"entities": "&lt;&gt;&amp;&apos;&quot;"}`),
+				},
+			},
+		},
+		{
+			name:  "order preservation with multiple parameters",
+			tools: []api.Tool{},
+			rawToolCall: `test-function
+<arg_key>first</arg_key>
+<arg_value>value1</arg_value>
+<arg_key>second</arg_key>
+<arg_value>value2</arg_value>
+<arg_key>third</arg_key>
+<arg_value>value3</arg_value>
+<arg_key>fourth</arg_key>
+<arg_value>value4</arg_value>
+<arg_key>fifth</arg_key>
+<arg_value>value5</arg_value>`,
+			wantToolCall: api.ToolCall{
+				Function: api.ToolCallFunction{
+					Name:      "test-function",
+					Arguments: args(`{"fifth": "value5", "first": "value1", "fourth": "value4", "second": "value2", "third": "value3"}`),
+				},
+			},
+		},
+		{
+			name:  "order preservation with identical key names but different positions",
+			tools: []api.Tool{},
+			rawToolCall: `test-function
+<arg_key>param</arg_key>
+<arg_value>first occurrence</arg_value>
+<arg_key>other</arg_key>
+<arg_value>middle</arg_value>
+<arg_key>param</arg_key>
+<arg_value>second occurrence</arg_value>`,
+			wantToolCall: api.ToolCall{
+				Function: api.ToolCallFunction{
+					Name: "test-function",
+					// Later occurrence should overwrite earlier one
+					Arguments: args(`{"other": "middle", "param": "second occurrence"}`),
+				},
+			},
+		},
+		{
+			name: "array with mixed types",
+			tools: []api.Tool{
+				tool("process", map[string]api.ToolProperty{
+					"items": {Type: api.PropertyType{"array"}},
+				}),
+			},
+			rawToolCall: `process
+<arg_key>items</arg_key>
+<arg_value>[1, "hello", true, null]</arg_value>`,
+			wantToolCall: api.ToolCall{
+				Function: api.ToolCallFunction{
+					Name:      "process",
+					Arguments: args(`{"items": [1, "hello", true, null]}`),
+				},
+			},
+		},
+		{
+			name: "empty array",
+			tools: []api.Tool{
+				tool("test", map[string]api.ToolProperty{
+					"tags": {Type: api.PropertyType{"array"}},
+				}),
+			},
+			rawToolCall: `test
+<arg_key>tags</arg_key>
+<arg_value>[]</arg_value>`,
+			wantToolCall: api.ToolCall{
+				Function: api.ToolCallFunction{
+					Name:      "test",
+					Arguments: args(`{"tags": []}`),
+				},
+			},
+		},
+		{
+			name: "anyOf array or string - with array of objects",
+			tools: []api.Tool{
+				tool("TodoWrite", map[string]api.ToolProperty{
+					"todos": {AnyOf: []api.ToolProperty{{Type: api.PropertyType{"array"}}, {Type: api.PropertyType{"string"}}}},
+				}),
+			},
+			// <tool_call>TodoWrite
+			// <arg_key>todos</arg_key>
+			// <arg_value>[{"content": "Set up HTML file and basic structure", "id": "1", "priority": "high", "status": "pending"}, {"content": "Create 3D scene with Three.js", "id": "2", "priority": "high", "status": "pending"}, {"content": "Implement terrain generation with blocks", "id": "3", "priority": "high", "status": "pending"}, {"content": "Add player controls (movement, camera)", "id": "4", "priority": "high", "status": "pending"}, {"content": "Implement block placement/destruction", "id": "5", "priority": "medium", "status": "pending"}, {"content": "Add lighting and textures", "id": "6", "priority": "medium", "status": "pending"}, {"content": "Test and optimize performance", "id": "7", "priority": "low", "status": "pending"}]</arg_value>
+			// </tool_call>
+			rawToolCall: `TodoWrite
+<arg_key>todos</arg_key>
+<arg_value>[{"content": "task 1", "status": "pending", "priority": "high", "id": "1"}, {"content": "task 2", "status": "completed", "priority": "low", "id": "2"}]</arg_value>`,
+			wantToolCall: api.ToolCall{
+				Function: api.ToolCallFunction{
+					Name:      "TodoWrite",
+					Arguments: args(`{"todos": [{"content": "task 1", "id": "1", "priority": "high", "status": "pending"}, {"content": "task 2", "id": "2", "priority": "low", "status": "completed"}]}`),
+				},
+			},
+		},
+		{
+			name: "anyOf array or string - with plain string",
+			tools: []api.Tool{
+				tool("TodoWrite", map[string]api.ToolProperty{
+					"todos": {Type: api.PropertyType{"array", "string"}},
+				}),
+			},
+			rawToolCall: `TodoWrite
+<arg_key>todos</arg_key>
+<arg_value>Error: could not load todos</arg_value>`,
+			wantToolCall: api.ToolCall{
+				Function: api.ToolCallFunction{
+					Name:      "TodoWrite",
+					Arguments: args(`{"todos": "Error: could not load todos"}`),
+				},
+			},
+		},
+	}
+
+	for i, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			gotToolCall, err := parseGLM46ToolCall(glm46EventRawToolCall{raw: tc.rawToolCall}, tc.tools)
+			if err != nil {
+				t.Errorf("case %d (%s): %v", i, tc.name, err)
+			}
+			if !toolCallEqual(gotToolCall, tc.wantToolCall) {
+				t.Errorf("case %d (%s): got tool call %#v, want %#v", i, tc.name, gotToolCall, tc.wantToolCall)
+			}
+		})
+	}
+}

model/parsers/glm47.go

@@ -0,0 +1,20 @@
+package parsers
+
+import "github.com/ollama/ollama/api"
+
+// GLM47Parser extends GLM46Parser with thinking-aware initialization.
+// GLM-4.7's prompt ends with <think> when thinking is enabled, so the parser
+// must start in CollectingThinking state (the model outputs thinking content directly).
+type GLM47Parser struct {
+	GLM46Parser
+}
+
+func (p *GLM47Parser) Init(tools []api.Tool, lastMessage *api.Message, thinkValue *api.ThinkValue) []api.Tool {
+	p.tools = tools
+	// When thinking is enabled (nil or true), the prompt ends with <think>,
+	// so model output starts directly with thinking content (no opening tag).
+	if thinkValue == nil || thinkValue.Bool() {
+		p.state = glm46ParserState_CollectingThinking
+	}
+	return tools
+}

model/parsers/glm47_test.go

@@ -0,0 +1,99 @@
+package parsers
+
+import (
+	"reflect"
+	"testing"
+
+	"github.com/ollama/ollama/api"
+)
+
+func TestGLM47ParserAdd(t *testing.T) {
+	parser := GLM47Parser{}
+	parser.Init([]api.Tool{
+		tool("calculate", map[string]api.ToolProperty{
+			"count":   {Type: api.PropertyType{"integer"}},
+			"enabled": {Type: api.PropertyType{"boolean"}},
+		}),
+	}, nil, nil)
+
+	// When thinking is enabled (thinkValue nil), the prompt ends with <think>,
+	// so the model output does NOT include the opening <think> tag.
+	content, thinking, calls, err := parser.Add("plan</think>Answer<tool_call>calculate<arg_key>count</arg_key><arg_value>3</arg_value><arg_key>enabled</arg_key><arg_value>true</arg_value></tool_call>", true)
+	if err != nil {
+		t.Fatalf("parse failed: %v", err)
+	}
+	if thinking != "plan" {
+		t.Fatalf("expected thinking 'plan', got %q", thinking)
+	}
+	if content != "Answer" {
+		t.Fatalf("expected content 'Answer', got %q", content)
+	}
+	if len(calls) != 1 {
+		t.Fatalf("expected 1 tool call, got %d", len(calls))
+	}
+	expectedArgs := args(`{"count": 3, "enabled": true}`)
+	if !toolCallEqual(api.ToolCall{Function: api.ToolCallFunction{Arguments: calls[0].Function.Arguments}}, api.ToolCall{Function: api.ToolCallFunction{Arguments: expectedArgs}}) {
+		t.Fatalf("expected args %#v, got %#v", expectedArgs.ToMap(), calls[0].Function.Arguments.ToMap())
+	}
+}
+
+func TestGLM47ParserNoThinkingContent(t *testing.T) {
+	parser := GLM47Parser{}
+	parser.Init(nil, nil, nil)
+
+	// When thinking is enabled but model has no thinking to output,
+	// it should output </think> immediately followed by content.
+	content, thinking, calls, err := parser.Add("</think>Plain answer", true)
+	if err != nil {
+		t.Fatalf("parse failed: %v", err)
+	}
+	if thinking != "" {
+		t.Fatalf("expected empty thinking, got %q", thinking)
+	}
+	if content != "Plain answer" {
+		t.Fatalf("expected content 'Plain answer', got %q", content)
+	}
+	if len(calls) != 0 {
+		t.Fatalf("expected no tool calls, got %d", len(calls))
+	}
+}
+
+func TestGLM47ParserThinkingDisabled(t *testing.T) {
+	parser := GLM47Parser{}
+	// When thinking is disabled, parser stays in LookingForThinkingOpen state
+	parser.Init(nil, nil, &api.ThinkValue{Value: false})
+
+	// Model outputs plain content (prompt ended with </think>)
+	content, thinking, calls, err := parser.Add("Plain answer", true)
+	if err != nil {
+		t.Fatalf("parse failed: %v", err)
+	}
+	if thinking != "" {
+		t.Fatalf("expected empty thinking, got %q", thinking)
+	}
+	if content != "Plain answer" {
+		t.Fatalf("expected content 'Plain answer', got %q", content)
+	}
+	if len(calls) != 0 {
+		t.Fatalf("expected no tool calls, got %d", len(calls))
+	}
+}
+
+func TestGLM47ParserToolCallEscaping(t *testing.T) {
+	toolCall, err := parseGLM46ToolCall(glm46EventRawToolCall{raw: `exec
+<arg_key>expr</arg_key>
+<arg_value>a < b && c > d</arg_value>`}, nil)
+	if err != nil {
+		t.Fatalf("parse failed: %v", err)
+	}
+
+	expected := api.ToolCall{
+		Function: api.ToolCallFunction{
+			Name:      "exec",
+			Arguments: args(`{"expr": "a < b && c > d"}`),
+		},
+	}
+	if !reflect.DeepEqual(toolCall, expected) {
+		t.Fatalf("expected %#v, got %#v", expected, toolCall)
+	}
+}

model/parsers/nemotron3nano.go

@@ -1,7 +1,6 @@
 package parsers
 
 import (
-	"regexp"
 	"strings"
 	"unicode"
 
@@ -14,243 +13,114 @@ const (
 	Nemotron3NanoCollectingThinking Nemotron3NanoParserState = iota
 	Nemotron3NanoSkipWhitespaceAfterThinking
 	Nemotron3NanoCollectingContent
-	Nemotron3NanoCollectingToolCalls
 )
 
 const (
-	nemotronThinkClose    = "</think>"
-	nemotronToolCallOpen  = "<tool_call>"
-	nemotronToolCallClose = "</tool_call>"
+	nemotronThinkClose   = "</think>"
+	nemotronToolCallOpen = "<tool_call>"
 )
 
 type Nemotron3NanoParser struct {
-	state  Nemotron3NanoParserState
-	buffer strings.Builder
-	tools  []api.Tool
+	state      Nemotron3NanoParserState
+	buffer     strings.Builder
+	toolParser *Qwen3CoderParser
 }
 
 func (p *Nemotron3NanoParser) HasToolSupport() bool     { return true }
 func (p *Nemotron3NanoParser) HasThinkingSupport() bool { return true }
 
 func (p *Nemotron3NanoParser) Init(tools []api.Tool, lastMessage *api.Message, thinkValue *api.ThinkValue) []api.Tool {
-	p.tools = tools
+	p.toolParser = &Qwen3CoderParser{}
+	p.toolParser.Init(tools, nil, nil)
 
-	// thinking is enabled if user requests it
 	thinkingEnabled := thinkValue != nil && thinkValue.Bool()
-
 	prefill := lastMessage != nil && lastMessage.Role == "assistant"
 
-	if !thinkingEnabled {
+	if !thinkingEnabled || (prefill && lastMessage.Content != "") {
 		p.state = Nemotron3NanoCollectingContent
-		return tools
+	} else {
+		p.state = Nemotron3NanoCollectingThinking
 	}
 
-	if prefill && lastMessage.Content != "" {
-		p.state = Nemotron3NanoCollectingContent
-		return tools
-	}
-
-	p.state = Nemotron3NanoCollectingThinking
 	return tools
 }
 
-type nemotronEvent interface {
-	isNemotronEvent()
-}
-
-type nemotronEventThinkingContent struct {
-	content string
-}
-
-type nemotronEventContent struct {
-	content string
-}
-
-type nemotronEventToolCall struct {
-	toolCall api.ToolCall
-}
-
-func (nemotronEventThinkingContent) isNemotronEvent() {}
-func (nemotronEventContent) isNemotronEvent()         {}
-func (nemotronEventToolCall) isNemotronEvent()        {}
-
 func (p *Nemotron3NanoParser) Add(s string, done bool) (content string, thinking string, calls []api.ToolCall, err error) {
-	p.buffer.WriteString(s)
-	events := p.parseEvents()
-
-	var toolCalls []api.ToolCall
-	var contentSb strings.Builder
-	var thinkingSb strings.Builder
-	for _, event := range events {
-		switch event := event.(type) {
-		case nemotronEventToolCall:
-			toolCalls = append(toolCalls, event.toolCall)
-		case nemotronEventThinkingContent:
-			thinkingSb.WriteString(event.content)
-		case nemotronEventContent:
-			contentSb.WriteString(event.content)
-		}
+	if p.state == Nemotron3NanoCollectingContent {
+		return p.toolParser.Add(s, done)
 	}
 
-	return contentSb.String(), thinkingSb.String(), toolCalls, nil
-}
-
-func (p *Nemotron3NanoParser) parseEvents() []nemotronEvent {
-	var all []nemotronEvent
-
-	keepLooping := true
-	for keepLooping {
-		var events []nemotronEvent
-		events, keepLooping = p.eat()
-		if len(events) > 0 {
-			all = append(all, events...)
-		}
-	}
-
-	return all
-}
-
-// emitWithPartialCheck extracts unambiguous content before a potential partial tag
-func (p *Nemotron3NanoParser) emitWithPartialCheck(bufStr, tag string) (unambiguous, ambiguous string) {
-	if overlapLen := overlap(bufStr, tag); overlapLen > 0 {
-		beforePartialTag := bufStr[:len(bufStr)-overlapLen]
-		trailingLen := trailingWhitespaceLen(beforePartialTag)
-		return bufStr[:len(beforePartialTag)-trailingLen], bufStr[len(beforePartialTag)-trailingLen:]
-	}
-	wsLen := trailingWhitespaceLen(bufStr)
-	return bufStr[:len(bufStr)-wsLen], bufStr[len(bufStr)-wsLen:]
-}
-
-func (p *Nemotron3NanoParser) eat() ([]nemotronEvent, bool) {
-	bufStr := p.buffer.String()
-	if bufStr == "" {
-		return nil, false
-	}
-
-	switch p.state {
-	case Nemotron3NanoCollectingThinking:
-		if strings.Contains(bufStr, nemotronThinkClose) {
-			split := strings.SplitN(bufStr, nemotronThinkClose, 2)
-			thinking := strings.TrimRightFunc(split[0], unicode.IsSpace)
-			p.buffer.Reset()
-			remainder := strings.TrimLeftFunc(split[1], unicode.IsSpace)
-			p.buffer.WriteString(remainder)
-			// Transition to whitespace-skipping state if buffer is empty,
-			// otherwise go directly to content collection
-			if remainder == "" {
-				p.state = Nemotron3NanoSkipWhitespaceAfterThinking
-			} else {
-				p.state = Nemotron3NanoCollectingContent
-			}
-			if thinking != "" {
-				return []nemotronEvent{nemotronEventThinkingContent{content: thinking}}, true
-			}
-			return nil, true
-		}
-		unambig, ambig := p.emitWithPartialCheck(bufStr, nemotronThinkClose)
-		p.buffer.Reset()
-		p.buffer.WriteString(ambig)
-		if unambig != "" {
-			return []nemotronEvent{nemotronEventThinkingContent{content: unambig}}, false
-		}
-		return nil, false
-
-	// We only want to skip whitespace between thinking and content
-	case Nemotron3NanoSkipWhitespaceAfterThinking:
-		bufStr = strings.TrimLeftFunc(bufStr, unicode.IsSpace)
-		p.buffer.Reset()
-		p.buffer.WriteString(bufStr)
-		if bufStr == "" {
-			return nil, false
+	if p.state == Nemotron3NanoSkipWhitespaceAfterThinking {
+		s = strings.TrimLeftFunc(s, unicode.IsSpace)
+		if s == "" {
+			return "", "", nil, nil
 		}
 		p.state = Nemotron3NanoCollectingContent
-		return nil, true
+		return p.toolParser.Add(s, done)
+	}
 
-	case Nemotron3NanoCollectingContent:
-		if strings.Contains(bufStr, nemotronToolCallOpen) {
-			split := strings.SplitN(bufStr, nemotronToolCallOpen, 2)
-			content := strings.TrimRightFunc(split[0], unicode.IsSpace)
-			p.buffer.Reset()
-			p.buffer.WriteString(split[1])
-			p.state = Nemotron3NanoCollectingToolCalls
-			if content != "" {
-				return []nemotronEvent{nemotronEventContent{content: content}}, true
-			}
-			return nil, true
-		}
-		unambig, ambig := p.emitWithPartialCheck(bufStr, nemotronToolCallOpen)
+	// Nemotron3NanoCollectingThinking - buffer and look for end markers
+	p.buffer.WriteString(s)
+	bufStr := p.buffer.String()
+
+	// Look for end of thinking: </think> or <tool_call> (model may skip </think>)
+	thinkIdx := strings.Index(bufStr, nemotronThinkClose)
+	toolIdx := strings.Index(bufStr, nemotronToolCallOpen)
+
+	var endIdx int = -1
+	var remainder string
+
+	if thinkIdx != -1 && (toolIdx == -1 || thinkIdx < toolIdx) {
+		endIdx = thinkIdx
+		remainder = strings.TrimLeftFunc(bufStr[thinkIdx+len(nemotronThinkClose):], unicode.IsSpace)
+	} else if toolIdx != -1 {
+		endIdx = toolIdx
+		remainder = bufStr[toolIdx:] // Include <tool_call> tag
+	}
+
+	if endIdx != -1 {
+		thinking = strings.TrimRightFunc(bufStr[:endIdx], unicode.IsSpace)
 		p.buffer.Reset()
-		p.buffer.WriteString(ambig)
-		if unambig != "" {
-			return []nemotronEvent{nemotronEventContent{content: unambig}}, false
+
+		if remainder == "" {
+			p.state = Nemotron3NanoSkipWhitespaceAfterThinking
+		} else {
+			p.state = Nemotron3NanoCollectingContent
+			content, _, calls, err = p.toolParser.Add(remainder, done)
 		}
-		return nil, false
-
-	case Nemotron3NanoCollectingToolCalls:
-		if strings.Contains(bufStr, nemotronToolCallClose) {
-			split := strings.SplitN(bufStr, nemotronToolCallClose, 2)
-			remaining := strings.TrimLeftFunc(split[1], unicode.IsSpace)
-			p.buffer.Reset()
-			p.buffer.WriteString(remaining)
-
-			var events []nemotronEvent
-			if tc, err := p.parseToolCall(split[0]); err == nil {
-				events = append(events, nemotronEventToolCall{toolCall: tc})
-			}
-
-			if !strings.Contains(remaining, nemotronToolCallOpen) {
-				p.state = Nemotron3NanoCollectingContent
-			}
-			return events, true
-		}
-		return nil, false
+		return content, thinking, calls, err
 	}
 
-	return nil, false
+	// No end marker - emit unambiguous thinking
+	thinking = p.emitThinking(bufStr)
+	return "", thinking, nil, nil
 }
 
-var (
-	nemotronFunctionRegex  = regexp.MustCompile(`<function=([^>]+)>`)
-	nemotronParameterRegex = regexp.MustCompile(`<parameter=([^>]+)>\n?([\s\S]*?)\n?</parameter>`)
-)
+// emitThinking returns unambiguous thinking content, keeping potential partial tags in buffer
+func (p *Nemotron3NanoParser) emitThinking(bufStr string) string {
+	// Check for partial </think> or <tool_call> at end
+	thinkOverlap := overlap(bufStr, nemotronThinkClose)
+	toolOverlap := overlap(bufStr, nemotronToolCallOpen)
+	maxOverlap := max(thinkOverlap, toolOverlap)
 
-func (p *Nemotron3NanoParser) parseToolCall(content string) (api.ToolCall, error) {
-	toolCall := api.ToolCall{}
-
-	// Extract function name
-	fnMatch := nemotronFunctionRegex.FindStringSubmatch(content)
-	if len(fnMatch) < 2 {
-		return toolCall, nil
-	}
-	toolCall.Function.Name = fnMatch[1]
-
-	// Extract parameters
-	toolCall.Function.Arguments = api.NewToolCallFunctionArguments()
-	paramMatches := nemotronParameterRegex.FindAllStringSubmatch(content, -1)
-	for _, match := range paramMatches {
-		if len(match) >= 3 {
-			paramName := match[1]
-			paramValue := strings.TrimSpace(match[2])
-
-			// Try to parse as typed value based on tool definition
-			toolCall.Function.Arguments.Set(paramName, p.parseParamValue(paramName, paramValue))
-		}
+	if maxOverlap > 0 {
+		unambiguous := bufStr[:len(bufStr)-maxOverlap]
+		unambiguous = strings.TrimRightFunc(unambiguous, unicode.IsSpace)
+		p.buffer.Reset()
+		p.buffer.WriteString(bufStr[len(bufStr)-maxOverlap:])
+		return unambiguous
 	}
 
-	return toolCall, nil
-}
-
-func (p *Nemotron3NanoParser) parseParamValue(paramName string, raw string) any {
-	// Find the matching tool to get parameter type
-	var paramType api.PropertyType
-	for _, tool := range p.tools {
-		if tool.Function.Parameters.Properties != nil {
-			if prop, ok := tool.Function.Parameters.Properties.Get(paramName); ok {
-				paramType = prop.Type
-				break
-			}
-		}
-	}
-
-	return parseValue(raw, paramType)
+	// No partial tags - emit all but trailing whitespace
+	wsLen := trailingWhitespaceLen(bufStr)
+	if wsLen > 0 {
+		unambiguous := bufStr[:len(bufStr)-wsLen]
+		p.buffer.Reset()
+		p.buffer.WriteString(bufStr[len(bufStr)-wsLen:])
+		return unambiguous
+	}
+
+	// Nothing to hold back
+	p.buffer.Reset()
+	return bufStr
 }

model/parsers/nemotron3nano_test.go

@@ -8,6 +8,8 @@ import (
 	"github.com/ollama/ollama/api"
 )
 
+// TestNemotron3NanoParser tests Nemotron-specific behavior (thinking support).
+// Tool call parsing is tested in qwen3coder_test.go since Nemotron delegates to Qwen3CoderParser.
 func TestNemotron3NanoParser(t *testing.T) {
 	tests := []struct {
 		name             string
@@ -17,18 +19,6 @@ func TestNemotron3NanoParser(t *testing.T) {
 		expectedThinking string
 		expectedCalls    []api.ToolCall
 	}{
-		{
-			name:            "simple content - no thinking",
-			input:           "Hello, how can I help you?",
-			thinkValue:      nil,
-			expectedContent: "Hello, how can I help you?",
-		},
-		{
-			name:            "simple content - thinking disabled",
-			input:           "Hello, how can I help you?",
-			thinkValue:      &api.ThinkValue{Value: false},
-			expectedContent: "Hello, how can I help you?",
-		},
 		{
 			name:             "thinking then content",
 			input:            "Let me think about this...</think>\nHere is my answer.",
@@ -43,69 +33,6 @@ func TestNemotron3NanoParser(t *testing.T) {
 			expectedThinking: "Step 1: Analyze\nStep 2: Process\nStep 3: Conclude",
 			expectedContent:  "The answer is 42.",
 		},
-		{
-			name:       "simple tool call",
-			input:      "<tool_call>\n<function=get_weather>\n<parameter=city>\nParis\n</parameter>\n</function>\n</tool_call>",
-			thinkValue: nil,
-			expectedCalls: []api.ToolCall{
-				{
-					Function: api.ToolCallFunction{
-						Name:      "get_weather",
-						Arguments: testArgs(map[string]any{"city": "Paris"}),
-					},
-				},
-			},
-		},
-		{
-			name:            "content then tool call",
-			input:           "Let me check the weather.\n<tool_call>\n<function=get_weather>\n<parameter=city>\nNYC\n</parameter>\n</function>\n</tool_call>",
-			thinkValue:      nil,
-			expectedContent: "Let me check the weather.",
-			expectedCalls: []api.ToolCall{
-				{
-					Function: api.ToolCallFunction{
-						Name:      "get_weather",
-						Arguments: testArgs(map[string]any{"city": "NYC"}),
-					},
-				},
-			},
-		},
-		{
-			name:       "tool call with multiple parameters",
-			input:      "<tool_call>\n<function=book_flight>\n<parameter=from>\nSFO\n</parameter>\n<parameter=to>\nNYC\n</parameter>\n</function>\n</tool_call>",
-			thinkValue: nil,
-			expectedCalls: []api.ToolCall{
-				{
-					Function: api.ToolCallFunction{
-						Name: "book_flight",
-						Arguments: testArgs(map[string]any{
-							"from": "SFO",
-							"to":   "NYC",
-						}),
-					},
-				},
-			},
-		},
-		{
-			name: "multiple tool calls",
-			input: "<tool_call>\n<function=get_weather>\n<parameter=city>\nSan Francisco\n</parameter>\n</function>\n</tool_call>\n" +
-				"<tool_call>\n<function=get_weather>\n<parameter=city>\nNew York\n</parameter>\n</function>\n</tool_call>",
-			thinkValue: nil,
-			expectedCalls: []api.ToolCall{
-				{
-					Function: api.ToolCallFunction{
-						Name:      "get_weather",
-						Arguments: testArgs(map[string]any{"city": "San Francisco"}),
-					},
-				},
-				{
-					Function: api.ToolCallFunction{
-						Name:      "get_weather",
-						Arguments: testArgs(map[string]any{"city": "New York"}),
-					},
-				},
-			},
-		},
 		{
 			name:             "thinking then tool call",
 			input:            "I should check the weather...</think>\n<tool_call>\n<function=get_weather>\n<parameter=city>\nParis\n</parameter>\n</function>\n</tool_call>",
@@ -135,19 +62,6 @@ func TestNemotron3NanoParser(t *testing.T) {
 				},
 			},
 		},
-		{
-			name:       "tool call with multiline parameter value",
-			input:      "<tool_call>\n<function=create_note>\n<parameter=content>\nLine 1\nLine 2\nLine 3\n</parameter>\n</function>\n</tool_call>",
-			thinkValue: nil,
-			expectedCalls: []api.ToolCall{
-				{
-					Function: api.ToolCallFunction{
-						Name:      "create_note",
-						Arguments: testArgs(map[string]any{"content": "Line 1\nLine 2\nLine 3"}),
-					},
-				},
-			},
-		},
 		{
 			name:             "empty thinking block - immediate close",
 			input:            "</think>\nHere is my answer.",
@@ -161,18 +75,6 @@ func TestNemotron3NanoParser(t *testing.T) {
 			thinkValue:      &api.ThinkValue{Value: false},
 			expectedContent: "</think>\nSome content after spurious tag.",
 		},
-		{
-			name:          "tool call with no function name - returns empty tool call",
-			input:         "<tool_call>\n<function=>\n</function>\n</tool_call>",
-			thinkValue:    nil,
-			expectedCalls: []api.ToolCall{{Function: api.ToolCallFunction{Name: "", Arguments: api.NewToolCallFunctionArguments()}}},
-		},
-		{
-			name:            "content with newlines preserved",
-			input:           "Line 1\n\nLine 2\n\n\nLine 3",
-			thinkValue:      nil,
-			expectedContent: "Line 1\n\nLine 2\n\n\nLine 3",
-		},
 		{
 			name:             "thinking with only whitespace after close tag",
 			input:            "My thoughts...</think>   \n\t\n   Content here.",
@@ -180,25 +82,6 @@ func TestNemotron3NanoParser(t *testing.T) {
 			expectedThinking: "My thoughts...",
 			expectedContent:  "Content here.",
 		},
-		{
-			name:            "unicode content",
-			input:           "Hello 世界! 🌍 Ñoño",
-			thinkValue:      nil,
-			expectedContent: "Hello 世界! 🌍 Ñoño",
-		},
-		{
-			name:       "tool call with numeric parameter",
-			input:      "<tool_call>\n<function=set_temp>\n<parameter=value>\n42\n</parameter>\n</function>\n</tool_call>",
-			thinkValue: nil,
-			expectedCalls: []api.ToolCall{
-				{
-					Function: api.ToolCallFunction{
-						Name:      "set_temp",
-						Arguments: testArgs(map[string]any{"value": "42"}),
-					},
-				},
-			},
-		},
 	}
 
 	for _, tt := range tests {
@@ -233,6 +116,8 @@ func TestNemotron3NanoParser(t *testing.T) {
 	}
 }
 
+// TestNemotron3NanoParser_Streaming tests streaming behavior for thinking support.
+// Tool call streaming is tested in qwen3coder_test.go.
 func TestNemotron3NanoParser_Streaming(t *testing.T) {
 	tests := []struct {
 		name             string
@@ -242,18 +127,6 @@ func TestNemotron3NanoParser_Streaming(t *testing.T) {
 		expectedThinking string
 		expectedCalls    []api.ToolCall
 	}{
-		{
-			name:            "streaming content character by character",
-			chunks:          []string{"H", "e", "l", "l", "o", ",", " ", "w", "o", "r", "l", "d", "!"},
-			thinkValue:      nil,
-			expectedContent: "Hello, world!",
-		},
-		{
-			name:            "streaming content small tokens",
-			chunks:          []string{"Hel", "lo", ", ", "how ", "can", " I", " help", " you", " today", "?"},
-			thinkValue:      nil,
-			expectedContent: "Hello, how can I help you today?",
-		},
 		{
 			name:             "streaming thinking then content - granular",
 			chunks:           []string{"Let", " me", " th", "ink", " about", " this", "...", "<", "/", "think", ">", "\n", "Here", " is", " my", " answer", "."},
@@ -268,45 +141,6 @@ func TestNemotron3NanoParser_Streaming(t *testing.T) {
 			expectedThinking: "Step 1: Analyze\nStep 2: Process",
 			expectedContent:  "The answer.",
 		},
-		{
-			name:       "streaming tool call - highly granular",
-			chunks:     []string{"<", "tool", "_", "call", ">", "\n", "<", "func", "tion", "=", "get", "_", "weather", ">", "\n", "<", "param", "eter", "=", "city", ">", "\n", "Par", "is", "\n", "</", "param", "eter", ">", "\n", "</", "func", "tion", ">", "\n", "</", "tool", "_", "call", ">"},
-			thinkValue: nil,
-			expectedCalls: []api.ToolCall{
-				{
-					Function: api.ToolCallFunction{
-						Name:      "get_weather",
-						Arguments: testArgs(map[string]any{"city": "Paris"}),
-					},
-				},
-			},
-		},
-		{
-			name:            "streaming content then tool call - granular",
-			chunks:          []string{"Let", " me", " check", " the", " weather", ".", "\n<", "tool_call", ">", "\n", "<function=", "get_weather", ">", "\n", "<parameter=", "city", ">", "\n", "NYC", "\n", "</parameter>", "\n", "</function>", "\n", "</tool_call>"},
-			thinkValue:      nil,
-			expectedContent: "Let me check the weather.",
-			expectedCalls: []api.ToolCall{
-				{
-					Function: api.ToolCallFunction{
-						Name:      "get_weather",
-						Arguments: testArgs(map[string]any{"city": "NYC"}),
-					},
-				},
-			},
-		},
-		{
-			name:   "tool call tag split character by character",
-			chunks: []string{"<", "t", "o", "o", "l", "_", "c", "a", "l", "l", ">", "\n", "<", "f", "u", "n", "c", "t", "i", "o", "n", "=", "t", "e", "s", "t", ">", "\n", "<", "/", "f", "u", "n", "c", "t", "i", "o", "n", ">", "\n", "<", "/", "t", "o", "o", "l", "_", "c", "a", "l", "l", ">"},
-			expectedCalls: []api.ToolCall{
-				{
-					Function: api.ToolCallFunction{
-						Name:      "test",
-						Arguments: api.NewToolCallFunctionArguments(),
-					},
-				},
-			},
-		},
 		{
 			name:             "thinking close tag split character by character",
 			chunks:           []string{"I", "'", "m", " ", "t", "h", "i", "n", "k", "i", "n", "g", ".", ".", ".", "<", "/", "t", "h", "i", "n", "k", ">", "\n", "D", "o", "n", "e", "!"},
@@ -321,22 +155,6 @@ func TestNemotron3NanoParser_Streaming(t *testing.T) {
 			expectedThinking: "Thinking...",
 			expectedContent:  "Content here.",
 		},
-		{
-			name:       "tool call with multiple parameters - streaming",
-			chunks:     []string{"<tool_", "call>\n", "<function", "=book_", "flight>", "\n<para", "meter=", "from>\n", "SFO\n", "</param", "eter>", "\n<param", "eter=to", ">\nNYC", "\n</para", "meter>", "\n</func", "tion>\n", "</tool_", "call>"},
-			thinkValue: nil,
-			expectedCalls: []api.ToolCall{
-				{
-					Function: api.ToolCallFunction{
-						Name: "book_flight",
-						Arguments: testArgs(map[string]any{
-							"from": "SFO",
-							"to":   "NYC",
-						}),
-					},
-				},
-			},
-		},
 		{
 			name:             "thinking then content then tool call - streaming",
 			chunks:           []string{"Ana", "lyzing", " your", " request", "...", "</", "think", ">\n", "I'll", " check", " that", " for", " you", ".", "\n", "<tool", "_call", ">\n", "<function", "=search", ">\n", "<parameter", "=query", ">\n", "test", " query", "\n</", "parameter", ">\n", "</function", ">\n", "</tool", "_call", ">"},
@@ -352,45 +170,6 @@ func TestNemotron3NanoParser_Streaming(t *testing.T) {
 				},
 			},
 		},
-		{
-			name: "multiple tool calls - streaming",
-			chunks: []string{
-				"<tool_call>", "\n", "<function=", "get_weather>", "\n",
-				"<parameter=", "city>\n", "San Fran", "cisco\n", "</parameter>", "\n",
-				"</function>", "\n", "</tool_call>", "\n",
-				"<tool_", "call>\n", "<function", "=get_weather", ">\n",
-				"<param", "eter=city", ">\nNew", " York\n", "</parameter>\n",
-				"</function>\n", "</tool_call>",
-			},
-			thinkValue: nil,
-			expectedCalls: []api.ToolCall{
-				{
-					Function: api.ToolCallFunction{
-						Name:      "get_weather",
-						Arguments: testArgs(map[string]any{"city": "San Francisco"}),
-					},
-				},
-				{
-					Function: api.ToolCallFunction{
-						Name:      "get_weather",
-						Arguments: testArgs(map[string]any{"city": "New York"}),
-					},
-				},
-			},
-		},
-		{
-			name:       "tool call with multiline parameter - streaming",
-			chunks:     []string{"<tool_call>\n", "<function=", "create_note>\n", "<parameter=", "content>\n", "Line 1", "\nLine", " 2\n", "Line 3", "\n</parameter>\n", "</function>\n", "</tool_call>"},
-			thinkValue: nil,
-			expectedCalls: []api.ToolCall{
-				{
-					Function: api.ToolCallFunction{
-						Name:      "create_note",
-						Arguments: testArgs(map[string]any{"content": "Line 1\nLine 2\nLine 3"}),
-					},
-				},
-			},
-		},
 		{
 			name:             "empty thinking block",
 			chunks:           []string{"</think>", "\n", "Just content."},
@@ -398,12 +177,6 @@ func TestNemotron3NanoParser_Streaming(t *testing.T) {
 			expectedThinking: "",
 			expectedContent:  "Just content.",
 		},
-		{
-			name:            "empty input chunks interspersed",
-			chunks:          []string{"Hello", "", " ", "", "world", "", "!"},
-			thinkValue:      nil,
-			expectedContent: "Hello world!",
-		},
 		{
 			name:             "tool call immediately after think close - no content",
 			chunks:           []string{"Analyzing...", "</think>", "\n", "<tool_call>", "\n<function=test>\n</function>\n", "</tool_call>"},
@@ -418,25 +191,6 @@ func TestNemotron3NanoParser_Streaming(t *testing.T) {
 				},
 			},
 		},
-		{
-			name:       "tool call with empty parameter value",
-			chunks:     []string{"<tool_call>\n<function=test>\n<parameter=name>\n", "\n</parameter>\n</function>\n</tool_call>"},
-			thinkValue: nil,
-			expectedCalls: []api.ToolCall{
-				{
-					Function: api.ToolCallFunction{
-						Name:      "test",
-						Arguments: testArgs(map[string]any{"name": ""}),
-					},
-				},
-			},
-		},
-		{
-			name:            "partial tool call tag at end - buffered",
-			chunks:          []string{"Here's some content", "<tool"},
-			thinkValue:      nil,
-			expectedContent: "Here's some content",
-		},
 	}
 
 	for _, tt := range tests {
@@ -572,3 +326,65 @@ func TestNemotron3NanoParser_WithTools(t *testing.T) {
 		t.Errorf("calls mismatch (-got +want):\n%s", diff)
 	}
 }
+
+// TestNemotron3NanoParser_ToolCallWithoutThinkClose tests the case where thinking is enabled
+// but the model outputs content + tool call WITHOUT the </think> tag.
+// The parser should still parse the tool call (content before is treated as thinking).
+func TestNemotron3NanoParser_ToolCallWithoutThinkClose(t *testing.T) {
+	chunks := []string{
+		"Let", " me", " analyze", " this", ".", "\n",
+		"<tool_call>", "\n",
+		"<function=get_weather>", "\n",
+		"<parameter=city>", "Paris", "</parameter>", "\n",
+		"</function>", "\n",
+		"</tool_call>",
+	}
+
+	p := &Nemotron3NanoParser{}
+	p.Init(nil, nil, &api.ThinkValue{Value: true}) // thinking ENABLED but model doesn't output </think>
+
+	var allContent string
+	var allThinking string
+	var allCalls []api.ToolCall
+
+	for _, chunk := range chunks {
+		content, thinking, calls, err := p.Add(chunk, false)
+		if err != nil {
+			t.Fatalf("unexpected error: %v", err)
+		}
+		allContent += content
+		allThinking += thinking
+		allCalls = append(allCalls, calls...)
+	}
+
+	// Drain
+	content, thinking, calls, err := p.Add("", true)
+	if err != nil {
+		t.Fatalf("unexpected error on done: %v", err)
+	}
+	allContent += content
+	allThinking += thinking
+	allCalls = append(allCalls, calls...)
+
+	// The parser was in thinking mode, so text before <tool_call> is emitted as thinking.
+	expectedThinking := "Let me analyze this."
+
+	expectedCalls := []api.ToolCall{
+		{
+			Function: api.ToolCallFunction{
+				Name:      "get_weather",
+				Arguments: testArgs(map[string]any{"city": "Paris"}),
+			},
+		},
+	}
+
+	if allContent != "" {
+		t.Errorf("expected no content (text was streamed as thinking), got: %q", allContent)
+	}
+	if diff := cmp.Diff(allThinking, expectedThinking); diff != "" {
+		t.Errorf("thinking mismatch (-got +want):\n%s", diff)
+	}
+	if diff := cmp.Diff(allCalls, expectedCalls, argsComparer); diff != "" {
+		t.Errorf("calls mismatch (-got +want):\n%s", diff)
+	}
+}

model/parsers/parsers.go

@@ -68,6 +68,8 @@ func ParserForName(name string) Parser {
 		return &Nemotron3NanoParser{}
 	case "functiongemma":
 		return &FunctionGemmaParser{}
+	case "glm-4.7":
+		return &GLM47Parser{}
 	default:
 		return nil
 	}

model/parsers/qwen3coder_test.go

@@ -91,6 +91,37 @@ func TestQwenParserStreaming(t *testing.T) {
 				},
 			},
 		},
+		{
+			desc: "tool call tags split character by character",
+			steps: []step{
+				{input: "<", wantEvents: []qwenEvent{}},
+				{input: "t", wantEvents: []qwenEvent{}},
+				{input: "o", wantEvents: []qwenEvent{}},
+				{input: "o", wantEvents: []qwenEvent{}},
+				{input: "l", wantEvents: []qwenEvent{}},
+				{input: "_", wantEvents: []qwenEvent{}},
+				{input: "c", wantEvents: []qwenEvent{}},
+				{input: "a", wantEvents: []qwenEvent{}},
+				{input: "l", wantEvents: []qwenEvent{}},
+				{input: "l", wantEvents: []qwenEvent{}},
+				{input: ">", wantEvents: []qwenEvent{}},
+				{input: "a", wantEvents: []qwenEvent{}},
+				{input: "b", wantEvents: []qwenEvent{}},
+				{input: "c", wantEvents: []qwenEvent{}},
+				{input: "<", wantEvents: []qwenEvent{}},
+				{input: "/", wantEvents: []qwenEvent{}},
+				{input: "t", wantEvents: []qwenEvent{}},
+				{input: "o", wantEvents: []qwenEvent{}},
+				{input: "o", wantEvents: []qwenEvent{}},
+				{input: "l", wantEvents: []qwenEvent{}},
+				{input: "_", wantEvents: []qwenEvent{}},
+				{input: "c", wantEvents: []qwenEvent{}},
+				{input: "a", wantEvents: []qwenEvent{}},
+				{input: "l", wantEvents: []qwenEvent{}},
+				{input: "l", wantEvents: []qwenEvent{}},
+				{input: ">", wantEvents: []qwenEvent{qwenEventRawToolCall{raw: "abc"}}},
+			},
+		},
 		{
 			desc: "trailing whitespace between content and tool call",
 			steps: []step{

model/parsers/testhelpers_test.go

@@ -96,3 +96,11 @@ func testArgs(m map[string]any) api.ToolCallFunctionArguments {
 	}
 	return args
 }
+
+func args(s string) api.ToolCallFunctionArguments {
+	var result api.ToolCallFunctionArguments
+	if err := json.Unmarshal([]byte(s), &result); err != nil {
+		panic("invalid JSON in args(): " + err.Error())
+	}
+	return result
+}

model/renderers/glm46.go

@@ -0,0 +1,110 @@
+package renderers
+
+import (
+	"encoding/json"
+	"fmt"
+	"strings"
+
+	"github.com/ollama/ollama/api"
+)
+
+type GLM46Renderer struct{}
+
+func (r *GLM46Renderer) Render(messages []api.Message, tools []api.Tool, thinkValue *api.ThinkValue) (string, error) {
+	var sb strings.Builder
+
+	sb.WriteString("[gMASK]<sop>")
+
+	var lastUserIndex int
+	for i, message := range messages {
+		if message.Role == "user" {
+			lastUserIndex = i
+		}
+	}
+
+	if len(tools) > 0 {
+		sb.WriteString("<|system|>\n")
+		sb.WriteString("# Tools\n\n")
+		sb.WriteString("You may call one or more functions to assist with the user query.\n\n")
+		sb.WriteString("You are provided with function signatures within <tools></tools> XML tags:\n")
+		sb.WriteString("<tools>\n")
+		for _, tool := range tools {
+			d, _ := json.Marshal(tool)
+			sb.WriteString(string(d) + "\n")
+		}
+		sb.WriteString("</tools>\n\n")
+		sb.WriteString("For each function call, output the function name and arguments within the following XML format:\n")
+		sb.WriteString("<tool_call>{function-name}\n")
+		sb.WriteString("<arg_key>{arg-key-1}</arg_key>\n")
+		sb.WriteString("<arg_value>{arg-value-1}</arg_value>\n")
+		sb.WriteString("<arg_key>{arg-key-2}</arg_key>\n")
+		sb.WriteString("<arg_value>{arg-value-2}</arg_value>\n")
+		sb.WriteString("...\n")
+		sb.WriteString("</tool_call>")
+	}
+
+	for i, message := range messages {
+		switch message.Role {
+		case "user":
+			sb.WriteString("<|user|>\n")
+			sb.WriteString(message.Content)
+			if thinkValue != nil && !thinkValue.Bool() && !strings.HasSuffix(message.Content, "/nothink") {
+				sb.WriteString("/nothink")
+			}
+		case "assistant":
+			sb.WriteString("<|assistant|>")
+			if i > lastUserIndex {
+				if message.Thinking != "" {
+					sb.WriteString("\n<think>" + message.Thinking + "</think>")
+				} else {
+					sb.WriteString("\n<think></think>")
+				}
+			}
+			if message.Content != "" {
+				sb.WriteString("\n" + message.Content)
+			}
+			if len(message.ToolCalls) > 0 {
+				for _, toolCall := range message.ToolCalls {
+					sb.WriteString("\n<tool_call>" + toolCall.Function.Name + "\n")
+					for key, value := range toolCall.Function.Arguments.All() {
+						sb.WriteString("<arg_key>" + key + "</arg_key>\n")
+
+						var valueStr string
+						if str, ok := value.(string); ok {
+							valueStr = str
+						} else {
+							jsonBytes, err := json.Marshal(value)
+							if err != nil {
+								valueStr = fmt.Sprintf("%v", value)
+							} else {
+								valueStr = string(jsonBytes)
+							}
+						}
+
+						sb.WriteString("<arg_value>" + valueStr + "</arg_value>\n")
+					}
+
+					sb.WriteString("</tool_call>")
+				}
+			}
+		case "tool":
+			if i == 0 || messages[i-1].Role != "tool" {
+				sb.WriteString("<|observation|>")
+			}
+			sb.WriteString("\n<tool_response>\n")
+			sb.WriteString(message.Content)
+			sb.WriteString("\n</tool_response>")
+		case "system":
+			sb.WriteString("<|system|>\n")
+			sb.WriteString(message.Content)
+		}
+	}
+
+	// Add generation prompt
+	sb.WriteString("<|assistant|>")
+	if thinkValue != nil && !thinkValue.Bool() {
+		sb.WriteString("\n<think></think>\n")
+	}
+
+	return sb.String(), nil
+}

model/renderers/glm46_test.go

@@ -0,0 +1,223 @@
+package renderers
+
+import (
+	"testing"
+
+	"github.com/google/go-cmp/cmp"
+	"github.com/ollama/ollama/api"
+)
+
+func TestGLM46Renderer(t *testing.T) {
+	tests := []struct {
+		name       string
+		messages   []api.Message
+		tools      []api.Tool
+		thinkValue *api.ThinkValue
+		expected   string
+		skip       string
+	}{
+		{
+			name: "basic",
+			messages: []api.Message{
+				{Role: "user", Content: "Hello, how are you?"},
+			},
+			expected: `[gMASK]<sop><|user|>
+Hello, how are you?<|assistant|>`,
+		},
+		{
+			name: "basic with system message",
+			messages: []api.Message{
+				{Role: "system", Content: "You are a helpful assistant."},
+				{Role: "user", Content: "Hello, how are you?"},
+			},
+			expected: `[gMASK]<sop><|system|>
+You are a helpful assistant.<|user|>
+Hello, how are you?<|assistant|>`,
+		},
+		{
+			name: "basic with user assistant user",
+			messages: []api.Message{
+				{Role: "user", Content: "What is the capital of France?"},
+				{Role: "assistant", Thinking: "Let me analyze the request...", Content: "The capital of France is Paris."},
+				{Role: "user", Content: "Fantastic!"},
+			},
+			expected: `[gMASK]<sop><|user|>
+What is the capital of France?<|assistant|>
+The capital of France is Paris.<|user|>
+Fantastic!<|assistant|>`,
+		},
+		{
+			skip: "tool call ordering not guaranteed yet",
+			name: "tools",
+			messages: []api.Message{
+				{Role: "system", Content: "You are a helpful assistant with access to tools."},
+				{Role: "user", Content: "What is the weather like in Tokyo?"},
+			},
+			tools: []api.Tool{
+				{
+					Type: "function",
+					Function: api.ToolFunction{
+						Name:        "get_weather",
+						Description: "Get the current weather in a given location",
+						Parameters: api.ToolFunctionParameters{
+							Type:       "object",
+							Required:   []string{"location"},
+							Properties: propsMap(`{"location": {"type": "string", "description": "The city and state, e.g. San Francisco, CA"}, "unit": {"type": "string", "enum": ["celsius", "fahrenheit"]}}`),
+						},
+					},
+				},
+			},
+			expected: `[gMASK]<sop><|system|>
+# Tools
+
+You may call one or more functions to assist with the user query.
+
+You are provided with function signatures within <tools></tools> XML tags:
+<tools>
+{"type":"function","function":{"name":"get_weather","description":"Get the current weather in a given location","parameters":{"type":"object","required":["location"],"properties":{"location":{"type":"string","description":"The city and state, e.g. San Francisco, CA"},"unit":{"type":"string","description":"","enum":["celsius","fahrenheit"]}}}}}
+</tools>
+
+For each function call, output the function name and arguments within the following XML format:
+<tool_call>{function-name}
+<arg_key>{arg-key-1}</arg_key>
+<arg_value>{arg-value-1}</arg_value>
+<arg_key>{arg-key-2}</arg_key>
+<arg_value>{arg-value-2}</arg_value>
+...
+</tool_call><|system|>
+You are a helpful assistant with access to tools.<|user|>
+What is the weather like in Tokyo?<|assistant|>`,
+		},
+		{
+			skip: "tool call ordering not guaranteed yet",
+			name: "tool calls",
+			messages: []api.Message{
+				{Role: "system", Content: "You are a helpful assistant with access to tools."},
+				{Role: "user", Content: "What is the weather like in Tokyo?"},
+				{
+					Role: "assistant",
+					ToolCalls: []api.ToolCall{
+						{
+							Function: api.ToolCallFunction{
+								Name:      "get_weather",
+								Arguments: args(`{"location": "Tokyo, Japan", "unit": "celsius"}`),
+							},
+						},
+						{
+							Function: api.ToolCallFunction{
+								Name:      "get_weather",
+								Arguments: args(`{"location": "Japan", "unit": "fahrenheit"}`),
+							},
+						},
+					},
+				},
+				{
+					Role:     "tool",
+					Content:  "{\"temperature\": 22, \"weather\": \"partly cloudy\", \"humidity\": 65}",
+					ToolName: "get_weather",
+				},
+				{
+					Role:     "tool",
+					Content:  "{\"temperature\": 68, \"weather\": \"sunny\", \"humidity\": 75}",
+					ToolName: "get_weather",
+				},
+				{
+					Role:    "assistant",
+					Content: "The weather in Tokyo is currently partly cloudy with a temperature of 22°C and 65% humidity. It's a pleasant day with moderate temperatures.",
+				},
+			},
+			tools: []api.Tool{
+				{
+					Type: "function",
+					Function: api.ToolFunction{
+						Name:        "get_weather",
+						Description: "Get the current weather in a given location",
+						Parameters: api.ToolFunctionParameters{
+							Type:       "object",
+							Required:   []string{"location"},
+							Properties: propsMap(`{"location": {"type": "string", "description": "The city and state, e.g. San Francisco, CA"}, "unit": {"type": "string", "enum": ["celsius", "fahrenheit"]}}`),
+						},
+					},
+				},
+			},
+			expected: `[gMASK]<sop><|system|>
+# Tools
+
+You may call one or more functions to assist with the user query.
+
+You are provided with function signatures within <tools></tools> XML tags:
+<tools>
+{"type":"function","function":{"name":"get_weather","description":"Get the current weather in a given location","parameters":{"type":"object","required":["location"],"properties":{"location":{"type":"string","description":"The city and state, e.g. San Francisco, CA"},"unit":{"type":"string","description":"","enum":["celsius","fahrenheit"]}}}}}
+</tools>
+
+For each function call, output the function name and arguments within the following XML format:
+<tool_call>{function-name}
+<arg_key>{arg-key-1}</arg_key>
+<arg_value>{arg-value-1}</arg_value>
+<arg_key>{arg-key-2}</arg_key>
+<arg_value>{arg-value-2}</arg_value>
+...
+</tool_call><|system|>
+You are a helpful assistant with access to tools.<|user|>
+What is the weather like in Tokyo?<|assistant|>
+<think></think>
+<tool_call>get_weather
+<arg_key>location</arg_key>
+<arg_value>Tokyo, Japan</arg_value>
+<arg_key>unit</arg_key>
+<arg_value>celsius</arg_value>
+</tool_call>
+<tool_call>get_weather
+<arg_key>location</arg_key>
+<arg_value>Japan</arg_value>
+<arg_key>unit</arg_key>
+<arg_value>fahrenheit</arg_value>
+</tool_call><|observation|>
+<tool_response>
+{"temperature": 22, "weather": "partly cloudy", "humidity": 65}
+</tool_response>
+<tool_response>
+{"temperature": 68, "weather": "sunny", "humidity": 75}
+</tool_response><|assistant|>
+<think></think>
+The weather in Tokyo is currently partly cloudy with a temperature of 22°C and 65% humidity. It's a pleasant day with moderate temperatures.<|assistant|>`,
+		},
+		{
+			name: "think true",
+			messages: []api.Message{
+				{Role: "user", Content: "Hello, how are you?"},
+			},
+			thinkValue: &api.ThinkValue{Value: true},
+			expected: `[gMASK]<sop><|user|>
+Hello, how are you?<|assistant|>`,
+		},
+		{
+			name: "think false",
+			messages: []api.Message{
+				{Role: "user", Content: "Hello, how are you?"},
+			},
+			thinkValue: &api.ThinkValue{Value: false},
+			expected: `[gMASK]<sop><|user|>
+Hello, how are you?/nothink<|assistant|>
+<think></think>
+`,
+		},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			if tt.skip != "" {
+				t.Skip(tt.skip)
+			}
+			renderer := &GLM46Renderer{}
+			rendered, err := renderer.Render(tt.messages, tt.tools, tt.thinkValue)
+			if err != nil {
+				t.Fatal(err)
+			}
+			if diff := cmp.Diff(rendered, tt.expected); diff != "" {
+				t.Errorf("mismatch (-got +want):\n%s", diff)
+				t.Logf("Got:\n%s", rendered)
+				t.Logf("Expected:\n%s", tt.expected)
+			}
+		})
+	}
+}

model/renderers/glm47.go

@@ -0,0 +1,170 @@
+package renderers
+
+import (
+	"encoding/json"
+	"fmt"
+	"strings"
+
+	"github.com/ollama/ollama/api"
+)
+
+// GLM47Renderer renders messages for GLM-4.7 models.
+//
+// GLM-4.7 Thinking Modes (ref: https://docs.z.ai/guides/capabilities/thinking-mode):
+//
+//  1. INTERLEAVED THINKING
+//     The model thinks between tool calls and after receiving tool results.
+//     This enables complex step-by-step reasoning: interpreting each tool output
+//     before deciding what to do next. Thinking blocks are preserved and returned
+//     with tool results to maintain reasoning continuity.
+//
+//  2. PRESERVED THINKING
+//     The model retains reasoning content from previous assistant turns in context.
+//     This preserves reasoning continuity across multi-turn conversations. The
+//     upstream API has a "clear_thinking" parameter to control this:
+//     - clear_thinking=true:  clears reasoning from previous turns (outputs </think>)
+//     - clear_thinking=false: preserves <think>...</think> blocks from previous turns
+//
+//  3. TURN-LEVEL THINKING
+//     Controls whether the model should reason on each turn. The upstream API
+//     uses "enable_thinking" parameter:
+//     - enable_thinking=true:  outputs <think> to start reasoning
+//     - enable_thinking=false: outputs </think> to skip reasoning
+//
+// OLLAMA DEFAULTS:
+//   - Thinking is ENABLED by default (thinkValue=nil or true outputs <think>)
+//   - Thinking is PRESERVED by default (reasoning content from previous turns is always
+//     included in <think>...</think> blocks, equivalent to clear_thinking=false)
+//   - Users can disable thinking per-turn via thinkValue=false
+type GLM47Renderer struct{}
+
+func (r *GLM47Renderer) Render(messages []api.Message, tools []api.Tool, thinkValue *api.ThinkValue) (string, error) {
+	var sb strings.Builder
+
+	sb.WriteString("[gMASK]<sop>")
+
+	if len(tools) > 0 {
+		sb.WriteString("<|system|>\n")
+		sb.WriteString("# Tools\n\n")
+		sb.WriteString("You may call one or more functions to assist with the user query.\n\n")
+		sb.WriteString("You are provided with function signatures within <tools></tools> XML tags:\n")
+		sb.WriteString("<tools>\n")
+		for _, tool := range tools {
+			d, _ := json.Marshal(tool)
+			sb.WriteString(formatGLM47ToolJSON(d))
+			sb.WriteString("\n")
+		}
+		sb.WriteString("</tools>\n\n")
+		sb.WriteString("For each function call, output the function name and arguments within the following XML format:\n")
+		sb.WriteString("<tool_call>{function-name}<arg_key>{arg-key-1}</arg_key><arg_value>{arg-value-1}</arg_value><arg_key>{arg-key-2}</arg_key><arg_value>{arg-value-2}</arg_value>...</tool_call>")
+	}
+
+	think := true
+	if thinkValue != nil && !thinkValue.Bool() {
+		think = false
+	}
+
+	for i, message := range messages {
+		switch message.Role {
+		case "user":
+			sb.WriteString("<|user|>")
+			sb.WriteString(message.Content)
+		case "assistant":
+			sb.WriteString("<|assistant|>")
+			if message.Thinking != "" {
+				sb.WriteString("<think>" + message.Thinking + "</think>")
+			} else {
+				sb.WriteString("</think>")
+			}
+			if message.Content != "" {
+				sb.WriteString(message.Content)
+			}
+			if len(message.ToolCalls) > 0 {
+				for _, toolCall := range message.ToolCalls {
+					sb.WriteString("<tool_call>" + toolCall.Function.Name)
+					sb.WriteString(renderGLM47ToolArguments(toolCall.Function.Arguments))
+					sb.WriteString("</tool_call>")
+				}
+			}
+		case "tool":
+			if i == 0 || messages[i-1].Role != "tool" {
+				sb.WriteString("<|observation|>")
+			}
+			sb.WriteString("<tool_response>")
+			sb.WriteString(message.Content)
+			sb.WriteString("</tool_response>")
+		case "system":
+			sb.WriteString("<|system|>")
+			sb.WriteString(message.Content)
+		}
+	}
+
+	sb.WriteString("<|assistant|>")
+	if think {
+		sb.WriteString("<think>")
+	} else {
+		sb.WriteString("</think>")
+	}
+
+	return sb.String(), nil
+}
+
+func renderGLM47ToolArguments(args api.ToolCallFunctionArguments) string {
+	var sb strings.Builder
+	for key, value := range args.All() {
+		sb.WriteString("<arg_key>" + key + "</arg_key>")
+		var valueStr string
+		if str, ok := value.(string); ok {
+			valueStr = str
+		} else {
+			jsonBytes, err := json.Marshal(value)
+			if err != nil {
+				valueStr = fmt.Sprintf("%v", value)
+			} else {
+				valueStr = string(jsonBytes)
+			}
+		}
+
+		sb.WriteString("<arg_value>" + valueStr + "</arg_value>")
+	}
+
+	return sb.String()
+}
+
+func formatGLM47ToolJSON(raw []byte) string {
+	var sb strings.Builder
+	sb.Grow(len(raw) + len(raw)/10)
+
+	inString := false
+	escaped := false
+	for i := range raw {
+		ch := raw[i]
+		sb.WriteByte(ch)
+
+		if inString {
+			if escaped {
+				escaped = false
+				continue
+			}
+			if ch == '\\' {
+				escaped = true
+				continue
+			}
+			if ch == '"' {
+				inString = false
+			}
+			continue
+		}
+
+		if ch == '"' {
+			inString = true
+			continue
+		}
+
+		if ch == ':' || ch == ',' {
+			sb.WriteByte(' ')
+		}
+	}
+
+	return sb.String()
+}

model/renderers/glm47_test.go

@@ -0,0 +1,191 @@
+package renderers
+
+import (
+	"testing"
+
+	"github.com/google/go-cmp/cmp"
+	"github.com/ollama/ollama/api"
+)
+
+func TestGLM47Renderer(t *testing.T) {
+	tests := []struct {
+		name       string
+		messages   []api.Message
+		tools      []api.Tool
+		thinkValue *api.ThinkValue
+		expected   string
+	}{
+		{
+			name: "basic user message",
+			messages: []api.Message{
+				{Role: "user", Content: "Hello"},
+			},
+			expected: "[gMASK]<sop><|user|>Hello<|assistant|><think>",
+		},
+		{
+			name: "thinking disabled",
+			messages: []api.Message{
+				{Role: "user", Content: "Hello"},
+			},
+			thinkValue: &api.ThinkValue{Value: false},
+			expected:   "[gMASK]<sop><|user|>Hello<|assistant|></think>",
+		},
+		{
+			name: "system and user",
+			messages: []api.Message{
+				{Role: "system", Content: "You are helpful."},
+				{Role: "user", Content: "Hello"},
+			},
+			expected: "[gMASK]<sop><|system|>You are helpful.<|user|>Hello<|assistant|><think>",
+		},
+		{
+			name: "multi-turn conversation",
+			messages: []api.Message{
+				{Role: "user", Content: "Hi"},
+				{Role: "assistant", Content: "Hello there"},
+				{Role: "user", Content: "How are you?"},
+			},
+			expected: "[gMASK]<sop><|user|>Hi<|assistant|></think>Hello there<|user|>How are you?<|assistant|><think>",
+		},
+		{
+			name: "assistant with reasoning_content",
+			messages: []api.Message{
+				{Role: "user", Content: "Answer with reasoning."},
+				{Role: "assistant", Thinking: "Plan.", Content: "Done."},
+			},
+			expected: "[gMASK]<sop><|user|>Answer with reasoning.<|assistant|><think>Plan.</think>Done.<|assistant|><think>",
+		},
+		{
+			name: "tool call with empty content",
+			messages: []api.Message{
+				{Role: "user", Content: "Weather?"},
+				{
+					Role: "assistant",
+					ToolCalls: []api.ToolCall{
+						{
+							Function: api.ToolCallFunction{
+								Name:      "get_weather",
+								Arguments: args(`{"location": "Tokyo", "unit": "celsius"}`),
+							},
+						},
+					},
+				},
+				{Role: "tool", Content: `{"temperature":22}`},
+			},
+			tools: []api.Tool{
+				{
+					Type: "function",
+					Function: api.ToolFunction{
+						Name:        "get_weather",
+						Description: "Get weather",
+						Parameters: api.ToolFunctionParameters{
+							Type:       "object",
+							Required:   []string{"location"},
+							Properties: propsMap(`{"location": {"type": "string"}}`),
+						},
+					},
+				},
+			},
+			expected: "[gMASK]<sop><|system|>\n# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>\n{\"type\": \"function\", \"function\": {\"name\": \"get_weather\", \"description\": \"Get weather\", \"parameters\": {\"type\": \"object\", \"required\": [\"location\"], \"properties\": {\"location\": {\"type\": \"string\"}}}}}\n</tools>\n\nFor each function call, output the function name and arguments within the following XML format:\n<tool_call>{function-name}<arg_key>{arg-key-1}</arg_key><arg_value>{arg-value-1}</arg_value><arg_key>{arg-key-2}</arg_key><arg_value>{arg-value-2}</arg_value>...</tool_call><|user|>Weather?<|assistant|></think><tool_call>get_weather<arg_key>location</arg_key><arg_value>Tokyo</arg_value><arg_key>unit</arg_key><arg_value>celsius</arg_value></tool_call><|observation|><tool_response>{\"temperature\":22}</tool_response><|assistant|><think>",
+		},
+		{
+			name: "tool call with content",
+			messages: []api.Message{
+				{Role: "user", Content: "Weather?"},
+				{
+					Role:    "assistant",
+					Content: "Let me check",
+					ToolCalls: []api.ToolCall{
+						{
+							Function: api.ToolCallFunction{
+								Name:      "get_weather",
+								Arguments: args(`{"location": "Tokyo"}`),
+							},
+						},
+					},
+				},
+				{Role: "tool", Content: `{"temperature":22}`},
+				{Role: "assistant", Content: "It is 22C."},
+			},
+			tools: []api.Tool{
+				{
+					Type: "function",
+					Function: api.ToolFunction{
+						Name:        "get_weather",
+						Description: "Get weather",
+						Parameters: api.ToolFunctionParameters{
+							Type:       "object",
+							Required:   []string{"location"},
+							Properties: propsMap(`{"location": {"type": "string"}}`),
+						},
+					},
+				},
+			},
+			expected: "[gMASK]<sop><|system|>\n# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>\n{\"type\": \"function\", \"function\": {\"name\": \"get_weather\", \"description\": \"Get weather\", \"parameters\": {\"type\": \"object\", \"required\": [\"location\"], \"properties\": {\"location\": {\"type\": \"string\"}}}}}\n</tools>\n\nFor each function call, output the function name and arguments within the following XML format:\n<tool_call>{function-name}<arg_key>{arg-key-1}</arg_key><arg_value>{arg-value-1}</arg_value><arg_key>{arg-key-2}</arg_key><arg_value>{arg-value-2}</arg_value>...</tool_call><|user|>Weather?<|assistant|></think>Let me check<tool_call>get_weather<arg_key>location</arg_key><arg_value>Tokyo</arg_value></tool_call><|observation|><tool_response>{\"temperature\":22}</tool_response><|assistant|></think>It is 22C.<|assistant|><think>",
+		},
+		{
+			name: "multiple tool calls and responses",
+			messages: []api.Message{
+				{Role: "user", Content: "Compare weather"},
+				{
+					Role: "assistant",
+					ToolCalls: []api.ToolCall{
+						{
+							Function: api.ToolCallFunction{
+								Name:      "get_weather",
+								Arguments: args(`{"location": "Tokyo"}`),
+							},
+						},
+						{
+							Function: api.ToolCallFunction{
+								Name:      "get_weather",
+								Arguments: args(`{"location": "Paris"}`),
+							},
+						},
+					},
+				},
+				{Role: "tool", Content: `{"temperature":22}`},
+				{Role: "tool", Content: `{"temperature":18}`},
+			},
+			tools: []api.Tool{
+				{
+					Type: "function",
+					Function: api.ToolFunction{
+						Name:        "get_weather",
+						Description: "Get weather",
+						Parameters: api.ToolFunctionParameters{
+							Type:       "object",
+							Required:   []string{"location"},
+							Properties: propsMap(`{"location": {"type": "string"}}`),
+						},
+					},
+				},
+			},
+			expected: "[gMASK]<sop><|system|>\n# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>\n{\"type\": \"function\", \"function\": {\"name\": \"get_weather\", \"description\": \"Get weather\", \"parameters\": {\"type\": \"object\", \"required\": [\"location\"], \"properties\": {\"location\": {\"type\": \"string\"}}}}}\n</tools>\n\nFor each function call, output the function name and arguments within the following XML format:\n<tool_call>{function-name}<arg_key>{arg-key-1}</arg_key><arg_value>{arg-value-1}</arg_value><arg_key>{arg-key-2}</arg_key><arg_value>{arg-value-2}</arg_value>...</tool_call><|user|>Compare weather<|assistant|></think><tool_call>get_weather<arg_key>location</arg_key><arg_value>Tokyo</arg_value></tool_call><tool_call>get_weather<arg_key>location</arg_key><arg_value>Paris</arg_value></tool_call><|observation|><tool_response>{\"temperature\":22}</tool_response><tool_response>{\"temperature\":18}</tool_response><|assistant|><think>",
+		},
+		{
+			name: "preserved thinking in multi-turn",
+			messages: []api.Message{
+				{Role: "user", Content: "Think step by step"},
+				{Role: "assistant", Thinking: "Let me think...", Content: "Here's my answer."},
+				{Role: "user", Content: "Continue"},
+			},
+			expected: "[gMASK]<sop><|user|>Think step by step<|assistant|><think>Let me think...</think>Here's my answer.<|user|>Continue<|assistant|><think>",
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			renderer := &GLM47Renderer{}
+			rendered, err := renderer.Render(tt.messages, tt.tools, tt.thinkValue)
+			if err != nil {
+				t.Fatal(err)
+			}
+			if diff := cmp.Diff(rendered, tt.expected); diff != "" {
+				t.Errorf("mismatch (-got +want):\n%s", diff)
+				t.Logf("Got:\n%s", rendered)
+				t.Logf("Expected:\n%s", tt.expected)
+			}
+		})
+	}
+}

model/renderers/renderer.go

@@ -80,6 +80,8 @@ func rendererForName(name string) Renderer {
 		return &Nemotron3NanoRenderer{}
 	case "functiongemma":
 		return &FunctionGemmaRenderer{}
+	case "glm-4.7":
+		return &GLM47Renderer{}
 	default:
 		return nil
 	}

model/renderers/testhelpers_test.go

@@ -1,6 +1,26 @@
 package renderers
 
-import "github.com/ollama/ollama/api"
+import (
+	"encoding/json"
+
+	"github.com/ollama/ollama/api"
+)
+
+func args(s string) api.ToolCallFunctionArguments {
+	var result api.ToolCallFunctionArguments
+	if err := json.Unmarshal([]byte(s), &result); err != nil {
+		panic("invalid JSON in args(): " + err.Error())
+	}
+	return result
+}
+
+func propsMap(s string) *api.ToolPropertiesMap {
+	var result api.ToolPropertiesMap
+	if err := json.Unmarshal([]byte(s), &result); err != nil {
+		panic("invalid JSON in propsMap(): " + err.Error())
+	}
+	return &result
+}
 
 // testPropsMap creates a ToolPropertiesMap from a map (convenience function for tests, order not preserved)
 func testPropsMap(m map[string]api.ToolProperty) *api.ToolPropertiesMap {

openai/openai.go

@@ -630,6 +630,10 @@ func nameFromToolCallID(messages []Message, toolCallID string) string {
 
 // decodeImageURL decodes a base64 data URI into raw image bytes.
 func decodeImageURL(url string) (api.ImageData, error) {
+	if strings.HasPrefix(url, "http://") || strings.HasPrefix(url, "https://") {
+		return nil, errors.New("image URLs are not currently supported, please use base64 encoded data instead")
+	}
+
 	types := []string{"jpeg", "jpg", "png", "webp"}
 
 	// Support blank mime type to match /api/chat's behavior of taking just unadorned base64
@@ -733,3 +737,60 @@ func FromCompleteRequest(r CompletionRequest) (api.GenerateRequest, error) {
 		DebugRenderOnly: r.DebugRenderOnly,
 	}, nil
 }
+
+// ImageGenerationRequest is an OpenAI-compatible image generation request.
+type ImageGenerationRequest struct {
+	Model          string `json:"model"`
+	Prompt         string `json:"prompt"`
+	N              int    `json:"n,omitempty"`
+	Size           string `json:"size,omitempty"`
+	ResponseFormat string `json:"response_format,omitempty"`
+	Seed           *int64 `json:"seed,omitempty"`
+}
+
+// ImageGenerationResponse is an OpenAI-compatible image generation response.
+type ImageGenerationResponse struct {
+	Created int64            `json:"created"`
+	Data    []ImageURLOrData `json:"data"`
+}
+
+// ImageURLOrData contains either a URL or base64-encoded image data.
+type ImageURLOrData struct {
+	URL     string `json:"url,omitempty"`
+	B64JSON string `json:"b64_json,omitempty"`
+}
+
+// FromImageGenerationRequest converts an OpenAI image generation request to an Ollama GenerateRequest.
+func FromImageGenerationRequest(r ImageGenerationRequest) api.GenerateRequest {
+	req := api.GenerateRequest{
+		Model:  r.Model,
+		Prompt: r.Prompt,
+	}
+	// Parse size if provided (e.g., "1024x768")
+	if r.Size != "" {
+		var w, h int32
+		if _, err := fmt.Sscanf(r.Size, "%dx%d", &w, &h); err == nil {
+			req.Width = w
+			req.Height = h
+		}
+	}
+	if r.Seed != nil {
+		if req.Options == nil {
+			req.Options = map[string]any{}
+		}
+		req.Options["seed"] = *r.Seed
+	}
+	return req
+}
+
+// ToImageGenerationResponse converts an Ollama GenerateResponse to an OpenAI ImageGenerationResponse.
+func ToImageGenerationResponse(resp api.GenerateResponse) ImageGenerationResponse {
+	var data []ImageURLOrData
+	if resp.Image != "" {
+		data = []ImageURLOrData{{B64JSON: resp.Image}}
+	}
+	return ImageGenerationResponse{
+		Created: resp.CreatedAt.Unix(),
+		Data:    data,
+	}
+}

openai/responses.go

@@ -4,6 +4,7 @@ import (
 	"encoding/json"
 	"fmt"
 	"math/rand"
+	"time"
 
 	"github.com/ollama/ollama/api"
 )
@@ -265,9 +266,9 @@ type ResponsesText struct {
 type ResponsesTool struct {
 	Type        string         `json:"type"` // "function"
 	Name        string         `json:"name"`
-	Description string         `json:"description,omitempty"`
-	Strict      bool           `json:"strict,omitempty"`
-	Parameters  map[string]any `json:"parameters,omitempty"`
+	Description *string        `json:"description"` // nullable but required
+	Strict      *bool          `json:"strict"`      // nullable but required
+	Parameters  map[string]any `json:"parameters"`  // nullable but required
 }
 
 type ResponsesRequest struct {
@@ -475,11 +476,16 @@ func convertTool(t ResponsesTool) (api.Tool, error) {
 		}
 	}
 
+	var description string
+	if t.Description != nil {
+		description = *t.Description
+	}
+
 	return api.Tool{
 		Type: t.Type,
 		Function: api.ToolFunction{
 			Name:        t.Name,
-			Description: t.Description,
+			Description: description,
 			Parameters:  params,
 		},
 	}, nil
@@ -516,17 +522,60 @@ func convertInputMessage(m ResponsesInputMessage) (api.Message, error) {
 
 // Response types for the Responses API
 
+// ResponsesTextField represents the text output configuration in the response.
+type ResponsesTextField struct {
+	Format ResponsesTextFormat `json:"format"`
+}
+
+// ResponsesReasoningOutput represents reasoning configuration in the response.
+type ResponsesReasoningOutput struct {
+	Effort  *string `json:"effort,omitempty"`
+	Summary *string `json:"summary,omitempty"`
+}
+
+// ResponsesError represents an error in the response.
+type ResponsesError struct {
+	Code    string `json:"code"`
+	Message string `json:"message"`
+}
+
+// ResponsesIncompleteDetails represents details about why a response was incomplete.
+type ResponsesIncompleteDetails struct {
+	Reason string `json:"reason"`
+}
+
 type ResponsesResponse struct {
-	ID        string                `json:"id"`
-	Object    string                `json:"object"`
-	CreatedAt int64                 `json:"created_at"`
-	Status    string                `json:"status"`
-	Model     string                `json:"model"`
-	Output    []ResponsesOutputItem `json:"output"`
-	Usage     *ResponsesUsage       `json:"usage,omitempty"`
-	// TODO(drifkin): add `temperature` and `top_p` to the response, but this
-	// requires additional plumbing to find the effective values since the
-	// defaults can come from the model or the request
+	ID                 string                      `json:"id"`
+	Object             string                      `json:"object"`
+	CreatedAt          int64                       `json:"created_at"`
+	CompletedAt        *int64                      `json:"completed_at"`
+	Status             string                      `json:"status"`
+	IncompleteDetails  *ResponsesIncompleteDetails `json:"incomplete_details"`
+	Model              string                      `json:"model"`
+	PreviousResponseID *string                     `json:"previous_response_id"`
+	Instructions       *string                     `json:"instructions"`
+	Output             []ResponsesOutputItem       `json:"output"`
+	Error              *ResponsesError             `json:"error"`
+	Tools              []ResponsesTool             `json:"tools"`
+	ToolChoice         any                         `json:"tool_choice"`
+	Truncation         string                      `json:"truncation"`
+	ParallelToolCalls  bool                        `json:"parallel_tool_calls"`
+	Text               ResponsesTextField          `json:"text"`
+	TopP               float64                     `json:"top_p"`
+	PresencePenalty    float64                     `json:"presence_penalty"`
+	FrequencyPenalty   float64                     `json:"frequency_penalty"`
+	TopLogprobs        int                         `json:"top_logprobs"`
+	Temperature        float64                     `json:"temperature"`
+	Reasoning          *ResponsesReasoningOutput   `json:"reasoning"`
+	Usage              *ResponsesUsage             `json:"usage"`
+	MaxOutputTokens    *int                        `json:"max_output_tokens"`
+	MaxToolCalls       *int                        `json:"max_tool_calls"`
+	Store              bool                        `json:"store"`
+	Background         bool                        `json:"background"`
+	ServiceTier        string                      `json:"service_tier"`
+	Metadata           map[string]any              `json:"metadata"`
+	SafetyIdentifier   *string                     `json:"safety_identifier"`
+	PromptCacheKey     *string                     `json:"prompt_cache_key"`
 }
 
 type ResponsesOutputItem struct {
@@ -550,18 +599,39 @@ type ResponsesReasoningSummary struct {
 }
 
 type ResponsesOutputContent struct {
-	Type string `json:"type"` // "output_text"
-	Text string `json:"text"`
+	Type        string `json:"type"` // "output_text"
+	Text        string `json:"text"`
+	Annotations []any  `json:"annotations"`
+	Logprobs    []any  `json:"logprobs"`
+}
+
+type ResponsesInputTokensDetails struct {
+	CachedTokens int `json:"cached_tokens"`
+}
+
+type ResponsesOutputTokensDetails struct {
+	ReasoningTokens int `json:"reasoning_tokens"`
 }
 
 type ResponsesUsage struct {
-	InputTokens  int `json:"input_tokens"`
-	OutputTokens int `json:"output_tokens"`
-	TotalTokens  int `json:"total_tokens"`
+	InputTokens         int                          `json:"input_tokens"`
+	OutputTokens        int                          `json:"output_tokens"`
+	TotalTokens         int                          `json:"total_tokens"`
+	InputTokensDetails  ResponsesInputTokensDetails  `json:"input_tokens_details"`
+	OutputTokensDetails ResponsesOutputTokensDetails `json:"output_tokens_details"`
 }
 
-// ToResponse converts an api.ChatResponse to a Responses API response
-func ToResponse(model, responseID, itemID string, chatResponse api.ChatResponse) ResponsesResponse {
+// derefFloat64 returns the value of a float64 pointer, or a default if nil.
+func derefFloat64(p *float64, def float64) float64 {
+	if p != nil {
+		return *p
+	}
+	return def
+}
+
+// ToResponse converts an api.ChatResponse to a Responses API response.
+// The request is used to echo back request parameters in the response.
+func ToResponse(model, responseID, itemID string, chatResponse api.ChatResponse, request ResponsesRequest) ResponsesResponse {
 	var output []ResponsesOutputItem
 
 	// Add reasoning item if thinking is present
@@ -585,6 +655,7 @@ func ToResponse(model, responseID, itemID string, chatResponse api.ChatResponse)
 			output = append(output, ResponsesOutputItem{
 				ID:        fmt.Sprintf("fc_%s_%d", responseID, i),
 				Type:      "function_call",
+				Status:    "completed",
 				CallID:    tc.ID,
 				Name:      tc.Function.Name,
 				Arguments: tc.Function.Arguments,
@@ -598,25 +669,90 @@ func ToResponse(model, responseID, itemID string, chatResponse api.ChatResponse)
 			Role:   "assistant",
 			Content: []ResponsesOutputContent{
 				{
-					Type: "output_text",
-					Text: chatResponse.Message.Content,
+					Type:        "output_text",
+					Text:        chatResponse.Message.Content,
+					Annotations: []any{},
+					Logprobs:    []any{},
 				},
 			},
 		})
 	}
 
+	var instructions *string
+	if request.Instructions != "" {
+		instructions = &request.Instructions
+	}
+
+	// Build truncation with default
+	truncation := "disabled"
+	if request.Truncation != nil {
+		truncation = *request.Truncation
+	}
+
+	tools := request.Tools
+	if tools == nil {
+		tools = []ResponsesTool{}
+	}
+
+	text := ResponsesTextField{
+		Format: ResponsesTextFormat{Type: "text"},
+	}
+	if request.Text != nil && request.Text.Format != nil {
+		text.Format = *request.Text.Format
+	}
+
+	// Build reasoning output from request
+	var reasoning *ResponsesReasoningOutput
+	if request.Reasoning.Effort != "" || request.Reasoning.Summary != "" {
+		reasoning = &ResponsesReasoningOutput{}
+		if request.Reasoning.Effort != "" {
+			reasoning.Effort = &request.Reasoning.Effort
+		}
+		if request.Reasoning.Summary != "" {
+			reasoning.Summary = &request.Reasoning.Summary
+		}
+	}
+
 	return ResponsesResponse{
-		ID:        responseID,
-		Object:    "response",
-		CreatedAt: chatResponse.CreatedAt.Unix(),
-		Status:    "completed",
-		Model:     model,
-		Output:    output,
+		ID:                 responseID,
+		Object:             "response",
+		CreatedAt:          chatResponse.CreatedAt.Unix(),
+		CompletedAt:        nil, // Set by middleware when writing final response
+		Status:             "completed",
+		IncompleteDetails:  nil, // Only populated if response incomplete
+		Model:              model,
+		PreviousResponseID: nil, // Not supported
+		Instructions:       instructions,
+		Output:             output,
+		Error:              nil, // Only populated on failure
+		Tools:              tools,
+		ToolChoice:         "auto", // Default value
+		Truncation:         truncation,
+		ParallelToolCalls:  true, // Default value
+		Text:               text,
+		TopP:               derefFloat64(request.TopP, 1.0),
+		PresencePenalty:    0, // Default value
+		FrequencyPenalty:   0, // Default value
+		TopLogprobs:        0, // Default value
+		Temperature:        derefFloat64(request.Temperature, 1.0),
+		Reasoning:          reasoning,
 		Usage: &ResponsesUsage{
 			InputTokens:  chatResponse.PromptEvalCount,
 			OutputTokens: chatResponse.EvalCount,
 			TotalTokens:  chatResponse.PromptEvalCount + chatResponse.EvalCount,
+			// TODO(drifkin): wire through the actual values
+			InputTokensDetails: ResponsesInputTokensDetails{CachedTokens: 0},
+			// TODO(drifkin): wire through the actual values
+			OutputTokensDetails: ResponsesOutputTokensDetails{ReasoningTokens: 0},
 		},
+		MaxOutputTokens:  request.MaxOutputTokens,
+		MaxToolCalls:     nil,   // Not supported
+		Store:            false, // We don't store responses
+		Background:       request.Background,
+		ServiceTier:      "default", // Default value
+		Metadata:         map[string]any{},
+		SafetyIdentifier: nil, // Not supported
+		PromptCacheKey:   nil, // Not supported
 	}
 }
 
@@ -636,6 +772,7 @@ type ResponsesStreamConverter struct {
 	responseID string
 	itemID     string
 	model      string
+	request    ResponsesRequest
 
 	// State tracking (mutated across Process calls)
 	firstWrite      bool
@@ -668,11 +805,12 @@ func (c *ResponsesStreamConverter) newEvent(eventType string, data map[string]an
 }
 
 // NewResponsesStreamConverter creates a new converter with the given configuration.
-func NewResponsesStreamConverter(responseID, itemID, model string) *ResponsesStreamConverter {
+func NewResponsesStreamConverter(responseID, itemID, model string, request ResponsesRequest) *ResponsesStreamConverter {
 	return &ResponsesStreamConverter{
 		responseID: responseID,
 		itemID:     itemID,
 		model:      model,
+		request:    request,
 		firstWrite: true,
 	}
 }
@@ -717,25 +855,120 @@ func (c *ResponsesStreamConverter) Process(r api.ChatResponse) []ResponsesStream
 	return events
 }
 
+// buildResponseObject creates a full response object with all required fields for streaming events.
+func (c *ResponsesStreamConverter) buildResponseObject(status string, output []any, usage map[string]any) map[string]any {
+	var instructions any = nil
+	if c.request.Instructions != "" {
+		instructions = c.request.Instructions
+	}
+
+	truncation := "disabled"
+	if c.request.Truncation != nil {
+		truncation = *c.request.Truncation
+	}
+
+	var tools []any
+	if c.request.Tools != nil {
+		for _, t := range c.request.Tools {
+			tools = append(tools, map[string]any{
+				"type":        t.Type,
+				"name":        t.Name,
+				"description": t.Description,
+				"strict":      t.Strict,
+				"parameters":  t.Parameters,
+			})
+		}
+	}
+	if tools == nil {
+		tools = []any{}
+	}
+
+	textFormat := map[string]any{"type": "text"}
+	if c.request.Text != nil && c.request.Text.Format != nil {
+		textFormat = map[string]any{
+			"type": c.request.Text.Format.Type,
+		}
+		if c.request.Text.Format.Name != "" {
+			textFormat["name"] = c.request.Text.Format.Name
+		}
+		if c.request.Text.Format.Schema != nil {
+			textFormat["schema"] = c.request.Text.Format.Schema
+		}
+		if c.request.Text.Format.Strict != nil {
+			textFormat["strict"] = *c.request.Text.Format.Strict
+		}
+	}
+
+	var reasoning any = nil
+	if c.request.Reasoning.Effort != "" || c.request.Reasoning.Summary != "" {
+		r := map[string]any{}
+		if c.request.Reasoning.Effort != "" {
+			r["effort"] = c.request.Reasoning.Effort
+		} else {
+			r["effort"] = nil
+		}
+		if c.request.Reasoning.Summary != "" {
+			r["summary"] = c.request.Reasoning.Summary
+		} else {
+			r["summary"] = nil
+		}
+		reasoning = r
+	}
+
+	// Build top_p and temperature with defaults
+	topP := 1.0
+	if c.request.TopP != nil {
+		topP = *c.request.TopP
+	}
+	temperature := 1.0
+	if c.request.Temperature != nil {
+		temperature = *c.request.Temperature
+	}
+
+	return map[string]any{
+		"id":                   c.responseID,
+		"object":               "response",
+		"created_at":           time.Now().Unix(),
+		"completed_at":         nil,
+		"status":               status,
+		"incomplete_details":   nil,
+		"model":                c.model,
+		"previous_response_id": nil,
+		"instructions":         instructions,
+		"output":               output,
+		"error":                nil,
+		"tools":                tools,
+		"tool_choice":          "auto",
+		"truncation":           truncation,
+		"parallel_tool_calls":  true,
+		"text":                 map[string]any{"format": textFormat},
+		"top_p":                topP,
+		"presence_penalty":     0,
+		"frequency_penalty":    0,
+		"top_logprobs":         0,
+		"temperature":          temperature,
+		"reasoning":            reasoning,
+		"usage":                usage,
+		"max_output_tokens":    c.request.MaxOutputTokens,
+		"max_tool_calls":       nil,
+		"store":                false,
+		"background":           c.request.Background,
+		"service_tier":         "default",
+		"metadata":             map[string]any{},
+		"safety_identifier":    nil,
+		"prompt_cache_key":     nil,
+	}
+}
+
 func (c *ResponsesStreamConverter) createResponseCreatedEvent() ResponsesStreamEvent {
 	return c.newEvent("response.created", map[string]any{
-		"response": map[string]any{
-			"id":     c.responseID,
-			"object": "response",
-			"status": "in_progress",
-			"output": []any{},
-		},
+		"response": c.buildResponseObject("in_progress", []any{}, nil),
 	})
 }
 
 func (c *ResponsesStreamConverter) createResponseInProgressEvent() ResponsesStreamEvent {
 	return c.newEvent("response.in_progress", map[string]any{
-		"response": map[string]any{
-			"id":     c.responseID,
-			"object": "response",
-			"status": "in_progress",
-			"output": []any{},
-		},
+		"response": c.buildResponseObject("in_progress", []any{}, nil),
 	})
 }
 
@@ -762,9 +995,10 @@ func (c *ResponsesStreamConverter) processThinking(thinking string) []ResponsesS
 
 	// Emit delta
 	events = append(events, c.newEvent("response.reasoning_summary_text.delta", map[string]any{
-		"item_id":      c.reasoningItemID,
-		"output_index": c.outputIndex,
-		"delta":        thinking,
+		"item_id":       c.reasoningItemID,
+		"output_index":  c.outputIndex,
+		"summary_index": 0,
+		"delta":         thinking,
 	}))
 
 	// TODO(drifkin): consider adding
@@ -783,9 +1017,10 @@ func (c *ResponsesStreamConverter) finishReasoning() []ResponsesStreamEvent {
 
 	events := []ResponsesStreamEvent{
 		c.newEvent("response.reasoning_summary_text.done", map[string]any{
-			"item_id":      c.reasoningItemID,
-			"output_index": c.outputIndex,
-			"text":         c.accumulatedThinking,
+			"item_id":       c.reasoningItemID,
+			"output_index":  c.outputIndex,
+			"summary_index": 0,
+			"text":          c.accumulatedThinking,
 		}),
 		c.newEvent("response.output_item.done", map[string]any{
 			"output_index": c.outputIndex,
@@ -898,8 +1133,10 @@ func (c *ResponsesStreamConverter) processTextContent(content string) []Response
 			"output_index":  c.outputIndex,
 			"content_index": c.contentIndex,
 			"part": map[string]any{
-				"type": "output_text",
-				"text": "",
+				"type":        "output_text",
+				"text":        "",
+				"annotations": []any{},
+				"logprobs":    []any{},
 			},
 		}))
 	}
@@ -913,6 +1150,7 @@ func (c *ResponsesStreamConverter) processTextContent(content string) []Response
 		"output_index":  c.outputIndex,
 		"content_index": 0,
 		"delta":         content,
+		"logprobs":      []any{},
 	}))
 
 	return events
@@ -944,8 +1182,10 @@ func (c *ResponsesStreamConverter) buildFinalOutput() []any {
 			"status": "completed",
 			"role":   "assistant",
 			"content": []map[string]any{{
-				"type": "output_text",
-				"text": c.accumulatedText,
+				"type":        "output_text",
+				"text":        c.accumulatedText,
+				"annotations": []any{},
+				"logprobs":    []any{},
 			}},
 		})
 	}
@@ -967,6 +1207,7 @@ func (c *ResponsesStreamConverter) processCompletion(r api.ChatResponse) []Respo
 			"output_index":  c.outputIndex,
 			"content_index": 0,
 			"text":          c.accumulatedText,
+			"logprobs":      []any{},
 		}))
 
 		// response.content_part.done
@@ -975,8 +1216,10 @@ func (c *ResponsesStreamConverter) processCompletion(r api.ChatResponse) []Respo
 			"output_index":  c.outputIndex,
 			"content_index": 0,
 			"part": map[string]any{
-				"type": "output_text",
-				"text": c.accumulatedText,
+				"type":        "output_text",
+				"text":        c.accumulatedText,
+				"annotations": []any{},
+				"logprobs":    []any{},
 			},
 		}))
 
@@ -989,26 +1232,31 @@ func (c *ResponsesStreamConverter) processCompletion(r api.ChatResponse) []Respo
 				"status": "completed",
 				"role":   "assistant",
 				"content": []map[string]any{{
-					"type": "output_text",
-					"text": c.accumulatedText,
+					"type":        "output_text",
+					"text":        c.accumulatedText,
+					"annotations": []any{},
+					"logprobs":    []any{},
 				}},
 			},
 		}))
 	}
 
 	// response.completed
-	events = append(events, c.newEvent("response.completed", map[string]any{
-		"response": map[string]any{
-			"id":     c.responseID,
-			"object": "response",
-			"status": "completed",
-			"output": c.buildFinalOutput(),
-			"usage": map[string]any{
-				"input_tokens":  r.PromptEvalCount,
-				"output_tokens": r.EvalCount,
-				"total_tokens":  r.PromptEvalCount + r.EvalCount,
-			},
+	usage := map[string]any{
+		"input_tokens":  r.PromptEvalCount,
+		"output_tokens": r.EvalCount,
+		"total_tokens":  r.PromptEvalCount + r.EvalCount,
+		"input_tokens_details": map[string]any{
+			"cached_tokens": 0,
 		},
+		"output_tokens_details": map[string]any{
+			"reasoning_tokens": 0,
+		},
+	}
+	response := c.buildResponseObject("completed", c.buildFinalOutput(), usage)
+	response["completed_at"] = time.Now().Unix()
+	events = append(events, c.newEvent("response.completed", map[string]any{
+		"response": response,
 	}))
 
 	return events

openai/responses_test.go

@@ -850,7 +850,7 @@ func TestFromResponsesRequest_Images(t *testing.T) {
 }
 
 func TestResponsesStreamConverter_TextOnly(t *testing.T) {
-	converter := NewResponsesStreamConverter("resp_123", "msg_456", "gpt-oss:20b")
+	converter := NewResponsesStreamConverter("resp_123", "msg_456", "gpt-oss:20b", ResponsesRequest{})
 
 	// First chunk with content
 	events := converter.Process(api.ChatResponse{
@@ -916,7 +916,7 @@ func TestResponsesStreamConverter_TextOnly(t *testing.T) {
 }
 
 func TestResponsesStreamConverter_ToolCalls(t *testing.T) {
-	converter := NewResponsesStreamConverter("resp_123", "msg_456", "gpt-oss:20b")
+	converter := NewResponsesStreamConverter("resp_123", "msg_456", "gpt-oss:20b", ResponsesRequest{})
 
 	events := converter.Process(api.ChatResponse{
 		Message: api.Message{
@@ -952,7 +952,7 @@ func TestResponsesStreamConverter_ToolCalls(t *testing.T) {
 }
 
 func TestResponsesStreamConverter_Reasoning(t *testing.T) {
-	converter := NewResponsesStreamConverter("resp_123", "msg_456", "gpt-oss:20b")
+	converter := NewResponsesStreamConverter("resp_123", "msg_456", "gpt-oss:20b", ResponsesRequest{})
 
 	// First chunk with thinking
 	events := converter.Process(api.ChatResponse{
@@ -1267,7 +1267,7 @@ func TestToResponse_WithReasoning(t *testing.T) {
 			Content:  "The answer is 42",
 		},
 		Done: true,
-	})
+	}, ResponsesRequest{})
 
 	// Should have 2 output items: reasoning + message
 	if len(response.Output) != 2 {
@@ -1638,7 +1638,7 @@ func TestFromResponsesRequest_ShorthandFormats(t *testing.T) {
 
 func TestResponsesStreamConverter_OutputIncludesContent(t *testing.T) {
 	// Verify that response.output_item.done includes content field for messages
-	converter := NewResponsesStreamConverter("resp_123", "msg_456", "gpt-oss:20b")
+	converter := NewResponsesStreamConverter("resp_123", "msg_456", "gpt-oss:20b", ResponsesRequest{})
 
 	// First chunk
 	converter.Process(api.ChatResponse{
@@ -1686,7 +1686,7 @@ func TestResponsesStreamConverter_OutputIncludesContent(t *testing.T) {
 
 func TestResponsesStreamConverter_ResponseCompletedIncludesOutput(t *testing.T) {
 	// Verify that response.completed includes the output array
-	converter := NewResponsesStreamConverter("resp_123", "msg_456", "gpt-oss:20b")
+	converter := NewResponsesStreamConverter("resp_123", "msg_456", "gpt-oss:20b", ResponsesRequest{})
 
 	// Process some content
 	converter.Process(api.ChatResponse{
@@ -1730,7 +1730,7 @@ func TestResponsesStreamConverter_ResponseCompletedIncludesOutput(t *testing.T)
 
 func TestResponsesStreamConverter_ResponseCreatedIncludesOutput(t *testing.T) {
 	// Verify that response.created includes an empty output array
-	converter := NewResponsesStreamConverter("resp_123", "msg_456", "gpt-oss:20b")
+	converter := NewResponsesStreamConverter("resp_123", "msg_456", "gpt-oss:20b", ResponsesRequest{})
 
 	events := converter.Process(api.ChatResponse{
 		Message: api.Message{Content: "Hi"},
@@ -1757,7 +1757,7 @@ func TestResponsesStreamConverter_ResponseCreatedIncludesOutput(t *testing.T) {
 
 func TestResponsesStreamConverter_SequenceNumbers(t *testing.T) {
 	// Verify that events include incrementing sequence numbers
-	converter := NewResponsesStreamConverter("resp_123", "msg_456", "gpt-oss:20b")
+	converter := NewResponsesStreamConverter("resp_123", "msg_456", "gpt-oss:20b", ResponsesRequest{})
 
 	events := converter.Process(api.ChatResponse{
 		Message: api.Message{Content: "Hello"},
@@ -1791,7 +1791,7 @@ func TestResponsesStreamConverter_SequenceNumbers(t *testing.T) {
 
 func TestResponsesStreamConverter_FunctionCallStatus(t *testing.T) {
 	// Verify that function call items include status field
-	converter := NewResponsesStreamConverter("resp_123", "msg_456", "gpt-oss:20b")
+	converter := NewResponsesStreamConverter("resp_123", "msg_456", "gpt-oss:20b", ResponsesRequest{})
 
 	events := converter.Process(api.ChatResponse{
 		Message: api.Message{

parser/parser_test.go

@@ -21,6 +21,7 @@ import (
 	"golang.org/x/text/encoding/unicode"
 
 	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/convert"
 	"github.com/ollama/ollama/fs/ggml"
 )
 
@@ -801,7 +802,7 @@ func createBinFile(t *testing.T, kv map[string]any, ti []*ggml.Tensor) (string,
 	}
 	defer f.Close()
 
-	base := map[string]any{"general.architecture": "test"}
+	var base convert.KV = map[string]any{"general.architecture": "test"}
 	maps.Copy(base, kv)
 
 	if err := ggml.WriteGGUF(f, base, ti); err != nil {

progress/stepbar.go

@@ -0,0 +1,33 @@
+package progress
+
+import (
+	"fmt"
+	"strings"
+)
+
+// StepBar displays step-based progress (e.g., for image generation steps).
+type StepBar struct {
+	message string
+	current int
+	total   int
+}
+
+func NewStepBar(message string, total int) *StepBar {
+	return &StepBar{message: message, total: total}
+}
+
+func (s *StepBar) Set(current int) {
+	s.current = current
+}
+
+func (s *StepBar) String() string {
+	percent := float64(s.current) / float64(s.total) * 100
+	barWidth := s.total
+	empty := barWidth - s.current
+
+	// "Generating   0% ▕         ▏ 0/9"
+	return fmt.Sprintf("%s %3.0f%% ▕%s%s▏ %d/%d",
+		s.message, percent,
+		strings.Repeat("█", s.current), strings.Repeat(" ", empty),
+		s.current, s.total)
+}

readline/errors.go

@@ -6,9 +6,6 @@ import (
 
 var ErrInterrupt = errors.New("Interrupt")
 
-// ErrExpandOutput is returned when user presses Ctrl+O to expand tool output
-var ErrExpandOutput = errors.New("ExpandOutput")
-
 type InterruptError struct {
 	Line []rune
 }

readline/readline.go

@@ -5,6 +5,7 @@ import (
 	"fmt"
 	"io"
 	"os"
+	"strings"
 )
 
 type Prompt struct {
@@ -36,10 +37,11 @@ type Terminal struct {
 }
 
 type Instance struct {
-	Prompt   *Prompt
-	Terminal *Terminal
-	History  *History
-	Pasting  bool
+	Prompt      *Prompt
+	Terminal    *Terminal
+	History     *History
+	Pasting     bool
+	pastedLines []string
 }
 
 func New(prompt Prompt) (*Instance, error) {
@@ -174,6 +176,8 @@ func (i *Instance) Readline() (string, error) {
 		case CharEsc:
 			esc = true
 		case CharInterrupt:
+			i.pastedLines = nil
+			i.Prompt.UseAlt = false
 			return "", ErrInterrupt
 		case CharPrev:
 			i.historyPrev(buf, &currentLineBuf)
@@ -188,7 +192,23 @@ func (i *Instance) Readline() (string, error) {
 		case CharForward:
 			buf.MoveRight()
 		case CharBackspace, CharCtrlH:
-			buf.Remove()
+			if buf.IsEmpty() && len(i.pastedLines) > 0 {
+				lastIdx := len(i.pastedLines) - 1
+				prevLine := i.pastedLines[lastIdx]
+				i.pastedLines = i.pastedLines[:lastIdx]
+				fmt.Print(CursorBOL + ClearToEOL + CursorUp + CursorBOL + ClearToEOL)
+				if len(i.pastedLines) == 0 {
+					fmt.Print(i.Prompt.Prompt)
+					i.Prompt.UseAlt = false
+				} else {
+					fmt.Print(i.Prompt.AltPrompt)
+				}
+				for _, r := range prevLine {
+					buf.Add(r)
+				}
+			} else {
+				buf.Remove()
+			}
 		case CharTab:
 			// todo: convert back to real tabs
 			for range 8 {
@@ -206,21 +226,33 @@ func (i *Instance) Readline() (string, error) {
 			buf.DeleteBefore()
 		case CharCtrlL:
 			buf.ClearScreen()
-		case CharCtrlO:
-			// Ctrl+O - expand tool output
-			return "", ErrExpandOutput
 		case CharCtrlW:
 			buf.DeleteWord()
 		case CharCtrlZ:
 			fd := os.Stdin.Fd()
 			return handleCharCtrlZ(fd, i.Terminal.termios)
-		case CharEnter, CharCtrlJ:
+		case CharCtrlJ:
+			i.pastedLines = append(i.pastedLines, buf.String())
+			buf.Buf.Clear()
+			buf.Pos = 0
+			buf.DisplayPos = 0
+			buf.LineHasSpace.Clear()
+			fmt.Println()
+			fmt.Print(i.Prompt.AltPrompt)
+			i.Prompt.UseAlt = true
+			continue
+		case CharEnter:
 			output := buf.String()
+			if len(i.pastedLines) > 0 {
+				output = strings.Join(i.pastedLines, "\n") + "\n" + output
+				i.pastedLines = nil
+			}
 			if output != "" {
 				i.History.Add(output)
 			}
 			buf.MoveToEnd()
 			fmt.Println()
+			i.Prompt.UseAlt = false
 
 			return output, nil
 		default:

runner/runner.go

@@ -3,6 +3,7 @@ package runner
 import (
 	"github.com/ollama/ollama/runner/llamarunner"
 	"github.com/ollama/ollama/runner/ollamarunner"
+	imagerunner "github.com/ollama/ollama/x/imagegen/runner"
 )
 
 func Execute(args []string) error {
@@ -11,12 +12,19 @@ func Execute(args []string) error {
 	}
 
 	var newRunner bool
-	if args[0] == "--ollama-engine" {
+	var imageRunner bool
+	if len(args) > 0 && args[0] == "--ollama-engine" {
 		args = args[1:]
 		newRunner = true
 	}
+	if len(args) > 0 && args[0] == "--image-engine" {
+		args = args[1:]
+		imageRunner = true
+	}
 
-	if newRunner {
+	if imageRunner {
+		return imagerunner.Execute(args)
+	} else if newRunner {
 		return ollamarunner.Execute(args)
 	} else {
 		return llamarunner.Execute(args)

scripts/build_darwin.sh

@@ -42,18 +42,41 @@ shift $(( $OPTIND - 1 ))
 _build_darwin() {
     for ARCH in $ARCHS; do
         status "Building darwin $ARCH"
-        INSTALL_PREFIX=dist/darwin-$ARCH/
-        GOOS=darwin GOARCH=$ARCH CGO_ENABLED=1 go build -o $INSTALL_PREFIX .
+        INSTALL_PREFIX=dist/darwin-$ARCH/        
 
         if [ "$ARCH" = "amd64" ]; then
             status "Building darwin $ARCH dynamic backends"
-            cmake -B build/darwin-$ARCH \
+            BUILD_DIR=build/darwin-$ARCH
+            cmake -B $BUILD_DIR \
                 -DCMAKE_OSX_ARCHITECTURES=x86_64 \
-                -DCMAKE_OSX_DEPLOYMENT_TARGET=11.3 \
+                -DCMAKE_OSX_DEPLOYMENT_TARGET=14.0 \
+                -DCMAKE_INSTALL_PREFIX=$INSTALL_PREFIX \
+                -DMLX_ENGINE=ON \
+                -DMLX_ENABLE_X64_MAC=ON \
+                -DOLLAMA_RUNNER_DIR=./
+            cmake --build $BUILD_DIR --target ggml-cpu -j
+            cmake --build $BUILD_DIR --target mlx mlxc -j
+            cmake --install $BUILD_DIR --component CPU
+            cmake --install $BUILD_DIR --component MLX
+            # Override CGO flags to point to the amd64 build directory
+            MLX_CGO_CFLAGS="-O3 -I$(pwd)/$BUILD_DIR/_deps/mlx-c-src -mmacosx-version-min=14.0"
+            MLX_CGO_LDFLAGS="-ldl -lc++ -framework Accelerate -mmacosx-version-min=14.0"
+        else
+            BUILD_DIR=build
+            cmake --preset MLX \
+                -DOLLAMA_RUNNER_DIR=./ \
+                -DCMAKE_OSX_DEPLOYMENT_TARGET=14.0 \
                 -DCMAKE_INSTALL_PREFIX=$INSTALL_PREFIX
-            cmake --build build/darwin-$ARCH --target ggml-cpu -j
-            cmake --install build/darwin-$ARCH --component CPU
+            cmake --build --preset MLX --parallel
+            cmake --install $BUILD_DIR --component MLX
+            # Use default CGO flags from mlx.go for arm64
+            MLX_CGO_CFLAGS="-O3 -I$(pwd)/$BUILD_DIR/_deps/mlx-c-src -mmacosx-version-min=14.0"
+            MLX_CGO_LDFLAGS="-lc++ -framework Metal -framework Foundation -framework Accelerate -mmacosx-version-min=14.0"
         fi
+        GOOS=darwin GOARCH=$ARCH CGO_ENABLED=1 CGO_CFLAGS="$MLX_CGO_CFLAGS" CGO_LDFLAGS="$MLX_CGO_LDFLAGS" go build -tags mlx -o $INSTALL_PREFIX .
+        # Copy MLX libraries to same directory as executable for dlopen
+        cp $INSTALL_PREFIX/lib/ollama/libmlxc.dylib $INSTALL_PREFIX/
+        cp $INSTALL_PREFIX/lib/ollama/libmlx.dylib $INSTALL_PREFIX/
     done
 }
 
@@ -64,14 +87,14 @@ _sign_darwin() {
     chmod +x dist/darwin/ollama
 
     if [ -n "$APPLE_IDENTITY" ]; then
-        for F in dist/darwin/ollama dist/darwin-amd64/lib/ollama/*; do
+        for F in dist/darwin/ollama dist/darwin-*/lib/ollama/*; do
             codesign -f --timestamp -s "$APPLE_IDENTITY" --identifier ai.ollama.ollama --options=runtime $F
         done
 
         # create a temporary zip for notarization
         TEMP=$(mktemp -u).zip
         ditto -c -k --keepParent dist/darwin/ollama "$TEMP"
-        xcrun notarytool submit "$TEMP" --wait --timeout 10m --apple-id $APPLE_ID --password $APPLE_PASSWORD --team-id $APPLE_TEAM_ID
+        xcrun notarytool submit "$TEMP" --wait --timeout 20m --apple-id $APPLE_ID --password $APPLE_PASSWORD --team-id $APPLE_TEAM_ID
         rm -f "$TEMP"
     fi
 
@@ -131,7 +154,13 @@ _build_macapp() {
     mkdir -p dist/Ollama.app/Contents/Resources
     if [ -d dist/darwin-amd64 ]; then
         lipo -create -output dist/Ollama.app/Contents/Resources/ollama dist/darwin-amd64/ollama dist/darwin-arm64/ollama
-        cp dist/darwin-amd64/lib/ollama/*.so dist/darwin-amd64/lib/ollama/*.dylib dist/Ollama.app/Contents/Resources/
+        for F in dist/darwin-amd64/lib/ollama/*mlx*.dylib ; do
+            lipo -create -output dist/darwin/$(basename $F) $F dist/darwin-arm64/lib/ollama/$(basename $F)
+        done
+        cp dist/darwin-*/lib/ollama/*.so dist/darwin-*/lib/ollama/*.dylib dist/Ollama.app/Contents/Resources/
+        cp dist/darwin/*.dylib dist/Ollama.app/Contents/Resources/
+        # Copy MLX metallib (architecture-independent, just use arm64 version)
+        cp dist/darwin-arm64/lib/ollama/*.metallib dist/Ollama.app/Contents/Resources/ 2>/dev/null || true
     else
         cp -a dist/darwin/ollama dist/Ollama.app/Contents/Resources/ollama
         cp dist/darwin/*.so dist/darwin/*.dylib dist/Ollama.app/Contents/Resources/
@@ -141,22 +170,22 @@ _build_macapp() {
     # Sign
     if [ -n "$APPLE_IDENTITY" ]; then
         codesign -f --timestamp -s "$APPLE_IDENTITY" --identifier ai.ollama.ollama --options=runtime dist/Ollama.app/Contents/Resources/ollama
-        for lib in dist/Ollama.app/Contents/Resources/*.so dist/Ollama.app/Contents/Resources/*.dylib ; do
+        for lib in dist/Ollama.app/Contents/Resources/*.so dist/Ollama.app/Contents/Resources/*.dylib dist/Ollama.app/Contents/Resources/*.metallib ; do
             codesign -f --timestamp -s "$APPLE_IDENTITY" --identifier ai.ollama.ollama --options=runtime ${lib}
         done
         codesign -f --timestamp -s "$APPLE_IDENTITY" --identifier com.electron.ollama --deep --options=runtime dist/Ollama.app
     fi
 
     rm -f dist/Ollama-darwin.zip
-    ditto -c -k --keepParent dist/Ollama.app dist/Ollama-darwin.zip
-    (cd dist/Ollama.app/Contents/Resources/; tar -cf - ollama *.so *.dylib) | gzip -9vc > dist/ollama-darwin.tgz
+    ditto -c -k --norsrc --keepParent dist/Ollama.app dist/Ollama-darwin.zip
+    (cd dist/Ollama.app/Contents/Resources/; tar -cf - ollama *.so *.dylib *.metallib 2>/dev/null) | gzip -9vc > dist/ollama-darwin.tgz
 
     # Notarize and Staple
     if [ -n "$APPLE_IDENTITY" ]; then
-        $(xcrun -f notarytool) submit dist/Ollama-darwin.zip --wait --timeout 10m --apple-id "$APPLE_ID" --password "$APPLE_PASSWORD" --team-id "$APPLE_TEAM_ID"
+        $(xcrun -f notarytool) submit dist/Ollama-darwin.zip --wait --timeout 20m --apple-id "$APPLE_ID" --password "$APPLE_PASSWORD" --team-id "$APPLE_TEAM_ID"
         rm -f dist/Ollama-darwin.zip
         $(xcrun -f stapler) staple dist/Ollama.app
-        ditto -c -k --keepParent dist/Ollama.app dist/Ollama-darwin.zip
+        ditto -c -k --norsrc --keepParent dist/Ollama.app dist/Ollama-darwin.zip
 
         rm -f dist/Ollama.dmg
 
@@ -177,7 +206,7 @@ _build_macapp() {
         rm -f dist/rw*.dmg
 
         codesign -f --timestamp -s "$APPLE_IDENTITY" --identifier ai.ollama.ollama --options=runtime dist/Ollama.dmg
-        $(xcrun -f notarytool) submit dist/Ollama.dmg --wait --timeout 10m --apple-id "$APPLE_ID" --password "$APPLE_PASSWORD" --team-id "$APPLE_TEAM_ID"
+        $(xcrun -f notarytool) submit dist/Ollama.dmg --wait --timeout 20m --apple-id "$APPLE_ID" --password "$APPLE_PASSWORD" --team-id "$APPLE_TEAM_ID"
         $(xcrun -f stapler) staple dist/Ollama.dmg
     else
         echo "WARNING: Code signing disabled, this bundle will not work for upgrade testing"

scripts/build_linux.sh

@@ -12,6 +12,17 @@ set -eu
 
 . $(dirname $0)/env.sh
 
+# Check for required tools
+if ! command -v zstd >/dev/null 2>&1; then
+    echo "ERROR: zstd is required but not installed." >&2
+    echo "Please install zstd:" >&2
+    echo "  - macOS: brew install zstd" >&2
+    echo "  - Debian/Ubuntu: sudo apt-get install zstd" >&2
+    echo "  - RHEL/CentOS/Fedora: sudo dnf install zstd" >&2
+    echo "  - Arch: sudo pacman -S zstd" >&2
+    exit 1
+fi
+
 mkdir -p dist
 
 docker buildx build \
@@ -37,19 +48,27 @@ if echo $PLATFORM | grep "amd64" > /dev/null; then
         .
 fi
 
+# Run deduplication for each platform output directory
+if echo $PLATFORM | grep "," > /dev/null ; then
+    $(dirname $0)/deduplicate_cuda_libs.sh "./dist/linux_amd64"
+    $(dirname $0)/deduplicate_cuda_libs.sh "./dist/linux_arm64"
+elif echo $PLATFORM | grep "amd64\|arm64" > /dev/null ; then
+    $(dirname $0)/deduplicate_cuda_libs.sh "./dist"
+fi
+
 # buildx behavior changes for single vs. multiplatform
 echo "Compressing linux tar bundles..."
 if echo $PLATFORM | grep "," > /dev/null ; then
-        tar c -C ./dist/linux_arm64 --exclude cuda_jetpack5 --exclude cuda_jetpack6 . | pigz -9vc >./dist/ollama-linux-arm64.tgz
-        tar c -C ./dist/linux_arm64 ./lib/ollama/cuda_jetpack5  | pigz -9vc >./dist/ollama-linux-arm64-jetpack5.tgz
-        tar c -C ./dist/linux_arm64 ./lib/ollama/cuda_jetpack6  | pigz -9vc >./dist/ollama-linux-arm64-jetpack6.tgz
-        tar c -C ./dist/linux_amd64 --exclude rocm . | pigz -9vc >./dist/ollama-linux-amd64.tgz
-        tar c -C ./dist/linux_amd64 ./lib/ollama/rocm  | pigz -9vc >./dist/ollama-linux-amd64-rocm.tgz
+        tar c -C ./dist/linux_arm64 --exclude cuda_jetpack5 --exclude cuda_jetpack6 . | zstd --ultra -22 -T0 >./dist/ollama-linux-arm64.tar.zst
+        tar c -C ./dist/linux_arm64 ./lib/ollama/cuda_jetpack5  | zstd --ultra -22 -T0 >./dist/ollama-linux-arm64-jetpack5.tar.zst
+        tar c -C ./dist/linux_arm64 ./lib/ollama/cuda_jetpack6  | zstd --ultra -22 -T0 >./dist/ollama-linux-arm64-jetpack6.tar.zst
+        tar c -C ./dist/linux_amd64 --exclude rocm . | zstd --ultra -22 -T0 >./dist/ollama-linux-amd64.tar.zst
+        tar c -C ./dist/linux_amd64 ./lib/ollama/rocm  | zstd --ultra -22 -T0 >./dist/ollama-linux-amd64-rocm.tar.zst
 elif echo $PLATFORM | grep "arm64" > /dev/null ; then
-        tar c -C ./dist/ --exclude cuda_jetpack5 --exclude cuda_jetpack6 bin lib | pigz -9vc >./dist/ollama-linux-arm64.tgz
-        tar c -C ./dist/ ./lib/ollama/cuda_jetpack5  | pigz -9vc >./dist/ollama-linux-arm64-jetpack5.tgz
-        tar c -C ./dist/ ./lib/ollama/cuda_jetpack6  | pigz -9vc >./dist/ollama-linux-arm64-jetpack6.tgz
+        tar c -C ./dist/ --exclude cuda_jetpack5 --exclude cuda_jetpack6 bin lib | zstd --ultra -22 -T0 >./dist/ollama-linux-arm64.tar.zst
+        tar c -C ./dist/ ./lib/ollama/cuda_jetpack5  | zstd --ultra -22 -T0 >./dist/ollama-linux-arm64-jetpack5.tar.zst
+        tar c -C ./dist/ ./lib/ollama/cuda_jetpack6  | zstd --ultra -22 -T0 >./dist/ollama-linux-arm64-jetpack6.tar.zst
 elif echo $PLATFORM | grep "amd64" > /dev/null ; then
-        tar c -C ./dist/ --exclude rocm bin lib | pigz -9vc >./dist/ollama-linux-amd64.tgz
-        tar c -C ./dist/ ./lib/ollama/rocm  | pigz -9vc >./dist/ollama-linux-amd64-rocm.tgz
+        tar c -C ./dist/ --exclude rocm bin lib | zstd --ultra -22 -T0 >./dist/ollama-linux-amd64.tar.zst
+        tar c -C ./dist/ ./lib/ollama/rocm  | zstd --ultra -22 -T0 >./dist/ollama-linux-amd64-rocm.tar.zst
 fi

scripts/deduplicate_cuda_libs.sh

@@ -0,0 +1,60 @@
+#!/bin/sh
+#
+# Deduplicate CUDA libraries across mlx_* and cuda_* directories
+# This script finds identical .so* files in mlx_cuda_* directories that exist
+# in corresponding cuda_* directories and replaces them with symlinks.
+#
+
+set -eu
+
+if [ $# -eq 0 ]; then
+    echo "ERROR: No directory specified" >&2
+    echo "Usage: $0 <base_directory>" >&2
+    exit 1
+fi
+
+base_dir="$1"
+
+if [ ! -d "${base_dir}" ]; then
+    echo "ERROR: Directory ${base_dir} does not exist" >&2
+    exit 1
+fi
+
+echo "Deduplicating CUDA libraries in ${base_dir}..."
+
+# Find all mlx_cuda_* directories
+for mlx_dir in "${base_dir}"/lib/ollama/mlx_cuda_*; do
+    [ -d "${mlx_dir}" ] || continue
+
+    # Extract CUDA version (e.g., v12, v13)
+    cuda_version=$(basename "${mlx_dir}" | sed 's/mlx_cuda_//')
+    cuda_dir="${base_dir}/lib/ollama/cuda_${cuda_version}"
+
+    # Skip if corresponding cuda_* directory doesn't exist
+    [ -d "${cuda_dir}" ] || continue
+
+    echo "  Checking ${mlx_dir} against ${cuda_dir}..."
+
+    # Find all .so* files in mlx directory
+    find "${mlx_dir}" -type f -name "*.so*" | while read mlx_file; do
+        filename=$(basename "${mlx_file}")
+        cuda_file="${cuda_dir}/${filename}"
+
+        # Skip if file doesn't exist in cuda directory
+        [ -f "${cuda_file}" ] || continue
+
+        # Compare checksums
+        mlx_sum=$(sha256sum "${mlx_file}" | awk '{print $1}')
+        cuda_sum=$(sha256sum "${cuda_file}" | awk '{print $1}')
+
+        if [ "${mlx_sum}" = "${cuda_sum}" ]; then
+            echo "    Deduplicating ${filename}"
+            # Calculate relative path from mlx_dir to cuda_dir
+            rel_path="../cuda_${cuda_version}/${filename}"
+            rm -f "${mlx_file}"
+            ln -s "${rel_path}" "${mlx_file}"
+        fi
+    done
+done
+
+echo "Deduplication complete"