build: add -O3 optimization to CGO flags (#13877)

CGO_CFLAGS and CGO_CXXFLAGS were being set without optimization flags,
which overrides Go's default -O2 and results in unoptimized C++ code.

This caused significant performance degradation in release builds
compared to local `go build` which uses the default optimization.

- build_darwin.sh: add -O3 to CGO_CFLAGS and CGO_CXXFLAGS exports
- Dockerfile: preserve CGO_CFLAGS/CGO_CXXFLAGS from build args instead
  of overwriting them
- app/README.md: update documentation to include -O3

.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

@@ -372,13 +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 ;;

CMakeLists.txt

@@ -48,9 +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)
@@ -189,13 +190,21 @@ if(MLX_ENGINE)
     install(TARGETS mlx mlxc
         RUNTIME_DEPENDENCIES
             DIRECTORIES ${CUDAToolkit_BIN_DIR} ${CUDAToolkit_BIN_DIR}/x64 ${CUDAToolkit_LIBRARY_DIR}
-            PRE_INCLUDE_REGEXES cublas cublasLt cudart nvrtc cudnn nccl
+            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

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++
 
@@ -149,6 +149,7 @@ 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
@@ -156,11 +157,6 @@ 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}
-COPY . .
-ARG GOFLAGS="'-ldflags=-w -s'"
-ENV CGO_ENABLED=1
-ARG CGO_CFLAGS
-ARG CGO_CXXFLAGS
 
 FROM base AS build
 WORKDIR /go/src/github.com/ollama/ollama
@@ -169,12 +165,16 @@ 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 "$(cat MLX_VERSION)" https://github.com/ml-explore/mlx-c.git build/_deps/mlx-c-src
 ARG GOFLAGS="'-ldflags=-w -s'"
 ENV CGO_ENABLED=1
 ARG CGO_CFLAGS
 ARG CGO_CXXFLAGS
+ENV CGO_CFLAGS="${CGO_CFLAGS} -I/go/src/github.com/ollama/ollama/build/_deps/mlx-c-src"
+ENV CGO_CXXFLAGS="${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/

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)
@@ -525,7 +558,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [LiteLLM](https://github.com/BerriAI/litellm)
 - [OllamaFarm for Go](https://github.com/presbrey/ollamafarm)
 - [OllamaSharp for .NET](https://github.com/awaescher/OllamaSharp)
-- [Ollama for Ruby](https://github.com/gbaptista/ollama-ai)
+- [Ollama for Ruby](https://github.com/crmne/ruby_llm)
 - [Ollama-rs for Rust](https://github.com/pepperoni21/ollama-rs)
 - [Ollama-hpp for C++](https://github.com/jmont-dev/ollama-hpp)
 - [Ollama4j for Java](https://github.com/ollama4j/ollama4j)
@@ -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)

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].
@@ -735,7 +749,7 @@ type ShowResponse struct {
 	Messages      []Message          `json:"messages,omitempty"`
 	RemoteModel   string             `json:"remote_model,omitempty"`
 	RemoteHost    string             `json:"remote_host,omitempty"`
-	ModelInfo     map[string]any     `json:"model_info,omitempty"`
+	ModelInfo     map[string]any     `json:"model_info"`
 	ProjectorInfo map[string]any     `json:"projector_info,omitempty"`
 	Tensors       []Tensor           `json:"tensors,omitempty"`
 	Capabilities  []model.Capability `json:"capabilities,omitempty"`
@@ -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/README.md

@@ -75,9 +75,9 @@ The `-dev` flag enables:
 CI builds with Xcode 14.1 for OS compatibility prior to v13.  If you want to manually build v11+ support, you can download the older Xcode [here](https://developer.apple.com/services-account/download?path=/Developer_Tools/Xcode_14.1/Xcode_14.1.xip), extract, then `mv ./Xcode.app /Applications/Xcode_14.1.0.app` then activate with:
 
 ```
-export CGO_CFLAGS=-mmacosx-version-min=12.0
-export CGO_CXXFLAGS=-mmacosx-version-min=12.0
-export CGO_LDFLAGS=-mmacosx-version-min=12.0
+export CGO_CFLAGS="-O3 -mmacosx-version-min=12.0"
+export CGO_CXXFLAGS="-O3 -mmacosx-version-min=12.0"
+export CGO_LDFLAGS="-mmacosx-version-min=12.0"
 export SDKROOT=/Applications/Xcode_14.1.0.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk
 export DEVELOPER_DIR=/Applications/Xcode_14.1.0.app/Contents/Developer
 ```

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

@@ -35,6 +35,7 @@ import (
 	"golang.org/x/term"
 
 	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/cmd/config"
 	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/format"
 	"github.com/ollama/ollama/parser"
@@ -46,8 +47,9 @@ 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"
-	imagegenclient "github.com/ollama/ollama/x/imagegen/client"
 )
 
 const ConnectInstructions = "To sign in, navigate to:\n    %s\n\n"
@@ -93,14 +95,87 @@ 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)
+	}
+
+	// Check for --experimental flag for safetensors model creation
+	experimental, _ := cmd.Flags().GetBool("experimental")
+	if experimental {
+		// Get Modelfile content - either from -f flag or default to "FROM ."
+		var reader io.Reader
+		filename, err := getModelfileName(cmd)
+		if os.IsNotExist(err) || filename == "" {
+			// No Modelfile specified or found - use default
+			reader = strings.NewReader("FROM .\n")
+		} else if err != nil {
+			return err
+		} else {
+			f, err := os.Open(filename)
+			if err != nil {
+				return err
+			}
+			defer f.Close()
+			reader = f
+		}
+
+		// Parse the Modelfile
+		modelfile, err := parser.ParseFile(reader)
+		if err != nil {
+			return fmt.Errorf("failed to parse Modelfile: %w", err)
+		}
+
+		// Extract FROM path and configuration
+		var modelDir string
+		mfConfig := &xcreateclient.ModelfileConfig{}
+
+		for _, cmd := range modelfile.Commands {
+			switch cmd.Name {
+			case "model":
+				modelDir = cmd.Args
+			case "template":
+				mfConfig.Template = cmd.Args
+			case "system":
+				mfConfig.System = cmd.Args
+			case "license":
+				mfConfig.License = cmd.Args
+			}
+		}
+
+		if modelDir == "" {
+			modelDir = "."
+		}
+
+		// Resolve relative paths based on Modelfile location
+		if !filepath.IsAbs(modelDir) && filename != "" {
+			modelDir = filepath.Join(filepath.Dir(filename), modelDir)
+		}
+
+		quantize, _ := cmd.Flags().GetString("quantize")
+		return xcreateclient.CreateModel(xcreateclient.CreateOptions{
+			ModelName: modelName,
+			ModelDir:  modelDir,
+			Quantize:  quantize,
+			Modelfile: mfConfig,
+		}, p)
+	}
+
 	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 imagegen.IsTensorModelDir(".") {
-				return imagegenclient.CreateModel(args[0], ".", p)
+			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 {
@@ -133,7 +208,7 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 	}
 	spinner.Stop()
 
-	req.Model = args[0]
+	req.Model = modelName
 	quantize, _ := cmd.Flags().GetString("quantize")
 	if quantize != "" {
 		req.Quantize = quantize
@@ -464,14 +539,6 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 
 	name := args[0]
 
-	// Check if this is a known image generation model (skip Show/Pull)
-	if imagegen.HasTensorLayers(name) {
-		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)
-	}
-
 	info, err := func() (*api.ShowResponse, error) {
 		showReq := &api.ShowRequest{Name: name}
 		info, err := client.Show(cmd.Context(), showReq)
@@ -533,9 +600,18 @@ 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("experimental-yolo")
+	enableWebsearch, _ := cmd.Flags().GetBool("experimental-websearch")
 
 	if interactive {
 		if err := loadOrUnloadModel(cmd, &opts); err != nil {
@@ -565,7 +641,7 @@ 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)
@@ -671,7 +747,11 @@ 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)
 			}
@@ -820,11 +900,11 @@ 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)
 			}
 		}
 
@@ -837,11 +917,6 @@ func DeleteHandler(cmd *cobra.Command, args []string) error {
 }
 
 func ShowHandler(cmd *cobra.Command, args []string) error {
-	// Check if this is an image generation model
-	if imagegen.HasTensorLayers(args[0]) {
-		return imagegen.Show(args[0], os.Stdout)
-	}
-
 	client, err := api.ClientFromEnvironment()
 	if err != nil {
 		return err
@@ -944,8 +1019,10 @@ func showInfo(resp *api.ShowResponse, verbose bool, w io.Writer) error {
 		}
 
 		if resp.ModelInfo != nil {
-			arch := resp.ModelInfo["general.architecture"].(string)
-			rows = append(rows, []string{"", "architecture", arch})
+			arch, _ := resp.ModelInfo["general.architecture"].(string)
+			if arch != "" {
+				rows = append(rows, []string{"", "architecture", arch})
+			}
 
 			var paramStr string
 			if resp.Details.ParameterSize != "" {
@@ -955,7 +1032,9 @@ func showInfo(resp *api.ShowResponse, verbose bool, w io.Writer) error {
 					paramStr = format.HumanNumber(uint64(f))
 				}
 			}
-			rows = append(rows, []string{"", "parameters", paramStr})
+			if paramStr != "" {
+				rows = append(rows, []string{"", "parameters", paramStr})
+			}
 
 			if v, ok := resp.ModelInfo[fmt.Sprintf("%s.context_length", arch)]; ok {
 				if f, ok := v.(float64); ok {
@@ -1741,15 +1820,22 @@ func NewCLI() *cobra.Command {
 	rootCmd.Flags().BoolP("version", "v", false, "Show version information")
 
 	createCmd := &cobra.Command{
-		Use:     "create MODEL",
-		Short:   "Create a model",
-		Args:    cobra.ExactArgs(1),
-		PreRunE: checkServerHeartbeat,
-		RunE:    CreateHandler,
+		Use:   "create MODEL",
+		Short: "Create a model",
+		Args:  cobra.ExactArgs(1),
+		PreRunE: func(cmd *cobra.Command, args []string) error {
+			// Skip server check for experimental mode (writes directly to disk)
+			if experimental, _ := cmd.Flags().GetBool("experimental"); experimental {
+				return nil
+			}
+			return checkServerHeartbeat(cmd, args)
+		},
+		RunE: CreateHandler,
 	}
 
 	createCmd.Flags().StringP("file", "f", "", "Name of the Modelfile (default \"Modelfile\")")
 	createCmd.Flags().StringP("quantize", "q", "", "Quantize model to this level (e.g. q4_K_M)")
+	createCmd.Flags().Bool("experimental", false, "Enable experimental safetensors model creation")
 
 	showCmd := &cobra.Command{
 		Use:     "show MODEL",
@@ -1786,6 +1872,7 @@ func NewCLI() *cobra.Command {
 	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().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)
@@ -1903,6 +1990,7 @@ 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{
@@ -1943,6 +2031,7 @@ func NewCLI() *cobra.Command {
 		copyCmd,
 		deleteCmd,
 		runnerCmd,
+		config.LaunchCmd(checkServerHeartbeat),
 	)
 
 	return rootCmd

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/config/claude.go

@@ -0,0 +1,36 @@
+package config
+
+import (
+	"fmt"
+	"os"
+	"os/exec"
+)
+
+// Claude implements Runner for Claude Code integration
+type Claude struct{}
+
+func (c *Claude) String() string { return "Claude Code" }
+
+func (c *Claude) args(model string) []string {
+	if model != "" {
+		return []string{"--model", model}
+	}
+	return nil
+}
+
+func (c *Claude) Run(model string) error {
+	if _, err := exec.LookPath("claude"); err != nil {
+		return fmt.Errorf("claude is not installed, install from https://code.claude.com/docs/en/quickstart")
+	}
+
+	cmd := exec.Command("claude", c.args(model)...)
+	cmd.Stdin = os.Stdin
+	cmd.Stdout = os.Stdout
+	cmd.Stderr = os.Stderr
+	cmd.Env = append(os.Environ(),
+		"ANTHROPIC_BASE_URL=http://localhost:11434",
+		"ANTHROPIC_API_KEY=",
+		"ANTHROPIC_AUTH_TOKEN=ollama",
+	)
+	return cmd.Run()
+}

cmd/config/claude_test.go

@@ -0,0 +1,42 @@
+package config
+
+import (
+	"slices"
+	"testing"
+)
+
+func TestClaudeIntegration(t *testing.T) {
+	c := &Claude{}
+
+	t.Run("String", func(t *testing.T) {
+		if got := c.String(); got != "Claude Code" {
+			t.Errorf("String() = %q, want %q", got, "Claude Code")
+		}
+	})
+
+	t.Run("implements Runner", func(t *testing.T) {
+		var _ Runner = c
+	})
+}
+
+func TestClaudeArgs(t *testing.T) {
+	c := &Claude{}
+
+	tests := []struct {
+		name  string
+		model string
+		want  []string
+	}{
+		{"with model", "llama3.2", []string{"--model", "llama3.2"}},
+		{"empty model", "", nil},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			got := c.args(tt.model)
+			if !slices.Equal(got, tt.want) {
+				t.Errorf("args(%q) = %v, want %v", tt.model, got, tt.want)
+			}
+		})
+	}
+}

cmd/config/codex.go

@@ -0,0 +1,61 @@
+package config
+
+import (
+	"fmt"
+	"os"
+	"os/exec"
+	"strings"
+
+	"golang.org/x/mod/semver"
+)
+
+// Codex implements Runner for Codex integration
+type Codex struct{}
+
+func (c *Codex) String() string { return "Codex" }
+
+func (c *Codex) args(model string) []string {
+	args := []string{"--oss"}
+	if model != "" {
+		args = append(args, "-m", model)
+	}
+	return args
+}
+
+func (c *Codex) Run(model string) error {
+	if err := checkCodexVersion(); err != nil {
+		return err
+	}
+
+	cmd := exec.Command("codex", c.args(model)...)
+	cmd.Stdin = os.Stdin
+	cmd.Stdout = os.Stdout
+	cmd.Stderr = os.Stderr
+	return cmd.Run()
+}
+
+func checkCodexVersion() error {
+	if _, err := exec.LookPath("codex"); err != nil {
+		return fmt.Errorf("codex is not installed, install with: npm install -g @openai/codex")
+	}
+
+	out, err := exec.Command("codex", "--version").Output()
+	if err != nil {
+		return fmt.Errorf("failed to get codex version: %w", err)
+	}
+
+	// Parse output like "codex-cli 0.87.0"
+	fields := strings.Fields(strings.TrimSpace(string(out)))
+	if len(fields) < 2 {
+		return fmt.Errorf("unexpected codex version output: %s", string(out))
+	}
+
+	version := "v" + fields[len(fields)-1]
+	minVersion := "v0.81.0"
+
+	if semver.Compare(version, minVersion) < 0 {
+		return fmt.Errorf("codex version %s is too old, minimum required is %s, update with: npm update -g @openai/codex", fields[len(fields)-1], "0.81.0")
+	}
+
+	return nil
+}

cmd/config/codex_test.go

@@ -0,0 +1,28 @@
+package config
+
+import (
+	"slices"
+	"testing"
+)
+
+func TestCodexArgs(t *testing.T) {
+	c := &Codex{}
+
+	tests := []struct {
+		name  string
+		model string
+		want  []string
+	}{
+		{"with model", "llama3.2", []string{"--oss", "-m", "llama3.2"}},
+		{"empty model", "", []string{"--oss"}},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			got := c.args(tt.model)
+			if !slices.Equal(got, tt.want) {
+				t.Errorf("args(%q) = %v, want %v", tt.model, got, tt.want)
+			}
+		})
+	}
+}

cmd/config/config.go

@@ -0,0 +1,115 @@
+// Package config provides integration configuration for external coding tools
+// (Claude Code, Codex, Droid, OpenCode) to use Ollama models.
+package config
+
+import (
+	"encoding/json"
+	"errors"
+	"fmt"
+	"os"
+	"path/filepath"
+	"strings"
+)
+
+type integration struct {
+	Models []string `json:"models"`
+}
+
+type config struct {
+	Integrations map[string]*integration `json:"integrations"`
+}
+
+func configPath() (string, error) {
+	home, err := os.UserHomeDir()
+	if err != nil {
+		return "", err
+	}
+	return filepath.Join(home, ".ollama", "config", "config.json"), nil
+}
+
+func load() (*config, error) {
+	path, err := configPath()
+	if err != nil {
+		return nil, err
+	}
+
+	data, err := os.ReadFile(path)
+	if err != nil {
+		if os.IsNotExist(err) {
+			return &config{Integrations: make(map[string]*integration)}, nil
+		}
+		return nil, err
+	}
+
+	var cfg config
+	if err := json.Unmarshal(data, &cfg); err != nil {
+		return nil, fmt.Errorf("failed to parse config: %w, at: %s", err, path)
+	}
+	if cfg.Integrations == nil {
+		cfg.Integrations = make(map[string]*integration)
+	}
+	return &cfg, nil
+}
+
+func save(cfg *config) error {
+	path, err := configPath()
+	if err != nil {
+		return err
+	}
+
+	if err := os.MkdirAll(filepath.Dir(path), 0o755); err != nil {
+		return err
+	}
+
+	data, err := json.MarshalIndent(cfg, "", "  ")
+	if err != nil {
+		return err
+	}
+
+	return writeWithBackup(path, data)
+}
+
+func saveIntegration(appName string, models []string) error {
+	if appName == "" {
+		return errors.New("app name cannot be empty")
+	}
+
+	cfg, err := load()
+	if err != nil {
+		return err
+	}
+
+	cfg.Integrations[strings.ToLower(appName)] = &integration{
+		Models: models,
+	}
+
+	return save(cfg)
+}
+
+func loadIntegration(appName string) (*integration, error) {
+	cfg, err := load()
+	if err != nil {
+		return nil, err
+	}
+
+	ic, ok := cfg.Integrations[strings.ToLower(appName)]
+	if !ok {
+		return nil, os.ErrNotExist
+	}
+
+	return ic, nil
+}
+
+func listIntegrations() ([]integration, error) {
+	cfg, err := load()
+	if err != nil {
+		return nil, err
+	}
+
+	result := make([]integration, 0, len(cfg.Integrations))
+	for _, ic := range cfg.Integrations {
+		result = append(result, *ic)
+	}
+
+	return result, nil
+}

cmd/config/config_test.go

@@ -0,0 +1,373 @@
+package config
+
+import (
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+// setTestHome sets both HOME (Unix) and USERPROFILE (Windows) for cross-platform tests
+func setTestHome(t *testing.T, dir string) {
+	t.Setenv("HOME", dir)
+	t.Setenv("USERPROFILE", dir)
+}
+
+// editorPaths is a test helper that safely calls Paths if the runner implements Editor
+func editorPaths(r Runner) []string {
+	if editor, ok := r.(Editor); ok {
+		return editor.Paths()
+	}
+	return nil
+}
+
+func TestIntegrationConfig(t *testing.T) {
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	t.Run("save and load round-trip", func(t *testing.T) {
+		models := []string{"llama3.2", "mistral", "qwen2.5"}
+		if err := saveIntegration("claude", models); err != nil {
+			t.Fatal(err)
+		}
+
+		config, err := loadIntegration("claude")
+		if err != nil {
+			t.Fatal(err)
+		}
+
+		if len(config.Models) != len(models) {
+			t.Errorf("expected %d models, got %d", len(models), len(config.Models))
+		}
+		for i, m := range models {
+			if config.Models[i] != m {
+				t.Errorf("model %d: expected %s, got %s", i, m, config.Models[i])
+			}
+		}
+	})
+
+	t.Run("defaultModel returns first model", func(t *testing.T) {
+		saveIntegration("codex", []string{"model-a", "model-b"})
+
+		config, _ := loadIntegration("codex")
+		defaultModel := ""
+		if len(config.Models) > 0 {
+			defaultModel = config.Models[0]
+		}
+		if defaultModel != "model-a" {
+			t.Errorf("expected model-a, got %s", defaultModel)
+		}
+	})
+
+	t.Run("defaultModel returns empty for no models", func(t *testing.T) {
+		config := &integration{Models: []string{}}
+		defaultModel := ""
+		if len(config.Models) > 0 {
+			defaultModel = config.Models[0]
+		}
+		if defaultModel != "" {
+			t.Errorf("expected empty string, got %s", defaultModel)
+		}
+	})
+
+	t.Run("app name is case-insensitive", func(t *testing.T) {
+		saveIntegration("Claude", []string{"model-x"})
+
+		config, err := loadIntegration("claude")
+		if err != nil {
+			t.Fatal(err)
+		}
+		defaultModel := ""
+		if len(config.Models) > 0 {
+			defaultModel = config.Models[0]
+		}
+		if defaultModel != "model-x" {
+			t.Errorf("expected model-x, got %s", defaultModel)
+		}
+	})
+
+	t.Run("multiple integrations in single file", func(t *testing.T) {
+		saveIntegration("app1", []string{"model-1"})
+		saveIntegration("app2", []string{"model-2"})
+
+		config1, _ := loadIntegration("app1")
+		config2, _ := loadIntegration("app2")
+
+		defaultModel1 := ""
+		if len(config1.Models) > 0 {
+			defaultModel1 = config1.Models[0]
+		}
+		defaultModel2 := ""
+		if len(config2.Models) > 0 {
+			defaultModel2 = config2.Models[0]
+		}
+		if defaultModel1 != "model-1" {
+			t.Errorf("expected model-1, got %s", defaultModel1)
+		}
+		if defaultModel2 != "model-2" {
+			t.Errorf("expected model-2, got %s", defaultModel2)
+		}
+	})
+}
+
+func TestListIntegrations(t *testing.T) {
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	t.Run("returns empty when no integrations", func(t *testing.T) {
+		configs, err := listIntegrations()
+		if err != nil {
+			t.Fatal(err)
+		}
+		if len(configs) != 0 {
+			t.Errorf("expected 0 integrations, got %d", len(configs))
+		}
+	})
+
+	t.Run("returns all saved integrations", func(t *testing.T) {
+		saveIntegration("claude", []string{"model-1"})
+		saveIntegration("droid", []string{"model-2"})
+
+		configs, err := listIntegrations()
+		if err != nil {
+			t.Fatal(err)
+		}
+		if len(configs) != 2 {
+			t.Errorf("expected 2 integrations, got %d", len(configs))
+		}
+	})
+}
+
+func TestEditorPaths(t *testing.T) {
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	t.Run("returns empty for claude (no Editor)", func(t *testing.T) {
+		r := integrations["claude"]
+		paths := editorPaths(r)
+		if len(paths) != 0 {
+			t.Errorf("expected no paths for claude, got %v", paths)
+		}
+	})
+
+	t.Run("returns empty for codex (no Editor)", func(t *testing.T) {
+		r := integrations["codex"]
+		paths := editorPaths(r)
+		if len(paths) != 0 {
+			t.Errorf("expected no paths for codex, got %v", paths)
+		}
+	})
+
+	t.Run("returns empty for droid when no config exists", func(t *testing.T) {
+		r := integrations["droid"]
+		paths := editorPaths(r)
+		if len(paths) != 0 {
+			t.Errorf("expected no paths, got %v", paths)
+		}
+	})
+
+	t.Run("returns path for droid when config exists", func(t *testing.T) {
+		settingsDir, _ := os.UserHomeDir()
+		settingsDir = filepath.Join(settingsDir, ".factory")
+		os.MkdirAll(settingsDir, 0o755)
+		os.WriteFile(filepath.Join(settingsDir, "settings.json"), []byte(`{}`), 0o644)
+
+		r := integrations["droid"]
+		paths := editorPaths(r)
+		if len(paths) != 1 {
+			t.Errorf("expected 1 path, got %d", len(paths))
+		}
+	})
+
+	t.Run("returns paths for opencode when configs exist", func(t *testing.T) {
+		home, _ := os.UserHomeDir()
+		configDir := filepath.Join(home, ".config", "opencode")
+		stateDir := filepath.Join(home, ".local", "state", "opencode")
+		os.MkdirAll(configDir, 0o755)
+		os.MkdirAll(stateDir, 0o755)
+		os.WriteFile(filepath.Join(configDir, "opencode.json"), []byte(`{}`), 0o644)
+		os.WriteFile(filepath.Join(stateDir, "model.json"), []byte(`{}`), 0o644)
+
+		r := integrations["opencode"]
+		paths := editorPaths(r)
+		if len(paths) != 2 {
+			t.Errorf("expected 2 paths, got %d: %v", len(paths), paths)
+		}
+	})
+}
+
+func TestLoadIntegration_CorruptedJSON(t *testing.T) {
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	// Create corrupted config.json file
+	dir := filepath.Join(tmpDir, ".ollama", "config")
+	os.MkdirAll(dir, 0o755)
+	os.WriteFile(filepath.Join(dir, "config.json"), []byte(`{corrupted json`), 0o644)
+
+	// Corrupted file is treated as empty, so loadIntegration returns not found
+	_, err := loadIntegration("test")
+	if err == nil {
+		t.Error("expected error for nonexistent integration in corrupted file")
+	}
+}
+
+func TestSaveIntegration_NilModels(t *testing.T) {
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	if err := saveIntegration("test", nil); err != nil {
+		t.Fatalf("saveIntegration with nil models failed: %v", err)
+	}
+
+	config, err := loadIntegration("test")
+	if err != nil {
+		t.Fatalf("loadIntegration failed: %v", err)
+	}
+
+	if config.Models == nil {
+		// nil is acceptable
+	} else if len(config.Models) != 0 {
+		t.Errorf("expected empty or nil models, got %v", config.Models)
+	}
+}
+
+func TestSaveIntegration_EmptyAppName(t *testing.T) {
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	err := saveIntegration("", []string{"model"})
+	if err == nil {
+		t.Error("expected error for empty app name, got nil")
+	}
+	if err != nil && !strings.Contains(err.Error(), "app name cannot be empty") {
+		t.Errorf("expected 'app name cannot be empty' error, got: %v", err)
+	}
+}
+
+func TestLoadIntegration_NonexistentIntegration(t *testing.T) {
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	_, err := loadIntegration("nonexistent")
+	if err == nil {
+		t.Error("expected error for nonexistent integration, got nil")
+	}
+	if !os.IsNotExist(err) {
+		t.Logf("error type is os.ErrNotExist as expected: %v", err)
+	}
+}
+
+func TestConfigPath(t *testing.T) {
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	path, err := configPath()
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	expected := filepath.Join(tmpDir, ".ollama", "config", "config.json")
+	if path != expected {
+		t.Errorf("expected %s, got %s", expected, path)
+	}
+}
+
+func TestLoad(t *testing.T) {
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	t.Run("returns empty config when file does not exist", func(t *testing.T) {
+		cfg, err := load()
+		if err != nil {
+			t.Fatal(err)
+		}
+		if cfg == nil {
+			t.Fatal("expected non-nil config")
+		}
+		if cfg.Integrations == nil {
+			t.Error("expected non-nil Integrations map")
+		}
+		if len(cfg.Integrations) != 0 {
+			t.Errorf("expected empty Integrations, got %d", len(cfg.Integrations))
+		}
+	})
+
+	t.Run("loads existing config", func(t *testing.T) {
+		path, _ := configPath()
+		os.MkdirAll(filepath.Dir(path), 0o755)
+		os.WriteFile(path, []byte(`{"integrations":{"test":{"models":["model-a"]}}}`), 0o644)
+
+		cfg, err := load()
+		if err != nil {
+			t.Fatal(err)
+		}
+		if cfg.Integrations["test"] == nil {
+			t.Fatal("expected test integration")
+		}
+		if len(cfg.Integrations["test"].Models) != 1 {
+			t.Errorf("expected 1 model, got %d", len(cfg.Integrations["test"].Models))
+		}
+	})
+
+	t.Run("returns error for corrupted JSON", func(t *testing.T) {
+		path, _ := configPath()
+		os.MkdirAll(filepath.Dir(path), 0o755)
+		os.WriteFile(path, []byte(`{corrupted`), 0o644)
+
+		_, err := load()
+		if err == nil {
+			t.Error("expected error for corrupted JSON")
+		}
+	})
+}
+
+func TestSave(t *testing.T) {
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	t.Run("creates config file", func(t *testing.T) {
+		cfg := &config{
+			Integrations: map[string]*integration{
+				"test": {Models: []string{"model-a", "model-b"}},
+			},
+		}
+
+		if err := save(cfg); err != nil {
+			t.Fatal(err)
+		}
+
+		path, _ := configPath()
+		if _, err := os.Stat(path); os.IsNotExist(err) {
+			t.Error("config file was not created")
+		}
+	})
+
+	t.Run("round-trip preserves data", func(t *testing.T) {
+		cfg := &config{
+			Integrations: map[string]*integration{
+				"claude": {Models: []string{"llama3.2", "mistral"}},
+				"codex":  {Models: []string{"qwen2.5"}},
+			},
+		}
+
+		if err := save(cfg); err != nil {
+			t.Fatal(err)
+		}
+
+		loaded, err := load()
+		if err != nil {
+			t.Fatal(err)
+		}
+
+		if len(loaded.Integrations) != 2 {
+			t.Errorf("expected 2 integrations, got %d", len(loaded.Integrations))
+		}
+		if loaded.Integrations["claude"] == nil {
+			t.Error("missing claude integration")
+		}
+		if len(loaded.Integrations["claude"].Models) != 2 {
+			t.Errorf("expected 2 models for claude, got %d", len(loaded.Integrations["claude"].Models))
+		}
+	})
+}

cmd/config/droid.go

@@ -0,0 +1,184 @@
+package config
+
+import (
+	"encoding/json"
+	"fmt"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"slices"
+)
+
+// Droid implements Runner and Editor for Droid integration
+type Droid struct{}
+
+// droidSettings represents the Droid settings.json file (only fields we use)
+type droidSettings struct {
+	CustomModels           []modelEntry    `json:"customModels"`
+	SessionDefaultSettings sessionSettings `json:"sessionDefaultSettings"`
+}
+
+type sessionSettings struct {
+	Model           string `json:"model"`
+	ReasoningEffort string `json:"reasoningEffort"`
+}
+
+type modelEntry struct {
+	Model           string `json:"model"`
+	DisplayName     string `json:"displayName"`
+	BaseURL         string `json:"baseUrl"`
+	APIKey          string `json:"apiKey"`
+	Provider        string `json:"provider"`
+	MaxOutputTokens int    `json:"maxOutputTokens"`
+	SupportsImages  bool   `json:"supportsImages"`
+	ID              string `json:"id"`
+	Index           int    `json:"index"`
+}
+
+func (d *Droid) String() string { return "Droid" }
+
+func (d *Droid) Run(model string) error {
+	if _, err := exec.LookPath("droid"); err != nil {
+		return fmt.Errorf("droid is not installed, install from https://docs.factory.ai/cli/getting-started/quickstart")
+	}
+
+	// Call Edit() to ensure config is up-to-date before launch
+	models := []string{model}
+	if config, err := loadIntegration("droid"); err == nil && len(config.Models) > 0 {
+		models = config.Models
+	}
+	if err := d.Edit(models); err != nil {
+		return fmt.Errorf("setup failed: %w", err)
+	}
+
+	cmd := exec.Command("droid")
+	cmd.Stdin = os.Stdin
+	cmd.Stdout = os.Stdout
+	cmd.Stderr = os.Stderr
+	return cmd.Run()
+}
+
+func (d *Droid) Paths() []string {
+	home, err := os.UserHomeDir()
+	if err != nil {
+		return nil
+	}
+	p := filepath.Join(home, ".factory", "settings.json")
+	if _, err := os.Stat(p); err == nil {
+		return []string{p}
+	}
+	return nil
+}
+
+func (d *Droid) Edit(models []string) error {
+	if len(models) == 0 {
+		return nil
+	}
+
+	home, err := os.UserHomeDir()
+	if err != nil {
+		return err
+	}
+
+	settingsPath := filepath.Join(home, ".factory", "settings.json")
+	if err := os.MkdirAll(filepath.Dir(settingsPath), 0o755); err != nil {
+		return err
+	}
+
+	// Read file once, unmarshal twice:
+	// map preserves unknown fields for writing back (including extra fields in model entries)
+	settingsMap := make(map[string]any)
+	var settings droidSettings
+	if data, err := os.ReadFile(settingsPath); err == nil {
+		if err := json.Unmarshal(data, &settingsMap); err != nil {
+			return fmt.Errorf("failed to parse settings file: %w, at: %s", err, settingsPath)
+		}
+		json.Unmarshal(data, &settings) // ignore error, zero values are fine
+	}
+
+	// Keep only non-Ollama models from the raw map (preserves extra fields)
+	// Rebuild Ollama models
+	var nonOllamaModels []any
+	if rawModels, ok := settingsMap["customModels"].([]any); ok {
+		for _, raw := range rawModels {
+			if m, ok := raw.(map[string]any); ok {
+				if m["apiKey"] != "ollama" {
+					nonOllamaModels = append(nonOllamaModels, raw)
+				}
+			}
+		}
+	}
+
+	// Build new Ollama model entries with sequential indices (0, 1, 2, ...)
+	var newModels []any
+	var defaultModelID string
+	for i, model := range models {
+		modelID := fmt.Sprintf("custom:%s-%d", model, i)
+		newModels = append(newModels, modelEntry{
+			Model:           model,
+			DisplayName:     model,
+			BaseURL:         "http://localhost:11434/v1",
+			APIKey:          "ollama",
+			Provider:        "generic-chat-completion-api",
+			MaxOutputTokens: 64000,
+			SupportsImages:  false,
+			ID:              modelID,
+			Index:           i,
+		})
+		if i == 0 {
+			defaultModelID = modelID
+		}
+	}
+
+	settingsMap["customModels"] = append(newModels, nonOllamaModels...)
+
+	// Update session default settings (preserve unknown fields in the nested object)
+	sessionSettings, ok := settingsMap["sessionDefaultSettings"].(map[string]any)
+	if !ok {
+		sessionSettings = make(map[string]any)
+	}
+	sessionSettings["model"] = defaultModelID
+
+	if !isValidReasoningEffort(settings.SessionDefaultSettings.ReasoningEffort) {
+		sessionSettings["reasoningEffort"] = "none"
+	}
+
+	settingsMap["sessionDefaultSettings"] = sessionSettings
+
+	data, err := json.MarshalIndent(settingsMap, "", "  ")
+	if err != nil {
+		return err
+	}
+	return writeWithBackup(settingsPath, data)
+}
+
+func (d *Droid) Models() []string {
+	home, err := os.UserHomeDir()
+	if err != nil {
+		return nil
+	}
+
+	data, err := os.ReadFile(filepath.Join(home, ".factory", "settings.json"))
+	if err != nil {
+		return nil
+	}
+
+	var settings droidSettings
+	if err := json.Unmarshal(data, &settings); err != nil {
+		return nil
+	}
+
+	var result []string
+	for _, m := range settings.CustomModels {
+		if m.APIKey == "ollama" {
+			result = append(result, m.Model)
+		}
+	}
+	return result
+}
+
+var validReasoningEfforts = []string{"high", "medium", "low", "none"}
+
+func isValidReasoningEffort(effort string) bool {
+	return slices.Contains(validReasoningEfforts, effort)
+}

cmd/config/files.go

@@ -0,0 +1,99 @@
+package config
+
+import (
+	"bytes"
+	"encoding/json"
+	"fmt"
+	"os"
+	"path/filepath"
+	"time"
+)
+
+func readJSONFile(path string) (map[string]any, error) {
+	data, err := os.ReadFile(path)
+	if err != nil {
+		return nil, err
+	}
+	var result map[string]any
+	if err := json.Unmarshal(data, &result); err != nil {
+		return nil, err
+	}
+	return result, nil
+}
+
+func copyFile(src, dst string) error {
+	info, err := os.Stat(src)
+	if err != nil {
+		return err
+	}
+	data, err := os.ReadFile(src)
+	if err != nil {
+		return err
+	}
+	return os.WriteFile(dst, data, info.Mode().Perm())
+}
+
+func backupDir() string {
+	return filepath.Join(os.TempDir(), "ollama-backups")
+}
+
+func backupToTmp(srcPath string) (string, error) {
+	dir := backupDir()
+	if err := os.MkdirAll(dir, 0o755); err != nil {
+		return "", err
+	}
+
+	backupPath := filepath.Join(dir, fmt.Sprintf("%s.%d", filepath.Base(srcPath), time.Now().Unix()))
+	if err := copyFile(srcPath, backupPath); err != nil {
+		return "", err
+	}
+	return backupPath, nil
+}
+
+// writeWithBackup writes data to path via temp file + rename, backing up any existing file first
+func writeWithBackup(path string, data []byte) error {
+	var backupPath string
+	// backup must be created before any writes to the target file
+	if existingContent, err := os.ReadFile(path); err == nil {
+		if !bytes.Equal(existingContent, data) {
+			backupPath, err = backupToTmp(path)
+			if err != nil {
+				return fmt.Errorf("backup failed: %w", err)
+			}
+		}
+	} else if !os.IsNotExist(err) {
+		return fmt.Errorf("read existing file: %w", err)
+	}
+
+	dir := filepath.Dir(path)
+	tmp, err := os.CreateTemp(dir, ".tmp-*")
+	if err != nil {
+		return fmt.Errorf("create temp failed: %w", err)
+	}
+	tmpPath := tmp.Name()
+
+	if _, err := tmp.Write(data); err != nil {
+		_ = tmp.Close()
+		_ = os.Remove(tmpPath)
+		return fmt.Errorf("write failed: %w", err)
+	}
+	if err := tmp.Sync(); err != nil {
+		_ = tmp.Close()
+		_ = os.Remove(tmpPath)
+		return fmt.Errorf("sync failed: %w", err)
+	}
+	if err := tmp.Close(); err != nil {
+		_ = os.Remove(tmpPath)
+		return fmt.Errorf("close failed: %w", err)
+	}
+
+	if err := os.Rename(tmpPath, path); err != nil {
+		_ = os.Remove(tmpPath)
+		if backupPath != "" {
+			_ = copyFile(backupPath, path)
+		}
+		return fmt.Errorf("rename failed: %w", err)
+	}
+
+	return nil
+}

cmd/config/files_test.go

@@ -0,0 +1,502 @@
+package config
+
+import (
+	"encoding/json"
+	"fmt"
+	"os"
+	"path/filepath"
+	"runtime"
+	"testing"
+)
+
+func mustMarshal(t *testing.T, v any) []byte {
+	t.Helper()
+	data, err := json.MarshalIndent(v, "", "  ")
+	if err != nil {
+		t.Fatal(err)
+	}
+	return data
+}
+
+func TestWriteWithBackup(t *testing.T) {
+	tmpDir := t.TempDir()
+
+	t.Run("creates file", func(t *testing.T) {
+		path := filepath.Join(tmpDir, "new.json")
+		data := mustMarshal(t, map[string]string{"key": "value"})
+
+		if err := writeWithBackup(path, data); err != nil {
+			t.Fatal(err)
+		}
+
+		content, err := os.ReadFile(path)
+		if err != nil {
+			t.Fatal(err)
+		}
+
+		var result map[string]string
+		if err := json.Unmarshal(content, &result); err != nil {
+			t.Fatal(err)
+		}
+		if result["key"] != "value" {
+			t.Errorf("expected value, got %s", result["key"])
+		}
+	})
+
+	t.Run("creates backup in /tmp/ollama-backups", func(t *testing.T) {
+		path := filepath.Join(tmpDir, "backup.json")
+
+		os.WriteFile(path, []byte(`{"original": true}`), 0o644)
+
+		data := mustMarshal(t, map[string]bool{"updated": true})
+		if err := writeWithBackup(path, data); err != nil {
+			t.Fatal(err)
+		}
+
+		entries, err := os.ReadDir(backupDir())
+		if err != nil {
+			t.Fatal("backup directory not created")
+		}
+
+		var foundBackup bool
+		for _, entry := range entries {
+			if filepath.Ext(entry.Name()) != ".json" {
+				name := entry.Name()
+				if len(name) > len("backup.json.") && name[:len("backup.json.")] == "backup.json." {
+					backupPath := filepath.Join(backupDir(), name)
+					backup, err := os.ReadFile(backupPath)
+					if err == nil {
+						var backupData map[string]bool
+						json.Unmarshal(backup, &backupData)
+						if backupData["original"] {
+							foundBackup = true
+							os.Remove(backupPath)
+							break
+						}
+					}
+				}
+			}
+		}
+
+		if !foundBackup {
+			t.Error("backup file not created in /tmp/ollama-backups")
+		}
+
+		current, _ := os.ReadFile(path)
+		var currentData map[string]bool
+		json.Unmarshal(current, &currentData)
+		if !currentData["updated"] {
+			t.Error("file doesn't contain updated data")
+		}
+	})
+
+	t.Run("no backup for new file", func(t *testing.T) {
+		path := filepath.Join(tmpDir, "nobak.json")
+
+		data := mustMarshal(t, map[string]string{"new": "file"})
+		if err := writeWithBackup(path, data); err != nil {
+			t.Fatal(err)
+		}
+
+		entries, _ := os.ReadDir(backupDir())
+		for _, entry := range entries {
+			if len(entry.Name()) > len("nobak.json.") && entry.Name()[:len("nobak.json.")] == "nobak.json." {
+				t.Error("backup should not exist for new file")
+			}
+		}
+	})
+
+	t.Run("no backup when content unchanged", func(t *testing.T) {
+		path := filepath.Join(tmpDir, "unchanged.json")
+
+		data := mustMarshal(t, map[string]string{"key": "value"})
+
+		if err := writeWithBackup(path, data); err != nil {
+			t.Fatal(err)
+		}
+
+		entries1, _ := os.ReadDir(backupDir())
+		countBefore := 0
+		for _, e := range entries1 {
+			if len(e.Name()) > len("unchanged.json.") && e.Name()[:len("unchanged.json.")] == "unchanged.json." {
+				countBefore++
+			}
+		}
+
+		if err := writeWithBackup(path, data); err != nil {
+			t.Fatal(err)
+		}
+
+		entries2, _ := os.ReadDir(backupDir())
+		countAfter := 0
+		for _, e := range entries2 {
+			if len(e.Name()) > len("unchanged.json.") && e.Name()[:len("unchanged.json.")] == "unchanged.json." {
+				countAfter++
+			}
+		}
+
+		if countAfter != countBefore {
+			t.Errorf("backup was created when content unchanged (before=%d, after=%d)", countBefore, countAfter)
+		}
+	})
+
+	t.Run("backup filename contains unix timestamp", func(t *testing.T) {
+		path := filepath.Join(tmpDir, "timestamped.json")
+
+		os.WriteFile(path, []byte(`{"v": 1}`), 0o644)
+		data := mustMarshal(t, map[string]int{"v": 2})
+		if err := writeWithBackup(path, data); err != nil {
+			t.Fatal(err)
+		}
+
+		entries, _ := os.ReadDir(backupDir())
+		var found bool
+		for _, entry := range entries {
+			name := entry.Name()
+			if len(name) > len("timestamped.json.") && name[:len("timestamped.json.")] == "timestamped.json." {
+				timestamp := name[len("timestamped.json."):]
+				for _, c := range timestamp {
+					if c < '0' || c > '9' {
+						t.Errorf("backup filename timestamp contains non-numeric character: %s", name)
+					}
+				}
+				found = true
+				os.Remove(filepath.Join(backupDir(), name))
+				break
+			}
+		}
+		if !found {
+			t.Error("backup file with timestamp not found")
+		}
+	})
+}
+
+// Edge case tests for files.go
+
+// TestWriteWithBackup_FailsIfBackupFails documents critical behavior: if backup fails, we must not proceed.
+// User could lose their config with no way to recover.
+func TestWriteWithBackup_FailsIfBackupFails(t *testing.T) {
+	if runtime.GOOS == "windows" {
+		t.Skip("permission tests unreliable on Windows")
+	}
+
+	tmpDir := t.TempDir()
+	path := filepath.Join(tmpDir, "config.json")
+
+	// Create original file
+	originalContent := []byte(`{"original": true}`)
+	os.WriteFile(path, originalContent, 0o644)
+
+	// Make backup directory read-only to force backup failure
+	backupDir := backupDir()
+	os.MkdirAll(backupDir, 0o755)
+	os.Chmod(backupDir, 0o444) // Read-only
+	defer os.Chmod(backupDir, 0o755)
+
+	newContent := []byte(`{"updated": true}`)
+	err := writeWithBackup(path, newContent)
+
+	// Should fail because backup couldn't be created
+	if err == nil {
+		t.Error("expected error when backup fails, got nil")
+	}
+
+	// Original file should be preserved
+	current, _ := os.ReadFile(path)
+	if string(current) != string(originalContent) {
+		t.Errorf("original file was modified despite backup failure: got %s", string(current))
+	}
+}
+
+// TestWriteWithBackup_PermissionDenied verifies clear error when target file has wrong permissions.
+// Common issue when config owned by root or wrong perms.
+func TestWriteWithBackup_PermissionDenied(t *testing.T) {
+	if runtime.GOOS == "windows" {
+		t.Skip("permission tests unreliable on Windows")
+	}
+
+	tmpDir := t.TempDir()
+
+	// Create a read-only directory
+	readOnlyDir := filepath.Join(tmpDir, "readonly")
+	os.MkdirAll(readOnlyDir, 0o755)
+	os.Chmod(readOnlyDir, 0o444)
+	defer os.Chmod(readOnlyDir, 0o755)
+
+	path := filepath.Join(readOnlyDir, "config.json")
+	err := writeWithBackup(path, []byte(`{"test": true}`))
+
+	if err == nil {
+		t.Error("expected permission error, got nil")
+	}
+}
+
+// TestWriteWithBackup_DirectoryDoesNotExist verifies behavior when target directory doesn't exist.
+// writeWithBackup doesn't create directories - caller is responsible.
+func TestWriteWithBackup_DirectoryDoesNotExist(t *testing.T) {
+	tmpDir := t.TempDir()
+	path := filepath.Join(tmpDir, "nonexistent", "subdir", "config.json")
+
+	err := writeWithBackup(path, []byte(`{"test": true}`))
+
+	// Should fail because directory doesn't exist
+	if err == nil {
+		t.Error("expected error for nonexistent directory, got nil")
+	}
+}
+
+// TestWriteWithBackup_SymlinkTarget documents behavior when target is a symlink.
+// Documents what happens if user symlinks their config file.
+func TestWriteWithBackup_SymlinkTarget(t *testing.T) {
+	if runtime.GOOS == "windows" {
+		t.Skip("symlink tests may require admin on Windows")
+	}
+
+	tmpDir := t.TempDir()
+	realFile := filepath.Join(tmpDir, "real.json")
+	symlink := filepath.Join(tmpDir, "link.json")
+
+	// Create real file and symlink
+	os.WriteFile(realFile, []byte(`{"v": 1}`), 0o644)
+	os.Symlink(realFile, symlink)
+
+	// Write through symlink
+	err := writeWithBackup(symlink, []byte(`{"v": 2}`))
+	if err != nil {
+		t.Fatalf("writeWithBackup through symlink failed: %v", err)
+	}
+
+	// The real file should be updated (symlink followed for temp file creation)
+	content, _ := os.ReadFile(symlink)
+	if string(content) != `{"v": 2}` {
+		t.Errorf("symlink target not updated correctly: got %s", string(content))
+	}
+}
+
+// TestBackupToTmp_SpecialCharsInFilename verifies backup works with special characters.
+// User may have config files with unusual names.
+func TestBackupToTmp_SpecialCharsInFilename(t *testing.T) {
+	tmpDir := t.TempDir()
+
+	// File with spaces and special chars
+	path := filepath.Join(tmpDir, "my config (backup).json")
+	os.WriteFile(path, []byte(`{"test": true}`), 0o644)
+
+	backupPath, err := backupToTmp(path)
+	if err != nil {
+		t.Fatalf("backupToTmp with special chars failed: %v", err)
+	}
+
+	// Verify backup exists and has correct content
+	content, err := os.ReadFile(backupPath)
+	if err != nil {
+		t.Fatalf("could not read backup: %v", err)
+	}
+	if string(content) != `{"test": true}` {
+		t.Errorf("backup content mismatch: got %s", string(content))
+	}
+
+	os.Remove(backupPath)
+}
+
+// TestCopyFile_PreservesPermissions verifies that copyFile preserves file permissions.
+func TestCopyFile_PreservesPermissions(t *testing.T) {
+	if runtime.GOOS == "windows" {
+		t.Skip("permission preservation tests unreliable on Windows")
+	}
+
+	tmpDir := t.TempDir()
+	src := filepath.Join(tmpDir, "src.json")
+	dst := filepath.Join(tmpDir, "dst.json")
+
+	// Create source with specific permissions
+	os.WriteFile(src, []byte(`{"test": true}`), 0o600)
+
+	err := copyFile(src, dst)
+	if err != nil {
+		t.Fatalf("copyFile failed: %v", err)
+	}
+
+	srcInfo, _ := os.Stat(src)
+	dstInfo, _ := os.Stat(dst)
+
+	if srcInfo.Mode().Perm() != dstInfo.Mode().Perm() {
+		t.Errorf("permissions not preserved: src=%v, dst=%v", srcInfo.Mode().Perm(), dstInfo.Mode().Perm())
+	}
+}
+
+// TestCopyFile_SourceNotFound verifies clear error when source doesn't exist.
+func TestCopyFile_SourceNotFound(t *testing.T) {
+	tmpDir := t.TempDir()
+	src := filepath.Join(tmpDir, "nonexistent.json")
+	dst := filepath.Join(tmpDir, "dst.json")
+
+	err := copyFile(src, dst)
+	if err == nil {
+		t.Error("expected error for nonexistent source, got nil")
+	}
+}
+
+// TestWriteWithBackup_TargetIsDirectory verifies error when path points to a directory.
+func TestWriteWithBackup_TargetIsDirectory(t *testing.T) {
+	tmpDir := t.TempDir()
+	dirPath := filepath.Join(tmpDir, "actualdir")
+	os.MkdirAll(dirPath, 0o755)
+
+	err := writeWithBackup(dirPath, []byte(`{"test": true}`))
+	if err == nil {
+		t.Error("expected error when target is a directory, got nil")
+	}
+}
+
+// TestWriteWithBackup_EmptyData verifies writing zero bytes works correctly.
+func TestWriteWithBackup_EmptyData(t *testing.T) {
+	tmpDir := t.TempDir()
+	path := filepath.Join(tmpDir, "empty.json")
+
+	err := writeWithBackup(path, []byte{})
+	if err != nil {
+		t.Fatalf("writeWithBackup with empty data failed: %v", err)
+	}
+
+	content, err := os.ReadFile(path)
+	if err != nil {
+		t.Fatalf("could not read file: %v", err)
+	}
+	if len(content) != 0 {
+		t.Errorf("expected empty file, got %d bytes", len(content))
+	}
+}
+
+// TestWriteWithBackup_FileUnreadableButDirWritable verifies behavior when existing file
+// cannot be read (for backup comparison) but directory is writable.
+func TestWriteWithBackup_FileUnreadableButDirWritable(t *testing.T) {
+	if runtime.GOOS == "windows" {
+		t.Skip("permission tests unreliable on Windows")
+	}
+
+	tmpDir := t.TempDir()
+	path := filepath.Join(tmpDir, "unreadable.json")
+
+	// Create file and make it unreadable
+	os.WriteFile(path, []byte(`{"original": true}`), 0o644)
+	os.Chmod(path, 0o000)
+	defer os.Chmod(path, 0o644)
+
+	// Should fail because we can't read the file to compare/backup
+	err := writeWithBackup(path, []byte(`{"updated": true}`))
+	if err == nil {
+		t.Error("expected error when file is unreadable, got nil")
+	}
+}
+
+// TestWriteWithBackup_RapidSuccessiveWrites verifies backup works with multiple writes
+// within the same second (timestamp collision scenario).
+func TestWriteWithBackup_RapidSuccessiveWrites(t *testing.T) {
+	tmpDir := t.TempDir()
+	path := filepath.Join(tmpDir, "rapid.json")
+
+	// Create initial file
+	os.WriteFile(path, []byte(`{"v": 0}`), 0o644)
+
+	// Rapid successive writes
+	for i := 1; i <= 3; i++ {
+		data := []byte(fmt.Sprintf(`{"v": %d}`, i))
+		if err := writeWithBackup(path, data); err != nil {
+			t.Fatalf("write %d failed: %v", i, err)
+		}
+	}
+
+	// Verify final content
+	content, _ := os.ReadFile(path)
+	if string(content) != `{"v": 3}` {
+		t.Errorf("expected final content {\"v\": 3}, got %s", string(content))
+	}
+
+	// Verify at least one backup exists
+	entries, _ := os.ReadDir(backupDir())
+	var backupCount int
+	for _, e := range entries {
+		if len(e.Name()) > len("rapid.json.") && e.Name()[:len("rapid.json.")] == "rapid.json." {
+			backupCount++
+		}
+	}
+	if backupCount == 0 {
+		t.Error("expected at least one backup file from rapid writes")
+	}
+}
+
+// TestWriteWithBackup_BackupDirIsFile verifies error when backup directory path is a file.
+func TestWriteWithBackup_BackupDirIsFile(t *testing.T) {
+	if runtime.GOOS == "windows" {
+		t.Skip("test modifies system temp directory")
+	}
+
+	// Create a file at the backup directory path
+	backupPath := backupDir()
+	// Clean up any existing directory first
+	os.RemoveAll(backupPath)
+	// Create a file instead of directory
+	os.WriteFile(backupPath, []byte("not a directory"), 0o644)
+	defer func() {
+		os.Remove(backupPath)
+		os.MkdirAll(backupPath, 0o755)
+	}()
+
+	tmpDir := t.TempDir()
+	path := filepath.Join(tmpDir, "test.json")
+	os.WriteFile(path, []byte(`{"original": true}`), 0o644)
+
+	err := writeWithBackup(path, []byte(`{"updated": true}`))
+	if err == nil {
+		t.Error("expected error when backup dir is a file, got nil")
+	}
+}
+
+// TestWriteWithBackup_NoOrphanTempFiles verifies temp files are cleaned up on failure.
+func TestWriteWithBackup_NoOrphanTempFiles(t *testing.T) {
+	if runtime.GOOS == "windows" {
+		t.Skip("permission tests unreliable on Windows")
+	}
+
+	tmpDir := t.TempDir()
+
+	// Count existing temp files
+	countTempFiles := func() int {
+		entries, _ := os.ReadDir(tmpDir)
+		count := 0
+		for _, e := range entries {
+			if len(e.Name()) > 4 && e.Name()[:4] == ".tmp" {
+				count++
+			}
+		}
+		return count
+	}
+
+	before := countTempFiles()
+
+	// Create a file, then make directory read-only to cause rename failure
+	path := filepath.Join(tmpDir, "orphan.json")
+	os.WriteFile(path, []byte(`{"v": 1}`), 0o644)
+
+	// Make a subdirectory and try to write there after making parent read-only
+	subDir := filepath.Join(tmpDir, "subdir")
+	os.MkdirAll(subDir, 0o755)
+	subPath := filepath.Join(subDir, "config.json")
+	os.WriteFile(subPath, []byte(`{"v": 1}`), 0o644)
+
+	// Make subdir read-only after creating temp file would succeed but rename would fail
+	// This is tricky to test - the temp file is created in the same dir, so if we can't
+	// rename, we also couldn't create. Let's just verify normal failure cleanup works.
+
+	// Force a failure by making the target a directory
+	badPath := filepath.Join(tmpDir, "isdir")
+	os.MkdirAll(badPath, 0o755)
+
+	_ = writeWithBackup(badPath, []byte(`{"test": true}`))
+
+	after := countTempFiles()
+	if after > before {
+		t.Errorf("orphan temp files left behind: before=%d, after=%d", before, after)
+	}
+}

cmd/config/integrations.go

@@ -0,0 +1,353 @@
+package config
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"maps"
+	"os"
+	"os/exec"
+	"runtime"
+	"slices"
+	"strings"
+	"time"
+
+	"github.com/ollama/ollama/api"
+	"github.com/spf13/cobra"
+)
+
+// Runners execute the launching of a model with the integration - claude, codex
+// Editors can edit config files (supports multi-model selection) - opencode, droid
+// They are composable interfaces where in some cases an editor is also a runner - opencode, droid
+// Runner can run an integration with a model.
+
+type Runner interface {
+	Run(model string) error
+	// String returns the human-readable name of the integration
+	String() string
+}
+
+// Editor can edit config files (supports multi-model selection)
+type Editor interface {
+	// Paths returns the paths to the config files for the integration
+	Paths() []string
+	// Edit updates the config files for the integration with the given models
+	Edit(models []string) error
+	// Models returns the models currently configured for the integration
+	// TODO(parthsareen): add error return to Models()
+	Models() []string
+}
+
+// integrations is the registry of available integrations.
+var integrations = map[string]Runner{
+	"claude":   &Claude{},
+	"codex":    &Codex{},
+	"droid":    &Droid{},
+	"opencode": &OpenCode{},
+}
+
+func selectIntegration() (string, error) {
+	if len(integrations) == 0 {
+		return "", fmt.Errorf("no integrations available")
+	}
+
+	names := slices.Sorted(maps.Keys(integrations))
+	var items []selectItem
+	for _, name := range names {
+		r := integrations[name]
+		description := r.String()
+		if conn, err := loadIntegration(name); err == nil && len(conn.Models) > 0 {
+			description = fmt.Sprintf("%s (%s)", r.String(), conn.Models[0])
+		}
+		items = append(items, selectItem{Name: name, Description: description})
+	}
+
+	return selectPrompt("Select integration:", items)
+}
+
+// selectModels lets the user select models for an integration
+func selectModels(ctx context.Context, name, current string) ([]string, error) {
+	r, ok := integrations[name]
+	if !ok {
+		return nil, fmt.Errorf("unknown integration: %s", name)
+	}
+
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		return nil, err
+	}
+
+	models, err := client.List(ctx)
+	if err != nil {
+		return nil, err
+	}
+
+	if len(models.Models) == 0 {
+		return nil, fmt.Errorf("no models available, run 'ollama pull <model>' first")
+	}
+
+	var items []selectItem
+	cloudModels := make(map[string]bool)
+	for _, m := range models.Models {
+		if m.RemoteModel != "" {
+			cloudModels[m.Name] = true
+		}
+		items = append(items, selectItem{Name: m.Name})
+	}
+
+	if len(items) == 0 {
+		return nil, fmt.Errorf("no local models available, run 'ollama pull <model>' first")
+	}
+
+	// Get previously configured models (saved config takes precedence)
+	var preChecked []string
+	if saved, err := loadIntegration(name); err == nil {
+		preChecked = saved.Models
+	} else if editor, ok := r.(Editor); ok {
+		preChecked = editor.Models()
+	}
+	checked := make(map[string]bool, len(preChecked))
+	for _, n := range preChecked {
+		checked[n] = true
+	}
+
+	// Resolve current to full name (e.g., "llama3.2" -> "llama3.2:latest")
+	for _, item := range items {
+		if item.Name == current || strings.HasPrefix(item.Name, current+":") {
+			current = item.Name
+			break
+		}
+	}
+
+	// If current model is configured, move to front of preChecked
+	if checked[current] {
+		preChecked = append([]string{current}, slices.DeleteFunc(preChecked, func(m string) bool { return m == current })...)
+	}
+
+	// Sort: checked first, then alphabetical
+	slices.SortFunc(items, func(a, b selectItem) int {
+		ac, bc := checked[a.Name], checked[b.Name]
+		if ac != bc {
+			if ac {
+				return -1
+			}
+			return 1
+		}
+		return strings.Compare(strings.ToLower(a.Name), strings.ToLower(b.Name))
+	})
+
+	var selected []string
+	// only editors support multi-model selection
+	if _, ok := r.(Editor); ok {
+		selected, err = multiSelectPrompt(fmt.Sprintf("Select models for %s:", r), items, preChecked)
+		if err != nil {
+			return nil, err
+		}
+	} else {
+		model, err := selectPrompt(fmt.Sprintf("Select model for %s:", r), items)
+		if err != nil {
+			return nil, err
+		}
+		selected = []string{model}
+	}
+
+	// if any model in selected is a cloud model, ensure signed in
+	var selectedCloudModels []string
+	for _, m := range selected {
+		if cloudModels[m] {
+			selectedCloudModels = append(selectedCloudModels, m)
+		}
+	}
+	if len(selectedCloudModels) > 0 {
+		// ensure user is signed in
+		user, err := client.Whoami(ctx)
+		if err == nil && user != nil && user.Name != "" {
+			return selected, nil
+		}
+
+		var aErr api.AuthorizationError
+		if !errors.As(err, &aErr) || aErr.SigninURL == "" {
+			return nil, err
+		}
+
+		modelList := strings.Join(selectedCloudModels, ", ")
+		yes, err := confirmPrompt(fmt.Sprintf("sign in to use %s?", modelList))
+		if err != nil || !yes {
+			return nil, fmt.Errorf("%s requires sign in", modelList)
+		}
+
+		fmt.Fprintf(os.Stderr, "\nTo sign in, navigate to:\n    %s\n\n", aErr.SigninURL)
+
+		// TODO(parthsareen): extract into auth package for cmd
+		// Auto-open browser (best effort, fail silently)
+		switch runtime.GOOS {
+		case "darwin":
+			_ = exec.Command("open", aErr.SigninURL).Start()
+		case "linux":
+			_ = exec.Command("xdg-open", aErr.SigninURL).Start()
+		case "windows":
+			_ = exec.Command("rundll32", "url.dll,FileProtocolHandler", aErr.SigninURL).Start()
+		}
+
+		spinnerFrames := []string{"|", "/", "-", "\\"}
+		frame := 0
+
+		fmt.Fprintf(os.Stderr, "\033[90mwaiting for sign in to complete... %s\033[0m", spinnerFrames[0])
+
+		ticker := time.NewTicker(200 * time.Millisecond)
+		defer ticker.Stop()
+
+		for {
+			select {
+			case <-ctx.Done():
+				fmt.Fprintf(os.Stderr, "\r\033[K")
+				return nil, ctx.Err()
+			case <-ticker.C:
+				frame++
+				fmt.Fprintf(os.Stderr, "\r\033[90mwaiting for sign in to complete... %s\033[0m", spinnerFrames[frame%len(spinnerFrames)])
+
+				// poll every 10th frame (~2 seconds)
+				if frame%10 == 0 {
+					u, err := client.Whoami(ctx)
+					if err == nil && u != nil && u.Name != "" {
+						fmt.Fprintf(os.Stderr, "\r\033[K\033[A\r\033[K\033[1msigned in:\033[0m %s\n", u.Name)
+						return selected, nil
+					}
+				}
+			}
+		}
+	}
+
+	return selected, nil
+}
+
+func runIntegration(name, modelName string) error {
+	r, ok := integrations[name]
+	if !ok {
+		return fmt.Errorf("unknown integration: %s", name)
+	}
+	fmt.Fprintf(os.Stderr, "\nLaunching %s with %s...\n", r, modelName)
+	return r.Run(modelName)
+}
+
+// LaunchCmd returns the cobra command for launching integrations.
+func LaunchCmd(checkServerHeartbeat func(cmd *cobra.Command, args []string) error) *cobra.Command {
+	var modelFlag string
+	var configFlag bool
+
+	cmd := &cobra.Command{
+		Use:   "launch [INTEGRATION]",
+		Short: "Launch an integration with Ollama",
+		Long: `Launch an integration configured with Ollama models.
+
+Supported integrations:
+  claude    Claude Code
+  codex     Codex
+  droid     Droid
+  opencode  OpenCode
+
+Examples:
+  ollama launch
+  ollama launch claude
+  ollama launch claude --model <model>
+  ollama launch droid --config (does not auto-launch)`,
+		Args:    cobra.MaximumNArgs(1),
+		PreRunE: checkServerHeartbeat,
+		RunE: func(cmd *cobra.Command, args []string) error {
+			var name string
+			if len(args) > 0 {
+				name = args[0]
+			} else {
+				var err error
+				name, err = selectIntegration()
+				if errors.Is(err, errCancelled) {
+					return nil
+				}
+				if err != nil {
+					return err
+				}
+			}
+
+			r, ok := integrations[strings.ToLower(name)]
+			if !ok {
+				return fmt.Errorf("unknown integration: %s", name)
+			}
+
+			// If launching without --model, use saved config if available
+			if !configFlag && modelFlag == "" {
+				if config, err := loadIntegration(name); err == nil && len(config.Models) > 0 {
+					return runIntegration(name, config.Models[0])
+				}
+			}
+
+			var models []string
+			if modelFlag != "" {
+				// When --model is specified, merge with existing models (new model becomes default)
+				models = []string{modelFlag}
+				if existing, err := loadIntegration(name); err == nil && len(existing.Models) > 0 {
+					for _, m := range existing.Models {
+						if m != modelFlag {
+							models = append(models, m)
+						}
+					}
+				}
+			} else {
+				var err error
+				models, err = selectModels(cmd.Context(), name, "")
+				if errors.Is(err, errCancelled) {
+					return nil
+				}
+				if err != nil {
+					return err
+				}
+			}
+
+			if editor, isEditor := r.(Editor); isEditor {
+				paths := editor.Paths()
+				if len(paths) > 0 {
+					fmt.Fprintf(os.Stderr, "This will modify your %s configuration:\n", r)
+					for _, p := range paths {
+						fmt.Fprintf(os.Stderr, "  %s\n", p)
+					}
+					fmt.Fprintf(os.Stderr, "Backups will be saved to %s/\n\n", backupDir())
+
+					if ok, _ := confirmPrompt("Proceed?"); !ok {
+						return nil
+					}
+				}
+			}
+
+			if err := saveIntegration(name, models); err != nil {
+				return fmt.Errorf("failed to save: %w", err)
+			}
+
+			if editor, isEditor := r.(Editor); isEditor {
+				if err := editor.Edit(models); err != nil {
+					return fmt.Errorf("setup failed: %w", err)
+				}
+			}
+
+			if _, isEditor := r.(Editor); isEditor {
+				if len(models) == 1 {
+					fmt.Fprintf(os.Stderr, "Added %s to %s\n", models[0], r)
+				} else {
+					fmt.Fprintf(os.Stderr, "Added %d models to %s (default: %s)\n", len(models), r, models[0])
+				}
+			}
+
+			if configFlag {
+				if launch, _ := confirmPrompt(fmt.Sprintf("\nLaunch %s now?", r)); launch {
+					return runIntegration(name, models[0])
+				}
+				fmt.Fprintf(os.Stderr, "Run 'ollama launch %s' to start with %s\n", strings.ToLower(name), models[0])
+				return nil
+			}
+
+			return runIntegration(name, models[0])
+		},
+	}
+
+	cmd.Flags().StringVar(&modelFlag, "model", "", "Model to use")
+	cmd.Flags().BoolVar(&configFlag, "config", false, "Configure without launching")
+	return cmd
+}

cmd/config/integrations_test.go

@@ -0,0 +1,188 @@
+package config
+
+import (
+	"slices"
+	"strings"
+	"testing"
+
+	"github.com/spf13/cobra"
+)
+
+func TestIntegrationLookup(t *testing.T) {
+	tests := []struct {
+		name      string
+		input     string
+		wantFound bool
+		wantName  string
+	}{
+		{"claude lowercase", "claude", true, "Claude Code"},
+		{"claude uppercase", "CLAUDE", true, "Claude Code"},
+		{"claude mixed case", "Claude", true, "Claude Code"},
+		{"codex", "codex", true, "Codex"},
+		{"droid", "droid", true, "Droid"},
+		{"opencode", "opencode", true, "OpenCode"},
+		{"unknown integration", "unknown", false, ""},
+		{"empty string", "", false, ""},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			r, found := integrations[strings.ToLower(tt.input)]
+			if found != tt.wantFound {
+				t.Errorf("integrations[%q] found = %v, want %v", tt.input, found, tt.wantFound)
+			}
+			if found && r.String() != tt.wantName {
+				t.Errorf("integrations[%q].String() = %q, want %q", tt.input, r.String(), tt.wantName)
+			}
+		})
+	}
+}
+
+func TestIntegrationRegistry(t *testing.T) {
+	expectedIntegrations := []string{"claude", "codex", "droid", "opencode"}
+
+	for _, name := range expectedIntegrations {
+		t.Run(name, func(t *testing.T) {
+			r, ok := integrations[name]
+			if !ok {
+				t.Fatalf("integration %q not found in registry", name)
+			}
+			if r.String() == "" {
+				t.Error("integration.String() should not be empty")
+			}
+		})
+	}
+}
+
+func TestHasLocalModel(t *testing.T) {
+	tests := []struct {
+		name   string
+		models []string
+		want   bool
+	}{
+		{"empty list", []string{}, false},
+		{"single local model", []string{"llama3.2"}, true},
+		{"single cloud model", []string{"cloud-model"}, false},
+		{"mixed models", []string{"cloud-model", "llama3.2"}, true},
+		{"multiple local models", []string{"llama3.2", "qwen2.5"}, true},
+		{"multiple cloud models", []string{"cloud-a", "cloud-b"}, false},
+		{"local model first", []string{"llama3.2", "cloud-model"}, true},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			got := slices.ContainsFunc(tt.models, func(m string) bool {
+				return !strings.Contains(m, "cloud")
+			})
+			if got != tt.want {
+				t.Errorf("hasLocalModel(%v) = %v, want %v", tt.models, got, tt.want)
+			}
+		})
+	}
+}
+
+func TestLaunchCmd(t *testing.T) {
+	// Mock checkServerHeartbeat that always succeeds
+	mockCheck := func(cmd *cobra.Command, args []string) error {
+		return nil
+	}
+
+	cmd := LaunchCmd(mockCheck)
+
+	t.Run("command structure", func(t *testing.T) {
+		if cmd.Use != "launch [INTEGRATION]" {
+			t.Errorf("Use = %q, want %q", cmd.Use, "launch [INTEGRATION]")
+		}
+		if cmd.Short == "" {
+			t.Error("Short description should not be empty")
+		}
+		if cmd.Long == "" {
+			t.Error("Long description should not be empty")
+		}
+	})
+
+	t.Run("flags exist", func(t *testing.T) {
+		modelFlag := cmd.Flags().Lookup("model")
+		if modelFlag == nil {
+			t.Error("--model flag should exist")
+		}
+
+		configFlag := cmd.Flags().Lookup("config")
+		if configFlag == nil {
+			t.Error("--config flag should exist")
+		}
+	})
+
+	t.Run("PreRunE is set", func(t *testing.T) {
+		if cmd.PreRunE == nil {
+			t.Error("PreRunE should be set to checkServerHeartbeat")
+		}
+	})
+}
+
+func TestRunIntegration_UnknownIntegration(t *testing.T) {
+	err := runIntegration("unknown-integration", "model")
+	if err == nil {
+		t.Error("expected error for unknown integration, got nil")
+	}
+	if !strings.Contains(err.Error(), "unknown integration") {
+		t.Errorf("error should mention 'unknown integration', got: %v", err)
+	}
+}
+
+func TestHasLocalModel_DocumentsHeuristic(t *testing.T) {
+	tests := []struct {
+		name   string
+		models []string
+		want   bool
+		reason string
+	}{
+		{"empty list", []string{}, false, "empty list has no local models"},
+		{"contains-cloud-substring", []string{"deepseek-r1:cloud"}, false, "model with 'cloud' substring is considered cloud"},
+		{"cloud-in-name", []string{"my-cloud-model"}, false, "'cloud' anywhere in name = cloud model"},
+		{"cloudless", []string{"cloudless-model"}, false, "'cloudless' still contains 'cloud'"},
+		{"local-model", []string{"llama3.2"}, true, "no 'cloud' = local"},
+		{"mixed", []string{"cloud-model", "llama3.2"}, true, "one local model = hasLocalModel true"},
+		{"all-cloud", []string{"cloud-a", "cloud-b"}, false, "all contain 'cloud'"},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			got := slices.ContainsFunc(tt.models, func(m string) bool {
+				return !strings.Contains(m, "cloud")
+			})
+			if got != tt.want {
+				t.Errorf("hasLocalModel(%v) = %v, want %v (%s)", tt.models, got, tt.want, tt.reason)
+			}
+		})
+	}
+}
+
+func TestLaunchCmd_NilHeartbeat(t *testing.T) {
+	// This should not panic - cmd creation should work even with nil
+	cmd := LaunchCmd(nil)
+	if cmd == nil {
+		t.Fatal("LaunchCmd returned nil")
+	}
+
+	// PreRunE should be nil when passed nil
+	if cmd.PreRunE != nil {
+		t.Log("Note: PreRunE is set even when nil is passed (acceptable)")
+	}
+}
+
+func TestAllIntegrations_HaveRequiredMethods(t *testing.T) {
+	for name, r := range integrations {
+		t.Run(name, func(t *testing.T) {
+			// Test String() doesn't panic and returns non-empty
+			displayName := r.String()
+			if displayName == "" {
+				t.Error("String() should not return empty")
+			}
+
+			// Test Run() exists (we can't call it without actually running the command)
+			// Just verify the method is available
+			var _ func(string) error = r.Run
+		})
+	}
+}

cmd/config/opencode.go

@@ -0,0 +1,203 @@
+package config
+
+import (
+	"encoding/json"
+	"fmt"
+	"maps"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"slices"
+	"strings"
+)
+
+// OpenCode implements Runner and Editor for OpenCode integration
+type OpenCode struct{}
+
+func (o *OpenCode) String() string { return "OpenCode" }
+
+func (o *OpenCode) Run(model string) error {
+	if _, err := exec.LookPath("opencode"); err != nil {
+		return fmt.Errorf("opencode is not installed, install from https://opencode.ai")
+	}
+
+	// Call Edit() to ensure config is up-to-date before launch
+	models := []string{model}
+	if config, err := loadIntegration("opencode"); err == nil && len(config.Models) > 0 {
+		models = config.Models
+	}
+	if err := o.Edit(models); err != nil {
+		return fmt.Errorf("setup failed: %w", err)
+	}
+
+	cmd := exec.Command("opencode")
+	cmd.Stdin = os.Stdin
+	cmd.Stdout = os.Stdout
+	cmd.Stderr = os.Stderr
+	return cmd.Run()
+}
+
+func (o *OpenCode) Paths() []string {
+	home, err := os.UserHomeDir()
+	if err != nil {
+		return nil
+	}
+
+	var paths []string
+	p := filepath.Join(home, ".config", "opencode", "opencode.json")
+	if _, err := os.Stat(p); err == nil {
+		paths = append(paths, p)
+	}
+	sp := filepath.Join(home, ".local", "state", "opencode", "model.json")
+	if _, err := os.Stat(sp); err == nil {
+		paths = append(paths, sp)
+	}
+	return paths
+}
+
+func (o *OpenCode) Edit(modelList []string) error {
+	if len(modelList) == 0 {
+		return nil
+	}
+
+	home, err := os.UserHomeDir()
+	if err != nil {
+		return err
+	}
+
+	configPath := filepath.Join(home, ".config", "opencode", "opencode.json")
+	if err := os.MkdirAll(filepath.Dir(configPath), 0o755); err != nil {
+		return err
+	}
+
+	config := make(map[string]any)
+	if data, err := os.ReadFile(configPath); err == nil {
+		_ = json.Unmarshal(data, &config) // Ignore parse errors; treat missing/corrupt files as empty
+	}
+
+	config["$schema"] = "https://opencode.ai/config.json"
+
+	provider, ok := config["provider"].(map[string]any)
+	if !ok {
+		provider = make(map[string]any)
+	}
+
+	ollama, ok := provider["ollama"].(map[string]any)
+	if !ok {
+		ollama = map[string]any{
+			"npm":  "@ai-sdk/openai-compatible",
+			"name": "Ollama (local)",
+			"options": map[string]any{
+				"baseURL": "http://localhost:11434/v1",
+			},
+		}
+	}
+
+	models, ok := ollama["models"].(map[string]any)
+	if !ok {
+		models = make(map[string]any)
+	}
+
+	selectedSet := make(map[string]bool)
+	for _, m := range modelList {
+		selectedSet[m] = true
+	}
+
+	for name, cfg := range models {
+		if cfgMap, ok := cfg.(map[string]any); ok {
+			if displayName, ok := cfgMap["name"].(string); ok {
+				if strings.HasSuffix(displayName, "[Ollama]") && !selectedSet[name] {
+					delete(models, name)
+				}
+			}
+		}
+	}
+
+	for _, model := range modelList {
+		models[model] = map[string]any{
+			"name": fmt.Sprintf("%s [Ollama]", model),
+		}
+	}
+
+	ollama["models"] = models
+	provider["ollama"] = ollama
+	config["provider"] = provider
+
+	configData, err := json.MarshalIndent(config, "", "  ")
+	if err != nil {
+		return err
+	}
+	if err := writeWithBackup(configPath, configData); err != nil {
+		return err
+	}
+
+	statePath := filepath.Join(home, ".local", "state", "opencode", "model.json")
+	if err := os.MkdirAll(filepath.Dir(statePath), 0o755); err != nil {
+		return err
+	}
+
+	state := map[string]any{
+		"recent":   []any{},
+		"favorite": []any{},
+		"variant":  map[string]any{},
+	}
+	if data, err := os.ReadFile(statePath); err == nil {
+		_ = json.Unmarshal(data, &state) // Ignore parse errors; use defaults
+	}
+
+	recent, _ := state["recent"].([]any)
+
+	modelSet := make(map[string]bool)
+	for _, m := range modelList {
+		modelSet[m] = true
+	}
+
+	// Filter out existing Ollama models we're about to re-add
+	newRecent := slices.DeleteFunc(slices.Clone(recent), func(entry any) bool {
+		e, ok := entry.(map[string]any)
+		if !ok || e["providerID"] != "ollama" {
+			return false
+		}
+		modelID, _ := e["modelID"].(string)
+		return modelSet[modelID]
+	})
+
+	// Prepend models in reverse order so first model ends up first
+	for _, model := range slices.Backward(modelList) {
+		newRecent = slices.Insert(newRecent, 0, any(map[string]any{
+			"providerID": "ollama",
+			"modelID":    model,
+		}))
+	}
+
+	const maxRecentModels = 10
+	newRecent = newRecent[:min(len(newRecent), maxRecentModels)]
+
+	state["recent"] = newRecent
+
+	stateData, err := json.MarshalIndent(state, "", "  ")
+	if err != nil {
+		return err
+	}
+	return writeWithBackup(statePath, stateData)
+}
+
+func (o *OpenCode) Models() []string {
+	home, err := os.UserHomeDir()
+	if err != nil {
+		return nil
+	}
+	config, err := readJSONFile(filepath.Join(home, ".config", "opencode", "opencode.json"))
+	if err != nil {
+		return nil
+	}
+	provider, _ := config["provider"].(map[string]any)
+	ollama, _ := provider["ollama"].(map[string]any)
+	models, _ := ollama["models"].(map[string]any)
+	if len(models) == 0 {
+		return nil
+	}
+	keys := slices.Collect(maps.Keys(models))
+	slices.Sort(keys)
+	return keys
+}

cmd/config/opencode_test.go

@@ -0,0 +1,437 @@
+package config
+
+import (
+	"encoding/json"
+	"os"
+	"path/filepath"
+	"testing"
+)
+
+func TestOpenCodeIntegration(t *testing.T) {
+	o := &OpenCode{}
+
+	t.Run("String", func(t *testing.T) {
+		if got := o.String(); got != "OpenCode" {
+			t.Errorf("String() = %q, want %q", got, "OpenCode")
+		}
+	})
+
+	t.Run("implements Runner", func(t *testing.T) {
+		var _ Runner = o
+	})
+
+	t.Run("implements Editor", func(t *testing.T) {
+		var _ Editor = o
+	})
+}
+
+func TestOpenCodeEdit(t *testing.T) {
+	o := &OpenCode{}
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	configDir := filepath.Join(tmpDir, ".config", "opencode")
+	configPath := filepath.Join(configDir, "opencode.json")
+	stateDir := filepath.Join(tmpDir, ".local", "state", "opencode")
+	statePath := filepath.Join(stateDir, "model.json")
+
+	cleanup := func() {
+		os.RemoveAll(configDir)
+		os.RemoveAll(stateDir)
+	}
+
+	t.Run("fresh install", func(t *testing.T) {
+		cleanup()
+		if err := o.Edit([]string{"llama3.2"}); err != nil {
+			t.Fatal(err)
+		}
+		assertOpenCodeModelExists(t, configPath, "llama3.2")
+		assertOpenCodeRecentModel(t, statePath, 0, "ollama", "llama3.2")
+	})
+
+	t.Run("preserve other providers", func(t *testing.T) {
+		cleanup()
+		os.MkdirAll(configDir, 0o755)
+		os.WriteFile(configPath, []byte(`{"provider":{"anthropic":{"apiKey":"xxx"}}}`), 0o644)
+		if err := o.Edit([]string{"llama3.2"}); err != nil {
+			t.Fatal(err)
+		}
+		data, _ := os.ReadFile(configPath)
+		var cfg map[string]any
+		json.Unmarshal(data, &cfg)
+		provider := cfg["provider"].(map[string]any)
+		if provider["anthropic"] == nil {
+			t.Error("anthropic provider was removed")
+		}
+		assertOpenCodeModelExists(t, configPath, "llama3.2")
+	})
+
+	t.Run("preserve other models", func(t *testing.T) {
+		cleanup()
+		os.MkdirAll(configDir, 0o755)
+		os.WriteFile(configPath, []byte(`{"provider":{"ollama":{"models":{"mistral":{"name":"Mistral"}}}}}`), 0o644)
+		if err := o.Edit([]string{"llama3.2"}); err != nil {
+			t.Fatal(err)
+		}
+		assertOpenCodeModelExists(t, configPath, "mistral")
+		assertOpenCodeModelExists(t, configPath, "llama3.2")
+	})
+
+	t.Run("update existing model", func(t *testing.T) {
+		cleanup()
+		o.Edit([]string{"llama3.2"})
+		o.Edit([]string{"llama3.2"})
+		assertOpenCodeModelExists(t, configPath, "llama3.2")
+	})
+
+	t.Run("preserve top-level keys", func(t *testing.T) {
+		cleanup()
+		os.MkdirAll(configDir, 0o755)
+		os.WriteFile(configPath, []byte(`{"theme":"dark","keybindings":{}}`), 0o644)
+		if err := o.Edit([]string{"llama3.2"}); err != nil {
+			t.Fatal(err)
+		}
+		data, _ := os.ReadFile(configPath)
+		var cfg map[string]any
+		json.Unmarshal(data, &cfg)
+		if cfg["theme"] != "dark" {
+			t.Error("theme was removed")
+		}
+		if cfg["keybindings"] == nil {
+			t.Error("keybindings was removed")
+		}
+	})
+
+	t.Run("model state - insert at index 0", func(t *testing.T) {
+		cleanup()
+		os.MkdirAll(stateDir, 0o755)
+		os.WriteFile(statePath, []byte(`{"recent":[{"providerID":"anthropic","modelID":"claude"}],"favorite":[],"variant":{}}`), 0o644)
+		if err := o.Edit([]string{"llama3.2"}); err != nil {
+			t.Fatal(err)
+		}
+		assertOpenCodeRecentModel(t, statePath, 0, "ollama", "llama3.2")
+		assertOpenCodeRecentModel(t, statePath, 1, "anthropic", "claude")
+	})
+
+	t.Run("model state - preserve favorites and variants", func(t *testing.T) {
+		cleanup()
+		os.MkdirAll(stateDir, 0o755)
+		os.WriteFile(statePath, []byte(`{"recent":[],"favorite":[{"providerID":"x","modelID":"y"}],"variant":{"a":"b"}}`), 0o644)
+		if err := o.Edit([]string{"llama3.2"}); err != nil {
+			t.Fatal(err)
+		}
+		data, _ := os.ReadFile(statePath)
+		var state map[string]any
+		json.Unmarshal(data, &state)
+		if len(state["favorite"].([]any)) != 1 {
+			t.Error("favorite was modified")
+		}
+		if state["variant"].(map[string]any)["a"] != "b" {
+			t.Error("variant was modified")
+		}
+	})
+
+	t.Run("model state - deduplicate on re-add", func(t *testing.T) {
+		cleanup()
+		os.MkdirAll(stateDir, 0o755)
+		os.WriteFile(statePath, []byte(`{"recent":[{"providerID":"ollama","modelID":"llama3.2"},{"providerID":"anthropic","modelID":"claude"}],"favorite":[],"variant":{}}`), 0o644)
+		if err := o.Edit([]string{"llama3.2"}); err != nil {
+			t.Fatal(err)
+		}
+		data, _ := os.ReadFile(statePath)
+		var state map[string]any
+		json.Unmarshal(data, &state)
+		recent := state["recent"].([]any)
+		if len(recent) != 2 {
+			t.Errorf("expected 2 recent entries, got %d", len(recent))
+		}
+		assertOpenCodeRecentModel(t, statePath, 0, "ollama", "llama3.2")
+	})
+
+	t.Run("remove model", func(t *testing.T) {
+		cleanup()
+		// First add two models
+		o.Edit([]string{"llama3.2", "mistral"})
+		assertOpenCodeModelExists(t, configPath, "llama3.2")
+		assertOpenCodeModelExists(t, configPath, "mistral")
+
+		// Then remove one by only selecting the other
+		o.Edit([]string{"llama3.2"})
+		assertOpenCodeModelExists(t, configPath, "llama3.2")
+		assertOpenCodeModelNotExists(t, configPath, "mistral")
+	})
+
+	t.Run("remove model preserves non-ollama models", func(t *testing.T) {
+		cleanup()
+		os.MkdirAll(configDir, 0o755)
+		// Add a non-Ollama model manually
+		os.WriteFile(configPath, []byte(`{"provider":{"ollama":{"models":{"external":{"name":"External Model"}}}}}`), 0o644)
+
+		o.Edit([]string{"llama3.2"})
+		assertOpenCodeModelExists(t, configPath, "llama3.2")
+		assertOpenCodeModelExists(t, configPath, "external") // Should be preserved
+	})
+}
+
+func assertOpenCodeModelExists(t *testing.T, path, model string) {
+	t.Helper()
+	data, err := os.ReadFile(path)
+	if err != nil {
+		t.Fatal(err)
+	}
+	var cfg map[string]any
+	if err := json.Unmarshal(data, &cfg); err != nil {
+		t.Fatal(err)
+	}
+	provider, ok := cfg["provider"].(map[string]any)
+	if !ok {
+		t.Fatal("provider not found")
+	}
+	ollama, ok := provider["ollama"].(map[string]any)
+	if !ok {
+		t.Fatal("ollama provider not found")
+	}
+	models, ok := ollama["models"].(map[string]any)
+	if !ok {
+		t.Fatal("models not found")
+	}
+	if models[model] == nil {
+		t.Errorf("model %s not found", model)
+	}
+}
+
+func assertOpenCodeModelNotExists(t *testing.T, path, model string) {
+	t.Helper()
+	data, err := os.ReadFile(path)
+	if err != nil {
+		t.Fatal(err)
+	}
+	var cfg map[string]any
+	if err := json.Unmarshal(data, &cfg); err != nil {
+		t.Fatal(err)
+	}
+	provider, ok := cfg["provider"].(map[string]any)
+	if !ok {
+		return // No provider means no model
+	}
+	ollama, ok := provider["ollama"].(map[string]any)
+	if !ok {
+		return // No ollama means no model
+	}
+	models, ok := ollama["models"].(map[string]any)
+	if !ok {
+		return // No models means no model
+	}
+	if models[model] != nil {
+		t.Errorf("model %s should not exist but was found", model)
+	}
+}
+
+func assertOpenCodeRecentModel(t *testing.T, path string, index int, providerID, modelID string) {
+	t.Helper()
+	data, err := os.ReadFile(path)
+	if err != nil {
+		t.Fatal(err)
+	}
+	var state map[string]any
+	if err := json.Unmarshal(data, &state); err != nil {
+		t.Fatal(err)
+	}
+	recent, ok := state["recent"].([]any)
+	if !ok {
+		t.Fatal("recent not found")
+	}
+	if index >= len(recent) {
+		t.Fatalf("index %d out of range (len=%d)", index, len(recent))
+	}
+	entry, ok := recent[index].(map[string]any)
+	if !ok {
+		t.Fatal("entry is not a map")
+	}
+	if entry["providerID"] != providerID {
+		t.Errorf("expected providerID %s, got %s", providerID, entry["providerID"])
+	}
+	if entry["modelID"] != modelID {
+		t.Errorf("expected modelID %s, got %s", modelID, entry["modelID"])
+	}
+}
+
+// Edge case tests for opencode.go
+
+func TestOpenCodeEdit_CorruptedConfigJSON(t *testing.T) {
+	o := &OpenCode{}
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	configDir := filepath.Join(tmpDir, ".config", "opencode")
+	configPath := filepath.Join(configDir, "opencode.json")
+
+	os.MkdirAll(configDir, 0o755)
+	os.WriteFile(configPath, []byte(`{corrupted json content`), 0o644)
+
+	// Should not panic - corrupted JSON should be treated as empty
+	err := o.Edit([]string{"llama3.2"})
+	if err != nil {
+		t.Fatalf("Edit failed with corrupted config: %v", err)
+	}
+
+	// Verify valid JSON was created
+	data, _ := os.ReadFile(configPath)
+	var cfg map[string]any
+	if err := json.Unmarshal(data, &cfg); err != nil {
+		t.Errorf("resulting config is not valid JSON: %v", err)
+	}
+}
+
+func TestOpenCodeEdit_CorruptedStateJSON(t *testing.T) {
+	o := &OpenCode{}
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	stateDir := filepath.Join(tmpDir, ".local", "state", "opencode")
+	statePath := filepath.Join(stateDir, "model.json")
+
+	os.MkdirAll(stateDir, 0o755)
+	os.WriteFile(statePath, []byte(`{corrupted state`), 0o644)
+
+	err := o.Edit([]string{"llama3.2"})
+	if err != nil {
+		t.Fatalf("Edit failed with corrupted state: %v", err)
+	}
+
+	// Verify valid state was created
+	data, _ := os.ReadFile(statePath)
+	var state map[string]any
+	if err := json.Unmarshal(data, &state); err != nil {
+		t.Errorf("resulting state is not valid JSON: %v", err)
+	}
+}
+
+func TestOpenCodeEdit_WrongTypeProvider(t *testing.T) {
+	o := &OpenCode{}
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	configDir := filepath.Join(tmpDir, ".config", "opencode")
+	configPath := filepath.Join(configDir, "opencode.json")
+
+	os.MkdirAll(configDir, 0o755)
+	os.WriteFile(configPath, []byte(`{"provider": "not a map"}`), 0o644)
+
+	err := o.Edit([]string{"llama3.2"})
+	if err != nil {
+		t.Fatalf("Edit with wrong type provider failed: %v", err)
+	}
+
+	// Verify provider is now correct type
+	data, _ := os.ReadFile(configPath)
+	var cfg map[string]any
+	json.Unmarshal(data, &cfg)
+
+	provider, ok := cfg["provider"].(map[string]any)
+	if !ok {
+		t.Fatalf("provider should be map after setup, got %T", cfg["provider"])
+	}
+	if provider["ollama"] == nil {
+		t.Error("ollama provider should be created")
+	}
+}
+
+func TestOpenCodeEdit_WrongTypeRecent(t *testing.T) {
+	o := &OpenCode{}
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	stateDir := filepath.Join(tmpDir, ".local", "state", "opencode")
+	statePath := filepath.Join(stateDir, "model.json")
+
+	os.MkdirAll(stateDir, 0o755)
+	os.WriteFile(statePath, []byte(`{"recent": "not an array", "favorite": [], "variant": {}}`), 0o644)
+
+	err := o.Edit([]string{"llama3.2"})
+	if err != nil {
+		t.Fatalf("Edit with wrong type recent failed: %v", err)
+	}
+
+	// The function should handle this gracefully
+	data, _ := os.ReadFile(statePath)
+	var state map[string]any
+	json.Unmarshal(data, &state)
+
+	// recent should be properly set after setup
+	recent, ok := state["recent"].([]any)
+	if !ok {
+		t.Logf("Note: recent type after setup is %T (documenting behavior)", state["recent"])
+	} else if len(recent) == 0 {
+		t.Logf("Note: recent is empty (documenting behavior)")
+	}
+}
+
+func TestOpenCodeEdit_EmptyModels(t *testing.T) {
+	o := &OpenCode{}
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	configDir := filepath.Join(tmpDir, ".config", "opencode")
+	configPath := filepath.Join(configDir, "opencode.json")
+
+	os.MkdirAll(configDir, 0o755)
+	originalContent := `{"provider":{"ollama":{"models":{"existing":{}}}}}`
+	os.WriteFile(configPath, []byte(originalContent), 0o644)
+
+	// Empty models should be no-op
+	err := o.Edit([]string{})
+	if err != nil {
+		t.Fatalf("Edit with empty models failed: %v", err)
+	}
+
+	// Original content should be preserved (file not modified)
+	data, _ := os.ReadFile(configPath)
+	if string(data) != originalContent {
+		t.Errorf("empty models should not modify file, but content changed")
+	}
+}
+
+func TestOpenCodeEdit_SpecialCharsInModelName(t *testing.T) {
+	o := &OpenCode{}
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	// Model name with special characters (though unusual)
+	specialModel := `model-with-"quotes"`
+
+	err := o.Edit([]string{specialModel})
+	if err != nil {
+		t.Fatalf("Edit with special chars failed: %v", err)
+	}
+
+	// Verify it was stored correctly
+	configDir := filepath.Join(tmpDir, ".config", "opencode")
+	configPath := filepath.Join(configDir, "opencode.json")
+	data, _ := os.ReadFile(configPath)
+
+	var cfg map[string]any
+	if err := json.Unmarshal(data, &cfg); err != nil {
+		t.Fatalf("resulting config is invalid JSON: %v", err)
+	}
+
+	// Model should be accessible
+	provider, _ := cfg["provider"].(map[string]any)
+	ollama, _ := provider["ollama"].(map[string]any)
+	models, _ := ollama["models"].(map[string]any)
+
+	if models[specialModel] == nil {
+		t.Errorf("model with special chars not found in config")
+	}
+}
+
+func TestOpenCodeModels_NoConfig(t *testing.T) {
+	o := &OpenCode{}
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	models := o.Models()
+	if len(models) > 0 {
+		t.Errorf("expected nil/empty for missing config, got %v", models)
+	}
+}

cmd/config/selector.go

@@ -0,0 +1,499 @@
+package config
+
+import (
+	"errors"
+	"fmt"
+	"io"
+	"os"
+	"strings"
+
+	"golang.org/x/term"
+)
+
+// ANSI escape sequences for terminal formatting.
+const (
+	ansiHideCursor = "\033[?25l"
+	ansiShowCursor = "\033[?25h"
+	ansiBold       = "\033[1m"
+	ansiReset      = "\033[0m"
+	ansiGray       = "\033[37m"
+	ansiClearDown  = "\033[J"
+)
+
+const maxDisplayedItems = 10
+
+var errCancelled = errors.New("cancelled")
+
+type selectItem struct {
+	Name        string
+	Description string
+}
+
+type inputEvent int
+
+const (
+	eventNone inputEvent = iota
+	eventEnter
+	eventEscape
+	eventUp
+	eventDown
+	eventTab
+	eventBackspace
+	eventChar
+)
+
+type selectState struct {
+	items        []selectItem
+	filter       string
+	selected     int
+	scrollOffset int
+}
+
+func newSelectState(items []selectItem) *selectState {
+	return &selectState{items: items}
+}
+
+func (s *selectState) filtered() []selectItem {
+	return filterItems(s.items, s.filter)
+}
+
+func (s *selectState) handleInput(event inputEvent, char byte) (done bool, result string, err error) {
+	filtered := s.filtered()
+
+	switch event {
+	case eventEnter:
+		if len(filtered) > 0 && s.selected < len(filtered) {
+			return true, filtered[s.selected].Name, nil
+		}
+	case eventEscape:
+		return true, "", errCancelled
+	case eventBackspace:
+		if len(s.filter) > 0 {
+			s.filter = s.filter[:len(s.filter)-1]
+			s.selected = 0
+			s.scrollOffset = 0
+		}
+	case eventUp:
+		if s.selected > 0 {
+			s.selected--
+			if s.selected < s.scrollOffset {
+				s.scrollOffset = s.selected
+			}
+		}
+	case eventDown:
+		if s.selected < len(filtered)-1 {
+			s.selected++
+			if s.selected >= s.scrollOffset+maxDisplayedItems {
+				s.scrollOffset = s.selected - maxDisplayedItems + 1
+			}
+		}
+	case eventChar:
+		s.filter += string(char)
+		s.selected = 0
+		s.scrollOffset = 0
+	}
+
+	return false, "", nil
+}
+
+type multiSelectState struct {
+	items         []selectItem
+	itemIndex     map[string]int
+	filter        string
+	highlighted   int
+	scrollOffset  int
+	checked       map[int]bool
+	checkOrder    []int
+	focusOnButton bool
+}
+
+func newMultiSelectState(items []selectItem, preChecked []string) *multiSelectState {
+	s := &multiSelectState{
+		items:     items,
+		itemIndex: make(map[string]int, len(items)),
+		checked:   make(map[int]bool),
+	}
+
+	for i, item := range items {
+		s.itemIndex[item.Name] = i
+	}
+
+	for _, name := range preChecked {
+		if idx, ok := s.itemIndex[name]; ok {
+			s.checked[idx] = true
+			s.checkOrder = append(s.checkOrder, idx)
+		}
+	}
+
+	return s
+}
+
+func (s *multiSelectState) filtered() []selectItem {
+	return filterItems(s.items, s.filter)
+}
+
+func (s *multiSelectState) toggleItem() {
+	filtered := s.filtered()
+	if len(filtered) == 0 || s.highlighted >= len(filtered) {
+		return
+	}
+
+	item := filtered[s.highlighted]
+	origIdx := s.itemIndex[item.Name]
+
+	if s.checked[origIdx] {
+		delete(s.checked, origIdx)
+		for i, idx := range s.checkOrder {
+			if idx == origIdx {
+				s.checkOrder = append(s.checkOrder[:i], s.checkOrder[i+1:]...)
+				break
+			}
+		}
+	} else {
+		s.checked[origIdx] = true
+		s.checkOrder = append(s.checkOrder, origIdx)
+	}
+}
+
+func (s *multiSelectState) handleInput(event inputEvent, char byte) (done bool, result []string, err error) {
+	filtered := s.filtered()
+
+	switch event {
+	case eventEnter:
+		if s.focusOnButton && len(s.checkOrder) > 0 {
+			var res []string
+			for _, idx := range s.checkOrder {
+				res = append(res, s.items[idx].Name)
+			}
+			return true, res, nil
+		} else if !s.focusOnButton {
+			s.toggleItem()
+		}
+	case eventTab:
+		if len(s.checkOrder) > 0 {
+			s.focusOnButton = !s.focusOnButton
+		}
+	case eventEscape:
+		return true, nil, errCancelled
+	case eventBackspace:
+		if len(s.filter) > 0 {
+			s.filter = s.filter[:len(s.filter)-1]
+			s.highlighted = 0
+			s.scrollOffset = 0
+			s.focusOnButton = false
+		}
+	case eventUp:
+		if s.focusOnButton {
+			s.focusOnButton = false
+		} else if s.highlighted > 0 {
+			s.highlighted--
+			if s.highlighted < s.scrollOffset {
+				s.scrollOffset = s.highlighted
+			}
+		}
+	case eventDown:
+		if s.focusOnButton {
+			s.focusOnButton = false
+		} else if s.highlighted < len(filtered)-1 {
+			s.highlighted++
+			if s.highlighted >= s.scrollOffset+maxDisplayedItems {
+				s.scrollOffset = s.highlighted - maxDisplayedItems + 1
+			}
+		}
+	case eventChar:
+		s.filter += string(char)
+		s.highlighted = 0
+		s.scrollOffset = 0
+		s.focusOnButton = false
+	}
+
+	return false, nil, nil
+}
+
+func (s *multiSelectState) selectedCount() int {
+	return len(s.checkOrder)
+}
+
+// Terminal I/O handling
+
+type terminalState struct {
+	fd       int
+	oldState *term.State
+}
+
+func enterRawMode() (*terminalState, error) {
+	fd := int(os.Stdin.Fd())
+	oldState, err := term.MakeRaw(fd)
+	if err != nil {
+		return nil, err
+	}
+	fmt.Fprint(os.Stderr, ansiHideCursor)
+	return &terminalState{fd: fd, oldState: oldState}, nil
+}
+
+func (t *terminalState) restore() {
+	fmt.Fprint(os.Stderr, ansiShowCursor)
+	term.Restore(t.fd, t.oldState)
+}
+
+func clearLines(n int) {
+	if n > 0 {
+		fmt.Fprintf(os.Stderr, "\033[%dA", n)
+		fmt.Fprint(os.Stderr, ansiClearDown)
+	}
+}
+
+func parseInput(r io.Reader) (inputEvent, byte, error) {
+	buf := make([]byte, 3)
+	n, err := r.Read(buf)
+	if err != nil {
+		return 0, 0, err
+	}
+
+	switch {
+	case n == 1 && buf[0] == 13:
+		return eventEnter, 0, nil
+	case n == 1 && (buf[0] == 3 || buf[0] == 27):
+		return eventEscape, 0, nil
+	case n == 1 && buf[0] == 9:
+		return eventTab, 0, nil
+	case n == 1 && buf[0] == 127:
+		return eventBackspace, 0, nil
+	case n == 3 && buf[0] == 27 && buf[1] == 91 && buf[2] == 65:
+		return eventUp, 0, nil
+	case n == 3 && buf[0] == 27 && buf[1] == 91 && buf[2] == 66:
+		return eventDown, 0, nil
+	case n == 1 && buf[0] >= 32 && buf[0] < 127:
+		return eventChar, buf[0], nil
+	}
+
+	return eventNone, 0, nil
+}
+
+// Rendering
+
+func renderSelect(w io.Writer, prompt string, s *selectState) int {
+	filtered := s.filtered()
+
+	fmt.Fprintf(w, "%s %s\r\n", prompt, s.filter)
+	lineCount := 1
+
+	if len(filtered) == 0 {
+		fmt.Fprintf(w, "  %s(no matches)%s\r\n", ansiGray, ansiReset)
+		lineCount++
+	} else {
+		displayCount := min(len(filtered), maxDisplayedItems)
+
+		for i := range displayCount {
+			idx := s.scrollOffset + i
+			if idx >= len(filtered) {
+				break
+			}
+			item := filtered[idx]
+			prefix := "    "
+			if idx == s.selected {
+				prefix = "  " + ansiBold + "> "
+			}
+			if item.Description != "" {
+				fmt.Fprintf(w, "%s%s%s %s- %s%s\r\n", prefix, item.Name, ansiReset, ansiGray, item.Description, ansiReset)
+			} else {
+				fmt.Fprintf(w, "%s%s%s\r\n", prefix, item.Name, ansiReset)
+			}
+			lineCount++
+		}
+
+		if remaining := len(filtered) - s.scrollOffset - displayCount; remaining > 0 {
+			fmt.Fprintf(w, "  %s... and %d more%s\r\n", ansiGray, remaining, ansiReset)
+			lineCount++
+		}
+	}
+
+	return lineCount
+}
+
+func renderMultiSelect(w io.Writer, prompt string, s *multiSelectState) int {
+	filtered := s.filtered()
+
+	fmt.Fprintf(w, "%s %s\r\n", prompt, s.filter)
+	lineCount := 1
+
+	if len(filtered) == 0 {
+		fmt.Fprintf(w, "  %s(no matches)%s\r\n", ansiGray, ansiReset)
+		lineCount++
+	} else {
+		displayCount := min(len(filtered), maxDisplayedItems)
+
+		for i := range displayCount {
+			idx := s.scrollOffset + i
+			if idx >= len(filtered) {
+				break
+			}
+			item := filtered[idx]
+			origIdx := s.itemIndex[item.Name]
+
+			checkbox := "[ ]"
+			if s.checked[origIdx] {
+				checkbox = "[x]"
+			}
+
+			prefix := "  "
+			suffix := ""
+			if idx == s.highlighted && !s.focusOnButton {
+				prefix = "> "
+			}
+			if len(s.checkOrder) > 0 && s.checkOrder[0] == origIdx {
+				suffix = " " + ansiGray + "(default)" + ansiReset
+			}
+
+			if idx == s.highlighted && !s.focusOnButton {
+				fmt.Fprintf(w, "  %s%s %s %s%s%s\r\n", ansiBold, prefix, checkbox, item.Name, ansiReset, suffix)
+			} else {
+				fmt.Fprintf(w, "  %s %s %s%s\r\n", prefix, checkbox, item.Name, suffix)
+			}
+			lineCount++
+		}
+
+		if remaining := len(filtered) - s.scrollOffset - displayCount; remaining > 0 {
+			fmt.Fprintf(w, "  %s... and %d more%s\r\n", ansiGray, remaining, ansiReset)
+			lineCount++
+		}
+	}
+
+	fmt.Fprintf(w, "\r\n")
+	lineCount++
+	count := s.selectedCount()
+	switch {
+	case count == 0:
+		fmt.Fprintf(w, "  %sSelect at least one model.%s\r\n", ansiGray, ansiReset)
+	case s.focusOnButton:
+		fmt.Fprintf(w, "  %s> [ Continue ]%s %s(%d selected)%s\r\n", ansiBold, ansiReset, ansiGray, count, ansiReset)
+	default:
+		fmt.Fprintf(w, "    %s[ Continue ] (%d selected) - press Tab%s\r\n", ansiGray, count, ansiReset)
+	}
+	lineCount++
+
+	return lineCount
+}
+
+// selectPrompt prompts the user to select a single item from a list.
+func selectPrompt(prompt string, items []selectItem) (string, error) {
+	if len(items) == 0 {
+		return "", fmt.Errorf("no items to select from")
+	}
+
+	ts, err := enterRawMode()
+	if err != nil {
+		return "", err
+	}
+	defer ts.restore()
+
+	state := newSelectState(items)
+	var lastLineCount int
+
+	render := func() {
+		clearLines(lastLineCount)
+		lastLineCount = renderSelect(os.Stderr, prompt, state)
+	}
+
+	render()
+
+	for {
+		event, char, err := parseInput(os.Stdin)
+		if err != nil {
+			return "", err
+		}
+
+		done, result, err := state.handleInput(event, char)
+		if done {
+			clearLines(lastLineCount)
+			if err != nil {
+				return "", err
+			}
+			return result, nil
+		}
+
+		render()
+	}
+}
+
+// multiSelectPrompt prompts the user to select multiple items from a list.
+func multiSelectPrompt(prompt string, items []selectItem, preChecked []string) ([]string, error) {
+	if len(items) == 0 {
+		return nil, fmt.Errorf("no items to select from")
+	}
+
+	ts, err := enterRawMode()
+	if err != nil {
+		return nil, err
+	}
+	defer ts.restore()
+
+	state := newMultiSelectState(items, preChecked)
+	var lastLineCount int
+
+	render := func() {
+		clearLines(lastLineCount)
+		lastLineCount = renderMultiSelect(os.Stderr, prompt, state)
+	}
+
+	render()
+
+	for {
+		event, char, err := parseInput(os.Stdin)
+		if err != nil {
+			return nil, err
+		}
+
+		done, result, err := state.handleInput(event, char)
+		if done {
+			clearLines(lastLineCount)
+			if err != nil {
+				return nil, err
+			}
+			return result, nil
+		}
+
+		render()
+	}
+}
+
+func confirmPrompt(prompt string) (bool, error) {
+	fd := int(os.Stdin.Fd())
+	oldState, err := term.MakeRaw(fd)
+	if err != nil {
+		return false, err
+	}
+	defer term.Restore(fd, oldState)
+
+	fmt.Fprintf(os.Stderr, "%s (\033[1my\033[0m/n) ", prompt)
+
+	buf := make([]byte, 1)
+	for {
+		if _, err := os.Stdin.Read(buf); err != nil {
+			return false, err
+		}
+
+		switch buf[0] {
+		case 'Y', 'y', 13:
+			fmt.Fprintf(os.Stderr, "yes\r\n")
+			return true, nil
+		case 'N', 'n', 27, 3:
+			fmt.Fprintf(os.Stderr, "no\r\n")
+			return false, nil
+		}
+	}
+}
+
+func filterItems(items []selectItem, filter string) []selectItem {
+	if filter == "" {
+		return items
+	}
+	var result []selectItem
+	filterLower := strings.ToLower(filter)
+	for _, item := range items {
+		if strings.Contains(strings.ToLower(item.Name), filterLower) {
+			result = append(result, item)
+		}
+	}
+	return result
+}

cmd/config/selector_test.go

@@ -0,0 +1,913 @@
+package config
+
+import (
+	"bytes"
+	"strings"
+	"testing"
+)
+
+func TestFilterItems(t *testing.T) {
+	items := []selectItem{
+		{Name: "llama3.2:latest"},
+		{Name: "qwen2.5:7b"},
+		{Name: "deepseek-v3:cloud"},
+		{Name: "GPT-OSS:20b"},
+	}
+
+	t.Run("EmptyFilter_ReturnsAllItems", func(t *testing.T) {
+		result := filterItems(items, "")
+		if len(result) != len(items) {
+			t.Errorf("expected %d items, got %d", len(items), len(result))
+		}
+	})
+
+	t.Run("CaseInsensitive_UppercaseFilterMatchesLowercase", func(t *testing.T) {
+		result := filterItems(items, "LLAMA")
+		if len(result) != 1 || result[0].Name != "llama3.2:latest" {
+			t.Errorf("expected llama3.2:latest, got %v", result)
+		}
+	})
+
+	t.Run("CaseInsensitive_LowercaseFilterMatchesUppercase", func(t *testing.T) {
+		result := filterItems(items, "gpt")
+		if len(result) != 1 || result[0].Name != "GPT-OSS:20b" {
+			t.Errorf("expected GPT-OSS:20b, got %v", result)
+		}
+	})
+
+	t.Run("PartialMatch", func(t *testing.T) {
+		result := filterItems(items, "deep")
+		if len(result) != 1 || result[0].Name != "deepseek-v3:cloud" {
+			t.Errorf("expected deepseek-v3:cloud, got %v", result)
+		}
+	})
+
+	t.Run("NoMatch_ReturnsEmpty", func(t *testing.T) {
+		result := filterItems(items, "nonexistent")
+		if len(result) != 0 {
+			t.Errorf("expected 0 items, got %d", len(result))
+		}
+	})
+}
+
+func TestSelectState(t *testing.T) {
+	items := []selectItem{
+		{Name: "item1"},
+		{Name: "item2"},
+		{Name: "item3"},
+	}
+
+	t.Run("InitialState", func(t *testing.T) {
+		s := newSelectState(items)
+		if s.selected != 0 {
+			t.Errorf("expected selected=0, got %d", s.selected)
+		}
+		if s.filter != "" {
+			t.Errorf("expected empty filter, got %q", s.filter)
+		}
+		if s.scrollOffset != 0 {
+			t.Errorf("expected scrollOffset=0, got %d", s.scrollOffset)
+		}
+	})
+
+	t.Run("Enter_SelectsCurrentItem", func(t *testing.T) {
+		s := newSelectState(items)
+		done, result, err := s.handleInput(eventEnter, 0)
+		if !done || result != "item1" || err != nil {
+			t.Errorf("expected (true, item1, nil), got (%v, %v, %v)", done, result, err)
+		}
+	})
+
+	t.Run("Enter_WithFilter_SelectsFilteredItem", func(t *testing.T) {
+		s := newSelectState(items)
+		s.filter = "item3"
+		done, result, err := s.handleInput(eventEnter, 0)
+		if !done || result != "item3" || err != nil {
+			t.Errorf("expected (true, item3, nil), got (%v, %v, %v)", done, result, err)
+		}
+	})
+
+	t.Run("Enter_EmptyFilteredList_DoesNothing", func(t *testing.T) {
+		s := newSelectState(items)
+		s.filter = "nonexistent"
+		done, result, err := s.handleInput(eventEnter, 0)
+		if done || result != "" || err != nil {
+			t.Errorf("expected (false, '', nil), got (%v, %v, %v)", done, result, err)
+		}
+	})
+
+	t.Run("Escape_ReturnsCancelledError", func(t *testing.T) {
+		s := newSelectState(items)
+		done, result, err := s.handleInput(eventEscape, 0)
+		if !done || result != "" || err != errCancelled {
+			t.Errorf("expected (true, '', errCancelled), got (%v, %v, %v)", done, result, err)
+		}
+	})
+
+	t.Run("Down_MovesSelection", func(t *testing.T) {
+		s := newSelectState(items)
+		s.handleInput(eventDown, 0)
+		if s.selected != 1 {
+			t.Errorf("expected selected=1, got %d", s.selected)
+		}
+	})
+
+	t.Run("Down_AtBottom_StaysAtBottom", func(t *testing.T) {
+		s := newSelectState(items)
+		s.selected = 2
+		s.handleInput(eventDown, 0)
+		if s.selected != 2 {
+			t.Errorf("expected selected=2 (stayed at bottom), got %d", s.selected)
+		}
+	})
+
+	t.Run("Up_MovesSelection", func(t *testing.T) {
+		s := newSelectState(items)
+		s.selected = 2
+		s.handleInput(eventUp, 0)
+		if s.selected != 1 {
+			t.Errorf("expected selected=1, got %d", s.selected)
+		}
+	})
+
+	t.Run("Up_AtTop_StaysAtTop", func(t *testing.T) {
+		s := newSelectState(items)
+		s.handleInput(eventUp, 0)
+		if s.selected != 0 {
+			t.Errorf("expected selected=0 (stayed at top), got %d", s.selected)
+		}
+	})
+
+	t.Run("Char_AppendsToFilter", func(t *testing.T) {
+		s := newSelectState(items)
+		s.handleInput(eventChar, 'i')
+		s.handleInput(eventChar, 't')
+		s.handleInput(eventChar, 'e')
+		s.handleInput(eventChar, 'm')
+		s.handleInput(eventChar, '2')
+		if s.filter != "item2" {
+			t.Errorf("expected filter='item2', got %q", s.filter)
+		}
+		filtered := s.filtered()
+		if len(filtered) != 1 || filtered[0].Name != "item2" {
+			t.Errorf("expected [item2], got %v", filtered)
+		}
+	})
+
+	t.Run("Char_ResetsSelectionToZero", func(t *testing.T) {
+		s := newSelectState(items)
+		s.selected = 2
+		s.handleInput(eventChar, 'x')
+		if s.selected != 0 {
+			t.Errorf("expected selected=0 after typing, got %d", s.selected)
+		}
+	})
+
+	t.Run("Backspace_RemovesLastFilterChar", func(t *testing.T) {
+		s := newSelectState(items)
+		s.filter = "test"
+		s.handleInput(eventBackspace, 0)
+		if s.filter != "tes" {
+			t.Errorf("expected filter='tes', got %q", s.filter)
+		}
+	})
+
+	t.Run("Backspace_EmptyFilter_DoesNothing", func(t *testing.T) {
+		s := newSelectState(items)
+		s.handleInput(eventBackspace, 0)
+		if s.filter != "" {
+			t.Errorf("expected filter='', got %q", s.filter)
+		}
+	})
+
+	t.Run("Backspace_ResetsSelectionToZero", func(t *testing.T) {
+		s := newSelectState(items)
+		s.filter = "test"
+		s.selected = 2
+		s.handleInput(eventBackspace, 0)
+		if s.selected != 0 {
+			t.Errorf("expected selected=0 after backspace, got %d", s.selected)
+		}
+	})
+
+	t.Run("Scroll_DownPastVisibleItems_ScrollsViewport", func(t *testing.T) {
+		// maxDisplayedItems is 10, so with 15 items we need to scroll
+		manyItems := make([]selectItem, 15)
+		for i := range manyItems {
+			manyItems[i] = selectItem{Name: string(rune('a' + i))}
+		}
+		s := newSelectState(manyItems)
+
+		// move down 12 times (past the 10-item viewport)
+		for range 12 {
+			s.handleInput(eventDown, 0)
+		}
+
+		if s.selected != 12 {
+			t.Errorf("expected selected=12, got %d", s.selected)
+		}
+		if s.scrollOffset != 3 {
+			t.Errorf("expected scrollOffset=3 (12-10+1), got %d", s.scrollOffset)
+		}
+	})
+
+	t.Run("Scroll_UpPastScrollOffset_ScrollsViewport", func(t *testing.T) {
+		manyItems := make([]selectItem, 15)
+		for i := range manyItems {
+			manyItems[i] = selectItem{Name: string(rune('a' + i))}
+		}
+		s := newSelectState(manyItems)
+		s.selected = 5
+		s.scrollOffset = 5
+
+		s.handleInput(eventUp, 0)
+
+		if s.selected != 4 {
+			t.Errorf("expected selected=4, got %d", s.selected)
+		}
+		if s.scrollOffset != 4 {
+			t.Errorf("expected scrollOffset=4, got %d", s.scrollOffset)
+		}
+	})
+}
+
+func TestMultiSelectState(t *testing.T) {
+	items := []selectItem{
+		{Name: "item1"},
+		{Name: "item2"},
+		{Name: "item3"},
+	}
+
+	t.Run("InitialState_NoPrechecked", func(t *testing.T) {
+		s := newMultiSelectState(items, nil)
+		if s.highlighted != 0 {
+			t.Errorf("expected highlighted=0, got %d", s.highlighted)
+		}
+		if s.selectedCount() != 0 {
+			t.Errorf("expected 0 selected, got %d", s.selectedCount())
+		}
+		if s.focusOnButton {
+			t.Error("expected focusOnButton=false initially")
+		}
+	})
+
+	t.Run("InitialState_WithPrechecked", func(t *testing.T) {
+		s := newMultiSelectState(items, []string{"item2", "item3"})
+		if s.selectedCount() != 2 {
+			t.Errorf("expected 2 selected, got %d", s.selectedCount())
+		}
+		if !s.checked[1] || !s.checked[2] {
+			t.Error("expected item2 and item3 to be checked")
+		}
+	})
+
+	t.Run("Prechecked_PreservesSelectionOrder", func(t *testing.T) {
+		// order matters: first checked = default model
+		s := newMultiSelectState(items, []string{"item3", "item1"})
+		if len(s.checkOrder) != 2 {
+			t.Fatalf("expected 2 in checkOrder, got %d", len(s.checkOrder))
+		}
+		if s.checkOrder[0] != 2 || s.checkOrder[1] != 0 {
+			t.Errorf("expected checkOrder=[2,0] (item3 first), got %v", s.checkOrder)
+		}
+	})
+
+	t.Run("Prechecked_IgnoresInvalidNames", func(t *testing.T) {
+		s := newMultiSelectState(items, []string{"item1", "nonexistent"})
+		if s.selectedCount() != 1 {
+			t.Errorf("expected 1 selected (nonexistent ignored), got %d", s.selectedCount())
+		}
+	})
+
+	t.Run("Toggle_ChecksUncheckedItem", func(t *testing.T) {
+		s := newMultiSelectState(items, nil)
+		s.toggleItem()
+		if !s.checked[0] {
+			t.Error("expected item1 to be checked after toggle")
+		}
+	})
+
+	t.Run("Toggle_UnchecksCheckedItem", func(t *testing.T) {
+		s := newMultiSelectState(items, []string{"item1"})
+		s.toggleItem()
+		if s.checked[0] {
+			t.Error("expected item1 to be unchecked after toggle")
+		}
+	})
+
+	t.Run("Toggle_RemovesFromCheckOrder", func(t *testing.T) {
+		s := newMultiSelectState(items, []string{"item1", "item2", "item3"})
+		s.highlighted = 1 // toggle item2
+		s.toggleItem()
+
+		if len(s.checkOrder) != 2 {
+			t.Fatalf("expected 2 in checkOrder, got %d", len(s.checkOrder))
+		}
+		// should be [0, 2] (item1, item3) with item2 removed
+		if s.checkOrder[0] != 0 || s.checkOrder[1] != 2 {
+			t.Errorf("expected checkOrder=[0,2], got %v", s.checkOrder)
+		}
+	})
+
+	t.Run("Enter_TogglesWhenNotOnButton", func(t *testing.T) {
+		s := newMultiSelectState(items, nil)
+		s.handleInput(eventEnter, 0)
+		if !s.checked[0] {
+			t.Error("expected item1 to be checked after enter")
+		}
+	})
+
+	t.Run("Enter_OnButton_ReturnsSelection", func(t *testing.T) {
+		s := newMultiSelectState(items, []string{"item2", "item1"})
+		s.focusOnButton = true
+
+		done, result, err := s.handleInput(eventEnter, 0)
+
+		if !done || err != nil {
+			t.Errorf("expected done=true, err=nil, got done=%v, err=%v", done, err)
+		}
+		// result should preserve selection order
+		if len(result) != 2 || result[0] != "item2" || result[1] != "item1" {
+			t.Errorf("expected [item2, item1], got %v", result)
+		}
+	})
+
+	t.Run("Enter_OnButton_EmptySelection_DoesNothing", func(t *testing.T) {
+		s := newMultiSelectState(items, nil)
+		s.focusOnButton = true
+		done, result, err := s.handleInput(eventEnter, 0)
+		if done || result != nil || err != nil {
+			t.Errorf("expected (false, nil, nil), got (%v, %v, %v)", done, result, err)
+		}
+	})
+
+	t.Run("Tab_SwitchesToButton_WhenHasSelection", func(t *testing.T) {
+		s := newMultiSelectState(items, []string{"item1"})
+		s.handleInput(eventTab, 0)
+		if !s.focusOnButton {
+			t.Error("expected focus on button after tab")
+		}
+	})
+
+	t.Run("Tab_DoesNothing_WhenNoSelection", func(t *testing.T) {
+		s := newMultiSelectState(items, nil)
+		s.handleInput(eventTab, 0)
+		if s.focusOnButton {
+			t.Error("tab should not focus button when nothing selected")
+		}
+	})
+
+	t.Run("Tab_TogglesButtonFocus", func(t *testing.T) {
+		s := newMultiSelectState(items, []string{"item1"})
+		s.handleInput(eventTab, 0)
+		if !s.focusOnButton {
+			t.Error("expected focus on button after first tab")
+		}
+		s.handleInput(eventTab, 0)
+		if s.focusOnButton {
+			t.Error("expected focus back on list after second tab")
+		}
+	})
+
+	t.Run("Escape_ReturnsCancelledError", func(t *testing.T) {
+		s := newMultiSelectState(items, []string{"item1"})
+		done, result, err := s.handleInput(eventEscape, 0)
+		if !done || result != nil || err != errCancelled {
+			t.Errorf("expected (true, nil, errCancelled), got (%v, %v, %v)", done, result, err)
+		}
+	})
+
+	t.Run("IsDefault_TrueForFirstChecked", func(t *testing.T) {
+		s := newMultiSelectState(items, []string{"item2", "item1"})
+		if !(len(s.checkOrder) > 0 && s.checkOrder[0] == 1) {
+			t.Error("expected item2 (idx 1) to be default (first checked)")
+		}
+		if len(s.checkOrder) > 0 && s.checkOrder[0] == 0 {
+			t.Error("expected item1 (idx 0) to NOT be default")
+		}
+	})
+
+	t.Run("IsDefault_FalseWhenNothingChecked", func(t *testing.T) {
+		s := newMultiSelectState(items, nil)
+		if len(s.checkOrder) > 0 && s.checkOrder[0] == 0 {
+			t.Error("expected isDefault=false when nothing checked")
+		}
+	})
+
+	t.Run("Down_MovesHighlight", func(t *testing.T) {
+		s := newMultiSelectState(items, nil)
+		s.handleInput(eventDown, 0)
+		if s.highlighted != 1 {
+			t.Errorf("expected highlighted=1, got %d", s.highlighted)
+		}
+	})
+
+	t.Run("Up_MovesHighlight", func(t *testing.T) {
+		s := newMultiSelectState(items, nil)
+		s.highlighted = 1
+		s.handleInput(eventUp, 0)
+		if s.highlighted != 0 {
+			t.Errorf("expected highlighted=0, got %d", s.highlighted)
+		}
+	})
+
+	t.Run("Arrow_ReturnsFocusFromButton", func(t *testing.T) {
+		s := newMultiSelectState(items, []string{"item1"})
+		s.focusOnButton = true
+		s.handleInput(eventDown, 0)
+		if s.focusOnButton {
+			t.Error("expected focus to return to list on arrow key")
+		}
+	})
+
+	t.Run("Char_AppendsToFilter", func(t *testing.T) {
+		s := newMultiSelectState(items, nil)
+		s.handleInput(eventChar, 'x')
+		if s.filter != "x" {
+			t.Errorf("expected filter='x', got %q", s.filter)
+		}
+	})
+
+	t.Run("Char_ResetsHighlightAndScroll", func(t *testing.T) {
+		manyItems := make([]selectItem, 15)
+		for i := range manyItems {
+			manyItems[i] = selectItem{Name: string(rune('a' + i))}
+		}
+		s := newMultiSelectState(manyItems, nil)
+		s.highlighted = 10
+		s.scrollOffset = 5
+
+		s.handleInput(eventChar, 'x')
+
+		if s.highlighted != 0 {
+			t.Errorf("expected highlighted=0, got %d", s.highlighted)
+		}
+		if s.scrollOffset != 0 {
+			t.Errorf("expected scrollOffset=0, got %d", s.scrollOffset)
+		}
+	})
+
+	t.Run("Backspace_RemovesLastFilterChar", func(t *testing.T) {
+		s := newMultiSelectState(items, nil)
+		s.filter = "test"
+		s.handleInput(eventBackspace, 0)
+		if s.filter != "tes" {
+			t.Errorf("expected filter='tes', got %q", s.filter)
+		}
+	})
+
+	t.Run("Backspace_RemovesFocusFromButton", func(t *testing.T) {
+		s := newMultiSelectState(items, []string{"item1"})
+		s.filter = "x"
+		s.focusOnButton = true
+		s.handleInput(eventBackspace, 0)
+		if s.focusOnButton {
+			t.Error("expected focusOnButton=false after backspace")
+		}
+	})
+}
+
+func TestParseInput(t *testing.T) {
+	t.Run("Enter", func(t *testing.T) {
+		event, char, err := parseInput(bytes.NewReader([]byte{13}))
+		if err != nil || event != eventEnter || char != 0 {
+			t.Errorf("expected (eventEnter, 0, nil), got (%v, %v, %v)", event, char, err)
+		}
+	})
+
+	t.Run("Escape", func(t *testing.T) {
+		event, _, err := parseInput(bytes.NewReader([]byte{27}))
+		if err != nil || event != eventEscape {
+			t.Errorf("expected eventEscape, got %v", event)
+		}
+	})
+
+	t.Run("CtrlC_TreatedAsEscape", func(t *testing.T) {
+		event, _, err := parseInput(bytes.NewReader([]byte{3}))
+		if err != nil || event != eventEscape {
+			t.Errorf("expected eventEscape for Ctrl+C, got %v", event)
+		}
+	})
+
+	t.Run("Tab", func(t *testing.T) {
+		event, _, err := parseInput(bytes.NewReader([]byte{9}))
+		if err != nil || event != eventTab {
+			t.Errorf("expected eventTab, got %v", event)
+		}
+	})
+
+	t.Run("Backspace", func(t *testing.T) {
+		event, _, err := parseInput(bytes.NewReader([]byte{127}))
+		if err != nil || event != eventBackspace {
+			t.Errorf("expected eventBackspace, got %v", event)
+		}
+	})
+
+	t.Run("UpArrow", func(t *testing.T) {
+		event, _, err := parseInput(bytes.NewReader([]byte{27, 91, 65}))
+		if err != nil || event != eventUp {
+			t.Errorf("expected eventUp, got %v", event)
+		}
+	})
+
+	t.Run("DownArrow", func(t *testing.T) {
+		event, _, err := parseInput(bytes.NewReader([]byte{27, 91, 66}))
+		if err != nil || event != eventDown {
+			t.Errorf("expected eventDown, got %v", event)
+		}
+	})
+
+	t.Run("PrintableChars", func(t *testing.T) {
+		tests := []struct {
+			name string
+			char byte
+		}{
+			{"lowercase", 'a'},
+			{"uppercase", 'Z'},
+			{"digit", '5'},
+			{"space", ' '},
+			{"tilde", '~'},
+		}
+		for _, tt := range tests {
+			t.Run(tt.name, func(t *testing.T) {
+				event, char, err := parseInput(bytes.NewReader([]byte{tt.char}))
+				if err != nil || event != eventChar || char != tt.char {
+					t.Errorf("expected (eventChar, %q), got (%v, %q)", tt.char, event, char)
+				}
+			})
+		}
+	})
+}
+
+func TestRenderSelect(t *testing.T) {
+	items := []selectItem{
+		{Name: "item1", Description: "first item"},
+		{Name: "item2"},
+	}
+
+	t.Run("ShowsPromptAndItems", func(t *testing.T) {
+		s := newSelectState(items)
+		var buf bytes.Buffer
+		lineCount := renderSelect(&buf, "Select:", s)
+
+		output := buf.String()
+		if !strings.Contains(output, "Select:") {
+			t.Error("expected prompt in output")
+		}
+		if !strings.Contains(output, "item1") {
+			t.Error("expected item1 in output")
+		}
+		if !strings.Contains(output, "first item") {
+			t.Error("expected description in output")
+		}
+		if !strings.Contains(output, "item2") {
+			t.Error("expected item2 in output")
+		}
+		if lineCount != 3 { // 1 prompt + 2 items
+			t.Errorf("expected 3 lines, got %d", lineCount)
+		}
+	})
+
+	t.Run("EmptyFilteredList_ShowsNoMatches", func(t *testing.T) {
+		s := newSelectState(items)
+		s.filter = "xyz"
+		var buf bytes.Buffer
+		renderSelect(&buf, "Select:", s)
+
+		if !strings.Contains(buf.String(), "no matches") {
+			t.Error("expected 'no matches' message")
+		}
+	})
+
+	t.Run("LongList_ShowsRemainingCount", func(t *testing.T) {
+		manyItems := make([]selectItem, 15)
+		for i := range manyItems {
+			manyItems[i] = selectItem{Name: string(rune('a' + i))}
+		}
+		s := newSelectState(manyItems)
+		var buf bytes.Buffer
+		renderSelect(&buf, "Select:", s)
+
+		// 15 items - 10 displayed = 5 more
+		if !strings.Contains(buf.String(), "5 more") {
+			t.Error("expected '5 more' indicator")
+		}
+	})
+}
+
+func TestRenderMultiSelect(t *testing.T) {
+	items := []selectItem{
+		{Name: "item1"},
+		{Name: "item2"},
+	}
+
+	t.Run("ShowsCheckboxes", func(t *testing.T) {
+		s := newMultiSelectState(items, []string{"item1"})
+		var buf bytes.Buffer
+		renderMultiSelect(&buf, "Select:", s)
+
+		output := buf.String()
+		if !strings.Contains(output, "[x]") {
+			t.Error("expected checked checkbox [x]")
+		}
+		if !strings.Contains(output, "[ ]") {
+			t.Error("expected unchecked checkbox [ ]")
+		}
+	})
+
+	t.Run("ShowsDefaultMarker", func(t *testing.T) {
+		s := newMultiSelectState(items, []string{"item1"})
+		var buf bytes.Buffer
+		renderMultiSelect(&buf, "Select:", s)
+
+		if !strings.Contains(buf.String(), "(default)") {
+			t.Error("expected (default) marker for first checked item")
+		}
+	})
+
+	t.Run("ShowsSelectedCount", func(t *testing.T) {
+		s := newMultiSelectState(items, []string{"item1", "item2"})
+		var buf bytes.Buffer
+		renderMultiSelect(&buf, "Select:", s)
+
+		if !strings.Contains(buf.String(), "2 selected") {
+			t.Error("expected '2 selected' in output")
+		}
+	})
+
+	t.Run("NoSelection_ShowsHelperText", func(t *testing.T) {
+		s := newMultiSelectState(items, nil)
+		var buf bytes.Buffer
+		renderMultiSelect(&buf, "Select:", s)
+
+		if !strings.Contains(buf.String(), "Select at least one") {
+			t.Error("expected 'Select at least one' helper text")
+		}
+	})
+}
+
+func TestErrCancelled(t *testing.T) {
+	t.Run("NotNil", func(t *testing.T) {
+		if errCancelled == nil {
+			t.Error("errCancelled should not be nil")
+		}
+	})
+
+	t.Run("Message", func(t *testing.T) {
+		if errCancelled.Error() != "cancelled" {
+			t.Errorf("expected 'cancelled', got %q", errCancelled.Error())
+		}
+	})
+}
+
+// Edge case tests for selector.go
+
+// TestSelectState_SingleItem verifies that single item list works without crash.
+// List with only one item should still work.
+func TestSelectState_SingleItem(t *testing.T) {
+	items := []selectItem{{Name: "only-one"}}
+
+	s := newSelectState(items)
+
+	// Down should do nothing (already at bottom)
+	s.handleInput(eventDown, 0)
+	if s.selected != 0 {
+		t.Errorf("down on single item: expected selected=0, got %d", s.selected)
+	}
+
+	// Up should do nothing (already at top)
+	s.handleInput(eventUp, 0)
+	if s.selected != 0 {
+		t.Errorf("up on single item: expected selected=0, got %d", s.selected)
+	}
+
+	// Enter should select the only item
+	done, result, err := s.handleInput(eventEnter, 0)
+	if !done || result != "only-one" || err != nil {
+		t.Errorf("enter on single item: expected (true, 'only-one', nil), got (%v, %q, %v)", done, result, err)
+	}
+}
+
+// TestSelectState_ExactlyMaxItems verifies boundary condition at maxDisplayedItems.
+// List with exactly maxDisplayedItems items should not scroll.
+func TestSelectState_ExactlyMaxItems(t *testing.T) {
+	items := make([]selectItem, maxDisplayedItems)
+	for i := range items {
+		items[i] = selectItem{Name: string(rune('a' + i))}
+	}
+
+	s := newSelectState(items)
+
+	// Move to last item
+	for range maxDisplayedItems - 1 {
+		s.handleInput(eventDown, 0)
+	}
+
+	if s.selected != maxDisplayedItems-1 {
+		t.Errorf("expected selected=%d, got %d", maxDisplayedItems-1, s.selected)
+	}
+
+	// Should not scroll when exactly at max
+	if s.scrollOffset != 0 {
+		t.Errorf("expected scrollOffset=0 for exactly maxDisplayedItems, got %d", s.scrollOffset)
+	}
+
+	// One more down should do nothing
+	s.handleInput(eventDown, 0)
+	if s.selected != maxDisplayedItems-1 {
+		t.Errorf("down at max: expected selected=%d, got %d", maxDisplayedItems-1, s.selected)
+	}
+}
+
+// TestFilterItems_RegexSpecialChars verifies that filter is literal, not regex.
+// User typing "model.v1" shouldn't match "modelsv1".
+func TestFilterItems_RegexSpecialChars(t *testing.T) {
+	items := []selectItem{
+		{Name: "model.v1"},
+		{Name: "modelsv1"},
+		{Name: "model-v1"},
+	}
+
+	// Filter with dot should only match literal dot
+	result := filterItems(items, "model.v1")
+	if len(result) != 1 {
+		t.Errorf("expected 1 exact match, got %d", len(result))
+	}
+	if len(result) > 0 && result[0].Name != "model.v1" {
+		t.Errorf("expected 'model.v1', got %s", result[0].Name)
+	}
+
+	// Other regex special chars should be literal too
+	items2 := []selectItem{
+		{Name: "test[0]"},
+		{Name: "test0"},
+		{Name: "test(1)"},
+	}
+
+	result2 := filterItems(items2, "test[0]")
+	if len(result2) != 1 || result2[0].Name != "test[0]" {
+		t.Errorf("expected only 'test[0]', got %v", result2)
+	}
+}
+
+// TestMultiSelectState_DuplicateNames documents handling of duplicate item names.
+// itemIndex uses name as key - duplicates cause collision. This documents
+// the current behavior: the last index for a duplicate name is stored
+func TestMultiSelectState_DuplicateNames(t *testing.T) {
+	// Duplicate names - this is an edge case that shouldn't happen in practice
+	items := []selectItem{
+		{Name: "duplicate"},
+		{Name: "duplicate"},
+		{Name: "unique"},
+	}
+
+	s := newMultiSelectState(items, nil)
+
+	// DOCUMENTED BEHAVIOR: itemIndex maps name to LAST index
+	// When there are duplicates, only the last occurrence's index is stored
+	if s.itemIndex["duplicate"] != 1 {
+		t.Errorf("itemIndex should map 'duplicate' to last index (1), got %d", s.itemIndex["duplicate"])
+	}
+
+	// Toggle item at highlighted=0 (first "duplicate")
+	// Due to name collision, toggleItem uses itemIndex["duplicate"] = 1
+	// So it actually toggles the SECOND duplicate item, not the first
+	s.toggleItem()
+
+	// This documents the potentially surprising behavior:
+	// We toggled at highlighted=0, but itemIndex lookup returned 1
+	if !s.checked[1] {
+		t.Error("toggle should check index 1 (due to name collision in itemIndex)")
+	}
+	if s.checked[0] {
+		t.Log("Note: index 0 is NOT checked, even though highlighted=0 (name collision behavior)")
+	}
+}
+
+// TestSelectState_FilterReducesBelowSelection verifies selection resets when filter reduces list.
+// Prevents index-out-of-bounds on next keystroke
+func TestSelectState_FilterReducesBelowSelection(t *testing.T) {
+	items := []selectItem{
+		{Name: "apple"},
+		{Name: "banana"},
+		{Name: "cherry"},
+	}
+
+	s := newSelectState(items)
+	s.selected = 2 // Select "cherry"
+
+	// Type a filter that removes cherry from results
+	s.handleInput(eventChar, 'a') // Filter to "a" - matches "apple" and "banana"
+
+	// Selection should reset to 0
+	if s.selected != 0 {
+		t.Errorf("expected selected=0 after filter, got %d", s.selected)
+	}
+
+	filtered := s.filtered()
+	if len(filtered) != 2 {
+		t.Errorf("expected 2 filtered items, got %d", len(filtered))
+	}
+}
+
+// TestFilterItems_UnicodeCharacters verifies filtering works with UTF-8.
+// Model names might contain unicode characters
+func TestFilterItems_UnicodeCharacters(t *testing.T) {
+	items := []selectItem{
+		{Name: "llama-日本語"},
+		{Name: "模型-chinese"},
+		{Name: "émoji-🦙"},
+		{Name: "regular-model"},
+	}
+
+	t.Run("filter japanese", func(t *testing.T) {
+		result := filterItems(items, "日本")
+		if len(result) != 1 || result[0].Name != "llama-日本語" {
+			t.Errorf("expected llama-日本語, got %v", result)
+		}
+	})
+
+	t.Run("filter chinese", func(t *testing.T) {
+		result := filterItems(items, "模型")
+		if len(result) != 1 || result[0].Name != "模型-chinese" {
+			t.Errorf("expected 模型-chinese, got %v", result)
+		}
+	})
+
+	t.Run("filter emoji", func(t *testing.T) {
+		result := filterItems(items, "🦙")
+		if len(result) != 1 || result[0].Name != "émoji-🦙" {
+			t.Errorf("expected émoji-🦙, got %v", result)
+		}
+	})
+
+	t.Run("filter accented char", func(t *testing.T) {
+		result := filterItems(items, "émoji")
+		if len(result) != 1 || result[0].Name != "émoji-🦙" {
+			t.Errorf("expected émoji-🦙, got %v", result)
+		}
+	})
+}
+
+// TestMultiSelectState_FilterReducesBelowHighlight verifies highlight resets when filter reduces list.
+func TestMultiSelectState_FilterReducesBelowHighlight(t *testing.T) {
+	items := []selectItem{
+		{Name: "apple"},
+		{Name: "banana"},
+		{Name: "cherry"},
+	}
+
+	s := newMultiSelectState(items, nil)
+	s.highlighted = 2 // Highlight "cherry"
+
+	// Type a filter that removes cherry
+	s.handleInput(eventChar, 'a')
+
+	if s.highlighted != 0 {
+		t.Errorf("expected highlighted=0 after filter, got %d", s.highlighted)
+	}
+}
+
+// TestMultiSelectState_EmptyItems verifies handling of empty item list.
+// Empty list should be handled gracefully.
+func TestMultiSelectState_EmptyItems(t *testing.T) {
+	s := newMultiSelectState([]selectItem{}, nil)
+
+	// Toggle should not panic on empty list
+	s.toggleItem()
+
+	if s.selectedCount() != 0 {
+		t.Errorf("expected 0 selected for empty list, got %d", s.selectedCount())
+	}
+
+	// Render should handle empty list
+	var buf bytes.Buffer
+	lineCount := renderMultiSelect(&buf, "Select:", s)
+	if lineCount == 0 {
+		t.Error("renderMultiSelect should produce output even for empty list")
+	}
+	if !strings.Contains(buf.String(), "no matches") {
+		t.Error("expected 'no matches' for empty list")
+	}
+}
+
+// TestSelectState_RenderWithDescriptions verifies rendering items with descriptions.
+func TestSelectState_RenderWithDescriptions(t *testing.T) {
+	items := []selectItem{
+		{Name: "item1", Description: "First item description"},
+		{Name: "item2", Description: ""},
+		{Name: "item3", Description: "Third item"},
+	}
+
+	s := newSelectState(items)
+	var buf bytes.Buffer
+	renderSelect(&buf, "Select:", s)
+
+	output := buf.String()
+	if !strings.Contains(output, "First item description") {
+		t.Error("expected description to be rendered")
+	}
+	if !strings.Contains(output, "item2") {
+		t.Error("expected item without description to be rendered")
+	}
+}

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
@@ -159,6 +159,7 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 			sb.WriteString(before)
 			if !ok {
 				fmt.Fprintln(&sb)
+				scanner.Prompt.UseAlt = true
 				continue
 			}
 

convert/convert.go

@@ -311,6 +311,10 @@ func LoadModelMetadata(fsys fs.FS) (ModelKV, *Tokenizer, error) {
 		conv = &deepseekocr{}
 	case "DeepseekV3ForCausalLM":
 		conv = &deepseek2Model{}
+	case "Glm4MoeLiteForCausalLM":
+		conv = &glm4MoeLiteModel{}
+	case "Lfm2ForCausalLM":
+		conv = &lfm2Model{}
 	default:
 		return nil, nil, fmt.Errorf("unsupported architecture %q", p.Architectures[0])
 	}

convert/convert_glm4moelite.go

@@ -0,0 +1,264 @@
+package convert
+
+import (
+	"cmp"
+	"fmt"
+	"log/slog"
+	"regexp"
+	"strconv"
+	"strings"
+
+	"github.com/pdevine/tensor"
+	"github.com/pdevine/tensor/native"
+
+	"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["glm4moelite.attention.key_length_mla"] = p.KVLoraRank + p.QKRopeHeadDim
+	kv["glm4moelite.attention.value_length_mla"] = p.KVLoraRank
+
+	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",
+	}
+}
+
+// repackKVB extracts K or V from the combined KV_B tensor for MLA absorption.
+// K output row-major: [n_head, kv_lora_rank, qk_nope] -> GGML ne[]={qk_nope, kv_lora_rank, n_head}
+// V output row-major: [n_head, v_head, kv_lora_rank] -> GGML ne[]={kv_lora_rank, v_head, n_head}
+func (p *glm4MoeLiteModel) repackKVB(extractK bool, kvFirst bool, numHeads int) Repacker {
+	qkNope := int(p.QKNopeHeadDim)
+	vHeadDim := int(p.VHeadDim)
+	kvLoraRank := int(p.KVLoraRank)
+	kvPerHead := qkNope + vHeadDim
+
+	return func(_ string, data []float32, shape []uint64) ([]float32, error) {
+		dims := make([]int, len(shape))
+		for i := range shape {
+			dims[i] = int(shape[i])
+		}
+
+		var tt tensor.Tensor = tensor.New(tensor.WithShape(dims...), tensor.WithBacking(data))
+		var err error
+
+		// Normalize to [n_head * (qk_nope + v_head), kv_lora_rank] layout
+		if kvFirst {
+			tt, err = tensor.Transpose(tt, 1, 0)
+			if err != nil {
+				return nil, err
+			}
+			tt = tensor.Materialize(tt)
+		}
+
+		// Reshape to [n_head, qk_nope + v_head, kv_lora_rank]
+		if err := tt.Reshape(numHeads, kvPerHead, kvLoraRank); err != nil {
+			return nil, err
+		}
+
+		if extractK {
+			// Slice K: [n_head, qk_nope, kv_lora_rank]
+			tt, err = tt.Slice(nil, tensor.S(0, qkNope), nil)
+			if err != nil {
+				return nil, err
+			}
+			tt = tensor.Materialize(tt)
+			// Transpose to [n_head, kv_lora_rank, qk_nope]
+			tt, err = tensor.Transpose(tt, 0, 2, 1)
+			if err != nil {
+				return nil, err
+			}
+			tt = tensor.Materialize(tt)
+		} else {
+			// Slice V: [n_head, v_head, kv_lora_rank] - already correct layout
+			tt, err = tt.Slice(nil, tensor.S(qkNope, kvPerHead), nil)
+			if err != nil {
+				return nil, err
+			}
+			tt = tensor.Materialize(tt)
+		}
+
+		if err := tt.Reshape(tt.Shape().TotalSize()); err != nil {
+			return nil, err
+		}
+		return native.VectorF32(tt.(*tensor.Dense))
+	}
+}
+
+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
+		}
+
+		// Split attn_kv_b into separate attn_k_b and attn_v_b for MLA absorption
+		if strings.HasSuffix(t.Name(), ".attn_kv_b.weight") {
+			qkNope := int(p.QKNopeHeadDim)
+			vHeadDim := int(p.VHeadDim)
+			kvLoraRank := int(p.KVLoraRank)
+			kvPerHead := qkNope + vHeadDim
+			numHeads := int(p.NumAttentionHeads)
+			kvFirst := true
+			if len(t.Shape()) == 2 {
+				switch {
+				case int(t.Shape()[0]) == kvLoraRank:
+					if kvPerHead > 0 && int(t.Shape()[1])%kvPerHead == 0 {
+						numHeads = int(t.Shape()[1]) / kvPerHead
+					}
+					kvFirst = true
+				case int(t.Shape()[1]) == kvLoraRank:
+					if kvPerHead > 0 && int(t.Shape()[0])%kvPerHead == 0 {
+						numHeads = int(t.Shape()[0]) / kvPerHead
+					}
+					kvFirst = false
+				default:
+					slog.Warn("glm4moelite: unexpected attn_kv_b layout", "name", t.Name(), "shape", t.Shape())
+				}
+			}
+
+			kTensor := t.Clone()
+			kTensor.SetRepacker(p.repackKVB(true, kvFirst, numHeads))
+			out = append(out, &ggml.Tensor{
+				Name:     strings.Replace(t.Name(), "attn_kv_b", "attn_k_b", 1),
+				Kind:     t.Kind(),
+				Shape:    []uint64{uint64(numHeads), uint64(kvLoraRank), uint64(qkNope)},
+				WriterTo: kTensor,
+			})
+
+			vTensor := t.Clone()
+			vTensor.SetRepacker(p.repackKVB(false, kvFirst, numHeads))
+			out = append(out, &ggml.Tensor{
+				Name:     strings.Replace(t.Name(), "attn_kv_b", "attn_v_b", 1),
+				Kind:     t.Kind(),
+				Shape:    []uint64{uint64(numHeads), uint64(vHeadDim), uint64(kvLoraRank)},
+				WriterTo: vTensor,
+			})
+			continue
+		}
+
+		out = append(out, &ggml.Tensor{
+			Name:     t.Name(),
+			Kind:     t.Kind(),
+			Shape:    t.Shape(),
+			WriterTo: t,
+		})
+	}
+	return out
+}

convert/convert_lfm2.go

@@ -0,0 +1,100 @@
+package convert
+
+import (
+	"slices"
+	"strings"
+
+	"github.com/ollama/ollama/fs/ggml"
+)
+
+type lfm2Model struct {
+	ModelParameters
+	HiddenSize            uint32   `json:"hidden_size"`
+	NumHiddenLayers       uint32   `json:"num_hidden_layers"`
+	MaxPositionEmbeddings uint32   `json:"max_position_embeddings"`
+	IntermediateSize      uint32   `json:"intermediate_size"`
+	NumAttentionHeads     uint32   `json:"num_attention_heads"`
+	NumKeyValueHeads      uint32   `json:"num_key_value_heads"`
+	RopeTheta             float32  `json:"rope_theta"`
+	NormEps               float32  `json:"norm_eps"`
+	ConvLCache            uint32   `json:"conv_L_cache"`
+	LayerTypes            []string `json:"layer_types"`
+	TieEmbedding          bool     `json:"tie_embedding"`
+}
+
+var _ ModelConverter = (*lfm2Model)(nil)
+
+func (p *lfm2Model) KV(t *Tokenizer) KV {
+	kv := p.ModelParameters.KV(t)
+	kv["general.architecture"] = "lfm2"
+	kv["lfm2.vocab_size"] = p.VocabSize
+	kv["lfm2.block_count"] = p.NumHiddenLayers
+	kv["lfm2.embedding_length"] = p.HiddenSize
+	kv["lfm2.feed_forward_length"] = p.IntermediateSize
+	kv["lfm2.context_length"] = p.MaxPositionEmbeddings
+
+	// Build per-layer KV head count array based on layer_types
+	// (0 = shortconv layer, non-zero = attention layer with that many KV heads)
+	kvHeadCounts := make([]uint32, p.NumHiddenLayers)
+	for i := range p.NumHiddenLayers {
+		if int(i) < len(p.LayerTypes) && p.LayerTypes[i] == "full_attention" {
+			kvHeadCounts[i] = p.NumKeyValueHeads
+		}
+	}
+
+	kv["lfm2.attention.head_count"] = p.NumAttentionHeads
+	kv["lfm2.attention.head_count_kv"] = kvHeadCounts
+	kv["lfm2.attention.key_length"] = p.HiddenSize / p.NumAttentionHeads
+	kv["lfm2.attention.value_length"] = p.HiddenSize / p.NumAttentionHeads
+	kv["lfm2.attention.layer_norm_rms_epsilon"] = p.NormEps
+	kv["lfm2.rope.freq_base"] = p.RopeTheta
+	kv["lfm2.shortconv.l_cache"] = p.ConvLCache
+
+	return kv
+}
+
+func (p *lfm2Model) Tensors(ts []Tensor) []*ggml.Tensor {
+	var out []*ggml.Tensor
+
+	for _, t := range ts {
+		shape := t.Shape()
+
+		// Squeeze conv weights: [D, 1, K] -> [D, K]
+		if strings.HasSuffix(t.Name(), "shortconv.conv.weight") {
+			if len(shape) == 3 && shape[1] == 1 {
+				shape = []uint64{shape[0], shape[2]}
+			}
+		}
+
+		out = append(out, &ggml.Tensor{
+			Name:     t.Name(),
+			Kind:     t.Kind(),
+			Shape:    slices.Clone(shape),
+			WriterTo: t,
+		})
+	}
+
+	return out
+}
+
+func (p *lfm2Model) Replacements() []string {
+	return []string{
+		"model.embed_tokens", "token_embd",
+		"model.embedding_norm", "output_norm",
+		"model.layers", "blk",
+		"operator_norm", "attn_norm",
+		"self_attn.q_proj", "attn_q",
+		"self_attn.k_proj", "attn_k",
+		"self_attn.v_proj", "attn_v",
+		"self_attn.out_proj", "attn_output",
+		"self_attn.q_layernorm", "attn_q_norm",
+		"self_attn.k_layernorm", "attn_k_norm",
+		"conv.conv", "shortconv.conv",
+		"conv.in_proj", "shortconv.in_proj",
+		"conv.out_proj", "shortconv.out_proj",
+		"feed_forward.w1", "ffn_gate",
+		"feed_forward.w2", "ffn_down",
+		"feed_forward.w3", "ffn_up",
+		"ffn_norm", "ffn_norm",
+	}
+}

convert/reader.go

@@ -40,6 +40,7 @@ const (
 func (t tensorBase) Kind() uint32 {
 	if strings.HasSuffix(t.name, ".ffn_gate_inp.weight") ||
 		strings.HasSuffix(t.name, ".bias") ||
+		strings.HasSuffix(t.name, ".shortconv.conv.weight") ||
 		t.name == "token_types.weight" ||
 		t.name == "v.positional_embedding_vlm" ||
 		t.name == "v.tile_position_embd.weight" ||

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

@@ -4,16 +4,6 @@ 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
@@ -21,8 +11,9 @@ ollama pull glm-4.7:cloud
 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_API_KEY="" # required but ignored
 export ANTHROPIC_BASE_URL=http://localhost:11434
-export ANTHROPIC_API_KEY=ollama  # required but ignored
 ```
 
 ### Simple `/v1/messages` example
@@ -244,29 +235,55 @@ curl -X POST http://localhost:11434/v1/messages \
 
 ## Using with Claude Code
 
-[Claude Code](https://code.claude.com/docs/en/overview) can be configured to use Ollama as its backend:
+[Claude Code](https://code.claude.com/docs/en/overview) can be configured to use Ollama as its backend. 
+
+### Recommended models
+
+For coding use cases, models like `glm-4.7`, `minimax-m2.1`, and `qwen3-coder` are recommended.
+
+Download a model before use:
 
 ```shell
-ANTHROPIC_BASE_URL=http://localhost:11434 ANTHROPIC_API_KEY=ollama claude --model qwen3-coder
+ollama pull qwen3-coder
+```
+> Note: Qwen 3 coder is a 30B parameter model requiring at least 24GB of VRAM to run smoothly. More is required for longer context lengths. 
+
+```shell
+ollama pull glm-4.7:cloud
+```
+
+### Quick setup
+
+```shell
+ollama launch claude
+```
+
+This will prompt you to select a model, configure Claude Code automatically, and launch it. To configure without launching:
+
+```shell
+ollama launch claude --config
+```
+
+### Manual setup
+
+Set the environment variables and run Claude Code:
+
+```shell
+ANTHROPIC_AUTH_TOKEN=ollama ANTHROPIC_BASE_URL=http://localhost:11434 ANTHROPIC_API_KEY="" 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
+export ANTHROPIC_API_KEY=""
 ```
 
 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

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/cli.mdx

@@ -8,6 +8,47 @@ title: CLI Reference
 ollama run gemma3
 ```
 
+### Launch integrations
+
+```
+ollama launch
+```
+
+Configure and launch external applications to use Ollama models. This provides an interactive way to set up and start integrations with supported apps.
+
+#### Supported integrations
+
+- **OpenCode** - Open-source coding assistant
+- **Claude Code** - Anthropic's agentic coding tool
+- **Codex** - OpenAI's coding assistant
+- **Droid** - Factory's AI coding agent
+
+#### Examples
+
+Launch an integration interactively:
+
+```
+ollama launch
+```
+
+Launch a specific integration:
+
+```
+ollama launch claude
+```
+
+Launch with a specific model:
+
+```
+ollama launch claude --model qwen3-coder
+```
+
+Configure without launching:
+
+```
+ollama launch droid --config
+```
+
 #### Multiline input
 
 For multiline input, you can wrap text with `"""`:

docs/cloud.mdx

@@ -3,8 +3,6 @@ title: Cloud
 sidebarTitle: Cloud
 ---
 
-<Info>Ollama's cloud is currently in preview.</Info>
-
 ## Cloud Models
 
 Ollama's cloud models are a new kind of model in Ollama that can run without a powerful GPU. Instead, cloud models are automatically offloaded to Ollama's cloud service while offering the same capabilities as local models, making it possible to keep using your local tools while running larger models that wouldn't fit on a personal computer.

docs/context-length.mdx

@@ -8,7 +8,7 @@ Context length is the maximum number of tokens that the model has access to in m
   The default context length in Ollama is 4096 tokens.
 </Note>
 
-Tasks which require large context like web search, agents, and coding tools should be set to at least 32000 tokens.
+Tasks which require large context like web search, agents, and coding tools should be set to at least 64000 tokens.
 
 ## Setting context length
 
@@ -24,7 +24,7 @@ Change the slider in the Ollama app under settings to your desired context lengt
 ### CLI
 If editing the context length for Ollama is not possible, the context length can also be updated when serving Ollama.  
 ```
-OLLAMA_CONTEXT_LENGTH=32000 ollama serve
+OLLAMA_CONTEXT_LENGTH=64000 ollama serve
 ```
 
 ### Check allocated context length and model offloading

docs/docs.json

@@ -102,16 +102,19 @@
             "group": "Integrations",
             "pages": [
               "/integrations/claude-code",
-              "/integrations/vscode",
-              "/integrations/jetbrains",
-              "/integrations/codex",
               "/integrations/cline",
+              "/integrations/codex",
               "/integrations/droid",
               "/integrations/goose",
-              "/integrations/zed",
-              "/integrations/roo-code",
+              "/integrations/jetbrains",
+              "/integrations/marimo",
               "/integrations/n8n",
-              "/integrations/xcode"
+              "/integrations/onyx",
+              "/integrations/opencode",
+              "/integrations/roo-code",
+              "/integrations/vscode",
+              "/integrations/xcode",
+              "/integrations/zed"
             ]
           },
           {

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/index.mdx

@@ -9,7 +9,7 @@ sidebarTitle: Welcome
 
 <CardGroup cols={2}>
   <Card title="Quickstart" icon="rocket" href="/quickstart">
-    Get up and running with your first model
+    Get up and running with your first model or integrate Ollama with your favorite tools
   </Card>
   <Card
     title="Download Ollama"

docs/integrations/claude-code.mdx

@@ -2,6 +2,12 @@
 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 `glm-4.7`, `qwen3-coder`, `gpt-oss`. 
+
+![Claude Code with Ollama](https://files.ollama.com/claude-code.png)
+
 ## Install
 
 Install [Claude Code](https://code.claude.com/docs/en/overview):
@@ -20,50 +26,50 @@ irm https://claude.ai/install.ps1 | iex
 
 ## Usage with Ollama
 
+### Quick setup
+
+```shell
+ollama launch claude
+```
+
+To configure without launching:
+
+```shell
+ollama launch claude --config
+```
+
+### Manual setup
+
 Claude Code connects to Ollama using the Anthropic-compatible API.
 
 1. Set the environment variables:
 
 ```shell
+export ANTHROPIC_AUTH_TOKEN=ollama
+export ANTHROPIC_API_KEY=""
 export ANTHROPIC_BASE_URL=http://localhost:11434
-export ANTHROPIC_API_KEY=ollama
 ```
 
 2. Run Claude Code with an Ollama model:
 
 ```shell
-claude --model qwen3-coder
+claude --model gpt-oss:20b
 ```
 
 Or run with environment variables inline:
 
 ```shell
-ANTHROPIC_BASE_URL=http://localhost:11434 ANTHROPIC_API_KEY=ollama claude --model qwen3-coder
+ANTHROPIC_AUTH_TOKEN=ollama ANTHROPIC_BASE_URL=http://localhost:11434 ANTHROPIC_API_KEY="" claude --model qwen3-coder 
 ```
 
-## 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
-```
+**Note:** Claude Code requires a large context window. We recommend at least 64k tokens. See the [context length documentation](/context-length) for how to adjust context length in Ollama.
 
 ## 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
+- `qwen3-coder` 
+- `glm-4.7`
+- `gpt-oss:20b`
+- `gpt-oss:120b`
+
+Cloud models are also available at [ollama.com/search?c=cloud](https://ollama.com/search?c=cloud).
 
-### Local models
-- `qwen3-coder` - Excellent for coding tasks
-- `gpt-oss:20b` - Strong general-purpose model

docs/integrations/codex.mdx

@@ -13,7 +13,21 @@ npm install -g @openai/codex
 
 ## Usage with Ollama
 
-<Note>Codex requires a larger context window. It is recommended to use a context window of at least 32K tokens.</Note>
+<Note>Codex requires a larger context window. It is recommended to use a context window of at least 64k tokens.</Note>
+
+### Quick setup
+
+```
+ollama launch codex
+```
+
+To configure without launching:
+
+```shell
+ollama launch codex --config
+```
+
+### Manual setup
 
 To use `codex` with Ollama, use the `--oss` flag:
 

docs/integrations/droid.mdx

@@ -11,10 +11,24 @@ Install the [Droid CLI](https://factory.ai/):
 curl -fsSL https://app.factory.ai/cli | sh
 ```
 
-<Note>Droid requires a larger context window. It is recommended to use a context window of at least 32K tokens. See [Context length](/context-length) for more information.</Note>
+<Note>Droid requires a larger context window. It is recommended to use a context window of at least 64k tokens. See [Context length](/context-length) for more information.</Note>
 
 ## Usage with Ollama
 
+### Quick setup
+
+```bash
+ollama launch droid
+```
+
+To configure without launching:
+
+```shell
+ollama launch droid --config
+```
+
+### Manual setup
+
 Add a local configuration block to `~/.factory/config.json`:
 
 ```json
@@ -73,4 +87,4 @@ Add the cloud configuration block to `~/.factory/config.json`:
    }
    ```
 
-Run `droid` in a new terminal to load the new settings.
+Run `droid` in a new terminal to load the new settings.

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/integrations/opencode.mdx

@@ -0,0 +1,106 @@
+---
+title: OpenCode
+---
+
+OpenCode is an open-source AI coding assistant that runs in your terminal.
+
+## Install
+
+Install the [OpenCode CLI](https://opencode.ai):
+
+```bash
+curl -fsSL https://opencode.ai/install.sh | bash
+```
+
+<Note>OpenCode requires a larger context window. It is recommended to use a context window of at least 64k tokens. See [Context length](/context-length) for more information.</Note>
+
+## Usage with Ollama
+
+### Quick setup
+
+```bash
+ollama launch opencode
+```
+
+To configure without launching:
+
+```shell
+ollama launch opencode --config
+```
+
+### Manual setup
+
+Add a configuration block to `~/.config/opencode/opencode.json`:
+
+```json
+{
+  "$schema": "https://opencode.ai/config.json",
+  "provider": {
+    "ollama": {
+      "npm": "@ai-sdk/openai-compatible",
+      "name": "Ollama",
+      "options": {
+        "baseURL": "http://localhost:11434/v1"
+      },
+      "models": {
+        "qwen3-coder": {
+          "name": "qwen3-coder"
+        }
+      }
+    }
+  }
+}
+```
+
+## Cloud Models
+
+`glm-4.7:cloud` is the recommended model for use with OpenCode.
+
+Add the cloud configuration to `~/.config/opencode/opencode.json`:
+
+```json
+{
+  "$schema": "https://opencode.ai/config.json",
+  "provider": {
+    "ollama": {
+      "npm": "@ai-sdk/openai-compatible",
+      "name": "Ollama",
+      "options": {
+        "baseURL": "http://localhost:11434/v1"
+      },
+      "models": {
+        "glm-4.7:cloud": {
+          "name": "glm-4.7:cloud"
+        }
+      }
+    }
+  }
+}
+```
+
+## Connecting to ollama.com
+
+1. Create an [API key](https://ollama.com/settings/keys) from ollama.com and export it as `OLLAMA_API_KEY`.
+2. Update `~/.config/opencode/opencode.json` to point to ollama.com:
+
+```json
+{
+  "$schema": "https://opencode.ai/config.json",
+  "provider": {
+    "ollama": {
+      "npm": "@ai-sdk/openai-compatible",
+      "name": "Ollama Cloud",
+      "options": {
+        "baseURL": "https://ollama.com/v1"
+      },
+      "models": {
+        "glm-4.7:cloud": {
+          "name": "glm-4.7:cloud"
+        }
+      }
+    }
+  }
+}
+```
+
+Run `opencode` in a new terminal to load the new settings.

docs/linux.mdx

@@ -1,5 +1,5 @@
 ---
-title: "Linux"
+title: Linux
 ---
 
 ## Install
@@ -13,14 +13,15 @@ curl -fsSL https://ollama.com/install.sh | sh
 ## Manual install
 
 <Note>
-  If you are upgrading from a prior version, you should remove the old libraries with `sudo rm -rf /usr/lib/ollama` first.
+  If you are upgrading from a prior version, you should remove the old libraries
+  with `sudo rm -rf /usr/lib/ollama` first.
 </Note>
 
 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:
@@ -40,8 +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
@@ -49,8 +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)
@@ -112,7 +113,11 @@ sudo systemctl status ollama
 ```
 
 <Note>
-  While AMD has contributed the `amdgpu` driver upstream to the official linux kernel source, the version is older and may not support all ROCm features. We recommend you install the latest driver from https://www.amd.com/en/support/linux-drivers for best support of your Radeon GPU.
+  While AMD has contributed the `amdgpu` driver upstream to the official linux
+  kernel source, the version is older and may not support all ROCm features. We
+  recommend you install the latest driver from
+  https://www.amd.com/en/support/linux-drivers for best support of your Radeon
+  GPU.
 </Note>
 
 ## Customizing
@@ -141,8 +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
@@ -191,4 +196,4 @@ Remove the downloaded models and Ollama service user and group:
 sudo userdel ollama
 sudo groupdel ollama
 sudo rm -r /usr/share/ollama
-```
+```

docs/quickstart.mdx

@@ -18,13 +18,13 @@ This quickstart will walk your through running your first model with Ollama. To
   <Tab title="CLI">
     Open a terminal and run the command:
 
-    ```
+    ```sh
     ollama run gemma3
     ```
 
   </Tab>
   <Tab title="cURL">
-    ```
+    ```sh
     ollama pull gemma3
     ```
 
@@ -45,13 +45,13 @@ This quickstart will walk your through running your first model with Ollama. To
   <Tab title="Python">
     Start by downloading a model:
 
-    ```
+    ```sh
     ollama pull gemma3
     ```
 
     Then install Ollama's Python library:
 
-    ```
+    ```sh
     pip install ollama
     ```
 
@@ -101,3 +101,42 @@ This quickstart will walk your through running your first model with Ollama. To
 </Tabs>
 
 See a full list of available models [here](https://ollama.com/models).
+
+## Coding 
+
+For coding use cases, we recommend using the `glm-4.7-flash` model. 
+
+Note: this model requires 23 GB of VRAM with 64000 tokens context length.
+```sh
+ollama pull glm-4.7-flash 
+```
+
+Alternatively, you can use a more powerful cloud model (with full context length):
+```sh
+ollama pull glm-4.7:cloud
+```
+
+Use `ollama launch` to quickly set up a coding tool with Ollama models:
+
+```sh
+ollama launch
+```
+
+### Supported integrations
+
+- [OpenCode](/integrations/opencode) - Open-source coding assistant
+- [Claude Code](/integrations/claude-code) - Anthropic's agentic coding tool
+- [Codex](/integrations/codex) - OpenAI's coding assistant
+- [Droid](/integrations/droid) - Factory's AI coding agent
+
+### Launch with a specific model
+
+```sh
+ollama launch claude --model glm-4.7-flash
+```
+
+### Configure without launching
+
+```sh
+ollama launch claude --config
+```

docs/troubleshooting.md

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

fs/ggml/ggml.go

@@ -269,6 +269,8 @@ func (kv KV) OllamaEngineRequired() bool {
 		"qwen25vl",
 		"qwen3", "qwen3moe",
 		"qwen3vl", "qwen3vlmoe",
+		"glm4moelite",
+		"lfm2",
 	}, kv.Architecture())
 }
 
@@ -856,7 +858,9 @@ func (f GGML) FlashAttention() bool {
 	return slices.Contains([]string{
 		"bert",
 		"gemma3",
+		"glm4moelite",
 		"gptoss", "gpt-oss",
+		"lfm2",
 		"mistral3",
 		"olmo3",
 		"qwen3", "qwen3moe",

go.mod

@@ -15,8 +15,8 @@ require (
 	github.com/spf13/cobra v1.7.0
 	github.com/stretchr/testify v1.9.0
 	github.com/x448/float16 v0.8.4
-	golang.org/x/sync v0.19.0
-	golang.org/x/sys v0.39.0
+	golang.org/x/sync v0.17.0
+	golang.org/x/sys v0.37.0
 )
 
 require (
@@ -30,8 +30,8 @@ require (
 	github.com/tkrajina/typescriptify-golang-structs v0.2.0
 	github.com/wk8/go-ordered-map/v2 v2.1.8
 	golang.org/x/image v0.22.0
-	golang.org/x/mod v0.31.0
-	golang.org/x/tools v0.40.0
+	golang.org/x/mod v0.30.0
+	golang.org/x/tools v0.38.0
 	gonum.org/v1/gonum v0.15.0
 )
 
@@ -81,11 +81,11 @@ require (
 	github.com/twitchyliquid64/golang-asm v0.15.1 // indirect
 	github.com/ugorji/go/codec v1.2.12 // indirect
 	golang.org/x/arch v0.8.0 // indirect
-	golang.org/x/crypto v0.46.0
-	golang.org/x/exp v0.0.0-20251219203646-944ab1f22d93
-	golang.org/x/net v0.48.0 // indirect
-	golang.org/x/term v0.38.0
-	golang.org/x/text v0.32.0
+	golang.org/x/crypto v0.43.0
+	golang.org/x/exp v0.0.0-20250218142911-aa4b98e5adaa // indirect
+	golang.org/x/net v0.46.0 // indirect
+	golang.org/x/term v0.36.0
+	golang.org/x/text v0.30.0
 	google.golang.org/protobuf v1.34.1
 	gopkg.in/yaml.v3 v3.0.1 // indirect
 )

go.sum

@@ -233,16 +233,16 @@ golang.org/x/crypto v0.0.0-20190308221718-c2843e01d9a2/go.mod h1:djNgcEr1/C05ACk
 golang.org/x/crypto v0.0.0-20190510104115-cbcb75029529/go.mod h1:yigFU9vqHzYiE8UmvKecakEJjdnWj3jj499lnFckfCI=
 golang.org/x/crypto v0.0.0-20191011191535-87dc89f01550/go.mod h1:yigFU9vqHzYiE8UmvKecakEJjdnWj3jj499lnFckfCI=
 golang.org/x/crypto v0.0.0-20200622213623-75b288015ac9/go.mod h1:LzIPMQfyMNhhGPhUkYOs5KpL4U8rLKemX1yGLhDgUto=
-golang.org/x/crypto v0.46.0 h1:cKRW/pmt1pKAfetfu+RCEvjvZkA9RimPbh7bhFjGVBU=
-golang.org/x/crypto v0.46.0/go.mod h1:Evb/oLKmMraqjZ2iQTwDwvCtJkczlDuTmdJXoZVzqU0=
+golang.org/x/crypto v0.43.0 h1:dduJYIi3A3KOfdGOHX8AVZ/jGiyPa3IbBozJ5kNuE04=
+golang.org/x/crypto v0.43.0/go.mod h1:BFbav4mRNlXJL4wNeejLpWxB7wMbc79PdRGhWKncxR0=
 golang.org/x/exp v0.0.0-20180321215751-8460e604b9de/go.mod h1:CJ0aWSM057203Lf6IL+f9T1iT9GByDxfZKAQTCR3kQA=
 golang.org/x/exp v0.0.0-20180807140117-3d87b88a115f/go.mod h1:CJ0aWSM057203Lf6IL+f9T1iT9GByDxfZKAQTCR3kQA=
 golang.org/x/exp v0.0.0-20190121172915-509febef88a4/go.mod h1:CJ0aWSM057203Lf6IL+f9T1iT9GByDxfZKAQTCR3kQA=
 golang.org/x/exp v0.0.0-20190125153040-c74c464bbbf2/go.mod h1:CJ0aWSM057203Lf6IL+f9T1iT9GByDxfZKAQTCR3kQA=
 golang.org/x/exp v0.0.0-20190306152737-a1d7652674e8/go.mod h1:CJ0aWSM057203Lf6IL+f9T1iT9GByDxfZKAQTCR3kQA=
 golang.org/x/exp v0.0.0-20191002040644-a1355ae1e2c3/go.mod h1:NOZ3BPKG0ec/BKJQgnvsSFpcKLM5xXVWnvZS97DWHgE=
-golang.org/x/exp v0.0.0-20251219203646-944ab1f22d93 h1:fQsdNF2N+/YewlRZiricy4P1iimyPKZ/xwniHj8Q2a0=
-golang.org/x/exp v0.0.0-20251219203646-944ab1f22d93/go.mod h1:EPRbTFwzwjXj9NpYyyrvenVh9Y+GFeEvMNh7Xuz7xgU=
+golang.org/x/exp v0.0.0-20250218142911-aa4b98e5adaa h1:t2QcU6V556bFjYgu4L6C+6VrCPyJZ+eyRsABUPs1mz4=
+golang.org/x/exp v0.0.0-20250218142911-aa4b98e5adaa/go.mod h1:BHOTPb3L19zxehTsLoJXVaTktb06DFgmdW6Wb9s8jqk=
 golang.org/x/image v0.0.0-20180708004352-c73c2afc3b81/go.mod h1:ux5Hcp/YLpHSI86hEcLt0YII63i6oz57MZXIpbrjZUs=
 golang.org/x/image v0.0.0-20190227222117-0694c2d4d067/go.mod h1:kZ7UVZpmo3dzQBMxlp+ypCbDeSB+sBbTgSJuh5dn5js=
 golang.org/x/image v0.0.0-20190802002840-cff245a6509b/go.mod h1:FeLwcggjj3mMvU+oOTbSwawSJRM1uh48EjtB4UJZlP0=
@@ -264,8 +264,8 @@ golang.org/x/mod v0.1.1-0.20191105210325-c90efee705ee/go.mod h1:QqPTAvyqsEbceGzB
 golang.org/x/mod v0.2.0/go.mod h1:s0Qsj1ACt9ePp/hMypM3fl4fZqREWJwdYDEqhRiZZUA=
 golang.org/x/mod v0.3.0/go.mod h1:s0Qsj1ACt9ePp/hMypM3fl4fZqREWJwdYDEqhRiZZUA=
 golang.org/x/mod v0.4.2/go.mod h1:s0Qsj1ACt9ePp/hMypM3fl4fZqREWJwdYDEqhRiZZUA=
-golang.org/x/mod v0.31.0 h1:HaW9xtz0+kOcWKwli0ZXy79Ix+UW/vOfmWI5QVd2tgI=
-golang.org/x/mod v0.31.0/go.mod h1:43JraMp9cGx1Rx3AqioxrbrhNsLl2l/iNAvuBkrezpg=
+golang.org/x/mod v0.30.0 h1:fDEXFVZ/fmCKProc/yAXXUijritrDzahmwwefnjoPFk=
+golang.org/x/mod v0.30.0/go.mod h1:lAsf5O2EvJeSFMiBxXDki7sCgAxEUcZHXoXMKT4GJKc=
 golang.org/x/net v0.0.0-20180724234803-3673e40ba225/go.mod h1:mL1N/T3taQHkDXs73rZJwtUhF3w3ftmwwsq0BUmARs4=
 golang.org/x/net v0.0.0-20180826012351-8a410e7b638d/go.mod h1:mL1N/T3taQHkDXs73rZJwtUhF3w3ftmwwsq0BUmARs4=
 golang.org/x/net v0.0.0-20190108225652-1e06a53dbb7e/go.mod h1:mL1N/T3taQHkDXs73rZJwtUhF3w3ftmwwsq0BUmARs4=
@@ -278,8 +278,8 @@ golang.org/x/net v0.0.0-20200822124328-c89045814202/go.mod h1:/O7V0waA8r7cgGh81R
 golang.org/x/net v0.0.0-20201021035429-f5854403a974/go.mod h1:sp8m0HH+o8qH0wwXwYZr8TS3Oi6o0r6Gce1SSxlDquU=
 golang.org/x/net v0.0.0-20210405180319-a5a99cb37ef4/go.mod h1:p54w0d4576C0XHj96bSt6lcn1PtDYWL6XObtHCRCNQM=
 golang.org/x/net v0.0.0-20210614182718-04defd469f4e/go.mod h1:9nx3DQGgdP8bBQD5qxJ1jj9UTztislL4KSBs9R2vV5Y=
-golang.org/x/net v0.48.0 h1:zyQRTTrjc33Lhh0fBgT/H3oZq9WuvRR5gPC70xpDiQU=
-golang.org/x/net v0.48.0/go.mod h1:+ndRgGjkh8FGtu1w1FGbEC31if4VrNVMuKTgcAAnQRY=
+golang.org/x/net v0.46.0 h1:giFlY12I07fugqwPuWJi68oOnpfqFnJIJzaIIm2JVV4=
+golang.org/x/net v0.46.0/go.mod h1:Q9BGdFy1y4nkUwiLvT5qtyhAnEHgnQ/zd8PfU6nc210=
 golang.org/x/oauth2 v0.0.0-20180821212333-d2e6202438be/go.mod h1:N/0e6XlmueqKjAGxoOufVs8QHGRruUQn6yWY3a++T0U=
 golang.org/x/oauth2 v0.0.0-20200107190931-bf48bf16ab8d/go.mod h1:gOpvHmFTYa4IltrdGE7lF6nIHvwfUNPOp7c8zoXwtLw=
 golang.org/x/sync v0.0.0-20180314180146-1d60e4601c6f/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
@@ -289,8 +289,8 @@ golang.org/x/sync v0.0.0-20190423024810-112230192c58/go.mod h1:RxMgew5VJxzue5/jJ
 golang.org/x/sync v0.0.0-20190911185100-cd5d95a43a6e/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
 golang.org/x/sync v0.0.0-20201020160332-67f06af15bc9/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
 golang.org/x/sync v0.0.0-20210220032951-036812b2e83c/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
-golang.org/x/sync v0.19.0 h1:vV+1eWNmZ5geRlYjzm2adRgW2/mcpevXNg50YZtPCE4=
-golang.org/x/sync v0.19.0/go.mod h1:9KTHXmSnoGruLpwFjVSX0lNNA75CykiMECbovNTZqGI=
+golang.org/x/sync v0.17.0 h1:l60nONMj9l5drqw6jlhIELNv9I0A4OFgRsG9k2oT9Ug=
+golang.org/x/sync v0.17.0/go.mod h1:9KTHXmSnoGruLpwFjVSX0lNNA75CykiMECbovNTZqGI=
 golang.org/x/sys v0.0.0-20180830151530-49385e6e1522/go.mod h1:STP8DvDyc/dI5b8T5hshtkjS+E42TnysNCUPdjciGhY=
 golang.org/x/sys v0.0.0-20190215142949-d0b11bdaac8a/go.mod h1:STP8DvDyc/dI5b8T5hshtkjS+E42TnysNCUPdjciGhY=
 golang.org/x/sys v0.0.0-20190312061237-fead79001313/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=
@@ -306,17 +306,17 @@ golang.org/x/sys v0.0.0-20210510120138-977fb7262007/go.mod h1:oPkhp1MJrh7nUepCBc
 golang.org/x/sys v0.0.0-20210630005230-0f9fa26af87c/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.5.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.6.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
-golang.org/x/sys v0.39.0 h1:CvCKL8MeisomCi6qNZ+wbb0DN9E5AATixKsvNtMoMFk=
-golang.org/x/sys v0.39.0/go.mod h1:OgkHotnGiDImocRcuBABYBEXf8A9a87e/uXjp9XT3ks=
+golang.org/x/sys v0.37.0 h1:fdNQudmxPjkdUTPnLn5mdQv7Zwvbvpaxqs831goi9kQ=
+golang.org/x/sys v0.37.0/go.mod h1:OgkHotnGiDImocRcuBABYBEXf8A9a87e/uXjp9XT3ks=
 golang.org/x/term v0.0.0-20201126162022-7de9c90e9dd1/go.mod h1:bj7SfCRtBDWHUb9snDiAeCFNEtKQo2Wmx5Cou7ajbmo=
-golang.org/x/term v0.38.0 h1:PQ5pkm/rLO6HnxFR7N2lJHOZX6Kez5Y1gDSJla6jo7Q=
-golang.org/x/term v0.38.0/go.mod h1:bSEAKrOT1W+VSu9TSCMtoGEOUcKxOKgl3LE5QEF/xVg=
+golang.org/x/term v0.36.0 h1:zMPR+aF8gfksFprF/Nc/rd1wRS1EI6nDBGyWAvDzx2Q=
+golang.org/x/term v0.36.0/go.mod h1:Qu394IJq6V6dCBRgwqshf3mPF85AqzYEzofzRdZkWss=
 golang.org/x/text v0.3.0/go.mod h1:NqM8EUOU14njkJ3fqMW+pc6Ldnwhi/IjpwHt7yyuwOQ=
 golang.org/x/text v0.3.3/go.mod h1:5Zoc/QRtKVWzQhOtBMvqHzDpF6irO9z98xDceosuGiQ=
 golang.org/x/text v0.3.5/go.mod h1:5Zoc/QRtKVWzQhOtBMvqHzDpF6irO9z98xDceosuGiQ=
 golang.org/x/text v0.3.6/go.mod h1:5Zoc/QRtKVWzQhOtBMvqHzDpF6irO9z98xDceosuGiQ=
-golang.org/x/text v0.32.0 h1:ZD01bjUt1FQ9WJ0ClOL5vxgxOI/sVCNgX1YtKwcY0mU=
-golang.org/x/text v0.32.0/go.mod h1:o/rUWzghvpD5TXrTIBuJU77MTaN0ljMWE47kxGJQ7jY=
+golang.org/x/text v0.30.0 h1:yznKA/E9zq54KzlzBEAWn1NXSQ8DIp/NYMy88xJjl4k=
+golang.org/x/text v0.30.0/go.mod h1:yDdHFIX9t+tORqspjENWgzaCVXgk0yYnYuSZ8UzzBVM=
 golang.org/x/tools v0.0.0-20180525024113-a5b4c53f6e8b/go.mod h1:n7NCudcB/nEzxVGmLbDWY5pfWTLqBcC2KZ6jyYvM4mQ=
 golang.org/x/tools v0.0.0-20180917221912-90fa682c2a6e/go.mod h1:n7NCudcB/nEzxVGmLbDWY5pfWTLqBcC2KZ6jyYvM4mQ=
 golang.org/x/tools v0.0.0-20190114222345-bf090417da8b/go.mod h1:n7NCudcB/nEzxVGmLbDWY5pfWTLqBcC2KZ6jyYvM4mQ=
@@ -330,8 +330,8 @@ golang.org/x/tools v0.0.0-20200130002326-2f3ba24bd6e7/go.mod h1:TB2adYChydJhpapK
 golang.org/x/tools v0.0.0-20200619180055-7c47624df98f/go.mod h1:EkVYQZoAsY45+roYkvgYkIh4xh/qjgUK9TdY2XT94GE=
 golang.org/x/tools v0.0.0-20210106214847-113979e3529a/go.mod h1:emZCQorbCU4vsT4fOWvOPXz4eW1wZW4PmDk9uLelYpA=
 golang.org/x/tools v0.1.4/go.mod h1:o0xws9oXOQQZyjljx8fwUC0k7L1pTE6eaCbjGeHmOkk=
-golang.org/x/tools v0.40.0 h1:yLkxfA+Qnul4cs9QA3KnlFu0lVmd8JJfoq+E41uSutA=
-golang.org/x/tools v0.40.0/go.mod h1:Ik/tzLRlbscWpqqMRjyWYDisX8bG13FrdXp3o4Sr9lc=
+golang.org/x/tools v0.38.0 h1:Hx2Xv8hISq8Lm16jvBZ2VQf+RLmbd7wVUsALibYI/IQ=
+golang.org/x/tools v0.38.0/go.mod h1:yEsQ/d/YK8cjh0L6rZlY8tgtlKiBNTL14pGDJPJpYQs=
 golang.org/x/xerrors v0.0.0-20190717185122-a985d3407aa7/go.mod h1:I/5z698sn9Ka8TeJc9MKroUUfqBBauWjQqLJ2OPfmY0=
 golang.org/x/xerrors v0.0.0-20191011141410-1b5146add898/go.mod h1:I/5z698sn9Ka8TeJc9MKroUUfqBBauWjQqLJ2OPfmY0=
 golang.org/x/xerrors v0.0.0-20191204190536-9bdfabe68543/go.mod h1:I/5z698sn9Ka8TeJc9MKroUUfqBBauWjQqLJ2OPfmY0=

integration/imagegen_test.go

@@ -0,0 +1,148 @@
+//go:build integration
+
+package integration
+
+import (
+	"context"
+	"encoding/base64"
+	"fmt"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/ollama/ollama/api"
+)
+
+func TestImageGeneration(t *testing.T) {
+	skipUnderMinVRAM(t, 8)
+
+	type testCase struct {
+		imageGenModel string
+		visionModel   string
+		prompt        string
+		expectedWords []string
+	}
+
+	testCases := []testCase{
+		{
+			imageGenModel: "jmorgan/z-image-turbo",
+			visionModel:   "llama3.2-vision",
+			prompt:        "A cartoon style llama flying like a superhero through the air with clouds in the background",
+			expectedWords: []string{"llama", "flying", "cartoon", "cloud", "sky", "superhero", "air", "animal", "camelid"},
+		},
+	}
+
+	for _, tc := range testCases {
+		t.Run(fmt.Sprintf("%s->%s", tc.imageGenModel, tc.visionModel), func(t *testing.T) {
+			ctx, cancel := context.WithTimeout(context.Background(), 10*time.Minute)
+			defer cancel()
+
+			client, _, cleanup := InitServerConnection(ctx, t)
+			defer cleanup()
+
+			// Pull both models
+			if err := PullIfMissing(ctx, client, tc.imageGenModel); err != nil {
+				t.Fatalf("failed to pull image gen model: %v", err)
+			}
+			if err := PullIfMissing(ctx, client, tc.visionModel); err != nil {
+				t.Fatalf("failed to pull vision model: %v", err)
+			}
+
+			// Generate the image
+			t.Logf("Generating image with prompt: %s", tc.prompt)
+			imageBase64, err := generateImage(ctx, client, tc.imageGenModel, tc.prompt)
+			if err != nil {
+				if strings.Contains(err.Error(), "image generation not available") {
+					t.Skip("Target system does not support image generation")
+				} else if strings.Contains(err.Error(), "executable file not found in") { // Windows pattern, not yet supported
+					t.Skip("Windows does not support image generation yet")
+				} else if strings.Contains(err.Error(), "CUDA driver version is insufficient") {
+					t.Skip("Driver is too old")
+				} else if strings.Contains(err.Error(), "insufficient memory for image generation") {
+					t.Skip("insufficient memory for image generation")
+				} else if strings.Contains(err.Error(), "error while loading shared libraries: libcuda.so.1") { // AMD GPU or CPU
+					t.Skip("CUDA GPU is not available")
+				} else if strings.Contains(err.Error(), "ollama-mlx: no such file or directory") {
+					// most likely linux arm - not supported yet
+					t.Skip("unsupported architecture")
+				}
+				t.Fatalf("failed to generate image: %v", err)
+			}
+
+			imageData, err := base64.StdEncoding.DecodeString(imageBase64)
+			if err != nil {
+				t.Fatalf("failed to decode image: %v", err)
+			}
+			t.Logf("Generated image: %d bytes", len(imageData))
+
+			// Preload vision model and check GPU loading
+			err = client.Generate(ctx, &api.GenerateRequest{Model: tc.visionModel}, func(response api.GenerateResponse) error { return nil })
+			if err != nil {
+				t.Fatalf("failed to load vision model: %v", err)
+			}
+
+			// Use vision model to describe the image
+			chatReq := api.ChatRequest{
+				Model: tc.visionModel,
+				Messages: []api.Message{
+					{
+						Role:    "user",
+						Content: "Describe this image in detail. What is shown? What style is it? What is the main subject doing?",
+						Images:  []api.ImageData{imageData},
+					},
+				},
+				Stream: &stream,
+				Options: map[string]any{
+					"seed":        42,
+					"temperature": 0.0,
+				},
+			}
+
+			// Verify the vision model's response contains expected keywords
+			response := DoChat(ctx, t, client, chatReq, tc.expectedWords, 240*time.Second, 30*time.Second)
+			if response != nil {
+				t.Logf("Vision model response: %s", response.Content)
+
+				// Additional detailed check for keywords
+				content := strings.ToLower(response.Content)
+				foundWords := []string{}
+				missingWords := []string{}
+				for _, word := range tc.expectedWords {
+					if strings.Contains(content, word) {
+						foundWords = append(foundWords, word)
+					} else {
+						missingWords = append(missingWords, word)
+					}
+				}
+				t.Logf("Found keywords: %v", foundWords)
+				if len(missingWords) > 0 {
+					t.Logf("Missing keywords (at least one was found so test passed): %v", missingWords)
+				}
+			}
+		})
+	}
+}
+
+// generateImage calls the Ollama API to generate an image and returns the base64 image data
+func generateImage(ctx context.Context, client *api.Client, model, prompt string) (string, error) {
+	var imageBase64 string
+
+	err := client.Generate(ctx, &api.GenerateRequest{
+		Model:  model,
+		Prompt: prompt,
+	}, func(resp api.GenerateResponse) error {
+		if resp.Image != "" {
+			imageBase64 = resp.Image
+		}
+		return nil
+	})
+	if err != nil {
+		return "", fmt.Errorf("failed to generate image: %w", err)
+	}
+
+	if imageBase64 == "" {
+		return "", fmt.Errorf("no image data in response")
+	}
+
+	return imageBase64, nil
+}

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():

integration/utils_test.go

@@ -38,6 +38,7 @@ var (
 
 	// Note: add newer models at the top of the list to test them first
 	ollamaEngineChatModels = []string{
+		"lfm2.5-thinking",
 		"ministral-3",
 		"qwen3-coder:30b",
 		"gpt-oss:20b",
@@ -143,6 +144,7 @@ var (
 		"granite3.3",
 		"hermes3",
 		"internlm2",
+		"lfm2.5-thinking",
 		"llama-guard3",
 		"llama-pro",
 		"llama2-chinese",
@@ -263,6 +265,7 @@ var (
 		"snowflake-arctic-embed2",
 	}
 	libraryToolsModels = []string{
+		"lfm2.5-thinking",
 		"qwen3-vl",
 		"gpt-oss:20b",
 		"gpt-oss:120b",

llama/patches/0032-ggml-enable-MLA-flash-attention-for-GLM-4.7-flash.patch

@@ -0,0 +1,309 @@
+From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
+From: nobody <>
+Date: Sat, 24 Jan 2026 02:31:01 +0000
+Subject: [PATCH] ggml: enable MLA flash attention for GLM-4.7-flash
+
+Add support for gqa_ratio 4 in MLA flash attention kernels. GLM-4.7-flash
+uses head size 576 with gqa_ratio 4, which was previously only supported
+for gqa_ratio 16 (DeepSeek).
+
+Metal changes:
+- Enable head size 576 for flash attention
+- Increase simdgroups to 8 for large heads (>=512)
+- Add case 8 kernel dispatch for 8 simdgroups
+
+CUDA changes:
+- Add gqa_ratio 4 support for head 576/512
+- Add tile configs for (576, 512, 4) and (576, 512, 8)
+- Add MMA config cases for ncols 4
+- Add template instances for ncols2=4
+- Fix nbatch_fa values in nvidia_fp32 config (32->64)
+---
+ ggml/src/ggml-cuda/fattn-mma-f16.cuh          | 40 +++++++++++++++----
+ ggml/src/ggml-cuda/fattn-tile.cuh             | 16 ++++++++
+ ggml/src/ggml-cuda/fattn.cu                   | 12 ++++--
+ ...ttn-mma-f16-instance-ncols1_16-ncols2_4.cu |  1 +
+ ...attn-mma-f16-instance-ncols1_2-ncols2_4.cu |  1 +
+ ...attn-mma-f16-instance-ncols1_4-ncols2_4.cu |  1 +
+ ...attn-mma-f16-instance-ncols1_8-ncols2_4.cu |  1 +
+ ggml/src/ggml-metal/ggml-metal-device.m       |  8 +---
+ ggml/src/ggml-metal/ggml-metal-ops.cpp        |  2 +-
+ ggml/src/ggml-metal/ggml-metal.metal          |  1 +
+ 10 files changed, 64 insertions(+), 19 deletions(-)
+
+diff --git a/ggml/src/ggml-cuda/fattn-mma-f16.cuh b/ggml/src/ggml-cuda/fattn-mma-f16.cuh
+index 7bd1044c1..3dea2205e 100644
+--- a/ggml/src/ggml-cuda/fattn-mma-f16.cuh
++++ b/ggml/src/ggml-cuda/fattn-mma-f16.cuh
+@@ -66,7 +66,8 @@ static constexpr __host__ __device__ fattn_mma_config ggml_cuda_fattn_mma_get_co
+     GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 32, 128, 2,  32, 128, 128, 128, 2, true);
+     GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 64, 128, 2,  32, 128, 128, 128, 2, true);
+ 
+-    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  8,  64, 4,  32, 288, 256, 128, 1, false);
++    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  4,  64, 4,  32, 288, 256, 128, 1, false);
++    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  8,  64, 4,  32, 288, 256, 128, 1, true);
+     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 16,  64, 4,  32, 288, 256, 128, 1, false);
+     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 32, 128, 2,  32, 160, 128, 128, 1, false);
+     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 64, 256, 1,  32, 160, 128, 128, 1, false);
+@@ -80,7 +81,8 @@ static constexpr __host__ __device__ fattn_mma_config ggml_cuda_fattn_mma_get_co
+     GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 32, 128, 2,  64, 128, 128,  64, 2, true);
+     GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 64, 128, 2,  64, 128, 128,  64, 2, true);
+ 
+-    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  8,  64, 4,  32,  96,  64, 128, 1, false);
++    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  4,  64, 4,  32,  96,  64, 128, 1, false);
++    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  8,  64, 4,  32,  96,  64, 128, 1, true);
+     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 16,  64, 4,  32,  96,  64, 128, 1, false);
+     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 32, 128, 2,  32, 160, 128, 128, 1, false);
+     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 64, 256, 1,  32, 160, 128, 128, 1, false);
+@@ -89,7 +91,8 @@ static constexpr __host__ __device__ fattn_mma_config ggml_cuda_fattn_mma_get_co
+ }
+ 
+ static constexpr __host__ __device__ fattn_mma_config ggml_cuda_fattn_mma_get_config_volta(const int DKQ, const int DV, const int ncols) {
+-    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  8,  64, 4,  32, 288, 256,  64, 1, false);
++    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  4,  64, 4,  32, 288, 256,  64, 1, false);
++    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  8,  64, 4,  32, 288, 256,  64, 1, true);
+     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 16,  64, 4,  32, 288, 256,  64, 1, false);
+     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 32, 128, 2,  32, 160, 128,  64, 1, false);
+     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 64, 256, 1,  32, 160, 128,  64, 1, false);
+@@ -397,7 +400,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
+     constexpr int  ncols           = ncols1 * ncols2;
+     constexpr int  cols_per_warp   = T_B_KQ::I;
+     constexpr int  cols_per_thread = 2; // This is specifically KQ columns, Volta only has a single VKQ column.
+-    constexpr int  np              = nwarps * (cols_per_warp/ncols2) / ncols1; // Number of parallel CUDA warps per Q column.
++    constexpr int  np              = cols_per_warp > ncols ? nwarps : nwarps * cols_per_warp/ncols; // Number of parallel CUDA warps per Q column.
+     constexpr int  nbatch_fa       = ggml_cuda_fattn_mma_get_nbatch_fa(DKQ, DV, ncols);
+     constexpr int  nbatch_K2       = ggml_cuda_fattn_mma_get_nbatch_K2(DKQ, DV, ncols);
+     constexpr int  nbatch_V2       = ggml_cuda_fattn_mma_get_nbatch_V2(DKQ, DV, ncols);
+@@ -467,7 +470,6 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
+                 }
+             }
+         } else {
+-            static_assert(cols_per_warp != 8, "cols_per_warp == 8 not implemented");
+ #pragma unroll
+             for (int k_KQ_0 = k0_start; k_KQ_0 < k0_stop; k_KQ_0 += T_A_KQ::J) {
+                 load_ldmatrix(Q_B[0], tile_Q + (threadIdx.y / np)*(T_B_KQ::I*stride_tile_Q) + k_KQ_0, stride_tile_Q);
+@@ -479,8 +481,18 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
+                     T_A_KQ K_A;
+                     load_ldmatrix(K_A, tile_K + i_KQ_0*stride_tile_K + (k_KQ_0 - k0_start), stride_tile_K);
+ 
+-                    // Wide version of KQ_C is column-major => swap A and B.
+-                    mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], Q_B[0], K_A);
++                    if constexpr (cols_per_warp == 8) {
++                        mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], K_A, Q_B[0]);
++                    } else {
++                        // Wide version of KQ_C is column-major
++#if defined(AMD_WMMA_AVAILABLE)
++                        // RDNA matrix C is column-major.
++                        mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], K_A, Q_B[0]);
++#else
++                        // swap A and B for CUDA.
++                        mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], Q_B[0], K_A);
++#endif // defined(AMD_WMMA_AVAILABLE)
++                    }
+                 }
+             }
+         }
+@@ -841,7 +853,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
+ 
+     constexpr int  cols_per_warp   = T_B_KQ::I;
+     constexpr int  cols_per_thread = 2; // This is specifically KQ columns, Volta only has a single VKQ column.
+-    constexpr int  np              = nwarps * (cols_per_warp/ncols2) / ncols1; // Number of parallel CUDA warps per Q column.
++    constexpr int  np              = cols_per_warp > ncols ? nwarps : nwarps * cols_per_warp/ncols; // Number of parallel CUDA warps per Q column.
+     constexpr int  nbatch_fa       = ggml_cuda_fattn_mma_get_nbatch_fa     (DKQ, DV, ncols);
+     constexpr int  nbatch_K2       = ggml_cuda_fattn_mma_get_nbatch_K2     (DKQ, DV, ncols);
+     constexpr int  nbatch_V2       = ggml_cuda_fattn_mma_get_nbatch_V2     (DKQ, DV, ncols);
+@@ -1353,6 +1365,13 @@ static __global__ void flash_attn_ext_f16(
+         NO_DEVICE_CODE;
+         return;
+     }
++#ifdef VOLTA_MMA_AVAILABLE
++    if (ncols1*ncols2 < 32) {
++        NO_DEVICE_CODE;
++        return;
++    }
++#endif // VOLTA_MMA_AVAILABLE
++
+ #if __CUDA_ARCH__ == GGML_CUDA_CC_TURING
+     if (ncols1*ncols2 > 32) {
+         NO_DEVICE_CODE;
+@@ -1585,3 +1604,8 @@ DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(256, 256,  64)
+ extern DECL_FATTN_MMA_F16_CASE(576, 512, 1, 16);
+ extern DECL_FATTN_MMA_F16_CASE(576, 512, 2, 16);
+ extern DECL_FATTN_MMA_F16_CASE(576, 512, 4, 16);
++
++// For GLM 4.7 Flash
++extern DECL_FATTN_MMA_F16_CASE(576, 512,  4,  4);
++extern DECL_FATTN_MMA_F16_CASE(576, 512,  8,  4);
++extern DECL_FATTN_MMA_F16_CASE(576, 512, 16,  4);
+diff --git a/ggml/src/ggml-cuda/fattn-tile.cuh b/ggml/src/ggml-cuda/fattn-tile.cuh
+index 7c4d6fe67..371be7442 100644
+--- a/ggml/src/ggml-cuda/fattn-tile.cuh
++++ b/ggml/src/ggml-cuda/fattn-tile.cuh
+@@ -68,6 +68,8 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_nv
+     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 2,  64,  64)
+     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 2,  64,  64)
+ 
++    GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  4, 128, 2,  64,  64)
++    GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  8, 256, 2,  64,  64)
+     GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 2,  64,  64)
+ 
+     return 0;
+@@ -122,6 +124,8 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_nv
+     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 2,  32, 128)
+     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 2,  32,  64)
+ 
++    GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  4, 128, 2,  32,  64)
++    GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  8, 256, 2,  32,  64)
+     GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 2,  32,  64)
+ 
+     return 0;
+@@ -183,6 +187,8 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_am
+     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 2,  32, 128)
+     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 2,  32, 128)
+ 
++    GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  4, 128, 2,  64,  64)
++    GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  8, 256, 2,  64,  64)
+     GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 2,  64,  64)
+     GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 32, 512, 1, 128,  64)
+ 
+@@ -245,6 +251,8 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_am
+     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 5,  32, 256)
+     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 3,  64, 128)
+ 
++    GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  4, 128, 2,  64,  64)
++    GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  8, 256, 2,  64,  64)
+     GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 4,  64,  64)
+     GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 32, 256, 2, 128,  64)
+ 
+@@ -1187,6 +1195,14 @@ static void launch_fattn_tile_switch_ncols2(ggml_backend_cuda_context & ctx, ggm
+             launch_fattn_tile_switch_ncols1<DKQ, DV, 16, use_logit_softcap>(ctx, dst);
+             return;
+         }
++        if (use_gqa_opt && gqa_ratio % 8 == 0) {
++            launch_fattn_tile_switch_ncols1<DKQ, DV, 8, use_logit_softcap>(ctx, dst);
++            return;
++        }
++        if (use_gqa_opt && gqa_ratio % 4 == 0) {
++            launch_fattn_tile_switch_ncols1<DKQ, DV, 4, use_logit_softcap>(ctx, dst);
++            return;
++        }
+     }
+ 
+     if constexpr (DV <= 256) {
+diff --git a/ggml/src/ggml-cuda/fattn.cu b/ggml/src/ggml-cuda/fattn.cu
+index 015540666..1693479cb 100644
+--- a/ggml/src/ggml-cuda/fattn.cu
++++ b/ggml/src/ggml-cuda/fattn.cu
+@@ -111,7 +111,7 @@ static void ggml_cuda_flash_attn_ext_mma_f16(ggml_backend_cuda_context & ctx, gg
+             ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2<256, 256>(ctx, dst);
+             break;
+         case 576: {
+-            // For Deepseek, go straight to the ncols1 switch to avoid compiling unnecessary kernels.
++            // For Deepseek/GLM4, go straight to the ncols1 switch to avoid compiling unnecessary kernels.
+             GGML_ASSERT(V->ne[0] == 512);
+             float max_bias = 0.0f;
+             memcpy(&max_bias, (const float *) KQV->op_params + 1, sizeof(float));
+@@ -121,8 +121,12 @@ static void ggml_cuda_flash_attn_ext_mma_f16(ggml_backend_cuda_context & ctx, gg
+ 
+             GGML_ASSERT(Q->ne[2] % K->ne[2] == 0);
+             const int gqa_ratio = Q->ne[2] / K->ne[2];
+-            GGML_ASSERT(gqa_ratio % 16 == 0);
+-            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512, 16>(ctx, dst);
++            GGML_ASSERT(gqa_ratio % 4 == 0);
++            if (gqa_ratio % 16 == 0) {
++                ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512, 16>(ctx, dst);
++            } else {
++                ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512,  4>(ctx, dst);
++            }
+         } break;
+         default:
+             GGML_ABORT("fatal error");
+@@ -251,7 +255,7 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
+             if (V->ne[0] != 512) {
+                 return BEST_FATTN_KERNEL_NONE;
+             }
+-            if (!gqa_opt_applies || gqa_ratio % 16 != 0) {
++            if (!gqa_opt_applies || gqa_ratio % 4 != 0) {
+                 return BEST_FATTN_KERNEL_NONE;
+             }
+             break;
+diff --git a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu
+index 2074e954a..517993cb0 100644
+--- a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu
++++ b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu
+@@ -8,3 +8,4 @@ DECL_FATTN_MMA_F16_CASE(96, 96, 16, 4);
+ DECL_FATTN_MMA_F16_CASE(112, 112, 16, 4);
+ DECL_FATTN_MMA_F16_CASE(128, 128, 16, 4);
+ DECL_FATTN_MMA_F16_CASE(256, 256, 16, 4);
++DECL_FATTN_MMA_F16_CASE(576, 512, 16, 4);
+diff --git a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu
+index 24c64cf00..97b19c67a 100644
+--- a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu
++++ b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu
+@@ -8,3 +8,4 @@ DECL_FATTN_MMA_F16_CASE(96, 96, 2, 4);
+ DECL_FATTN_MMA_F16_CASE(112, 112, 2, 4);
+ DECL_FATTN_MMA_F16_CASE(128, 128, 2, 4);
+ DECL_FATTN_MMA_F16_CASE(256, 256, 2, 4);
++DECL_FATTN_MMA_F16_CASE(576, 512, 2, 4);
+diff --git a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu
+index 1ada657f1..989626dfa 100644
+--- a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu
++++ b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu
+@@ -8,3 +8,4 @@ DECL_FATTN_MMA_F16_CASE(96, 96, 4, 4);
+ DECL_FATTN_MMA_F16_CASE(112, 112, 4, 4);
+ DECL_FATTN_MMA_F16_CASE(128, 128, 4, 4);
+ DECL_FATTN_MMA_F16_CASE(256, 256, 4, 4);
++DECL_FATTN_MMA_F16_CASE(576, 512, 4, 4);
+diff --git a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu
+index 86d4ffae2..173de7aac 100644
+--- a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu
++++ b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu
+@@ -8,3 +8,4 @@ DECL_FATTN_MMA_F16_CASE(96, 96, 8, 4);
+ DECL_FATTN_MMA_F16_CASE(112, 112, 8, 4);
+ DECL_FATTN_MMA_F16_CASE(128, 128, 8, 4);
+ DECL_FATTN_MMA_F16_CASE(256, 256, 8, 4);
++DECL_FATTN_MMA_F16_CASE(576, 512, 8, 4);
+diff --git a/ggml/src/ggml-metal/ggml-metal-device.m b/ggml/src/ggml-metal/ggml-metal-device.m
+index f24270bb1..7b5ee968c 100644
+--- a/ggml/src/ggml-metal/ggml-metal-device.m
++++ b/ggml/src/ggml-metal/ggml-metal-device.m
+@@ -1071,12 +1071,8 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
+                 op->src[0]->ne[0] != 112 &&
+                 op->src[0]->ne[0] != 128 &&
+                 op->src[0]->ne[0] != 192 &&
+-                op->src[0]->ne[0] != 256) {
+-                return false;
+-            }
+-            if (op->src[0]->ne[0] == 576) {
+-                // DeepSeek sizes
+-                // TODO: disabled for now, until optmized
++                op->src[0]->ne[0] != 256 &&
++                op->src[0]->ne[0] != 576) {
+                 return false;
+             }
+             if (op->src[1]->type != op->src[2]->type) {
+diff --git a/ggml/src/ggml-metal/ggml-metal-ops.cpp b/ggml/src/ggml-metal/ggml-metal-ops.cpp
+index e99c1763f..80864f303 100644
+--- a/ggml/src/ggml-metal/ggml-metal-ops.cpp
++++ b/ggml/src/ggml-metal/ggml-metal-ops.cpp
+@@ -2456,7 +2456,7 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
+ 
+         // simdgroups per threadgroup (a.k.a. warps)
+         //nsg = ne01 <= nqptg ? MAX(4, MIN(nsgmax, MIN(ne11/ncpsg, (int64_t) pipeline.maxTotalThreadsPerThreadgroup/32))) : 4;
+-        int32_t nsg = 4;
++        int32_t nsg = ne00 >= 512 ? 8 : 4;
+ 
+         const size_t smem = FATTN_SMEM(nsg);
+ 
+diff --git a/ggml/src/ggml-metal/ggml-metal.metal b/ggml/src/ggml-metal/ggml-metal.metal
+index c98d269d1..d33c16079 100644
+--- a/ggml/src/ggml-metal/ggml-metal.metal
++++ b/ggml/src/ggml-metal/ggml-metal.metal
+@@ -6166,6 +6166,7 @@ kernel void kernel_flash_attn_ext(
+       //case 1: kernel_flash_attn_ext_impl<FWD_TMPL, 1>(FWD_ARGS); break;
+       //case 2: kernel_flash_attn_ext_impl<FWD_TMPL, 2>(FWD_ARGS); break;
+         case 4: kernel_flash_attn_ext_impl<FWD_TMPL, 4>(FWD_ARGS); break;
++        case 8: kernel_flash_attn_ext_impl<FWD_TMPL, 8>(FWD_ARGS); break;
+     }
+ #undef FWD_TMPL
+ #undef FWD_ARGS

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 {

manifest/layer.go

@@ -1,4 +1,4 @@
-package server
+package manifest
 
 import (
 	"crypto/sha256"
@@ -14,7 +14,7 @@ type Layer struct {
 	Size      int64  `json:"size"`
 	From      string `json:"from,omitempty"`
 	Name      string `json:"name,omitempty"` // tensor name, e.g., "text_encoder/model.embed_tokens.weight"
-	status    string
+	Status    string `json:"-"`
 }
 
 const (
@@ -22,7 +22,7 @@ const (
 )
 
 func NewLayer(r io.Reader, mediatype string) (Layer, error) {
-	blobs, err := GetBlobsPath("")
+	blobs, err := BlobsPath("")
 	if err != nil {
 		return Layer{}, err
 	}
@@ -45,7 +45,7 @@ func NewLayer(r io.Reader, mediatype string) (Layer, error) {
 	}
 
 	digest := fmt.Sprintf("sha256:%x", sha256sum.Sum(nil))
-	blob, err := GetBlobsPath(digest)
+	blob, err := BlobsPath(digest)
 	if err != nil {
 		return Layer{}, err
 	}
@@ -65,7 +65,7 @@ func NewLayer(r io.Reader, mediatype string) (Layer, error) {
 		MediaType: mediatype,
 		Digest:    digest,
 		Size:      n,
-		status:    fmt.Sprintf("%s %s", status, digest),
+		Status:    fmt.Sprintf("%s %s", status, digest),
 	}, nil
 }
 
@@ -74,7 +74,7 @@ func NewLayerFromLayer(digest, mediatype, from string) (Layer, error) {
 		return Layer{}, errors.New("creating new layer from layer with empty digest")
 	}
 
-	blob, err := GetBlobsPath(digest)
+	blob, err := BlobsPath(digest)
 	if err != nil {
 		return Layer{}, err
 	}
@@ -89,7 +89,7 @@ func NewLayerFromLayer(digest, mediatype, from string) (Layer, error) {
 		Digest:    digest,
 		Size:      fi.Size(),
 		From:      from,
-		status:    fmt.Sprintf("using existing layer %s", digest),
+		Status:    fmt.Sprintf("using existing layer %s", digest),
 	}, nil
 }
 
@@ -98,7 +98,7 @@ func (l *Layer) Open() (io.ReadSeekCloser, error) {
 		return nil, errors.New("opening layer with empty digest")
 	}
 
-	blob, err := GetBlobsPath(l.Digest)
+	blob, err := BlobsPath(l.Digest)
 	if err != nil {
 		return nil, err
 	}
@@ -126,7 +126,7 @@ func (l *Layer) Remove() error {
 		}
 	}
 
-	blob, err := GetBlobsPath(l.Digest)
+	blob, err := BlobsPath(l.Digest)
 	if err != nil {
 		return err
 	}

manifest/manifest.go

@@ -1,10 +1,9 @@
-package server
+package manifest
 
 import (
 	"crypto/sha256"
 	"encoding/hex"
 	"encoding/json"
-	"errors"
 	"fmt"
 	"io"
 	"log/slog"
@@ -33,12 +32,38 @@ func (m *Manifest) Size() (size int64) {
 	return
 }
 
+func (m *Manifest) Digest() string {
+	return m.digest
+}
+
+func (m *Manifest) FileInfo() os.FileInfo {
+	return m.fi
+}
+
+// ReadConfigJSON reads and unmarshals a config layer as JSON.
+func (m *Manifest) ReadConfigJSON(configPath string, v any) error {
+	for _, layer := range m.Layers {
+		if layer.MediaType == "application/vnd.ollama.image.json" && layer.Name == configPath {
+			blobPath, err := BlobsPath(layer.Digest)
+			if err != nil {
+				return err
+			}
+			data, err := os.ReadFile(blobPath)
+			if err != nil {
+				return err
+			}
+			return json.Unmarshal(data, v)
+		}
+	}
+	return fmt.Errorf("config %q not found in manifest", configPath)
+}
+
 func (m *Manifest) Remove() error {
 	if err := os.Remove(m.filepath); err != nil {
 		return err
 	}
 
-	manifests, err := GetManifestPath()
+	manifests, err := Path()
 	if err != nil {
 		return err
 	}
@@ -47,16 +72,40 @@ func (m *Manifest) Remove() error {
 }
 
 func (m *Manifest) RemoveLayers() error {
-	for _, layer := range append(m.Layers, m.Config) {
-		if layer.Digest != "" {
-			if err := layer.Remove(); errors.Is(err, os.ErrNotExist) {
-				slog.Debug("layer does not exist", "digest", layer.Digest)
-			} else if err != nil {
-				return err
+	ms, err := Manifests(true)
+	if err != nil {
+		return err
+	}
+
+	// Build set of digests still in use by other manifests
+	inUse := make(map[string]struct{})
+	for _, other := range ms {
+		for _, layer := range append(other.Layers, other.Config) {
+			if layer.Digest != "" {
+				inUse[layer.Digest] = struct{}{}
 			}
 		}
 	}
 
+	// Remove layers not used by any other manifest
+	for _, layer := range append(m.Layers, m.Config) {
+		if layer.Digest == "" {
+			continue
+		}
+		if _, used := inUse[layer.Digest]; used {
+			continue
+		}
+		blob, err := BlobsPath(layer.Digest)
+		if err != nil {
+			return err
+		}
+		if err := os.Remove(blob); os.IsNotExist(err) {
+			slog.Debug("layer does not exist", "digest", layer.Digest)
+		} else if err != nil {
+			return err
+		}
+	}
+
 	return nil
 }
 
@@ -65,7 +114,7 @@ func ParseNamedManifest(n model.Name) (*Manifest, error) {
 		return nil, model.Unqualified(n)
 	}
 
-	manifests, err := GetManifestPath()
+	manifests, err := Path()
 	if err != nil {
 		return nil, err
 	}
@@ -97,7 +146,7 @@ func ParseNamedManifest(n model.Name) (*Manifest, error) {
 }
 
 func WriteManifest(name model.Name, config Layer, layers []Layer) error {
-	manifests, err := GetManifestPath()
+	manifests, err := Path()
 	if err != nil {
 		return err
 	}
@@ -124,7 +173,7 @@ func WriteManifest(name model.Name, config Layer, layers []Layer) error {
 }
 
 func Manifests(continueOnError bool) (map[model.Name]*Manifest, error) {
-	manifests, err := GetManifestPath()
+	manifests, err := Path()
 	if err != nil {
 		return nil, err
 	}

manifest/manifest_test.go

@@ -1,4 +1,4 @@
-package server
+package manifest
 
 import (
 	"encoding/json"

manifest/paths.go

@@ -0,0 +1,95 @@
+package manifest
+
+import (
+	"errors"
+	"fmt"
+	"os"
+	"path/filepath"
+	"regexp"
+	"strings"
+
+	"github.com/ollama/ollama/envconfig"
+	"github.com/ollama/ollama/types/model"
+)
+
+var ErrInvalidDigestFormat = errors.New("invalid digest format")
+
+func Path() (string, error) {
+	path := filepath.Join(envconfig.Models(), "manifests")
+	if err := os.MkdirAll(path, 0o755); err != nil {
+		return "", fmt.Errorf("%w: ensure path elements are traversable", err)
+	}
+
+	return path, nil
+}
+
+// PathForName returns the path to the manifest file for a specific model name.
+func PathForName(n model.Name) (string, error) {
+	if !n.IsValid() {
+		return "", os.ErrNotExist
+	}
+
+	manifests, err := Path()
+	if err != nil {
+		return "", err
+	}
+
+	return filepath.Join(manifests, n.Filepath()), nil
+}
+
+func BlobsPath(digest string) (string, error) {
+	// only accept actual sha256 digests
+	pattern := "^sha256[:-][0-9a-fA-F]{64}$"
+	re := regexp.MustCompile(pattern)
+
+	if digest != "" && !re.MatchString(digest) {
+		return "", ErrInvalidDigestFormat
+	}
+
+	digest = strings.ReplaceAll(digest, ":", "-")
+	path := filepath.Join(envconfig.Models(), "blobs", digest)
+	dirPath := filepath.Dir(path)
+	if digest == "" {
+		dirPath = path
+	}
+
+	if err := os.MkdirAll(dirPath, 0o755); err != nil {
+		return "", fmt.Errorf("%w: ensure path elements are traversable", err)
+	}
+
+	return path, nil
+}
+
+// PruneDirectory removes empty directories recursively.
+func PruneDirectory(path string) error {
+	info, err := os.Lstat(path)
+	if err != nil {
+		return err
+	}
+
+	if info.IsDir() && info.Mode()&os.ModeSymlink == 0 {
+		entries, err := os.ReadDir(path)
+		if err != nil {
+			return err
+		}
+
+		for _, entry := range entries {
+			if err := PruneDirectory(filepath.Join(path, entry.Name())); err != nil {
+				return err
+			}
+		}
+
+		entries, err = os.ReadDir(path)
+		if err != nil {
+			return err
+		}
+
+		if len(entries) > 0 {
+			return nil
+		}
+
+		return os.Remove(path)
+	}
+
+	return nil
+}

middleware/anthropic.go

@@ -118,6 +118,9 @@ func AnthropicMessagesMiddleware() gin.HandlerFunc {
 			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()))

middleware/anthropic_test.go

@@ -582,3 +582,26 @@ func TestAnthropicWriter_ErrorFromRoutes(t *testing.T) {
 		})
 	}
 }
+
+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 +546,112 @@ 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()
+	}
+}
+
+func ImageEditsMiddleware() gin.HandlerFunc {
+	return func(c *gin.Context) {
+		var req openai.ImageEditRequest
+		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
+		}
+
+		if req.Image == "" {
+			c.AbortWithStatusJSON(http.StatusBadRequest, openai.NewError(http.StatusBadRequest, "image is required"))
+			return
+		}
+
+		genReq, err := openai.FromImageEditRequest(req)
+		if err != nil {
+			c.AbortWithStatusJSON(http.StatusBadRequest, openai.NewError(http.StatusBadRequest, err.Error()))
+			return
+		}
+
+		var b bytes.Buffer
+		if err := json.NewEncoder(&b).Encode(genReq); 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 +961,280 @@ 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)
+	}
+}
+
+func TestImageEditsMiddleware(t *testing.T) {
+	type testCase struct {
+		name string
+		body string
+		req  api.GenerateRequest
+		err  openai.ErrorResponse
+	}
+
+	var capturedRequest *api.GenerateRequest
+
+	// Base64-encoded test image (1x1 pixel PNG)
+	testImage := "data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAQAAAC1HAwCAAAAC0lEQVR42mNk+A8AAQUBAScY42YAAAAASUVORK5CYII="
+	decodedImage, _ := base64.StdEncoding.DecodeString("iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAQAAAC1HAwCAAAAC0lEQVR42mNk+A8AAQUBAScY42YAAAAASUVORK5CYII=")
+
+	testCases := []testCase{
+		{
+			name: "image edit basic",
+			body: `{
+				"model": "test-model",
+				"prompt": "make it blue",
+				"image": "` + testImage + `"
+			}`,
+			req: api.GenerateRequest{
+				Model:  "test-model",
+				Prompt: "make it blue",
+				Images: []api.ImageData{decodedImage},
+			},
+		},
+		{
+			name: "image edit with size",
+			body: `{
+				"model": "test-model",
+				"prompt": "make it blue",
+				"image": "` + testImage + `",
+				"size": "512x768"
+			}`,
+			req: api.GenerateRequest{
+				Model:  "test-model",
+				Prompt: "make it blue",
+				Images: []api.ImageData{decodedImage},
+				Width:  512,
+				Height: 768,
+			},
+		},
+		{
+			name: "image edit missing prompt",
+			body: `{
+				"model": "test-model",
+				"image": "` + testImage + `"
+			}`,
+			err: openai.ErrorResponse{
+				Error: openai.Error{
+					Message: "prompt is required",
+					Type:    "invalid_request_error",
+				},
+			},
+		},
+		{
+			name: "image edit missing model",
+			body: `{
+				"prompt": "make it blue",
+				"image": "` + testImage + `"
+			}`,
+			err: openai.ErrorResponse{
+				Error: openai.Error{
+					Message: "model is required",
+					Type:    "invalid_request_error",
+				},
+			},
+		},
+		{
+			name: "image edit missing image",
+			body: `{
+				"model": "test-model",
+				"prompt": "make it blue"
+			}`,
+			err: openai.ErrorResponse{
+				Error: openai.Error{
+					Message: "image is required",
+					Type:    "invalid_request_error",
+				},
+			},
+		},
+	}
+
+	endpoint := func(c *gin.Context) {
+		c.Status(http.StatusOK)
+	}
+
+	gin.SetMode(gin.TestMode)
+	router := gin.New()
+	router.Use(ImageEditsMiddleware(), 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)
+			}
+		})
+	}
+}

ml/backend.go

@@ -162,6 +162,7 @@ type Tensor interface {
 	AvgPool2D(ctx Context, k, s int, p float32) Tensor
 	Conv2D(ctx Context, weight Tensor, s0, s1, p0, p1, d0, d1 int) Tensor
 	Conv3D(ctx Context, weight Tensor, c, s0, s1, s2, p0, p1, p2, d0, d1, d2 int) Tensor
+	SSMConv(ctx Context, kernel Tensor) Tensor
 
 	IM2Col(ctx Context, weight Tensor, s0, s1, p0, p1, d0, d1 int) Tensor
 

ml/backend/ggml/ggml.go

@@ -1641,6 +1641,13 @@ func (t *Tensor) Conv3D(ctx ml.Context, t2 ml.Tensor, c, s0, s1, s2, p0, p1, p2,
 	return tt
 }
 
+func (t *Tensor) SSMConv(ctx ml.Context, kernel ml.Tensor) ml.Tensor {
+	return &Tensor{
+		b: t.b,
+		t: C.ggml_ssm_conv(ctx.(*Context).ctx, t.t, kernel.(*Tensor).t),
+	}
+}
+
 func (t *Tensor) AvgPool2D(ctx ml.Context, k, s int, p float32) ml.Tensor {
 	return &Tensor{
 		b: t.b,

ml/backend/ggml/ggml/src/ggml-cuda/fattn-mma-f16.cuh

@@ -66,7 +66,8 @@ static constexpr __host__ __device__ fattn_mma_config ggml_cuda_fattn_mma_get_co
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 32, 128, 2,  32, 128, 128, 128, 2, true);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 64, 128, 2,  32, 128, 128, 128, 2, true);
 
-    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  8,  64, 4,  32, 288, 256, 128, 1, false);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  4,  64, 4,  32, 288, 256, 128, 1, false);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  8,  64, 4,  32, 288, 256, 128, 1, true);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 16,  64, 4,  32, 288, 256, 128, 1, false);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 32, 128, 2,  32, 160, 128, 128, 1, false);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 64, 256, 1,  32, 160, 128, 128, 1, false);
@@ -80,7 +81,8 @@ static constexpr __host__ __device__ fattn_mma_config ggml_cuda_fattn_mma_get_co
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 32, 128, 2,  64, 128, 128,  64, 2, true);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 64, 128, 2,  64, 128, 128,  64, 2, true);
 
-    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  8,  64, 4,  32,  96,  64, 128, 1, false);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  4,  64, 4,  32,  96,  64, 128, 1, false);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  8,  64, 4,  32,  96,  64, 128, 1, true);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 16,  64, 4,  32,  96,  64, 128, 1, false);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 32, 128, 2,  32, 160, 128, 128, 1, false);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 64, 256, 1,  32, 160, 128, 128, 1, false);
@@ -89,7 +91,8 @@ static constexpr __host__ __device__ fattn_mma_config ggml_cuda_fattn_mma_get_co
 }
 
 static constexpr __host__ __device__ fattn_mma_config ggml_cuda_fattn_mma_get_config_volta(const int DKQ, const int DV, const int ncols) {
-    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  8,  64, 4,  32, 288, 256,  64, 1, false);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  4,  64, 4,  32, 288, 256,  64, 1, false);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  8,  64, 4,  32, 288, 256,  64, 1, true);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 16,  64, 4,  32, 288, 256,  64, 1, false);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 32, 128, 2,  32, 160, 128,  64, 1, false);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 64, 256, 1,  32, 160, 128,  64, 1, false);
@@ -397,7 +400,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
     constexpr int  ncols           = ncols1 * ncols2;
     constexpr int  cols_per_warp   = T_B_KQ::I;
     constexpr int  cols_per_thread = 2; // This is specifically KQ columns, Volta only has a single VKQ column.
-    constexpr int  np              = nwarps * (cols_per_warp/ncols2) / ncols1; // Number of parallel CUDA warps per Q column.
+    constexpr int  np              = cols_per_warp > ncols ? nwarps : nwarps * cols_per_warp/ncols; // Number of parallel CUDA warps per Q column.
     constexpr int  nbatch_fa       = ggml_cuda_fattn_mma_get_nbatch_fa(DKQ, DV, ncols);
     constexpr int  nbatch_K2       = ggml_cuda_fattn_mma_get_nbatch_K2(DKQ, DV, ncols);
     constexpr int  nbatch_V2       = ggml_cuda_fattn_mma_get_nbatch_V2(DKQ, DV, ncols);
@@ -467,7 +470,6 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
                 }
             }
         } else {
-            static_assert(cols_per_warp != 8, "cols_per_warp == 8 not implemented");
 #pragma unroll
             for (int k_KQ_0 = k0_start; k_KQ_0 < k0_stop; k_KQ_0 += T_A_KQ::J) {
                 load_ldmatrix(Q_B[0], tile_Q + (threadIdx.y / np)*(T_B_KQ::I*stride_tile_Q) + k_KQ_0, stride_tile_Q);
@@ -479,8 +481,18 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
                     T_A_KQ K_A;
                     load_ldmatrix(K_A, tile_K + i_KQ_0*stride_tile_K + (k_KQ_0 - k0_start), stride_tile_K);
 
-                    // Wide version of KQ_C is column-major => swap A and B.
-                    mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], Q_B[0], K_A);
+                    if constexpr (cols_per_warp == 8) {
+                        mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], K_A, Q_B[0]);
+                    } else {
+                        // Wide version of KQ_C is column-major
+#if defined(AMD_WMMA_AVAILABLE)
+                        // RDNA matrix C is column-major.
+                        mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], K_A, Q_B[0]);
+#else
+                        // swap A and B for CUDA.
+                        mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], Q_B[0], K_A);
+#endif // defined(AMD_WMMA_AVAILABLE)
+                    }
                 }
             }
         }
@@ -841,7 +853,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
 
     constexpr int  cols_per_warp   = T_B_KQ::I;
     constexpr int  cols_per_thread = 2; // This is specifically KQ columns, Volta only has a single VKQ column.
-    constexpr int  np              = nwarps * (cols_per_warp/ncols2) / ncols1; // Number of parallel CUDA warps per Q column.
+    constexpr int  np              = cols_per_warp > ncols ? nwarps : nwarps * cols_per_warp/ncols; // Number of parallel CUDA warps per Q column.
     constexpr int  nbatch_fa       = ggml_cuda_fattn_mma_get_nbatch_fa     (DKQ, DV, ncols);
     constexpr int  nbatch_K2       = ggml_cuda_fattn_mma_get_nbatch_K2     (DKQ, DV, ncols);
     constexpr int  nbatch_V2       = ggml_cuda_fattn_mma_get_nbatch_V2     (DKQ, DV, ncols);
@@ -1353,6 +1365,13 @@ static __global__ void flash_attn_ext_f16(
         NO_DEVICE_CODE;
         return;
     }
+#ifdef VOLTA_MMA_AVAILABLE
+    if (ncols1*ncols2 < 32) {
+        NO_DEVICE_CODE;
+        return;
+    }
+#endif // VOLTA_MMA_AVAILABLE
+
 #if __CUDA_ARCH__ == GGML_CUDA_CC_TURING
     if (ncols1*ncols2 > 32) {
         NO_DEVICE_CODE;
@@ -1585,3 +1604,8 @@ DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(256, 256,  64)
 extern DECL_FATTN_MMA_F16_CASE(576, 512, 1, 16);
 extern DECL_FATTN_MMA_F16_CASE(576, 512, 2, 16);
 extern DECL_FATTN_MMA_F16_CASE(576, 512, 4, 16);
+
+// For GLM 4.7 Flash
+extern DECL_FATTN_MMA_F16_CASE(576, 512,  4,  4);
+extern DECL_FATTN_MMA_F16_CASE(576, 512,  8,  4);
+extern DECL_FATTN_MMA_F16_CASE(576, 512, 16,  4);

ml/backend/ggml/ggml/src/ggml-cuda/fattn-tile.cuh

@@ -68,6 +68,8 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_nv
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 2,  64,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 2,  64,  64)
 
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  4, 128, 2,  64,  64)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  8, 256, 2,  64,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 2,  64,  64)
 
     return 0;
@@ -122,6 +124,8 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_nv
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 2,  32, 128)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 2,  32,  64)
 
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  4, 128, 2,  32,  64)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  8, 256, 2,  32,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 2,  32,  64)
 
     return 0;
@@ -183,6 +187,8 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_am
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 2,  32, 128)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 2,  32, 128)
 
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  4, 128, 2,  64,  64)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  8, 256, 2,  64,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 2,  64,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 32, 512, 1, 128,  64)
 
@@ -245,6 +251,8 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_am
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 5,  32, 256)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 3,  64, 128)
 
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  4, 128, 2,  64,  64)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  8, 256, 2,  64,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 4,  64,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 32, 256, 2, 128,  64)
 
@@ -1187,6 +1195,14 @@ static void launch_fattn_tile_switch_ncols2(ggml_backend_cuda_context & ctx, ggm
             launch_fattn_tile_switch_ncols1<DKQ, DV, 16, use_logit_softcap>(ctx, dst);
             return;
         }
+        if (use_gqa_opt && gqa_ratio % 8 == 0) {
+            launch_fattn_tile_switch_ncols1<DKQ, DV, 8, use_logit_softcap>(ctx, dst);
+            return;
+        }
+        if (use_gqa_opt && gqa_ratio % 4 == 0) {
+            launch_fattn_tile_switch_ncols1<DKQ, DV, 4, use_logit_softcap>(ctx, dst);
+            return;
+        }
     }
 
     if constexpr (DV <= 256) {

ml/backend/ggml/ggml/src/ggml-cuda/fattn.cu

@@ -111,7 +111,7 @@ static void ggml_cuda_flash_attn_ext_mma_f16(ggml_backend_cuda_context & ctx, gg
             ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2<256, 256>(ctx, dst);
             break;
         case 576: {
-            // For Deepseek, go straight to the ncols1 switch to avoid compiling unnecessary kernels.
+            // For Deepseek/GLM4, go straight to the ncols1 switch to avoid compiling unnecessary kernels.
             GGML_ASSERT(V->ne[0] == 512);
             float max_bias = 0.0f;
             memcpy(&max_bias, (const float *) KQV->op_params + 1, sizeof(float));
@@ -121,8 +121,12 @@ static void ggml_cuda_flash_attn_ext_mma_f16(ggml_backend_cuda_context & ctx, gg
 
             GGML_ASSERT(Q->ne[2] % K->ne[2] == 0);
             const int gqa_ratio = Q->ne[2] / K->ne[2];
-            GGML_ASSERT(gqa_ratio % 16 == 0);
-            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512, 16>(ctx, dst);
+            GGML_ASSERT(gqa_ratio % 4 == 0);
+            if (gqa_ratio % 16 == 0) {
+                ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512, 16>(ctx, dst);
+            } else {
+                ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512,  4>(ctx, dst);
+            }
         } break;
         default:
             GGML_ABORT("fatal error");
@@ -251,7 +255,7 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
             if (V->ne[0] != 512) {
                 return BEST_FATTN_KERNEL_NONE;
             }
-            if (!gqa_opt_applies || gqa_ratio % 16 != 0) {
+            if (!gqa_opt_applies || gqa_ratio % 4 != 0) {
                 return BEST_FATTN_KERNEL_NONE;
             }
             break;

ml/backend/ggml/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu

@@ -8,3 +8,4 @@ DECL_FATTN_MMA_F16_CASE(96, 96, 16, 4);
 DECL_FATTN_MMA_F16_CASE(112, 112, 16, 4);
 DECL_FATTN_MMA_F16_CASE(128, 128, 16, 4);
 DECL_FATTN_MMA_F16_CASE(256, 256, 16, 4);
+DECL_FATTN_MMA_F16_CASE(576, 512, 16, 4);

ml/backend/ggml/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu

@@ -8,3 +8,4 @@ DECL_FATTN_MMA_F16_CASE(96, 96, 2, 4);
 DECL_FATTN_MMA_F16_CASE(112, 112, 2, 4);
 DECL_FATTN_MMA_F16_CASE(128, 128, 2, 4);
 DECL_FATTN_MMA_F16_CASE(256, 256, 2, 4);
+DECL_FATTN_MMA_F16_CASE(576, 512, 2, 4);

ml/backend/ggml/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu

@@ -8,3 +8,4 @@ DECL_FATTN_MMA_F16_CASE(96, 96, 4, 4);
 DECL_FATTN_MMA_F16_CASE(112, 112, 4, 4);
 DECL_FATTN_MMA_F16_CASE(128, 128, 4, 4);
 DECL_FATTN_MMA_F16_CASE(256, 256, 4, 4);
+DECL_FATTN_MMA_F16_CASE(576, 512, 4, 4);

ml/backend/ggml/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu

@@ -8,3 +8,4 @@ DECL_FATTN_MMA_F16_CASE(96, 96, 8, 4);
 DECL_FATTN_MMA_F16_CASE(112, 112, 8, 4);
 DECL_FATTN_MMA_F16_CASE(128, 128, 8, 4);
 DECL_FATTN_MMA_F16_CASE(256, 256, 8, 4);
+DECL_FATTN_MMA_F16_CASE(576, 512, 8, 4);

ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-device.m

@@ -1071,12 +1071,8 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
                 op->src[0]->ne[0] != 112 &&
                 op->src[0]->ne[0] != 128 &&
                 op->src[0]->ne[0] != 192 &&
-                op->src[0]->ne[0] != 256) {
-                return false;
-            }
-            if (op->src[0]->ne[0] == 576) {
-                // DeepSeek sizes
-                // TODO: disabled for now, until optmized
+                op->src[0]->ne[0] != 256 &&
+                op->src[0]->ne[0] != 576) {
                 return false;
             }
             if (op->src[1]->type != op->src[2]->type) {

ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-embed.metal

@@ -8967,6 +8967,7 @@ kernel void kernel_flash_attn_ext(
       //case 1: kernel_flash_attn_ext_impl<FWD_TMPL, 1>(FWD_ARGS); break;
       //case 2: kernel_flash_attn_ext_impl<FWD_TMPL, 2>(FWD_ARGS); break;
         case 4: kernel_flash_attn_ext_impl<FWD_TMPL, 4>(FWD_ARGS); break;
+        case 8: kernel_flash_attn_ext_impl<FWD_TMPL, 8>(FWD_ARGS); break;
     }
 #undef FWD_TMPL
 #undef FWD_ARGS

ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-ops.cpp

@@ -2456,7 +2456,7 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
 
         // simdgroups per threadgroup (a.k.a. warps)
         //nsg = ne01 <= nqptg ? MAX(4, MIN(nsgmax, MIN(ne11/ncpsg, (int64_t) pipeline.maxTotalThreadsPerThreadgroup/32))) : 4;
-        int32_t nsg = 4;
+        int32_t nsg = ne00 >= 512 ? 8 : 4;
 
         const size_t smem = FATTN_SMEM(nsg);
 

ml/backend/ggml/ggml/src/ggml-metal/ggml-metal.metal

@@ -6166,6 +6166,7 @@ kernel void kernel_flash_attn_ext(
       //case 1: kernel_flash_attn_ext_impl<FWD_TMPL, 1>(FWD_ARGS); break;
       //case 2: kernel_flash_attn_ext_impl<FWD_TMPL, 2>(FWD_ARGS); break;
         case 4: kernel_flash_attn_ext_impl<FWD_TMPL, 4>(FWD_ARGS); break;
+        case 8: kernel_flash_attn_ext_impl<FWD_TMPL, 8>(FWD_ARGS); break;
     }
 #undef FWD_TMPL
 #undef FWD_ARGS

model/model.go

@@ -39,6 +39,13 @@ type Model interface {
 	Config() config
 }
 
+// Validator is an optional interface that models can implement to perform
+// validation after tensors have been loaded. If validation fails, model
+// loading will fail with the returned error.
+type Validator interface {
+	Validate() error
+}
+
 // MultimodalProcessor must be implemented by multimodal models.
 type MultimodalProcessor interface {
 	// EncodeMultimodal processes a single input (such as an image) and
@@ -116,6 +123,13 @@ func New(modelPath string, params ml.BackendParams) (Model, error) {
 	base := Base{b: b, config: m.Config()}
 	v := reflect.ValueOf(m)
 	v.Elem().Set(populateFields(base, v.Elem()))
+
+	if validator, ok := m.(Validator); ok {
+		if err := validator.Validate(); err != nil {
+			return nil, err
+		}
+	}
+
 	return m, nil
 }
 

model/models/glm4moelite/model.go

@@ -0,0 +1,323 @@
+package glm4moelite
+
+import (
+	"errors"
+	"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"
+)
+
+var ErrOldModelFormat = errors.New("this model uses a weight format that is no longer supported; please re-download it")
+
+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"`
+
+	KB *nn.Linear `gguf:"attn_k_b"`
+	VB *nn.Linear `gguf:"attn_v_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)
+
+	// MLA absorption: absorb K projection into query
+	qPass := queryChunks[0].Permute(ctx, 0, 2, 1, 3)
+	qPassAbsorb := attn.KB.Forward(ctx, qPass).Permute(ctx, 0, 2, 1, 3)
+	query = qRot.Concat(ctx, qPassAbsorb, 0)
+
+	kPass = kPass.Reshape(ctx, opts.kvLoraRank, 1, seqLength)
+	key := kRot.Concat(ctx, kPass, 0)
+
+	attention := nn.AttentionWithVMLA(ctx, query, key, kPass, nil, attn.VB.Weight, 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"))
+	kvLoraRank := int(c.Uint("attention.kv_lora_rank"))
+	qkRopeHeadDim := int(c.Uint("rope.dimension_count"))
+
+	// For MLA absorption, the effective key dimension is kvLoraRank + qkRopeHeadDim
+	mlaKeyLength := kvLoraRank + qkRopeHeadDim
+	kqScale := 1.0 / math.Sqrt(float64(mlaKeyLength))
+
+	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) Validate() error {
+	for _, layer := range m.Layers {
+		if layer.Attention != nil && (layer.Attention.KB == nil || layer.Attention.VB == nil) {
+			return ErrOldModelFormat
+		}
+	}
+	return 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/glm4moelite/model_test.go

@@ -0,0 +1,73 @@
+package glm4moelite
+
+import (
+	"testing"
+
+	"github.com/ollama/ollama/ml/nn"
+)
+
+func TestValidate(t *testing.T) {
+	tests := []struct {
+		name    string
+		model   *Model
+		wantErr bool
+	}{
+		{
+			name: "valid model with KB and VB",
+			model: &Model{
+				Layers: []Layer{
+					{Attention: &Attention{KB: &nn.Linear{}, VB: &nn.Linear{}}},
+				},
+			},
+			wantErr: false,
+		},
+		{
+			name: "missing KB",
+			model: &Model{
+				Layers: []Layer{
+					{Attention: &Attention{VB: &nn.Linear{}}},
+				},
+			},
+			wantErr: true,
+		},
+		{
+			name: "missing VB",
+			model: &Model{
+				Layers: []Layer{
+					{Attention: &Attention{KB: &nn.Linear{}}},
+				},
+			},
+			wantErr: true,
+		},
+		{
+			name: "missing both KB and VB",
+			model: &Model{
+				Layers: []Layer{
+					{Attention: &Attention{}},
+				},
+			},
+			wantErr: true,
+		},
+		{
+			name: "nil Attention is ok",
+			model: &Model{
+				Layers: []Layer{
+					{Attention: nil},
+				},
+			},
+			wantErr: false,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			err := tt.model.Validate()
+			if (err != nil) != tt.wantErr {
+				t.Errorf("Validate() error = %v, wantErr %v", err, tt.wantErr)
+			}
+			if tt.wantErr && err != ErrOldModelFormat {
+				t.Errorf("Validate() error = %v, want %v", err, ErrOldModelFormat)
+			}
+		})
+	}
+}

model/models/lfm2/cache.go

@@ -0,0 +1,410 @@
+package lfm2
+
+import (
+	"slices"
+
+	"github.com/ollama/ollama/kvcache"
+	"github.com/ollama/ollama/ml"
+	"github.com/ollama/ollama/model/input"
+)
+
+var _ kvcache.Cache = (*HybridCache)(nil)
+
+// HybridCache stores:
+// - a standard causal KV cache for attention layers
+// - a per-sequence recurrent conv state for shortconv layers
+//
+// Conv state shape (per layer, per sequence): [dConv, hiddenSize] where dConv = L_cache - 1.
+// Stored internally as a tensor of shape [dConv * hiddenSize, maxSlots].
+type HybridCache struct {
+	kv *kvcache.Causal
+
+	backend      ml.Backend
+	dtype        ml.DType
+	maxSequences int
+
+	hiddenSize int
+	dConv      int
+
+	// slot mapping for recurrent state
+	slotForSeq map[int]int
+	refCount   []int
+	freeSlots  []int
+
+	// per-layer conv state buffers (allocated lazily)
+	convCtxs   map[int]ml.Context
+	convStates map[int]ml.Tensor // [dConv*hiddenSize, maxSlots]
+
+	// current forward batch (derived in StartForward)
+	curSeqs       []int
+	curSlots      []int
+	curSlotsInput ml.Tensor
+	curSeqTokens  int
+
+	// track if EnsureWritable has been called for this forward pass
+	writableEnsured bool
+	// track any error from EnsureWritable to propagate later
+	writableError error
+}
+
+func NewHybridCache(shift func(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error), hiddenSize, dConv int) *HybridCache {
+	return &HybridCache{
+		kv:         kvcache.NewCausalCache(shift),
+		hiddenSize: hiddenSize,
+		dConv:      dConv,
+		slotForSeq: make(map[int]int),
+		convCtxs:   make(map[int]ml.Context),
+		convStates: make(map[int]ml.Tensor),
+	}
+}
+
+func (c *HybridCache) Init(backend ml.Backend, dtype ml.DType, maxSequences, capacity, maxBatch int) {
+	c.backend = backend
+	c.dtype = dtype
+	c.maxSequences = maxSequences
+
+	// initialize slot allocator
+	c.refCount = make([]int, maxSequences)
+	c.freeSlots = c.freeSlots[:0]
+	for i := maxSequences - 1; i >= 0; i-- {
+		c.freeSlots = append(c.freeSlots, i)
+	}
+
+	c.kv.Init(backend, dtype, maxSequences, capacity, maxBatch)
+}
+
+func (c *HybridCache) Close() {
+	for _, ctx := range c.convCtxs {
+		ctx.Close()
+	}
+	c.kv.Close()
+}
+
+func (c *HybridCache) SetConfig(config ml.CacheConfig) {
+	c.kv.SetConfig(config)
+}
+
+func (c *HybridCache) SetLayer(layer int) {
+	c.kv.SetLayer(layer)
+}
+
+func (c *HybridCache) Get(ctx ml.Context) (ml.Tensor, ml.Tensor, ml.Tensor) {
+	return c.kv.Get(ctx)
+}
+
+func (c *HybridCache) Put(ctx ml.Context, key, value ml.Tensor) {
+	c.kv.Put(ctx, key, value)
+}
+
+func (c *HybridCache) StartForward(ctx ml.Context, batch input.Batch, reserve bool) error {
+	if err := c.kv.StartForward(ctx, batch, reserve); err != nil {
+		return err
+	}
+
+	// Derive equal-length sequence layout for shortconv.
+	// LFM2 shortconv assumes tokens form a [seq_tokens, seqs] grid.
+	seqCounts := make(map[int]int)
+	c.curSeqs = c.curSeqs[:0]
+	for _, s := range batch.Sequences {
+		if _, ok := seqCounts[s]; !ok {
+			c.curSeqs = append(c.curSeqs, s)
+		}
+		seqCounts[s]++
+	}
+
+	if len(c.curSeqs) == 0 {
+		return nil
+	}
+
+	nTokens := len(batch.Sequences)
+	nSeqs := len(c.curSeqs)
+	want := nTokens / nSeqs
+	for _, s := range c.curSeqs {
+		if seqCounts[s] != want {
+			return kvcache.ErrNotSupported
+		}
+	}
+
+	c.curSeqTokens = want
+
+	// When reserving memory for estimation, use fake slot assignments
+	// without modifying permanent state (slotForSeq, refCount)
+	if reserve {
+		c.curSlots = c.curSlots[:0]
+		slots := make([]int32, nSeqs)
+		for i := range nSeqs {
+			c.curSlots = append(c.curSlots, i)
+			slots[i] = int32(i)
+		}
+		c.curSlotsInput = ctx.Input().FromInts(slots, len(slots))
+		return nil
+	}
+
+	// Ensure slots exist for sequences in this batch
+	c.curSlots = c.curSlots[:0]
+	var newSlots []int // track newly allocated slots that need zeroing
+	for _, s := range c.curSeqs {
+		slot, ok := c.slotForSeq[s]
+		if !ok {
+			var err error
+			slot, err = c.allocSlot()
+			if err != nil {
+				return err
+			}
+			c.slotForSeq[s] = slot
+			c.refCount[slot] = 1
+			newSlots = append(newSlots, slot)
+		}
+		c.curSlots = append(c.curSlots, slot)
+	}
+
+	// Zero conv state for newly allocated slots to clear stale data from previous sequences
+	if len(newSlots) > 0 {
+		c.zeroConvSlots(ctx, newSlots)
+	}
+
+	// Create a tensor for the current slots
+	slots := make([]int32, len(c.curSlots))
+	for i, v := range c.curSlots {
+		slots[i] = int32(v)
+	}
+	c.curSlotsInput = ctx.Input().FromInts(slots, len(slots))
+
+	// Reset writable state for new forward pass
+	c.writableEnsured = false
+	c.writableError = nil
+
+	return nil
+}
+
+func (c *HybridCache) allocSlot() (int, error) {
+	if len(c.freeSlots) == 0 {
+		return 0, kvcache.ErrKvCacheFull
+	}
+	slot := c.freeSlots[len(c.freeSlots)-1]
+	c.freeSlots = c.freeSlots[:len(c.freeSlots)-1]
+	return slot, nil
+}
+
+func (c *HybridCache) freeSlot(slot int) {
+	// Bounds check before freeing
+	if slot >= 0 && slot < c.maxSequences {
+		c.freeSlots = append(c.freeSlots, slot)
+	}
+}
+
+// zeroConvSlots zeros the conv state for the given slots across all layers.
+// This must be called when recycling slots to prevent stale state from affecting new sequences.
+func (c *HybridCache) zeroConvSlots(ctx ml.Context, slots []int) {
+	if len(slots) == 0 || len(c.convStates) == 0 {
+		return
+	}
+
+	// Use input context for creating tensors
+	inputCtx := ctx.Input()
+
+	// Create slot indices tensor
+	slotIndices := make([]int32, len(slots))
+	for i, s := range slots {
+		slotIndices[i] = int32(s)
+	}
+	slotsTensor := inputCtx.FromInts(slotIndices, len(slotIndices))
+
+	// Create zero tensor for the slots (SetRows requires F32 source)
+	zeros := inputCtx.Zeros(ml.DTypeF32, c.dConv*c.hiddenSize, len(slots))
+
+	// Zero each layer's conv state for these slots
+	for _, buf := range c.convStates {
+		ctx.Forward(buf.SetRows(ctx, zeros, slotsTensor))
+	}
+}
+
+// EnsureWritable ensures that sequences in the current batch have private (non-shared) conv slots.
+// Returns an error if slot allocation fails.
+func (c *HybridCache) EnsureWritable(ctx ml.Context) error {
+	for i, seq := range c.curSeqs {
+		slot, ok := c.slotForSeq[seq]
+		if !ok {
+			continue
+		}
+
+		// Bounds check
+		if slot < 0 || slot >= len(c.refCount) {
+			continue
+		}
+
+		if c.refCount[slot] <= 1 {
+			continue
+		}
+
+		newSlot, err := c.allocSlot()
+		if err != nil {
+			return err
+		}
+		c.refCount[slot]--
+		c.refCount[newSlot] = 1
+		c.slotForSeq[seq] = newSlot
+		c.curSlots[i] = newSlot
+
+		// Copy existing conv state for all initialized layers
+		for _, buf := range c.convStates {
+			// buf: [dConv*hiddenSize, maxSlots]
+			src := buf.Rows(ctx, ctx.Input().FromInts([]int32{int32(slot)}, 1))
+			// SetRows requires F32 source
+			srcF32 := src.Cast(ctx, ml.DTypeF32)
+			ctx.Forward(buf.SetRows(ctx, srcF32, ctx.Input().FromInts([]int32{int32(newSlot)}, 1)))
+		}
+	}
+
+	// Rebuild current slots tensor
+	slots := make([]int32, len(c.curSlots))
+	for i, v := range c.curSlots {
+		slots[i] = int32(v)
+	}
+	c.curSlotsInput = ctx.Input().FromInts(slots, len(slots))
+
+	return nil
+}
+
+func (c *HybridCache) CopyPrefix(srcSeq, dstSeq int, prefixLen int32) {
+	// KV cache shares prefix metadata (no copy) which is correct for prefix reuse.
+	c.kv.CopyPrefix(srcSeq, dstSeq, prefixLen)
+
+	// For shortconv state we implement copy-on-write: dst shares the same slot as src.
+	// On the first write to dst, EnsureWritable will create a private slot.
+	if dstSlot, ok := c.slotForSeq[dstSeq]; ok {
+		// Bounds check before decrementing
+		if dstSlot >= 0 && dstSlot < len(c.refCount) {
+			c.refCount[dstSlot]--
+			if c.refCount[dstSlot] <= 0 {
+				c.refCount[dstSlot] = 0
+				c.freeSlot(dstSlot)
+			}
+		}
+		delete(c.slotForSeq, dstSeq)
+	}
+
+	srcSlot, ok := c.slotForSeq[srcSeq]
+	if !ok {
+		// src may not have a slot yet; dst will allocate on demand
+		return
+	}
+
+	// Bounds check before incrementing
+	if srcSlot >= 0 && srcSlot < len(c.refCount) {
+		c.slotForSeq[dstSeq] = srcSlot
+		c.refCount[srcSlot]++
+	}
+}
+
+func (c *HybridCache) CanResume(seq int, pos int32) bool {
+	return c.kv.CanResume(seq, pos)
+}
+
+func (c *HybridCache) Remove(seq int, beginIndex, endIndex int32) error {
+	if err := c.kv.Remove(seq, beginIndex, endIndex); err != nil {
+		return err
+	}
+
+	// For recurrent state, any removal invalidates the state because
+	// the state at position N depends on all previous positions.
+	// Drop the slot mapping so it resets on next use.
+	slot, ok := c.slotForSeq[seq]
+	if !ok {
+		return nil
+	}
+
+	// Bounds check
+	if slot < 0 || slot >= len(c.refCount) {
+		delete(c.slotForSeq, seq)
+		return nil
+	}
+
+	c.refCount[slot]--
+	if c.refCount[slot] <= 0 {
+		c.refCount[slot] = 0
+		c.freeSlot(slot)
+	}
+	delete(c.slotForSeq, seq)
+
+	return nil
+}
+
+func (c *HybridCache) slotsTensor() ml.Tensor {
+	return c.curSlotsInput
+}
+
+func (c *HybridCache) seqTokens() int {
+	return c.curSeqTokens
+}
+
+func (c *HybridCache) numSeqs() int {
+	return len(c.curSeqs)
+}
+
+func (c *HybridCache) convBuffer(ctx ml.Context, layer int) ml.Tensor {
+	if buf, ok := c.convStates[layer]; ok {
+		return buf
+	}
+
+	if _, ok := c.convCtxs[layer]; !ok {
+		c.convCtxs[layer] = c.backend.NewContextSize(1).Layer(layer)
+	}
+
+	buf := c.convCtxs[layer].Zeros(c.dtype, c.dConv*c.hiddenSize, c.maxSequences)
+	c.convStates[layer] = buf
+	return buf
+}
+
+// ConvState returns the conv state for current batch sequences as shape [dConv, hiddenSize, nSeqs].
+// Returns an error if copy-on-write allocation fails.
+func (c *HybridCache) ConvState(ctx ml.Context, layer int) (ml.Tensor, error) {
+	if !c.writableEnsured {
+		needsWritable := false
+		for _, seq := range c.curSeqs {
+			slot, ok := c.slotForSeq[seq]
+			if !ok {
+				continue
+			}
+			if slot >= 0 && slot < len(c.refCount) && c.refCount[slot] > 1 {
+				needsWritable = true
+				break
+			}
+		}
+
+		if needsWritable {
+			if err := c.EnsureWritable(ctx); err != nil {
+				c.writableError = err
+			}
+		}
+		c.writableEnsured = true
+	}
+
+	if c.writableError != nil {
+		return nil, c.writableError
+	}
+
+	buf := c.convBuffer(ctx, layer)
+	cur := buf.Rows(ctx, c.slotsTensor())
+	return cur.Reshape(ctx, c.dConv, c.hiddenSize, c.numSeqs()), nil
+}
+
+// UpdateConvState writes a new conv state for current batch sequences.
+// newState must have shape [dConv, hiddenSize, nSeqs].
+func (c *HybridCache) UpdateConvState(ctx ml.Context, layer int, newState ml.Tensor) {
+	buf := c.convBuffer(ctx, layer)
+	src := newState.Reshape(ctx, c.dConv*c.hiddenSize, c.numSeqs())
+	// SetRows requires F32 source
+	srcF32 := src.Cast(ctx, ml.DTypeF32)
+	ctx.Forward(buf.SetRows(ctx, srcF32, c.slotsTensor()))
+}
+
+// IsSupportedForBatch returns true if the current batch layout supports shortconv.
+func (c *HybridCache) IsSupportedForBatch() bool {
+	return c.curSeqTokens > 0 && len(c.curSeqs) > 0
+}
+
+// Seqs returns the ordered unique sequences for the current forward pass.
+func (c *HybridCache) Seqs() []int {
+	return slices.Clone(c.curSeqs)
+}

model/models/lfm2/cache_test.go

@@ -0,0 +1,444 @@
+package lfm2
+
+import (
+	"testing"
+
+	"github.com/ollama/ollama/kvcache"
+	"github.com/ollama/ollama/ml"
+)
+
+// TestHybridCache tests verify the slot management logic of HybridCache.
+// These tests focus on the recurrent state slot allocation, reference counting,
+// and copy-on-write semantics without requiring a full ML backend.
+
+// createSlotOnlyCache creates a HybridCache with only the slot management
+// fields initialized. Used to test slot logic in isolation.
+func createSlotOnlyCache(maxSequences int) *HybridCache {
+	return &HybridCache{
+		hiddenSize:   256,
+		dConv:        3,
+		maxSequences: maxSequences,
+		refCount:     make([]int, maxSequences),
+		freeSlots:    initFreeSlots(maxSequences),
+		slotForSeq:   make(map[int]int),
+		convCtxs:     make(map[int]ml.Context),
+		convStates:   make(map[int]ml.Tensor),
+	}
+}
+
+func initFreeSlots(n int) []int {
+	slots := make([]int, 0, n)
+	for i := n - 1; i >= 0; i-- {
+		slots = append(slots, i)
+	}
+	return slots
+}
+
+func TestHybridCache_SlotAllocation(t *testing.T) {
+	cache := createSlotOnlyCache(4)
+
+	// Verify initial state
+	if len(cache.freeSlots) != 4 {
+		t.Errorf("expected 4 free slots, got %d", len(cache.freeSlots))
+	}
+
+	// Allocate all slots
+	for range 4 {
+		slot, err := cache.allocSlot()
+		if err != nil {
+			t.Fatalf("allocSlot failed: %v", err)
+		}
+		cache.refCount[slot] = 1
+	}
+
+	// Should be full now
+	if len(cache.freeSlots) != 0 {
+		t.Errorf("expected 0 free slots, got %d", len(cache.freeSlots))
+	}
+
+	// Trying to allocate another should fail
+	_, err := cache.allocSlot()
+	if err != kvcache.ErrKvCacheFull {
+		t.Errorf("expected ErrKvCacheFull, got %v", err)
+	}
+}
+
+func TestHybridCache_SlotReuse(t *testing.T) {
+	cache := createSlotOnlyCache(4)
+
+	// Allocate a slot
+	slot1, _ := cache.allocSlot()
+	cache.refCount[slot1] = 1
+
+	// Free it
+	cache.refCount[slot1] = 0
+	cache.freeSlot(slot1)
+
+	// Allocate again - should get the same slot back (LIFO)
+	slot2, _ := cache.allocSlot()
+	if slot2 != slot1 {
+		t.Errorf("expected slot %d to be reused, got %d", slot1, slot2)
+	}
+}
+
+func TestHybridCache_SlotRefCounting_ShareSlot(t *testing.T) {
+	cache := createSlotOnlyCache(4)
+
+	// Allocate slot for seq 1
+	slot1, _ := cache.allocSlot()
+	cache.slotForSeq[1] = slot1
+	cache.refCount[slot1] = 1
+
+	// Simulate sharing slot with seq 2 (copy-on-write style)
+	cache.slotForSeq[2] = slot1
+	cache.refCount[slot1]++
+
+	// Should share the same slot
+	if cache.slotForSeq[2] != slot1 {
+		t.Errorf("expected seq 2 to share slot %d, got %d", slot1, cache.slotForSeq[2])
+	}
+
+	// Ref count should be 2
+	if cache.refCount[slot1] != 2 {
+		t.Errorf("expected refCount 2, got %d", cache.refCount[slot1])
+	}
+}
+
+func TestHybridCache_SlotRefCounting_DecRef(t *testing.T) {
+	cache := createSlotOnlyCache(4)
+
+	// Allocate slot for seq 1
+	slot1, _ := cache.allocSlot()
+	cache.slotForSeq[1] = slot1
+	cache.refCount[slot1] = 1
+
+	// Share with seq 2
+	cache.slotForSeq[2] = slot1
+	cache.refCount[slot1]++
+
+	// Unshare seq 2
+	cache.refCount[slot1]--
+	delete(cache.slotForSeq, 2)
+
+	// Ref count should be back to 1
+	if cache.refCount[slot1] != 1 {
+		t.Errorf("expected refCount 1 after unshare, got %d", cache.refCount[slot1])
+	}
+
+	// Seq 2 should no longer have a slot
+	if _, ok := cache.slotForSeq[2]; ok {
+		t.Error("seq 2 should not have a slot after unshare")
+	}
+}
+
+func TestHybridCache_SlotFreeWhenUnused(t *testing.T) {
+	cache := createSlotOnlyCache(4)
+
+	initialFreeSlots := len(cache.freeSlots)
+
+	// Allocate slot for seq 1
+	slot1, _ := cache.allocSlot()
+	cache.slotForSeq[1] = slot1
+	cache.refCount[slot1] = 1
+
+	// Free the slot when refCount drops to 0
+	cache.refCount[slot1]--
+	if cache.refCount[slot1] <= 0 {
+		cache.refCount[slot1] = 0
+		cache.freeSlot(slot1)
+	}
+	delete(cache.slotForSeq, 1)
+
+	// Slot should be freed
+	if len(cache.freeSlots) != initialFreeSlots {
+		t.Errorf("expected %d free slots, got %d", initialFreeSlots, len(cache.freeSlots))
+	}
+
+	// Ref count should be 0
+	if cache.refCount[slot1] != 0 {
+		t.Errorf("expected refCount 0, got %d", cache.refCount[slot1])
+	}
+}
+
+func TestHybridCache_SlotOverwrite(t *testing.T) {
+	cache := createSlotOnlyCache(4)
+
+	// Allocate slots for seq 1 and seq 2
+	slot1, _ := cache.allocSlot()
+	cache.slotForSeq[1] = slot1
+	cache.refCount[slot1] = 1
+
+	slot2, _ := cache.allocSlot()
+	cache.slotForSeq[2] = slot2
+	cache.refCount[slot2] = 1
+
+	initialFreeSlots := len(cache.freeSlots)
+
+	// Simulate overwriting seq 2's slot with slot1 (sharing)
+	// First free the old slot
+	cache.refCount[slot2]--
+	if cache.refCount[slot2] <= 0 {
+		cache.refCount[slot2] = 0
+		cache.freeSlot(slot2)
+	}
+	// Then share slot1
+	cache.slotForSeq[2] = slot1
+	cache.refCount[slot1]++
+
+	// Seq 2 should now share slot1
+	if cache.slotForSeq[2] != slot1 {
+		t.Errorf("expected seq 2 to share slot %d, got %d", slot1, cache.slotForSeq[2])
+	}
+
+	// Old slot2 should be freed
+	if len(cache.freeSlots) != initialFreeSlots+1 {
+		t.Errorf("expected %d free slots, got %d", initialFreeSlots+1, len(cache.freeSlots))
+	}
+}
+
+func TestHybridCache_BoundsChecking(t *testing.T) {
+	cache := createSlotOnlyCache(4)
+
+	// Test freeing invalid slot (should not panic)
+	cache.freeSlot(-1)
+	cache.freeSlot(100) // out of bounds
+
+	// freeSlot does bounds checking, so invalid slots should be ignored
+	if len(cache.freeSlots) != 4 {
+		t.Errorf("invalid slots should not affect free list, got %d slots", len(cache.freeSlots))
+	}
+}
+
+func TestHybridCache_MultipleSequences_RefCounting(t *testing.T) {
+	cache := createSlotOnlyCache(8)
+
+	// Allocate slot for seq 1
+	slot1, _ := cache.allocSlot()
+	cache.slotForSeq[1] = slot1
+	cache.refCount[slot1] = 1
+
+	// Fork to seq 2, 3, 4 (all share slot1)
+	for _, seq := range []int{2, 3, 4} {
+		cache.slotForSeq[seq] = slot1
+		cache.refCount[slot1]++
+	}
+
+	// Ref count should be 4
+	if cache.refCount[slot1] != 4 {
+		t.Errorf("expected refCount 4, got %d", cache.refCount[slot1])
+	}
+
+	// Remove seq 2, 3
+	for _, seq := range []int{2, 3} {
+		delete(cache.slotForSeq, seq)
+		cache.refCount[slot1]--
+	}
+
+	if cache.refCount[slot1] != 2 {
+		t.Errorf("expected refCount 2, got %d", cache.refCount[slot1])
+	}
+
+	// Slot should still be allocated (not in free list)
+	found := false
+	for _, s := range cache.freeSlots {
+		if s == slot1 {
+			found = true
+			break
+		}
+	}
+	if found {
+		t.Error("slot1 should not be in free list yet")
+	}
+
+	// Remove remaining sequences
+	for _, seq := range []int{1, 4} {
+		delete(cache.slotForSeq, seq)
+		cache.refCount[slot1]--
+	}
+
+	if cache.refCount[slot1] != 0 {
+		t.Errorf("expected refCount 0, got %d", cache.refCount[slot1])
+	}
+}
+
+func TestHybridCache_ChainedSharing(t *testing.T) {
+	cache := createSlotOnlyCache(8)
+
+	// Create seq 1
+	slot1, _ := cache.allocSlot()
+	cache.slotForSeq[1] = slot1
+	cache.refCount[slot1] = 1
+
+	// Share 1 -> 2
+	cache.slotForSeq[2] = slot1
+	cache.refCount[slot1]++
+
+	// Share 2 -> 3 (should still share slot1)
+	cache.slotForSeq[3] = cache.slotForSeq[2] // which is slot1
+	cache.refCount[slot1]++
+
+	// All should share slot1
+	if cache.slotForSeq[1] != slot1 || cache.slotForSeq[2] != slot1 || cache.slotForSeq[3] != slot1 {
+		t.Error("all sequences should share slot1")
+	}
+
+	if cache.refCount[slot1] != 3 {
+		t.Errorf("expected refCount 3, got %d", cache.refCount[slot1])
+	}
+}
+
+func TestHybridCache_CacheParameters(t *testing.T) {
+	cache := NewHybridCache(nil, 512, 5) // hiddenSize=512, dConv=5
+
+	if cache.hiddenSize != 512 {
+		t.Errorf("expected hiddenSize 512, got %d", cache.hiddenSize)
+	}
+	if cache.dConv != 5 {
+		t.Errorf("expected dConv 5, got %d", cache.dConv)
+	}
+}
+
+func TestHybridCache_NumSeqs(t *testing.T) {
+	cache := createSlotOnlyCache(4)
+
+	// Initially no sequences
+	if cache.numSeqs() != 0 {
+		t.Errorf("expected 0 seqs, got %d", cache.numSeqs())
+	}
+
+	// Manually set up current batch state
+	cache.curSeqs = []int{1, 2, 3}
+
+	if cache.numSeqs() != 3 {
+		t.Errorf("expected 3 seqs, got %d", cache.numSeqs())
+	}
+}
+
+func TestHybridCache_SeqTokens(t *testing.T) {
+	cache := createSlotOnlyCache(4)
+
+	// Initially 0
+	if cache.seqTokens() != 0 {
+		t.Errorf("expected 0 seqTokens, got %d", cache.seqTokens())
+	}
+
+	// Manually set up current batch state
+	cache.curSeqTokens = 16
+
+	if cache.seqTokens() != 16 {
+		t.Errorf("expected 16 seqTokens, got %d", cache.seqTokens())
+	}
+}
+
+// Test that Seqs returns a clone of curSeqs
+func TestHybridCache_Seqs_ReturnsClone(t *testing.T) {
+	cache := createSlotOnlyCache(4)
+
+	cache.curSeqs = []int{1, 2, 3}
+
+	seqs := cache.Seqs()
+
+	// Modify returned slice
+	seqs[0] = 999
+
+	// Original should be unchanged
+	if cache.curSeqs[0] != 1 {
+		t.Error("Seqs should return a clone, not the original slice")
+	}
+}
+
+func TestHybridCache_IsSupportedForBatch(t *testing.T) {
+	cache := createSlotOnlyCache(4)
+
+	// Initially not supported (no batch set up)
+	if cache.IsSupportedForBatch() {
+		t.Error("expected IsSupportedForBatch to be false initially")
+	}
+
+	// Set up a valid batch
+	cache.curSeqTokens = 1
+	cache.curSeqs = []int{1}
+
+	if !cache.IsSupportedForBatch() {
+		t.Error("expected IsSupportedForBatch to be true with valid batch")
+	}
+}
+
+func TestHybridCache_ZeroConvSlots_EmptyInputs(t *testing.T) {
+	cache := createSlotOnlyCache(4)
+
+	// zeroConvSlots should handle empty slots without panicking
+	cache.zeroConvSlots(nil, nil)
+	cache.zeroConvSlots(nil, []int{})
+
+	// zeroConvSlots should handle empty convStates without panicking
+	cache.zeroConvSlots(nil, []int{0, 1, 2})
+}
+
+func TestHybridCache_SlotRecycling_TracksNewSlots(t *testing.T) {
+	cache := createSlotOnlyCache(4)
+
+	// Allocate slot for seq 1
+	slot1, _ := cache.allocSlot()
+	cache.slotForSeq[1] = slot1
+	cache.refCount[slot1] = 1
+
+	// Free the slot (simulating sequence removal)
+	cache.refCount[slot1]--
+	cache.freeSlot(slot1)
+	delete(cache.slotForSeq, 1)
+
+	// Verify slot is in free list
+	if len(cache.freeSlots) != 4 {
+		t.Errorf("expected 4 free slots after freeing, got %d", len(cache.freeSlots))
+	}
+
+	// Allocate for new seq 2 - should get recycled slot
+	slot2, _ := cache.allocSlot()
+	if slot2 != slot1 {
+		t.Errorf("expected recycled slot %d, got %d", slot1, slot2)
+	}
+
+	// This recycled slot would need zeroing in the real implementation
+	// The actual zeroing is tested via integration tests since it requires ML context
+}
+
+func TestHybridCache_NewSequence_GetsTrackedForZeroing(t *testing.T) {
+	cache := createSlotOnlyCache(4)
+
+	// Simulate the slot allocation flow from StartForward
+	// When a sequence doesn't have a slot, it gets allocated and tracked as "new"
+
+	newSlots := []int{}
+
+	// Seq 1 doesn't have a slot - allocate and track
+	seq := 1
+	if _, ok := cache.slotForSeq[seq]; !ok {
+		slot, err := cache.allocSlot()
+		if err != nil {
+			t.Fatalf("allocSlot failed: %v", err)
+		}
+		cache.slotForSeq[seq] = slot
+		cache.refCount[slot] = 1
+		newSlots = append(newSlots, slot)
+	}
+
+	// Verify newSlots contains the allocated slot
+	if len(newSlots) != 1 {
+		t.Errorf("expected 1 new slot, got %d", len(newSlots))
+	}
+
+	// Seq 1 already has a slot - should NOT be tracked as new
+	newSlots2 := []int{}
+	if _, ok := cache.slotForSeq[seq]; !ok {
+		slot, _ := cache.allocSlot()
+		cache.slotForSeq[seq] = slot
+		cache.refCount[slot] = 1
+		newSlots2 = append(newSlots2, slot)
+	}
+
+	// Verify no new slots for existing sequence
+	if len(newSlots2) != 0 {
+		t.Errorf("expected 0 new slots for existing sequence, got %d", len(newSlots2))
+	}
+}

model/models/lfm2/model.go

@@ -0,0 +1,253 @@
+package lfm2
+
+import (
+	"cmp"
+	"math"
+
+	"github.com/ollama/ollama/fs"
+	"github.com/ollama/ollama/ml"
+	"github.com/ollama/ollama/ml/nn"
+	"github.com/ollama/ollama/ml/nn/rope"
+	"github.com/ollama/ollama/model"
+	"github.com/ollama/ollama/model/input"
+)
+
+type Options struct {
+	hiddenSize       int
+	headDim, ropeDim int
+
+	eps, ropeBase, ropeScale float32
+
+	ropeType              string
+	originalContextLength int
+
+	// per-layer head counts (LFM2 alternates attention and recurrent layers)
+	numHeadsByLayer   []int
+	numKVHeadsByLayer []int
+}
+
+func (o Options) headDimValue() int {
+	// Head dim is shared across layers; fall back to first attention layer head count.
+	for _, h := range o.numHeadsByLayer {
+		if h > 0 {
+			return cmp.Or(o.headDim, o.hiddenSize/h)
+		}
+	}
+	return cmp.Or(o.headDim, o.hiddenSize)
+}
+
+func (o Options) applyRotaryPositionEmbeddings(ctx ml.Context, states, positions ml.Tensor) ml.Tensor {
+	opts := []func(*rope.Options){rope.WithTypeNeoX()}
+	if o.ropeType == "yarn" {
+		attnFactor := float32(1.0 / (1.0 + 0.1*math.Log(float64(o.ropeScale))))
+		opts = append(opts,
+			rope.WithOriginalContextLength(o.originalContextLength),
+			rope.WithExtrapolationFactor(1.),
+			rope.WithAttentionFactor(attnFactor),
+		)
+	}
+
+	headCount := 1
+	for _, h := range o.numHeadsByLayer {
+		if h > 0 {
+			headCount = h
+			break
+		}
+	}
+	return nn.RoPE(ctx, states, positions, cmp.Or(o.ropeDim, o.headDim, o.hiddenSize/headCount), o.ropeBase, 1./o.ropeScale, opts...)
+}
+
+type Model struct {
+	model.Base
+	model.TextProcessor
+
+	TokenEmbedding *nn.Embedding `gguf:"token_embd"`
+	Layers         []Layer       `gguf:"blk"`
+	OutputNorm     *nn.RMSNorm   `gguf:"output_norm,alt:token_embd_norm"`
+	Output         *nn.Linear    `gguf:"output,alt:token_embd"`
+
+	Options
+}
+
+func New(c fs.Config) (model.Model, error) {
+	if c.Uint("expert_count") > 0 {
+		return nil, model.ErrUnsupportedModel
+	}
+
+	if c.String("tokenizer.ggml.model") != "gpt2" {
+		return nil, model.ErrUnsupportedTokenizer
+	}
+
+	vocabulary := model.Vocabulary{
+		Values: c.Strings("tokenizer.ggml.tokens"),
+		Scores: c.Floats("tokenizer.ggml.scores"),
+		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")...,
+		),
+	}
+
+	var pretokenizers []string
+	switch c.String("tokenizer.ggml.pre") {
+	case "default":
+		// use default BPE pretokenizer
+	default:
+		// llama-bpe style (default for LFM2)
+		pretokenizers = []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+`,
+		}
+	}
+
+	m := Model{
+		TextProcessor: model.NewBytePairEncoding(&vocabulary, pretokenizers...),
+		Layers:        make([]Layer, c.Uint("block_count")),
+		Options: Options{
+			hiddenSize:            int(c.Uint("embedding_length")),
+			headDim:               int(c.Uint("attention.key_length")),
+			ropeDim:               int(c.Uint("rope.dimension_count")),
+			eps:                   c.Float("attention.layer_norm_rms_epsilon"),
+			ropeType:              c.String("rope.scaling.type"),
+			ropeBase:              c.Float("rope.freq_base"),
+			ropeScale:             c.Float("rope.scaling.factor", 1),
+			originalContextLength: int(c.Uint("rope.scaling.original_context_length")),
+		},
+	}
+
+	type headCounts interface {
+		HeadCount() []uint64
+		HeadCountKV() []uint64
+	}
+	hc, ok := c.(headCounts)
+	if !ok {
+		return nil, model.ErrUnsupportedModel
+	}
+
+	headCount := hc.HeadCount()
+	headCountKV := hc.HeadCountKV()
+
+	m.numHeadsByLayer = make([]int, len(m.Layers))
+	m.numKVHeadsByLayer = make([]int, len(m.Layers))
+	for i := range m.Layers {
+		m.numHeadsByLayer[i] = int(headCount[i])
+		m.numKVHeadsByLayer[i] = int(headCountKV[i])
+
+		if m.numKVHeadsByLayer[i] == 0 {
+			m.Layers[i].Operator = &ShortConv{}
+		} else {
+			m.Layers[i].Operator = &Attention{}
+		}
+	}
+
+	lCache := int(c.Uint("shortconv.l_cache"))
+	dConv := max(0, lCache-1)
+	m.Cache = NewHybridCache(m.Shift, m.hiddenSize, dConv)
+	return &m, nil
+}
+
+type Operator interface {
+	Forward(ctx ml.Context, hiddenStates, positions ml.Tensor, cache *HybridCache, layer int, opts *Options) ml.Tensor
+}
+
+type Attention struct {
+	Query     *nn.Linear  `gguf:"attn_q"`
+	QueryNorm *nn.RMSNorm `gguf:"attn_q_norm"`
+	Key       *nn.Linear  `gguf:"attn_k"`
+	KeyNorm   *nn.RMSNorm `gguf:"attn_k_norm"`
+	Value     *nn.Linear  `gguf:"attn_v"`
+	Output    *nn.Linear  `gguf:"attn_output,alt:attn_out"`
+}
+
+func (sa *Attention) Forward(ctx ml.Context, hiddenStates, positions ml.Tensor, cache *HybridCache, layer int, opts *Options) ml.Tensor {
+	batchSize := hiddenStates.Dim(1)
+	headDim := opts.headDimValue()
+	numHeads := opts.numHeadsByLayer[layer]
+	numKVHeads := opts.numKVHeadsByLayer[layer]
+
+	query := sa.Query.Forward(ctx, hiddenStates)
+	key := sa.Key.Forward(ctx, hiddenStates)
+	value := sa.Value.Forward(ctx, hiddenStates)
+
+	query = query.Reshape(ctx, headDim, numHeads, batchSize)
+	key = key.Reshape(ctx, headDim, numKVHeads, batchSize)
+	value = value.Reshape(ctx, headDim, numKVHeads, batchSize)
+
+	query = sa.QueryNorm.Forward(ctx, query, opts.eps)
+	key = sa.KeyNorm.Forward(ctx, key, opts.eps)
+
+	query = opts.applyRotaryPositionEmbeddings(ctx, query, positions)
+	key = opts.applyRotaryPositionEmbeddings(ctx, key, positions)
+
+	attention := nn.Attention(ctx, query, key, value, 1./math.Sqrt(float64(headDim)), cache)
+	attention = attention.Reshape(ctx, attention.Dim(0)*attention.Dim(1), batchSize)
+	return sa.Output.Forward(ctx, attention)
+}
+
+type MLP struct {
+	Up   *nn.Linear `gguf:"ffn_up"`
+	Down *nn.Linear `gguf:"ffn_down"`
+	Gate *nn.Linear `gguf:"ffn_gate"`
+}
+
+func (mlp *MLP) Forward(ctx ml.Context, hiddenState ml.Tensor, opts *Options) ml.Tensor {
+	hiddenState = mlp.Gate.Forward(ctx, hiddenState).SILU(ctx, mlp.Up.Forward(ctx, hiddenState))
+	return mlp.Down.Forward(ctx, hiddenState)
+}
+
+type Layer struct {
+	AttentionNorm *nn.RMSNorm `gguf:"attn_norm"`
+	Operator      Operator
+	MLPNorm       *nn.RMSNorm `gguf:"ffn_norm"`
+	MLP           *MLP
+}
+
+func (l *Layer) Forward(ctx ml.Context, layer int, hiddenState, positions, outputs ml.Tensor, cache *HybridCache, opts *Options) ml.Tensor {
+	residual := hiddenState
+
+	hiddenState = l.AttentionNorm.Forward(ctx, hiddenState, opts.eps)
+	hiddenState = l.Operator.Forward(ctx, hiddenState, positions, cache, layer, opts)
+
+	if outputs != nil {
+		hiddenState = hiddenState.Rows(ctx, outputs)
+		residual = residual.Rows(ctx, outputs)
+	}
+
+	hiddenState = hiddenState.Add(ctx, residual)
+	residual = hiddenState
+
+	hiddenState = l.MLPNorm.Forward(ctx, hiddenState, opts.eps)
+	hiddenState = l.MLP.Forward(ctx, hiddenState, opts)
+	return hiddenState.Add(ctx, residual)
+}
+
+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))
+
+	hiddenState := 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
+		}
+
+		hiddenState = layer.Forward(ctx, i, hiddenState, positions, outputs, m.Cache.(*HybridCache), &m.Options)
+	}
+
+	hiddenState = m.OutputNorm.Forward(ctx, hiddenState, m.eps)
+	return m.Output.Forward(ctx, hiddenState), nil
+}
+
+func init() {
+	model.Register("lfm2", New)
+}

model/models/lfm2/shortconv.go

@@ -0,0 +1,50 @@
+package lfm2
+
+import (
+	"github.com/ollama/ollama/ml"
+	"github.com/ollama/ollama/ml/nn"
+)
+
+type shortConvKernel struct {
+	Weight ml.Tensor `gguf:"weight"`
+}
+
+// ShortConv implements the LFM2 short-convolution block (GGML_OP_SSM_CONV) with a recurrent
+// state stored in the HybridCache.
+type ShortConv struct {
+	Conv    *shortConvKernel `gguf:"shortconv.conv"`
+	InProj  *nn.Linear       `gguf:"shortconv.in_proj"`
+	OutProj *nn.Linear       `gguf:"shortconv.out_proj"`
+}
+
+func (sc *ShortConv) Forward(ctx ml.Context, hiddenStates ml.Tensor, _ ml.Tensor, cache *HybridCache, layer int, opts *Options) ml.Tensor {
+	nSeqs := cache.numSeqs()
+	seqTokens := cache.seqTokens()
+	hiddenSize := hiddenStates.Dim(0)
+	if nSeqs <= 0 || seqTokens <= 0 || hiddenStates.Dim(1) != nSeqs*seqTokens {
+		panic("lfm2: unsupported batch layout for shortconv")
+	}
+
+	bcx := sc.InProj.Forward(ctx, hiddenStates).Reshape(ctx, 3*hiddenSize, seqTokens, nSeqs)
+
+	elementSize := bcx.Stride(0)
+	b := bcx.View(ctx, 0*hiddenSize*elementSize, hiddenSize, bcx.Stride(1), seqTokens, bcx.Stride(2), nSeqs)
+	c := bcx.View(ctx, 1*hiddenSize*elementSize, hiddenSize, bcx.Stride(1), seqTokens, bcx.Stride(2), nSeqs)
+	x := bcx.View(ctx, 2*hiddenSize*elementSize, hiddenSize, bcx.Stride(1), seqTokens, bcx.Stride(2), nSeqs)
+
+	bx := b.Mul(ctx, x).Permute(ctx, 1, 0, 2, 3)
+
+	state, err := cache.ConvState(ctx, layer)
+	if err != nil {
+		panic("lfm2: failed to get conv state: " + err.Error())
+	}
+	sx := state.Concat(ctx, bx, 0)
+
+	convOut := sx.SSMConv(ctx, sc.Conv.Weight)
+	y := c.Mul(ctx, convOut)
+
+	dConv := sx.Dim(0) - seqTokens
+	cache.UpdateConvState(ctx, layer, sx.Slice(ctx, 0, sx.Dim(0)-dConv, sx.Dim(0), 1))
+
+	return sc.OutProj.Forward(ctx, y.Reshape(ctx, hiddenSize, seqTokens*nSeqs))
+}

model/models/models.go

@@ -7,7 +7,9 @@ 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/lfm2"
 	_ "github.com/ollama/ollama/model/models/llama"
 	_ "github.com/ollama/ollama/model/models/llama4"
 	_ "github.com/ollama/ollama/model/models/mistral3"

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)
+			}
+		})
+	}
+}