fix tests

Add --quantize fp4 support to ollama create for image generation models
(flux2, z-image-turbo), using MLX's affine 4-bit quantization.

Changes:
- Add fp4 to validation in CreateImageGenModel
- Add FP4 case to quantizeTensor (group_size=32, bits=4, affine mode)
- Add GetQuantization() to WeightSource interface for dynamic params
- Update LoadLinearLayer to use quantization params from model metadata

CMakeLists.txt

@@ -190,7 +190,7 @@ 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

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
@@ -156,14 +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
-RUN mkdir -p dist/bin
-RUN --mount=type=cache,target=/root/.cache/go-build \
-    go build -tags mlx -trimpath -buildmode=pie -o dist/bin/ollama-mlx .
 
 FROM base AS build
 WORKDIR /go/src/github.com/ollama/ollama
@@ -172,12 +165,14 @@ RUN curl -fsSL https://golang.org/dl/go$(awk '/^go/ { print $2 }' go.mod).linux-
 ENV PATH=/usr/local/go/bin:$PATH
 RUN go mod download
 COPY . .
+# Clone mlx-c headers for CGO (version from MLX_VERSION file)
+RUN git clone --depth 1 --branch "$(cat MLX_VERSION)" https://github.com/ml-explore/mlx-c.git build/_deps/mlx-c-src
 ARG GOFLAGS="'-ldflags=-w -s'"
 ENV CGO_ENABLED=1
-ARG CGO_CFLAGS
+ENV CGO_CFLAGS="-I/go/src/github.com/ollama/ollama/build/_deps/mlx-c-src"
 ARG CGO_CXXFLAGS
 RUN --mount=type=cache,target=/root/.cache/go-build \
-    go build -trimpath -buildmode=pie -o /bin/ollama .
+    go build -tags mlx -trimpath -buildmode=pie -o /bin/ollama .
 
 FROM --platform=linux/amd64 scratch AS amd64
 # COPY --from=cuda-11 dist/lib/ollama/ /lib/ollama/
@@ -185,7 +180,6 @@ COPY --from=cuda-12 dist/lib/ollama /lib/ollama/
 COPY --from=cuda-13 dist/lib/ollama /lib/ollama/
 COPY --from=vulkan  dist/lib/ollama  /lib/ollama/
 COPY --from=mlx     /go/src/github.com/ollama/ollama/dist/lib/ollama /lib/ollama/
-COPY --from=mlx     /go/src/github.com/ollama/ollama/dist/bin/ /bin/
 
 FROM --platform=linux/arm64 scratch AS arm64
 # 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)
@@ -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/types.go

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

app/cmd/app/app_darwin.m

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

cmd/cmd.go

@@ -46,8 +46,9 @@ import (
 	"github.com/ollama/ollama/types/syncmap"
 	"github.com/ollama/ollama/version"
 	xcmd "github.com/ollama/ollama/x/cmd"
+	"github.com/ollama/ollama/x/create"
+	xcreateclient "github.com/ollama/ollama/x/create/client"
 	"github.com/ollama/ollama/x/imagegen"
-	imagegenclient "github.com/ollama/ollama/x/imagegen/client"
 )
 
 const ConnectInstructions = "To sign in, navigate to:\n    %s\n\n"
@@ -93,15 +94,87 @@ 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(".") {
+			if create.IsTensorModelDir(".") {
 				quantize, _ := cmd.Flags().GetString("quantize")
-				return imagegenclient.CreateModel(args[0], ".", quantize, p)
+				return xcreateclient.CreateModel(xcreateclient.CreateOptions{
+					ModelName: modelName,
+					ModelDir:  ".",
+					Quantize:  quantize,
+				}, p)
 			}
 			reader = strings.NewReader("FROM .\n")
 		} else {
@@ -134,7 +207,7 @@ 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
@@ -527,7 +600,7 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 	}
 
 	// Check if this is an image generation model
-	if slices.Contains(info.Capabilities, model.CapabilityImageGeneration) {
+	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\"")
 		}
@@ -1742,15 +1815,22 @@ 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",
@@ -1905,6 +1985,7 @@ func NewCLI() *cobra.Command {
 	} {
 		switch cmd {
 		case runCmd:
+			imagegen.AppendFlagsDocs(cmd)
 			appendEnvDocs(cmd, []envconfig.EnvVar{envVars["OLLAMA_HOST"], envVars["OLLAMA_NOHISTORY"]})
 		case serveCmd:
 			appendEnvDocs(cmd, []envconfig.EnvVar{

cmd/cmd_test.go

@@ -1555,7 +1555,7 @@ func TestShowInfoImageGen(t *testing.T) {
 			ParameterSize:     "10.3B",
 			QuantizationLevel: "FP8",
 		},
-		Capabilities: []model.Capability{model.CapabilityImageGeneration},
+		Capabilities: []model.Capability{model.CapabilityImage},
 		Requires:     "0.14.0",
 	}, false, &b)
 	if err != nil {

cmd/interactive.go

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

convert/convert.go

@@ -311,6 +311,8 @@ func LoadModelMetadata(fsys fs.FS) (ModelKV, *Tokenizer, error) {
 		conv = &deepseekocr{}
 	case "DeepseekV3ForCausalLM":
 		conv = &deepseek2Model{}
+	case "Glm4MoeLiteForCausalLM":
+		conv = &glm4MoeLiteModel{}
 	default:
 		return nil, nil, fmt.Errorf("unsupported architecture %q", p.Architectures[0])
 	}

convert/convert_glm4moelite.go

@@ -0,0 +1,151 @@
+package convert
+
+import (
+	"cmp"
+	"fmt"
+	"log/slog"
+	"regexp"
+	"strconv"
+
+	"github.com/ollama/ollama/fs/ggml"
+)
+
+type glm4MoeLiteModel struct {
+	ModelParameters
+	MaxPositionEmbeddings uint32  `json:"max_position_embeddings"`
+	HiddenSize            uint32  `json:"hidden_size"`
+	HiddenLayers          uint32  `json:"num_hidden_layers"`
+	IntermediateSize      uint32  `json:"intermediate_size"`
+	NumAttentionHeads     uint32  `json:"num_attention_heads"`
+	NumKeyValueHeads      uint32  `json:"num_key_value_heads"`
+	RMSNormEPS            float32 `json:"rms_norm_eps"`
+
+	RopeTheta     float32 `json:"rope_theta"`
+	QKNopeHeadDim uint32  `json:"qk_nope_head_dim"`
+	QKRopeHeadDim uint32  `json:"qk_rope_head_dim"`
+	KVLoraRank    uint32  `json:"kv_lora_rank"`
+	QLoraRank     uint32  `json:"q_lora_rank"`
+	VHeadDim      uint32  `json:"v_head_dim"`
+
+	ExpertCount            uint32  `json:"n_routed_experts"`
+	ExpertSharedCount      uint32  `json:"n_shared_experts"`
+	ExpertIntermediateSize uint32  `json:"moe_intermediate_size"`
+	ExpertUsedCount        uint32  `json:"num_experts_per_tok"`
+	ExpertWeightsNorm      bool    `json:"norm_topk_prob"`
+	ExpertWeightsScale     float32 `json:"routed_scaling_factor"`
+
+	LeadingDenseBlockCount uint32 `json:"first_k_dense_replace"`
+}
+
+func (p *glm4MoeLiteModel) KV(t *Tokenizer) KV {
+	kv := p.ModelParameters.KV(t)
+	kv["general.architecture"] = "glm4moelite"
+	kv["general.type"] = "model"
+	kv["glm4moelite.block_count"] = p.HiddenLayers
+
+	numHeads := p.NumAttentionHeads
+	numKVHeads := p.NumKeyValueHeads
+
+	kv["glm4moelite.attention.head_count"] = numHeads
+	kv["glm4moelite.attention.head_count_kv"] = numKVHeads
+	kv["glm4moelite.attention.key_length"] = p.QKNopeHeadDim + p.QKRopeHeadDim
+	kv["glm4moelite.attention.kv_lora_rank"] = p.KVLoraRank
+	kv["glm4moelite.attention.layer_norm_rms_epsilon"] = p.RMSNormEPS
+	kv["glm4moelite.attention.q_lora_rank"] = p.QLoraRank
+	kv["glm4moelite.attention.value_length"] = p.VHeadDim
+	kv["glm4moelite.context_length"] = p.MaxPositionEmbeddings
+	kv["glm4moelite.embedding_length"] = p.HiddenSize
+	kv["glm4moelite.expert_count"] = p.ExpertCount
+	kv["glm4moelite.expert_feed_forward_length"] = p.ExpertIntermediateSize
+	kv["glm4moelite.expert_shared_count"] = p.ExpertSharedCount
+
+	// GLM-4.7 MOE Lite uses sigmoid gating
+	kv["glm4moelite.expert_gating_func"] = uint32(2) // sigmoid
+	kv["glm4moelite.expert_used_count"] = p.ExpertUsedCount
+	kv["glm4moelite.expert_weights_norm"] = p.ExpertWeightsNorm
+	kv["glm4moelite.expert_weights_scale"] = p.ExpertWeightsScale
+	kv["glm4moelite.feed_forward_length"] = p.IntermediateSize
+	kv["glm4moelite.leading_dense_block_count"] = p.LeadingDenseBlockCount
+
+	kv["glm4moelite.rope.dimension_count"] = p.QKRopeHeadDim
+	kv["glm4moelite.rope.freq_base"] = cmp.Or(p.RopeTheta, float32(1000000.0))
+
+	kv["tokenizer.ggml.pre"] = "glm4"
+
+	return kv
+}
+
+func (p *glm4MoeLiteModel) Replacements() []string {
+	return []string{
+		"lm_head", "output",
+		"model.embed_tokens", "token_embd",
+		"model.norm", "output_norm",
+		"model.layers", "blk",
+		"input_layernorm", "attn_norm",
+		"self_attn.kv_a_proj_with_mqa", "attn_kv_a_mqa",
+		"self_attn.kv_a_layernorm", "attn_kv_a_norm",
+		"self_attn.kv_b_proj", "attn_kv_b",
+		"self_attn.q_a_proj", "attn_q_a",
+		"self_attn.q_a_layernorm", "attn_q_a_norm",
+		"self_attn.q_b_proj", "attn_q_b",
+		"self_attn.o_proj", "attn_output",
+		"post_attention_layernorm", "ffn_norm",
+		"mlp.shared_experts.down_proj", "ffn_down_shexp",
+		"mlp.shared_experts.gate_proj", "ffn_gate_shexp",
+		"mlp.shared_experts.up_proj", "ffn_up_shexp",
+		"mlp.gate_proj", "ffn_gate",
+		"mlp.down_proj", "ffn_down",
+		"mlp.up_proj", "ffn_up",
+		"mlp.gate.e_score_correction_bias", "exp_probs_b.bias",
+		"mlp.gate", "ffn_gate_inp",
+	}
+}
+
+func (p *glm4MoeLiteModel) Tensors(s []Tensor) (out []*ggml.Tensor) {
+	merges := make([]merge, p.HiddenLayers*3)
+	for i := range p.HiddenLayers {
+		merges[i*3+0] = merge{
+			fmt.Sprintf("blk.%d.mlp.experts.*.gate_proj.weight", i),
+			fmt.Sprintf("blk.%d.ffn_gate_exps.weight", i),
+		}
+		merges[i*3+1] = merge{
+			fmt.Sprintf("blk.%d.mlp.experts.*.up_proj.weight", i),
+			fmt.Sprintf("blk.%d.ffn_up_exps.weight", i),
+		}
+		merges[i*3+2] = merge{
+			fmt.Sprintf("blk.%d.mlp.experts.*.down_proj.weight", i),
+			fmt.Sprintf("blk.%d.ffn_down_exps.weight", i),
+		}
+	}
+
+	skipLayer := func(n string, minValue uint32) bool {
+		re := regexp.MustCompile(`^blk\.(\d+)`)
+		matches := re.FindStringSubmatch(n)
+		if matches == nil {
+			return false
+		}
+
+		blkNum, err := strconv.Atoi(matches[1])
+		if err != nil {
+			return false
+		}
+
+		return uint32(blkNum) >= minValue
+	}
+
+	out, s = mergeTensors(s, merges...)
+	for _, t := range s {
+		// skip any additional layers (such as the Multi-Token Prediction layer)
+		if skipLayer(t.Name(), p.HiddenLayers) {
+			slog.Debug("skipping layer", "name", t.Name())
+			continue
+		}
+		out = append(out, &ggml.Tensor{
+			Name:     t.Name(),
+			Kind:     t.Kind(),
+			Shape:    t.Shape(),
+			WriterTo: t,
+		})
+	}
+	return out
+}

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

@@ -21,6 +21,7 @@ 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_BASE_URL=http://localhost:11434
 export ANTHROPIC_API_KEY=ollama  # required but ignored
 ```
@@ -247,12 +248,13 @@ curl -X POST http://localhost:11434/v1/messages \
 [Claude Code](https://code.claude.com/docs/en/overview) can be configured to use Ollama as its backend:
 
 ```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=ollama claude --model qwen3-coder
 ```
 
 Or set the environment variables in your shell profile:
 
 ```shell
+export ANTHROPIC_AUTH_TOKEN=ollama
 export ANTHROPIC_BASE_URL=http://localhost:11434
 export ANTHROPIC_API_KEY=ollama
 ```

docs/api/openai-compatibility.mdx

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

docs/capabilities/web-search.mdx

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

docs/docs.json

@@ -111,7 +111,9 @@
               "/integrations/zed",
               "/integrations/roo-code",
               "/integrations/n8n",
-              "/integrations/xcode"
+              "/integrations/xcode",
+              "/integrations/onyx",
+              "/integrations/marimo"
             ]
           },
           {

docs/faq.mdx

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

docs/integrations/claude-code.mdx

@@ -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 `qwen3-coder`, `gpt-oss:20b`, or other models.
+
+![Claude Code with Ollama](https://files.ollama.com/claude-code.png)
+
 ## Install
 
 Install [Claude Code](https://code.claude.com/docs/en/overview):
@@ -25,22 +31,24 @@ Claude Code connects to Ollama using the Anthropic-compatible API.
 1. Set the environment variables:
 
 ```shell
+export ANTHROPIC_AUTH_TOKEN=ollama
 export ANTHROPIC_BASE_URL=http://localhost:11434
-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 claude --model gpt-oss:20b
 ```
 
+**Note:** Claude Code requires a large context window. We recommend at least 32K tokens. See the [context length documentation](/context-length) for how to adjust context length in Ollama.
+
 ## Connecting to ollama.com
 
 1. Create an [API key](https://ollama.com/settings/keys) on ollama.com
@@ -67,3 +75,4 @@ claude --model glm-4.7:cloud
 ### Local models
 - `qwen3-coder` - Excellent for coding tasks
 - `gpt-oss:20b` - Strong general-purpose model
+- `gpt-oss:120b` - Larger general-purpose model for more complex tasks

docs/integrations/marimo.mdx

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

docs/integrations/onyx.mdx

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

docs/linux.mdx

@@ -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
-```
+```

fs/ggml/ggml.go

@@ -269,6 +269,7 @@ func (kv KV) OllamaEngineRequired() bool {
 		"qwen25vl",
 		"qwen3", "qwen3moe",
 		"qwen3vl", "qwen3vlmoe",
+		"glm4moelite",
 	}, kv.Architecture())
 }
 
@@ -856,6 +857,7 @@ func (f GGML) FlashAttention() bool {
 	return slices.Contains([]string{
 		"bert",
 		"gemma3",
+		"glm4moelite",
 		"gptoss", "gpt-oss",
 		"mistral3",
 		"olmo3",

integration/tools_test.go

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

llm/server.go

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

middleware/openai.go

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

middleware/openai_test.go

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

model/models/glm4moelite/model.go

@@ -0,0 +1,308 @@
+package glm4moelite
+
+import (
+	"math"
+
+	"github.com/ollama/ollama/fs"
+	"github.com/ollama/ollama/kvcache"
+	"github.com/ollama/ollama/ml"
+	"github.com/ollama/ollama/ml/nn"
+	"github.com/ollama/ollama/model"
+	"github.com/ollama/ollama/model/input"
+)
+
+type Options struct {
+	numExpertsUsed      int
+	numExperts          int
+	normTopKProb        bool
+	routedScalingFactor float32
+
+	kvLoraRank,
+	qkNopeHeadDim,
+	qkRopeHeadDim,
+	kqNopeHeadDim,
+	qkHeadDim int
+	qLoraRank int
+	vHeadDim  int
+
+	hiddenSize,
+	numHeads,
+	numKVHeads int
+
+	eps,
+	ropeBase float32
+	kqScale float64
+}
+
+func (o Options) applyRotaryPositionEmbeddings(ctx ml.Context, t, p ml.Tensor) ml.Tensor {
+	// Standard RoPE without YARN scaling
+	return nn.RoPE(ctx, t, p, o.qkRopeHeadDim, o.ropeBase, 1.0)
+}
+
+type Attention struct {
+	Q *nn.Linear `gguf:"attn_q"`
+
+	QA     *nn.Linear  `gguf:"attn_q_a"`
+	QANorm *nn.RMSNorm `gguf:"attn_q_a_norm"`
+	QB     *nn.Linear  `gguf:"attn_q_b"`
+
+	KVA     *nn.Linear  `gguf:"attn_kv_a_mqa"`
+	KVANorm *nn.RMSNorm `gguf:"attn_kv_a_norm"`
+	KVB     *nn.Linear  `gguf:"attn_kv_b"`
+
+	Output *nn.Linear `gguf:"attn_out,alt:attn_output"`
+}
+
+func (attn *Attention) Forward(ctx ml.Context, hiddenStates, positions ml.Tensor, cache kvcache.Cache, opts *Options) ml.Tensor {
+	seqLength := hiddenStates.Dim(1)
+
+	var query ml.Tensor
+	if opts.qLoraRank == 0 {
+		query = attn.Q.Forward(ctx, hiddenStates)
+	} else {
+		query = attn.QA.Forward(ctx, hiddenStates)
+		query = attn.QANorm.Forward(ctx, query, opts.eps)
+		query = attn.QB.Forward(ctx, query)
+	}
+
+	query = query.Reshape(ctx, query.Dim(0)/opts.numHeads, opts.numHeads, seqLength)
+	queryChunks := query.ChunkSections(ctx, 0, opts.qkNopeHeadDim, opts.qkRopeHeadDim)
+
+	compressedKV := attn.KVA.Forward(ctx, hiddenStates)
+	kPass := compressedKV.Slice(ctx, 0, 0, opts.kvLoraRank, 1)
+	kRot := compressedKV.View(ctx,
+		opts.kvLoraRank*compressedKV.Stride(0), opts.qkRopeHeadDim,
+		compressedKV.Stride(1), 1,
+		compressedKV.Stride(1), compressedKV.Dim(1),
+	)
+
+	qRot := opts.applyRotaryPositionEmbeddings(ctx, queryChunks[1], positions)
+	kRot = opts.applyRotaryPositionEmbeddings(ctx, kRot, positions)
+	kPass = attn.KVANorm.Forward(ctx, kPass, opts.eps)
+
+	// GLM-4.7 MOE Lite uses v3-style MLA (no v3.1 absorbed mode)
+	kPass = attn.KVB.Forward(ctx, kPass)
+
+	kv := kPass.Reshape(ctx, kPass.Dim(0)/opts.numKVHeads, opts.numKVHeads, seqLength)
+	kvChunks := kv.ChunkSections(ctx, 0, opts.kqNopeHeadDim, opts.vHeadDim)
+
+	kRot = kRot.Repeat(ctx, 1, queryChunks[0].Dim(1))
+	query = qRot.Concat(ctx, queryChunks[0], 0)
+	key := kRot.Concat(ctx, kvChunks[0], 0)
+	attention := nn.Attention(ctx, query, key, kvChunks[1], opts.kqScale, cache)
+
+	attention = attention.Reshape(ctx, attention.Dim(0)*attention.Dim(1), seqLength)
+	return attn.Output.Forward(ctx, attention)
+}
+
+type MLP interface {
+	Forward(ml.Context, ml.Tensor, *Options) ml.Tensor
+}
+
+type sparse struct {
+	Router       *nn.Linear `gguf:"ffn_gate_inp"`
+	Gate         *nn.Linear `gguf:"ffn_gate_exps"`
+	Up           *nn.Linear `gguf:"ffn_up_exps"`
+	Down         *nn.Linear `gguf:"ffn_down_exps"`
+	SharedExpert *dense     `gguf:",suf:_shexp"`
+	ExpProbsBias ml.Tensor  `gguf:"exp_probs_b.bias,alt:exp_probs_b"`
+}
+
+func (moe *sparse) Moe(ctx ml.Context, hiddenStates, topKIndices, topKWeights ml.Tensor, opts *Options) ml.Tensor {
+	hiddenStates = hiddenStates.Reshape(ctx, hiddenStates.Dim(0), 1, hiddenStates.Dim(1))
+
+	upStates := moe.Up.Weight.MulmatID(ctx, hiddenStates, topKIndices)
+	hiddenStates = moe.Gate.Weight.MulmatID(ctx, hiddenStates, topKIndices)
+	hiddenStates = hiddenStates.SILU(ctx, upStates)
+
+	experts := moe.Down.Weight.MulmatID(ctx, hiddenStates, topKIndices)
+	experts = experts.Mul(ctx, topKWeights)
+
+	nextStates := experts.View(ctx, 0, experts.Dim(0), experts.Stride(2), experts.Dim(2))
+	for i := 1; i < opts.numExpertsUsed; i++ {
+		nextStates = nextStates.Add(ctx, experts.View(ctx, i*experts.Stride(1), experts.Dim(0), experts.Stride(2), experts.Dim(2)))
+	}
+	return nextStates
+}
+
+func (moe *sparse) topKIndices(ctx ml.Context, scores ml.Tensor, opts *Options) ml.Tensor {
+	if moe.ExpProbsBias != nil {
+		scores = scores.Add(ctx, moe.ExpProbsBias)
+	}
+	topKIndices := scores.TopK(ctx, opts.numExpertsUsed)
+	return topKIndices
+}
+
+func (moe *sparse) Forward(ctx ml.Context, hiddenStates ml.Tensor, opts *Options) ml.Tensor {
+	residuals := hiddenStates
+
+	routerLogits := moe.Router.Forward(ctx, hiddenStates)
+	scores := routerLogits.Sigmoid(ctx)
+	topKIndices := moe.topKIndices(ctx, scores, opts)
+	topKWeights := scores.Reshape(ctx, 1, opts.numExperts, hiddenStates.Dim(1)).Rows(ctx, topKIndices)
+
+	if opts.normTopKProb {
+		topKWeights = topKWeights.Reshape(ctx, opts.numExpertsUsed, hiddenStates.Dim(1))
+		topKWeights = topKWeights.Div(ctx, topKWeights.SumRows(ctx))
+		topKWeights = topKWeights.Reshape(ctx, 1, opts.numExpertsUsed, hiddenStates.Dim(1))
+	}
+
+	topKWeights = topKWeights.Scale(ctx, float64(opts.routedScalingFactor))
+	hiddenStates = moe.Moe(ctx, hiddenStates, topKIndices, topKWeights, opts)
+	sharedExpertResult := moe.SharedExpert.Forward(ctx, residuals, opts)
+
+	hiddenStates = hiddenStates.Add(ctx, sharedExpertResult)
+	return hiddenStates
+}
+
+type dense struct {
+	Gate *nn.Linear `gguf:"ffn_gate"`
+	Up   *nn.Linear `gguf:"ffn_up"`
+	Down *nn.Linear `gguf:"ffn_down"`
+}
+
+func (mlp *dense) Forward(ctx ml.Context, hiddenStates ml.Tensor, opts *Options) ml.Tensor {
+	hiddenStates = mlp.Gate.Forward(ctx, hiddenStates).SILU(ctx, mlp.Up.Forward(ctx, hiddenStates))
+	return mlp.Down.Forward(ctx, hiddenStates)
+}
+
+type Layer struct {
+	AttentionNorm *nn.RMSNorm `gguf:"attn_norm"`
+	Attention     *Attention
+
+	MLPNorm *nn.RMSNorm `gguf:"ffn_norm"`
+	MLP     MLP
+}
+
+func (t *Layer) Forward(ctx ml.Context, hiddenStates, positions, outputs ml.Tensor, cache kvcache.Cache, opts *Options) ml.Tensor {
+	residual := hiddenStates
+	hiddenStates = t.AttentionNorm.Forward(ctx, hiddenStates, opts.eps)
+	hiddenStates = t.Attention.Forward(ctx, hiddenStates, positions, cache, opts)
+
+	if outputs != nil {
+		hiddenStates = hiddenStates.Rows(ctx, outputs)
+		residual = residual.Rows(ctx, outputs)
+	}
+
+	hiddenStates = hiddenStates.Add(ctx, residual)
+	residual = hiddenStates
+
+	hiddenStates = t.MLPNorm.Forward(ctx, hiddenStates, opts.eps)
+	hiddenStates = t.MLP.Forward(ctx, hiddenStates, opts)
+	hiddenStates = hiddenStates.Add(ctx, residual)
+	return hiddenStates
+}
+
+type Model struct {
+	model.Base
+	model.BytePairEncoding
+
+	TokenEmbedding *nn.Embedding `gguf:"token_embd"`
+	Layers         []Layer       `gguf:"blk"`
+
+	OutputNorm *nn.RMSNorm `gguf:"output_norm"`
+	Output     *nn.Linear  `gguf:"output,alt:token_embd"`
+
+	*Options
+}
+
+func New(c fs.Config) (model.Model, error) {
+	layers := make([]Layer, c.Uint("block_count"))
+
+	firstDenseLayerIndex := int(c.Uint("leading_dense_block_count"))
+	for i := range layers {
+		if i < firstDenseLayerIndex {
+			layers[i].MLP = &dense{}
+		} else {
+			layers[i].MLP = &sparse{}
+		}
+	}
+
+	keyLength := int(c.Uint("attention.key_length"))
+	valueLength := int(c.Uint("attention.value_length"))
+
+	// Simple kqScale for GLM-4.7 MOE Lite (no mScale factor like DeepSeek)
+	kqScale := 1.0 / math.Sqrt(float64(keyLength))
+
+	var pre []string
+	switch c.String("tokenizer.ggml.pre") {
+	case "glm4":
+		pre = []string{
+			`(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\r\n\p{L}\p{N}]?\p{L}+|\p{N}{1,3}| ?[^\s\p{L}\p{N}]+[\r\n]*|\s*[\r\n]+|\s+(?!\S)|\s+`,
+		}
+	default:
+		return nil, model.ErrUnsupportedTokenizer
+	}
+
+	m := Model{
+		BytePairEncoding: model.NewBytePairEncoding(
+			&model.Vocabulary{
+				Values: c.Strings("tokenizer.ggml.tokens"),
+				Types:  c.Ints("tokenizer.ggml.token_type"),
+				Merges: c.Strings("tokenizer.ggml.merges"),
+				AddBOS: c.Bool("tokenizer.ggml.add_bos_token", true),
+				BOS:    []int32{int32(c.Uint("tokenizer.ggml.bos_token_id"))},
+				AddEOS: c.Bool("tokenizer.ggml.add_eos_token", false),
+				EOS: append(
+					[]int32{int32(c.Uint("tokenizer.ggml.eos_token_id"))},
+					c.Ints("tokenizer.ggml.eos_token_ids")...,
+				),
+			},
+			pre...,
+		),
+		Layers: layers,
+		Options: &Options{
+			hiddenSize:     int(c.Uint("embedding_length")),
+			numHeads:       int(c.Uint("attention.head_count")),
+			numKVHeads:     int(c.Uint("attention.head_count_kv")),
+			eps:            c.Float("attention.layer_norm_rms_epsilon"),
+			ropeBase:       c.Float("rope.freq_base"),
+			numExperts:     int(c.Uint("expert_count")),
+			numExpertsUsed: int(c.Uint("expert_used_count")),
+			normTopKProb:   c.Bool("expert_weights_norm", true),
+
+			qLoraRank:     int(c.Uint("attention.q_lora_rank")),
+			kvLoraRank:    int(c.Uint("attention.kv_lora_rank")),
+			qkHeadDim:     keyLength,
+			vHeadDim:      valueLength,
+			qkRopeHeadDim: int(c.Uint("rope.dimension_count")),
+			qkNopeHeadDim: keyLength - int(c.Uint("rope.dimension_count")),
+			kqNopeHeadDim: keyLength - int(c.Uint("rope.dimension_count")),
+
+			routedScalingFactor: c.Float("expert_weights_scale"),
+
+			kqScale: kqScale,
+		},
+	}
+
+	m.Cache = kvcache.NewCausalCache(m.Shift)
+	return &m, nil
+}
+
+func (m Model) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) {
+	return m.applyRotaryPositionEmbeddings(ctx, key, shift), nil
+}
+
+func (m *Model) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
+	positions := ctx.Input().FromInts(batch.Positions, len(batch.Positions))
+
+	hiddenStates := m.TokenEmbedding.Forward(ctx, batch.Inputs)
+
+	for i, layer := range m.Layers {
+		m.Cache.SetLayer(i)
+
+		var outputs ml.Tensor
+		if i == len(m.Layers)-1 {
+			outputs = batch.Outputs
+		}
+
+		hiddenStates = layer.Forward(ctx, hiddenStates, positions, outputs, m.Cache, m.Options)
+	}
+
+	hiddenStates = m.OutputNorm.Forward(ctx, hiddenStates, m.eps)
+	return m.Output.Forward(ctx, hiddenStates), nil
+}
+
+func init() {
+	model.Register("glm4moelite", New)
+}

model/models/models.go

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

model/parsers/glm46.go

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

model/parsers/glm46_test.go

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

model/parsers/glm47.go

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

model/parsers/glm47_test.go

@@ -0,0 +1,101 @@
+package parsers
+
+import (
+	"reflect"
+	"testing"
+
+	"github.com/ollama/ollama/api"
+)
+
+// Edge cases covered: assistant output with thinking enabled, typed argument coercion,
+// tool-call content containing XML-sensitive characters, and thinking disabled mode.
+func TestGLM47ParserAdd(t *testing.T) {
+	parser := GLM47Parser{}
+	parser.Init([]api.Tool{
+		tool("calculate", map[string]api.ToolProperty{
+			"count":   {Type: api.PropertyType{"integer"}},
+			"enabled": {Type: api.PropertyType{"boolean"}},
+		}),
+	}, nil, nil)
+
+	// When thinking is enabled (thinkValue nil), the prompt ends with <think>,
+	// so the model output does NOT include the opening <think> tag.
+	content, thinking, calls, err := parser.Add("plan</think>Answer<tool_call>calculate<arg_key>count</arg_key><arg_value>3</arg_value><arg_key>enabled</arg_key><arg_value>true</arg_value></tool_call>", true)
+	if err != nil {
+		t.Fatalf("parse failed: %v", err)
+	}
+	if thinking != "plan" {
+		t.Fatalf("expected thinking 'plan', got %q", thinking)
+	}
+	if content != "Answer" {
+		t.Fatalf("expected content 'Answer', got %q", content)
+	}
+	if len(calls) != 1 {
+		t.Fatalf("expected 1 tool call, got %d", len(calls))
+	}
+	expectedArgs := args(`{"count": 3, "enabled": true}`)
+	if !toolCallEqual(api.ToolCall{Function: api.ToolCallFunction{Arguments: calls[0].Function.Arguments}}, api.ToolCall{Function: api.ToolCallFunction{Arguments: expectedArgs}}) {
+		t.Fatalf("expected args %#v, got %#v", expectedArgs.ToMap(), calls[0].Function.Arguments.ToMap())
+	}
+}
+
+func TestGLM47ParserNoThinkingContent(t *testing.T) {
+	parser := GLM47Parser{}
+	parser.Init(nil, nil, nil)
+
+	// When thinking is enabled but model has no thinking to output,
+	// it should output </think> immediately followed by content.
+	content, thinking, calls, err := parser.Add("</think>Plain answer", true)
+	if err != nil {
+		t.Fatalf("parse failed: %v", err)
+	}
+	if thinking != "" {
+		t.Fatalf("expected empty thinking, got %q", thinking)
+	}
+	if content != "Plain answer" {
+		t.Fatalf("expected content 'Plain answer', got %q", content)
+	}
+	if len(calls) != 0 {
+		t.Fatalf("expected no tool calls, got %d", len(calls))
+	}
+}
+
+func TestGLM47ParserThinkingDisabled(t *testing.T) {
+	parser := GLM47Parser{}
+	// When thinking is disabled, parser stays in LookingForThinkingOpen state
+	parser.Init(nil, nil, &api.ThinkValue{Value: false})
+
+	// Model outputs plain content (prompt ended with </think>)
+	content, thinking, calls, err := parser.Add("Plain answer", true)
+	if err != nil {
+		t.Fatalf("parse failed: %v", err)
+	}
+	if thinking != "" {
+		t.Fatalf("expected empty thinking, got %q", thinking)
+	}
+	if content != "Plain answer" {
+		t.Fatalf("expected content 'Plain answer', got %q", content)
+	}
+	if len(calls) != 0 {
+		t.Fatalf("expected no tool calls, got %d", len(calls))
+	}
+}
+
+func TestGLM47ParserToolCallEscaping(t *testing.T) {
+	toolCall, err := parseGLM46ToolCall(glm46EventRawToolCall{raw: `exec
+<arg_key>expr</arg_key>
+<arg_value>a < b && c > d</arg_value>`}, nil)
+	if err != nil {
+		t.Fatalf("parse failed: %v", err)
+	}
+
+	expected := api.ToolCall{
+		Function: api.ToolCallFunction{
+			Name:      "exec",
+			Arguments: args(`{"expr": "a < b && c > d"}`),
+		},
+	}
+	if !reflect.DeepEqual(toolCall, expected) {
+		t.Fatalf("expected %#v, got %#v", expected, toolCall)
+	}
+}

model/parsers/nemotron3nano.go

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

model/parsers/nemotron3nano_test.go

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

model/parsers/parsers.go

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

model/parsers/qwen3coder_test.go

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

model/parsers/testhelpers_test.go

@@ -96,3 +96,19 @@ func testArgs(m map[string]any) api.ToolCallFunctionArguments {
 	}
 	return args
 }
+
+func args(s string) api.ToolCallFunctionArguments {
+	var result api.ToolCallFunctionArguments
+	if err := json.Unmarshal([]byte(s), &result); err != nil {
+		panic("invalid JSON in args(): " + err.Error())
+	}
+	return result
+}
+
+func propsMap(s string) *api.ToolPropertiesMap {
+	var result api.ToolPropertiesMap
+	if err := json.Unmarshal([]byte(s), &result); err != nil {
+		panic("invalid JSON in propsMap(): " + err.Error())
+	}
+	return &result
+}

model/renderers/glm46.go

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

model/renderers/glm46_test.go

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

model/renderers/glm47.go

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

model/renderers/glm47_test.go

@@ -0,0 +1,205 @@
+package renderers
+
+import (
+	"testing"
+
+	"github.com/google/go-cmp/cmp"
+	"github.com/ollama/ollama/api"
+)
+
+// Test cases validated against zai-org/GLM-4.7 chat template from HuggingFace.
+// Edge cases covered: tool JSON spacing, think disabling, tool calls without content,
+// sorted tool argument keys, multiple tool calls, and multiple tool responses under
+// a single observation tag.
+
+func TestGLM47Renderer(t *testing.T) {
+	tests := []struct {
+		name       string
+		messages   []api.Message
+		tools      []api.Tool
+		thinkValue *api.ThinkValue
+		expected   string
+	}{
+		// Validated against: tokenizer.apply_chat_template([{"role": "user", "content": "Hello"}], add_generation_prompt=True, tokenize=False, enable_thinking=True)
+		{
+			name: "basic user message",
+			messages: []api.Message{
+				{Role: "user", Content: "Hello"},
+			},
+			expected: "[gMASK]<sop><|user|>Hello<|assistant|><think>",
+		},
+		// Validated against: tokenizer.apply_chat_template([{"role": "user", "content": "Hello"}], add_generation_prompt=True, tokenize=False, enable_thinking=False)
+		{
+			name: "thinking disabled",
+			messages: []api.Message{
+				{Role: "user", Content: "Hello"},
+			},
+			thinkValue: &api.ThinkValue{Value: false},
+			expected:   "[gMASK]<sop><|user|>Hello<|assistant|></think>",
+		},
+		// Validated against: tokenizer.apply_chat_template([{"role": "system", "content": "You are helpful."}, {"role": "user", "content": "Hello"}], ...)
+		{
+			name: "system and user",
+			messages: []api.Message{
+				{Role: "system", Content: "You are helpful."},
+				{Role: "user", Content: "Hello"},
+			},
+			expected: "[gMASK]<sop><|system|>You are helpful.<|user|>Hello<|assistant|><think>",
+		},
+		// Validated against multi-turn conversation
+		{
+			name: "multi-turn conversation",
+			messages: []api.Message{
+				{Role: "user", Content: "Hi"},
+				{Role: "assistant", Content: "Hello there"},
+				{Role: "user", Content: "How are you?"},
+			},
+			expected: "[gMASK]<sop><|user|>Hi<|assistant|></think>Hello there<|user|>How are you?<|assistant|><think>",
+		},
+		// Validated against: message with reasoning_content field
+		{
+			name: "assistant with reasoning_content",
+			messages: []api.Message{
+				{Role: "user", Content: "Answer with reasoning."},
+				{Role: "assistant", Thinking: "Plan.", Content: "Done."},
+			},
+			expected: "[gMASK]<sop><|user|>Answer with reasoning.<|assistant|><think>Plan.</think>Done.<|assistant|><think>",
+		},
+		// Validated against: tool call with empty content - no "None" should be output
+		{
+			name: "tool call with empty content",
+			messages: []api.Message{
+				{Role: "user", Content: "Weather?"},
+				{
+					Role: "assistant",
+					ToolCalls: []api.ToolCall{
+						{
+							Function: api.ToolCallFunction{
+								Name:      "get_weather",
+								Arguments: args(`{"location": "Tokyo", "unit": "celsius"}`),
+							},
+						},
+					},
+				},
+				{Role: "tool", Content: `{"temperature":22}`},
+			},
+			tools: []api.Tool{
+				{
+					Type: "function",
+					Function: api.ToolFunction{
+						Name:        "get_weather",
+						Description: "Get weather",
+						Parameters: api.ToolFunctionParameters{
+							Type:       "object",
+							Required:   []string{"location"},
+							Properties: propsMap(`{"location": {"type": "string"}}`),
+						},
+					},
+				},
+			},
+			expected: "[gMASK]<sop><|system|>\n# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>\n{\"type\": \"function\", \"function\": {\"name\": \"get_weather\", \"description\": \"Get weather\", \"parameters\": {\"type\": \"object\", \"required\": [\"location\"], \"properties\": {\"location\": {\"type\": \"string\"}}}}}\n</tools>\n\nFor each function call, output the function name and arguments within the following XML format:\n<tool_call>{function-name}<arg_key>{arg-key-1}</arg_key><arg_value>{arg-value-1}</arg_value><arg_key>{arg-key-2}</arg_key><arg_value>{arg-value-2}</arg_value>...</tool_call><|user|>Weather?<|assistant|></think><tool_call>get_weather<arg_key>location</arg_key><arg_value>Tokyo</arg_value><arg_key>unit</arg_key><arg_value>celsius</arg_value></tool_call><|observation|><tool_response>{\"temperature\":22}</tool_response><|assistant|><think>",
+		},
+		// Validated against: tool call with content before tool call
+		{
+			name: "tool call with content",
+			messages: []api.Message{
+				{Role: "user", Content: "Weather?"},
+				{
+					Role:    "assistant",
+					Content: "Let me check",
+					ToolCalls: []api.ToolCall{
+						{
+							Function: api.ToolCallFunction{
+								Name:      "get_weather",
+								Arguments: args(`{"location": "Tokyo"}`),
+							},
+						},
+					},
+				},
+				{Role: "tool", Content: `{"temperature":22}`},
+				{Role: "assistant", Content: "It is 22C."},
+			},
+			tools: []api.Tool{
+				{
+					Type: "function",
+					Function: api.ToolFunction{
+						Name:        "get_weather",
+						Description: "Get weather",
+						Parameters: api.ToolFunctionParameters{
+							Type:       "object",
+							Required:   []string{"location"},
+							Properties: propsMap(`{"location": {"type": "string"}}`),
+						},
+					},
+				},
+			},
+			expected: "[gMASK]<sop><|system|>\n# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>\n{\"type\": \"function\", \"function\": {\"name\": \"get_weather\", \"description\": \"Get weather\", \"parameters\": {\"type\": \"object\", \"required\": [\"location\"], \"properties\": {\"location\": {\"type\": \"string\"}}}}}\n</tools>\n\nFor each function call, output the function name and arguments within the following XML format:\n<tool_call>{function-name}<arg_key>{arg-key-1}</arg_key><arg_value>{arg-value-1}</arg_value><arg_key>{arg-key-2}</arg_key><arg_value>{arg-value-2}</arg_value>...</tool_call><|user|>Weather?<|assistant|></think>Let me check<tool_call>get_weather<arg_key>location</arg_key><arg_value>Tokyo</arg_value></tool_call><|observation|><tool_response>{\"temperature\":22}</tool_response><|assistant|></think>It is 22C.<|assistant|><think>",
+		},
+		// Validated against: multiple tool calls and multiple consecutive tool responses
+		{
+			name: "multiple tool calls and responses",
+			messages: []api.Message{
+				{Role: "user", Content: "Compare weather"},
+				{
+					Role: "assistant",
+					ToolCalls: []api.ToolCall{
+						{
+							Function: api.ToolCallFunction{
+								Name:      "get_weather",
+								Arguments: args(`{"location": "Tokyo"}`),
+							},
+						},
+						{
+							Function: api.ToolCallFunction{
+								Name:      "get_weather",
+								Arguments: args(`{"location": "Paris"}`),
+							},
+						},
+					},
+				},
+				{Role: "tool", Content: `{"temperature":22}`},
+				{Role: "tool", Content: `{"temperature":18}`},
+			},
+			tools: []api.Tool{
+				{
+					Type: "function",
+					Function: api.ToolFunction{
+						Name:        "get_weather",
+						Description: "Get weather",
+						Parameters: api.ToolFunctionParameters{
+							Type:       "object",
+							Required:   []string{"location"},
+							Properties: propsMap(`{"location": {"type": "string"}}`),
+						},
+					},
+				},
+			},
+			expected: "[gMASK]<sop><|system|>\n# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>\n{\"type\": \"function\", \"function\": {\"name\": \"get_weather\", \"description\": \"Get weather\", \"parameters\": {\"type\": \"object\", \"required\": [\"location\"], \"properties\": {\"location\": {\"type\": \"string\"}}}}}\n</tools>\n\nFor each function call, output the function name and arguments within the following XML format:\n<tool_call>{function-name}<arg_key>{arg-key-1}</arg_key><arg_value>{arg-value-1}</arg_value><arg_key>{arg-key-2}</arg_key><arg_value>{arg-value-2}</arg_value>...</tool_call><|user|>Compare weather<|assistant|></think><tool_call>get_weather<arg_key>location</arg_key><arg_value>Tokyo</arg_value></tool_call><tool_call>get_weather<arg_key>location</arg_key><arg_value>Paris</arg_value></tool_call><|observation|><tool_response>{\"temperature\":22}</tool_response><tool_response>{\"temperature\":18}</tool_response><|assistant|><think>",
+		},
+		// Validated against: preserved thinking (clear_thinking=false) - reasoning_content is included
+		{
+			name: "preserved thinking in multi-turn",
+			messages: []api.Message{
+				{Role: "user", Content: "Think step by step"},
+				{Role: "assistant", Thinking: "Let me think...", Content: "Here's my answer."},
+				{Role: "user", Content: "Continue"},
+			},
+			expected: "[gMASK]<sop><|user|>Think step by step<|assistant|><think>Let me think...</think>Here's my answer.<|user|>Continue<|assistant|><think>",
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			renderer := &GLM47Renderer{}
+			rendered, err := renderer.Render(tt.messages, tt.tools, tt.thinkValue)
+			if err != nil {
+				t.Fatal(err)
+			}
+			if diff := cmp.Diff(rendered, tt.expected); diff != "" {
+				t.Errorf("mismatch (-got +want):\n%s", diff)
+				t.Logf("Got:\n%s", rendered)
+				t.Logf("Expected:\n%s", tt.expected)
+			}
+		})
+	}
+}

model/renderers/renderer.go

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

model/renderers/testhelpers_test.go

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

openai/openai.go

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

openai/responses.go

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

openai/responses_test.go

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

readline/readline.go

@@ -5,6 +5,7 @@ import (
 	"fmt"
 	"io"
 	"os"
+	"strings"
 )
 
 type Prompt struct {
@@ -36,10 +37,11 @@ type Terminal struct {
 }
 
 type Instance struct {
-	Prompt   *Prompt
-	Terminal *Terminal
-	History  *History
-	Pasting  bool
+	Prompt      *Prompt
+	Terminal    *Terminal
+	History     *History
+	Pasting     bool
+	pastedLines []string
 }
 
 func New(prompt Prompt) (*Instance, error) {
@@ -174,6 +176,8 @@ func (i *Instance) Readline() (string, error) {
 		case CharEsc:
 			esc = true
 		case CharInterrupt:
+			i.pastedLines = nil
+			i.Prompt.UseAlt = false
 			return "", ErrInterrupt
 		case CharPrev:
 			i.historyPrev(buf, &currentLineBuf)
@@ -188,7 +192,23 @@ func (i *Instance) Readline() (string, error) {
 		case CharForward:
 			buf.MoveRight()
 		case CharBackspace, CharCtrlH:
-			buf.Remove()
+			if buf.IsEmpty() && len(i.pastedLines) > 0 {
+				lastIdx := len(i.pastedLines) - 1
+				prevLine := i.pastedLines[lastIdx]
+				i.pastedLines = i.pastedLines[:lastIdx]
+				fmt.Print(CursorBOL + ClearToEOL + CursorUp + CursorBOL + ClearToEOL)
+				if len(i.pastedLines) == 0 {
+					fmt.Print(i.Prompt.Prompt)
+					i.Prompt.UseAlt = false
+				} else {
+					fmt.Print(i.Prompt.AltPrompt)
+				}
+				for _, r := range prevLine {
+					buf.Add(r)
+				}
+			} else {
+				buf.Remove()
+			}
 		case CharTab:
 			// todo: convert back to real tabs
 			for range 8 {
@@ -211,13 +231,28 @@ func (i *Instance) Readline() (string, error) {
 		case CharCtrlZ:
 			fd := os.Stdin.Fd()
 			return handleCharCtrlZ(fd, i.Terminal.termios)
-		case CharEnter, CharCtrlJ:
+		case CharCtrlJ:
+			i.pastedLines = append(i.pastedLines, buf.String())
+			buf.Buf.Clear()
+			buf.Pos = 0
+			buf.DisplayPos = 0
+			buf.LineHasSpace.Clear()
+			fmt.Println()
+			fmt.Print(i.Prompt.AltPrompt)
+			i.Prompt.UseAlt = true
+			continue
+		case CharEnter:
 			output := buf.String()
+			if len(i.pastedLines) > 0 {
+				output = strings.Join(i.pastedLines, "\n") + "\n" + output
+				i.pastedLines = nil
+			}
 			if output != "" {
 				i.History.Add(output)
 			}
 			buf.MoveToEnd()
 			fmt.Println()
+			i.Prompt.UseAlt = false
 
 			return output, nil
 		default:

scripts/build_darwin.sh

@@ -60,7 +60,7 @@ _build_darwin() {
             cmake --install $BUILD_DIR --component MLX
             # Override CGO flags to point to the amd64 build directory
             MLX_CGO_CFLAGS="-O3 -I$(pwd)/$BUILD_DIR/_deps/mlx-c-src -mmacosx-version-min=14.0"
-            MLX_CGO_LDFLAGS="-L$(pwd)/$BUILD_DIR/lib/ollama -lmlxc -lmlx -Wl,-rpath,@executable_path -lc++ -framework Accelerate -mmacosx-version-min=14.0"
+            MLX_CGO_LDFLAGS="-ldl -lc++ -framework Accelerate -mmacosx-version-min=14.0"
         else
             BUILD_DIR=build
             cmake --preset MLX \
@@ -71,10 +71,12 @@ _build_darwin() {
             cmake --install $BUILD_DIR --component MLX
             # Use default CGO flags from mlx.go for arm64
             MLX_CGO_CFLAGS="-O3 -I$(pwd)/$BUILD_DIR/_deps/mlx-c-src -mmacosx-version-min=14.0"
-            MLX_CGO_LDFLAGS="-L$(pwd)/$BUILD_DIR/lib/ollama -lmlxc -lmlx -Wl,-rpath,@executable_path -lc++ -framework Metal -framework Foundation -framework Accelerate -mmacosx-version-min=14.0"
+            MLX_CGO_LDFLAGS="-lc++ -framework Metal -framework Foundation -framework Accelerate -mmacosx-version-min=14.0"
         fi
-        GOOS=darwin GOARCH=$ARCH CGO_ENABLED=1 CGO_CFLAGS="$MLX_CGO_CFLAGS" CGO_LDFLAGS="$MLX_CGO_LDFLAGS" go build -tags mlx -o $INSTALL_PREFIX/ollama-mlx .
-        GOOS=darwin GOARCH=$ARCH CGO_ENABLED=1 go build -o $INSTALL_PREFIX .
+        GOOS=darwin GOARCH=$ARCH CGO_ENABLED=1 CGO_CFLAGS="$MLX_CGO_CFLAGS" CGO_LDFLAGS="$MLX_CGO_LDFLAGS" go build -tags mlx -o $INSTALL_PREFIX .
+        # Copy MLX libraries to same directory as executable for dlopen
+        cp $INSTALL_PREFIX/lib/ollama/libmlxc.dylib $INSTALL_PREFIX/
+        cp $INSTALL_PREFIX/lib/ollama/libmlx.dylib $INSTALL_PREFIX/
     done
 }
 
@@ -82,12 +84,10 @@ _sign_darwin() {
     status "Creating universal binary..."
     mkdir -p dist/darwin
     lipo -create -output dist/darwin/ollama dist/darwin-*/ollama
-    lipo -create -output dist/darwin/ollama-mlx dist/darwin-*/ollama-mlx
     chmod +x dist/darwin/ollama
-    chmod +x dist/darwin/ollama-mlx
 
     if [ -n "$APPLE_IDENTITY" ]; then
-        for F in dist/darwin/ollama dist/darwin-*/lib/ollama/* dist/darwin/ollama-mlx; do
+        for F in dist/darwin/ollama dist/darwin-*/lib/ollama/*; do
             codesign -f --timestamp -s "$APPLE_IDENTITY" --identifier ai.ollama.ollama --options=runtime $F
         done
 
@@ -154,7 +154,6 @@ _build_macapp() {
     mkdir -p dist/Ollama.app/Contents/Resources
     if [ -d dist/darwin-amd64 ]; then
         lipo -create -output dist/Ollama.app/Contents/Resources/ollama dist/darwin-amd64/ollama dist/darwin-arm64/ollama
-        lipo -create -output dist/Ollama.app/Contents/Resources/ollama-mlx dist/darwin-amd64/ollama-mlx dist/darwin-arm64/ollama-mlx
         for F in dist/darwin-amd64/lib/ollama/*mlx*.dylib ; do
             lipo -create -output dist/darwin/$(basename $F) $F dist/darwin-arm64/lib/ollama/$(basename $F)
         done
@@ -166,28 +165,27 @@ _build_macapp() {
         cp -a dist/darwin/ollama dist/Ollama.app/Contents/Resources/ollama
         cp dist/darwin/*.so dist/darwin/*.dylib dist/Ollama.app/Contents/Resources/
     fi
-    cp -a dist/darwin/ollama-mlx dist/Ollama.app/Contents/Resources/ollama-mlx
     chmod a+x dist/Ollama.app/Contents/Resources/ollama
 
     # Sign
     if [ -n "$APPLE_IDENTITY" ]; then
         codesign -f --timestamp -s "$APPLE_IDENTITY" --identifier ai.ollama.ollama --options=runtime dist/Ollama.app/Contents/Resources/ollama
-        for lib in dist/Ollama.app/Contents/Resources/*.so dist/Ollama.app/Contents/Resources/*.dylib dist/Ollama.app/Contents/Resources/*.metallib dist/Ollama.app/Contents/Resources/ollama-mlx ; do
+        for lib in dist/Ollama.app/Contents/Resources/*.so dist/Ollama.app/Contents/Resources/*.dylib dist/Ollama.app/Contents/Resources/*.metallib ; do
             codesign -f --timestamp -s "$APPLE_IDENTITY" --identifier ai.ollama.ollama --options=runtime ${lib}
         done
         codesign -f --timestamp -s "$APPLE_IDENTITY" --identifier com.electron.ollama --deep --options=runtime dist/Ollama.app
     fi
 
     rm -f dist/Ollama-darwin.zip
-    ditto -c -k --keepParent dist/Ollama.app dist/Ollama-darwin.zip
-    (cd dist/Ollama.app/Contents/Resources/; tar -cf - ollama ollama-mlx *.so *.dylib *.metallib 2>/dev/null) | gzip -9vc > dist/ollama-darwin.tgz
+    ditto -c -k --norsrc --keepParent dist/Ollama.app dist/Ollama-darwin.zip
+    (cd dist/Ollama.app/Contents/Resources/; tar -cf - ollama *.so *.dylib *.metallib 2>/dev/null) | gzip -9vc > dist/ollama-darwin.tgz
 
     # Notarize and Staple
     if [ -n "$APPLE_IDENTITY" ]; then
         $(xcrun -f notarytool) submit dist/Ollama-darwin.zip --wait --timeout 20m --apple-id "$APPLE_ID" --password "$APPLE_PASSWORD" --team-id "$APPLE_TEAM_ID"
         rm -f dist/Ollama-darwin.zip
         $(xcrun -f stapler) staple dist/Ollama.app
-        ditto -c -k --keepParent dist/Ollama.app dist/Ollama-darwin.zip
+        ditto -c -k --norsrc --keepParent dist/Ollama.app dist/Ollama-darwin.zip
 
         rm -f dist/Ollama.dmg
 

server/auth.go

@@ -50,12 +50,17 @@ func (r registryChallenge) URL() (*url.URL, error) {
 	return redirectURL, nil
 }
 
-func getAuthorizationToken(ctx context.Context, challenge registryChallenge) (string, error) {
+func getAuthorizationToken(ctx context.Context, challenge registryChallenge, originalHost string) (string, error) {
 	redirectURL, err := challenge.URL()
 	if err != nil {
 		return "", err
 	}
 
+	// Validate that the realm host matches the original request host to prevent sending tokens cross-origin.
+	if redirectURL.Host != originalHost {
+		return "", fmt.Errorf("realm host %q does not match original host %q", redirectURL.Host, originalHost)
+	}
+
 	sha256sum := sha256.Sum256(nil)
 	data := []byte(fmt.Sprintf("%s,%s,%s", http.MethodGet, redirectURL.String(), base64.StdEncoding.EncodeToString([]byte(hex.EncodeToString(sha256sum[:])))))
 

server/auth_test.go

@@ -0,0 +1,113 @@
+package server
+
+import (
+	"context"
+	"strings"
+	"testing"
+	"time"
+)
+
+func TestGetAuthorizationTokenRejectsCrossDomain(t *testing.T) {
+	tests := []struct {
+		realm        string
+		originalHost string
+		wantMismatch bool
+	}{
+		{"https://example.com/token", "example.com", false},
+		{"https://example.com/token", "other.com", true},
+		{"https://example.com/token", "localhost:8000", true},
+		{"https://localhost:5000/token", "localhost:5000", false},
+		{"https://localhost:5000/token", "localhost:6000", true},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.originalHost, func(t *testing.T) {
+			ctx, cancel := context.WithTimeout(context.Background(), 100*time.Millisecond)
+			defer cancel()
+
+			challenge := registryChallenge{Realm: tt.realm, Service: "test", Scope: "repo:x:pull"}
+			_, err := getAuthorizationToken(ctx, challenge, tt.originalHost)
+
+			isMismatch := err != nil && strings.Contains(err.Error(), "does not match")
+			if tt.wantMismatch && !isMismatch {
+				t.Errorf("expected domain mismatch error, got: %v", err)
+			}
+			if !tt.wantMismatch && isMismatch {
+				t.Errorf("unexpected domain mismatch error: %v", err)
+			}
+		})
+	}
+}
+
+func TestParseRegistryChallenge(t *testing.T) {
+	tests := []struct {
+		input                             string
+		wantRealm, wantService, wantScope string
+	}{
+		{
+			`Bearer realm="https://auth.example.com/token",service="registry",scope="repo:foo:pull"`,
+			"https://auth.example.com/token", "registry", "repo:foo:pull",
+		},
+		{
+			`Bearer realm="https://r.ollama.ai/v2/token",service="ollama",scope="-"`,
+			"https://r.ollama.ai/v2/token", "ollama", "-",
+		},
+		{"", "", "", ""},
+	}
+
+	for _, tt := range tests {
+		result := parseRegistryChallenge(tt.input)
+		if result.Realm != tt.wantRealm || result.Service != tt.wantService || result.Scope != tt.wantScope {
+			t.Errorf("parseRegistryChallenge(%q) = {%q, %q, %q}, want {%q, %q, %q}",
+				tt.input, result.Realm, result.Service, result.Scope,
+				tt.wantRealm, tt.wantService, tt.wantScope)
+		}
+	}
+}
+
+func TestRegistryChallengeURL(t *testing.T) {
+	challenge := registryChallenge{
+		Realm:   "https://auth.example.com/token",
+		Service: "registry",
+		Scope:   "repo:foo:pull repo:bar:push",
+	}
+
+	u, err := challenge.URL()
+	if err != nil {
+		t.Fatalf("URL() error: %v", err)
+	}
+
+	if u.Host != "auth.example.com" {
+		t.Errorf("host = %q, want %q", u.Host, "auth.example.com")
+	}
+	if u.Path != "/token" {
+		t.Errorf("path = %q, want %q", u.Path, "/token")
+	}
+
+	q := u.Query()
+	if q.Get("service") != "registry" {
+		t.Errorf("service = %q, want %q", q.Get("service"), "registry")
+	}
+	if scopes := q["scope"]; len(scopes) != 2 {
+		t.Errorf("scope count = %d, want 2", len(scopes))
+	}
+	if q.Get("ts") == "" {
+		t.Error("missing ts")
+	}
+	if q.Get("nonce") == "" {
+		t.Error("missing nonce")
+	}
+
+	// Nonces should differ between calls
+	u2, _ := challenge.URL()
+	if q.Get("nonce") == u2.Query().Get("nonce") {
+		t.Error("nonce should be unique per call")
+	}
+}
+
+func TestRegistryChallengeURLInvalid(t *testing.T) {
+	challenge := registryChallenge{Realm: "://invalid"}
+	if _, err := challenge.URL(); err == nil {
+		t.Error("expected error for invalid URL")
+	}
+}

server/download.go

@@ -95,48 +95,11 @@ func (p *blobDownloadPart) UnmarshalJSON(b []byte) error {
 }
 
 const (
-	// numDownloadParts is the default number of concurrent download parts for standard downloads
-	numDownloadParts = 16
-	// numHFDownloadParts is the reduced number of concurrent download parts for HuggingFace
-	// downloads to avoid triggering rate limits (HTTP 429 errors). See GitHub issue #13297.
-	numHFDownloadParts        = 4
+	numDownloadParts          = 16
 	minDownloadPartSize int64 = 100 * format.MegaByte
 	maxDownloadPartSize int64 = 1000 * format.MegaByte
 )
 
-// isHuggingFaceURL returns true if the URL is from a HuggingFace domain.
-// This includes:
-//   - huggingface.co (main domain)
-//   - *.huggingface.co (subdomains like cdn-lfs.huggingface.co)
-//   - hf.co (shortlink domain)
-//   - *.hf.co (CDN domains like cdn-lfs.hf.co, cdn-lfs3.hf.co)
-func isHuggingFaceURL(u *url.URL) bool {
-	if u == nil {
-		return false
-	}
-	host := strings.ToLower(u.Hostname())
-	return host == "huggingface.co" ||
-		strings.HasSuffix(host, ".huggingface.co") ||
-		host == "hf.co" ||
-		strings.HasSuffix(host, ".hf.co")
-}
-
-// getNumDownloadParts returns the number of concurrent download parts to use
-// for the given URL. HuggingFace URLs use reduced concurrency (default 4) to
-// avoid triggering rate limits. This can be overridden via the OLLAMA_HF_CONCURRENCY
-// environment variable. For non-HuggingFace URLs, returns the standard concurrency (16).
-func getNumDownloadParts(u *url.URL) int {
-	if isHuggingFaceURL(u) {
-		if v := os.Getenv("OLLAMA_HF_CONCURRENCY"); v != "" {
-			if n, err := strconv.Atoi(v); err == nil && n > 0 {
-				return n
-			}
-		}
-		return numHFDownloadParts
-	}
-	return numDownloadParts
-}
-
 func (p *blobDownloadPart) Name() string {
 	return strings.Join([]string{
 		p.blobDownload.Name, "partial", strconv.Itoa(p.N),
@@ -308,11 +271,7 @@ func (b *blobDownload) run(ctx context.Context, requestURL *url.URL, opts *regis
 	}
 
 	g, inner := errgroup.WithContext(ctx)
-	concurrency := getNumDownloadParts(directURL)
-	if concurrency != numDownloadParts {
-		slog.Info(fmt.Sprintf("using reduced concurrency (%d) for HuggingFace download", concurrency))
-	}
-	g.SetLimit(concurrency)
+	g.SetLimit(numDownloadParts)
 	for i := range b.Parts {
 		part := b.Parts[i]
 		if part.Completed.Load() == part.Size {

server/download_test.go

@@ -1,194 +0,0 @@
-package server
-
-import (
-	"net/url"
-	"testing"
-
-	"github.com/stretchr/testify/assert"
-)
-
-func TestIsHuggingFaceURL(t *testing.T) {
-	tests := []struct {
-		name     string
-		url      string
-		expected bool
-	}{
-		{
-			name:     "nil url",
-			url:      "",
-			expected: false,
-		},
-		{
-			name:     "huggingface.co main domain",
-			url:      "https://huggingface.co/some/model",
-			expected: true,
-		},
-		{
-			name:     "cdn-lfs.huggingface.co subdomain",
-			url:      "https://cdn-lfs.huggingface.co/repos/abc/123",
-			expected: true,
-		},
-		{
-			name:     "cdn-lfs3.hf.co CDN domain",
-			url:      "https://cdn-lfs3.hf.co/repos/abc/123",
-			expected: true,
-		},
-		{
-			name:     "hf.co shortlink domain",
-			url:      "https://hf.co/model",
-			expected: true,
-		},
-		{
-			name:     "uppercase HuggingFace domain",
-			url:      "https://HUGGINGFACE.CO/model",
-			expected: true,
-		},
-		{
-			name:     "mixed case HF domain",
-			url:      "https://Cdn-Lfs.HF.Co/repos",
-			expected: true,
-		},
-		{
-			name:     "ollama registry",
-			url:      "https://registry.ollama.ai/v2/library/llama3",
-			expected: false,
-		},
-		{
-			name:     "github.com",
-			url:      "https://github.com/ollama/ollama",
-			expected: false,
-		},
-		{
-			name:     "fake huggingface domain",
-			url:      "https://nothuggingface.co/model",
-			expected: false,
-		},
-		{
-			name:     "fake hf domain",
-			url:      "https://nothf.co/model",
-			expected: false,
-		},
-		{
-			name:     "huggingface in path not host",
-			url:      "https://example.com/huggingface.co/model",
-			expected: false,
-		},
-	}
-
-	for _, tc := range tests {
-		t.Run(tc.name, func(t *testing.T) {
-			var u *url.URL
-			if tc.url != "" {
-				var err error
-				u, err = url.Parse(tc.url)
-				if err != nil {
-					t.Fatalf("failed to parse URL: %v", err)
-				}
-			}
-			got := isHuggingFaceURL(u)
-			assert.Equal(t, tc.expected, got)
-		})
-	}
-}
-
-func TestGetNumDownloadParts(t *testing.T) {
-	tests := []struct {
-		name        string
-		url         string
-		envValue    string
-		expected    int
-		description string
-	}{
-		{
-			name:        "nil url returns default",
-			url:         "",
-			envValue:    "",
-			expected:    numDownloadParts,
-			description: "nil URL should return standard concurrency",
-		},
-		{
-			name:        "ollama registry returns default",
-			url:         "https://registry.ollama.ai/v2/library/llama3",
-			envValue:    "",
-			expected:    numDownloadParts,
-			description: "Ollama registry should use standard concurrency",
-		},
-		{
-			name:        "huggingface returns reduced default",
-			url:         "https://huggingface.co/model/repo",
-			envValue:    "",
-			expected:    numHFDownloadParts,
-			description: "HuggingFace should use reduced concurrency",
-		},
-		{
-			name:        "hf.co CDN returns reduced default",
-			url:         "https://cdn-lfs3.hf.co/repos/abc/123",
-			envValue:    "",
-			expected:    numHFDownloadParts,
-			description: "HuggingFace CDN should use reduced concurrency",
-		},
-		{
-			name:        "huggingface with env override",
-			url:         "https://huggingface.co/model/repo",
-			envValue:    "2",
-			expected:    2,
-			description: "OLLAMA_HF_CONCURRENCY should override default",
-		},
-		{
-			name:        "huggingface with higher env override",
-			url:         "https://huggingface.co/model/repo",
-			envValue:    "8",
-			expected:    8,
-			description: "OLLAMA_HF_CONCURRENCY can be set higher than default",
-		},
-		{
-			name:        "huggingface with invalid env (non-numeric)",
-			url:         "https://huggingface.co/model/repo",
-			envValue:    "invalid",
-			expected:    numHFDownloadParts,
-			description: "Invalid OLLAMA_HF_CONCURRENCY should fall back to default",
-		},
-		{
-			name:        "huggingface with invalid env (zero)",
-			url:         "https://huggingface.co/model/repo",
-			envValue:    "0",
-			expected:    numHFDownloadParts,
-			description: "Zero OLLAMA_HF_CONCURRENCY should fall back to default",
-		},
-		{
-			name:        "huggingface with invalid env (negative)",
-			url:         "https://huggingface.co/model/repo",
-			envValue:    "-1",
-			expected:    numHFDownloadParts,
-			description: "Negative OLLAMA_HF_CONCURRENCY should fall back to default",
-		},
-		{
-			name:        "non-huggingface ignores env",
-			url:         "https://registry.ollama.ai/v2/library/llama3",
-			envValue:    "2",
-			expected:    numDownloadParts,
-			description: "OLLAMA_HF_CONCURRENCY should not affect non-HF URLs",
-		},
-	}
-
-	for _, tc := range tests {
-		t.Run(tc.name, func(t *testing.T) {
-			// Set or clear the environment variable
-			if tc.envValue != "" {
-				t.Setenv("OLLAMA_HF_CONCURRENCY", tc.envValue)
-			}
-
-			var u *url.URL
-			if tc.url != "" {
-				var err error
-				u, err = url.Parse(tc.url)
-				if err != nil {
-					t.Fatalf("failed to parse URL: %v", err)
-				}
-			}
-
-			got := getNumDownloadParts(u)
-			assert.Equal(t, tc.expected, got, tc.description)
-		})
-	}
-}

server/images.go

@@ -41,6 +41,7 @@ var (
 	errCapabilityVision     = errors.New("vision")
 	errCapabilityEmbedding  = errors.New("embedding")
 	errCapabilityThinking   = errors.New("thinking")
+	errCapabilityImage      = errors.New("image generation")
 	errInsecureProtocol     = errors.New("insecure protocol http")
 )
 
@@ -76,7 +77,7 @@ func (m *Model) Capabilities() []model.Capability {
 
 	// Check for image generation model via config capabilities
 	if slices.Contains(m.Config.Capabilities, "image") {
-		return []model.Capability{model.CapabilityImageGeneration}
+		return []model.Capability{model.CapabilityImage}
 	}
 
 	// Check for completion capability
@@ -159,6 +160,7 @@ func (m *Model) CheckCapabilities(want ...model.Capability) error {
 		model.CapabilityVision:     errCapabilityVision,
 		model.CapabilityEmbedding:  errCapabilityEmbedding,
 		model.CapabilityThinking:   errCapabilityThinking,
+		model.CapabilityImage:      errCapabilityImage,
 	}
 
 	for _, cap := range want {
@@ -775,7 +777,7 @@ func pullWithTransfer(ctx context.Context, mp ModelPath, layers []Layer, manifes
 			Realm:   challenge.Realm,
 			Service: challenge.Service,
 			Scope:   challenge.Scope,
-		})
+		}, base.Host)
 	}
 
 	if err := transfer.Download(ctx, transfer.DownloadOptions{
@@ -850,7 +852,7 @@ func pushWithTransfer(ctx context.Context, mp ModelPath, layers []Layer, manifes
 			Realm:   challenge.Realm,
 			Service: challenge.Service,
 			Scope:   challenge.Scope,
-		})
+		}, base.Host)
 	}
 
 	return transfer.Upload(ctx, transfer.UploadOptions{
@@ -916,7 +918,7 @@ func makeRequestWithRetry(ctx context.Context, method string, requestURL *url.UR
 
 			// Handle authentication error with one retry
 			challenge := parseRegistryChallenge(resp.Header.Get("www-authenticate"))
-			token, err := getAuthorizationToken(ctx, challenge)
+			token, err := getAuthorizationToken(ctx, challenge, requestURL.Host)
 			if err != nil {
 				return nil, err
 			}

server/images_test.go

@@ -54,7 +54,7 @@ func TestModelCapabilities(t *testing.T) {
 					Capabilities: []string{"image"},
 				},
 			},
-			expectedCaps: []model.Capability{model.CapabilityImageGeneration},
+			expectedCaps: []model.Capability{model.CapabilityImage},
 		},
 		{
 			name: "model with completion capability",
@@ -242,6 +242,24 @@ func TestModelCheckCapabilities(t *testing.T) {
 			checkCaps:      []model.Capability{"unknown"},
 			expectedErrMsg: "unknown capability",
 		},
+		{
+			name: "model missing image generation capability",
+			model: Model{
+				ModelPath: completionModelPath,
+				Template:  chatTemplate,
+			},
+			checkCaps:      []model.Capability{model.CapabilityImage},
+			expectedErrMsg: "does not support image generation",
+		},
+		{
+			name: "model with image generation capability",
+			model: Model{
+				Config: model.ConfigV2{
+					Capabilities: []string{"image"},
+				},
+			},
+			checkCaps: []model.Capability{model.CapabilityImage},
+		},
 	}
 
 	for _, tt := range tests {

server/routes.go

@@ -51,7 +51,7 @@ import (
 	"github.com/ollama/ollama/types/model"
 	"github.com/ollama/ollama/version"
 	"github.com/ollama/ollama/x/imagegen"
-	imagegenapi "github.com/ollama/ollama/x/imagegen/api"
+	xserver "github.com/ollama/ollama/x/server"
 )
 
 const signinURLStr = "https://ollama.com/connect?name=%s&key=%s"
@@ -164,29 +164,6 @@ func (s *Server) scheduleRunner(ctx context.Context, name string, caps []model.C
 	return runner.llama, model, &opts, nil
 }
 
-// ScheduleImageGenRunner schedules an image generation model runner.
-// This implements the imagegenapi.RunnerScheduler interface.
-func (s *Server) ScheduleImageGenRunner(c *gin.Context, modelName string, opts api.Options, keepAlive *api.Duration) (llm.LlamaServer, error) {
-	m := &Model{
-		Name:      modelName,
-		ShortName: modelName,
-		ModelPath: modelName, // For image gen, ModelPath is just the model name
-		Config: model.ConfigV2{
-			Capabilities: []string{"image"},
-		},
-	}
-
-	runnerCh, errCh := s.sched.GetRunner(c.Request.Context(), m, opts, keepAlive)
-	var runner *runnerRef
-	select {
-	case runner = <-runnerCh:
-	case err := <-errCh:
-		return nil, err
-	}
-
-	return runner.llama, nil
-}
-
 func signinURL() (string, error) {
 	pubKey, err := auth.GetPublicKey()
 	if err != nil {
@@ -214,12 +191,6 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 		return
 	}
 
-	// Check if this is a known image generation model
-	if imagegen.ResolveModelName(req.Model) != "" {
-		imagegenapi.HandleGenerateRequest(c, s, req.Model, req.Prompt, req.KeepAlive, streamResponse)
-		return
-	}
-
 	name := model.ParseName(req.Model)
 	if !name.IsValid() {
 		// Ideally this is "invalid model name" but we're keeping with
@@ -249,6 +220,12 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 		return
 	}
 
+	// Handle image generation models
+	if slices.Contains(m.Capabilities(), model.CapabilityImage) {
+		s.handleImageGenerate(c, req, name.String(), checkpointStart)
+		return
+	}
+
 	if req.TopLogprobs < 0 || req.TopLogprobs > 20 {
 		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "top_logprobs must be between 0 and 20"})
 		return
@@ -1125,7 +1102,7 @@ func GetModelInfo(req api.ShowRequest) (*api.ShowResponse, error) {
 	}
 
 	// For image generation models, populate details from imagegen package
-	if slices.Contains(m.Capabilities(), model.CapabilityImageGeneration) {
+	if slices.Contains(m.Capabilities(), model.CapabilityImage) {
 		if info, err := imagegen.GetModelInfo(name.String()); err == nil {
 			modelDetails.Family = info.Architecture
 			modelDetails.ParameterSize = format.HumanNumber(uint64(info.ParameterCount))
@@ -1133,6 +1110,22 @@ func GetModelInfo(req api.ShowRequest) (*api.ShowResponse, error) {
 		}
 	}
 
+	// For safetensors LLM models (experimental), populate details from config.json
+	if m.Config.ModelFormat == "safetensors" && slices.Contains(m.Config.Capabilities, "completion") {
+		if info, err := xserver.GetSafetensorsLLMInfo(name.String()); err == nil {
+			if arch, ok := info["general.architecture"].(string); ok && arch != "" {
+				modelDetails.Family = arch
+			}
+			if paramCount, ok := info["general.parameter_count"].(int64); ok && paramCount > 0 {
+				modelDetails.ParameterSize = format.HumanNumber(uint64(paramCount))
+			}
+		}
+		// Get torch_dtype directly from config.json for quantization level
+		if dtype, err := xserver.GetSafetensorsDtype(name.String()); err == nil && dtype != "" {
+			modelDetails.QuantizationLevel = dtype
+		}
+	}
+
 	if req.System != "" {
 		m.System = req.System
 	}
@@ -1215,7 +1208,27 @@ func GetModelInfo(req api.ShowRequest) (*api.ShowResponse, error) {
 		return resp, nil
 	}
 
-	if slices.Contains(m.Capabilities(), model.CapabilityImageGeneration) {
+	if slices.Contains(m.Capabilities(), model.CapabilityImage) {
+		// Populate tensor info if verbose
+		if req.Verbose {
+			if tensors, err := xserver.GetSafetensorsTensorInfo(name.String()); err == nil {
+				resp.Tensors = tensors
+			}
+		}
+		return resp, nil
+	}
+
+	// For safetensors LLM models (experimental), populate ModelInfo from config.json
+	if m.Config.ModelFormat == "safetensors" && slices.Contains(m.Config.Capabilities, "completion") {
+		if info, err := xserver.GetSafetensorsLLMInfo(name.String()); err == nil {
+			resp.ModelInfo = info
+		}
+		// Populate tensor info if verbose
+		if req.Verbose {
+			if tensors, err := xserver.GetSafetensorsTensorInfo(name.String()); err == nil {
+				resp.Tensors = tensors
+			}
+		}
 		return resp, nil
 	}
 
@@ -1587,13 +1600,12 @@ func (s *Server) GenerateRoutes(rc *ollama.Registry) (http.Handler, error) {
 	r.GET("/v1/models", middleware.ListMiddleware(), s.ListHandler)
 	r.GET("/v1/models/:model", middleware.RetrieveMiddleware(), s.ShowHandler)
 	r.POST("/v1/responses", middleware.ResponsesMiddleware(), s.ChatHandler)
+	// OpenAI-compatible image generation endpoint
+	r.POST("/v1/images/generations", middleware.ImageGenerationsMiddleware(), s.GenerateHandler)
 
 	// Inference (Anthropic compatibility)
 	r.POST("/v1/messages", middleware.AnthropicMessagesMiddleware(), s.ChatHandler)
 
-	// Experimental image generation support
-	imagegenapi.RegisterRoutes(r, s)
-
 	if rc != nil {
 		// wrap old with new
 		rs := &registry.Local{
@@ -2460,3 +2472,91 @@ func filterThinkTags(msgs []api.Message, m *Model) []api.Message {
 	}
 	return msgs
 }
+
+// handleImageGenerate handles image generation requests within GenerateHandler.
+// This is called when the model has the Image capability.
+func (s *Server) handleImageGenerate(c *gin.Context, req api.GenerateRequest, modelName string, checkpointStart time.Time) {
+	// Validate image dimensions
+	const maxDimension int32 = 4096
+	if req.Width > maxDimension || req.Height > maxDimension {
+		c.JSON(http.StatusBadRequest, gin.H{"error": fmt.Sprintf("width and height must be <= %d", maxDimension)})
+		return
+	}
+
+	// Schedule the runner for image generation
+	runner, _, _, err := s.scheduleRunner(c.Request.Context(), modelName, []model.Capability{model.CapabilityImage}, nil, req.KeepAlive)
+	if err != nil {
+		handleScheduleError(c, req.Model, err)
+		return
+	}
+
+	checkpointLoaded := time.Now()
+
+	// Handle load-only request (empty prompt)
+	if req.Prompt == "" {
+		c.JSON(http.StatusOK, api.GenerateResponse{
+			Model:      req.Model,
+			CreatedAt:  time.Now().UTC(),
+			Done:       true,
+			DoneReason: "load",
+		})
+		return
+	}
+
+	// Set headers for streaming response
+	c.Header("Content-Type", "application/x-ndjson")
+
+	// Get seed from options if provided
+	var seed int64
+	if s, ok := req.Options["seed"]; ok {
+		switch v := s.(type) {
+		case int:
+			seed = int64(v)
+		case int64:
+			seed = v
+		case float64:
+			seed = int64(v)
+		}
+	}
+
+	var streamStarted bool
+	if err := runner.Completion(c.Request.Context(), llm.CompletionRequest{
+		Prompt: req.Prompt,
+		Width:  req.Width,
+		Height: req.Height,
+		Steps:  req.Steps,
+		Seed:   seed,
+	}, func(cr llm.CompletionResponse) {
+		streamStarted = true
+		res := api.GenerateResponse{
+			Model:     req.Model,
+			CreatedAt: time.Now().UTC(),
+			Done:      cr.Done,
+		}
+
+		if cr.TotalSteps > 0 {
+			res.Completed = int64(cr.Step)
+			res.Total = int64(cr.TotalSteps)
+		}
+
+		if cr.Image != "" {
+			res.Image = cr.Image
+		}
+
+		if cr.Done {
+			res.DoneReason = cr.DoneReason.String()
+			res.Metrics.TotalDuration = time.Since(checkpointStart)
+			res.Metrics.LoadDuration = checkpointLoaded.Sub(checkpointStart)
+		}
+
+		data, _ := json.Marshal(res)
+		c.Writer.Write(append(data, '\n'))
+		c.Writer.Flush()
+	}); err != nil {
+		// Only send JSON error if streaming hasn't started yet
+		// (once streaming starts, headers are committed and we can't change status code)
+		if !streamStarted {
+			c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+		}
+	}
+}

server/sched.go

@@ -574,7 +574,8 @@ func (s *Scheduler) loadImageGen(req *LlmRequest) bool {
 		Options:         &req.opts,
 		loading:         false,
 		sessionDuration: sessionDuration,
-		refCount:        1,
+		totalSize:       server.TotalSize(),
+		vramSize:        server.VRAMSize(),
 	}
 
 	s.loadedMu.Lock()

server/sched_test.go

@@ -6,7 +6,6 @@ import (
 	"errors"
 	"log/slog"
 	"os"
-	"slices"
 	"testing"
 	"time"
 
@@ -17,7 +16,6 @@ import (
 	"github.com/ollama/ollama/fs/ggml"
 	"github.com/ollama/ollama/llm"
 	"github.com/ollama/ollama/ml"
-	"github.com/ollama/ollama/types/model"
 )
 
 func TestMain(m *testing.M) {
@@ -807,32 +805,8 @@ func (s *mockLlm) GetDeviceInfos(ctx context.Context) []ml.DeviceInfo { return n
 func (s *mockLlm) HasExited() bool                                    { return false }
 func (s *mockLlm) GetActiveDeviceIDs() []ml.DeviceID                  { return nil }
 
-// TestImageGenCapabilityDetection verifies that models with "image" capability
-// are correctly identified and routed differently from language models.
-func TestImageGenCapabilityDetection(t *testing.T) {
-	// Model with image capability should be detected
-	imageModel := &Model{
-		Config: model.ConfigV2{
-			Capabilities: []string{"image"},
-		},
-	}
-	require.True(t, slices.Contains(imageModel.Config.Capabilities, "image"))
-
-	// Model without image capability should not be detected
-	langModel := &Model{
-		Config: model.ConfigV2{
-			Capabilities: []string{"completion"},
-		},
-	}
-	require.False(t, slices.Contains(langModel.Config.Capabilities, "image"))
-
-	// Empty capabilities should not match
-	emptyModel := &Model{}
-	require.False(t, slices.Contains(emptyModel.Config.Capabilities, "image"))
-}
-
 // TestImageGenRunnerCanBeEvicted verifies that an image generation model
-// loaded in the scheduler can be evicted by a language model request.
+// loaded in the scheduler can be evicted when idle.
 func TestImageGenRunnerCanBeEvicted(t *testing.T) {
 	ctx, done := context.WithTimeout(t.Context(), 500*time.Millisecond)
 	defer done()
@@ -864,3 +838,59 @@ func TestImageGenRunnerCanBeEvicted(t *testing.T) {
 	require.NotNil(t, runner)
 	require.Equal(t, "/fake/image/model", runner.modelPath)
 }
+
+// TestImageGenSchedulerCoexistence verifies that image generation models
+// can coexist with language models in the scheduler and VRAM is tracked correctly.
+func TestImageGenSchedulerCoexistence(t *testing.T) {
+	ctx, done := context.WithTimeout(t.Context(), 500*time.Millisecond)
+	defer done()
+
+	s := InitScheduler(ctx)
+	s.getGpuFn = getGpuFn
+	s.getSystemInfoFn = getSystemInfoFn
+
+	// Load both an imagegen runner and a language model runner
+	imageGenRunner := &runnerRef{
+		model:           &Model{Name: "flux", ModelPath: "/fake/flux/model"},
+		modelPath:       "/fake/flux/model",
+		llama:           &mockLlm{vramSize: 8 * format.GigaByte, vramByGPU: map[ml.DeviceID]uint64{{Library: "Metal"}: 8 * format.GigaByte}},
+		sessionDuration: 10 * time.Millisecond,
+		numParallel:     1,
+		refCount:        0,
+	}
+
+	langModelRunner := &runnerRef{
+		model:           &Model{Name: "llama3", ModelPath: "/fake/llama3/model"},
+		modelPath:       "/fake/llama3/model",
+		llama:           &mockLlm{vramSize: 4 * format.GigaByte, vramByGPU: map[ml.DeviceID]uint64{{Library: "Metal"}: 4 * format.GigaByte}},
+		sessionDuration: 10 * time.Millisecond,
+		numParallel:     1,
+		refCount:        0,
+	}
+
+	s.loadedMu.Lock()
+	s.loaded["/fake/flux/model"] = imageGenRunner
+	s.loaded["/fake/llama3/model"] = langModelRunner
+	s.loadedMu.Unlock()
+
+	// Verify both are loaded
+	s.loadedMu.Lock()
+	require.Len(t, s.loaded, 2)
+	require.NotNil(t, s.loaded["/fake/flux/model"])
+	require.NotNil(t, s.loaded["/fake/llama3/model"])
+	s.loadedMu.Unlock()
+
+	// Verify updateFreeSpace accounts for both
+	gpus := []ml.DeviceInfo{
+		{
+			DeviceID:    ml.DeviceID{Library: "Metal"},
+			TotalMemory: 24 * format.GigaByte,
+			FreeMemory:  24 * format.GigaByte,
+		},
+	}
+	s.updateFreeSpace(gpus)
+
+	// Free memory should be reduced by both models
+	expectedFree := uint64(24*format.GigaByte) - uint64(8*format.GigaByte) - uint64(4*format.GigaByte)
+	require.Equal(t, expectedFree, gpus[0].FreeMemory)
+}

server/upload.go

@@ -279,7 +279,7 @@ func (b *blobUpload) uploadPart(ctx context.Context, method string, requestURL *
 	case resp.StatusCode == http.StatusUnauthorized:
 		w.Rollback()
 		challenge := parseRegistryChallenge(resp.Header.Get("www-authenticate"))
-		token, err := getAuthorizationToken(ctx, challenge)
+		token, err := getAuthorizationToken(ctx, challenge, requestURL.Host)
 		if err != nil {
 			return err
 		}

types/model/capability.go

@@ -9,7 +9,7 @@ const (
 	CapabilityVision          = Capability("vision")
 	CapabilityEmbedding       = Capability("embedding")
 	CapabilityThinking        = Capability("thinking")
-	CapabilityImageGeneration = Capability("image")
+	CapabilityImage = Capability("image")
 )
 
 func (c Capability) String() string {

x/README.md

@@ -1,50 +0,0 @@
-# Experimental Features
-
-## MLX Backend
-
-We're working on a new experimental backend based on the [MLX project](https://github.com/ml-explore/mlx)
-
-Support is currently limited to MacOS and Linux with CUDA GPUs. We're looking to add support for Windows CUDA soon, and other GPU vendors.
-
-### Building ollama-mlx
-
-The `ollama-mlx` binary is a separate build of Ollama with MLX support enabled. This enables experimental features like image generation.
-
-#### macOS (Apple Silicon and Intel)
-
-```bash
-# Build MLX backend libraries
-cmake --preset MLX
-cmake --build --preset MLX --parallel
-cmake --install build --component MLX
-
-# Build ollama-mlx binary
-go build -tags mlx -o ollama-mlx .
-```
-
-#### Linux (CUDA)
-
-On Linux, use the preset "MLX CUDA 13" or "MLX CUDA 12" to enable CUDA with the default Ollama NVIDIA GPU architectures enabled:
-
-```bash
-# Build MLX backend libraries with CUDA support
-cmake --preset 'MLX CUDA 13'
-cmake --build --preset 'MLX CUDA 13' --parallel
-cmake --install build --component MLX
-
-# Build ollama-mlx binary
-CGO_CFLAGS="-O3 -I$(pwd)/build/_deps/mlx-c-src" \
-CGO_LDFLAGS="-L$(pwd)/build/lib/ollama -lmlxc -lmlx" \
-go build -tags mlx -o ollama-mlx .
-```
-
-#### Using build scripts
-
-The build scripts automatically create the `ollama-mlx` binary:
-
-- **macOS**: `./scripts/build_darwin.sh` produces `dist/darwin/ollama-mlx`
-- **Linux**: `./scripts/build_linux.sh` produces `ollama-mlx` in the output archives
-
-## Image Generation
-
-Image generation is built into the `ollama-mlx` binary. Run `ollama-mlx serve` to start the server with image generation support enabled.

x/cmd/run.go

@@ -25,14 +25,6 @@ import (
 	"github.com/ollama/ollama/x/tools"
 )
 
-// MultilineState tracks the state of multiline input
-type MultilineState int
-
-const (
-	MultilineNone MultilineState = iota
-	MultilineSystem
-)
-
 // Tool output capping constants
 const (
 	// localModelTokenLimit is the token limit for local models (smaller context).
@@ -656,7 +648,7 @@ func GenerateInteractive(cmd *cobra.Command, modelName string, wordWrap bool, op
 		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
@@ -707,7 +699,6 @@ func GenerateInteractive(cmd *cobra.Command, modelName string, wordWrap bool, op
 	var sb strings.Builder
 	var format string
 	var system string
-	var multiline MultilineState = MultilineNone
 
 	for {
 		line, err := scanner.Readline()
@@ -721,37 +712,12 @@ func GenerateInteractive(cmd *cobra.Command, modelName string, wordWrap bool, op
 			}
 			scanner.Prompt.UseAlt = false
 			sb.Reset()
-			multiline = MultilineNone
 			continue
 		case err != nil:
 			return err
 		}
 
 		switch {
-		case multiline != MultilineNone:
-			// check if there's a multiline terminating string
-			before, ok := strings.CutSuffix(line, `"""`)
-			sb.WriteString(before)
-			if !ok {
-				fmt.Fprintln(&sb)
-				continue
-			}
-
-			switch multiline {
-			case MultilineSystem:
-				system = sb.String()
-				newMessage := api.Message{Role: "system", Content: system}
-				if len(messages) > 0 && messages[len(messages)-1].Role == "system" {
-					messages[len(messages)-1] = newMessage
-				} else {
-					messages = append(messages, newMessage)
-				}
-				fmt.Println("Set system message.")
-				sb.Reset()
-			}
-
-			multiline = MultilineNone
-			scanner.Prompt.UseAlt = false
 		case strings.HasPrefix(line, "/exit"), strings.HasPrefix(line, "/bye"):
 			return nil
 		case strings.HasPrefix(line, "/clear"):
@@ -860,41 +826,18 @@ func GenerateInteractive(cmd *cobra.Command, modelName string, wordWrap bool, op
 					options[args[2]] = fp[args[2]]
 				case "system":
 					if len(args) < 3 {
-						fmt.Println("Usage: /set system <message> or /set system \"\"\"<multi-line message>\"\"\"")
+						fmt.Println("Usage: /set system <message>")
 						continue
 					}
 
-					multiline = MultilineSystem
-
-					line := strings.Join(args[2:], " ")
-					line, ok := strings.CutPrefix(line, `"""`)
-					if !ok {
-						multiline = MultilineNone
-					} else {
-						// only cut suffix if the line is multiline
-						line, ok = strings.CutSuffix(line, `"""`)
-						if ok {
-							multiline = MultilineNone
-						}
-					}
-
-					sb.WriteString(line)
-					if multiline != MultilineNone {
-						scanner.Prompt.UseAlt = true
-						continue
-					}
-
-					system = sb.String()
-					newMessage := api.Message{Role: "system", Content: sb.String()}
-					// Check if the slice is not empty and the last message is from 'system'
+					system = strings.Join(args[2:], " ")
+					newMessage := api.Message{Role: "system", Content: system}
 					if len(messages) > 0 && messages[len(messages)-1].Role == "system" {
-						// Replace the last message
 						messages[len(messages)-1] = newMessage
 					} else {
 						messages = append(messages, newMessage)
 					}
 					fmt.Println("Set system message.")
-					sb.Reset()
 					continue
 				default:
 					fmt.Printf("Unknown command '/set %s'. Type /? for help\n", args[1])
@@ -1081,7 +1024,7 @@ func GenerateInteractive(cmd *cobra.Command, modelName string, wordWrap bool, op
 			sb.WriteString(line)
 		}
 
-		if sb.Len() > 0 && multiline == MultilineNone {
+		if sb.Len() > 0 {
 			newMessage := api.Message{Role: "user", Content: sb.String()}
 			messages = append(messages, newMessage)
 

x/create/client/create.go

@@ -0,0 +1,282 @@
+// Package client provides client-side model creation for safetensors-based models.
+//
+// This package is in x/ because the safetensors model storage format is under development.
+// It also exists to break an import cycle: server imports x/create, so x/create
+// cannot import server. This sub-package can import server because server doesn't
+// import it.
+package client
+
+import (
+	"bytes"
+	"encoding/json"
+	"fmt"
+	"io"
+
+	"github.com/ollama/ollama/progress"
+	"github.com/ollama/ollama/server"
+	"github.com/ollama/ollama/types/model"
+	"github.com/ollama/ollama/x/create"
+)
+
+// MinOllamaVersion is the minimum Ollama version required for safetensors models.
+const MinOllamaVersion = "0.14.0"
+
+// ModelfileConfig holds configuration extracted from a Modelfile.
+type ModelfileConfig struct {
+	Template string
+	System   string
+	License  string
+}
+
+// CreateOptions holds all options for model creation.
+type CreateOptions struct {
+	ModelName string
+	ModelDir  string
+	Quantize  string           // "fp8" for quantization
+	Modelfile *ModelfileConfig // template/system/license from Modelfile
+}
+
+// CreateModel imports a model from a local directory.
+// This creates blobs and manifest directly on disk, bypassing the HTTP API.
+// Automatically detects model type (safetensors LLM vs image gen) and routes accordingly.
+func CreateModel(opts CreateOptions, p *progress.Progress) error {
+	// Detect model type
+	isSafetensors := create.IsSafetensorsModelDir(opts.ModelDir)
+	isImageGen := create.IsTensorModelDir(opts.ModelDir)
+
+	if !isSafetensors && !isImageGen {
+		return fmt.Errorf("%s is not a supported model directory (needs config.json + *.safetensors or model_index.json)", opts.ModelDir)
+	}
+
+	// Determine model type settings
+	var modelType, spinnerKey string
+	var capabilities []string
+	if isSafetensors {
+		modelType = "safetensors model"
+		spinnerKey = "create"
+		capabilities = []string{"completion"}
+	} else {
+		modelType = "image generation model"
+		spinnerKey = "imagegen"
+		capabilities = []string{"image"}
+	}
+
+	// Set up progress spinner
+	statusMsg := "importing " + modelType
+	spinner := progress.NewSpinner(statusMsg)
+	p.Add(spinnerKey, spinner)
+
+	progressFn := func(msg string) {
+		spinner.Stop()
+		statusMsg = msg
+		spinner = progress.NewSpinner(statusMsg)
+		p.Add(spinnerKey, spinner)
+	}
+
+	// Create the model using shared callbacks
+	var err error
+	if isSafetensors {
+		err = create.CreateSafetensorsModel(
+			opts.ModelName, opts.ModelDir, opts.Quantize,
+			newLayerCreator(), newTensorLayerCreator(),
+			newManifestWriter(opts, capabilities),
+			progressFn,
+		)
+	} else {
+		err = create.CreateImageGenModel(
+			opts.ModelName, opts.ModelDir, opts.Quantize,
+			newLayerCreator(), newTensorLayerCreator(),
+			newManifestWriter(opts, capabilities),
+			progressFn,
+		)
+	}
+
+	spinner.Stop()
+	if err != nil {
+		return err
+	}
+
+	fmt.Printf("Created %s '%s'\n", modelType, opts.ModelName)
+	return nil
+}
+
+// newLayerCreator returns a LayerCreator callback for creating config/JSON layers.
+func newLayerCreator() create.LayerCreator {
+	return func(r io.Reader, mediaType, name string) (create.LayerInfo, error) {
+		layer, err := server.NewLayer(r, mediaType)
+		if err != nil {
+			return create.LayerInfo{}, err
+		}
+
+		return create.LayerInfo{
+			Digest:    layer.Digest,
+			Size:      layer.Size,
+			MediaType: layer.MediaType,
+			Name:      name,
+		}, nil
+	}
+}
+
+// newTensorLayerCreator returns a QuantizingTensorLayerCreator callback for creating tensor layers.
+// When quantize is non-empty, returns multiple layers (weight + scales + optional qbias).
+func newTensorLayerCreator() create.QuantizingTensorLayerCreator {
+	return func(r io.Reader, name, dtype string, shape []int32, quantize string) ([]create.LayerInfo, error) {
+		if quantize != "" {
+			return createQuantizedLayers(r, name, dtype, shape, quantize)
+		}
+		return createUnquantizedLayer(r, name)
+	}
+}
+
+// createQuantizedLayers quantizes a tensor and returns the resulting layers.
+func createQuantizedLayers(r io.Reader, name, dtype string, shape []int32, quantize string) ([]create.LayerInfo, error) {
+	if !QuantizeSupported() {
+		return nil, fmt.Errorf("quantization requires MLX support")
+	}
+
+	// Quantize the tensor
+	qweightData, scalesData, qbiasData, _, _, _, err := quantizeTensor(r, name, dtype, shape, quantize)
+	if err != nil {
+		return nil, fmt.Errorf("failed to quantize %s: %w", name, err)
+	}
+
+	// Create layer for quantized weight
+	weightLayer, err := server.NewLayer(bytes.NewReader(qweightData), server.MediaTypeImageTensor)
+	if err != nil {
+		return nil, err
+	}
+
+	// Create layer for scales
+	scalesLayer, err := server.NewLayer(bytes.NewReader(scalesData), server.MediaTypeImageTensor)
+	if err != nil {
+		return nil, err
+	}
+
+	layers := []create.LayerInfo{
+		{
+			Digest:    weightLayer.Digest,
+			Size:      weightLayer.Size,
+			MediaType: weightLayer.MediaType,
+			Name:      name,
+		},
+		{
+			Digest:    scalesLayer.Digest,
+			Size:      scalesLayer.Size,
+			MediaType: scalesLayer.MediaType,
+			Name:      name + "_scale",
+		},
+	}
+
+	// Add qbiases layer if present (affine mode)
+	if qbiasData != nil {
+		qbiasLayer, err := server.NewLayer(bytes.NewReader(qbiasData), server.MediaTypeImageTensor)
+		if err != nil {
+			return nil, err
+		}
+		layers = append(layers, create.LayerInfo{
+			Digest:    qbiasLayer.Digest,
+			Size:      qbiasLayer.Size,
+			MediaType: qbiasLayer.MediaType,
+			Name:      name + "_qbias",
+		})
+	}
+
+	return layers, nil
+}
+
+// createUnquantizedLayer creates a single tensor layer without quantization.
+func createUnquantizedLayer(r io.Reader, name string) ([]create.LayerInfo, error) {
+	layer, err := server.NewLayer(r, server.MediaTypeImageTensor)
+	if err != nil {
+		return nil, err
+	}
+
+	return []create.LayerInfo{
+		{
+			Digest:    layer.Digest,
+			Size:      layer.Size,
+			MediaType: layer.MediaType,
+			Name:      name,
+		},
+	}, nil
+}
+
+// newManifestWriter returns a ManifestWriter callback for writing the model manifest.
+func newManifestWriter(opts CreateOptions, capabilities []string) create.ManifestWriter {
+	return func(modelName string, config create.LayerInfo, layers []create.LayerInfo) error {
+		name := model.ParseName(modelName)
+		if !name.IsValid() {
+			return fmt.Errorf("invalid model name: %s", modelName)
+		}
+
+		// Create config blob with version requirement
+		configData := model.ConfigV2{
+			ModelFormat:  "safetensors",
+			Capabilities: capabilities,
+			Requires:     MinOllamaVersion,
+		}
+		configJSON, err := json.Marshal(configData)
+		if err != nil {
+			return fmt.Errorf("failed to marshal config: %w", err)
+		}
+
+		// Create config layer blob
+		configLayer, err := server.NewLayer(bytes.NewReader(configJSON), "application/vnd.docker.container.image.v1+json")
+		if err != nil {
+			return fmt.Errorf("failed to create config layer: %w", err)
+		}
+
+		// Convert LayerInfo to server.Layer
+		serverLayers := make([]server.Layer, 0, len(layers))
+		for _, l := range layers {
+			serverLayers = append(serverLayers, server.Layer{
+				MediaType: l.MediaType,
+				Digest:    l.Digest,
+				Size:      l.Size,
+				Name:      l.Name,
+			})
+		}
+
+		// Add Modelfile layers if present
+		if opts.Modelfile != nil {
+			modelfileLayers, err := createModelfileLayers(opts.Modelfile)
+			if err != nil {
+				return err
+			}
+			serverLayers = append(serverLayers, modelfileLayers...)
+		}
+
+		return server.WriteManifest(name, configLayer, serverLayers)
+	}
+}
+
+// createModelfileLayers creates layers for template, system, and license from Modelfile config.
+func createModelfileLayers(mf *ModelfileConfig) ([]server.Layer, error) {
+	var layers []server.Layer
+
+	if mf.Template != "" {
+		layer, err := server.NewLayer(bytes.NewReader([]byte(mf.Template)), "application/vnd.ollama.image.template")
+		if err != nil {
+			return nil, fmt.Errorf("failed to create template layer: %w", err)
+		}
+		layers = append(layers, layer)
+	}
+
+	if mf.System != "" {
+		layer, err := server.NewLayer(bytes.NewReader([]byte(mf.System)), "application/vnd.ollama.image.system")
+		if err != nil {
+			return nil, fmt.Errorf("failed to create system layer: %w", err)
+		}
+		layers = append(layers, layer)
+	}
+
+	if mf.License != "" {
+		layer, err := server.NewLayer(bytes.NewReader([]byte(mf.License)), "application/vnd.ollama.image.license")
+		if err != nil {
+			return nil, fmt.Errorf("failed to create license layer: %w", err)
+		}
+		layers = append(layers, layer)
+	}
+
+	return layers, nil
+}

x/create/client/create_test.go

@@ -0,0 +1,146 @@
+package client
+
+import (
+	"testing"
+)
+
+func TestModelfileConfig(t *testing.T) {
+	// Test that ModelfileConfig struct works as expected
+	config := &ModelfileConfig{
+		Template: "{{ .Prompt }}",
+		System:   "You are a helpful assistant.",
+		License:  "MIT",
+	}
+
+	if config.Template != "{{ .Prompt }}" {
+		t.Errorf("Template = %q, want %q", config.Template, "{{ .Prompt }}")
+	}
+	if config.System != "You are a helpful assistant." {
+		t.Errorf("System = %q, want %q", config.System, "You are a helpful assistant.")
+	}
+	if config.License != "MIT" {
+		t.Errorf("License = %q, want %q", config.License, "MIT")
+	}
+}
+
+func TestModelfileConfig_Empty(t *testing.T) {
+	config := &ModelfileConfig{}
+
+	if config.Template != "" {
+		t.Errorf("Template should be empty, got %q", config.Template)
+	}
+	if config.System != "" {
+		t.Errorf("System should be empty, got %q", config.System)
+	}
+	if config.License != "" {
+		t.Errorf("License should be empty, got %q", config.License)
+	}
+}
+
+func TestModelfileConfig_PartialFields(t *testing.T) {
+	// Test config with only some fields set
+	config := &ModelfileConfig{
+		Template: "{{ .Prompt }}",
+		// System and License intentionally empty
+	}
+
+	if config.Template == "" {
+		t.Error("Template should not be empty")
+	}
+	if config.System != "" {
+		t.Error("System should be empty")
+	}
+	if config.License != "" {
+		t.Error("License should be empty")
+	}
+}
+
+func TestMinOllamaVersion(t *testing.T) {
+	// Verify the minimum version constant is set
+	if MinOllamaVersion == "" {
+		t.Error("MinOllamaVersion should not be empty")
+	}
+	if MinOllamaVersion != "0.14.0" {
+		t.Errorf("MinOllamaVersion = %q, want %q", MinOllamaVersion, "0.14.0")
+	}
+}
+
+func TestCreateModel_InvalidDir(t *testing.T) {
+	// Test that CreateModel returns error for invalid directory
+	err := CreateModel(CreateOptions{
+		ModelName: "test-model",
+		ModelDir:  "/nonexistent/path",
+	}, nil)
+	if err == nil {
+		t.Error("expected error for nonexistent directory, got nil")
+	}
+}
+
+func TestCreateModel_NotSafetensorsDir(t *testing.T) {
+	// Test that CreateModel returns error for directory without safetensors
+	dir := t.TempDir()
+
+	err := CreateModel(CreateOptions{
+		ModelName: "test-model",
+		ModelDir:  dir,
+	}, nil)
+	if err == nil {
+		t.Error("expected error for empty directory, got nil")
+	}
+}
+
+func TestCreateOptions(t *testing.T) {
+	opts := CreateOptions{
+		ModelName: "my-model",
+		ModelDir:  "/path/to/model",
+		Quantize:  "fp8",
+		Modelfile: &ModelfileConfig{
+			Template: "test",
+			System:   "system",
+			License:  "MIT",
+		},
+	}
+
+	if opts.ModelName != "my-model" {
+		t.Errorf("ModelName = %q, want %q", opts.ModelName, "my-model")
+	}
+	if opts.ModelDir != "/path/to/model" {
+		t.Errorf("ModelDir = %q, want %q", opts.ModelDir, "/path/to/model")
+	}
+	if opts.Quantize != "fp8" {
+		t.Errorf("Quantize = %q, want %q", opts.Quantize, "fp8")
+	}
+	if opts.Modelfile == nil {
+		t.Error("Modelfile should not be nil")
+	}
+	if opts.Modelfile.Template != "test" {
+		t.Errorf("Modelfile.Template = %q, want %q", opts.Modelfile.Template, "test")
+	}
+}
+
+func TestCreateOptions_Defaults(t *testing.T) {
+	opts := CreateOptions{
+		ModelName: "test",
+		ModelDir:  "/tmp",
+	}
+
+	// Quantize should default to empty
+	if opts.Quantize != "" {
+		t.Errorf("Quantize should be empty by default, got %q", opts.Quantize)
+	}
+
+	// Modelfile should default to nil
+	if opts.Modelfile != nil {
+		t.Error("Modelfile should be nil by default")
+	}
+}
+
+func TestQuantizeSupported(t *testing.T) {
+	// This just verifies the function exists and returns a boolean
+	// The actual value depends on build tags (mlx vs non-mlx)
+	supported := QuantizeSupported()
+
+	// In non-mlx builds, this should be false
+	// We can't easily test both cases, so just verify it returns something
+	_ = supported
+}

x/create/client/quantize.go

@@ -11,10 +11,11 @@ import (
 	"github.com/ollama/ollama/x/imagegen/mlx"
 )
 
-// quantizeTensor loads a tensor from safetensors format, quantizes it to affine int8,
+// quantizeTensor loads a tensor from safetensors format, quantizes it,
 // and returns safetensors data for the quantized weights, scales, and biases.
+// Supported quantization types: "fp8" (affine 8-bit)
 // Uses MLX's native SaveSafetensors to ensure correct dtype handling (especially uint32 for quantized weights).
-func quantizeTensor(r io.Reader, name, dtype string, shape []int32) (qweightData, scalesData, qbiasData []byte, qweightShape, scalesShape, qbiasShape []int32, err error) {
+func quantizeTensor(r io.Reader, name, dtype string, shape []int32, quantize string) (qweightData, scalesData, qbiasData []byte, qweightShape, scalesShape, qbiasShape []int32, err error) {
 	tmpDir := ensureTempDir()
 
 	// Read safetensors data to a temp file (LoadSafetensorsNative needs a path)
@@ -50,9 +51,18 @@ func quantizeTensor(r io.Reader, name, dtype string, shape []int32) (qweightData
 		mlx.Eval(arr)
 	}
 
-	// Quantize with affine mode: group_size=32, bits=8
-	// Note: mxfp8 mode doesn't have matmul kernels in MLX, affine mode does
-	qweight, scales, qbiases := mlx.Quantize(arr, 32, 8, "affine")
+	// Quantize based on quantization type
+	var qweight, scales, qbiases *mlx.Array
+	switch quantize {
+	case "fp4":
+		// affine mode: group_size=32, bits=4
+		qweight, scales, qbiases = mlx.Quantize(arr, 32, 4, "affine")
+	case "fp8":
+		// affine mode: group_size=32, bits=8
+		qweight, scales, qbiases = mlx.Quantize(arr, 32, 8, "affine")
+	default:
+		return nil, nil, nil, nil, nil, nil, fmt.Errorf("unsupported quantization type: %s", quantize)
+	}
 
 	// Eval and make contiguous for data access
 	qweight = mlx.Contiguous(qweight)

x/create/client/quantize_stub.go

@@ -8,7 +8,7 @@ import (
 )
 
 // quantizeTensor is not available without MLX
-func quantizeTensor(r io.Reader, name, dtype string, shape []int32) (qweightData, scalesData, qbiasData []byte, qweightShape, scalesShape, qbiasShape []int32, err error) {
+func quantizeTensor(r io.Reader, name, dtype string, shape []int32, quantize string) (qweightData, scalesData, qbiasData []byte, qweightShape, scalesShape, qbiasShape []int32, err error) {
 	return nil, nil, nil, nil, nil, nil, fmt.Errorf("quantization requires MLX support (build with mlx tag)")
 }
 

x/create/create.go

@@ -0,0 +1,399 @@
+package create
+
+import (
+	"encoding/json"
+	"fmt"
+	"io"
+	"os"
+	"path/filepath"
+	"slices"
+	"strings"
+
+	"github.com/ollama/ollama/envconfig"
+	"github.com/ollama/ollama/x/imagegen/safetensors"
+)
+
+// ModelConfig represents the config blob stored with a model.
+type ModelConfig struct {
+	ModelFormat  string   `json:"model_format"`
+	Capabilities []string `json:"capabilities"`
+}
+
+// Manifest represents the manifest JSON structure.
+type Manifest struct {
+	SchemaVersion int             `json:"schemaVersion"`
+	MediaType     string          `json:"mediaType"`
+	Config        ManifestLayer   `json:"config"`
+	Layers        []ManifestLayer `json:"layers"`
+}
+
+// ManifestLayer represents a layer in the manifest.
+type ManifestLayer struct {
+	MediaType string `json:"mediaType"`
+	Digest    string `json:"digest"`
+	Size      int64  `json:"size"`
+	Name      string `json:"name,omitempty"`
+}
+
+// defaultManifestDir returns the manifest storage directory.
+func defaultManifestDir() string {
+	return filepath.Join(envconfig.Models(), "manifests")
+}
+
+// defaultBlobDir returns the blob storage directory.
+func defaultBlobDir() string {
+	return filepath.Join(envconfig.Models(), "blobs")
+}
+
+// resolveManifestPath converts a model name to a manifest file path.
+func resolveManifestPath(modelName string) string {
+	host := "registry.ollama.ai"
+	namespace := "library"
+	name := modelName
+	tag := "latest"
+
+	if idx := strings.LastIndex(name, ":"); idx != -1 {
+		tag = name[idx+1:]
+		name = name[:idx]
+	}
+
+	parts := strings.Split(name, "/")
+	switch len(parts) {
+	case 3:
+		host = parts[0]
+		namespace = parts[1]
+		name = parts[2]
+	case 2:
+		namespace = parts[0]
+		name = parts[1]
+	}
+
+	return filepath.Join(defaultManifestDir(), host, namespace, name, tag)
+}
+
+// loadManifest loads a manifest for the given model name.
+func loadManifest(modelName string) (*Manifest, error) {
+	manifestPath := resolveManifestPath(modelName)
+
+	data, err := os.ReadFile(manifestPath)
+	if err != nil {
+		return nil, err
+	}
+
+	var manifest Manifest
+	if err := json.Unmarshal(data, &manifest); err != nil {
+		return nil, err
+	}
+
+	return &manifest, nil
+}
+
+// loadModelConfig loads the config blob for a model.
+func loadModelConfig(modelName string) (*ModelConfig, error) {
+	manifest, err := loadManifest(modelName)
+	if err != nil {
+		return nil, err
+	}
+
+	// Read the config blob
+	blobName := strings.Replace(manifest.Config.Digest, ":", "-", 1)
+	blobPath := filepath.Join(defaultBlobDir(), blobName)
+
+	data, err := os.ReadFile(blobPath)
+	if err != nil {
+		return nil, err
+	}
+
+	var config ModelConfig
+	if err := json.Unmarshal(data, &config); err != nil {
+		return nil, err
+	}
+
+	return &config, nil
+}
+
+// IsSafetensorsModel checks if a model was created with the experimental
+// safetensors builder by checking the model format in the config.
+func IsSafetensorsModel(modelName string) bool {
+	config, err := loadModelConfig(modelName)
+	if err != nil {
+		return false
+	}
+	return config.ModelFormat == "safetensors"
+}
+
+// IsSafetensorsLLMModel checks if a model is a safetensors LLM model
+// (has completion capability, not image generation).
+func IsSafetensorsLLMModel(modelName string) bool {
+	config, err := loadModelConfig(modelName)
+	if err != nil {
+		return false
+	}
+	return config.ModelFormat == "safetensors" && slices.Contains(config.Capabilities, "completion")
+}
+
+// IsImageGenModel checks if a model is an image generation model
+// (has image capability).
+func IsImageGenModel(modelName string) bool {
+	config, err := loadModelConfig(modelName)
+	if err != nil {
+		return false
+	}
+	return config.ModelFormat == "safetensors" && slices.Contains(config.Capabilities, "image")
+}
+
+// GetModelArchitecture returns the architecture from the model's config.json layer.
+func GetModelArchitecture(modelName string) (string, error) {
+	manifest, err := loadManifest(modelName)
+	if err != nil {
+		return "", err
+	}
+
+	// Find the config.json layer
+	for _, layer := range manifest.Layers {
+		if layer.Name == "config.json" && layer.MediaType == "application/vnd.ollama.image.json" {
+			blobName := strings.Replace(layer.Digest, ":", "-", 1)
+			blobPath := filepath.Join(defaultBlobDir(), blobName)
+
+			data, err := os.ReadFile(blobPath)
+			if err != nil {
+				return "", err
+			}
+
+			var cfg struct {
+				Architectures []string `json:"architectures"`
+				ModelType     string   `json:"model_type"`
+			}
+			if err := json.Unmarshal(data, &cfg); err != nil {
+				return "", err
+			}
+
+			// Prefer model_type, fall back to first architecture
+			if cfg.ModelType != "" {
+				return cfg.ModelType, nil
+			}
+			if len(cfg.Architectures) > 0 {
+				return cfg.Architectures[0], nil
+			}
+		}
+	}
+
+	return "", fmt.Errorf("architecture not found in model config")
+}
+
+// IsTensorModelDir checks if the directory contains a diffusers-style tensor model
+// by looking for model_index.json, which is the standard diffusers pipeline config.
+func IsTensorModelDir(dir string) bool {
+	_, err := os.Stat(filepath.Join(dir, "model_index.json"))
+	return err == nil
+}
+
+// IsSafetensorsModelDir checks if the directory contains a standard safetensors model
+// by looking for config.json and at least one .safetensors file.
+func IsSafetensorsModelDir(dir string) bool {
+	// Must have config.json
+	if _, err := os.Stat(filepath.Join(dir, "config.json")); err != nil {
+		return false
+	}
+
+	// Must have at least one .safetensors file
+	entries, err := os.ReadDir(dir)
+	if err != nil {
+		return false
+	}
+
+	for _, entry := range entries {
+		if strings.HasSuffix(entry.Name(), ".safetensors") {
+			return true
+		}
+	}
+
+	return false
+}
+
+// LayerInfo holds metadata for a created layer.
+type LayerInfo struct {
+	Digest    string
+	Size      int64
+	MediaType string
+	Name      string // Path-style name: "component/tensor" or "path/to/config.json"
+}
+
+// LayerCreator is called to create a blob layer.
+// name is the path-style name (e.g., "tokenizer/tokenizer.json")
+type LayerCreator func(r io.Reader, mediaType, name string) (LayerInfo, error)
+
+// TensorLayerCreator creates a tensor blob layer with metadata.
+// name is the path-style name including component (e.g., "text_encoder/model.embed_tokens.weight")
+type TensorLayerCreator func(r io.Reader, name, dtype string, shape []int32) (LayerInfo, error)
+
+// QuantizingTensorLayerCreator creates tensor layers with optional quantization.
+// When quantize is non-empty (e.g., "fp8"), returns multiple layers (weight + scales + biases).
+type QuantizingTensorLayerCreator func(r io.Reader, name, dtype string, shape []int32, quantize string) ([]LayerInfo, error)
+
+// ManifestWriter writes the manifest file.
+type ManifestWriter func(modelName string, config LayerInfo, layers []LayerInfo) error
+
+// ShouldQuantize returns true if a tensor should be quantized.
+// For image gen models (component non-empty): quantizes linear weights, skipping VAE, embeddings, norms.
+// For LLM models (component empty): quantizes linear weights, skipping embeddings, norms, and small tensors.
+func ShouldQuantize(name, component string) bool {
+	// Image gen specific: skip VAE entirely
+	if component == "vae" {
+		return false
+	}
+
+	// Skip embeddings
+	if strings.Contains(name, "embed") {
+		return false
+	}
+
+	// Skip layer norms and RMS norms
+	if strings.Contains(name, "norm") || strings.Contains(name, "ln_") || strings.Contains(name, "layernorm") {
+		return false
+	}
+
+	// Skip biases
+	if strings.HasSuffix(name, ".bias") {
+		return false
+	}
+
+	// Only quantize weights
+	return strings.HasSuffix(name, ".weight")
+}
+
+// ShouldQuantizeTensor returns true if a tensor should be quantized based on name and shape.
+// This is a more detailed check that also considers tensor dimensions.
+func ShouldQuantizeTensor(name string, shape []int32) bool {
+	// Use basic name-based check first
+	if !ShouldQuantize(name, "") {
+		return false
+	}
+
+	// Only quantize 2D tensors (linear layers) - skip 1D (biases, norms) and higher-D (convolutions if any)
+	if len(shape) != 2 {
+		return false
+	}
+
+	// Skip small tensors (less than 1024 elements) - not worth quantizing
+	if len(shape) >= 2 && int64(shape[0])*int64(shape[1]) < 1024 {
+		return false
+	}
+
+	// MLX quantization requires last dimension to be divisible by group size (32)
+	if shape[len(shape)-1]%32 != 0 {
+		return false
+	}
+
+	return true
+}
+
+// CreateSafetensorsModel imports a standard safetensors model from a directory.
+// This handles Hugging Face style models with config.json and *.safetensors files.
+// Stores each tensor as a separate blob for fine-grained deduplication.
+// If quantize is non-empty (e.g., "fp8"), eligible tensors will be quantized.
+func CreateSafetensorsModel(modelName, modelDir, quantize string, createLayer LayerCreator, createTensorLayer QuantizingTensorLayerCreator, writeManifest ManifestWriter, fn func(status string)) error {
+	var layers []LayerInfo
+	var configLayer LayerInfo
+
+	entries, err := os.ReadDir(modelDir)
+	if err != nil {
+		return fmt.Errorf("failed to read directory: %w", err)
+	}
+
+	// Process all safetensors files
+	for _, entry := range entries {
+		if entry.IsDir() || !strings.HasSuffix(entry.Name(), ".safetensors") {
+			continue
+		}
+
+		stPath := filepath.Join(modelDir, entry.Name())
+
+		// Extract individual tensors from safetensors file
+		extractor, err := safetensors.OpenForExtraction(stPath)
+		if err != nil {
+			return fmt.Errorf("failed to open %s: %w", stPath, err)
+		}
+
+		tensorNames := extractor.ListTensors()
+		quantizeMsg := ""
+		if quantize != "" {
+			quantizeMsg = fmt.Sprintf(", quantizing to %s", quantize)
+		}
+		fn(fmt.Sprintf("importing %s (%d tensors%s)", entry.Name(), len(tensorNames), quantizeMsg))
+
+		for _, tensorName := range tensorNames {
+			td, err := extractor.GetTensor(tensorName)
+			if err != nil {
+				extractor.Close()
+				return fmt.Errorf("failed to get tensor %s: %w", tensorName, err)
+			}
+
+			// Determine quantization type for this tensor (empty string if not quantizing)
+			quantizeType := ""
+			if quantize != "" && ShouldQuantizeTensor(tensorName, td.Shape) {
+				quantizeType = quantize
+			}
+
+			// Store as minimal safetensors format (88 bytes header overhead)
+			// This enables native mmap loading via mlx_load_safetensors
+			// createTensorLayer returns multiple layers if quantizing (weight + scales)
+			newLayers, err := createTensorLayer(td.SafetensorsReader(), tensorName, td.Dtype, td.Shape, quantizeType)
+			if err != nil {
+				extractor.Close()
+				return fmt.Errorf("failed to create layer for %s: %w", tensorName, err)
+			}
+			layers = append(layers, newLayers...)
+		}
+
+		extractor.Close()
+	}
+
+	// Process all JSON config files
+	for _, entry := range entries {
+		if entry.IsDir() || !strings.HasSuffix(entry.Name(), ".json") {
+			continue
+		}
+
+		// Skip the index file as we don't need it after extraction
+		if entry.Name() == "model.safetensors.index.json" {
+			continue
+		}
+
+		cfgPath := entry.Name()
+		fullPath := filepath.Join(modelDir, cfgPath)
+
+		fn(fmt.Sprintf("importing config %s", cfgPath))
+
+		f, err := os.Open(fullPath)
+		if err != nil {
+			return fmt.Errorf("failed to open %s: %w", cfgPath, err)
+		}
+
+		layer, err := createLayer(f, "application/vnd.ollama.image.json", cfgPath)
+		f.Close()
+		if err != nil {
+			return fmt.Errorf("failed to create layer for %s: %w", cfgPath, err)
+		}
+
+		// Use config.json as the config layer
+		if cfgPath == "config.json" {
+			configLayer = layer
+		}
+
+		layers = append(layers, layer)
+	}
+
+	if configLayer.Digest == "" {
+		return fmt.Errorf("config.json not found in %s", modelDir)
+	}
+
+	fn(fmt.Sprintf("writing manifest for %s", modelName))
+
+	if err := writeManifest(modelName, configLayer, layers); err != nil {
+		return fmt.Errorf("failed to write manifest: %w", err)
+	}
+
+	fn(fmt.Sprintf("successfully imported %s with %d layers", modelName, len(layers)))
+	return nil
+}

x/create/create_test.go

@@ -0,0 +1,752 @@
+package create
+
+import (
+	"bytes"
+	"encoding/binary"
+	"encoding/json"
+	"io"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+func TestIsTensorModelDir(t *testing.T) {
+	tests := []struct {
+		name     string
+		setup    func(dir string) error
+		expected bool
+	}{
+		{
+			name: "valid diffusers model with model_index.json",
+			setup: func(dir string) error {
+				return os.WriteFile(filepath.Join(dir, "model_index.json"), []byte(`{"_class_name": "FluxPipeline"}`), 0o644)
+			},
+			expected: true,
+		},
+		{
+			name: "empty directory",
+			setup: func(dir string) error {
+				return nil
+			},
+			expected: false,
+		},
+		{
+			name: "directory with other files but no model_index.json",
+			setup: func(dir string) error {
+				return os.WriteFile(filepath.Join(dir, "config.json"), []byte(`{}`), 0o644)
+			},
+			expected: false,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			dir := t.TempDir()
+			if err := tt.setup(dir); err != nil {
+				t.Fatalf("setup failed: %v", err)
+			}
+
+			got := IsTensorModelDir(dir)
+			if got != tt.expected {
+				t.Errorf("IsTensorModelDir() = %v, want %v", got, tt.expected)
+			}
+		})
+	}
+}
+
+func TestIsSafetensorsModelDir(t *testing.T) {
+	tests := []struct {
+		name     string
+		setup    func(dir string) error
+		expected bool
+	}{
+		{
+			name: "valid safetensors model with config.json and .safetensors file",
+			setup: func(dir string) error {
+				if err := os.WriteFile(filepath.Join(dir, "config.json"), []byte(`{"model_type": "gemma3"}`), 0o644); err != nil {
+					return err
+				}
+				return os.WriteFile(filepath.Join(dir, "model.safetensors"), []byte("dummy"), 0o644)
+			},
+			expected: true,
+		},
+		{
+			name: "config.json only, no safetensors files",
+			setup: func(dir string) error {
+				return os.WriteFile(filepath.Join(dir, "config.json"), []byte(`{}`), 0o644)
+			},
+			expected: false,
+		},
+		{
+			name: "safetensors file only, no config.json",
+			setup: func(dir string) error {
+				return os.WriteFile(filepath.Join(dir, "model.safetensors"), []byte("dummy"), 0o644)
+			},
+			expected: false,
+		},
+		{
+			name: "empty directory",
+			setup: func(dir string) error {
+				return nil
+			},
+			expected: false,
+		},
+		{
+			name: "multiple safetensors files with config.json",
+			setup: func(dir string) error {
+				if err := os.WriteFile(filepath.Join(dir, "config.json"), []byte(`{}`), 0o644); err != nil {
+					return err
+				}
+				if err := os.WriteFile(filepath.Join(dir, "model-00001-of-00002.safetensors"), []byte("dummy"), 0o644); err != nil {
+					return err
+				}
+				return os.WriteFile(filepath.Join(dir, "model-00002-of-00002.safetensors"), []byte("dummy"), 0o644)
+			},
+			expected: true,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			dir := t.TempDir()
+			if err := tt.setup(dir); err != nil {
+				t.Fatalf("setup failed: %v", err)
+			}
+
+			got := IsSafetensorsModelDir(dir)
+			if got != tt.expected {
+				t.Errorf("IsSafetensorsModelDir() = %v, want %v", got, tt.expected)
+			}
+		})
+	}
+}
+
+func TestIsSafetensorsModelDir_NonexistentDir(t *testing.T) {
+	got := IsSafetensorsModelDir("/nonexistent/path/that/does/not/exist")
+	if got != false {
+		t.Errorf("IsSafetensorsModelDir() = %v for nonexistent dir, want false", got)
+	}
+}
+
+// createMinimalSafetensors creates a minimal valid safetensors file with one tensor
+func createMinimalSafetensors(t *testing.T, path string) {
+	t.Helper()
+
+	// Create a minimal safetensors file with a single float32 tensor
+	header := map[string]interface{}{
+		"test_tensor": map[string]interface{}{
+			"dtype":        "F32",
+			"shape":        []int{2, 2},
+			"data_offsets": []int{0, 16}, // 4 float32 values = 16 bytes
+		},
+	}
+	headerJSON, err := json.Marshal(header)
+	if err != nil {
+		t.Fatalf("failed to marshal header: %v", err)
+	}
+
+	// Pad header to 8-byte alignment
+	padding := (8 - len(headerJSON)%8) % 8
+	headerJSON = append(headerJSON, bytes.Repeat([]byte(" "), padding)...)
+
+	// Write file
+	f, err := os.Create(path)
+	if err != nil {
+		t.Fatalf("failed to create file: %v", err)
+	}
+	defer f.Close()
+
+	// Write header size (8 bytes, little endian)
+	if err := binary.Write(f, binary.LittleEndian, uint64(len(headerJSON))); err != nil {
+		t.Fatalf("failed to write header size: %v", err)
+	}
+
+	// Write header
+	if _, err := f.Write(headerJSON); err != nil {
+		t.Fatalf("failed to write header: %v", err)
+	}
+
+	// Write tensor data (16 bytes of zeros for 4 float32 values)
+	if _, err := f.Write(make([]byte, 16)); err != nil {
+		t.Fatalf("failed to write tensor data: %v", err)
+	}
+}
+
+func TestCreateSafetensorsModel(t *testing.T) {
+	dir := t.TempDir()
+
+	// Create config.json
+	configJSON := `{"model_type": "test", "architectures": ["TestModel"]}`
+	if err := os.WriteFile(filepath.Join(dir, "config.json"), []byte(configJSON), 0o644); err != nil {
+		t.Fatalf("failed to write config.json: %v", err)
+	}
+
+	// Create a minimal safetensors file
+	createMinimalSafetensors(t, filepath.Join(dir, "model.safetensors"))
+
+	// Track what was created
+	var createdLayers []LayerInfo
+	var manifestWritten bool
+	var manifestModelName string
+	var manifestConfigLayer LayerInfo
+	var manifestLayers []LayerInfo
+	var statusMessages []string
+
+	// Mock callbacks
+	createLayer := func(r io.Reader, mediaType, name string) (LayerInfo, error) {
+		data, err := io.ReadAll(r)
+		if err != nil {
+			return LayerInfo{}, err
+		}
+		layer := LayerInfo{
+			Digest:    "sha256:test",
+			Size:      int64(len(data)),
+			MediaType: mediaType,
+			Name:      name,
+		}
+		createdLayers = append(createdLayers, layer)
+		return layer, nil
+	}
+
+	createTensorLayer := func(r io.Reader, name, dtype string, shape []int32, quantize string) ([]LayerInfo, error) {
+		data, err := io.ReadAll(r)
+		if err != nil {
+			return nil, err
+		}
+		layer := LayerInfo{
+			Digest:    "sha256:tensor_" + name,
+			Size:      int64(len(data)),
+			MediaType: "application/vnd.ollama.image.tensor",
+			Name:      name,
+		}
+		createdLayers = append(createdLayers, layer)
+		return []LayerInfo{layer}, nil
+	}
+
+	writeManifest := func(modelName string, config LayerInfo, layers []LayerInfo) error {
+		manifestWritten = true
+		manifestModelName = modelName
+		manifestConfigLayer = config
+		manifestLayers = layers
+		return nil
+	}
+
+	progressFn := func(status string) {
+		statusMessages = append(statusMessages, status)
+	}
+
+	// Run CreateSafetensorsModel
+	err := CreateSafetensorsModel("test-model", dir, "", createLayer, createTensorLayer, writeManifest, progressFn)
+	if err != nil {
+		t.Fatalf("CreateSafetensorsModel failed: %v", err)
+	}
+
+	// Verify manifest was written
+	if !manifestWritten {
+		t.Error("manifest was not written")
+	}
+
+	if manifestModelName != "test-model" {
+		t.Errorf("manifest model name = %q, want %q", manifestModelName, "test-model")
+	}
+
+	// Verify config layer was set
+	if manifestConfigLayer.Name != "config.json" {
+		t.Errorf("config layer name = %q, want %q", manifestConfigLayer.Name, "config.json")
+	}
+
+	// Verify we have at least one tensor and one config layer
+	hasTensor := false
+	hasConfig := false
+	for _, layer := range manifestLayers {
+		if layer.Name == "test_tensor" {
+			hasTensor = true
+		}
+		if layer.Name == "config.json" {
+			hasConfig = true
+		}
+	}
+
+	if !hasTensor {
+		t.Error("no tensor layer found in manifest")
+	}
+	if !hasConfig {
+		t.Error("no config layer found in manifest")
+	}
+
+	// Verify status messages were sent
+	if len(statusMessages) == 0 {
+		t.Error("no status messages received")
+	}
+}
+
+func TestCreateSafetensorsModel_NoConfigJson(t *testing.T) {
+	dir := t.TempDir()
+
+	// Create only a safetensors file, no config.json
+	createMinimalSafetensors(t, filepath.Join(dir, "model.safetensors"))
+
+	// Mock callbacks (minimal)
+	createLayer := func(r io.Reader, mediaType, name string) (LayerInfo, error) {
+		io.ReadAll(r)
+		return LayerInfo{Name: name}, nil
+	}
+	createTensorLayer := func(r io.Reader, name, dtype string, shape []int32, quantize string) ([]LayerInfo, error) {
+		io.ReadAll(r)
+		return []LayerInfo{{Name: name}}, nil
+	}
+	writeManifest := func(modelName string, config LayerInfo, layers []LayerInfo) error {
+		return nil
+	}
+	progressFn := func(status string) {}
+
+	err := CreateSafetensorsModel("test-model", dir, "", createLayer, createTensorLayer, writeManifest, progressFn)
+	if err == nil {
+		t.Error("expected error for missing config.json, got nil")
+	}
+}
+
+func TestCreateSafetensorsModel_EmptyDir(t *testing.T) {
+	dir := t.TempDir()
+
+	// Mock callbacks
+	createLayer := func(r io.Reader, mediaType, name string) (LayerInfo, error) {
+		return LayerInfo{}, nil
+	}
+	createTensorLayer := func(r io.Reader, name, dtype string, shape []int32, quantize string) ([]LayerInfo, error) {
+		return []LayerInfo{{}}, nil
+	}
+	writeManifest := func(modelName string, config LayerInfo, layers []LayerInfo) error {
+		return nil
+	}
+	progressFn := func(status string) {}
+
+	err := CreateSafetensorsModel("test-model", dir, "", createLayer, createTensorLayer, writeManifest, progressFn)
+	if err == nil {
+		t.Error("expected error for empty directory, got nil")
+	}
+}
+
+func TestCreateSafetensorsModel_SkipsIndexJson(t *testing.T) {
+	dir := t.TempDir()
+
+	// Create config.json
+	if err := os.WriteFile(filepath.Join(dir, "config.json"), []byte(`{}`), 0o644); err != nil {
+		t.Fatalf("failed to write config.json: %v", err)
+	}
+
+	// Create model.safetensors.index.json (should be skipped)
+	indexJSON := `{"metadata": {"total_size": 100}, "weight_map": {}}`
+	if err := os.WriteFile(filepath.Join(dir, "model.safetensors.index.json"), []byte(indexJSON), 0o644); err != nil {
+		t.Fatalf("failed to write index.json: %v", err)
+	}
+
+	// Create a minimal safetensors file
+	createMinimalSafetensors(t, filepath.Join(dir, "model.safetensors"))
+
+	var configNames []string
+
+	createLayer := func(r io.Reader, mediaType, name string) (LayerInfo, error) {
+		io.ReadAll(r)
+		configNames = append(configNames, name)
+		return LayerInfo{Name: name, Digest: "sha256:test"}, nil
+	}
+	createTensorLayer := func(r io.Reader, name, dtype string, shape []int32, quantize string) ([]LayerInfo, error) {
+		io.ReadAll(r)
+		return []LayerInfo{{Name: name}}, nil
+	}
+	writeManifest := func(modelName string, config LayerInfo, layers []LayerInfo) error {
+		return nil
+	}
+	progressFn := func(status string) {}
+
+	err := CreateSafetensorsModel("test-model", dir, "", createLayer, createTensorLayer, writeManifest, progressFn)
+	if err != nil {
+		t.Fatalf("CreateSafetensorsModel failed: %v", err)
+	}
+
+	// Verify model.safetensors.index.json was not included
+	for _, name := range configNames {
+		if name == "model.safetensors.index.json" {
+			t.Error("model.safetensors.index.json should have been skipped")
+		}
+	}
+}
+
+func TestResolveManifestPath(t *testing.T) {
+	tests := []struct {
+		name      string
+		modelName string
+		wantParts []string // Parts that should appear in the path
+	}{
+		{
+			name:      "simple model name",
+			modelName: "llama2",
+			wantParts: []string{"registry.ollama.ai", "library", "llama2", "latest"},
+		},
+		{
+			name:      "model name with tag",
+			modelName: "llama2:7b",
+			wantParts: []string{"registry.ollama.ai", "library", "llama2", "7b"},
+		},
+		{
+			name:      "model name with namespace",
+			modelName: "myuser/mymodel",
+			wantParts: []string{"registry.ollama.ai", "myuser", "mymodel", "latest"},
+		},
+		{
+			name:      "model name with namespace and tag",
+			modelName: "myuser/mymodel:v1",
+			wantParts: []string{"registry.ollama.ai", "myuser", "mymodel", "v1"},
+		},
+		{
+			name:      "fully qualified model name",
+			modelName: "registry.example.com/namespace/model:tag",
+			wantParts: []string{"registry.example.com", "namespace", "model", "tag"},
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			got := resolveManifestPath(tt.modelName)
+
+			for _, part := range tt.wantParts {
+				if !strings.Contains(got, part) {
+					t.Errorf("resolveManifestPath(%q) = %q, missing part %q", tt.modelName, got, part)
+				}
+			}
+		})
+	}
+}
+
+func TestLayerInfo(t *testing.T) {
+	layer := LayerInfo{
+		Digest:    "sha256:abc123",
+		Size:      1024,
+		MediaType: "application/vnd.ollama.image.tensor",
+		Name:      "model.weight",
+	}
+
+	if layer.Digest != "sha256:abc123" {
+		t.Errorf("Digest = %q, want %q", layer.Digest, "sha256:abc123")
+	}
+	if layer.Size != 1024 {
+		t.Errorf("Size = %d, want %d", layer.Size, 1024)
+	}
+	if layer.MediaType != "application/vnd.ollama.image.tensor" {
+		t.Errorf("MediaType = %q, want %q", layer.MediaType, "application/vnd.ollama.image.tensor")
+	}
+	if layer.Name != "model.weight" {
+		t.Errorf("Name = %q, want %q", layer.Name, "model.weight")
+	}
+}
+
+func TestModelConfig(t *testing.T) {
+	config := ModelConfig{
+		ModelFormat:  "safetensors",
+		Capabilities: []string{"completion", "chat"},
+	}
+
+	if config.ModelFormat != "safetensors" {
+		t.Errorf("ModelFormat = %q, want %q", config.ModelFormat, "safetensors")
+	}
+	if len(config.Capabilities) != 2 {
+		t.Errorf("Capabilities length = %d, want %d", len(config.Capabilities), 2)
+	}
+}
+
+func TestManifest(t *testing.T) {
+	manifest := Manifest{
+		SchemaVersion: 2,
+		MediaType:     "application/vnd.oci.image.manifest.v1+json",
+		Config: ManifestLayer{
+			MediaType: "application/vnd.docker.container.image.v1+json",
+			Digest:    "sha256:config",
+			Size:      100,
+		},
+		Layers: []ManifestLayer{
+			{
+				MediaType: "application/vnd.ollama.image.tensor",
+				Digest:    "sha256:layer1",
+				Size:      1000,
+				Name:      "weight.bin",
+			},
+		},
+	}
+
+	if manifest.SchemaVersion != 2 {
+		t.Errorf("SchemaVersion = %d, want %d", manifest.SchemaVersion, 2)
+	}
+	if manifest.Config.Digest != "sha256:config" {
+		t.Errorf("Config.Digest = %q, want %q", manifest.Config.Digest, "sha256:config")
+	}
+	if len(manifest.Layers) != 1 {
+		t.Errorf("Layers length = %d, want %d", len(manifest.Layers), 1)
+	}
+	if manifest.Layers[0].Name != "weight.bin" {
+		t.Errorf("Layers[0].Name = %q, want %q", manifest.Layers[0].Name, "weight.bin")
+	}
+}
+
+func TestShouldQuantize(t *testing.T) {
+	tests := []struct {
+		name      string
+		tensor    string
+		component string
+		want      bool
+	}{
+		// VAE component should never be quantized
+		{"vae weight", "decoder.weight", "vae", false},
+		{"vae bias", "decoder.bias", "vae", false},
+
+		// Embeddings should not be quantized
+		{"embedding weight", "embed_tokens.weight", "", false},
+		{"embedding in name", "token_embedding.weight", "", false},
+
+		// Norms should not be quantized
+		{"layer norm", "layer_norm.weight", "", false},
+		{"rms norm", "rms_norm.weight", "", false},
+		{"ln prefix", "ln_1.weight", "", false},
+		{"layernorm in name", "input_layernorm.weight", "", false},
+
+		// Biases should not be quantized
+		{"bias tensor", "attention.bias", "", false},
+		{"proj bias", "o_proj.bias", "", false},
+
+		// Linear weights should be quantized
+		{"linear weight", "q_proj.weight", "", true},
+		{"attention weight", "self_attn.weight", "", true},
+		{"mlp weight", "mlp.gate_proj.weight", "", true},
+
+		// Transformer component weights should be quantized
+		{"transformer weight", "layers.0.weight", "transformer", true},
+		{"text_encoder weight", "encoder.weight", "text_encoder", true},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			got := ShouldQuantize(tt.tensor, tt.component)
+			if got != tt.want {
+				t.Errorf("ShouldQuantize(%q, %q) = %v, want %v", tt.tensor, tt.component, got, tt.want)
+			}
+		})
+	}
+}
+
+func TestShouldQuantizeTensor(t *testing.T) {
+	tests := []struct {
+		name   string
+		tensor string
+		shape  []int32
+		want   bool
+	}{
+		// 2D tensors with sufficient size should be quantized
+		{"large 2D weight", "q_proj.weight", []int32{4096, 4096}, true},
+		{"medium 2D weight", "small_proj.weight", []int32{128, 128}, true},
+
+		// Small tensors should not be quantized (< 1024 elements)
+		{"tiny 2D weight", "tiny.weight", []int32{16, 16}, false},
+		{"small 2D weight", "small.weight", []int32{31, 31}, false},
+
+		// 1D tensors should not be quantized
+		{"1D tensor", "layer_norm.weight", []int32{4096}, false},
+
+		// 3D+ tensors should not be quantized
+		{"3D tensor", "conv.weight", []int32{64, 64, 3}, false},
+		{"4D tensor", "conv2d.weight", []int32{64, 64, 3, 3}, false},
+
+		// Embeddings should not be quantized regardless of shape
+		{"embedding 2D", "embed_tokens.weight", []int32{32000, 4096}, false},
+
+		// Norms should not be quantized regardless of shape
+		{"norm 2D", "layer_norm.weight", []int32{4096, 1}, false},
+
+		// Biases should not be quantized
+		{"bias 2D", "proj.bias", []int32{4096, 1}, false},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			got := ShouldQuantizeTensor(tt.tensor, tt.shape)
+			if got != tt.want {
+				t.Errorf("ShouldQuantizeTensor(%q, %v) = %v, want %v", tt.tensor, tt.shape, got, tt.want)
+			}
+		})
+	}
+}
+
+func TestCreateSafetensorsModel_WithQuantize(t *testing.T) {
+	dir := t.TempDir()
+
+	// Create config.json
+	configJSON := `{"model_type": "test", "architectures": ["TestModel"]}`
+	if err := os.WriteFile(filepath.Join(dir, "config.json"), []byte(configJSON), 0o644); err != nil {
+		t.Fatalf("failed to write config.json: %v", err)
+	}
+
+	// Create a minimal safetensors file
+	createMinimalSafetensors(t, filepath.Join(dir, "model.safetensors"))
+
+	var quantizeRequested []string
+
+	createLayer := func(r io.Reader, mediaType, name string) (LayerInfo, error) {
+		io.ReadAll(r)
+		return LayerInfo{Name: name, Digest: "sha256:test"}, nil
+	}
+
+	createTensorLayer := func(r io.Reader, name, dtype string, shape []int32, quantize string) ([]LayerInfo, error) {
+		io.ReadAll(r)
+		quantizeRequested = append(quantizeRequested, quantize)
+		return []LayerInfo{{Name: name}}, nil
+	}
+
+	writeManifest := func(modelName string, config LayerInfo, layers []LayerInfo) error {
+		return nil
+	}
+
+	progressFn := func(status string) {}
+
+	// Run with quantize enabled
+	err := CreateSafetensorsModel("test-model", dir, "fp8", createLayer, createTensorLayer, writeManifest, progressFn)
+	if err != nil {
+		t.Fatalf("CreateSafetensorsModel failed: %v", err)
+	}
+
+	// Verify quantize was passed to callback (will be false for small test tensor)
+	if len(quantizeRequested) == 0 {
+		t.Error("no tensors processed")
+	}
+}
+
+// createMinimalImageGenModel creates a minimal diffusers-style model directory
+func createMinimalImageGenModel(t *testing.T, dir string) {
+	t.Helper()
+
+	// Create model_index.json
+	modelIndex := `{"_class_name": "FluxPipeline", "_diffusers_version": "0.30.0"}`
+	if err := os.WriteFile(filepath.Join(dir, "model_index.json"), []byte(modelIndex), 0o644); err != nil {
+		t.Fatalf("failed to write model_index.json: %v", err)
+	}
+
+	// Create transformer directory with a safetensors file
+	transformerDir := filepath.Join(dir, "transformer")
+	if err := os.MkdirAll(transformerDir, 0o755); err != nil {
+		t.Fatalf("failed to create transformer dir: %v", err)
+	}
+	createMinimalSafetensors(t, filepath.Join(transformerDir, "model.safetensors"))
+
+	// Create transformer config
+	transformerConfig := `{"hidden_size": 3072}`
+	if err := os.WriteFile(filepath.Join(transformerDir, "config.json"), []byte(transformerConfig), 0o644); err != nil {
+		t.Fatalf("failed to write transformer config: %v", err)
+	}
+}
+
+func TestCreateImageGenModel(t *testing.T) {
+	dir := t.TempDir()
+	createMinimalImageGenModel(t, dir)
+
+	var manifestWritten bool
+	var manifestModelName string
+	var statusMessages []string
+
+	createLayer := func(r io.Reader, mediaType, name string) (LayerInfo, error) {
+		io.ReadAll(r)
+		return LayerInfo{Name: name, Digest: "sha256:test"}, nil
+	}
+
+	createTensorLayer := func(r io.Reader, name, dtype string, shape []int32, quantize string) ([]LayerInfo, error) {
+		io.ReadAll(r)
+		return []LayerInfo{{Name: name, Digest: "sha256:tensor"}}, nil
+	}
+
+	writeManifest := func(modelName string, config LayerInfo, layers []LayerInfo) error {
+		manifestWritten = true
+		manifestModelName = modelName
+		return nil
+	}
+
+	progressFn := func(status string) {
+		statusMessages = append(statusMessages, status)
+	}
+
+	err := CreateImageGenModel("test-imagegen", dir, "", createLayer, createTensorLayer, writeManifest, progressFn)
+	if err != nil {
+		t.Fatalf("CreateImageGenModel failed: %v", err)
+	}
+
+	if !manifestWritten {
+		t.Error("manifest was not written")
+	}
+
+	if manifestModelName != "test-imagegen" {
+		t.Errorf("manifest model name = %q, want %q", manifestModelName, "test-imagegen")
+	}
+
+	if len(statusMessages) == 0 {
+		t.Error("no status messages received")
+	}
+}
+
+func TestCreateImageGenModel_NoModelIndex(t *testing.T) {
+	dir := t.TempDir()
+
+	// Create only transformer without model_index.json
+	transformerDir := filepath.Join(dir, "transformer")
+	if err := os.MkdirAll(transformerDir, 0o755); err != nil {
+		t.Fatalf("failed to create transformer dir: %v", err)
+	}
+	createMinimalSafetensors(t, filepath.Join(transformerDir, "model.safetensors"))
+
+	createLayer := func(r io.Reader, mediaType, name string) (LayerInfo, error) {
+		io.ReadAll(r)
+		return LayerInfo{Name: name}, nil
+	}
+	createTensorLayer := func(r io.Reader, name, dtype string, shape []int32, quantize string) ([]LayerInfo, error) {
+		io.ReadAll(r)
+		return []LayerInfo{{Name: name}}, nil
+	}
+	writeManifest := func(modelName string, config LayerInfo, layers []LayerInfo) error {
+		return nil
+	}
+	progressFn := func(status string) {}
+
+	err := CreateImageGenModel("test-imagegen", dir, "", createLayer, createTensorLayer, writeManifest, progressFn)
+	if err == nil {
+		t.Error("expected error for missing model_index.json, got nil")
+	}
+}
+
+func TestCreateImageGenModel_WithQuantize(t *testing.T) {
+	dir := t.TempDir()
+	createMinimalImageGenModel(t, dir)
+
+	var quantizeRequested []string
+
+	createLayer := func(r io.Reader, mediaType, name string) (LayerInfo, error) {
+		io.ReadAll(r)
+		return LayerInfo{Name: name, Digest: "sha256:test"}, nil
+	}
+
+	createTensorLayer := func(r io.Reader, name, dtype string, shape []int32, quantize string) ([]LayerInfo, error) {
+		io.ReadAll(r)
+		quantizeRequested = append(quantizeRequested, quantize)
+		return []LayerInfo{{Name: name}}, nil
+	}
+
+	writeManifest := func(modelName string, config LayerInfo, layers []LayerInfo) error {
+		return nil
+	}
+
+	progressFn := func(status string) {}
+
+	err := CreateImageGenModel("test-imagegen", dir, "fp8", createLayer, createTensorLayer, writeManifest, progressFn)
+	if err != nil {
+		t.Fatalf("CreateImageGenModel failed: %v", err)
+	}
+
+	if len(quantizeRequested) == 0 {
+		t.Error("no tensors processed")
+	}
+}

x/create/imagegen.go

@@ -1,4 +1,4 @@
-package imagegen
+package create
 
 import (
 	"bytes"
@@ -12,40 +12,24 @@ import (
 	"github.com/ollama/ollama/x/imagegen/safetensors"
 )
 
-// IsTensorModelDir checks if the directory contains a tensor model
-// by looking for model_index.json, which is the standard diffusers pipeline config.
-func IsTensorModelDir(dir string) bool {
-	_, err := os.Stat(filepath.Join(dir, "model_index.json"))
-	return err == nil
-}
-
-// LayerInfo holds metadata for a created layer.
-type LayerInfo struct {
-	Digest    string
-	Size      int64
-	MediaType string
-	Name      string // Path-style name: "component/tensor" or "path/to/config.json"
-}
-
-// LayerCreator is called to create a blob layer.
-// name is the path-style name (e.g., "tokenizer/tokenizer.json")
-type LayerCreator func(r io.Reader, mediaType, name string) (LayerInfo, error)
-
-// TensorLayerCreator creates a tensor blob layer with metadata.
-// name is the path-style name including component (e.g., "text_encoder/model.embed_tokens.weight")
-type TensorLayerCreator func(r io.Reader, name, dtype string, shape []int32) (LayerInfo, error)
-
-// ManifestWriter writes the manifest file.
-type ManifestWriter func(modelName string, config LayerInfo, layers []LayerInfo) error
-
-// CreateModel imports an image generation model from a directory.
+// CreateImageGenModel imports an image generation model from a directory.
 // Stores each tensor as a separate blob for fine-grained deduplication.
-// If quantize is "fp8", linear weights in transformer/text_encoder are quantized to mxfp8 format.
+// If quantize is specified, linear weights in transformer/text_encoder are quantized.
+// Supported quantization types: fp8 (or empty for no quantization).
 // Layer creation and manifest writing are done via callbacks to avoid import cycles.
-func CreateModel(modelName, modelDir, quantize string, createLayer LayerCreator, createTensorLayer QuantizingTensorLayerCreator, writeManifest ManifestWriter, fn func(status string)) error {
+func CreateImageGenModel(modelName, modelDir, quantize string, createLayer LayerCreator, createTensorLayer QuantizingTensorLayerCreator, writeManifest ManifestWriter, fn func(status string)) error {
+	// Validate quantization type
+	switch quantize {
+	case "", "fp4", "fp8":
+		// valid
+	default:
+		return fmt.Errorf("unsupported quantization type %q: supported types are fp4, fp8", quantize)
+	}
+
 	var layers []LayerInfo
 	var configLayer LayerInfo
 	var totalParams int64 // Count parameters from original tensor shapes
+	var torchDtype string // Read from component config for quantization display
 
 	// Components to process - extract individual tensors from each
 	components := []string{"text_encoder", "transformer", "vae"}
@@ -77,8 +61,8 @@ func CreateModel(modelName, modelDir, quantize string, createLayer LayerCreator,
 
 			tensorNames := extractor.ListTensors()
 			quantizeMsg := ""
-			if quantize == "fp8" && component != "vae" {
-				quantizeMsg = ", quantizing to fp8"
+			if quantize != "" && component != "vae" {
+				quantizeMsg = ", quantizing to " + quantize
 			}
 			fn(fmt.Sprintf("importing %s/%s (%d tensors%s)", component, entry.Name(), len(tensorNames), quantizeMsg))
 
@@ -103,11 +87,14 @@ func CreateModel(modelName, modelDir, quantize string, createLayer LayerCreator,
 				// Use path-style name: "component/tensor_name"
 				fullName := component + "/" + tensorName
 
-				// Determine if this tensor should be quantized
-				doQuantize := quantize == "fp8" && ShouldQuantize(tensorName, component)
+				// Determine quantization type for this tensor (empty string if not quantizing)
+				quantizeType := ""
+				if quantize != "" && ShouldQuantize(tensorName, component) && canQuantizeShape(td.Shape) {
+					quantizeType = quantize
+				}
 
 				// createTensorLayer returns multiple layers if quantizing (weight + scales)
-				newLayers, err := createTensorLayer(td.SafetensorsReader(), fullName, td.Dtype, td.Shape, doQuantize)
+				newLayers, err := createTensorLayer(td.SafetensorsReader(), fullName, td.Dtype, td.Shape, quantizeType)
 				if err != nil {
 					extractor.Close()
 					return fmt.Errorf("failed to create layer for %s: %w", fullName, err)
@@ -119,6 +106,19 @@ func CreateModel(modelName, modelDir, quantize string, createLayer LayerCreator,
 		}
 	}
 
+	// Read torch_dtype from text_encoder config for quantization display
+	if torchDtype == "" {
+		textEncoderConfig := filepath.Join(modelDir, "text_encoder/config.json")
+		if data, err := os.ReadFile(textEncoderConfig); err == nil {
+			var cfg struct {
+				TorchDtype string `json:"torch_dtype"`
+			}
+			if json.Unmarshal(data, &cfg) == nil && cfg.TorchDtype != "" {
+				torchDtype = cfg.TorchDtype
+			}
+		}
+	}
+
 	// Import config files
 	configFiles := []string{
 		"model_index.json",
@@ -164,11 +164,11 @@ func CreateModel(modelName, modelDir, quantize string, createLayer LayerCreator,
 			// Add parameter count (counted from tensor shapes during import)
 			cfg["parameter_count"] = totalParams
 
-			// Add quantization info
-			if quantize == "fp8" {
-				cfg["quantization"] = "FP8"
+			// Add quantization info - use quantize type if set, otherwise torch_dtype
+			if quantize != "" {
+				cfg["quantization"] = strings.ToUpper(quantize)
 			} else {
-				cfg["quantization"] = "BF16"
+				cfg["quantization"] = torchDtype
 			}
 
 			data, err = json.MarshalIndent(cfg, "", "    ")
@@ -211,3 +211,12 @@ func CreateModel(modelName, modelDir, quantize string, createLayer LayerCreator,
 	fn(fmt.Sprintf("successfully imported %s with %d layers", modelName, len(layers)))
 	return nil
 }
+
+// canQuantizeShape returns true if a tensor shape is compatible with MLX quantization.
+// MLX requires the last dimension to be divisible by the group size (32).
+func canQuantizeShape(shape []int32) bool {
+	if len(shape) < 2 {
+		return false
+	}
+	return shape[len(shape)-1]%32 == 0
+}

x/imagegen/api/handler.go

@@ -1,231 +0,0 @@
-package api
-
-import (
-	"fmt"
-	"net/http"
-	"strconv"
-	"strings"
-	"time"
-
-	"github.com/gin-gonic/gin"
-
-	"github.com/ollama/ollama/api"
-	"github.com/ollama/ollama/llm"
-	"github.com/ollama/ollama/x/imagegen"
-)
-
-// RunnerScheduler is the interface for scheduling a model runner.
-// This is implemented by server.Server to avoid circular imports.
-type RunnerScheduler interface {
-	ScheduleImageGenRunner(ctx *gin.Context, modelName string, opts api.Options, keepAlive *api.Duration) (llm.LlamaServer, error)
-}
-
-// RegisterRoutes registers the image generation API routes.
-func RegisterRoutes(r gin.IRouter, scheduler RunnerScheduler) {
-	r.POST("/v1/images/generations", func(c *gin.Context) {
-		ImageGenerationHandler(c, scheduler)
-	})
-}
-
-// ImageGenerationHandler handles OpenAI-compatible image generation requests.
-func ImageGenerationHandler(c *gin.Context, scheduler RunnerScheduler) {
-	var req ImageGenerationRequest
-	if err := c.BindJSON(&req); err != nil {
-		c.JSON(http.StatusBadRequest, gin.H{"error": gin.H{"message": err.Error()}})
-		return
-	}
-
-	// Validate required fields
-	if req.Model == "" {
-		c.JSON(http.StatusBadRequest, gin.H{"error": gin.H{"message": "model is required"}})
-		return
-	}
-	if req.Prompt == "" {
-		c.JSON(http.StatusBadRequest, gin.H{"error": gin.H{"message": "prompt is required"}})
-		return
-	}
-
-	// Apply defaults
-	if req.N == 0 {
-		req.N = 1
-	}
-	if req.Size == "" {
-		req.Size = "1024x1024"
-	}
-	if req.ResponseFormat == "" {
-		req.ResponseFormat = "b64_json"
-	}
-
-	// Verify model exists
-	if imagegen.ResolveModelName(req.Model) == "" {
-		c.JSON(http.StatusNotFound, gin.H{"error": gin.H{"message": fmt.Sprintf("model %q not found", req.Model)}})
-		return
-	}
-
-	// Parse size
-	width, height := parseSize(req.Size)
-
-	// Build options - we repurpose NumCtx/NumGPU for width/height
-	opts := api.Options{}
-	opts.NumCtx = int(width)
-	opts.NumGPU = int(height)
-
-	// Schedule runner
-	runner, err := scheduler.ScheduleImageGenRunner(c, req.Model, opts, nil)
-	if err != nil {
-		status := http.StatusInternalServerError
-		if strings.Contains(err.Error(), "not found") {
-			status = http.StatusNotFound
-		}
-		c.JSON(status, gin.H{"error": gin.H{"message": err.Error()}})
-		return
-	}
-
-	// Build completion request
-	completionReq := llm.CompletionRequest{
-		Prompt:  req.Prompt,
-		Options: &opts,
-	}
-
-	if req.Stream {
-		handleStreamingResponse(c, runner, completionReq, req.ResponseFormat)
-	} else {
-		handleNonStreamingResponse(c, runner, completionReq, req.ResponseFormat)
-	}
-}
-
-func handleStreamingResponse(c *gin.Context, runner llm.LlamaServer, req llm.CompletionRequest, format string) {
-	c.Header("Content-Type", "text/event-stream")
-	c.Header("Cache-Control", "no-cache")
-	c.Header("Connection", "keep-alive")
-
-	var imageBase64 string
-	err := runner.Completion(c.Request.Context(), req, func(resp llm.CompletionResponse) {
-		if resp.Done {
-			imageBase64 = extractBase64(resp.Content)
-		} else {
-			progress := parseProgress(resp.Content)
-			if progress.Total > 0 {
-				c.SSEvent("progress", progress)
-				c.Writer.Flush()
-			}
-		}
-	})
-	if err != nil {
-		c.SSEvent("error", gin.H{"error": err.Error()})
-		return
-	}
-
-	c.SSEvent("done", buildResponse(imageBase64, format))
-}
-
-func handleNonStreamingResponse(c *gin.Context, runner llm.LlamaServer, req llm.CompletionRequest, format string) {
-	var imageBase64 string
-	err := runner.Completion(c.Request.Context(), req, func(resp llm.CompletionResponse) {
-		if resp.Done {
-			imageBase64 = extractBase64(resp.Content)
-		}
-	})
-	if err != nil {
-		c.JSON(http.StatusInternalServerError, gin.H{"error": gin.H{"message": err.Error()}})
-		return
-	}
-
-	c.JSON(http.StatusOK, buildResponse(imageBase64, format))
-}
-
-func parseSize(size string) (int32, int32) {
-	parts := strings.Split(size, "x")
-	if len(parts) != 2 {
-		return 1024, 1024
-	}
-	w, _ := strconv.Atoi(parts[0])
-	h, _ := strconv.Atoi(parts[1])
-	if w == 0 {
-		w = 1024
-	}
-	if h == 0 {
-		h = 1024
-	}
-	return int32(w), int32(h)
-}
-
-func extractBase64(content string) string {
-	if strings.HasPrefix(content, "IMAGE_BASE64:") {
-		return content[13:]
-	}
-	return ""
-}
-
-func parseProgress(content string) ImageProgressEvent {
-	var step, total int
-	fmt.Sscanf(content, "\rGenerating: step %d/%d", &step, &total)
-	return ImageProgressEvent{Step: step, Total: total}
-}
-
-func buildResponse(imageBase64, format string) ImageGenerationResponse {
-	resp := ImageGenerationResponse{
-		Created: time.Now().Unix(),
-		Data:    make([]ImageData, 1),
-	}
-
-	if imageBase64 == "" {
-		return resp
-	}
-
-	if format == "url" {
-		// URL format not supported when using base64 transfer
-		resp.Data[0].B64JSON = imageBase64
-	} else {
-		resp.Data[0].B64JSON = imageBase64
-	}
-
-	return resp
-}
-
-// HandleGenerateRequest handles Ollama /api/generate requests for image gen models.
-// This allows routes.go to delegate image generation with minimal code.
-func HandleGenerateRequest(c *gin.Context, scheduler RunnerScheduler, modelName, prompt string, keepAlive *api.Duration, streamFn func(c *gin.Context, ch chan any)) {
-	opts := api.Options{}
-
-	// Schedule runner
-	runner, err := scheduler.ScheduleImageGenRunner(c, modelName, opts, keepAlive)
-	if err != nil {
-		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
-		return
-	}
-
-	// Build completion request
-	completionReq := llm.CompletionRequest{
-		Prompt:  prompt,
-		Options: &opts,
-	}
-
-	// Stream responses via channel
-	ch := make(chan any)
-	go func() {
-		defer close(ch)
-		err := runner.Completion(c.Request.Context(), completionReq, func(resp llm.CompletionResponse) {
-			ch <- GenerateResponse{
-				Model:     modelName,
-				CreatedAt: time.Now().UTC(),
-				Response:  resp.Content,
-				Done:      resp.Done,
-			}
-		})
-		if err != nil {
-			// Log error but don't block - channel is already being consumed
-			_ = err
-		}
-	}()
-
-	streamFn(c, ch)
-}
-
-// GenerateResponse matches api.GenerateResponse structure for streaming.
-type GenerateResponse struct {
-	Model     string    `json:"model"`
-	CreatedAt time.Time `json:"created_at"`
-	Response  string    `json:"response"`
-	Done      bool      `json:"done"`
-}

x/imagegen/api/types.go

@@ -1,31 +0,0 @@
-// Package api provides OpenAI-compatible image generation API types.
-package api
-
-// ImageGenerationRequest is an OpenAI-compatible image generation request.
-type ImageGenerationRequest struct {
-	Model          string `json:"model"`
-	Prompt         string `json:"prompt"`
-	N              int    `json:"n,omitempty"`
-	Size           string `json:"size,omitempty"`
-	ResponseFormat string `json:"response_format,omitempty"`
-	Stream         bool   `json:"stream,omitempty"`
-}
-
-// ImageGenerationResponse is an OpenAI-compatible image generation response.
-type ImageGenerationResponse struct {
-	Created int64       `json:"created"`
-	Data    []ImageData `json:"data"`
-}
-
-// ImageData contains the generated image data.
-type ImageData struct {
-	URL           string `json:"url,omitempty"`
-	B64JSON       string `json:"b64_json,omitempty"`
-	RevisedPrompt string `json:"revised_prompt,omitempty"`
-}
-
-// ImageProgressEvent is sent during streaming to indicate generation progress.
-type ImageProgressEvent struct {
-	Step  int `json:"step"`
-	Total int `json:"total"`
-}

x/imagegen/cli.go

@@ -7,7 +7,6 @@ package imagegen
 
 import (
 	"encoding/base64"
-	"encoding/json"
 	"errors"
 	"fmt"
 	"io"
@@ -39,79 +38,20 @@ func DefaultOptions() ImageGenOptions {
 	return ImageGenOptions{
 		Width:  1024,
 		Height: 1024,
-		Steps:  9,
+		Steps:  0, // 0 means model default
 		Seed:   0, // 0 means random
 	}
 }
 
-// ModelInfo contains metadata about an image generation model.
-type ModelInfo struct {
-	Architecture   string
-	ParameterCount int64
-	Quantization   string
-}
-
-// GetModelInfo returns metadata about an image generation model.
-func GetModelInfo(modelName string) (*ModelInfo, error) {
-	manifest, err := LoadManifest(modelName)
-	if err != nil {
-		return nil, fmt.Errorf("failed to load manifest: %w", err)
-	}
-
-	info := &ModelInfo{}
-
-	// Read model_index.json for architecture, parameter count, and quantization
-	if data, err := manifest.ReadConfig("model_index.json"); err == nil {
-		var index struct {
-			Architecture   string `json:"architecture"`
-			ParameterCount int64  `json:"parameter_count"`
-			Quantization   string `json:"quantization"`
-		}
-		if json.Unmarshal(data, &index) == nil {
-			info.Architecture = index.Architecture
-			info.ParameterCount = index.ParameterCount
-			info.Quantization = index.Quantization
-		}
-	}
-
-	// Fallback: detect quantization from tensor names if not in config
-	if info.Quantization == "" {
-		for _, layer := range manifest.Manifest.Layers {
-			if strings.HasSuffix(layer.Name, ".weight_scale") {
-				info.Quantization = "FP8"
-				break
-			}
-		}
-		if info.Quantization == "" {
-			info.Quantization = "BF16"
-		}
-	}
-
-	// Fallback: estimate parameter count if not in config
-	if info.ParameterCount == 0 {
-		var totalSize int64
-		for _, layer := range manifest.Manifest.Layers {
-			if layer.MediaType == "application/vnd.ollama.image.tensor" {
-				if !strings.HasSuffix(layer.Name, "_scale") && !strings.HasSuffix(layer.Name, "_qbias") {
-					totalSize += layer.Size
-				}
-			}
-		}
-		// Assume BF16 (2 bytes/param) as rough estimate
-		info.ParameterCount = totalSize / 2
-	}
-
-	return info, nil
-}
-
 // RegisterFlags adds image generation flags to the given command.
 // Flags are hidden since they only apply to image generation models.
 func RegisterFlags(cmd *cobra.Command) {
 	cmd.Flags().Int("width", 1024, "Image width")
 	cmd.Flags().Int("height", 1024, "Image height")
-	cmd.Flags().Int("steps", 9, "Denoising steps")
+	cmd.Flags().Int("steps", 0, "Denoising steps (0 = model default)")
 	cmd.Flags().Int("seed", 0, "Random seed (0 for random)")
 	cmd.Flags().String("negative", "", "Negative prompt")
+	// Hide from main flags section - shown in separate section via AppendFlagsDocs
 	cmd.Flags().MarkHidden("width")
 	cmd.Flags().MarkHidden("height")
 	cmd.Flags().MarkHidden("steps")
@@ -119,6 +59,19 @@ func RegisterFlags(cmd *cobra.Command) {
 	cmd.Flags().MarkHidden("negative")
 }
 
+// AppendFlagsDocs appends image generation flags documentation to the command's usage template.
+func AppendFlagsDocs(cmd *cobra.Command) {
+	usage := `
+Image Generation Flags (experimental):
+      --width int      Image width
+      --height int     Image height
+      --steps int      Denoising steps
+      --seed int       Random seed
+      --negative str   Negative prompt
+`
+	cmd.SetUsageTemplate(cmd.UsageTemplate() + usage)
+}
+
 // RunCLI handles the CLI for image generation models.
 // Returns true if it handled the request, false if the caller should continue with normal flow.
 // Supports flags: --width, --height, --steps, --seed, --negative
@@ -158,17 +111,15 @@ func generateImageWithOptions(cmd *cobra.Command, modelName, prompt string, keep
 		return err
 	}
 
-	// Build request with image gen options encoded in Options fields
-	// NumCtx=width, NumGPU=height, NumPredict=steps, Seed=seed
 	req := &api.GenerateRequest{
 		Model:  modelName,
 		Prompt: prompt,
-		Options: map[string]any{
-			"num_ctx":     opts.Width,
-			"num_gpu":     opts.Height,
-			"num_predict": opts.Steps,
-			"seed":        opts.Seed,
-		},
+		Width:  int32(opts.Width),
+		Height: int32(opts.Height),
+		Steps:  int32(opts.Steps),
+	}
+	if opts.Seed != 0 {
+		req.Options = map[string]any{"seed": opts.Seed}
 	}
 	if keepAlive != nil {
 		req.KeepAlive = keepAlive
@@ -182,32 +133,25 @@ func generateImageWithOptions(cmd *cobra.Command, modelName, prompt string, keep
 	var stepBar *progress.StepBar
 	var imageBase64 string
 	err = client.Generate(cmd.Context(), req, func(resp api.GenerateResponse) error {
-		content := resp.Response
-
-		// Handle progress updates - parse step info and switch to step bar
-		if strings.HasPrefix(content, "\rGenerating:") {
-			var step, total int
-			fmt.Sscanf(content, "\rGenerating: step %d/%d", &step, &total)
-			if stepBar == nil && total > 0 {
+		// Handle progress updates using structured fields
+		if resp.Total > 0 {
+			if stepBar == nil {
 				spinner.Stop()
-				stepBar = progress.NewStepBar("Generating", total)
+				stepBar = progress.NewStepBar("Generating", int(resp.Total))
 				p.Add("", stepBar)
 			}
-			if stepBar != nil {
-				stepBar.Set(step)
-			}
-			return nil
+			stepBar.Set(int(resp.Completed))
 		}
 
-		// Handle final response with base64 image data
-		if resp.Done && strings.HasPrefix(content, "IMAGE_BASE64:") {
-			imageBase64 = content[13:]
+		// Handle final response with image data
+		if resp.Done && resp.Image != "" {
+			imageBase64 = resp.Image
 		}
 
 		return nil
 	})
 
-	p.Stop()
+	p.StopAndClear()
 	if err != nil {
 		return err
 	}
@@ -245,6 +189,23 @@ func runInteractive(cmd *cobra.Command, modelName string, keepAlive *api.Duratio
 		return err
 	}
 
+	// Preload the model with the specified keepalive
+	p := progress.NewProgress(os.Stderr)
+	spinner := progress.NewSpinner("")
+	p.Add("", spinner)
+
+	preloadReq := &api.GenerateRequest{
+		Model:     modelName,
+		KeepAlive: keepAlive,
+	}
+	if err := client.Generate(cmd.Context(), preloadReq, func(resp api.GenerateResponse) error {
+		return nil
+	}); err != nil {
+		p.StopAndClear()
+		return fmt.Errorf("failed to load model: %w", err)
+	}
+	p.StopAndClear()
+
 	scanner, err := readline.New(readline.Prompt{
 		Prompt:      ">>> ",
 		Placeholder: "Describe an image to generate (/help for commands)",
@@ -282,7 +243,7 @@ func runInteractive(cmd *cobra.Command, modelName string, keepAlive *api.Duratio
 		case strings.HasPrefix(line, "/bye"):
 			return nil
 		case strings.HasPrefix(line, "/?"), strings.HasPrefix(line, "/help"):
-			printInteractiveHelp(opts)
+			printInteractiveHelp()
 			continue
 		case strings.HasPrefix(line, "/set "):
 			if err := handleSetCommand(line[5:], &opts); err != nil {
@@ -301,12 +262,12 @@ func runInteractive(cmd *cobra.Command, modelName string, keepAlive *api.Duratio
 		req := &api.GenerateRequest{
 			Model:  modelName,
 			Prompt: line,
-			Options: map[string]any{
-				"num_ctx":     opts.Width,
-				"num_gpu":     opts.Height,
-				"num_predict": opts.Steps,
-				"seed":        opts.Seed,
-			},
+			Width:  int32(opts.Width),
+			Height: int32(opts.Height),
+			Steps:  int32(opts.Steps),
+		}
+		if opts.Seed != 0 {
+			req.Options = map[string]any{"seed": opts.Seed}
 		}
 		if keepAlive != nil {
 			req.KeepAlive = keepAlive
@@ -321,32 +282,25 @@ func runInteractive(cmd *cobra.Command, modelName string, keepAlive *api.Duratio
 		var imageBase64 string
 
 		err = client.Generate(cmd.Context(), req, func(resp api.GenerateResponse) error {
-			content := resp.Response
-
-			// Handle progress updates - parse step info and switch to step bar
-			if strings.HasPrefix(content, "\rGenerating:") {
-				var step, total int
-				fmt.Sscanf(content, "\rGenerating: step %d/%d", &step, &total)
-				if stepBar == nil && total > 0 {
+			// Handle progress updates using structured fields
+			if resp.Total > 0 {
+				if stepBar == nil {
 					spinner.Stop()
-					stepBar = progress.NewStepBar("Generating", total)
+					stepBar = progress.NewStepBar("Generating", int(resp.Total))
 					p.Add("", stepBar)
 				}
-				if stepBar != nil {
-					stepBar.Set(step)
-				}
-				return nil
+				stepBar.Set(int(resp.Completed))
 			}
 
-			// Handle final response with base64 image data
-			if resp.Done && strings.HasPrefix(content, "IMAGE_BASE64:") {
-				imageBase64 = content[13:]
+			// Handle final response with image data
+			if resp.Done && resp.Image != "" {
+				imageBase64 = resp.Image
 			}
 
 			return nil
 		})
 
-		p.Stop()
+		p.StopAndClear()
 		if err != nil {
 			fmt.Fprintf(os.Stderr, "Error: %v\n", err)
 			continue
@@ -397,12 +351,13 @@ func sanitizeFilename(s string) string {
 }
 
 // printInteractiveHelp prints help for interactive mode commands.
-func printInteractiveHelp(opts ImageGenOptions) {
+// TODO: reconcile /set commands with /set parameter in text gen REPL (cmd/cmd.go)
+func printInteractiveHelp() {
 	fmt.Fprintln(os.Stderr, "Commands:")
-	fmt.Fprintln(os.Stderr, "  /set width <n>     Set image width (current:", opts.Width, ")")
-	fmt.Fprintln(os.Stderr, "  /set height <n>    Set image height (current:", opts.Height, ")")
-	fmt.Fprintln(os.Stderr, "  /set steps <n>     Set denoising steps (current:", opts.Steps, ")")
-	fmt.Fprintln(os.Stderr, "  /set seed <n>      Set random seed (current:", opts.Seed, ", 0=random)")
+	fmt.Fprintln(os.Stderr, "  /set width <n>     Set image width")
+	fmt.Fprintln(os.Stderr, "  /set height <n>    Set image height")
+	fmt.Fprintln(os.Stderr, "  /set steps <n>     Set denoising steps")
+	fmt.Fprintln(os.Stderr, "  /set seed <n>      Set random seed")
 	fmt.Fprintln(os.Stderr, "  /set negative <s>  Set negative prompt")
 	fmt.Fprintln(os.Stderr, "  /show              Show current settings")
 	fmt.Fprintln(os.Stderr, "  /bye               Exit")

x/imagegen/client/create.go

@@ -1,190 +0,0 @@
-// Package client provides client-side model creation for tensor-based models.
-//
-// This package is in x/ because the tensor model storage format is under development.
-// It also exists to break an import cycle: server imports x/imagegen, so x/imagegen
-// cannot import server. This sub-package can import server because server doesn't
-// import it.
-//
-// TODO (jmorganca): This is temporary. When tensor models are promoted to production:
-//  1. Add proper API endpoints for tensor model creation
-//  2. Move tensor extraction to server-side
-//  3. Remove this package
-//  4. Follow the same client→server pattern as regular model creation
-package client
-
-import (
-	"bytes"
-	"encoding/json"
-	"fmt"
-	"io"
-
-	"github.com/ollama/ollama/progress"
-	"github.com/ollama/ollama/server"
-	"github.com/ollama/ollama/types/model"
-	"github.com/ollama/ollama/x/imagegen"
-)
-
-// MinOllamaVersion is the minimum Ollama version required for image generation models.
-const MinOllamaVersion = "0.14.0"
-
-// CreateModel imports a tensor-based model from a local directory.
-// This creates blobs and manifest directly on disk, bypassing the HTTP API.
-// If quantize is "fp8", weights will be quantized to mxfp8 format during import.
-//
-// TODO (jmorganca): Replace with API-based creation when promoted to production.
-func CreateModel(modelName, modelDir, quantize string, p *progress.Progress) error {
-	if !imagegen.IsTensorModelDir(modelDir) {
-		return fmt.Errorf("%s is not an image generation model directory (model_index.json not found)", modelDir)
-	}
-
-	status := "importing image generation model"
-	spinner := progress.NewSpinner(status)
-	p.Add("imagegen", spinner)
-
-	// Create layer callback for config files
-	createLayer := func(r io.Reader, mediaType, name string) (imagegen.LayerInfo, error) {
-		layer, err := server.NewLayer(r, mediaType)
-		if err != nil {
-			return imagegen.LayerInfo{}, err
-		}
-		layer.Name = name
-
-		return imagegen.LayerInfo{
-			Digest:    layer.Digest,
-			Size:      layer.Size,
-			MediaType: layer.MediaType,
-			Name:      name,
-		}, nil
-	}
-
-	// Create tensor layer callback for individual tensors
-	// name is path-style: "component/tensor_name"
-	// When quantize is true, returns multiple layers (weight + scales)
-	createTensorLayer := func(r io.Reader, name, dtype string, shape []int32, doQuantize bool) ([]imagegen.LayerInfo, error) {
-		if doQuantize {
-			// Check if quantization is supported
-			if !QuantizeSupported() {
-				return nil, fmt.Errorf("quantization requires MLX support")
-			}
-
-			// Quantize the tensor (affine mode returns weight, scales, qbiases)
-			qweightData, scalesData, qbiasData, _, _, _, err := quantizeTensor(r, name, dtype, shape)
-			if err != nil {
-				return nil, fmt.Errorf("failed to quantize %s: %w", name, err)
-			}
-
-			// Create layer for quantized weight
-			weightLayer, err := server.NewLayer(bytes.NewReader(qweightData), server.MediaTypeImageTensor)
-			if err != nil {
-				return nil, err
-			}
-
-			// Create layer for scales (use _scale suffix convention)
-			scalesLayer, err := server.NewLayer(bytes.NewReader(scalesData), server.MediaTypeImageTensor)
-			if err != nil {
-				return nil, err
-			}
-
-			layers := []imagegen.LayerInfo{
-				{
-					Digest:    weightLayer.Digest,
-					Size:      weightLayer.Size,
-					MediaType: weightLayer.MediaType,
-					Name:      name, // Keep original name for weight
-				},
-				{
-					Digest:    scalesLayer.Digest,
-					Size:      scalesLayer.Size,
-					MediaType: scalesLayer.MediaType,
-					Name:      name + "_scale", // Add _scale suffix
-				},
-			}
-
-			// Add qbiases layer if present (affine mode)
-			if qbiasData != nil {
-				qbiasLayer, err := server.NewLayer(bytes.NewReader(qbiasData), server.MediaTypeImageTensor)
-				if err != nil {
-					return nil, err
-				}
-				layers = append(layers, imagegen.LayerInfo{
-					Digest:    qbiasLayer.Digest,
-					Size:      qbiasLayer.Size,
-					MediaType: qbiasLayer.MediaType,
-					Name:      name + "_qbias", // Add _qbias suffix
-				})
-			}
-
-			return layers, nil
-		}
-
-		// Non-quantized path: just create a single layer
-		layer, err := server.NewLayer(r, server.MediaTypeImageTensor)
-		if err != nil {
-			return nil, err
-		}
-
-		return []imagegen.LayerInfo{
-			{
-				Digest:    layer.Digest,
-				Size:      layer.Size,
-				MediaType: layer.MediaType,
-				Name:      name,
-			},
-		}, nil
-	}
-
-	// Create manifest writer callback
-	writeManifest := func(modelName string, config imagegen.LayerInfo, layers []imagegen.LayerInfo) error {
-		name := model.ParseName(modelName)
-		if !name.IsValid() {
-			return fmt.Errorf("invalid model name: %s", modelName)
-		}
-
-		// Create a proper config blob with version requirement
-		configData := model.ConfigV2{
-			ModelFormat:  "safetensors",
-			Capabilities: []string{"image"},
-			Requires:     MinOllamaVersion,
-		}
-		configJSON, err := json.Marshal(configData)
-		if err != nil {
-			return fmt.Errorf("failed to marshal config: %w", err)
-		}
-
-		// Create config layer blob
-		configLayer, err := server.NewLayer(bytes.NewReader(configJSON), "application/vnd.docker.container.image.v1+json")
-		if err != nil {
-			return fmt.Errorf("failed to create config layer: %w", err)
-		}
-
-		// Convert LayerInfo to server.Layer (include the original model_index.json in layers)
-		serverLayers := make([]server.Layer, len(layers))
-		for i, l := range layers {
-			serverLayers[i] = server.Layer{
-				MediaType: l.MediaType,
-				Digest:    l.Digest,
-				Size:      l.Size,
-				Name:      l.Name,
-			}
-		}
-
-		return server.WriteManifest(name, configLayer, serverLayers)
-	}
-
-	// Progress callback
-	progressFn := func(msg string) {
-		spinner.Stop()
-		status = msg
-		spinner = progress.NewSpinner(status)
-		p.Add("imagegen", spinner)
-	}
-
-	err := imagegen.CreateModel(modelName, modelDir, quantize, createLayer, createTensorLayer, writeManifest, progressFn)
-	spinner.Stop()
-	if err != nil {
-		return err
-	}
-
-	fmt.Printf("Created image generation model '%s'\n", modelName)
-	return nil
-}

x/imagegen/cmd/engine/generate.go

@@ -65,12 +65,12 @@ func (s *utf8Streamer) Flush() string {
 	return result
 }
 
-func init() {
-	generationStream = mlx.NewStream()
-}
-
 // withStream runs fn with the generation stream as default
 func withStream(fn func()) {
+	// Lazy initialization of generationStream
+	if generationStream == nil {
+		generationStream = mlx.NewStream()
+	}
 	orig := mlx.GetDefaultStream()
 	mlx.SetDefaultStream(generationStream)
 	fn()

x/imagegen/cmd/engine/main.go

@@ -7,12 +7,17 @@ import (
 	"encoding/json"
 	"flag"
 	"fmt"
+	"image"
+	_ "image/jpeg"
+	_ "image/png"
 	"log"
 	"os"
 	"path/filepath"
 	"runtime/pprof"
 
+	"github.com/ollama/ollama/x/imagegen"
 	"github.com/ollama/ollama/x/imagegen/mlx"
+	"github.com/ollama/ollama/x/imagegen/models/flux2"
 	"github.com/ollama/ollama/x/imagegen/models/gemma3"
 	"github.com/ollama/ollama/x/imagegen/models/gpt_oss"
 	"github.com/ollama/ollama/x/imagegen/models/llama"
@@ -46,9 +51,9 @@ func main() {
 	imagePath := flag.String("image", "", "Image path for multimodal models")
 
 	// Image generation params
-	width := flag.Int("width", 1024, "Image width")
-	height := flag.Int("height", 1024, "Image height")
-	steps := flag.Int("steps", 9, "Denoising steps")
+	width := flag.Int("width", 0, "Image width (0 = auto from input or 1024)")
+	height := flag.Int("height", 0, "Image height (0 = auto from input or 1024)")
+	steps := flag.Int("steps", 0, "Denoising steps (0 = model default)")
 	seed := flag.Int64("seed", 42, "Random seed")
 	out := flag.String("output", "output.png", "Output path")
 
@@ -61,6 +66,7 @@ func main() {
 
 	// Legacy mode flags
 	zimageFlag := flag.Bool("zimage", false, "Z-Image generation")
+	flux2Flag := flag.Bool("flux2", false, "FLUX.2 Klein generation")
 	qwenImage := flag.Bool("qwen-image", false, "Qwen-Image text-to-image generation")
 	qwenImageEdit := flag.Bool("qwen-image-edit", false, "Qwen-Image-Edit image editing")
 	var inputImages stringSlice
@@ -78,6 +84,11 @@ func main() {
 		return
 	}
 
+	// Check if MLX initialized successfully
+	if !mlx.IsMLXAvailable() {
+		log.Fatalf("MLX initialization failed: %v", mlx.GetMLXInitError())
+	}
+
 	// CPU profiling
 	if *cpuProfile != "" {
 		f, err := os.Create(*cpuProfile)
@@ -117,6 +128,44 @@ func main() {
 		if err == nil {
 			err = saveImageArray(img, *out)
 		}
+	case *flux2Flag:
+		m := &flux2.Model{}
+		if loadErr := m.Load(*modelPath); loadErr != nil {
+			log.Fatal(loadErr)
+		}
+		// Load input images with EXIF orientation correction
+		var loadedImages []image.Image
+		for _, path := range inputImages {
+			img, loadErr := loadImageWithEXIF(path)
+			if loadErr != nil {
+				log.Fatalf("Failed to load image %s: %v", path, loadErr)
+			}
+			loadedImages = append(loadedImages, img)
+		}
+		// When input images provided and user didn't override dimensions, use 0 to match input
+		fluxWidth := int32(*width)
+		fluxHeight := int32(*height)
+		if len(loadedImages) > 0 && *width == 0 && *height == 0 {
+			// Both unset, will auto-detect from input
+		} else if len(loadedImages) > 0 && *width == 0 {
+			fluxWidth = 0 // Compute from height + aspect ratio
+		} else if len(loadedImages) > 0 && *height == 0 {
+			fluxHeight = 0 // Compute from width + aspect ratio
+		}
+		var img *mlx.Array
+		img, err = m.GenerateFromConfig(context.Background(), &flux2.GenerateConfig{
+			Prompt:        *prompt,
+			Width:         fluxWidth,
+			Height:        fluxHeight,
+			Steps:         *steps,
+			GuidanceScale: float32(*cfgScale),
+			Seed:          *seed,
+			CapturePath:   *gpuCapture,
+			InputImages:   loadedImages,
+		})
+		if err == nil {
+			err = saveImageArray(img, *out)
+		}
 	case *qwenImage:
 		m, loadErr := qwen_image.LoadPersistent(*modelPath)
 		if loadErr != nil {
@@ -271,6 +320,8 @@ func detectModelKind(modelPath string) (string, error) {
 			switch index.ClassName {
 			case "FluxPipeline", "ZImagePipeline":
 				return "zimage", nil
+			case "Flux2KleinPipeline":
+				return "flux2", nil
 			}
 		}
 		return "zimage", nil
@@ -291,3 +342,12 @@ func detectModelKind(modelPath string) (string, error) {
 
 	return cfg.ModelType, nil
 }
+
+// loadImageWithEXIF loads an image from a file path with EXIF orientation correction.
+func loadImageWithEXIF(path string) (image.Image, error) {
+	data, err := os.ReadFile(path)
+	if err != nil {
+		return nil, fmt.Errorf("read file: %w", err)
+	}
+	return imagegen.DecodeImage(data)
+}

x/imagegen/image.go

@@ -7,6 +7,7 @@ import (
 	"encoding/base64"
 	"fmt"
 	"image"
+	_ "image/jpeg"
 	"image/png"
 	"os"
 	"path/filepath"
@@ -108,3 +109,160 @@ func clampF(v, min, max float32) float32 {
 	}
 	return v
 }
+
+// DecodeImage decodes image bytes with EXIF orientation applied.
+func DecodeImage(data []byte) (image.Image, error) {
+	orientation := readJPEGOrientation(data)
+
+	img, _, err := image.Decode(bytes.NewReader(data))
+	if err != nil {
+		return nil, err
+	}
+
+	return applyOrientation(img, orientation), nil
+}
+
+// readJPEGOrientation extracts EXIF orientation from JPEG bytes.
+// Returns 1 (normal) for non-JPEG or if orientation not found.
+func readJPEGOrientation(data []byte) int {
+	if len(data) < 2 || data[0] != 0xFF || data[1] != 0xD8 {
+		return 1 // Not JPEG
+	}
+
+	r := bytes.NewReader(data[2:])
+	for {
+		var marker [2]byte
+		if _, err := r.Read(marker[:]); err != nil || marker[0] != 0xFF {
+			return 1
+		}
+
+		if marker[1] == 0xE1 { // APP1 (EXIF)
+			var lenBytes [2]byte
+			if _, err := r.Read(lenBytes[:]); err != nil {
+				return 1
+			}
+			segLen := int(uint16(lenBytes[0])<<8|uint16(lenBytes[1])) - 2
+			if segLen < 14 {
+				r.Seek(int64(segLen), 1)
+				continue
+			}
+			seg := make([]byte, segLen)
+			if _, err := r.Read(seg); err != nil {
+				return 1
+			}
+			if string(seg[:4]) == "Exif" && seg[4] == 0 && seg[5] == 0 {
+				return parseTIFFOrientation(seg[6:])
+			}
+			continue
+		}
+
+		if marker[1] == 0xD9 || marker[1] == 0xDA {
+			return 1 // EOI or SOS
+		}
+		if marker[1] >= 0xD0 && marker[1] <= 0xD7 {
+			continue // RST markers
+		}
+
+		var lenBytes [2]byte
+		if _, err := r.Read(lenBytes[:]); err != nil {
+			return 1
+		}
+		segLen := int(uint16(lenBytes[0])<<8|uint16(lenBytes[1])) - 2
+		if segLen > 0 {
+			r.Seek(int64(segLen), 1)
+		}
+	}
+}
+
+func parseTIFFOrientation(tiff []byte) int {
+	if len(tiff) < 8 {
+		return 1
+	}
+
+	var big bool
+	switch string(tiff[:2]) {
+	case "MM":
+		big = true
+	case "II":
+		big = false
+	default:
+		return 1
+	}
+
+	u16 := func(b []byte) uint16 {
+		if big {
+			return uint16(b[0])<<8 | uint16(b[1])
+		}
+		return uint16(b[1])<<8 | uint16(b[0])
+	}
+	u32 := func(b []byte) uint32 {
+		if big {
+			return uint32(b[0])<<24 | uint32(b[1])<<16 | uint32(b[2])<<8 | uint32(b[3])
+		}
+		return uint32(b[3])<<24 | uint32(b[2])<<16 | uint32(b[1])<<8 | uint32(b[0])
+	}
+
+	if u16(tiff[2:4]) != 42 {
+		return 1
+	}
+
+	ifdOffset := u32(tiff[4:8])
+	if int(ifdOffset)+2 > len(tiff) {
+		return 1
+	}
+
+	numEntries := u16(tiff[ifdOffset : ifdOffset+2])
+	for i := range int(numEntries) {
+		offset := ifdOffset + 2 + uint32(i)*12
+		if int(offset)+12 > len(tiff) {
+			break
+		}
+		if u16(tiff[offset:offset+2]) == 0x0112 { // Orientation tag
+			o := int(u16(tiff[offset+8 : offset+10]))
+			if o >= 1 && o <= 8 {
+				return o
+			}
+			return 1
+		}
+	}
+	return 1
+}
+
+func applyOrientation(img image.Image, orientation int) image.Image {
+	if orientation <= 1 || orientation > 8 {
+		return img
+	}
+
+	bounds := img.Bounds()
+	w, h := bounds.Dx(), bounds.Dy()
+
+	outW, outH := w, h
+	if orientation >= 5 {
+		outW, outH = h, w
+	}
+
+	out := image.NewRGBA(image.Rect(0, 0, outW, outH))
+	for y := range h {
+		for x := range w {
+			var dx, dy int
+			switch orientation {
+			case 2:
+				dx, dy = w-1-x, y
+			case 3:
+				dx, dy = w-1-x, h-1-y
+			case 4:
+				dx, dy = x, h-1-y
+			case 5:
+				dx, dy = y, x
+			case 6:
+				dx, dy = h-1-y, x
+			case 7:
+				dx, dy = h-1-y, w-1-x
+			case 8:
+				dx, dy = y, w-1-x
+			}
+			out.Set(dx, dy, img.At(x+bounds.Min.X, y+bounds.Min.Y))
+		}
+	}
+	return out
+}

x/imagegen/manifest.go

@@ -175,3 +175,63 @@ func (m *ModelManifest) HasTensorLayers() bool {
 	}
 	return false
 }
+
+// ModelInfo contains metadata about an image generation model.
+type ModelInfo struct {
+	Architecture   string
+	ParameterCount int64
+	Quantization   string
+}
+
+// GetModelInfo returns metadata about an image generation model.
+func GetModelInfo(modelName string) (*ModelInfo, error) {
+	manifest, err := LoadManifest(modelName)
+	if err != nil {
+		return nil, fmt.Errorf("failed to load manifest: %w", err)
+	}
+
+	info := &ModelInfo{}
+
+	// Read model_index.json for architecture, parameter count, and quantization
+	if data, err := manifest.ReadConfig("model_index.json"); err == nil {
+		var index struct {
+			Architecture   string `json:"architecture"`
+			ParameterCount int64  `json:"parameter_count"`
+			Quantization   string `json:"quantization"`
+		}
+		if json.Unmarshal(data, &index) == nil {
+			info.Architecture = index.Architecture
+			info.ParameterCount = index.ParameterCount
+			info.Quantization = index.Quantization
+		}
+	}
+
+	// Fallback: detect quantization from tensor names if not in config
+	if info.Quantization == "" {
+		for _, layer := range manifest.Manifest.Layers {
+			if strings.HasSuffix(layer.Name, ".weight_scale") {
+				info.Quantization = "FP8"
+				break
+			}
+		}
+		if info.Quantization == "" {
+			info.Quantization = "BF16"
+		}
+	}
+
+	// Fallback: estimate parameter count if not in config
+	if info.ParameterCount == 0 {
+		var totalSize int64
+		for _, layer := range manifest.Manifest.Layers {
+			if layer.MediaType == "application/vnd.ollama.image.tensor" {
+				if !strings.HasSuffix(layer.Name, "_scale") && !strings.HasSuffix(layer.Name, "_qbias") {
+					totalSize += layer.Size
+				}
+			}
+		}
+		// Assume BF16 (2 bytes/param) as rough estimate
+		info.ParameterCount = totalSize / 2
+	}
+
+	return info, nil
+}

x/imagegen/memory.go

@@ -24,9 +24,8 @@ var SupportedBackends = []string{"metal", "cuda", "cpu"}
 
 // modelVRAMEstimates maps pipeline class names to their estimated VRAM requirements.
 var modelVRAMEstimates = map[string]uint64{
-	"ZImagePipeline":    21 * GB, // ~21GB for Z-Image (text encoder + transformer + VAE)
-	"FluxPipeline":      21 * GB, // ~21GB for Flux (same architecture)
-	"QwenImagePipeline": 80 * GB, // TODO: verify actual requirements, using conservative estimate for now
+	"ZImagePipeline": 21 * GB, // ~21GB for Z-Image (text encoder + transformer + VAE)
+	"FluxPipeline":   20 * GB, // ~20GB for Flux
 }
 
 // CheckPlatformSupport validates that image generation is supported on the current platform.
@@ -72,31 +71,38 @@ func ResolveModelName(modelName string) string {
 // EstimateVRAM returns the estimated VRAM needed for an image generation model.
 // Returns a conservative default of 21GB if the model type cannot be determined.
 func EstimateVRAM(modelName string) uint64 {
-	manifest, err := LoadManifest(modelName)
-	if err != nil {
-		return 21 * GB
-	}
-
-	data, err := manifest.ReadConfig("model_index.json")
-	if err != nil {
-		return 21 * GB
-	}
-
-	// Parse just the class name
-	var index struct {
-		ClassName string `json:"_class_name"`
-	}
-	if err := json.Unmarshal(data, &index); err != nil {
-		return 21 * GB
-	}
-
-	if estimate, ok := modelVRAMEstimates[index.ClassName]; ok {
+	className := DetectModelType(modelName)
+	if estimate, ok := modelVRAMEstimates[className]; ok {
 		return estimate
 	}
 	return 21 * GB
 }
 
-// HasTensorLayers checks if the given model has tensor layers.
-func HasTensorLayers(modelName string) bool {
-	return ResolveModelName(modelName) != ""
+// DetectModelType reads model_index.json and returns the model type.
+// Checks both "architecture" (Ollama format) and "_class_name" (diffusers format).
+// Returns empty string if detection fails.
+func DetectModelType(modelName string) string {
+	manifest, err := LoadManifest(modelName)
+	if err != nil {
+		return ""
+	}
+
+	data, err := manifest.ReadConfig("model_index.json")
+	if err != nil {
+		return ""
+	}
+
+	var index struct {
+		Architecture string `json:"architecture"`
+		ClassName    string `json:"_class_name"`
+	}
+	if err := json.Unmarshal(data, &index); err != nil {
+		return ""
+	}
+
+	// Prefer architecture (Ollama format), fall back to _class_name (diffusers)
+	if index.Architecture != "" {
+		return index.Architecture
+	}
+	return index.ClassName
 }

x/imagegen/memory_test.go

@@ -72,9 +72,8 @@ func TestCheckMemoryRequirements(t *testing.T) {
 func TestModelVRAMEstimates(t *testing.T) {
 	// Verify the VRAM estimates map has expected entries
 	expected := map[string]uint64{
-		"ZImagePipeline":    21 * GB,
-		"FluxPipeline":      21 * GB,
-		"QwenImagePipeline": 80 * GB,
+		"ZImagePipeline": 21 * GB,
+		"FluxPipeline":   20 * GB,
 	}
 
 	for name, expectedVRAM := range expected {
@@ -94,13 +93,6 @@ func TestEstimateVRAMDefault(t *testing.T) {
 	}
 }
 
-func TestHasTensorLayers(t *testing.T) {
-	// Non-existent model should return false
-	if HasTensorLayers("nonexistent-model") {
-		t.Error("HasTensorLayers() should return false for non-existent model")
-	}
-}
-
 func TestResolveModelName(t *testing.T) {
 	// Non-existent model should return empty string
 	result := ResolveModelName("nonexistent-model")

x/imagegen/mlx/compile.go

@@ -3,7 +3,7 @@
 package mlx
 
 /*
-#include "mlx/c/mlx.h"
+#include "mlx.h"
 #include <stdlib.h>
 
 // Forward declaration for Go callback

x/imagegen/mlx/doc.go

@@ -0,0 +1,6 @@
+//go:build mlx
+
+// Package mlx provides Go bindings for the MLX-C library with dynamic loading support.
+//
+//go:generate go run generate_wrappers.go ../../../build/_deps/mlx-c-src/mlx/c mlx.h mlx.c
+package mlx

x/imagegen/mlx/generate_wrappers.go

@@ -0,0 +1,439 @@
+//go:build ignore
+
+// This tool generates MLX-C dynamic loading wrappers.
+// Usage: go run generate_wrappers.go <mlx-c-include-dir> <output-header> [output-impl]
+package main
+
+import (
+	"bytes"
+	"flag"
+	"fmt"
+	"io/fs"
+	"os"
+	"path/filepath"
+	"regexp"
+	"strings"
+)
+
+type Function struct {
+	Name          string
+	ReturnType    string
+	Params        string
+	ParamNames    []string
+	NeedsARM64Guard bool
+}
+
+func findHeaders(directory string) ([]string, error) {
+	var headers []string
+	err := filepath.WalkDir(directory, func(path string, d fs.DirEntry, err error) error {
+		if err != nil {
+			return err
+		}
+		if !d.IsDir() && strings.HasSuffix(path, ".h") {
+			headers = append(headers, path)
+		}
+		return nil
+	})
+	return headers, err
+}
+
+func cleanContent(content string) string {
+	// Remove single-line comments
+	re := regexp.MustCompile(`//.*?\n`)
+	content = re.ReplaceAllString(content, "\n")
+
+	// Remove multi-line comments
+	re = regexp.MustCompile(`/\*.*?\*/`)
+	content = re.ReplaceAllString(content, "")
+
+	// Remove preprocessor directives (lines starting with #) - use multiline mode
+	re = regexp.MustCompile(`(?m)^\s*#.*?$`)
+	content = re.ReplaceAllString(content, "")
+
+	// Remove extern "C" { and } blocks more conservatively
+	// Only remove the extern "C" { line, not the content inside
+	re = regexp.MustCompile(`extern\s+"C"\s*\{\s*?\n`)
+	content = re.ReplaceAllString(content, "\n")
+	// Remove standalone closing braces that are not part of function declarations
+	re = regexp.MustCompile(`\n\s*\}\s*\n`)
+	content = re.ReplaceAllString(content, "\n")
+
+	// Collapse whitespace and newlines
+	re = regexp.MustCompile(`\s+`)
+	content = re.ReplaceAllString(content, " ")
+
+	return content
+}
+
+func extractParamNames(params string) []string {
+	if params == "" || strings.TrimSpace(params) == "void" {
+		return []string{}
+	}
+
+	var names []string
+
+	// Split by comma, but respect parentheses (for function pointers)
+	parts := splitParams(params)
+
+	// Remove array brackets
+	arrayBrackets := regexp.MustCompile(`\[.*?\]`)
+
+	// Function pointer pattern
+	funcPtrPattern := regexp.MustCompile(`\(\s*\*\s*(\w+)\s*\)`)
+
+	// Type keywords to skip
+	typeKeywords := map[string]bool{
+		"const":     true,
+		"struct":    true,
+		"unsigned":  true,
+		"signed":    true,
+		"long":      true,
+		"short":     true,
+		"int":       true,
+		"char":      true,
+		"float":     true,
+		"double":    true,
+		"void":      true,
+		"size_t":    true,
+		"uint8_t":   true,
+		"uint16_t":  true,
+		"uint32_t":  true,
+		"uint64_t":  true,
+		"int8_t":    true,
+		"int16_t":   true,
+		"int32_t":   true,
+		"int64_t":   true,
+		"intptr_t":  true,
+		"uintptr_t": true,
+	}
+
+	for _, part := range parts {
+		if part == "" {
+			continue
+		}
+
+		// Remove array brackets
+		part = arrayBrackets.ReplaceAllString(part, "")
+
+		// For function pointers like "void (*callback)(int)"
+		if matches := funcPtrPattern.FindStringSubmatch(part); len(matches) > 1 {
+			names = append(names, matches[1])
+			continue
+		}
+
+		// Regular parameter: last identifier
+		tokens := regexp.MustCompile(`\w+`).FindAllString(part, -1)
+		if len(tokens) > 0 {
+			// The last token is usually the parameter name
+			// Skip type keywords
+			for i := len(tokens) - 1; i >= 0; i-- {
+				if !typeKeywords[tokens[i]] {
+					names = append(names, tokens[i])
+					break
+				}
+			}
+		}
+	}
+
+	return names
+}
+
+func splitParams(params string) []string {
+	var parts []string
+	var current bytes.Buffer
+	depth := 0
+
+	for _, char := range params + "," {
+		switch char {
+		case '(':
+			depth++
+			current.WriteRune(char)
+		case ')':
+			depth--
+			current.WriteRune(char)
+		case ',':
+			if depth == 0 {
+				parts = append(parts, strings.TrimSpace(current.String()))
+				current.Reset()
+			} else {
+				current.WriteRune(char)
+			}
+		default:
+			current.WriteRune(char)
+		}
+	}
+
+	return parts
+}
+
+func parseFunctions(content string) []Function {
+	var functions []Function
+
+	// Match function declarations: return_type function_name(params);
+	// Matches both mlx_* and _mlx_* functions
+	pattern := regexp.MustCompile(`\b((?:const\s+)?(?:struct\s+)?[\w\s]+?[\*\s]*)\s+(_?mlx_\w+)\s*\(([^)]*(?:\([^)]*\)[^)]*)*)\)\s*;`)
+
+	matches := pattern.FindAllStringSubmatch(content, -1)
+	for _, match := range matches {
+		returnType := strings.TrimSpace(match[1])
+		funcName := strings.TrimSpace(match[2])
+		params := strings.TrimSpace(match[3])
+
+		// Skip if this looks like a variable declaration
+		if params == "" || strings.Contains(params, "{") {
+			continue
+		}
+
+		// Clean up return type
+		returnType = strings.Join(strings.Fields(returnType), " ")
+
+		// Extract parameter names
+		paramNames := extractParamNames(params)
+
+		// Check if ARM64 guard is needed
+		needsGuard := needsARM64Guard(funcName, returnType, params)
+
+		functions = append(functions, Function{
+			Name:           funcName,
+			ReturnType:     returnType,
+			Params:         params,
+			ParamNames:     paramNames,
+			NeedsARM64Guard: needsGuard,
+		})
+	}
+
+	return functions
+}
+
+func needsARM64Guard(name, retType, params string) bool {
+	return strings.Contains(name, "float16") ||
+		strings.Contains(name, "bfloat16") ||
+		strings.Contains(retType, "float16_t") ||
+		strings.Contains(retType, "bfloat16_t") ||
+		strings.Contains(params, "float16_t") ||
+		strings.Contains(params, "bfloat16_t")
+}
+
+func generateWrapperFiles(functions []Function, headerPath, implPath string) error {
+	// Generate header file
+	var headerBuf bytes.Buffer
+
+	headerBuf.WriteString("// AUTO-GENERATED by generate_wrappers.go - DO NOT EDIT\n")
+	headerBuf.WriteString("// This file provides wrapper declarations for MLX-C functions that use dlopen/dlsym\n")
+	headerBuf.WriteString("//\n")
+	headerBuf.WriteString("// Strategy: Include MLX-C headers for type definitions, then provide wrapper\n")
+	headerBuf.WriteString("// functions that shadow the originals, allowing Go code to call them directly (e.g., C.mlx_add).\n")
+	headerBuf.WriteString("// Function pointers are defined in mlx.c (single compilation unit).\n\n")
+	headerBuf.WriteString("#ifndef MLX_WRAPPERS_H\n")
+	headerBuf.WriteString("#define MLX_WRAPPERS_H\n\n")
+
+	headerBuf.WriteString("// Include MLX headers for type definitions and original declarations\n")
+	headerBuf.WriteString("#include \"mlx/c/mlx.h\"\n")
+	headerBuf.WriteString("#include \"mlx_dynamic.h\"\n")
+	headerBuf.WriteString("#include <stdio.h>\n\n")
+
+	// Undef all MLX functions to avoid conflicts
+	headerBuf.WriteString("// Undefine any existing MLX function macros\n")
+	for _, fn := range functions {
+		headerBuf.WriteString(fmt.Sprintf("#undef %s\n", fn.Name))
+	}
+	headerBuf.WriteString("\n")
+
+	// Function pointer extern declarations
+	headerBuf.WriteString("// Function pointer declarations (defined in mlx.c, loaded via dlsym)\n")
+	for _, fn := range functions {
+		if fn.NeedsARM64Guard {
+			headerBuf.WriteString("#if defined(__aarch64__) || defined(_M_ARM64)\n")
+		}
+		headerBuf.WriteString(fmt.Sprintf("extern %s (*%s_ptr)(%s);\n", fn.ReturnType, fn.Name, fn.Params))
+		if fn.NeedsARM64Guard {
+			headerBuf.WriteString("#endif\n")
+		}
+	}
+	headerBuf.WriteString("\n")
+
+	// Initialization function declaration
+	headerBuf.WriteString("// Initialize all function pointers via dlsym (defined in mlx.c)\n")
+	headerBuf.WriteString("int mlx_load_functions(void* handle);\n\n")
+
+	// Wrapper function declarations
+	headerBuf.WriteString("// Wrapper function declarations that call through function pointers\n")
+	headerBuf.WriteString("// Go code calls these directly as C.mlx_* (no #define redirection needed)\n")
+	for _, fn := range functions {
+		if fn.NeedsARM64Guard {
+			headerBuf.WriteString("#if defined(__aarch64__) || defined(_M_ARM64)\n")
+		}
+		headerBuf.WriteString(fmt.Sprintf("%s %s(%s);\n", fn.ReturnType, fn.Name, fn.Params))
+		if fn.NeedsARM64Guard {
+			headerBuf.WriteString("#endif\n")
+		}
+		headerBuf.WriteString("\n")
+	}
+
+	headerBuf.WriteString("#endif // MLX_WRAPPERS_H\n")
+
+	// Write header file
+	if err := os.WriteFile(headerPath, headerBuf.Bytes(), 0644); err != nil {
+		return fmt.Errorf("failed to write header file: %w", err)
+	}
+
+	// Generate implementation file
+	var implBuf bytes.Buffer
+
+	implBuf.WriteString("// AUTO-GENERATED by generate_wrappers.go - DO NOT EDIT\n")
+	implBuf.WriteString("// This file contains the function pointer definitions and initialization\n")
+	implBuf.WriteString("// All function pointers are in a single compilation unit to avoid duplication\n\n")
+
+	implBuf.WriteString("#include \"mlx/c/mlx.h\"\n")
+	implBuf.WriteString("#include \"mlx_dynamic.h\"\n")
+	implBuf.WriteString("#include <stdio.h>\n")
+	implBuf.WriteString("#include <dlfcn.h>\n\n")
+
+	// Function pointer definitions
+	implBuf.WriteString("// Function pointer definitions\n")
+	for _, fn := range functions {
+		if fn.NeedsARM64Guard {
+			implBuf.WriteString("#if defined(__aarch64__) || defined(_M_ARM64)\n")
+		}
+		implBuf.WriteString(fmt.Sprintf("%s (*%s_ptr)(%s) = NULL;\n", fn.ReturnType, fn.Name, fn.Params))
+		if fn.NeedsARM64Guard {
+			implBuf.WriteString("#endif\n")
+		}
+	}
+	implBuf.WriteString("\n")
+
+	// Initialization function
+	implBuf.WriteString("// Initialize all function pointers via dlsym\n")
+	implBuf.WriteString("int mlx_load_functions(void* handle) {\n")
+	implBuf.WriteString("    if (handle == NULL) {\n")
+	implBuf.WriteString("        fprintf(stderr, \"MLX: Invalid library handle\\n\");\n")
+	implBuf.WriteString("        return -1;\n")
+	implBuf.WriteString("    }\n\n")
+
+	for _, fn := range functions {
+		if fn.NeedsARM64Guard {
+			implBuf.WriteString("#if defined(__aarch64__) || defined(_M_ARM64)\n")
+		}
+		implBuf.WriteString(fmt.Sprintf("    %s_ptr = dlsym(handle, \"%s\");\n", fn.Name, fn.Name))
+		implBuf.WriteString(fmt.Sprintf("    if (%s_ptr == NULL) {\n", fn.Name))
+		implBuf.WriteString(fmt.Sprintf("        fprintf(stderr, \"MLX: Failed to load symbol: %s\\n\");\n", fn.Name))
+		implBuf.WriteString("        return -1;\n")
+		implBuf.WriteString("    }\n")
+		if fn.NeedsARM64Guard {
+			implBuf.WriteString("#endif\n")
+		}
+	}
+
+	implBuf.WriteString("    return 0;\n")
+	implBuf.WriteString("}\n\n")
+
+	// Wrapper function implementations
+	implBuf.WriteString("// Wrapper function implementations that call through function pointers\n")
+	for _, fn := range functions {
+		if fn.NeedsARM64Guard {
+			implBuf.WriteString("#if defined(__aarch64__) || defined(_M_ARM64)\n")
+		}
+		implBuf.WriteString(fmt.Sprintf("%s %s(%s) {\n", fn.ReturnType, fn.Name, fn.Params))
+
+		// Call through function pointer
+		if fn.ReturnType != "void" {
+			implBuf.WriteString(fmt.Sprintf("    return %s_ptr(", fn.Name))
+		} else {
+			implBuf.WriteString(fmt.Sprintf("    %s_ptr(", fn.Name))
+		}
+
+		// Pass parameters
+		implBuf.WriteString(strings.Join(fn.ParamNames, ", "))
+		implBuf.WriteString(");\n")
+		implBuf.WriteString("}\n")
+		if fn.NeedsARM64Guard {
+			implBuf.WriteString("#endif\n")
+		}
+		implBuf.WriteString("\n")
+	}
+
+	// Write implementation file
+	if err := os.WriteFile(implPath, implBuf.Bytes(), 0644); err != nil {
+		return fmt.Errorf("failed to write implementation file: %w", err)
+	}
+
+	return nil
+}
+
+func main() {
+	flag.Usage = func() {
+		fmt.Fprintf(flag.CommandLine.Output(), "Usage: go run generate_wrappers.go <mlx-c-include-dir> <output-header> [output-impl]\n")
+		fmt.Fprintf(flag.CommandLine.Output(), "Generate MLX-C dynamic loading wrappers.\n\n")
+		flag.PrintDefaults()
+	}
+	flag.Parse()
+
+	args := flag.Args()
+	if len(args) < 2 {
+		fmt.Fprintf(flag.CommandLine.Output(), "ERROR: Missing required arguments\n\n")
+		flag.Usage()
+		os.Exit(1)
+	}
+
+	headerDir := args[0]
+	outputHeader := args[1]
+	// Default implementation file is same name with .c extension
+	outputImpl := outputHeader
+	if len(args) > 2 {
+		outputImpl = args[2]
+	} else if strings.HasSuffix(outputHeader, ".h") {
+		outputImpl = outputHeader[:len(outputHeader)-2] + ".c"
+	}
+
+	// Check if header directory exists
+	if _, err := os.Stat(headerDir); os.IsNotExist(err) {
+		fmt.Fprintf(os.Stderr, "ERROR: MLX-C headers directory not found at: %s\n\n", headerDir)
+		fmt.Fprintf(os.Stderr, "Please run CMake first to download MLX-C dependencies:\n")
+		fmt.Fprintf(os.Stderr, "  cmake -B build\n\n")
+		fmt.Fprintf(os.Stderr, "The CMake build will download and extract MLX-C headers needed for wrapper generation.\n")
+		os.Exit(1)
+	}
+
+	fmt.Fprintf(os.Stderr, "Parsing MLX-C headers from: %s\n", headerDir)
+
+	// Find all headers
+	headers, err := findHeaders(headerDir)
+	if err != nil {
+		fmt.Fprintf(os.Stderr, "ERROR: Failed to find header files: %v\n", err)
+		os.Exit(1)
+	}
+	fmt.Fprintf(os.Stderr, "Found %d header files\n", len(headers))
+
+	// Parse all headers
+	var allFunctions []Function
+	seen := make(map[string]bool)
+
+	for _, header := range headers {
+		content, err := os.ReadFile(header)
+		if err != nil {
+			fmt.Fprintf(os.Stderr, "Error reading %s: %v\n", header, err)
+			continue
+		}
+
+		cleaned := cleanContent(string(content))
+		functions := parseFunctions(cleaned)
+
+		// Deduplicate
+		for _, fn := range functions {
+			if !seen[fn.Name] {
+				seen[fn.Name] = true
+				allFunctions = append(allFunctions, fn)
+			}
+		}
+	}
+
+	fmt.Fprintf(os.Stderr, "Found %d unique function declarations\n", len(allFunctions))
+
+	// Generate wrapper files
+	if err := generateWrapperFiles(allFunctions, outputHeader, outputImpl); err != nil {
+		fmt.Fprintf(os.Stderr, "ERROR: Failed to generate wrapper files: %v\n", err)
+		os.Exit(1)
+	}
+
+	fmt.Fprintf(os.Stderr, "Generated %s and %s successfully\n", outputHeader, outputImpl)
+}

x/imagegen/mlx/mlx.go

@@ -3,12 +3,13 @@
 package mlx
 
 /*
-#cgo CFLAGS: -O3 -I${SRCDIR}/../../../build/_deps/mlx-c-src
-#cgo LDFLAGS: -L${SRCDIR}/../../../build/lib/ollama/ -lmlxc -Wl,-rpath,${SRCDIR}/../../../build/lib/ollama/
+#cgo CFLAGS: -O3 -I${SRCDIR}/../../../build/_deps/mlx-c-src -I${SRCDIR}
 #cgo darwin LDFLAGS: -lc++ -framework Metal -framework Foundation -framework Accelerate
-#cgo linux LDFLAGS: -lstdc++ -lcuda -lcudart -lnvrtc
+#cgo linux LDFLAGS: -lstdc++ -ldl
+#cgo windows LDFLAGS: -lstdc++
 
-#include "mlx/c/mlx.h"
+// Use generated wrappers instead of direct MLX headers
+#include "mlx.h"
 #include <stdlib.h>
 #include <stdint.h>
 #include <string.h>
@@ -42,192 +43,6 @@ static inline mlx_stream cpu_stream() {
 // CGO noescape/nocallback hints to reduce CGO overhead
 // noescape: pointers won't escape, no heap allocation needed
 // nocallback: function won't call back into Go
-#cgo noescape mlx_add
-#cgo nocallback mlx_add
-#cgo noescape mlx_subtract
-#cgo nocallback mlx_subtract
-#cgo noescape mlx_multiply
-#cgo nocallback mlx_multiply
-#cgo noescape mlx_divide
-#cgo nocallback mlx_divide
-#cgo noescape mlx_negative
-#cgo nocallback mlx_negative
-#cgo noescape mlx_abs
-#cgo nocallback mlx_abs
-#cgo noescape mlx_exp
-#cgo nocallback mlx_exp
-#cgo noescape mlx_log
-#cgo nocallback mlx_log
-#cgo noescape mlx_sqrt
-#cgo nocallback mlx_sqrt
-#cgo noescape mlx_rsqrt
-#cgo nocallback mlx_rsqrt
-#cgo noescape mlx_square
-#cgo nocallback mlx_square
-#cgo noescape mlx_power
-#cgo nocallback mlx_power
-#cgo noescape mlx_erf
-#cgo nocallback mlx_erf
-#cgo noescape mlx_sigmoid
-#cgo nocallback mlx_sigmoid
-#cgo noescape mlx_tanh
-#cgo nocallback mlx_tanh
-#cgo noescape mlx_sin
-#cgo nocallback mlx_sin
-#cgo noescape mlx_cos
-#cgo nocallback mlx_cos
-#cgo noescape mlx_maximum
-#cgo nocallback mlx_maximum
-#cgo noescape mlx_minimum
-#cgo nocallback mlx_minimum
-#cgo noescape mlx_clip
-#cgo nocallback mlx_clip
-#cgo noescape mlx_sum
-#cgo nocallback mlx_sum
-#cgo noescape mlx_sum_axis
-#cgo nocallback mlx_sum_axis
-#cgo noescape mlx_mean
-#cgo nocallback mlx_mean
-#cgo noescape mlx_mean_axis
-#cgo nocallback mlx_mean_axis
-#cgo noescape mlx_var_axis
-#cgo nocallback mlx_var_axis
-#cgo noescape mlx_argmax
-#cgo nocallback mlx_argmax
-#cgo noescape mlx_argmax_axis
-#cgo nocallback mlx_argmax_axis
-#cgo noescape mlx_softmax_axis
-#cgo nocallback mlx_softmax_axis
-#cgo noescape mlx_cumsum
-#cgo nocallback mlx_cumsum
-#cgo noescape mlx_matmul
-#cgo nocallback mlx_matmul
-#cgo noescape mlx_addmm
-#cgo nocallback mlx_addmm
-#cgo noescape mlx_gather_mm
-#cgo nocallback mlx_gather_mm
-#cgo noescape mlx_gather_qmm
-#cgo nocallback mlx_gather_qmm
-#cgo noescape mlx_reshape
-#cgo nocallback mlx_reshape
-#cgo noescape mlx_transpose_axes
-#cgo nocallback mlx_transpose_axes
-#cgo noescape mlx_expand_dims
-#cgo nocallback mlx_expand_dims
-#cgo noescape mlx_squeeze_axis
-#cgo nocallback mlx_squeeze_axis
-#cgo noescape mlx_flatten
-#cgo nocallback mlx_flatten
-#cgo noescape mlx_concatenate_axis
-#cgo nocallback mlx_concatenate_axis
-#cgo noescape mlx_slice
-#cgo nocallback mlx_slice
-#cgo noescape mlx_slice_update
-#cgo nocallback mlx_slice_update
-#cgo noescape mlx_as_strided
-#cgo nocallback mlx_as_strided
-#cgo noescape mlx_view
-#cgo nocallback mlx_view
-#cgo noescape mlx_contiguous
-#cgo nocallback mlx_contiguous
-#cgo noescape mlx_pad
-#cgo nocallback mlx_pad
-#cgo noescape mlx_tile
-#cgo nocallback mlx_tile
-#cgo noescape mlx_take_axis
-#cgo nocallback mlx_take_axis
-#cgo noescape mlx_take_along_axis
-#cgo nocallback mlx_take_along_axis
-#cgo noescape mlx_put_along_axis
-#cgo nocallback mlx_put_along_axis
-#cgo noescape mlx_where
-#cgo nocallback mlx_where
-#cgo noescape mlx_argsort_axis
-#cgo nocallback mlx_argsort_axis
-#cgo noescape mlx_argpartition_axis
-#cgo nocallback mlx_argpartition_axis
-#cgo noescape mlx_topk_axis
-#cgo nocallback mlx_topk_axis
-#cgo noescape mlx_less
-#cgo nocallback mlx_less
-#cgo noescape mlx_greater_equal
-#cgo nocallback mlx_greater_equal
-#cgo noescape mlx_logical_and
-#cgo nocallback mlx_logical_and
-#cgo noescape mlx_zeros
-#cgo nocallback mlx_zeros
-#cgo noescape mlx_zeros_like
-#cgo nocallback mlx_zeros_like
-#cgo noescape mlx_ones
-#cgo nocallback mlx_ones
-#cgo noescape mlx_full
-#cgo nocallback mlx_full
-#cgo noescape mlx_arange
-#cgo nocallback mlx_arange
-#cgo noescape mlx_linspace
-#cgo nocallback mlx_linspace
-#cgo noescape mlx_tri
-#cgo nocallback mlx_tri
-#cgo noescape mlx_astype
-#cgo nocallback mlx_astype
-#cgo noescape mlx_fast_rms_norm
-#cgo nocallback mlx_fast_rms_norm
-#cgo noescape mlx_fast_rope
-#cgo nocallback mlx_fast_rope
-#cgo noescape mlx_fast_scaled_dot_product_attention
-#cgo nocallback mlx_fast_scaled_dot_product_attention
-#cgo noescape mlx_conv2d
-#cgo nocallback mlx_conv2d
-#cgo noescape mlx_conv3d
-#cgo nocallback mlx_conv3d
-#cgo noescape mlx_random_key
-#cgo nocallback mlx_random_key
-#cgo noescape mlx_random_split
-#cgo nocallback mlx_random_split
-#cgo noescape mlx_random_categorical_num_samples
-#cgo nocallback mlx_random_categorical_num_samples
-#cgo noescape mlx_random_normal
-#cgo nocallback mlx_random_normal
-#cgo noescape mlx_random_uniform
-#cgo nocallback mlx_random_uniform
-#cgo noescape mlx_array_eval
-#cgo nocallback mlx_array_eval
-#cgo noescape mlx_eval
-#cgo nocallback mlx_eval
-#cgo noescape mlx_async_eval
-#cgo nocallback mlx_async_eval
-#cgo noescape mlx_synchronize
-#cgo nocallback mlx_synchronize
-#cgo noescape mlx_array_new
-#cgo nocallback mlx_array_new
-#cgo noescape mlx_array_new_data
-#cgo nocallback mlx_array_new_data
-#cgo noescape mlx_array_new_float
-#cgo nocallback mlx_array_new_float
-#cgo noescape mlx_array_free
-#cgo nocallback mlx_array_free
-#cgo noescape mlx_array_size
-#cgo nocallback mlx_array_size
-#cgo noescape mlx_array_ndim
-#cgo nocallback mlx_array_ndim
-#cgo noescape mlx_array_dim
-#cgo nocallback mlx_array_dim
-#cgo noescape mlx_array_dtype
-#cgo nocallback mlx_array_dtype
-#cgo noescape mlx_array_item_int32
-#cgo nocallback mlx_array_item_int32
-#cgo noescape mlx_vector_array_new_data
-#cgo nocallback mlx_vector_array_new_data
-#cgo noescape mlx_vector_array_free
-#cgo nocallback mlx_vector_array_free
-#cgo noescape mlx_array_new_int
-#cgo nocallback mlx_array_new_int
-#cgo noescape mlx_stream_new_device
-#cgo nocallback mlx_stream_new_device
-#cgo noescape mlx_get_default_stream
-#cgo nocallback mlx_get_default_stream
-#cgo noescape mlx_set_default_stream
-#cgo nocallback mlx_set_default_stream
 */
 import "C"
 import (
@@ -1322,6 +1137,27 @@ func RMSNormNoWeight(x *Array, eps float32) *Array {
 	return RMSNorm(x, ones, eps)
 }
 
+// LayerNorm applies layer normalization without learnable params
+// (x - mean) / sqrt(var + eps)
+func LayerNorm(x *Array, eps float32) *Array {
+	return LayerNormWithWeightBias(x, nil, nil, eps)
+}
+
+// LayerNormWithWeightBias computes layer normalization using mlx.fast
+// weight and bias can be nil for elementwise_affine=False
+func LayerNormWithWeightBias(x, weight, bias *Array, eps float32) *Array {
+	res := C.mlx_array_new()
+	var wc, bc C.mlx_array
+	if weight != nil {
+		wc = weight.c
+	}
+	if bias != nil {
+		bc = bias.c
+	}
+	C.mlx_fast_layer_norm(&res, x.c, wc, bc, C.float(eps), C.default_stream())
+	return newArray(res)
+}
+
 // RoPE applies rotary position embeddings using mlx.fast
 func RoPE(x *Array, dims int, traditional bool, base, scale float32, offset int) *Array {
 	res := C.mlx_array_new()
@@ -1796,7 +1632,57 @@ func ArgmaxKeepArray(logits *Array) *Array {
 var RandomState = []*Array{nil}
 var randomStateMu sync.Mutex
 
+var mlxInitialized bool
+var mlxInitError error
+
+// InitMLX initializes the MLX library by dynamically loading libmlxc.
+// This must be called before using any MLX functions.
+// Returns an error if the library cannot be loaded.
+func InitMLX() error {
+	if mlxInitialized {
+		return mlxInitError
+	}
+
+	// Try to load the MLX dynamic library
+	ret := C.mlx_dynamic_init()
+	if ret != 0 {
+		errMsg := C.GoString(C.mlx_dynamic_error())
+		mlxInitError = fmt.Errorf("failed to initialize MLX: %s", errMsg)
+		return mlxInitError
+	}
+
+	// Initialize all function pointers via dlsym
+	handle := C.mlx_get_handle()
+	ret = C.mlx_load_functions(handle)
+	if ret != 0 {
+		mlxInitError = fmt.Errorf("failed to load MLX function symbols")
+		return mlxInitError
+	}
+
+	mlxInitialized = true
+	mlxInitError = nil
+	return nil
+}
+
+// IsMLXAvailable returns whether MLX was successfully initialized
+func IsMLXAvailable() bool {
+	return mlxInitialized && mlxInitError == nil
+}
+
+// GetMLXInitError returns any error that occurred during MLX initialization
+func GetMLXInitError() error {
+	return mlxInitError
+}
+
 func init() {
+	// Initialize MLX dynamic library first
+	if err := InitMLX(); err != nil {
+		// Don't panic in init - let the caller handle the error
+		// Store the error for later retrieval
+		mlxInitError = err
+		return
+	}
+
 	// Lock main goroutine to OS thread for CUDA context stability.
 	// CUDA contexts are bound to threads; Go can migrate goroutines between threads.
 	runtime.LockOSThread()

x/imagegen/mlx/mlx_dynamic.c

@@ -0,0 +1,144 @@
+// mlx_dynamic.c - Dynamic loading wrapper for MLX-C library
+// This file provides runtime dynamic loading of libmlxc instead of link-time binding
+
+#include "mlx_dynamic.h"
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#ifdef _WIN32
+#include <windows.h>
+typedef HMODULE lib_handle_t;
+#define LOAD_LIB(path) LoadLibraryA(path)
+#define GET_SYMBOL(handle, name) GetProcAddress(handle, name)
+#define CLOSE_LIB(handle) FreeLibrary(handle)
+#define LIB_ERROR() "LoadLibrary failed"
+#else
+#include <dlfcn.h>
+typedef void* lib_handle_t;
+#define LOAD_LIB(path) dlopen(path, RTLD_LAZY | RTLD_GLOBAL)
+#define GET_SYMBOL(handle, name) dlsym(handle, name)
+#define CLOSE_LIB(handle) dlclose(handle)
+#define LIB_ERROR() dlerror()
+#ifdef __APPLE__
+#include <mach-o/dyld.h>
+#include <libgen.h>
+#endif
+#endif
+
+static lib_handle_t mlx_handle = NULL;
+static int mlx_initialized = 0;
+static char mlx_error_buffer[512] = {0};
+
+#ifdef __APPLE__
+// Get path to library in same directory as executable
+static char* get_exe_relative_path(const char* libname) {
+    static char path[1024];
+    uint32_t size = sizeof(path);
+    if (_NSGetExecutablePath(path, &size) != 0) {
+        return NULL;
+    }
+    // Get directory of executable
+    char* dir = dirname(path);
+    static char fullpath[1024];
+    snprintf(fullpath, sizeof(fullpath), "%s/%s", dir, libname);
+    return fullpath;
+}
+#endif
+
+// Try to load library from a specific path
+static int try_load_lib(const char* path) {
+    if (!path) return 0;
+    mlx_handle = LOAD_LIB(path);
+    return mlx_handle != NULL;
+}
+
+// Initialize MLX dynamic library
+// Returns 0 on success, -1 on failure
+// On failure, call mlx_dynamic_error() to get error message
+int mlx_dynamic_init(void) {
+    if (mlx_initialized) {
+        return 0;  // Already initialized
+    }
+
+    const char* lib_path = NULL;
+    const char* tried_paths[8] = {0};
+    int num_tried = 0;
+
+#ifdef _WIN32
+    // Windows: try same directory as executable
+    lib_path = "libmlxc.dll";
+    tried_paths[num_tried++] = lib_path;
+    if (try_load_lib(lib_path)) goto success;
+#elif defined(__APPLE__)
+    // macOS: try executable directory first
+    lib_path = get_exe_relative_path("libmlxc.dylib");
+    if (lib_path) {
+        tried_paths[num_tried++] = lib_path;
+        if (try_load_lib(lib_path)) goto success;
+    }
+    // Try build directory (for tests run from repo root)
+    lib_path = "./build/lib/ollama/libmlxc.dylib";
+    tried_paths[num_tried++] = lib_path;
+    if (try_load_lib(lib_path)) goto success;
+    // Fallback to system paths
+    lib_path = "libmlxc.dylib";
+    tried_paths[num_tried++] = lib_path;
+    if (try_load_lib(lib_path)) goto success;
+#else
+    // Linux: try build directory first (for tests)
+    lib_path = "./build/lib/ollama/libmlxc.so";
+    tried_paths[num_tried++] = lib_path;
+    if (try_load_lib(lib_path)) goto success;
+    // Fallback to system paths
+    lib_path = "libmlxc.so";
+    tried_paths[num_tried++] = lib_path;
+    if (try_load_lib(lib_path)) goto success;
+#endif
+
+    // Failed to load library - build error message with all tried paths
+    {
+        const char* err = LIB_ERROR();
+        int offset = snprintf(mlx_error_buffer, sizeof(mlx_error_buffer),
+                     "MLX: Failed to load libmlxc library. Tried: ");
+        for (int i = 0; i < num_tried && offset < (int)sizeof(mlx_error_buffer) - 50; i++) {
+            offset += snprintf(mlx_error_buffer + offset, sizeof(mlx_error_buffer) - offset,
+                             "%s%s", i > 0 ? ", " : "", tried_paths[i]);
+        }
+        if (err) {
+            snprintf(mlx_error_buffer + offset, sizeof(mlx_error_buffer) - offset,
+                    ". Last error: %s", err);
+        }
+    }
+    return -1;
+
+success:
+    mlx_initialized = 1;
+    snprintf(mlx_error_buffer, sizeof(mlx_error_buffer),
+             "MLX: Successfully loaded %s", lib_path ? lib_path : "library");
+    return 0;
+}
+
+// Get the last error message
+const char* mlx_dynamic_error(void) {
+    return mlx_error_buffer;
+}
+
+// Check if MLX is initialized
+int mlx_dynamic_is_initialized(void) {
+    return mlx_initialized;
+}
+
+// Get the library handle (for use by generated wrappers)
+void* mlx_get_handle(void) {
+    return mlx_handle;
+}
+
+// Cleanup (optional, called at program exit)
+void mlx_dynamic_cleanup(void) {
+    if (mlx_handle != NULL) {
+        CLOSE_LIB(mlx_handle);
+        mlx_handle = NULL;
+        mlx_initialized = 0;
+    }
+}

x/imagegen/mlx/mlx_dynamic.h

@@ -0,0 +1,29 @@
+// mlx_dynamic.h - Dynamic loading interface for MLX-C library
+#ifndef MLX_DYNAMIC_H
+#define MLX_DYNAMIC_H
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// Initialize the MLX dynamic library
+// Returns 0 on success, -1 on failure
+int mlx_dynamic_init(void);
+
+// Get the last error message from dynamic loading
+const char* mlx_dynamic_error(void);
+
+// Check if MLX is initialized
+int mlx_dynamic_is_initialized(void);
+
+// Get the library handle (for use by generated wrappers)
+void* mlx_get_handle(void);
+
+// Cleanup resources (optional, for clean shutdown)
+void mlx_dynamic_cleanup(void);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif // MLX_DYNAMIC_H

x/imagegen/mlx/mlx_test.go

@@ -4,9 +4,30 @@ package mlx
 
 import (
 	"fmt"
+	"os"
+	"path/filepath"
+	"runtime"
 	"testing"
 )
 
+// TestMain initializes MLX before running tests.
+// If MLX libraries are not available, tests are skipped.
+func TestMain(m *testing.M) {
+	// Change to repo root so ./build/lib/ollama/ path works
+	_, thisFile, _, _ := runtime.Caller(0)
+	repoRoot := filepath.Join(filepath.Dir(thisFile), "..", "..", "..")
+	if err := os.Chdir(repoRoot); err != nil {
+		fmt.Printf("Failed to change to repo root: %v\n", err)
+		os.Exit(1)
+	}
+
+	if err := InitMLX(); err != nil {
+		fmt.Printf("Skipping MLX tests: %v\n", err)
+		os.Exit(0)
+	}
+	os.Exit(m.Run())
+}
+
 // TestBasicCleanup verifies non-kept arrays are freed and kept arrays survive.
 func TestBasicCleanup(t *testing.T) {
 	weight := NewArrayFloat32([]float32{1, 2, 3, 4}, []int32{2, 2})

x/imagegen/models/flux2/flux2.go

@@ -0,0 +1,539 @@
+//go:build mlx
+
+// Package flux2 implements the FLUX.2 Klein diffusion transformer model.
+// Klein is a 4B parameter distilled model that supports sub-second inference.
+package flux2
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"image"
+	"math"
+	"time"
+
+	"github.com/ollama/ollama/x/imagegen"
+	"github.com/ollama/ollama/x/imagegen/mlx"
+	"github.com/ollama/ollama/x/imagegen/models/qwen3"
+	"github.com/ollama/ollama/x/imagegen/tokenizer"
+	"golang.org/x/image/draw"
+)
+
+// GenerateConfig holds all options for image generation.
+type GenerateConfig struct {
+	Prompt        string
+	Width         int32                 // Image width (default: 1024)
+	Height        int32                 // Image height (default: 1024)
+	Steps         int                   // Denoising steps (default: 4 for Klein)
+	GuidanceScale float32               // Guidance scale (default: 1.0, Klein doesn't need CFG)
+	Seed          int64                 // Random seed
+	Progress      func(step, totalSteps int) // Optional progress callback
+	CapturePath   string                // GPU capture path (debug)
+	InputImages   []image.Image         // Reference images for image conditioning (already loaded)
+}
+
+// Model represents a FLUX.2 Klein model.
+type Model struct {
+	ModelName       string
+	Tokenizer       *tokenizer.Tokenizer
+	TextEncoder     *qwen3.TextEncoder
+	Transformer     *Flux2Transformer2DModel
+	VAE             *AutoencoderKLFlux2
+	SchedulerConfig *SchedulerConfig
+}
+
+// TextEncoderLayerIndices are the layers from which to extract text embeddings.
+// Diffusers uses hidden_states[9, 18, 27]. In Python, hidden_states[0] is the embedding
+// output before any layers, so hidden_states[9] = after layer 8 (0-indexed).
+// Go's ForwardWithLayerOutputs captures after layer i runs, so we use [8, 17, 26].
+var TextEncoderLayerIndices = []int{8, 17, 26}
+
+// Load loads the FLUX.2 Klein model from ollama blob storage.
+func (m *Model) Load(modelName string) error {
+	fmt.Printf("Loading FLUX.2 Klein model from manifest: %s...\n", modelName)
+	start := time.Now()
+
+	if mlx.GPUIsAvailable() {
+		mlx.SetDefaultDeviceGPU()
+		mlx.EnableCompile()
+	}
+
+	m.ModelName = modelName
+
+	// Load manifest
+	manifest, err := imagegen.LoadManifest(modelName)
+	if err != nil {
+		return fmt.Errorf("load manifest: %w", err)
+	}
+
+	// Load tokenizer
+	fmt.Print("  Loading tokenizer... ")
+	tokData, err := manifest.ReadConfig("tokenizer/tokenizer.json")
+	if err != nil {
+		return fmt.Errorf("tokenizer: %w", err)
+	}
+
+	tokConfig := &tokenizer.TokenizerConfig{}
+	if data, err := manifest.ReadConfig("tokenizer/tokenizer_config.json"); err == nil {
+		tokConfig.TokenizerConfigJSON = data
+	}
+	if data, err := manifest.ReadConfig("tokenizer/generation_config.json"); err == nil {
+		tokConfig.GenerationConfigJSON = data
+	}
+	if data, err := manifest.ReadConfig("tokenizer/special_tokens_map.json"); err == nil {
+		tokConfig.SpecialTokensMapJSON = data
+	}
+
+	tok, err := tokenizer.LoadFromBytesWithConfig(tokData, tokConfig)
+	if err != nil {
+		return fmt.Errorf("tokenizer: %w", err)
+	}
+	m.Tokenizer = tok
+	fmt.Println("✓")
+
+	// Load text encoder
+	m.TextEncoder = &qwen3.TextEncoder{}
+	if err := m.TextEncoder.Load(manifest, "text_encoder/config.json"); err != nil {
+		return fmt.Errorf("text encoder: %w", err)
+	}
+
+	// Load transformer
+	m.Transformer = &Flux2Transformer2DModel{}
+	if err := m.Transformer.Load(manifest); err != nil {
+		return fmt.Errorf("transformer: %w", err)
+	}
+
+	// Load VAE
+	m.VAE = &AutoencoderKLFlux2{}
+	if err := m.VAE.Load(manifest); err != nil {
+		return fmt.Errorf("VAE: %w", err)
+	}
+
+	// Evaluate all weights in a single batch (reduces GPU sync overhead)
+	fmt.Print("  Evaluating weights... ")
+	allWeights := mlx.Collect(m.TextEncoder)
+	allWeights = append(allWeights, mlx.Collect(m.Transformer)...)
+	allWeights = append(allWeights, mlx.Collect(m.VAE)...)
+	mlx.Eval(allWeights...)
+	fmt.Println("✓")
+
+	// Load scheduler config
+	m.SchedulerConfig = DefaultSchedulerConfig()
+	if schedData, err := manifest.ReadConfig("scheduler/scheduler_config.json"); err == nil {
+		if err := json.Unmarshal(schedData, m.SchedulerConfig); err != nil {
+			fmt.Printf("  Warning: failed to parse scheduler config: %v\n", err)
+		}
+	}
+
+	mem := mlx.MetalGetActiveMemory()
+	fmt.Printf("  Loaded in %.2fs (%.1f GB VRAM)\n", time.Since(start).Seconds(), float64(mem)/(1024*1024*1024))
+
+	return nil
+}
+
+// Generate creates an image from a prompt.
+func (m *Model) Generate(prompt string, width, height int32, steps int, seed int64) (*mlx.Array, error) {
+	return m.GenerateFromConfig(context.Background(), &GenerateConfig{
+		Prompt: prompt,
+		Width:  width,
+		Height: height,
+		Steps:  steps,
+		Seed:   seed,
+	})
+}
+
+// GenerateWithProgress creates an image with progress callback.
+func (m *Model) GenerateWithProgress(prompt string, width, height int32, steps int, seed int64, progress func(step, totalSteps int)) (*mlx.Array, error) {
+	return m.GenerateFromConfig(context.Background(), &GenerateConfig{
+		Prompt:   prompt,
+		Width:    width,
+		Height:   height,
+		Steps:    steps,
+		Seed:     seed,
+		Progress: progress,
+	})
+}
+
+// GenerateFromConfig generates an image using the unified config struct.
+func (m *Model) GenerateFromConfig(ctx context.Context, cfg *GenerateConfig) (*mlx.Array, error) {
+	start := time.Now()
+	result, err := m.generate(ctx, cfg)
+	if err != nil {
+		return nil, err
+	}
+	fmt.Printf("Generated in %.2fs (%d steps)\n", time.Since(start).Seconds(), cfg.Steps)
+	return result, nil
+}
+
+// GenerateImage implements runner.ImageModel interface.
+func (m *Model) GenerateImage(ctx context.Context, prompt string, width, height int32, steps int, seed int64, progress func(step, total int)) (*mlx.Array, error) {
+	return m.GenerateFromConfig(ctx, &GenerateConfig{
+		Prompt:   prompt,
+		Width:    width,
+		Height:   height,
+		Steps:    steps,
+		Seed:     seed,
+		Progress: progress,
+	})
+}
+
+// MaxOutputPixels is the maximum output resolution (4 megapixels, ~2048x2048)
+const MaxOutputPixels = 2048 * 2048
+
+// MaxRefPixels is the maximum resolution for reference images (smaller to reduce attention memory)
+const MaxRefPixels = 728 * 728
+
+// generate is the internal denoising pipeline.
+func (m *Model) generate(ctx context.Context, cfg *GenerateConfig) (*mlx.Array, error) {
+	// Enable MLX compilation for fused kernels
+	mlx.EnableCompile()
+
+	// Apply defaults
+	if cfg.Steps <= 0 {
+		cfg.Steps = 4 // Klein default: 4 steps for distilled model
+	}
+	if cfg.GuidanceScale <= 0 {
+		cfg.GuidanceScale = 1.0 // Klein doesn't need guidance
+	}
+
+	// Determine output dimensions
+	if len(cfg.InputImages) > 0 {
+		// With input images, compute missing dimension from aspect ratio
+		// Images are already EXIF-rotated by the caller
+		bounds := cfg.InputImages[0].Bounds()
+		imgW, imgH := bounds.Dx(), bounds.Dy()
+		aspectRatio := float64(imgH) / float64(imgW)
+		if cfg.Width > 0 && cfg.Height <= 0 {
+			// Width specified, compute height
+			cfg.Height = int32(math.Round(float64(cfg.Width)*aspectRatio/16) * 16)
+		} else if cfg.Height > 0 && cfg.Width <= 0 {
+			// Height specified, compute width
+			cfg.Width = int32(math.Round(float64(cfg.Height)/aspectRatio/16) * 16)
+		} else if cfg.Width <= 0 && cfg.Height <= 0 {
+			// Neither specified, use input dimensions
+			cfg.Width = int32(imgW)
+			cfg.Height = int32(imgH)
+		}
+	}
+	if cfg.Width <= 0 {
+		cfg.Width = 1024
+	}
+	if cfg.Height <= 0 {
+		cfg.Height = 1024
+	}
+
+	// Cap to max pixels, preserve aspect ratio, round to multiple of 16
+	pixels := int(cfg.Width) * int(cfg.Height)
+	if pixels > MaxOutputPixels {
+		scale := math.Sqrt(float64(MaxOutputPixels) / float64(pixels))
+		cfg.Width = int32(math.Round(float64(cfg.Width) * scale / 16) * 16)
+		cfg.Height = int32(math.Round(float64(cfg.Height) * scale / 16) * 16)
+	}
+	cfg.Height = int32((cfg.Height + 8) / 16 * 16) // round to nearest 16
+	cfg.Width = int32((cfg.Width + 8) / 16 * 16)
+	fmt.Printf("  Output: %dx%d\n", cfg.Width, cfg.Height)
+
+	tcfg := m.Transformer.TransformerConfig
+	patchSize := m.VAE.Config.PatchSize
+
+	// Latent dimensions: image / 8 (VAE downscale) / patch_size
+	latentH := cfg.Height / 8
+	latentW := cfg.Width / 8
+	patchH := latentH / patchSize[0]
+	patchW := latentW / patchSize[1]
+	imgSeqLen := patchH * patchW
+
+	// Text encoding with multi-layer extraction (no padding, use true sequence length)
+	fmt.Print("  Encoding prompt... ")
+	promptEmbeds, textLen := m.TextEncoder.EncodePromptWithLayers(m.Tokenizer, cfg.Prompt, 512, TextEncoderLayerIndices, false)
+	fmt.Println("✓")
+
+	// Encode reference images if provided
+	var refTokens *ImageCondTokens
+	var refHeights, refWidths []int32
+	if len(cfg.InputImages) > 0 {
+		fmt.Printf("  Encoding %d reference image(s):\n", len(cfg.InputImages))
+
+		var err error
+		refTokens, err = m.EncodeImageRefs(cfg.InputImages)
+		if err != nil {
+			return nil, fmt.Errorf("encode reference images: %w", err)
+		}
+
+		// Extract heights/widths for RoPE computation (same limits as EncodeImageRefs)
+		limitPixels := MaxRefPixels
+		if len(cfg.InputImages) > 1 {
+			limitPixels = MaxRefPixels / 2
+		}
+		for _, img := range cfg.InputImages {
+			_, w, h := PrepareImage(img, limitPixels)
+			refHeights = append(refHeights, int32(h/16))
+			refWidths = append(refWidths, int32(w/16))
+		}
+	}
+
+	// Scheduler
+	scheduler := NewFlowMatchScheduler(m.SchedulerConfig)
+	scheduler.SetTimestepsWithMu(cfg.Steps, CalculateShift(imgSeqLen, cfg.Steps))
+
+	// Init latents in packed form [B, C*4, H/2, W/2] like diffusers
+	// diffusers creates noise in [B, 128, 64, 64] and packs to [B, 4096, 128]
+	latentChannels := m.VAE.Config.LatentChannels
+	packedChannels := latentChannels * 4 // 32 * 4 = 128
+	latents := scheduler.InitNoise([]int32{1, packedChannels, patchH, patchW}, cfg.Seed)
+
+	// Pack latents (transpose): [B, C, H, W] -> [B, H*W, C]
+	// This matches diffusers' _pack_latents
+	patches := packLatents(latents)
+	noiseSeqLen := patches.Shape()[1]
+
+	// RoPE cache - includes reference images if present
+	rope := PrepareRoPECache(textLen, patchH, patchW, tcfg.AxesDimsRoPE, tcfg.RopeTheta, refHeights, refWidths, ImageRefScale)
+
+	// Cleanup setup arrays when done
+	defer func() {
+		rope.Cos.Free()
+		rope.Sin.Free()
+		promptEmbeds.Free()
+		if refTokens != nil {
+			refTokens.Tokens.Free()
+		}
+	}()
+
+	// Pre-compute all timesteps before the loop to avoid per-step tensor creation
+	timesteps := make([]*mlx.Array, cfg.Steps)
+	for i := 0; i < cfg.Steps; i++ {
+		tCurr := scheduler.Timesteps[i] / float32(m.SchedulerConfig.NumTrainTimesteps)
+		timesteps[i] = mlx.ToBFloat16(mlx.NewArray([]float32{tCurr}, []int32{1}))
+	}
+
+	// Evaluate setup arrays
+	fmt.Print("  Evaluating setup... ")
+	setupStart := time.Now()
+	toEval := []*mlx.Array{promptEmbeds, patches, rope.Cos, rope.Sin}
+	toEval = append(toEval, timesteps...)
+	if refTokens != nil {
+		toEval = append(toEval, refTokens.Tokens)
+	}
+	mlx.Eval(toEval...)
+	mlx.MetalResetPeakMemory() // Reset peak to measure generation separately
+	fmt.Printf("✓ (%.2fs, %.1f GB)\n", time.Since(setupStart).Seconds(),
+		float64(mlx.MetalGetActiveMemory())/(1024*1024*1024))
+
+	if cfg.Progress != nil {
+		cfg.Progress(0, cfg.Steps)
+	}
+
+	loopStart := time.Now()
+	stepStart := time.Now()
+
+	// Denoising loop
+	for i := 0; i < cfg.Steps; i++ {
+		// Check for cancellation
+		if ctx != nil {
+			select {
+			case <-ctx.Done():
+				return nil, ctx.Err()
+			default:
+			}
+		}
+
+		// GPU capture on step 2 if requested
+		if cfg.CapturePath != "" && i == 1 {
+			mlx.MetalStartCapture(cfg.CapturePath)
+		}
+
+		timestep := timesteps[i]
+
+		// Prepare input - concatenate noise patches with reference tokens if present
+		imgInput := patches
+		if refTokens != nil {
+			imgInput = mlx.Concatenate([]*mlx.Array{patches, refTokens.Tokens}, 1)
+		}
+
+		// Transformer forward pass
+		output := m.Transformer.Forward(imgInput, promptEmbeds, timestep, rope)
+
+		// If we concatenated reference tokens, slice to only get noise portion
+		if refTokens != nil {
+			output = mlx.Slice(output, []int32{0, 0, 0}, []int32{1, noiseSeqLen, output.Shape()[2]})
+		}
+
+		// Scheduler step (keep reference to old patches for the computation graph)
+		newPatches := scheduler.Step(output, patches, i)
+
+		if cfg.CapturePath != "" && i == 1 {
+			mlx.MetalStopCapture()
+		}
+
+		mlx.Eval(newPatches)
+		patches = newPatches
+
+		elapsed := time.Since(stepStart).Seconds()
+		peakGB := float64(mlx.MetalGetPeakMemory()) / (1024 * 1024 * 1024)
+		if i == 0 {
+			fmt.Printf("    step %d: %.2fs (JIT warmup), peak %.1f GB\n", i+1, elapsed, peakGB)
+		} else {
+			fmt.Printf("    step %d: %.2fs, peak %.1f GB\n", i+1, elapsed, peakGB)
+		}
+		stepStart = time.Now()
+		if cfg.Progress != nil {
+			cfg.Progress(i+1, cfg.Steps)
+		}
+	}
+
+	loopTime := time.Since(loopStart).Seconds()
+	peakMem := float64(mlx.MetalGetPeakMemory()) / (1024 * 1024 * 1024)
+	fmt.Printf("  Denoised %d steps in %.2fs (%.2fs/step), peak %.1f GB\n",
+		cfg.Steps, loopTime, loopTime/float64(cfg.Steps), peakMem)
+
+	// Free timesteps now that denoising is done
+	for _, ts := range timesteps {
+		ts.Free()
+	}
+
+	// VAE decode with tiling for larger images
+	fmt.Print("  Decoding VAE... ")
+	vaeStart := time.Now()
+	// Enable tiling for images > 512x512 (latent > 64x64)
+	// VAE attention is O(n²) on latent pixels, tiling reduces memory significantly
+	if patchH*2 > 64 || patchW*2 > 64 {
+		m.VAE.Tiling = DefaultTilingConfig()
+	}
+	decoded := m.VAE.Decode(patches, patchH, patchW)
+	mlx.Eval(decoded)
+
+	// Free patches now that decode is done
+	patches.Free()
+
+	fmt.Printf("✓ (%.2fs, peak %.1f GB)\n", time.Since(vaeStart).Seconds(),
+		float64(mlx.MetalGetPeakMemory())/(1024*1024*1024))
+
+	return decoded, nil
+}
+
+// packLatents converts [B, C, H, W] to [B, H*W, C] (matches diffusers _pack_latents)
+func packLatents(x *mlx.Array) *mlx.Array {
+	shape := x.Shape()
+	B := shape[0]
+	C := shape[1]
+	H := shape[2]
+	W := shape[3]
+	// [B, C, H, W] -> [B, C, H*W] -> [B, H*W, C]
+	x = mlx.Reshape(x, B, C, H*W)
+	return mlx.Transpose(x, 0, 2, 1)
+}
+
+// LoadPersistent loads the model and keeps it in memory for repeated use.
+func LoadPersistent(modelName string) (*Model, error) {
+	m := &Model{}
+	if err := m.Load(modelName); err != nil {
+		return nil, err
+	}
+	return m, nil
+}
+
+// ImageRefScale is the time coordinate offset between reference images (matches diffusers scale=10)
+const ImageRefScale = 10
+
+// PrepareImage resizes and crops an image to be a multiple of 16, with optional pixel limit.
+// Returns the processed image and its dimensions.
+func PrepareImage(img image.Image, limitPixels int) (image.Image, int, int) {
+	bounds := img.Bounds()
+	w, h := bounds.Dx(), bounds.Dy()
+
+	// Cap pixels if needed (like diffusers cap_pixels)
+	if limitPixels > 0 && w*h > limitPixels {
+		scale := math.Sqrt(float64(limitPixels) / float64(w*h))
+		w = int(float64(w) * scale)
+		h = int(float64(h) * scale)
+	}
+
+	// Round down to multiple of 16
+	w = (w / 16) * 16
+	h = (h / 16) * 16
+
+	if w < 16 {
+		w = 16
+	}
+	if h < 16 {
+		h = 16
+	}
+
+	// Resize using high-quality bicubic interpolation (matches diffusers' default lanczos)
+	resized := image.NewRGBA(image.Rect(0, 0, w, h))
+	draw.CatmullRom.Scale(resized, resized.Bounds(), img, img.Bounds(), draw.Over, nil)
+
+	return resized, w, h
+}
+
+// ImageToTensor converts an image to a tensor in [-1, 1] range with shape [1, C, H, W].
+func ImageToTensor(img image.Image) *mlx.Array {
+	bounds := img.Bounds()
+	w, h := bounds.Dx(), bounds.Dy()
+
+	// Convert to float32 array in NCHW format [1, 3, H, W] with values in [-1, 1]
+	data := make([]float32, 3*h*w)
+
+	for y := 0; y < h; y++ {
+		for x := 0; x < w; x++ {
+			r, g, b, _ := img.At(x+bounds.Min.X, y+bounds.Min.Y).RGBA()
+			// RGBA returns 16-bit values, convert to [-1, 1]
+			data[0*h*w+y*w+x] = float32(r>>8)/127.5 - 1.0
+			data[1*h*w+y*w+x] = float32(g>>8)/127.5 - 1.0
+			data[2*h*w+y*w+x] = float32(b>>8)/127.5 - 1.0
+		}
+	}
+
+	arr := mlx.NewArrayFloat32(data, []int32{1, 3, int32(h), int32(w)})
+	return arr
+}
+
+// ImageCondTokens holds encoded reference image tokens.
+type ImageCondTokens struct {
+	Tokens *mlx.Array // [1, total_tokens, C] - concatenated reference tokens
+}
+
+// EncodeImageRefs encodes reference images using the VAE.
+func (m *Model) EncodeImageRefs(images []image.Image) (*ImageCondTokens, error) {
+	if len(images) == 0 {
+		return nil, nil
+	}
+
+	// Limit reference images to reduce attention memory
+	limitPixels := MaxRefPixels
+	if len(images) > 1 {
+		limitPixels = MaxRefPixels / 2
+	}
+
+	var allTokens []*mlx.Array
+
+	for _, img := range images {
+		// Prepare image (resize, crop to multiple of 16)
+		prepared, prepW, prepH := PrepareImage(img, limitPixels)
+		fmt.Printf("    Encoding %dx%d image... ", prepW, prepH)
+
+		// Convert to tensor [-1, 1]
+		tensor := ImageToTensor(prepared)
+
+		// Encode with VAE - returns [1, L, 128]
+		encoded := m.VAE.EncodeImage(tensor)
+		squeezed := mlx.Squeeze(encoded, 0) // [L, C]
+
+		// Defer eval - will be done with other setup arrays
+		allTokens = append(allTokens, squeezed)
+		fmt.Println("✓")
+	}
+
+	// For single image, just add batch dimension directly
+	// For multiple images, concatenate first
+	var tokens *mlx.Array
+	if len(allTokens) == 1 {
+		tokens = mlx.ExpandDims(allTokens[0], 0) // [1, L, C]
+	} else {
+		tokens = mlx.Concatenate(allTokens, 0) // [total_L, C]
+		tokens = mlx.ExpandDims(tokens, 0)     // [1, total_L, C]
+	}
+
+	return &ImageCondTokens{Tokens: tokens}, nil
+}

x/imagegen/models/flux2/rope.go

@@ -0,0 +1,224 @@
+//go:build mlx
+
+package flux2
+
+import (
+	"math"
+
+	"github.com/ollama/ollama/x/imagegen/mlx"
+)
+
+// RoPEConfig holds 4D RoPE configuration for Flux2
+type RoPEConfig struct {
+	Theta    int32   // 2000 for Klein
+	AxesDims []int32 // [32, 32, 32, 32] - dimensions for T, H, W, L axes
+}
+
+// RoPECache holds precomputed RoPE cos/sin values
+type RoPECache struct {
+	Cos      *mlx.Array // [1, TotalSeqLen, 1, head_dim/2]
+	Sin      *mlx.Array // [1, TotalSeqLen, 1, head_dim/2]
+	TextLen  int32      // Length of text sequence
+	ImageLen int32      // Length of image sequence
+}
+
+// PrepareTextIDs creates position IDs for text tokens.
+// Text tokens use: T=0, H=0, W=0, L=0..seqLen-1
+// Returns: [seqLen, 4]
+func PrepareTextIDs(seqLen int32) *mlx.Array {
+	ids := make([]float32, seqLen*4)
+	for i := int32(0); i < seqLen; i++ {
+		idx := i * 4
+		ids[idx+0] = 0             // T = 0
+		ids[idx+1] = 0             // H = 0
+		ids[idx+2] = 0             // W = 0
+		ids[idx+3] = float32(i)    // L = sequence position
+	}
+	return mlx.NewArray(ids, []int32{seqLen, 4})
+}
+
+// PrepareLatentIDs creates position IDs for image latent tokens.
+// Latent tokens use: T=0, H=0..height-1, W=0..width-1, L=0
+// The latents are in row-major order (H then W).
+// Returns: [height*width, 4]
+func PrepareLatentIDs(height, width int32) *mlx.Array {
+	seqLen := height * width
+	ids := make([]float32, seqLen*4)
+	idx := 0
+	for h := int32(0); h < height; h++ {
+		for w := int32(0); w < width; w++ {
+			ids[idx*4+0] = 0           // T = 0
+			ids[idx*4+1] = float32(h)  // H = row
+			ids[idx*4+2] = float32(w)  // W = column
+			ids[idx*4+3] = 0           // L = 0
+			idx++
+		}
+	}
+	return mlx.NewArray(ids, []int32{seqLen, 4})
+}
+
+// PrepareImageIDs creates position IDs for reference image tokens (used in editing).
+// Reference images use: T=scale*(i+1), H=0..h-1, W=0..w-1, L=0
+// where i is the image index (0, 1, 2, ...) and scale separates images in T dimension.
+// Returns: [total_tokens, 4]
+func PrepareImageIDs(imageHeights, imageWidths []int32, scale int32) *mlx.Array {
+	// Calculate total tokens
+	totalTokens := int32(0)
+	for i := range imageHeights {
+		totalTokens += imageHeights[i] * imageWidths[i]
+	}
+
+	ids := make([]float32, totalTokens*4)
+	idx := int32(0)
+	for imgIdx, h := range imageHeights {
+		w := imageWidths[imgIdx]
+		tValue := float32(scale * int32(imgIdx+1))
+		for hi := int32(0); hi < h; hi++ {
+			for wi := int32(0); wi < w; wi++ {
+				ids[idx*4+0] = tValue       // T = scale * (imgIdx + 1)
+				ids[idx*4+1] = float32(hi)  // H = row
+				ids[idx*4+2] = float32(wi)  // W = column
+				ids[idx*4+3] = 0            // L = 0
+				idx++
+			}
+		}
+	}
+	return mlx.NewArray(ids, []int32{totalTokens, 4})
+}
+
+// ComputeRoPE computes cos and sin for 4D rotary position embeddings.
+// ids: [L, 4] with (T, H, W, L) coordinates
+// axesDims: [32, 32, 32, 32] - each axis has this many dimensions (total = head_dim = 128)
+// theta: base frequency (2000 for Klein)
+// Returns: cos, sin each [1, L, 1, head_dim] with repeat_interleave applied
+func ComputeRoPE(ids *mlx.Array, axesDims []int32, theta int32) (*mlx.Array, *mlx.Array) {
+	shape := ids.Shape()
+	seqLen := shape[0]
+
+	// Compute total head dim (sum of all axes dims)
+	headDim := int32(0)
+	for _, d := range axesDims {
+		headDim += d
+	}
+
+	// Extract each coordinate dimension
+	// ids[:, 0] = T, ids[:, 1] = H, ids[:, 2] = W, ids[:, 3] = L
+	posT := mlx.Slice(ids, []int32{0, 0}, []int32{seqLen, 1}) // [L, 1]
+	posH := mlx.Slice(ids, []int32{0, 1}, []int32{seqLen, 2}) // [L, 1]
+	posW := mlx.Slice(ids, []int32{0, 2}, []int32{seqLen, 3}) // [L, 1]
+	posL := mlx.Slice(ids, []int32{0, 3}, []int32{seqLen, 4}) // [L, 1]
+
+	// Compute frequencies for each axis
+	logTheta := float32(math.Log(float64(theta)))
+	cosArrs := make([]*mlx.Array, 4)
+	sinArrs := make([]*mlx.Array, 4)
+	positions := []*mlx.Array{posT, posH, posW, posL}
+
+	for i, axisDim := range axesDims {
+		half := axisDim / 2
+
+		// Create frequency array for this axis: theta^(-2j/dim) for j=0..half-1
+		// This matches diffusers: 1.0 / (theta ** (torch.arange(0, dim, 2) / dim))
+		freqs := make([]float32, half)
+		for j := int32(0); j < half; j++ {
+			freqs[j] = float32(math.Exp(float64(-logTheta * float32(2*j) / float32(axisDim))))
+		}
+		freqArr := mlx.NewArray(freqs, []int32{1, half})
+
+		// Compute pos * freq -> [L, half]
+		posExpanded := positions[i] // [L, 1]
+		args := mlx.Mul(posExpanded, freqArr) // [L, half]
+
+		// Compute cos and sin for this axis
+		cosAxis := mlx.Cos(args) // [L, half]
+		sinAxis := mlx.Sin(args) // [L, half]
+
+		// repeat_interleave(2): [c0, c1, ...] -> [c0, c0, c1, c1, ...]
+		// Reshape [L, half] -> [L, half, 1], tile to [L, half, 2], reshape to [L, axisDim]
+		cosAxis = mlx.ExpandDims(cosAxis, 2)                        // [L, half, 1]
+		cosAxis = mlx.Tile(cosAxis, []int32{1, 1, 2})               // [L, half, 2]
+		cosAxis = mlx.Reshape(cosAxis, seqLen, axisDim)             // [L, axisDim]
+
+		sinAxis = mlx.ExpandDims(sinAxis, 2)
+		sinAxis = mlx.Tile(sinAxis, []int32{1, 1, 2})
+		sinAxis = mlx.Reshape(sinAxis, seqLen, axisDim)
+
+		cosArrs[i] = cosAxis
+		sinArrs[i] = sinAxis
+	}
+
+	// Concatenate all axes: [L, headDim]
+	cos := mlx.Concatenate(cosArrs, 1)
+	sin := mlx.Concatenate(sinArrs, 1)
+
+	// Reshape to [1, L, 1, headDim] for broadcasting with attention
+	cos = mlx.Reshape(cos, 1, seqLen, 1, headDim)
+	sin = mlx.Reshape(sin, 1, seqLen, 1, headDim)
+
+	return cos, sin
+}
+
+// ApplyRoPE4D applies 4D rotary position embeddings to queries and keys.
+// x: [B, L, nheads, head_dim]
+// cos, sin: [1, L, 1, head_dim] (with repeat_interleave applied)
+// Returns: x with RoPE applied
+// Matches diffusers apply_rotary_emb with use_real=True, use_real_unbind_dim=-1
+func ApplyRoPE4D(x *mlx.Array, cos, sin *mlx.Array) *mlx.Array {
+	shape := x.Shape()
+	B := shape[0]
+	L := shape[1]
+	nheads := shape[2]
+	headDim := shape[3]
+	half := headDim / 2
+
+	// Reshape x to [B, L, nheads, half, 2] and split into real/imag
+	xReshaped := mlx.Reshape(x, B, L, nheads, half, 2)
+
+	// Extract real (index 0) and imag (index 1) parts
+	xReal := mlx.Slice(xReshaped, []int32{0, 0, 0, 0, 0}, []int32{B, L, nheads, half, 1})
+	xImag := mlx.Slice(xReshaped, []int32{0, 0, 0, 0, 1}, []int32{B, L, nheads, half, 2})
+	xReal = mlx.Squeeze(xReal, 4) // [B, L, nheads, half]
+	xImag = mlx.Squeeze(xImag, 4) // [B, L, nheads, half]
+
+	// x_rotated = stack([-x_imag, x_real], dim=-1).flatten(-2)
+	// This creates [-x_imag[0], x_real[0], -x_imag[1], x_real[1], ...]
+	negXImag := mlx.Neg(xImag)
+	negXImag = mlx.ExpandDims(negXImag, 4) // [B, L, nheads, half, 1]
+	xReal = mlx.ExpandDims(xReal, 4)       // [B, L, nheads, half, 1]
+	xRotated := mlx.Concatenate([]*mlx.Array{negXImag, xReal}, 4) // [B, L, nheads, half, 2]
+	xRotated = mlx.Reshape(xRotated, B, L, nheads, headDim)       // [B, L, nheads, headDim]
+
+	// out = x * cos + x_rotated * sin
+	return mlx.Add(mlx.Mul(x, cos), mlx.Mul(xRotated, sin))
+}
+
+// PrepareRoPECache creates RoPE cache for text + noise, optionally with reference images.
+// textLen: number of text tokens
+// noiseH, noiseW: dimensions of the noise latent in patch tokens
+// axesDims: [32, 32, 32, 32]
+// theta: 2000
+// refHeights, refWidths: optional reference image dimensions (pass nil/empty for no images)
+// scale: time coordinate offset between reference images (e.g., 10)
+func PrepareRoPECache(textLen, noiseH, noiseW int32, axesDims []int32, theta int32, refHeights, refWidths []int32, scale int32) *RoPECache {
+	textIDs := PrepareTextIDs(textLen)
+	noiseIDs := PrepareLatentIDs(noiseH, noiseW)
+
+	var allIDs *mlx.Array
+	imageLen := noiseH * noiseW
+
+	if len(refHeights) > 0 {
+		refIDs := PrepareImageIDs(refHeights, refWidths, scale)
+		allIDs = mlx.Concatenate([]*mlx.Array{textIDs, noiseIDs, refIDs}, 0)
+		for i := range refHeights {
+			imageLen += refHeights[i] * refWidths[i]
+		}
+	} else {
+		allIDs = mlx.Concatenate([]*mlx.Array{textIDs, noiseIDs}, 0)
+	}
+
+	cos, sin := ComputeRoPE(allIDs, axesDims, theta)
+	cos = mlx.ToBFloat16(cos)
+	sin = mlx.ToBFloat16(sin)
+
+	return &RoPECache{Cos: cos, Sin: sin, TextLen: textLen, ImageLen: imageLen}
+}

x/imagegen/models/flux2/scheduler.go

@@ -0,0 +1,149 @@
+//go:build mlx
+
+package flux2
+
+import (
+	"math"
+
+	"github.com/ollama/ollama/x/imagegen/mlx"
+)
+
+// SchedulerConfig holds Flow-Match scheduler configuration
+type SchedulerConfig struct {
+	NumTrainTimesteps  int32   `json:"num_train_timesteps"`  // 1000
+	Shift              float32 `json:"shift"`                // 3.0 for Klein
+	UseDynamicShifting bool    `json:"use_dynamic_shifting"` // true
+	TimeShiftType      string  `json:"time_shift_type"`      // "exponential" or "linear"
+}
+
+// DefaultSchedulerConfig returns default config for Klein
+func DefaultSchedulerConfig() *SchedulerConfig {
+	return &SchedulerConfig{
+		NumTrainTimesteps:  1000,
+		Shift:              3.0, // Klein uses 3.0
+		UseDynamicShifting: true,
+		TimeShiftType:      "exponential",
+	}
+}
+
+// FlowMatchScheduler implements the Flow-Match Euler discrete scheduler
+type FlowMatchScheduler struct {
+	Config    *SchedulerConfig
+	Timesteps []float32 // Discretized timesteps (t from 1 to 0)
+	Sigmas    []float32 // Noise levels at each timestep
+	NumSteps  int       // Number of inference steps
+}
+
+// NewFlowMatchScheduler creates a new scheduler
+func NewFlowMatchScheduler(cfg *SchedulerConfig) *FlowMatchScheduler {
+	return &FlowMatchScheduler{
+		Config: cfg,
+	}
+}
+
+// SetTimesteps sets up the scheduler for the given number of inference steps
+func (s *FlowMatchScheduler) SetTimesteps(numSteps int) {
+	s.SetTimestepsWithMu(numSteps, 0)
+}
+
+// SetTimestepsWithMu sets up scheduler matching diffusers set_timesteps(sigmas=..., mu=...)
+func (s *FlowMatchScheduler) SetTimestepsWithMu(numSteps int, mu float32) {
+	s.NumSteps = numSteps
+
+	// diffusers: sigmas = linspace(1, 1/num_steps, num_steps)
+	// Then applies time shift, appends 0.0 at end
+	s.Sigmas = make([]float32, numSteps+1)
+
+	for i := 0; i < numSteps; i++ {
+		// linspace(1, 1/num_steps, num_steps)
+		var sigma float32
+		if numSteps == 1 {
+			sigma = 1.0
+		} else {
+			sigma = 1.0 - float32(i)/float32(numSteps-1)*(1.0-1.0/float32(numSteps))
+		}
+
+		// Apply time shift if using dynamic shifting
+		if s.Config.UseDynamicShifting && mu != 0 {
+			sigma = s.timeShift(mu, sigma)
+		} else {
+			// If not dynamic shifting, apply fixed shift scaling like diffusers
+			shift := s.Config.Shift
+			sigma = shift * sigma / (1 + (shift-1)*sigma)
+		}
+		s.Sigmas[i] = sigma
+	}
+	// Append terminal zero
+	s.Sigmas[numSteps] = 0.0
+
+	// Timesteps scaled to training range (matches diffusers: timesteps = sigmas * num_train_timesteps)
+	s.Timesteps = make([]float32, numSteps+1)
+	for i, v := range s.Sigmas {
+		s.Timesteps[i] = v * float32(s.Config.NumTrainTimesteps)
+	}
+}
+
+// timeShift applies the dynamic time shift
+func (s *FlowMatchScheduler) timeShift(mu float32, t float32) float32 {
+	if t <= 0 {
+		return 0
+	}
+	if s.Config.TimeShiftType == "linear" {
+		return mu / (mu + (1.0/t-1.0))
+	}
+	// Default: exponential
+	expMu := float32(math.Exp(float64(mu)))
+	return expMu / (expMu + (1.0/t - 1.0))
+}
+
+// Step performs one denoising step
+func (s *FlowMatchScheduler) Step(modelOutput, sample *mlx.Array, timestepIdx int) *mlx.Array {
+	sigma := s.Sigmas[timestepIdx]
+	sigmaNext := s.Sigmas[timestepIdx+1]
+
+	// Euler step: x_{t-dt} = x_t + (sigma_next - sigma) * v_t
+	dt := sigmaNext - sigma
+
+	// Upcast to float32 for precision (matches diffusers)
+	sampleF32 := mlx.AsType(sample, mlx.DtypeFloat32)
+	outputF32 := mlx.AsType(modelOutput, mlx.DtypeFloat32)
+
+	scaledOutput := mlx.MulScalar(outputF32, dt)
+	result := mlx.Add(sampleF32, scaledOutput)
+
+	// Cast back to bfloat16
+	return mlx.ToBFloat16(result)
+}
+
+// GetTimestep returns the timestep value at the given index
+func (s *FlowMatchScheduler) GetTimestep(idx int) float32 {
+	if idx < len(s.Timesteps) {
+		return s.Timesteps[idx]
+	}
+	return 0.0
+}
+
+// InitNoise creates initial noise for sampling
+func (s *FlowMatchScheduler) InitNoise(shape []int32, seed int64) *mlx.Array {
+	return mlx.RandomNormalWithDtype(shape, uint64(seed), mlx.DtypeBFloat16)
+}
+
+// CalculateShift computes the mu shift value for dynamic scheduling
+// Matches diffusers compute_empirical_mu function
+func CalculateShift(imgSeqLen int32, numSteps int) float32 {
+	a1, b1 := float32(8.73809524e-05), float32(1.89833333)
+	a2, b2 := float32(0.00016927), float32(0.45666666)
+
+	seqLen := float32(imgSeqLen)
+
+	if imgSeqLen > 4300 {
+		return a2*seqLen + b2
+	}
+
+	m200 := a2*seqLen + b2
+	m10 := a1*seqLen + b1
+
+	a := (m200 - m10) / 190.0
+	b := m200 - 200.0*a
+	return a*float32(numSteps) + b
+}

x/imagegen/models/flux2/transformer.go

@@ -0,0 +1,562 @@
+//go:build mlx
+
+package flux2
+
+import (
+	"fmt"
+	"math"
+
+	"github.com/ollama/ollama/x/imagegen"
+	"github.com/ollama/ollama/x/imagegen/mlx"
+	"github.com/ollama/ollama/x/imagegen/nn"
+	"github.com/ollama/ollama/x/imagegen/safetensors"
+)
+
+// TransformerConfig holds Flux2 transformer configuration
+type TransformerConfig struct {
+	AttentionHeadDim         int32   `json:"attention_head_dim"`          // 128
+	AxesDimsRoPE             []int32 `json:"axes_dims_rope"`              // [32, 32, 32, 32]
+	Eps                      float32 `json:"eps"`                         // 1e-6
+	GuidanceEmbeds           bool    `json:"guidance_embeds"`             // false for Klein
+	InChannels               int32   `json:"in_channels"`                 // 128
+	JointAttentionDim        int32   `json:"joint_attention_dim"`         // 7680
+	MLPRatio                 float32 `json:"mlp_ratio"`                   // 3.0
+	NumAttentionHeads        int32   `json:"num_attention_heads"`         // 24
+	NumLayers                int32   `json:"num_layers"`                  // 5
+	NumSingleLayers          int32   `json:"num_single_layers"`           // 20
+	PatchSize                int32   `json:"patch_size"`                  // 1
+	RopeTheta                int32   `json:"rope_theta"`                  // 2000
+	TimestepGuidanceChannels int32   `json:"timestep_guidance_channels"`  // 256
+}
+
+// Computed dimensions
+func (c *TransformerConfig) InnerDim() int32 {
+	return c.NumAttentionHeads * c.AttentionHeadDim // 24 * 128 = 3072
+}
+
+func (c *TransformerConfig) MLPHiddenDim() int32 {
+	return int32(float32(c.InnerDim()) * c.MLPRatio) // 3072 * 3.0 = 9216
+}
+
+// TimestepEmbedder creates timestep embeddings
+// Weight names: time_guidance_embed.timestep_embedder.linear_1.weight, linear_2.weight
+type TimestepEmbedder struct {
+	Linear1  nn.LinearLayer `weight:"linear_1"`
+	Linear2  nn.LinearLayer `weight:"linear_2"`
+	EmbedDim int32          // 256
+}
+
+// Forward creates sinusoidal embeddings and projects them
+func (t *TimestepEmbedder) Forward(timesteps *mlx.Array) *mlx.Array {
+	half := t.EmbedDim / 2
+	freqs := make([]float32, half)
+	for i := int32(0); i < half; i++ {
+		freqs[i] = float32(math.Exp(-math.Log(10000.0) * float64(i) / float64(half)))
+	}
+	freqsArr := mlx.NewArray(freqs, []int32{1, half})
+
+	// timesteps: [B] -> [B, 1]
+	tExpanded := mlx.ExpandDims(timesteps, 1)
+	// args: [B, half]
+	args := mlx.Mul(tExpanded, freqsArr)
+
+	// [cos(args), sin(args)] -> [B, embed_dim]
+	sinEmbed := mlx.Concatenate([]*mlx.Array{mlx.Cos(args), mlx.Sin(args)}, 1)
+
+	// MLP: linear_1 -> silu -> linear_2
+	h := t.Linear1.Forward(sinEmbed)
+	h = mlx.SiLU(h)
+	return t.Linear2.Forward(h)
+}
+
+// TimeGuidanceEmbed wraps the timestep embedder
+// Weight names: time_guidance_embed.timestep_embedder.*
+type TimeGuidanceEmbed struct {
+	TimestepEmbedder *TimestepEmbedder `weight:"timestep_embedder"`
+}
+
+// Forward computes timestep embeddings
+func (t *TimeGuidanceEmbed) Forward(timesteps *mlx.Array) *mlx.Array {
+	return t.TimestepEmbedder.Forward(timesteps)
+}
+
+// Modulation computes adaptive modulation parameters
+// Weight names: double_stream_modulation_img.linear.weight, etc.
+type Modulation struct {
+	Linear nn.LinearLayer `weight:"linear"`
+}
+
+// Forward computes modulation parameters
+func (m *Modulation) Forward(temb *mlx.Array) *mlx.Array {
+	h := mlx.SiLU(temb)
+	return m.Linear.Forward(h)
+}
+
+// TransformerBlockAttn implements dual-stream attention
+// Weight names: transformer_blocks.N.attn.*
+type TransformerBlockAttn struct {
+	// Image stream (separate Q, K, V projections)
+	ToQ nn.LinearLayer `weight:"to_q"`
+	ToK nn.LinearLayer `weight:"to_k"`
+	ToV nn.LinearLayer `weight:"to_v"`
+	// Note: to_out has .0 suffix in weights, handled specially
+	ToOut0 nn.LinearLayer `weight:"to_out.0"`
+
+	// Text stream (add_ projections)
+	AddQProj nn.LinearLayer `weight:"add_q_proj"`
+	AddKProj nn.LinearLayer `weight:"add_k_proj"`
+	AddVProj nn.LinearLayer `weight:"add_v_proj"`
+	ToAddOut nn.LinearLayer `weight:"to_add_out"`
+
+	// QK norms for image stream
+	NormQ *mlx.Array `weight:"norm_q.weight"`
+	NormK *mlx.Array `weight:"norm_k.weight"`
+
+	// QK norms for text stream (added)
+	NormAddedQ *mlx.Array `weight:"norm_added_q.weight"`
+	NormAddedK *mlx.Array `weight:"norm_added_k.weight"`
+}
+
+// FeedForward implements SwiGLU MLP
+// Weight names: transformer_blocks.N.ff.linear_in.weight, linear_out.weight
+type FeedForward struct {
+	LinearIn  nn.LinearLayer `weight:"linear_in"`
+	LinearOut nn.LinearLayer `weight:"linear_out"`
+}
+
+// Forward applies SwiGLU MLP
+func (ff *FeedForward) Forward(x *mlx.Array) *mlx.Array {
+	// LinearIn outputs 2x hidden dim for SwiGLU
+	h := ff.LinearIn.Forward(x)
+	shape := h.Shape()
+	half := shape[len(shape)-1] / 2
+
+	// Split into gate and up
+	gate := mlx.Slice(h, []int32{0, 0, 0}, []int32{shape[0], shape[1], half})
+	up := mlx.Slice(h, []int32{0, 0, half}, []int32{shape[0], shape[1], shape[2]})
+
+	// SwiGLU: silu(gate) * up
+	h = mlx.Mul(mlx.SiLU(gate), up)
+	return ff.LinearOut.Forward(h)
+}
+
+// TransformerBlock implements a dual-stream transformer block
+// Weight names: transformer_blocks.N.*
+type TransformerBlock struct {
+	Attn      *TransformerBlockAttn `weight:"attn"`
+	FF        *FeedForward          `weight:"ff"`
+	FFContext *FeedForward          `weight:"ff_context"`
+
+	// Config (set after loading)
+	NHeads  int32
+	HeadDim int32
+	Scale   float32
+}
+
+// Forward applies the dual-stream block
+// imgHidden: [B, imgLen, dim]
+// txtHidden: [B, txtLen, dim]
+// imgMod, txtMod: modulation params [B, 6*dim] each
+// cos, sin: RoPE values
+func (block *TransformerBlock) Forward(imgHidden, txtHidden *mlx.Array, imgMod, txtMod *mlx.Array, cos, sin *mlx.Array) (*mlx.Array, *mlx.Array) {
+	imgShape := imgHidden.Shape()
+	B := imgShape[0]
+	imgLen := imgShape[1]
+	dim := imgShape[2]
+	txtLen := txtHidden.Shape()[1]
+
+	// Parse modulation: 6 params each (shift1, scale1, gate1, shift2, scale2, gate2)
+	imgShift1, imgScale1, imgGate1 := parseModulation3(imgMod, dim, 0)
+	imgShift2, imgScale2, imgGate2 := parseModulation3(imgMod, dim, 3)
+	txtShift1, txtScale1, txtGate1 := parseModulation3(txtMod, dim, 0)
+	txtShift2, txtScale2, txtGate2 := parseModulation3(txtMod, dim, 3)
+
+	// === Attention branch ===
+	// Modulate inputs
+	imgNorm := modulateLayerNorm(imgHidden, imgShift1, imgScale1)
+	txtNorm := modulateLayerNorm(txtHidden, txtShift1, txtScale1)
+
+	// Compute Q, K, V for image stream (separate projections)
+	imgQ := block.Attn.ToQ.Forward(imgNorm)
+	imgK := block.Attn.ToK.Forward(imgNorm)
+	imgV := block.Attn.ToV.Forward(imgNorm)
+
+	// Compute Q, K, V for text stream (add_ projections)
+	txtQ := block.Attn.AddQProj.Forward(txtNorm)
+	txtK := block.Attn.AddKProj.Forward(txtNorm)
+	txtV := block.Attn.AddVProj.Forward(txtNorm)
+
+	// Reshape for attention: [B, L, dim] -> [B, L, nheads, headDim]
+	imgQ = mlx.Reshape(imgQ, B, imgLen, block.NHeads, block.HeadDim)
+	imgK = mlx.Reshape(imgK, B, imgLen, block.NHeads, block.HeadDim)
+	imgV = mlx.Reshape(imgV, B, imgLen, block.NHeads, block.HeadDim)
+	txtQ = mlx.Reshape(txtQ, B, txtLen, block.NHeads, block.HeadDim)
+	txtK = mlx.Reshape(txtK, B, txtLen, block.NHeads, block.HeadDim)
+	txtV = mlx.Reshape(txtV, B, txtLen, block.NHeads, block.HeadDim)
+
+	// Apply QK norm (RMSNorm with learned scale)
+	imgQ = applyQKNorm(imgQ, block.Attn.NormQ)
+	imgK = applyQKNorm(imgK, block.Attn.NormK)
+	txtQ = applyQKNorm(txtQ, block.Attn.NormAddedQ)
+	txtK = applyQKNorm(txtK, block.Attn.NormAddedK)
+
+	// Concatenate for joint attention: text first, then image
+	q := mlx.Concatenate([]*mlx.Array{txtQ, imgQ}, 1)
+	k := mlx.Concatenate([]*mlx.Array{txtK, imgK}, 1)
+	v := mlx.Concatenate([]*mlx.Array{txtV, imgV}, 1)
+
+	// Apply RoPE
+	q = ApplyRoPE4D(q, cos, sin)
+	k = ApplyRoPE4D(k, cos, sin)
+
+	// Transpose for SDPA: [B, nheads, L, headDim]
+	q = mlx.Transpose(q, 0, 2, 1, 3)
+	k = mlx.Transpose(k, 0, 2, 1, 3)
+	v = mlx.Transpose(v, 0, 2, 1, 3)
+
+	// Scaled dot-product attention
+	out := mlx.ScaledDotProductAttention(q, k, v, block.Scale, false)
+
+	// Transpose back: [B, L, nheads, headDim]
+	out = mlx.Transpose(out, 0, 2, 1, 3)
+
+	// Split back into txt and img
+	totalLen := txtLen + imgLen
+	txtOut := mlx.Slice(out, []int32{0, 0, 0, 0}, []int32{B, txtLen, block.NHeads, block.HeadDim})
+	imgOut := mlx.Slice(out, []int32{0, txtLen, 0, 0}, []int32{B, totalLen, block.NHeads, block.HeadDim})
+
+	// Reshape and project
+	txtOut = mlx.Reshape(txtOut, B, txtLen, dim)
+	imgOut = mlx.Reshape(imgOut, B, imgLen, dim)
+	txtOut = block.Attn.ToAddOut.Forward(txtOut)
+	imgOut = block.Attn.ToOut0.Forward(imgOut)
+
+	// Apply gates and residual
+	imgHidden = mlx.Add(imgHidden, mlx.Mul(imgGate1, imgOut))
+	txtHidden = mlx.Add(txtHidden, mlx.Mul(txtGate1, txtOut))
+
+	// === MLP branch ===
+	imgNorm = modulateLayerNorm(imgHidden, imgShift2, imgScale2)
+	txtNorm = modulateLayerNorm(txtHidden, txtShift2, txtScale2)
+
+	imgFFOut := block.FF.Forward(imgNorm)
+	txtFFOut := block.FFContext.Forward(txtNorm)
+
+	imgHidden = mlx.Add(imgHidden, mlx.Mul(imgGate2, imgFFOut))
+	txtHidden = mlx.Add(txtHidden, mlx.Mul(txtGate2, txtFFOut))
+
+	return imgHidden, txtHidden
+}
+
+// SingleTransformerBlockAttn implements attention for single-stream blocks
+// Weight names: single_transformer_blocks.N.attn.*
+type SingleTransformerBlockAttn struct {
+	ToQKVMlpProj nn.LinearLayer `weight:"to_qkv_mlp_proj"` // Fused QKV + MLP input
+	ToOut        nn.LinearLayer `weight:"to_out"`          // Fused attn_out + MLP out
+	NormQ        *mlx.Array     `weight:"norm_q.weight"`
+	NormK        *mlx.Array     `weight:"norm_k.weight"`
+}
+
+// SingleTransformerBlock implements a single-stream transformer block
+// Weight names: single_transformer_blocks.N.*
+type SingleTransformerBlock struct {
+	Attn *SingleTransformerBlockAttn `weight:"attn"`
+
+	// Config
+	NHeads    int32
+	HeadDim   int32
+	InnerDim  int32
+	MLPHidDim int32
+	Scale     float32
+}
+
+// Forward applies the single-stream block
+// x: [B, L, dim] concatenated text+image
+// mod: modulation [B, 3*dim]
+func (block *SingleTransformerBlock) Forward(x *mlx.Array, mod *mlx.Array, cos, sin *mlx.Array) *mlx.Array {
+	shape := x.Shape()
+	B := shape[0]
+	L := shape[1]
+	dim := shape[2]
+
+	// Parse modulation: (shift, scale, gate)
+	shift, scale, gate := parseModulation3(mod, dim, 0)
+
+	// Modulate input
+	h := modulateLayerNorm(x, shift, scale)
+
+	// Fused projection: QKV + MLP gate/up
+	// linear1 outputs: [q, k, v, mlp_gate, mlp_up] = [dim, dim, dim, mlpHid, mlpHid]
+	qkvMlp := block.Attn.ToQKVMlpProj.Forward(h)
+
+	// Split: first 3*dim is QKV, rest is MLP
+	qkvDim := 3 * block.InnerDim
+	qkv := mlx.Slice(qkvMlp, []int32{0, 0, 0}, []int32{B, L, qkvDim})
+	mlpIn := mlx.Slice(qkvMlp, []int32{0, 0, qkvDim}, []int32{B, L, qkvMlp.Shape()[2]})
+
+	// Split QKV
+	q, k, v := splitQKV(qkv, B, L, block.InnerDim)
+
+	// Reshape for attention
+	q = mlx.Reshape(q, B, L, block.NHeads, block.HeadDim)
+	k = mlx.Reshape(k, B, L, block.NHeads, block.HeadDim)
+	v = mlx.Reshape(v, B, L, block.NHeads, block.HeadDim)
+
+	// QK norm
+	q = applyQKNorm(q, block.Attn.NormQ)
+	k = applyQKNorm(k, block.Attn.NormK)
+
+	// Apply RoPE
+	q = ApplyRoPE4D(q, cos, sin)
+	k = ApplyRoPE4D(k, cos, sin)
+
+	// Transpose for SDPA
+	q = mlx.Transpose(q, 0, 2, 1, 3)
+	k = mlx.Transpose(k, 0, 2, 1, 3)
+	v = mlx.Transpose(v, 0, 2, 1, 3)
+
+	// SDPA
+	attnOut := mlx.ScaledDotProductAttention(q, k, v, block.Scale, false)
+
+	// Transpose back and reshape
+	attnOut = mlx.Transpose(attnOut, 0, 2, 1, 3)
+	attnOut = mlx.Reshape(attnOut, B, L, block.InnerDim)
+
+	// MLP: SwiGLU
+	mlpShape := mlpIn.Shape()
+	half := mlpShape[2] / 2
+	mlpGate := mlx.Slice(mlpIn, []int32{0, 0, 0}, []int32{B, L, half})
+	mlpUp := mlx.Slice(mlpIn, []int32{0, 0, half}, []int32{B, L, mlpShape[2]})
+	mlpOut := mlx.Mul(mlx.SiLU(mlpGate), mlpUp)
+
+	// Concatenate attention and MLP for fused output
+	combined := mlx.Concatenate([]*mlx.Array{attnOut, mlpOut}, 2)
+
+	// Output projection
+	out := block.Attn.ToOut.Forward(combined)
+
+	// Apply gate and residual
+	return mlx.Add(x, mlx.Mul(gate, out))
+}
+
+// NormOut implements the output normalization with modulation
+// Weight names: norm_out.linear.weight
+type NormOut struct {
+	Linear nn.LinearLayer `weight:"linear"`
+}
+
+// Forward computes final modulated output
+func (n *NormOut) Forward(x *mlx.Array, temb *mlx.Array) *mlx.Array {
+	shape := x.Shape()
+	B := shape[0]
+	dim := shape[2]
+
+	// Modulation: temb -> silu -> linear -> [shift, scale]
+	mod := mlx.SiLU(temb)
+	mod = n.Linear.Forward(mod)
+
+	// Split into scale and shift (diffusers order: scale first, shift second)
+	scale := mlx.Slice(mod, []int32{0, 0}, []int32{B, dim})
+	shift := mlx.Slice(mod, []int32{0, dim}, []int32{B, 2 * dim})
+	shift = mlx.ExpandDims(shift, 1)
+	scale = mlx.ExpandDims(scale, 1)
+
+	// Modulate with RMSNorm
+	return modulateLayerNorm(x, shift, scale)
+}
+
+// Flux2Transformer2DModel is the main Flux2 transformer
+// Weight names at top level: time_guidance_embed.*, double_stream_modulation_*.*, etc.
+type Flux2Transformer2DModel struct {
+	// Timestep embedding
+	TimeGuidanceEmbed *TimeGuidanceEmbed `weight:"time_guidance_embed"`
+
+	// Shared modulation
+	DoubleStreamModulationImg *Modulation `weight:"double_stream_modulation_img"`
+	DoubleStreamModulationTxt *Modulation `weight:"double_stream_modulation_txt"`
+	SingleStreamModulation    *Modulation `weight:"single_stream_modulation"`
+
+	// Embedders
+	XEmbedder       nn.LinearLayer `weight:"x_embedder"`
+	ContextEmbedder nn.LinearLayer `weight:"context_embedder"`
+
+	// Transformer blocks
+	TransformerBlocks       []*TransformerBlock       `weight:"transformer_blocks"`
+	SingleTransformerBlocks []*SingleTransformerBlock `weight:"single_transformer_blocks"`
+
+	// Output
+	NormOut *NormOut       `weight:"norm_out"`
+	ProjOut nn.LinearLayer `weight:"proj_out"`
+
+	*TransformerConfig
+}
+
+// Load loads the Flux2 transformer from ollama blob storage.
+func (m *Flux2Transformer2DModel) Load(manifest *imagegen.ModelManifest) error {
+	fmt.Print("  Loading transformer... ")
+
+	// Load config from blob
+	var cfg TransformerConfig
+	if err := manifest.ReadConfigJSON("transformer/config.json", &cfg); err != nil {
+		return fmt.Errorf("config: %w", err)
+	}
+	m.TransformerConfig = &cfg
+
+	// Initialize slices
+	m.TransformerBlocks = make([]*TransformerBlock, cfg.NumLayers)
+	m.SingleTransformerBlocks = make([]*SingleTransformerBlock, cfg.NumSingleLayers)
+
+	// Initialize TimeGuidanceEmbed with embed dim
+	m.TimeGuidanceEmbed = &TimeGuidanceEmbed{
+		TimestepEmbedder: &TimestepEmbedder{EmbedDim: cfg.TimestepGuidanceChannels},
+	}
+
+	// Load weights from tensor blobs
+	weights, err := imagegen.LoadWeightsFromManifest(manifest, "transformer")
+	if err != nil {
+		return fmt.Errorf("weights: %w", err)
+	}
+	if err := weights.Load(0); err != nil {
+		return fmt.Errorf("load weights: %w", err)
+	}
+	defer weights.ReleaseAll()
+
+	return m.loadWeights(weights)
+}
+
+// loadWeights loads weights from any WeightSource into the model
+func (m *Flux2Transformer2DModel) loadWeights(weights safetensors.WeightSource) error {
+	if err := safetensors.LoadModule(m, weights, ""); err != nil {
+		return fmt.Errorf("load module: %w", err)
+	}
+	m.initComputedFields()
+	fmt.Println("✓")
+	return nil
+}
+
+// initComputedFields initializes computed fields after loading weights
+func (m *Flux2Transformer2DModel) initComputedFields() {
+	cfg := m.TransformerConfig
+	innerDim := cfg.InnerDim()
+	scale := float32(1.0 / math.Sqrt(float64(cfg.AttentionHeadDim)))
+
+	// Initialize transformer blocks
+	for _, block := range m.TransformerBlocks {
+		block.NHeads = cfg.NumAttentionHeads
+		block.HeadDim = cfg.AttentionHeadDim
+		block.Scale = scale
+	}
+
+	// Initialize single transformer blocks
+	for _, block := range m.SingleTransformerBlocks {
+		block.NHeads = cfg.NumAttentionHeads
+		block.HeadDim = cfg.AttentionHeadDim
+		block.InnerDim = innerDim
+		block.MLPHidDim = cfg.MLPHiddenDim()
+		block.Scale = scale
+	}
+}
+
+// Forward runs the Flux2 transformer
+func (m *Flux2Transformer2DModel) Forward(patches, txtEmbeds *mlx.Array, timesteps *mlx.Array, rope *RoPECache) *mlx.Array {
+	patchShape := patches.Shape()
+	B := patchShape[0]
+	imgLen := patchShape[1]
+	txtLen := txtEmbeds.Shape()[1]
+
+	// Scale timestep to 0-1000 range (diffusers multiplies by 1000)
+	scaledTimesteps := mlx.MulScalar(timesteps, 1000.0)
+
+	// Compute timestep embedding
+	temb := m.TimeGuidanceEmbed.Forward(scaledTimesteps)
+
+	// Embed patches and text
+	imgHidden := m.XEmbedder.Forward(patches)
+	txtHidden := m.ContextEmbedder.Forward(txtEmbeds)
+
+	// Compute shared modulation
+	imgMod := m.DoubleStreamModulationImg.Forward(temb)
+	txtMod := m.DoubleStreamModulationTxt.Forward(temb)
+	singleMod := m.SingleStreamModulation.Forward(temb)
+
+	// Double (dual-stream) blocks
+	for _, block := range m.TransformerBlocks {
+		imgHidden, txtHidden = block.Forward(imgHidden, txtHidden, imgMod, txtMod, rope.Cos, rope.Sin)
+	}
+
+	// Concatenate for single-stream: text first, then image
+	hidden := mlx.Concatenate([]*mlx.Array{txtHidden, imgHidden}, 1)
+
+	// Single-stream blocks
+	for _, block := range m.SingleTransformerBlocks {
+		hidden = block.Forward(hidden, singleMod, rope.Cos, rope.Sin)
+	}
+
+	// Extract image portion
+	totalLen := txtLen + imgLen
+	imgOut := mlx.Slice(hidden, []int32{0, txtLen, 0}, []int32{B, totalLen, hidden.Shape()[2]})
+
+	// Final norm and projection
+	imgOut = m.NormOut.Forward(imgOut, temb)
+	return m.ProjOut.Forward(imgOut)
+}
+
+// Note: QK normalization uses mlx.RMSNorm (the fast version) directly
+// See applyQKNorm function below
+
+// compiledSwiGLU fuses: silu(gate) * up
+// Called 30x per step (10 in dual-stream + 20 in single-stream blocks)
+var compiledSwiGLU *mlx.CompiledFunc
+
+func getCompiledSwiGLU() *mlx.CompiledFunc {
+	if compiledSwiGLU == nil {
+		compiledSwiGLU = mlx.CompileShapeless(func(inputs []*mlx.Array) []*mlx.Array {
+			gate, up := inputs[0], inputs[1]
+			return []*mlx.Array{mlx.Mul(mlx.SiLU(gate), up)}
+		}, true)
+	}
+	return compiledSwiGLU
+}
+
+// Helper functions
+
+// parseModulation3 extracts 3 modulation params (shift, scale, gate) starting at offset
+func parseModulation3(mod *mlx.Array, dim int32, offset int32) (*mlx.Array, *mlx.Array, *mlx.Array) {
+	B := mod.Shape()[0]
+	start := offset * dim
+	shift := mlx.Slice(mod, []int32{0, start}, []int32{B, start + dim})
+	scale := mlx.Slice(mod, []int32{0, start + dim}, []int32{B, start + 2*dim})
+	gate := mlx.Slice(mod, []int32{0, start + 2*dim}, []int32{B, start + 3*dim})
+
+	// Expand for broadcasting [B, dim] -> [B, 1, dim]
+	shift = mlx.ExpandDims(shift, 1)
+	scale = mlx.ExpandDims(scale, 1)
+	gate = mlx.ExpandDims(gate, 1)
+
+	return shift, scale, gate
+}
+
+// modulateLayerNorm applies LayerNorm then shift/scale modulation
+// Diffusers uses LayerNorm(elementwise_affine=False) which centers the data
+func modulateLayerNorm(x *mlx.Array, shift, scale *mlx.Array) *mlx.Array {
+	// Fast LayerNorm without learnable params
+	x = mlx.LayerNorm(x, 1e-6)
+
+	// Modulate: x * (1 + scale) + shift
+	x = mlx.Mul(x, mlx.AddScalar(scale, 1.0))
+	return mlx.Add(x, shift)
+}
+
+// splitQKV splits a fused QKV tensor into Q, K, V
+func splitQKV(qkv *mlx.Array, B, L, dim int32) (*mlx.Array, *mlx.Array, *mlx.Array) {
+	q := mlx.Slice(qkv, []int32{0, 0, 0}, []int32{B, L, dim})
+	k := mlx.Slice(qkv, []int32{0, 0, dim}, []int32{B, L, 2 * dim})
+	v := mlx.Slice(qkv, []int32{0, 0, 2 * dim}, []int32{B, L, 3 * dim})
+	return q, k, v
+}
+
+// applyQKNorm applies RMSNorm with learned scale (no bias)
+// Uses the optimized mlx_fast_rms_norm
+func applyQKNorm(x *mlx.Array, scale *mlx.Array) *mlx.Array {
+	return mlx.RMSNorm(x, scale, 1e-6)
+}