undo readme

TeaCache:
- Timestep embedding similarity caching for diffusion models
- Polynomial rescaling with configurable thresholds
- Reduces transformer forward passes by ~30-50%

FP8 quantization:
- Support for FP8 quantized models (8-bit weights with scales)
- QuantizedMatmul on Metal, Dequantize on CUDA
- Client-side quantization via ollama create --quantize fp8

Other improvements:
- Fix Show API for image generation models
- Server properly returns model info (architecture, parameters, quantization)
- Memory allocation optimizations
- CLI improvements for image generation

.github/ISSUE_TEMPLATE/10_bug_report.yml

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

.github/workflows/release.yaml

@@ -68,6 +68,7 @@ jobs:
           name: bundles-darwin
           path: |
             dist/*.tgz
+            dist/*.tar.zst
             dist/*.zip
             dist/*.dmg
 
@@ -392,13 +393,13 @@ jobs:
           done
       - run: |
           for ARCHIVE in dist/${{ matrix.os }}-${{ matrix.arch }}/*.tar.in; do
-            tar c -C dist/${{ matrix.os }}-${{ matrix.arch }} -T $ARCHIVE --owner 0 --group 0 | pigz -9vc >$(basename ${ARCHIVE//.*/}.tgz);
+            tar c -C dist/${{ matrix.os }}-${{ matrix.arch }} -T $ARCHIVE --owner 0 --group 0 | zstd --ultra -22 -T0 >$(basename ${ARCHIVE//.*/}.tar.zst);
           done
       - uses: actions/upload-artifact@v4
         with:
           name: bundles-${{ matrix.os }}-${{ matrix.arch }}-${{ matrix.target }}
           path: |
-            *.tgz
+            *.tar.zst
 
   # Build each Docker variant (OS, arch, and flavor) separately. Using QEMU is unreliable and slower.
   docker-build-push:
@@ -531,7 +532,7 @@ jobs:
       - name: Upload release artifacts
         run: |
           pids=()
-          for payload in dist/*.txt dist/*.zip dist/*.tgz dist/*.exe dist/*.dmg ; do
+          for payload in dist/*.txt dist/*.zip dist/*.tgz dist/*.tar.zst dist/*.exe dist/*.dmg ; do
             echo "Uploading $payload"
             gh release upload ${GITHUB_REF_NAME} $payload --clobber &
             pids[$!]=$!

CMakeLists.txt

@@ -2,6 +2,22 @@ cmake_minimum_required(VERSION 3.21)
 
 project(Ollama C CXX)
 
+# Handle cross-compilation on macOS: when CMAKE_OSX_ARCHITECTURES is set to a
+# single architecture different from the host, override CMAKE_SYSTEM_PROCESSOR
+# to match. This is necessary because CMAKE_SYSTEM_PROCESSOR defaults to the
+# host architecture, but downstream projects (like MLX) use it to detect the
+# target architecture.
+if(CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_OSX_ARCHITECTURES MATCHES ";")
+    # Single architecture specified
+    if(CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" AND NOT CMAKE_SYSTEM_PROCESSOR STREQUAL "x86_64")
+        message(STATUS "Cross-compiling for x86_64: overriding CMAKE_SYSTEM_PROCESSOR from ${CMAKE_SYSTEM_PROCESSOR} to x86_64")
+        set(CMAKE_SYSTEM_PROCESSOR "x86_64")
+    elseif(CMAKE_OSX_ARCHITECTURES STREQUAL "arm64" AND NOT CMAKE_SYSTEM_PROCESSOR STREQUAL "arm64")
+        message(STATUS "Cross-compiling for arm64: overriding CMAKE_SYSTEM_PROCESSOR from ${CMAKE_SYSTEM_PROCESSOR} to arm64")
+        set(CMAKE_SYSTEM_PROCESSOR "arm64")
+    endif()
+endif()
+
 include(CheckLanguage)
 include(GNUInstallDirs)
 
@@ -12,7 +28,7 @@ set(BUILD_SHARED_LIBS ON)
 
 set(CMAKE_CXX_STANDARD 17)
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
-set(CMAKE_CXX_EXTENSIONS OFF)
+set(CMAKE_CXX_EXTENSIONS ON) # Recent versions of MLX Requires gnu++17 extensions to compile properly
 
 set(GGML_BUILD ON)
 set(GGML_SHARED ON)
@@ -147,14 +163,48 @@ if(CMAKE_HIP_COMPILER)
     endif()
 endif()
 
-find_package(Vulkan)
-if(Vulkan_FOUND)
-    add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-vulkan)
-    install(TARGETS ggml-vulkan
-        RUNTIME_DEPENDENCIES
-            PRE_INCLUDE_REGEXES vulkan
-            PRE_EXCLUDE_REGEXES ".*"
-        RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT Vulkan
-        LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT Vulkan
-    )
+if(NOT APPLE)
+    find_package(Vulkan)
+    if(Vulkan_FOUND)
+        add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-vulkan)
+        install(TARGETS ggml-vulkan
+            RUNTIME_DEPENDENCIES
+                PRE_INCLUDE_REGEXES vulkan
+                PRE_EXCLUDE_REGEXES ".*"
+            RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT Vulkan
+            LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT Vulkan
+        )
+    endif()
 endif()
+
+option(MLX_ENGINE "Enable MLX backend" OFF)
+
+if(MLX_ENGINE)
+    message(STATUS "Setting up MLX (this takes a while...)")
+    add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/x/ml/backend/mlx)
+
+    # Find CUDA toolkit if MLX is built with CUDA support
+    find_package(CUDAToolkit)
+
+    install(TARGETS mlx mlxc
+        RUNTIME_DEPENDENCIES
+            DIRECTORIES ${CUDAToolkit_BIN_DIR} ${CUDAToolkit_BIN_DIR}/x64 ${CUDAToolkit_LIBRARY_DIR}
+            PRE_INCLUDE_REGEXES cublas cublasLt cudart nvrtc cudnn nccl
+            PRE_EXCLUDE_REGEXES ".*"
+        RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT MLX
+        LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT MLX
+        FRAMEWORK DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT MLX
+    )
+
+    # Manually install cudart and cublas since they might not be picked up as direct dependencies
+    if(CUDAToolkit_FOUND)
+        file(GLOB CUDART_LIBS
+            "${CUDAToolkit_LIBRARY_DIR}/libcudart.so*"
+            "${CUDAToolkit_LIBRARY_DIR}/libcublas.so*")
+        if(CUDART_LIBS)
+            install(FILES ${CUDART_LIBS}
+                DESTINATION ${OLLAMA_INSTALL_DIR}
+                COMPONENT MLX)
+        endif()
+    endif()
+endif()

CMakePresets.json

@@ -41,7 +41,7 @@
       "inherits": [ "CUDA" ],
       "cacheVariables": {
         "CMAKE_CUDA_ARCHITECTURES": "75-virtual;80-virtual;86-virtual;87-virtual;89-virtual;90-virtual;90a-virtual;100-virtual;103-virtual;110-virtual;120-virtual;121-virtual",
-        "CMAKE_CUDA_FLAGS": "-t 2",
+        "CMAKE_CUDA_FLAGS": "-t 4",
         "OLLAMA_RUNNER_DIR": "cuda_v13"
       }
     },
@@ -83,6 +83,28 @@
       "cacheVariables": {
         "OLLAMA_RUNNER_DIR": "vulkan"
       }
+    },
+    {
+      "name": "MLX",
+      "inherits": [ "Default" ],
+      "cacheVariables": {
+        "MLX_ENGINE": "ON",
+        "OLLAMA_RUNNER_DIR": "mlx"
+      }
+    },
+    {
+      "name": "MLX CUDA 12",
+      "inherits": [ "MLX", "CUDA 12" ],
+      "cacheVariables": {
+        "OLLAMA_RUNNER_DIR": "mlx_cuda_v12"
+      }
+    },
+    {
+      "name": "MLX CUDA 13",
+      "inherits": [ "MLX", "CUDA 13" ],
+      "cacheVariables": {
+        "OLLAMA_RUNNER_DIR": "mlx_cuda_v13"
+      }
     }
   ],
   "buildPresets": [
@@ -140,6 +162,21 @@
       "name": "Vulkan",
       "targets": [ "ggml-vulkan" ],
       "configurePreset": "Vulkan"
+    },
+    {
+      "name": "MLX",
+      "targets": [ "mlx", "mlxc" ],
+      "configurePreset": "MLX"
+    },
+    {
+      "name": "MLX CUDA 12",
+      "targets": [ "mlx", "mlxc" ],
+      "configurePreset": "MLX CUDA 12"
+    },
+    {
+      "name": "MLX CUDA 13",
+      "targets": [ "mlx", "mlxc" ],
+      "configurePreset": "MLX CUDA 13"
     }
   ]
 }

Dockerfile

@@ -131,8 +131,36 @@ COPY ml/backend/ggml/ggml ml/backend/ggml/ggml
 RUN --mount=type=cache,target=/root/.ccache \
     cmake --preset 'Vulkan' \
         && cmake --build --parallel --preset 'Vulkan' \
-        && cmake --install build --component Vulkan --strip --parallel 8 
+        && cmake --install build --component Vulkan --strip --parallel 8
 
+FROM base AS mlx
+ARG CUDA13VERSION=13.0
+RUN dnf install -y cuda-toolkit-${CUDA13VERSION//./-} \
+    && dnf install -y openblas-devel lapack-devel \
+    && dnf install -y libcudnn9-cuda-13 libcudnn9-devel-cuda-13 \
+    && dnf install -y libnccl libnccl-devel
+ENV PATH=/usr/local/cuda-13/bin:$PATH
+ENV BLAS_INCLUDE_DIRS=/usr/include/openblas
+ENV LAPACK_INCLUDE_DIRS=/usr/include/openblas
+ENV CGO_LDFLAGS="-L/usr/local/cuda-13/lib64 -L/usr/local/cuda-13/targets/x86_64-linux/lib/stubs"
+ARG PARALLEL
+WORKDIR /go/src/github.com/ollama/ollama
+COPY CMakeLists.txt CMakePresets.json .
+COPY ml/backend/ggml/ggml ml/backend/ggml/ggml
+COPY x/ml/backend/mlx x/ml/backend/mlx
+COPY go.mod go.sum .
+RUN curl -fsSL https://golang.org/dl/go$(awk '/^go/ { print $2 }' go.mod).linux-$(case $(uname -m) in x86_64) echo amd64 ;; aarch64) echo arm64 ;; esac).tar.gz | tar xz -C /usr/local
+ENV PATH=/usr/local/go/bin:$PATH
+RUN go mod download
+RUN --mount=type=cache,target=/root/.ccache \
+    cmake --preset 'MLX CUDA 13' -DBLAS_INCLUDE_DIRS=/usr/include/openblas -DLAPACK_INCLUDE_DIRS=/usr/include/openblas \
+        && cmake --build --parallel ${PARALLEL} --preset 'MLX CUDA 13' \
+        && cmake --install build --component MLX --strip --parallel ${PARALLEL}
+COPY . .
+ARG GOFLAGS="'-ldflags=-w -s'"
+ENV CGO_ENABLED=1
+ARG CGO_CFLAGS
+ARG CGO_CXXFLAGS
 
 FROM base AS build
 WORKDIR /go/src/github.com/ollama/ollama
@@ -153,6 +181,7 @@ FROM --platform=linux/amd64 scratch AS amd64
 COPY --from=cuda-12 dist/lib/ollama /lib/ollama/
 COPY --from=cuda-13 dist/lib/ollama /lib/ollama/
 COPY --from=vulkan  dist/lib/ollama  /lib/ollama/
+COPY --from=mlx     /go/src/github.com/ollama/ollama/dist/lib/ollama /lib/ollama/
 
 FROM --platform=linux/arm64 scratch AS arm64
 # COPY --from=cuda-11 dist/lib/ollama/ /lib/ollama/
@@ -171,7 +200,7 @@ COPY --from=build /bin/ollama /bin/ollama
 
 FROM ubuntu:24.04
 RUN apt-get update \
-    && apt-get install -y ca-certificates libvulkan1 \
+    && apt-get install -y ca-certificates libvulkan1 libopenblas0 \
     && apt-get clean \
     && rm -rf /var/lib/apt/lists/*
 COPY --from=archive /bin /usr/bin

anthropic/anthropic.go

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

anthropic/anthropic_test.go

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

api/client.go

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

cmd/cmd.go

@@ -46,6 +46,8 @@ import (
 	"github.com/ollama/ollama/types/syncmap"
 	"github.com/ollama/ollama/version"
 	xcmd "github.com/ollama/ollama/x/cmd"
+	"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"
@@ -96,6 +98,11 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 	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(".") {
+				quantize, _ := cmd.Flags().GetString("quantize")
+				return imagegenclient.CreateModel(args[0], ".", quantize, p)
+			}
 			reader = strings.NewReader("FROM .\n")
 		} else {
 			return errModelfileNotFound
@@ -457,6 +464,7 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 	}
 
 	name := args[0]
+
 	info, err := func() (*api.ShowResponse, error) {
 		showReq := &api.ShowRequest{Name: name}
 		info, err := client.Show(cmd.Context(), showReq)
@@ -518,8 +526,17 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 		return generateEmbedding(cmd, name, opts.Prompt, opts.KeepAlive, truncate, dimensions)
 	}
 
+	// Check if this is an image generation model
+	if slices.Contains(info.Capabilities, model.CapabilityImageGeneration) {
+		if opts.Prompt == "" && !interactive {
+			return errors.New("image generation models require a prompt. Usage: ollama run " + name + " \"your prompt here\"")
+		}
+		return imagegen.RunCLI(cmd, name, opts.Prompt, interactive, opts.KeepAlive)
+	}
+
 	// Check for experimental flag
 	isExperimental, _ := cmd.Flags().GetBool("experimental")
+	yoloMode, _ := cmd.Flags().GetBool("experimental-yolo")
 
 	if interactive {
 		if err := loadOrUnloadModel(cmd, &opts); err != nil {
@@ -547,9 +564,9 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 			}
 		}
 
-		// Use experimental agent loop with
+		// Use experimental agent loop with tools
 		if isExperimental {
-			return xcmd.GenerateInteractive(cmd, opts.Model, opts.WordWrap, opts.Options, opts.Think, opts.HideThinking, opts.KeepAlive)
+			return xcmd.GenerateInteractive(cmd, opts.Model, opts.WordWrap, opts.Options, opts.Think, opts.HideThinking, opts.KeepAlive, yoloMode)
 		}
 
 		return generateInteractive(cmd, opts)
@@ -655,7 +672,11 @@ func PushHandler(cmd *cobra.Command, args []string) error {
 
 			bar, ok := bars[resp.Digest]
 			if !ok {
-				bar = progress.NewBar(fmt.Sprintf("pushing %s...", resp.Digest[7:19]), resp.Total, resp.Completed)
+				msg := resp.Status
+				if msg == "" {
+					msg = fmt.Sprintf("pushing %s...", resp.Digest[7:19])
+				}
+				bar = progress.NewBar(msg, resp.Total, resp.Completed)
 				bars[resp.Digest] = bar
 				p.Add(resp.Digest, bar)
 			}
@@ -1764,6 +1785,10 @@ func NewCLI() *cobra.Command {
 	runCmd.Flags().Bool("truncate", false, "For embedding models: truncate inputs exceeding context length (default: true). Set --truncate=false to error instead")
 	runCmd.Flags().Int("dimensions", 0, "Truncate output embeddings to specified dimension (embedding models only)")
 	runCmd.Flags().Bool("experimental", false, "Enable experimental agent loop with tools")
+	runCmd.Flags().Bool("experimental-yolo", false, "Skip all tool approval prompts (use with caution)")
+
+	// Image generation flags (width, height, steps, seed, etc.)
+	imagegen.RegisterFlags(runCmd)
 
 	stopCmd := &cobra.Command{
 		Use:     "stop MODEL",

convert/convert.go

@@ -6,11 +6,14 @@ import (
 	"errors"
 	"fmt"
 	"io/fs"
+	"iter"
 	"log/slog"
+	"maps"
 	"os"
 	"slices"
 	"strings"
 
+	ofs "github.com/ollama/ollama/fs"
 	"github.com/ollama/ollama/fs/ggml"
 )
 
@@ -18,8 +21,13 @@ type ModelParameters struct {
 	Architectures []string `json:"architectures"`
 	VocabSize     uint32   `json:"vocab_size"`
 
+	// TODO is this needed?
+	ModelType string `json:"model_type"`
+
 	TextModel struct {
-		VocabSize uint32 `json:"vocab_size"`
+		VocabSize  uint32 `json:"vocab_size"`
+		HiddenSize uint32 `json:"hidden_size"`
+		ModelType  string `json:"model_type"`
 	} `json:"text_config"`
 }
 
@@ -33,8 +41,94 @@ type AdapterParameters struct {
 	} `json:"lora_parameters"`
 }
 
-func (ModelParameters) KV(t *Tokenizer) ggml.KV {
-	kv := ggml.KV{
+type KV map[string]any
+
+func (kv KV) Architecture() string {
+	return kv.String("general.architecture", "unknown")
+}
+
+type valueTypes interface {
+	uint8 | int8 | uint16 | int16 |
+		uint32 | int32 | uint64 | int64 |
+		string | float32 | float64 | bool
+}
+
+type arrayValueTypes interface {
+	[]uint8 | []int8 | []uint16 | []int16 |
+		[]uint32 | []int32 | []uint64 | []int64 |
+		[]string | []float32 | []float64 | []bool
+}
+
+func keyValue[T valueTypes | arrayValueTypes](kv KV, key string, defaultValue ...T) (T, bool) {
+	if !strings.HasPrefix(key, "tokenizer.") && !strings.HasPrefix(key, "general.") {
+		key = kv.Architecture() + "." + key
+	}
+
+	if val, ok := kv[key].(T); ok {
+		return val, true
+	}
+	return defaultValue[0], false
+}
+
+func (kv KV) String(key string, defaultValue ...string) string {
+	val, _ := keyValue(kv, key, append(defaultValue, "")...)
+	return val
+}
+
+func (kv KV) Uint(key string, defaultValue ...uint32) uint32 {
+	val, _ := keyValue(kv, key, append(defaultValue, 0)...)
+	return val
+}
+
+func (kv KV) Float(key string, defaultValue ...float32) float32 {
+	val, _ := keyValue(kv, key, append(defaultValue, 0)...)
+	return val
+}
+
+func (kv KV) Bool(key string, defaultValue ...bool) bool {
+	val, _ := keyValue(kv, key, append(defaultValue, false)...)
+	return val
+}
+
+func (kv KV) Strings(key string, defaultValue ...[]string) []string {
+	val, _ := keyValue(kv, key, append(defaultValue, []string{""})...)
+	return val
+}
+
+func (kv KV) Ints(key string, defaultValue ...[]int32) []int32 {
+	val, _ := keyValue(kv, key, append(defaultValue, []int32{0})...)
+	return val
+}
+
+func (kv KV) Uints(key string, defaultValue ...[]uint32) []uint32 {
+	val, _ := keyValue(kv, key, append(defaultValue, []uint32{0})...)
+	return val
+}
+
+func (kv KV) Floats(key string, defaultValue ...[]float32) []float32 {
+	val, _ := keyValue(kv, key, append(defaultValue, []float32{0})...)
+	return val
+}
+
+func (kv KV) Bools(key string, defaultValue ...[]bool) []bool {
+	val, _ := keyValue(kv, key, append(defaultValue, []bool{false})...)
+	return val
+}
+
+func (kv KV) Len() int {
+	return len(kv)
+}
+
+func (kv KV) Keys() iter.Seq[string] {
+	return maps.Keys(kv)
+}
+
+func (kv KV) Value(key string) any {
+	return kv[key]
+}
+
+func (ModelParameters) KV(t *Tokenizer) KV {
+	kv := KV{
 		"general.file_type":            uint32(1),
 		"general.quantization_version": uint32(2),
 		"tokenizer.ggml.pre":           t.Pre,
@@ -63,7 +157,7 @@ func (ModelParameters) KV(t *Tokenizer) ggml.KV {
 	return kv
 }
 
-func (p AdapterParameters) KV() ggml.KV {
+func (p AdapterParameters) KV() KV {
 	var alpha float32
 	if p.LoraParameters.Alpha == 0 {
 		alpha = float32(p.Alpha)
@@ -71,7 +165,7 @@ func (p AdapterParameters) KV() ggml.KV {
 		alpha = p.LoraParameters.Alpha
 	}
 
-	kv := ggml.KV{
+	kv := KV{
 		"adapter.lora.alpha": alpha,
 		"adapter.type":       "lora",
 		"general.file_type":  uint32(1),
@@ -88,9 +182,14 @@ func (ModelParameters) specialTokenTypes() []string {
 	}
 }
 
-type ModelConverter interface {
+type ModelKV interface {
 	// KV maps parameters to LLM key-values
-	KV(*Tokenizer) ggml.KV
+	KV(*Tokenizer) KV
+}
+
+type ModelConverter interface {
+	ModelKV
+
 	// Tensors maps input tensors to LLM tensors. Model specific modifications can be done here.
 	Tensors([]Tensor) []*ggml.Tensor
 	// Replacements returns a list of string pairs to replace in tensor names.
@@ -107,7 +206,7 @@ type moreParser interface {
 
 type AdapterConverter interface {
 	// KV maps parameters to LLM key-values
-	KV(ggml.KV) ggml.KV
+	KV(ofs.Config) KV
 	// Tensors maps input tensors to LLM tensors. Adapter specific modifications can be done here.
 	Tensors([]Tensor) []*ggml.Tensor
 	// Replacements returns a list of string pairs to replace in tensor names.
@@ -115,7 +214,7 @@ type AdapterConverter interface {
 	Replacements() []string
 }
 
-func ConvertAdapter(fsys fs.FS, f *os.File, baseKV ggml.KV) error {
+func ConvertAdapter(fsys fs.FS, f *os.File, baseKV ofs.Config) error {
 	bts, err := fs.ReadFile(fsys, "adapter_config.json")
 	if err != nil {
 		return err
@@ -126,8 +225,8 @@ func ConvertAdapter(fsys fs.FS, f *os.File, baseKV ggml.KV) error {
 		return err
 	}
 
-	arch, ok := baseKV["general.architecture"]
-	if !ok {
+	arch := baseKV.Architecture()
+	if arch == "" {
 		return errors.New("architecture not set for the base model")
 	}
 
@@ -153,23 +252,19 @@ func ConvertAdapter(fsys fs.FS, f *os.File, baseKV ggml.KV) error {
 	return writeFile(f, conv.KV(baseKV), conv.Tensors(ts))
 }
 
-// Convert writes an Ollama compatible model to the provided io.WriteSeeker based on configurations
-// and files it finds in the input path.
-// Supported input model formats include safetensors.
-// Supported input tokenizers files include tokenizer.json (preferred) and tokenizer.model.
-func ConvertModel(fsys fs.FS, f *os.File) error {
+func LoadModelMetadata(fsys fs.FS) (ModelKV, *Tokenizer, error) {
 	bts, err := fs.ReadFile(fsys, "config.json")
 	if err != nil {
-		return err
+		return nil, nil, err
 	}
 
 	var p ModelParameters
 	if err := json.Unmarshal(bts, &p); err != nil {
-		return err
+		return nil, nil, err
 	}
 
 	if len(p.Architectures) < 1 {
-		return errors.New("unknown architecture")
+		return nil, nil, errors.New("unknown architecture")
 	}
 
 	var conv ModelConverter
@@ -217,22 +312,22 @@ func ConvertModel(fsys fs.FS, f *os.File) error {
 	case "DeepseekV3ForCausalLM":
 		conv = &deepseek2Model{}
 	default:
-		return fmt.Errorf("unsupported architecture %q", p.Architectures[0])
+		return nil, nil, fmt.Errorf("unsupported architecture %q", p.Architectures[0])
 	}
 
 	if err := json.Unmarshal(bts, conv); err != nil {
-		return err
+		return nil, nil, err
 	}
 
 	if t, ok := conv.(moreParser); ok {
 		if err := t.parseMore(fsys); err != nil {
-			return err
+			return nil, nil, err
 		}
 	}
 
 	t, err := parseTokenizer(fsys, conv.specialTokenTypes())
 	if err != nil {
-		return err
+		return nil, nil, err
 	}
 
 	vocabSize := int(cmp.Or(p.VocabSize, p.TextModel.VocabSize))
@@ -254,6 +349,19 @@ func ConvertModel(fsys fs.FS, f *os.File) error {
 	default:
 		slog.Debug("vocabulary", "size", len(t.Vocabulary.Tokens))
 	}
+	return conv, t, nil
+}
+
+// Convert writes an Ollama compatible model to the provided io.WriteSeeker based on configurations
+// and files it finds in the input path.
+// Supported input model formats include safetensors.
+// Supported input tokenizers files include tokenizer.json (preferred) and tokenizer.model.
+func ConvertModel(fsys fs.FS, f *os.File) error {
+	kv, t, err := LoadModelMetadata(fsys)
+	if err != nil {
+		return err
+	}
+	conv := kv.(ModelConverter)
 
 	ts, err := parseTensors(fsys, strings.NewReplacer(conv.Replacements()...))
 	if err != nil {
@@ -263,7 +371,7 @@ func ConvertModel(fsys fs.FS, f *os.File) error {
 	return writeFile(f, conv.KV(t), conv.Tensors(ts))
 }
 
-func writeFile(f *os.File, kv ggml.KV, ts []*ggml.Tensor) error {
+func writeFile(f *os.File, kv KV, ts []*ggml.Tensor) error {
 	for i := range ts {
 		ts[i].Shape = slices.Clone(ts[i].Shape)
 		slices.Reverse(ts[i].Shape)

convert/convert_bert.go

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

convert/convert_commandr.go

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

convert/convert_deepseek2.go

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

convert/convert_deepseekocr.go

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

convert/convert_gemma.go

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

convert/convert_gemma2.go

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

convert/convert_gemma2_adapter.go

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

convert/convert_gemma3.go

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

convert/convert_gemma3n.go

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

convert/convert_gptoss.go

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

convert/convert_llama.go

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

convert/convert_llama4.go

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

convert/convert_llama_adapter.go

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

convert/convert_mistral.go

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

convert/convert_mistral_causal.go

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

convert/convert_mixtral.go

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

convert/convert_mllama.go

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

convert/convert_nomicbert.go

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

convert/convert_olmo.go

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

convert/convert_phi3.go

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

convert/convert_qwen2.go

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

convert/convert_qwen25vl.go

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

convert/convert_qwen3.go

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

convert/convert_qwen3vl.go

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

convert/convert_test.go

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

docs/README.md

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

docs/api/anthropic-compatibility.mdx

@@ -0,0 +1,406 @@
+---
+title: Anthropic compatibility
+---
+
+Ollama provides compatibility with the [Anthropic Messages API](https://docs.anthropic.com/en/api/messages) to help connect existing applications to Ollama, including tools like Claude Code.
+
+## Recommended models
+
+For coding use cases, models like `glm-4.7:cloud`, `minimax-m2.1:cloud`, and `qwen3-coder` are recommended.
+
+Pull a model before use:
+```shell
+ollama pull qwen3-coder
+ollama pull glm-4.7:cloud
+```
+
+## Usage
+
+### Environment variables
+
+To use Ollama with tools that expect the Anthropic API (like Claude Code), set these environment variables:
+
+```shell
+export ANTHROPIC_BASE_URL=http://localhost:11434
+export ANTHROPIC_API_KEY=ollama  # required but ignored
+```
+
+### Simple `/v1/messages` example
+
+<CodeGroup dropdown>
+
+```python basic.py
+import anthropic
+
+client = anthropic.Anthropic(
+    base_url='http://localhost:11434',
+    api_key='ollama',  # required but ignored
+)
+
+message = client.messages.create(
+    model='qwen3-coder',
+    max_tokens=1024,
+    messages=[
+        {'role': 'user', 'content': 'Hello, how are you?'}
+    ]
+)
+print(message.content[0].text)
+```
+
+```javascript basic.js
+import Anthropic from "@anthropic-ai/sdk";
+
+const anthropic = new Anthropic({
+  baseURL: "http://localhost:11434",
+  apiKey: "ollama", // required but ignored
+});
+
+const message = await anthropic.messages.create({
+  model: "qwen3-coder",
+  max_tokens: 1024,
+  messages: [{ role: "user", content: "Hello, how are you?" }],
+});
+
+console.log(message.content[0].text);
+```
+
+```shell basic.sh
+curl -X POST http://localhost:11434/v1/messages \
+-H "Content-Type: application/json" \
+-H "x-api-key: ollama" \
+-H "anthropic-version: 2023-06-01" \
+-d '{
+  "model": "qwen3-coder",
+  "max_tokens": 1024,
+  "messages": [{ "role": "user", "content": "Hello, how are you?" }]
+}'
+```
+
+</CodeGroup>
+
+### Streaming example
+
+<CodeGroup dropdown>
+
+```python streaming.py
+import anthropic
+
+client = anthropic.Anthropic(
+    base_url='http://localhost:11434',
+    api_key='ollama',
+)
+
+with client.messages.stream(
+    model='qwen3-coder',
+    max_tokens=1024,
+    messages=[{'role': 'user', 'content': 'Count from 1 to 10'}]
+) as stream:
+    for text in stream.text_stream:
+        print(text, end='', flush=True)
+```
+
+```javascript streaming.js
+import Anthropic from "@anthropic-ai/sdk";
+
+const anthropic = new Anthropic({
+  baseURL: "http://localhost:11434",
+  apiKey: "ollama",
+});
+
+const stream = await anthropic.messages.stream({
+  model: "qwen3-coder",
+  max_tokens: 1024,
+  messages: [{ role: "user", content: "Count from 1 to 10" }],
+});
+
+for await (const event of stream) {
+  if (
+    event.type === "content_block_delta" &&
+    event.delta.type === "text_delta"
+  ) {
+    process.stdout.write(event.delta.text);
+  }
+}
+```
+
+```shell streaming.sh
+curl -X POST http://localhost:11434/v1/messages \
+-H "Content-Type: application/json" \
+-d '{
+  "model": "qwen3-coder",
+  "max_tokens": 1024,
+  "stream": true,
+  "messages": [{ "role": "user", "content": "Count from 1 to 10" }]
+}'
+```
+
+</CodeGroup>
+
+### Tool calling example
+
+<CodeGroup dropdown>
+
+```python tools.py
+import anthropic
+
+client = anthropic.Anthropic(
+    base_url='http://localhost:11434',
+    api_key='ollama',
+)
+
+message = client.messages.create(
+    model='qwen3-coder',
+    max_tokens=1024,
+    tools=[
+        {
+            'name': 'get_weather',
+            'description': 'Get the current weather in a location',
+            'input_schema': {
+                'type': 'object',
+                'properties': {
+                    'location': {
+                        'type': 'string',
+                        'description': 'The city and state, e.g. San Francisco, CA'
+                    }
+                },
+                'required': ['location']
+            }
+        }
+    ],
+    messages=[{'role': 'user', 'content': "What's the weather in San Francisco?"}]
+)
+
+for block in message.content:
+    if block.type == 'tool_use':
+        print(f'Tool: {block.name}')
+        print(f'Input: {block.input}')
+```
+
+```javascript tools.js
+import Anthropic from "@anthropic-ai/sdk";
+
+const anthropic = new Anthropic({
+  baseURL: "http://localhost:11434",
+  apiKey: "ollama",
+});
+
+const message = await anthropic.messages.create({
+  model: "qwen3-coder",
+  max_tokens: 1024,
+  tools: [
+    {
+      name: "get_weather",
+      description: "Get the current weather in a location",
+      input_schema: {
+        type: "object",
+        properties: {
+          location: {
+            type: "string",
+            description: "The city and state, e.g. San Francisco, CA",
+          },
+        },
+        required: ["location"],
+      },
+    },
+  ],
+  messages: [{ role: "user", content: "What's the weather in San Francisco?" }],
+});
+
+for (const block of message.content) {
+  if (block.type === "tool_use") {
+    console.log("Tool:", block.name);
+    console.log("Input:", block.input);
+  }
+}
+```
+
+```shell tools.sh
+curl -X POST http://localhost:11434/v1/messages \
+-H "Content-Type: application/json" \
+-d '{
+  "model": "qwen3-coder",
+  "max_tokens": 1024,
+  "tools": [
+    {
+      "name": "get_weather",
+      "description": "Get the current weather in a location",
+      "input_schema": {
+        "type": "object",
+        "properties": {
+          "location": {
+            "type": "string",
+            "description": "The city and state"
+          }
+        },
+        "required": ["location"]
+      }
+    }
+  ],
+  "messages": [{ "role": "user", "content": "What is the weather in San Francisco?" }]
+}'
+```
+
+</CodeGroup>
+
+## Using with Claude Code
+
+[Claude Code](https://code.claude.com/docs/en/overview) can be configured to use Ollama as its backend:
+
+```shell
+ANTHROPIC_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_BASE_URL=http://localhost:11434
+export ANTHROPIC_API_KEY=ollama
+```
+
+Then run Claude Code with any Ollama model:
+
+```shell
+# Local models
+claude --model qwen3-coder
+claude --model gpt-oss:20b
+
+# Cloud models
+claude --model glm-4.7:cloud
+claude --model minimax-m2.1:cloud
+```
+
+## Endpoints
+
+### `/v1/messages`
+
+#### Supported features
+
+- [x] Messages
+- [x] Streaming
+- [x] System prompts
+- [x] Multi-turn conversations
+- [x] Vision (images)
+- [x] Tools (function calling)
+- [x] Tool results
+- [x] Thinking/extended thinking
+
+#### Supported request fields
+
+- [x] `model`
+- [x] `max_tokens`
+- [x] `messages`
+  - [x] Text `content`
+  - [x] Image `content` (base64)
+  - [x] Array of content blocks
+  - [x] `tool_use` blocks
+  - [x] `tool_result` blocks
+  - [x] `thinking` blocks
+- [x] `system` (string or array)
+- [x] `stream`
+- [x] `temperature`
+- [x] `top_p`
+- [x] `top_k`
+- [x] `stop_sequences`
+- [x] `tools`
+- [x] `thinking`
+- [ ] `tool_choice`
+- [ ] `metadata`
+
+#### Supported response fields
+
+- [x] `id`
+- [x] `type`
+- [x] `role`
+- [x] `model`
+- [x] `content` (text, tool_use, thinking blocks)
+- [x] `stop_reason` (end_turn, max_tokens, tool_use)
+- [x] `usage` (input_tokens, output_tokens)
+
+#### Streaming events
+
+- [x] `message_start`
+- [x] `content_block_start`
+- [x] `content_block_delta` (text_delta, input_json_delta, thinking_delta)
+- [x] `content_block_stop`
+- [x] `message_delta`
+- [x] `message_stop`
+- [x] `ping`
+- [x] `error`
+
+## Models
+
+Ollama supports both local and cloud models.
+
+### Local models
+
+Pull a local model before use:
+
+```shell
+ollama pull qwen3-coder
+```
+
+Recommended local models:
+- `qwen3-coder` - Excellent for coding tasks
+- `gpt-oss:20b` - Strong general-purpose model
+
+### Cloud models
+
+Cloud models are available immediately without pulling:
+
+- `glm-4.7:cloud` - High-performance cloud model
+- `minimax-m2.1:cloud` - Fast cloud model
+
+### Default model names
+
+For tooling that relies on default Anthropic model names such as `claude-3-5-sonnet`, use `ollama cp` to copy an existing model name:
+
+```shell
+ollama cp qwen3-coder claude-3-5-sonnet
+```
+
+Afterwards, this new model name can be specified in the `model` field:
+
+```shell
+curl http://localhost:11434/v1/messages \
+    -H "Content-Type: application/json" \
+    -d '{
+        "model": "claude-3-5-sonnet",
+        "max_tokens": 1024,
+        "messages": [
+            {
+                "role": "user",
+                "content": "Hello!"
+            }
+        ]
+    }'
+```
+
+## Differences from the Anthropic API
+
+### Behavior differences
+
+- API key is accepted but not validated
+- `anthropic-version` header is accepted but not used
+- Token counts are approximations based on the underlying model's tokenizer
+
+### Not supported
+
+The following Anthropic API features are not currently supported:
+
+| Feature | Description |
+|---------|-------------|
+| `/v1/messages/count_tokens` | Token counting endpoint |
+| `tool_choice` | Forcing specific tool use or disabling tools |
+| `metadata` | Request metadata (user_id) |
+| Prompt caching | `cache_control` blocks for caching prefixes |
+| Batches API | `/v1/messages/batches` for async batch processing |
+| Citations | `citations` content blocks |
+| PDF support | `document` content blocks with PDF files |
+| Server-sent errors | `error` events during streaming (errors return HTTP status) |
+
+### Partial support
+
+| Feature | Status |
+|---------|--------|
+| Image content | Base64 images supported; URL images not supported |
+| Extended thinking | Basic support; `budget_tokens` accepted but not enforced |

docs/docs.json

@@ -32,7 +32,9 @@
     "codeblocks": "system"
   },
   "contextual": {
-    "options": ["copy"]
+    "options": [
+      "copy"
+    ]
   },
   "navbar": {
     "links": [
@@ -52,7 +54,9 @@
       "display": "simple"
     },
     "examples": {
-      "languages": ["curl"]
+      "languages": [
+        "curl"
+      ]
     }
   },
   "redirects": [
@@ -97,6 +101,7 @@
           {
             "group": "Integrations",
             "pages": [
+              "/integrations/claude-code",
               "/integrations/vscode",
               "/integrations/jetbrains",
               "/integrations/codex",
@@ -139,7 +144,8 @@
               "/api/streaming",
               "/api/usage",
               "/api/errors",
-              "/api/openai-compatibility"
+              "/api/openai-compatibility",
+              "/api/anthropic-compatibility"
             ]
           },
           {

docs/integrations/claude-code.mdx

@@ -0,0 +1,69 @@
+---
+title: Claude Code
+---
+
+## Install
+
+Install [Claude Code](https://code.claude.com/docs/en/overview):
+
+<CodeGroup>
+
+```shell macOS / Linux
+curl -fsSL https://claude.ai/install.sh | bash
+```
+
+```powershell Windows
+irm https://claude.ai/install.ps1 | iex
+```
+
+</CodeGroup>
+
+## Usage with Ollama
+
+Claude Code connects to Ollama using the Anthropic-compatible API.
+
+1. Set the environment variables:
+
+```shell
+export ANTHROPIC_BASE_URL=http://localhost:11434
+export ANTHROPIC_API_KEY=ollama
+```
+
+2. Run Claude Code with an Ollama model:
+
+```shell
+claude --model qwen3-coder
+```
+
+Or run with environment variables inline:
+
+```shell
+ANTHROPIC_BASE_URL=http://localhost:11434 ANTHROPIC_API_KEY=ollama claude --model qwen3-coder
+```
+
+## Connecting to ollama.com
+
+1. Create an [API key](https://ollama.com/settings/keys) on ollama.com
+2. Set the environment variables:
+
+```shell
+export ANTHROPIC_BASE_URL=https://ollama.com
+export ANTHROPIC_API_KEY=<your-api-key>
+```
+
+3. Run Claude Code with a cloud model:
+
+```shell
+claude --model glm-4.7:cloud
+```
+
+## Recommended Models
+
+### Cloud models
+- `glm-4.7:cloud` - High-performance cloud model
+- `minimax-m2.1:cloud` - Fast cloud model
+- `qwen3-coder:480b` - Large coding model
+
+### Local models
+- `qwen3-coder` - Excellent for coding tasks
+- `gpt-oss:20b` - Strong general-purpose model

docs/linux.mdx

@@ -1,5 +1,5 @@
 ---
-title: Linux
+title: "Linux"
 ---
 
 ## Install
@@ -13,8 +13,7 @@ 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:
@@ -113,11 +112,7 @@ 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
@@ -196,4 +191,4 @@ Remove the downloaded models and Ollama service user and group:
 sudo userdel ollama
 sudo groupdel ollama
 sudo rm -r /usr/share/ollama
-```
+```

fs/config.go

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

fs/ggml/ggml.go

@@ -6,7 +6,9 @@ import (
 	"errors"
 	"fmt"
 	"io"
+	"iter"
 	"log/slog"
+	"maps"
 	"math"
 	"slices"
 	"strings"
@@ -239,6 +241,18 @@ func (kv KV) Bools(key string, defaultValue ...[]bool) []bool {
 	return val.values
 }
 
+func (kv KV) Len() int {
+	return len(kv)
+}
+
+func (kv KV) Keys() iter.Seq[string] {
+	return maps.Keys(kv)
+}
+
+func (kv KV) Value(key string) any {
+	return kv[key]
+}
+
 func (kv KV) OllamaEngineRequired() bool {
 	return slices.Contains([]string{
 		"bert",

fs/ggml/gguf.go

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

middleware/anthropic.go

@@ -0,0 +1,149 @@
+package middleware
+
+import (
+	"bytes"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+
+	"github.com/gin-gonic/gin"
+
+	"github.com/ollama/ollama/anthropic"
+	"github.com/ollama/ollama/api"
+)
+
+// AnthropicWriter wraps the response writer to transform Ollama responses to Anthropic format
+type AnthropicWriter struct {
+	BaseWriter
+	stream    bool
+	id        string
+	model     string
+	converter *anthropic.StreamConverter
+}
+
+func (w *AnthropicWriter) writeError(data []byte) (int, error) {
+	var errData struct {
+		Error string `json:"error"`
+	}
+	if err := json.Unmarshal(data, &errData); err != nil {
+		return 0, err
+	}
+
+	w.ResponseWriter.Header().Set("Content-Type", "application/json")
+	err := json.NewEncoder(w.ResponseWriter).Encode(anthropic.NewError(w.ResponseWriter.Status(), errData.Error))
+	if err != nil {
+		return 0, err
+	}
+
+	return len(data), nil
+}
+
+func (w *AnthropicWriter) writeEvent(eventType string, data any) error {
+	d, err := json.Marshal(data)
+	if err != nil {
+		return err
+	}
+	_, err = w.ResponseWriter.Write([]byte(fmt.Sprintf("event: %s\ndata: %s\n\n", eventType, d)))
+	if err != nil {
+		return err
+	}
+	if f, ok := w.ResponseWriter.(http.Flusher); ok {
+		f.Flush()
+	}
+	return nil
+}
+
+func (w *AnthropicWriter) writeResponse(data []byte) (int, error) {
+	var chatResponse api.ChatResponse
+	err := json.Unmarshal(data, &chatResponse)
+	if err != nil {
+		return 0, err
+	}
+
+	if w.stream {
+		w.ResponseWriter.Header().Set("Content-Type", "text/event-stream")
+
+		events := w.converter.Process(chatResponse)
+		for _, event := range events {
+			if err := w.writeEvent(event.Event, event.Data); err != nil {
+				return 0, err
+			}
+		}
+		return len(data), nil
+	}
+
+	w.ResponseWriter.Header().Set("Content-Type", "application/json")
+	response := anthropic.ToMessagesResponse(w.id, chatResponse)
+	return len(data), json.NewEncoder(w.ResponseWriter).Encode(response)
+}
+
+func (w *AnthropicWriter) Write(data []byte) (int, error) {
+	code := w.ResponseWriter.Status()
+	if code != http.StatusOK {
+		return w.writeError(data)
+	}
+
+	return w.writeResponse(data)
+}
+
+// AnthropicMessagesMiddleware handles Anthropic Messages API requests
+func AnthropicMessagesMiddleware() gin.HandlerFunc {
+	return func(c *gin.Context) {
+		var req anthropic.MessagesRequest
+		err := c.ShouldBindJSON(&req)
+		if err != nil {
+			c.AbortWithStatusJSON(http.StatusBadRequest, anthropic.NewError(http.StatusBadRequest, err.Error()))
+			return
+		}
+
+		if req.Model == "" {
+			c.AbortWithStatusJSON(http.StatusBadRequest, anthropic.NewError(http.StatusBadRequest, "model is required"))
+			return
+		}
+
+		if req.MaxTokens <= 0 {
+			c.AbortWithStatusJSON(http.StatusBadRequest, anthropic.NewError(http.StatusBadRequest, "max_tokens is required and must be positive"))
+			return
+		}
+
+		if len(req.Messages) == 0 {
+			c.AbortWithStatusJSON(http.StatusBadRequest, anthropic.NewError(http.StatusBadRequest, "messages is required"))
+			return
+		}
+
+		chatReq, err := anthropic.FromMessagesRequest(req)
+		if err != nil {
+			c.AbortWithStatusJSON(http.StatusBadRequest, anthropic.NewError(http.StatusBadRequest, err.Error()))
+			return
+		}
+
+		var b bytes.Buffer
+		if err := json.NewEncoder(&b).Encode(chatReq); err != nil {
+			c.AbortWithStatusJSON(http.StatusInternalServerError, anthropic.NewError(http.StatusInternalServerError, err.Error()))
+			return
+		}
+
+		c.Request.Body = io.NopCloser(&b)
+
+		messageID := anthropic.GenerateMessageID()
+
+		w := &AnthropicWriter{
+			BaseWriter: BaseWriter{ResponseWriter: c.Writer},
+			stream:     req.Stream,
+			id:         messageID,
+			model:      req.Model,
+			converter:  anthropic.NewStreamConverter(messageID, req.Model),
+		}
+
+		if req.Stream {
+			c.Writer.Header().Set("Content-Type", "text/event-stream")
+			c.Writer.Header().Set("Cache-Control", "no-cache")
+			c.Writer.Header().Set("Connection", "keep-alive")
+		}
+
+		c.Writer = w
+
+		c.Next()
+	}
+}

middleware/anthropic_test.go

@@ -0,0 +1,584 @@
+package middleware
+
+import (
+	"bytes"
+	"encoding/json"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	"github.com/gin-gonic/gin"
+	"github.com/google/go-cmp/cmp"
+	"github.com/google/go-cmp/cmp/cmpopts"
+
+	"github.com/ollama/ollama/anthropic"
+	"github.com/ollama/ollama/api"
+)
+
+func captureAnthropicRequest(capturedRequest any) gin.HandlerFunc {
+	return func(c *gin.Context) {
+		bodyBytes, _ := io.ReadAll(c.Request.Body)
+		c.Request.Body = io.NopCloser(bytes.NewReader(bodyBytes))
+		_ = json.Unmarshal(bodyBytes, capturedRequest)
+		c.Next()
+	}
+}
+
+// testProps creates ToolPropertiesMap from a map (convenience function for tests)
+func testProps(m map[string]api.ToolProperty) *api.ToolPropertiesMap {
+	props := api.NewToolPropertiesMap()
+	for k, v := range m {
+		props.Set(k, v)
+	}
+	return props
+}
+
+func TestAnthropicMessagesMiddleware(t *testing.T) {
+	type testCase struct {
+		name string
+		body string
+		req  api.ChatRequest
+		err  anthropic.ErrorResponse
+	}
+
+	var capturedRequest *api.ChatRequest
+	stream := true
+
+	testCases := []testCase{
+		{
+			name: "basic message",
+			body: `{
+				"model": "test-model",
+				"max_tokens": 1024,
+				"messages": [
+					{"role": "user", "content": "Hello"}
+				]
+			}`,
+			req: api.ChatRequest{
+				Model: "test-model",
+				Messages: []api.Message{
+					{Role: "user", Content: "Hello"},
+				},
+				Options: map[string]any{"num_predict": 1024},
+				Stream:  &False,
+			},
+		},
+		{
+			name: "with system prompt",
+			body: `{
+				"model": "test-model",
+				"max_tokens": 1024,
+				"system": "You are helpful.",
+				"messages": [
+					{"role": "user", "content": "Hello"}
+				]
+			}`,
+			req: api.ChatRequest{
+				Model: "test-model",
+				Messages: []api.Message{
+					{Role: "system", Content: "You are helpful."},
+					{Role: "user", Content: "Hello"},
+				},
+				Options: map[string]any{"num_predict": 1024},
+				Stream:  &False,
+			},
+		},
+		{
+			name: "with options",
+			body: `{
+				"model": "test-model",
+				"max_tokens": 2048,
+				"temperature": 0.7,
+				"top_p": 0.9,
+				"top_k": 40,
+				"stop_sequences": ["\n", "END"],
+				"messages": [
+					{"role": "user", "content": "Hello"}
+				]
+			}`,
+			req: api.ChatRequest{
+				Model: "test-model",
+				Messages: []api.Message{
+					{Role: "user", Content: "Hello"},
+				},
+				Options: map[string]any{
+					"num_predict": 2048,
+					"temperature": 0.7,
+					"top_p":       0.9,
+					"top_k":       40,
+					"stop":        []string{"\n", "END"},
+				},
+				Stream: &False,
+			},
+		},
+		{
+			name: "streaming",
+			body: `{
+				"model": "test-model",
+				"max_tokens": 1024,
+				"stream": true,
+				"messages": [
+					{"role": "user", "content": "Hello"}
+				]
+			}`,
+			req: api.ChatRequest{
+				Model: "test-model",
+				Messages: []api.Message{
+					{Role: "user", Content: "Hello"},
+				},
+				Options: map[string]any{"num_predict": 1024},
+				Stream:  &stream,
+			},
+		},
+		{
+			name: "with tools",
+			body: `{
+				"model": "test-model",
+				"max_tokens": 1024,
+				"messages": [
+					{"role": "user", "content": "What's the weather?"}
+				],
+				"tools": [{
+					"name": "get_weather",
+					"description": "Get current weather",
+					"input_schema": {
+						"type": "object",
+						"properties": {
+							"location": {"type": "string"}
+						},
+						"required": ["location"]
+					}
+				}]
+			}`,
+			req: api.ChatRequest{
+				Model: "test-model",
+				Messages: []api.Message{
+					{Role: "user", Content: "What's the weather?"},
+				},
+				Tools: []api.Tool{
+					{
+						Type: "function",
+						Function: api.ToolFunction{
+							Name:        "get_weather",
+							Description: "Get current weather",
+							Parameters: api.ToolFunctionParameters{
+								Type:     "object",
+								Required: []string{"location"},
+								Properties: testProps(map[string]api.ToolProperty{
+									"location": {Type: api.PropertyType{"string"}},
+								}),
+							},
+						},
+					},
+				},
+				Options: map[string]any{"num_predict": 1024},
+				Stream:  &False,
+			},
+		},
+		{
+			name: "with tool result",
+			body: `{
+				"model": "test-model",
+				"max_tokens": 1024,
+				"messages": [
+					{"role": "user", "content": "What's the weather?"},
+					{"role": "assistant", "content": [
+						{"type": "tool_use", "id": "call_123", "name": "get_weather", "input": {"location": "Paris"}}
+					]},
+					{"role": "user", "content": [
+						{"type": "tool_result", "tool_use_id": "call_123", "content": "Sunny, 22°C"}
+					]}
+				]
+			}`,
+			req: api.ChatRequest{
+				Model: "test-model",
+				Messages: []api.Message{
+					{Role: "user", Content: "What's the weather?"},
+					{
+						Role: "assistant",
+						ToolCalls: []api.ToolCall{
+							{
+								ID: "call_123",
+								Function: api.ToolCallFunction{
+									Name:      "get_weather",
+									Arguments: testArgs(map[string]any{"location": "Paris"}),
+								},
+							},
+						},
+					},
+					{Role: "tool", Content: "Sunny, 22°C", ToolCallID: "call_123"},
+				},
+				Options: map[string]any{"num_predict": 1024},
+				Stream:  &False,
+			},
+		},
+		{
+			name: "with thinking enabled",
+			body: `{
+				"model": "test-model",
+				"max_tokens": 1024,
+				"thinking": {"type": "enabled", "budget_tokens": 1000},
+				"messages": [
+					{"role": "user", "content": "Hello"}
+				]
+			}`,
+			req: api.ChatRequest{
+				Model: "test-model",
+				Messages: []api.Message{
+					{Role: "user", Content: "Hello"},
+				},
+				Options: map[string]any{"num_predict": 1024},
+				Stream:  &False,
+				Think:   &api.ThinkValue{Value: true},
+			},
+		},
+		{
+			name: "missing model error",
+			body: `{
+				"max_tokens": 1024,
+				"messages": [
+					{"role": "user", "content": "Hello"}
+				]
+			}`,
+			err: anthropic.ErrorResponse{
+				Type: "error",
+				Error: anthropic.Error{
+					Type:    "invalid_request_error",
+					Message: "model is required",
+				},
+			},
+		},
+		{
+			name: "missing max_tokens error",
+			body: `{
+				"model": "test-model",
+				"messages": [
+					{"role": "user", "content": "Hello"}
+				]
+			}`,
+			err: anthropic.ErrorResponse{
+				Type: "error",
+				Error: anthropic.Error{
+					Type:    "invalid_request_error",
+					Message: "max_tokens is required and must be positive",
+				},
+			},
+		},
+		{
+			name: "missing messages error",
+			body: `{
+				"model": "test-model",
+				"max_tokens": 1024
+			}`,
+			err: anthropic.ErrorResponse{
+				Type: "error",
+				Error: anthropic.Error{
+					Type:    "invalid_request_error",
+					Message: "messages is required",
+				},
+			},
+		},
+		{
+			name: "tool_use missing id error",
+			body: `{
+				"model": "test-model",
+				"max_tokens": 1024,
+				"messages": [
+					{"role": "assistant", "content": [
+						{"type": "tool_use", "name": "test"}
+					]}
+				]
+			}`,
+			err: anthropic.ErrorResponse{
+				Type: "error",
+				Error: anthropic.Error{
+					Type:    "invalid_request_error",
+					Message: "tool_use block missing required 'id' field",
+				},
+			},
+		},
+	}
+
+	endpoint := func(c *gin.Context) {
+		c.Status(http.StatusOK)
+	}
+
+	gin.SetMode(gin.TestMode)
+	router := gin.New()
+	router.Use(AnthropicMessagesMiddleware(), captureAnthropicRequest(&capturedRequest))
+	router.Handle(http.MethodPost, "/v1/messages", endpoint)
+
+	for _, tc := range testCases {
+		t.Run(tc.name, func(t *testing.T) {
+			req, _ := http.NewRequest(http.MethodPost, "/v1/messages", strings.NewReader(tc.body))
+			req.Header.Set("Content-Type", "application/json")
+
+			defer func() { capturedRequest = nil }()
+
+			resp := httptest.NewRecorder()
+			router.ServeHTTP(resp, req)
+
+			if tc.err.Type != "" {
+				// Expect error
+				if resp.Code == http.StatusOK {
+					t.Fatalf("expected error response, got 200 OK")
+				}
+				var errResp anthropic.ErrorResponse
+				if err := json.Unmarshal(resp.Body.Bytes(), &errResp); err != nil {
+					t.Fatalf("failed to unmarshal error: %v", err)
+				}
+				if errResp.Type != tc.err.Type {
+					t.Errorf("expected error type %q, got %q", tc.err.Type, errResp.Type)
+				}
+				if errResp.Error.Type != tc.err.Error.Type {
+					t.Errorf("expected error.type %q, got %q", tc.err.Error.Type, errResp.Error.Type)
+				}
+				if errResp.Error.Message != tc.err.Error.Message {
+					t.Errorf("expected error.message %q, got %q", tc.err.Error.Message, errResp.Error.Message)
+				}
+				return
+			}
+
+			if resp.Code != http.StatusOK {
+				t.Fatalf("unexpected status code: %d, body: %s", resp.Code, resp.Body.String())
+			}
+
+			if capturedRequest == nil {
+				t.Fatal("request was not captured")
+			}
+
+			// Compare relevant fields
+			if capturedRequest.Model != tc.req.Model {
+				t.Errorf("model mismatch: got %q, want %q", capturedRequest.Model, tc.req.Model)
+			}
+
+			if diff := cmp.Diff(tc.req.Messages, capturedRequest.Messages,
+				cmpopts.IgnoreUnexported(api.ToolCallFunctionArguments{}, api.ToolPropertiesMap{})); diff != "" {
+				t.Errorf("messages mismatch (-want +got):\n%s", diff)
+			}
+
+			if tc.req.Stream != nil && capturedRequest.Stream != nil {
+				if *tc.req.Stream != *capturedRequest.Stream {
+					t.Errorf("stream mismatch: got %v, want %v", *capturedRequest.Stream, *tc.req.Stream)
+				}
+			}
+
+			if tc.req.Think != nil {
+				if capturedRequest.Think == nil {
+					t.Error("expected Think to be set")
+				} else if capturedRequest.Think.Value != tc.req.Think.Value {
+					t.Errorf("Think mismatch: got %v, want %v", capturedRequest.Think.Value, tc.req.Think.Value)
+				}
+			}
+		})
+	}
+}
+
+func TestAnthropicMessagesMiddleware_Headers(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
+	t.Run("streaming sets correct headers", func(t *testing.T) {
+		router := gin.New()
+		router.Use(AnthropicMessagesMiddleware())
+		router.POST("/v1/messages", func(c *gin.Context) {
+			// Check headers were set
+			if c.Writer.Header().Get("Content-Type") != "text/event-stream" {
+				t.Errorf("expected Content-Type text/event-stream, got %q", c.Writer.Header().Get("Content-Type"))
+			}
+			if c.Writer.Header().Get("Cache-Control") != "no-cache" {
+				t.Errorf("expected Cache-Control no-cache, got %q", c.Writer.Header().Get("Cache-Control"))
+			}
+			c.Status(http.StatusOK)
+		})
+
+		body := `{"model": "test", "max_tokens": 100, "stream": true, "messages": [{"role": "user", "content": "Hi"}]}`
+		req, _ := http.NewRequest(http.MethodPost, "/v1/messages", strings.NewReader(body))
+		req.Header.Set("Content-Type", "application/json")
+
+		resp := httptest.NewRecorder()
+		router.ServeHTTP(resp, req)
+	})
+}
+
+func TestAnthropicMessagesMiddleware_InvalidJSON(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	router := gin.New()
+	router.Use(AnthropicMessagesMiddleware())
+	router.POST("/v1/messages", func(c *gin.Context) {
+		c.Status(http.StatusOK)
+	})
+
+	req, _ := http.NewRequest(http.MethodPost, "/v1/messages", strings.NewReader(`{invalid json`))
+	req.Header.Set("Content-Type", "application/json")
+
+	resp := httptest.NewRecorder()
+	router.ServeHTTP(resp, req)
+
+	if resp.Code != http.StatusBadRequest {
+		t.Errorf("expected status 400, got %d", resp.Code)
+	}
+
+	var errResp anthropic.ErrorResponse
+	if err := json.Unmarshal(resp.Body.Bytes(), &errResp); err != nil {
+		t.Fatalf("failed to unmarshal error: %v", err)
+	}
+
+	if errResp.Type != "error" {
+		t.Errorf("expected type 'error', got %q", errResp.Type)
+	}
+	if errResp.Error.Type != "invalid_request_error" {
+		t.Errorf("expected error type 'invalid_request_error', got %q", errResp.Error.Type)
+	}
+}
+
+func TestAnthropicWriter_NonStreaming(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
+	router := gin.New()
+	router.Use(AnthropicMessagesMiddleware())
+	router.POST("/v1/messages", func(c *gin.Context) {
+		// Simulate Ollama response
+		resp := api.ChatResponse{
+			Model: "test-model",
+			Message: api.Message{
+				Role:    "assistant",
+				Content: "Hello there!",
+			},
+			Done:       true,
+			DoneReason: "stop",
+			Metrics: api.Metrics{
+				PromptEvalCount: 10,
+				EvalCount:       5,
+			},
+		}
+		data, _ := json.Marshal(resp)
+		c.Writer.WriteHeader(http.StatusOK)
+		_, _ = c.Writer.Write(data)
+	})
+
+	body := `{"model": "test-model", "max_tokens": 100, "messages": [{"role": "user", "content": "Hi"}]}`
+	req, _ := http.NewRequest(http.MethodPost, "/v1/messages", strings.NewReader(body))
+	req.Header.Set("Content-Type", "application/json")
+
+	resp := httptest.NewRecorder()
+	router.ServeHTTP(resp, req)
+
+	if resp.Code != http.StatusOK {
+		t.Fatalf("expected status 200, got %d", resp.Code)
+	}
+
+	var result anthropic.MessagesResponse
+	if err := json.Unmarshal(resp.Body.Bytes(), &result); err != nil {
+		t.Fatalf("failed to unmarshal response: %v", err)
+	}
+
+	if result.Type != "message" {
+		t.Errorf("expected type 'message', got %q", result.Type)
+	}
+	if result.Role != "assistant" {
+		t.Errorf("expected role 'assistant', got %q", result.Role)
+	}
+	if len(result.Content) != 1 {
+		t.Fatalf("expected 1 content block, got %d", len(result.Content))
+	}
+	if result.Content[0].Text == nil || *result.Content[0].Text != "Hello there!" {
+		t.Errorf("expected text 'Hello there!', got %v", result.Content[0].Text)
+	}
+	if result.StopReason != "end_turn" {
+		t.Errorf("expected stop_reason 'end_turn', got %q", result.StopReason)
+	}
+	if result.Usage.InputTokens != 10 {
+		t.Errorf("expected input_tokens 10, got %d", result.Usage.InputTokens)
+	}
+	if result.Usage.OutputTokens != 5 {
+		t.Errorf("expected output_tokens 5, got %d", result.Usage.OutputTokens)
+	}
+}
+
+// TestAnthropicWriter_ErrorFromRoutes tests error handling when routes.go sends
+// gin.H{"error": "message"} without a StatusCode field (which is the common case)
+func TestAnthropicWriter_ErrorFromRoutes(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
+	tests := []struct {
+		name          string
+		statusCode    int
+		errorPayload  any
+		wantErrorType string
+		wantMessage   string
+	}{
+		// routes.go sends errors without StatusCode in JSON, so we must use HTTP status
+		{
+			name:          "404 with gin.H error (model not found)",
+			statusCode:    http.StatusNotFound,
+			errorPayload:  gin.H{"error": "model 'nonexistent' not found"},
+			wantErrorType: "not_found_error",
+			wantMessage:   "model 'nonexistent' not found",
+		},
+		{
+			name:          "400 with gin.H error (bad request)",
+			statusCode:    http.StatusBadRequest,
+			errorPayload:  gin.H{"error": "model is required"},
+			wantErrorType: "invalid_request_error",
+			wantMessage:   "model is required",
+		},
+		{
+			name:          "500 with gin.H error (internal error)",
+			statusCode:    http.StatusInternalServerError,
+			errorPayload:  gin.H{"error": "something went wrong"},
+			wantErrorType: "api_error",
+			wantMessage:   "something went wrong",
+		},
+		{
+			name:       "404 with api.StatusError",
+			statusCode: http.StatusNotFound,
+			errorPayload: api.StatusError{
+				StatusCode:   http.StatusNotFound,
+				ErrorMessage: "model not found via StatusError",
+			},
+			wantErrorType: "not_found_error",
+			wantMessage:   "model not found via StatusError",
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			router := gin.New()
+			router.Use(AnthropicMessagesMiddleware())
+			router.POST("/v1/messages", func(c *gin.Context) {
+				// Simulate what routes.go does - set status and write error JSON
+				data, _ := json.Marshal(tt.errorPayload)
+				c.Writer.WriteHeader(tt.statusCode)
+				_, _ = c.Writer.Write(data)
+			})
+
+			body := `{"model": "test-model", "max_tokens": 100, "messages": [{"role": "user", "content": "Hi"}]}`
+			req, _ := http.NewRequest(http.MethodPost, "/v1/messages", strings.NewReader(body))
+			req.Header.Set("Content-Type", "application/json")
+
+			resp := httptest.NewRecorder()
+			router.ServeHTTP(resp, req)
+
+			if resp.Code != tt.statusCode {
+				t.Errorf("expected status %d, got %d", tt.statusCode, resp.Code)
+			}
+
+			var errResp anthropic.ErrorResponse
+			if err := json.Unmarshal(resp.Body.Bytes(), &errResp); err != nil {
+				t.Fatalf("failed to unmarshal error response: %v\nbody: %s", err, resp.Body.String())
+			}
+
+			if errResp.Type != "error" {
+				t.Errorf("expected type 'error', got %q", errResp.Type)
+			}
+			if errResp.Error.Type != tt.wantErrorType {
+				t.Errorf("expected error type %q, got %q", tt.wantErrorType, errResp.Error.Type)
+			}
+			if errResp.Error.Message != tt.wantMessage {
+				t.Errorf("expected message %q, got %q", tt.wantMessage, errResp.Error.Message)
+			}
+		})
+	}
+}

parser/parser_test.go

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

progress/stepbar.go

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

readline/types.go

@@ -18,6 +18,7 @@ const (
 	CharCtrlL     = 12
 	CharEnter     = 13
 	CharNext      = 14
+	CharCtrlO     = 15 // Ctrl+O - used for expanding tool output
 	CharPrev      = 16
 	CharBckSearch = 18
 	CharFwdSearch = 19

runner/runner.go

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

scripts/build_darwin.sh

@@ -42,18 +42,39 @@ shift $(( $OPTIND - 1 ))
 _build_darwin() {
     for ARCH in $ARCHS; do
         status "Building darwin $ARCH"
-        INSTALL_PREFIX=dist/darwin-$ARCH/
-        GOOS=darwin GOARCH=$ARCH CGO_ENABLED=1 go build -o $INSTALL_PREFIX .
+        INSTALL_PREFIX=dist/darwin-$ARCH/        
 
         if [ "$ARCH" = "amd64" ]; then
             status "Building darwin $ARCH dynamic backends"
-            cmake -B build/darwin-$ARCH \
+            BUILD_DIR=build/darwin-$ARCH
+            cmake -B $BUILD_DIR \
                 -DCMAKE_OSX_ARCHITECTURES=x86_64 \
-                -DCMAKE_OSX_DEPLOYMENT_TARGET=11.3 \
+                -DCMAKE_OSX_DEPLOYMENT_TARGET=14.0 \
+                -DCMAKE_INSTALL_PREFIX=$INSTALL_PREFIX \
+                -DMLX_ENGINE=ON \
+                -DMLX_ENABLE_X64_MAC=ON \
+                -DOLLAMA_RUNNER_DIR=./
+            cmake --build $BUILD_DIR --target ggml-cpu -j
+            cmake --build $BUILD_DIR --target mlx mlxc -j
+            cmake --install $BUILD_DIR --component CPU
+            cmake --install $BUILD_DIR --component MLX
+            # Override CGO flags to point to the amd64 build directory
+            MLX_CGO_CFLAGS="-O3 -I$(pwd)/$BUILD_DIR/_deps/mlx-c-src -mmacosx-version-min=14.0"
+            MLX_CGO_LDFLAGS="-L$(pwd)/$BUILD_DIR/lib/ollama -lmlxc -lmlx -Wl,-rpath,@executable_path -lc++ -framework Accelerate -mmacosx-version-min=14.0"
+        else
+            BUILD_DIR=build
+            cmake --preset MLX \
+                -DOLLAMA_RUNNER_DIR=./ \
+                -DCMAKE_OSX_DEPLOYMENT_TARGET=14.0 \
                 -DCMAKE_INSTALL_PREFIX=$INSTALL_PREFIX
-            cmake --build build/darwin-$ARCH --target ggml-cpu -j
-            cmake --install build/darwin-$ARCH --component CPU
+            cmake --build --preset MLX --parallel
+            cmake --install $BUILD_DIR --component MLX
+            # Use default CGO flags from mlx.go for arm64
+            MLX_CGO_CFLAGS="-O3 -I$(pwd)/$BUILD_DIR/_deps/mlx-c-src -mmacosx-version-min=14.0"
+            MLX_CGO_LDFLAGS="-L$(pwd)/$BUILD_DIR/lib/ollama -lmlxc -lmlx -Wl,-rpath,@executable_path -lc++ -framework Metal -framework Foundation -framework Accelerate -mmacosx-version-min=14.0"
         fi
+        GOOS=darwin GOARCH=$ARCH CGO_ENABLED=1 CGO_CFLAGS="$MLX_CGO_CFLAGS" CGO_LDFLAGS="$MLX_CGO_LDFLAGS" go build -tags mlx -o $INSTALL_PREFIX/imagegen ./x/imagegen/cmd/engine
+        GOOS=darwin GOARCH=$ARCH CGO_ENABLED=1 go build -o $INSTALL_PREFIX .
     done
 }
 
@@ -61,10 +82,12 @@ _sign_darwin() {
     status "Creating universal binary..."
     mkdir -p dist/darwin
     lipo -create -output dist/darwin/ollama dist/darwin-*/ollama
+    lipo -create -output dist/darwin/imagegen dist/darwin-*/imagegen
     chmod +x dist/darwin/ollama
+    chmod +x dist/darwin/imagegen
 
     if [ -n "$APPLE_IDENTITY" ]; then
-        for F in dist/darwin/ollama dist/darwin-amd64/lib/ollama/*; do
+        for F in dist/darwin/ollama dist/darwin-*/lib/ollama/* dist/darwin/imagegen; do
             codesign -f --timestamp -s "$APPLE_IDENTITY" --identifier ai.ollama.ollama --options=runtime $F
         done
 
@@ -131,17 +154,23 @@ _build_macapp() {
     mkdir -p dist/Ollama.app/Contents/Resources
     if [ -d dist/darwin-amd64 ]; then
         lipo -create -output dist/Ollama.app/Contents/Resources/ollama dist/darwin-amd64/ollama dist/darwin-arm64/ollama
-        cp dist/darwin-amd64/lib/ollama/*.so dist/darwin-amd64/lib/ollama/*.dylib dist/Ollama.app/Contents/Resources/
+        lipo -create -output dist/Ollama.app/Contents/Resources/imagegen dist/darwin-amd64/imagegen dist/darwin-arm64/imagegen
+        for F in dist/darwin-amd64/lib/ollama/*mlx*.dylib ; do
+            lipo -create -output dist/darwin/$(basename $F) $F dist/darwin-arm64/lib/ollama/$(basename $F)
+        done
+        cp dist/darwin-*/lib/ollama/*.so dist/darwin-*/lib/ollama/*.dylib dist/Ollama.app/Contents/Resources/
+        cp dist/darwin/*.dylib dist/Ollama.app/Contents/Resources/
     else
         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/imagegen dist/Ollama.app/Contents/Resources/imagegen
     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 ; do
+        for lib in dist/Ollama.app/Contents/Resources/*.so dist/Ollama.app/Contents/Resources/*.dylib dist/Ollama.app/Contents/Resources/imagegen ; 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
@@ -149,7 +178,7 @@ _build_macapp() {
 
     rm -f dist/Ollama-darwin.zip
     ditto -c -k --keepParent dist/Ollama.app dist/Ollama-darwin.zip
-    (cd dist/Ollama.app/Contents/Resources/; tar -cf - ollama *.so *.dylib) | gzip -9vc > dist/ollama-darwin.tgz
+    (cd dist/Ollama.app/Contents/Resources/; tar -cf - ollama imagegen *.so *.dylib) | gzip -9vc > dist/ollama-darwin.tgz
 
     # Notarize and Staple
     if [ -n "$APPLE_IDENTITY" ]; then

scripts/build_linux.sh

@@ -12,6 +12,17 @@ set -eu
 
 . $(dirname $0)/env.sh
 
+# Check for required tools
+if ! command -v zstd >/dev/null 2>&1; then
+    echo "ERROR: zstd is required but not installed." >&2
+    echo "Please install zstd:" >&2
+    echo "  - macOS: brew install zstd" >&2
+    echo "  - Debian/Ubuntu: sudo apt-get install zstd" >&2
+    echo "  - RHEL/CentOS/Fedora: sudo dnf install zstd" >&2
+    echo "  - Arch: sudo pacman -S zstd" >&2
+    exit 1
+fi
+
 mkdir -p dist
 
 docker buildx build \
@@ -37,19 +48,68 @@ if echo $PLATFORM | grep "amd64" > /dev/null; then
         .
 fi
 
+# Deduplicate CUDA libraries across mlx_* and cuda_* directories
+deduplicate_cuda_libs() {
+    local base_dir="$1"
+    echo "Deduplicating CUDA libraries in ${base_dir}..."
+
+    # Find all mlx_cuda_* directories
+    for mlx_dir in "${base_dir}"/lib/ollama/mlx_cuda_*; do
+        [ -d "${mlx_dir}" ] || continue
+
+        # Extract CUDA version (e.g., v12, v13)
+        cuda_version=$(basename "${mlx_dir}" | sed 's/mlx_cuda_//')
+        cuda_dir="${base_dir}/lib/ollama/cuda_${cuda_version}"
+
+        # Skip if corresponding cuda_* directory doesn't exist
+        [ -d "${cuda_dir}" ] || continue
+
+        echo "  Checking ${mlx_dir} against ${cuda_dir}..."
+
+        # Find all .so* files in mlx directory
+        find "${mlx_dir}" -type f -name "*.so*" | while read mlx_file; do
+            filename=$(basename "${mlx_file}")
+            cuda_file="${cuda_dir}/${filename}"
+
+            # Skip if file doesn't exist in cuda directory
+            [ -f "${cuda_file}" ] || continue
+
+            # Compare checksums
+            mlx_sum=$(sha256sum "${mlx_file}" | awk '{print $1}')
+            cuda_sum=$(sha256sum "${cuda_file}" | awk '{print $1}')
+
+            if [ "${mlx_sum}" = "${cuda_sum}" ]; then
+                echo "    Deduplicating ${filename}"
+                # Calculate relative path from mlx_dir to cuda_dir
+                rel_path="../cuda_${cuda_version}/${filename}"
+                rm -f "${mlx_file}"
+                ln -s "${rel_path}" "${mlx_file}"
+            fi
+        done
+    done
+}
+
+# Run deduplication for each platform output directory
+if echo $PLATFORM | grep "," > /dev/null ; then
+    deduplicate_cuda_libs "./dist/linux_amd64"
+    deduplicate_cuda_libs "./dist/linux_arm64"
+elif echo $PLATFORM | grep "amd64\|arm64" > /dev/null ; then
+    deduplicate_cuda_libs "./dist"
+fi
+
 # buildx behavior changes for single vs. multiplatform
 echo "Compressing linux tar bundles..."
 if echo $PLATFORM | grep "," > /dev/null ; then
-        tar c -C ./dist/linux_arm64 --exclude cuda_jetpack5 --exclude cuda_jetpack6 . | pigz -9vc >./dist/ollama-linux-arm64.tgz
-        tar c -C ./dist/linux_arm64 ./lib/ollama/cuda_jetpack5  | pigz -9vc >./dist/ollama-linux-arm64-jetpack5.tgz
-        tar c -C ./dist/linux_arm64 ./lib/ollama/cuda_jetpack6  | pigz -9vc >./dist/ollama-linux-arm64-jetpack6.tgz
-        tar c -C ./dist/linux_amd64 --exclude rocm . | pigz -9vc >./dist/ollama-linux-amd64.tgz
-        tar c -C ./dist/linux_amd64 ./lib/ollama/rocm  | pigz -9vc >./dist/ollama-linux-amd64-rocm.tgz
+        tar c -C ./dist/linux_arm64 --exclude cuda_jetpack5 --exclude cuda_jetpack6 . | zstd --ultra -22 -T0 >./dist/ollama-linux-arm64.tar.zst
+        tar c -C ./dist/linux_arm64 ./lib/ollama/cuda_jetpack5  | zstd --ultra -22 -T0 >./dist/ollama-linux-arm64-jetpack5.tar.zst
+        tar c -C ./dist/linux_arm64 ./lib/ollama/cuda_jetpack6  | zstd --ultra -22 -T0 >./dist/ollama-linux-arm64-jetpack6.tar.zst
+        tar c -C ./dist/linux_amd64 --exclude rocm . | zstd --ultra -22 -T0 >./dist/ollama-linux-amd64.tar.zst
+        tar c -C ./dist/linux_amd64 ./lib/ollama/rocm  | zstd --ultra -22 -T0 >./dist/ollama-linux-amd64-rocm.tar.zst
 elif echo $PLATFORM | grep "arm64" > /dev/null ; then
-        tar c -C ./dist/ --exclude cuda_jetpack5 --exclude cuda_jetpack6 bin lib | pigz -9vc >./dist/ollama-linux-arm64.tgz
-        tar c -C ./dist/ ./lib/ollama/cuda_jetpack5  | pigz -9vc >./dist/ollama-linux-arm64-jetpack5.tgz
-        tar c -C ./dist/ ./lib/ollama/cuda_jetpack6  | pigz -9vc >./dist/ollama-linux-arm64-jetpack6.tgz
+        tar c -C ./dist/ --exclude cuda_jetpack5 --exclude cuda_jetpack6 bin lib | zstd --ultra -22 -T0 >./dist/ollama-linux-arm64.tar.zst
+        tar c -C ./dist/ ./lib/ollama/cuda_jetpack5  | zstd --ultra -22 -T0 >./dist/ollama-linux-arm64-jetpack5.tar.zst
+        tar c -C ./dist/ ./lib/ollama/cuda_jetpack6  | zstd --ultra -22 -T0 >./dist/ollama-linux-arm64-jetpack6.tar.zst
 elif echo $PLATFORM | grep "amd64" > /dev/null ; then
-        tar c -C ./dist/ --exclude rocm bin lib | pigz -9vc >./dist/ollama-linux-amd64.tgz
-        tar c -C ./dist/ ./lib/ollama/rocm  | pigz -9vc >./dist/ollama-linux-amd64-rocm.tgz
+        tar c -C ./dist/ --exclude rocm bin lib | zstd --ultra -22 -T0 >./dist/ollama-linux-amd64.tar.zst
+        tar c -C ./dist/ ./lib/ollama/rocm  | zstd --ultra -22 -T0 >./dist/ollama-linux-amd64-rocm.tar.zst
 fi

scripts/install.sh

@@ -66,6 +66,36 @@ if [ -n "$NEEDS" ]; then
     exit 1
 fi
 
+# Function to download and extract with fallback from zst to tgz
+download_and_extract() {
+    local url_base="$1"
+    local dest_dir="$2"
+    local filename="$3"
+
+    # Check if .tar.zst is available
+    if curl --fail --silent --head --location "${url_base}/${filename}.tar.zst${VER_PARAM}" >/dev/null 2>&1; then
+        # zst file exists - check if we have zstd tool
+        if ! available zstd; then
+            error "This version requires zstd for extraction. Please install zstd and try again:
+  - Debian/Ubuntu: sudo apt-get install zstd
+  - RHEL/CentOS/Fedora: sudo dnf install zstd
+  - Arch: sudo pacman -S zstd"
+        fi
+
+        status "Downloading ${filename}.tar.zst"
+        curl --fail --show-error --location --progress-bar \
+            "${url_base}/${filename}.tar.zst${VER_PARAM}" | \
+            zstd -d | $SUDO tar -xf - -C "${dest_dir}"
+        return 0
+    fi
+
+    # Fall back to .tgz for older versions
+    status "Downloading ${filename}.tgz"
+    curl --fail --show-error --location --progress-bar \
+        "${url_base}/${filename}.tgz${VER_PARAM}" | \
+        $SUDO tar -xzf - -C "${dest_dir}"
+}
+
 for BINDIR in /usr/local/bin /usr/bin /bin; do
     echo $PATH | grep -q $BINDIR && break || continue
 done
@@ -78,10 +108,7 @@ fi
 status "Installing ollama to $OLLAMA_INSTALL_DIR"
 $SUDO install -o0 -g0 -m755 -d $BINDIR
 $SUDO install -o0 -g0 -m755 -d "$OLLAMA_INSTALL_DIR/lib/ollama"
-status "Downloading Linux ${ARCH} bundle"
-curl --fail --show-error --location --progress-bar \
-    "https://ollama.com/download/ollama-linux-${ARCH}.tgz${VER_PARAM}" | \
-    $SUDO tar -xzf - -C "$OLLAMA_INSTALL_DIR"
+download_and_extract "https://ollama.com/download" "$OLLAMA_INSTALL_DIR" "ollama-linux-${ARCH}"
 
 if [ "$OLLAMA_INSTALL_DIR/bin/ollama" != "$BINDIR/ollama" ] ; then
     status "Making ollama accessible in the PATH in $BINDIR"
@@ -91,15 +118,9 @@ fi
 # Check for NVIDIA JetPack systems with additional downloads
 if [ -f /etc/nv_tegra_release ] ; then
     if grep R36 /etc/nv_tegra_release > /dev/null ; then
-        status "Downloading JetPack 6 components"
-        curl --fail --show-error --location --progress-bar \
-            "https://ollama.com/download/ollama-linux-${ARCH}-jetpack6.tgz${VER_PARAM}" | \
-            $SUDO tar -xzf - -C "$OLLAMA_INSTALL_DIR"
+        download_and_extract "https://ollama.com/download" "$OLLAMA_INSTALL_DIR" "ollama-linux-${ARCH}-jetpack6"
     elif grep R35 /etc/nv_tegra_release > /dev/null ; then
-        status "Downloading JetPack 5 components"
-        curl --fail --show-error --location --progress-bar \
-            "https://ollama.com/download/ollama-linux-${ARCH}-jetpack5.tgz${VER_PARAM}" | \
-            $SUDO tar -xzf - -C "$OLLAMA_INSTALL_DIR"
+        download_and_extract "https://ollama.com/download" "$OLLAMA_INSTALL_DIR" "ollama-linux-${ARCH}-jetpack5"
     else
         warning "Unsupported JetPack version detected.  GPU may not be supported"
     fi
@@ -222,10 +243,7 @@ if ! check_gpu lspci nvidia && ! check_gpu lshw nvidia && ! check_gpu lspci amdg
 fi
 
 if check_gpu lspci amdgpu || check_gpu lshw amdgpu; then
-    status "Downloading Linux ROCm ${ARCH} bundle"
-    curl --fail --show-error --location --progress-bar \
-        "https://ollama.com/download/ollama-linux-${ARCH}-rocm.tgz${VER_PARAM}" | \
-        $SUDO tar -xzf - -C "$OLLAMA_INSTALL_DIR"
+    download_and_extract "https://ollama.com/download" "$OLLAMA_INSTALL_DIR" "ollama-linux-${ARCH}-rocm"
 
     install_success
     status "AMD GPU ready."

server/create.go

@@ -26,6 +26,7 @@ import (
 	"github.com/ollama/ollama/convert"
 	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/format"
+	ofs "github.com/ollama/ollama/fs"
 	"github.com/ollama/ollama/fs/ggml"
 	"github.com/ollama/ollama/template"
 	"github.com/ollama/ollama/types/errtypes"
@@ -454,7 +455,7 @@ func convertFromSafetensors(files map[string]string, baseLayers []*layerGGML, is
 	return layers, nil
 }
 
-func kvFromLayers(baseLayers []*layerGGML) (ggml.KV, error) {
+func kvFromLayers(baseLayers []*layerGGML) (ofs.Config, error) {
 	for _, l := range baseLayers {
 		if l.GGML != nil {
 			return l.KV(), nil

server/images.go

@@ -30,6 +30,7 @@ import (
 	"github.com/ollama/ollama/thinking"
 	"github.com/ollama/ollama/types/model"
 	"github.com/ollama/ollama/version"
+	"github.com/ollama/ollama/x/imagegen/transfer"
 )
 
 var (
@@ -73,6 +74,11 @@ type Model struct {
 func (m *Model) Capabilities() []model.Capability {
 	capabilities := []model.Capability{}
 
+	// Check for image generation model via config capabilities
+	if slices.Contains(m.Config.Capabilities, "image") {
+		return []model.Capability{model.CapabilityImageGeneration}
+	}
+
 	// Check for completion capability
 	if m.ModelPath != "" {
 		f, err := gguf.Open(m.ModelPath)
@@ -555,6 +561,24 @@ func PushModel(ctx context.Context, name string, regOpts *registryOptions, fn fu
 		layers = append(layers, manifest.Config)
 	}
 
+	// Use fast transfer for models with tensor layers (many small blobs)
+	if hasTensorLayers(layers) {
+		// Read raw manifest JSON to preserve tensor metadata fields
+		manifestPath, err := mp.GetManifestPath()
+		if err != nil {
+			return err
+		}
+		manifestJSON, err := os.ReadFile(manifestPath)
+		if err != nil {
+			return err
+		}
+		if err := pushWithTransfer(ctx, mp, layers, manifestJSON, regOpts, fn); err != nil {
+			return err
+		}
+		fn(api.ProgressResponse{Status: "success"})
+		return nil
+	}
+
 	for _, layer := range layers {
 		if err := uploadBlob(ctx, mp, layer, regOpts, fn); err != nil {
 			slog.Info(fmt.Sprintf("error uploading blob: %v", err))
@@ -620,6 +644,15 @@ func PullModel(ctx context.Context, name string, regOpts *registryOptions, fn fu
 		layers = append(layers, manifest.Config)
 	}
 
+	// Use fast transfer for models with tensor layers (many small blobs)
+	if hasTensorLayers(layers) {
+		if err := pullWithTransfer(ctx, mp, layers, manifest, regOpts, fn); err != nil {
+			return err
+		}
+		fn(api.ProgressResponse{Status: "success"})
+		return nil
+	}
+
 	skipVerify := make(map[string]bool)
 	for _, layer := range layers {
 		cacheHit, err := downloadBlob(ctx, downloadOpts{
@@ -634,7 +667,6 @@ func PullModel(ctx context.Context, name string, regOpts *registryOptions, fn fu
 		skipVerify[layer.Digest] = cacheHit
 		delete(deleteMap, layer.Digest)
 	}
-	delete(deleteMap, manifest.Config.Digest)
 
 	fn(api.ProgressResponse{Status: "verifying sha256 digest"})
 	for _, layer := range layers {
@@ -643,13 +675,11 @@ func PullModel(ctx context.Context, name string, regOpts *registryOptions, fn fu
 		}
 		if err := verifyBlob(layer.Digest); err != nil {
 			if errors.Is(err, errDigestMismatch) {
-				// something went wrong, delete the blob
 				fp, err := GetBlobsPath(layer.Digest)
 				if err != nil {
 					return err
 				}
 				if err := os.Remove(fp); err != nil {
-					// log this, but return the original error
 					slog.Info(fmt.Sprintf("couldn't remove file with digest mismatch '%s': %v", fp, err))
 				}
 			}
@@ -657,6 +687,11 @@ func PullModel(ctx context.Context, name string, regOpts *registryOptions, fn fu
 		}
 	}
 
+	for _, layer := range layers {
+		delete(deleteMap, layer.Digest)
+	}
+	delete(deleteMap, manifest.Config.Digest)
+
 	fn(api.ProgressResponse{Status: "writing manifest"})
 
 	manifestJSON, err := json.Marshal(manifest)
@@ -690,6 +725,148 @@ func PullModel(ctx context.Context, name string, regOpts *registryOptions, fn fu
 	return nil
 }
 
+// hasTensorLayers checks if any layer has tensor media type.
+func hasTensorLayers(layers []Layer) bool {
+	for _, layer := range layers {
+		if layer.MediaType == MediaTypeImageTensor {
+			return true
+		}
+	}
+	return false
+}
+
+// pullWithTransfer uses the simplified x/transfer package for downloading blobs.
+func pullWithTransfer(ctx context.Context, mp ModelPath, layers []Layer, manifest *Manifest, regOpts *registryOptions, fn func(api.ProgressResponse)) error {
+	blobs := make([]transfer.Blob, len(layers))
+	for i, layer := range layers {
+		blobs[i] = transfer.Blob{
+			Digest: layer.Digest,
+			Size:   layer.Size,
+		}
+	}
+
+	destDir, err := GetBlobsPath("")
+	if err != nil {
+		return err
+	}
+
+	base := mp.BaseURL()
+	if base.Scheme != "http" && regOpts != nil && regOpts.Insecure {
+		base.Scheme = "http"
+	}
+	baseURL := base.String()
+
+	var totalSize int64
+	for _, blob := range blobs {
+		totalSize += blob.Size
+	}
+
+	progress := func(completed, total int64) {
+		fn(api.ProgressResponse{
+			Status:    "pulling model",
+			Digest:    "sha256:model",
+			Total:     total,
+			Completed: completed,
+		})
+	}
+
+	getToken := func(ctx context.Context, challenge transfer.AuthChallenge) (string, error) {
+		return getAuthorizationToken(ctx, registryChallenge{
+			Realm:   challenge.Realm,
+			Service: challenge.Service,
+			Scope:   challenge.Scope,
+		})
+	}
+
+	if err := transfer.Download(ctx, transfer.DownloadOptions{
+		Blobs:      blobs,
+		BaseURL:    baseURL,
+		DestDir:    destDir,
+		Repository: mp.GetNamespaceRepository(),
+		Progress:   progress,
+		Token:      regOpts.Token,
+		GetToken:   getToken,
+		Logger:     slog.Default(),
+	}); err != nil {
+		return err
+	}
+
+	// Write manifest
+	fn(api.ProgressResponse{Status: "writing manifest"})
+	manifestJSON, err := json.Marshal(manifest)
+	if err != nil {
+		return err
+	}
+
+	fp, err := mp.GetManifestPath()
+	if err != nil {
+		return err
+	}
+	if err := os.MkdirAll(filepath.Dir(fp), 0o755); err != nil {
+		return err
+	}
+
+	return os.WriteFile(fp, manifestJSON, 0o644)
+}
+
+// pushWithTransfer uses the simplified x/transfer package for uploading blobs and manifest.
+func pushWithTransfer(ctx context.Context, mp ModelPath, layers []Layer, manifestJSON []byte, regOpts *registryOptions, fn func(api.ProgressResponse)) error {
+	blobs := make([]transfer.Blob, len(layers))
+	for i, layer := range layers {
+		blobs[i] = transfer.Blob{
+			Digest: layer.Digest,
+			Size:   layer.Size,
+			From:   layer.From,
+		}
+	}
+
+	srcDir, err := GetBlobsPath("")
+	if err != nil {
+		return err
+	}
+
+	base := mp.BaseURL()
+	if base.Scheme != "http" && regOpts != nil && regOpts.Insecure {
+		base.Scheme = "http"
+	}
+	baseURL := base.String()
+
+	var totalSize int64
+	for _, blob := range blobs {
+		totalSize += blob.Size
+	}
+
+	progress := func(completed, total int64) {
+		fn(api.ProgressResponse{
+			Status:    "pushing model",
+			Digest:    "sha256:model",
+			Total:     total,
+			Completed: completed,
+		})
+	}
+
+	getToken := func(ctx context.Context, challenge transfer.AuthChallenge) (string, error) {
+		return getAuthorizationToken(ctx, registryChallenge{
+			Realm:   challenge.Realm,
+			Service: challenge.Service,
+			Scope:   challenge.Scope,
+		})
+	}
+
+	return transfer.Upload(ctx, transfer.UploadOptions{
+		Blobs:       blobs,
+		BaseURL:     baseURL,
+		SrcDir:      srcDir,
+		Progress:    progress,
+		Token:       regOpts.Token,
+		GetToken:    getToken,
+		Logger:      slog.Default(),
+		Manifest:    manifestJSON,
+		ManifestRef: mp.Tag,
+		Repository:  mp.GetNamespaceRepository(),
+	})
+}
+
 func pullModelManifest(ctx context.Context, mp ModelPath, regOpts *registryOptions) (*Manifest, error) {
 	requestURL := mp.BaseURL().JoinPath("v2", mp.GetNamespaceRepository(), "manifests", mp.Tag)
 

server/images_test.go

@@ -47,6 +47,15 @@ func TestModelCapabilities(t *testing.T) {
 		model        Model
 		expectedCaps []model.Capability
 	}{
+		{
+			name: "model with image generation capability via config",
+			model: Model{
+				Config: model.ConfigV2{
+					Capabilities: []string{"image"},
+				},
+			},
+			expectedCaps: []model.Capability{model.CapabilityImageGeneration},
+		},
 		{
 			name: "model with completion capability",
 			model: Model{

server/layer.go

@@ -13,9 +13,14 @@ type Layer struct {
 	Digest    string `json:"digest"`
 	Size      int64  `json:"size"`
 	From      string `json:"from,omitempty"`
+	Name      string `json:"name,omitempty"` // tensor name, e.g., "text_encoder/model.embed_tokens.weight"
 	status    string
 }
 
+const (
+	MediaTypeImageTensor = "application/vnd.ollama.image.tensor"
+)
+
 func NewLayer(r io.Reader, mediatype string) (Layer, error) {
 	blobs, err := GetBlobsPath("")
 	if err != nil {

server/routes.go

@@ -50,6 +50,8 @@ import (
 	"github.com/ollama/ollama/types/errtypes"
 	"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"
 )
 
 const signinURLStr = "https://ollama.com/connect?name=%s&key=%s"
@@ -162,6 +164,29 @@ 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 {
@@ -189,6 +214,12 @@ 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
@@ -1093,6 +1124,15 @@ func GetModelInfo(req api.ShowRequest) (*api.ShowResponse, error) {
 		QuantizationLevel: m.Config.FileType,
 	}
 
+	// For image generation models, populate details from imagegen package
+	if slices.Contains(m.Capabilities(), model.CapabilityImageGeneration) {
+		if info, err := imagegen.GetModelInfo(name.String()); err == nil {
+			modelDetails.Family = info.Architecture
+			modelDetails.ParameterSize = format.HumanNumber(uint64(info.ParameterCount))
+			modelDetails.QuantizationLevel = info.Quantization
+		}
+	}
+
 	if req.System != "" {
 		m.System = req.System
 	}
@@ -1175,6 +1215,10 @@ func GetModelInfo(req api.ShowRequest) (*api.ShowResponse, error) {
 		return resp, nil
 	}
 
+	if slices.Contains(m.Capabilities(), model.CapabilityImageGeneration) {
+		return resp, nil
+	}
+
 	kvData, tensors, err := getModelData(m.ModelPath, req.Verbose)
 	if err != nil {
 		return nil, err
@@ -1544,6 +1588,12 @@ func (s *Server) GenerateRoutes(rc *ollama.Registry) (http.Handler, error) {
 	r.GET("/v1/models/:model", middleware.RetrieveMiddleware(), s.ShowHandler)
 	r.POST("/v1/responses", middleware.ResponsesMiddleware(), s.ChatHandler)
 
+	// 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{

server/routes_create_test.go

@@ -22,6 +22,7 @@ import (
 	gocmpopts "github.com/google/go-cmp/cmp/cmpopts"
 
 	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/convert"
 	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/fs/ggml"
 	"github.com/ollama/ollama/types/model"
@@ -41,7 +42,7 @@ func createBinFile(t *testing.T, kv map[string]any, ti []*ggml.Tensor) (string,
 	}
 	defer f.Close()
 
-	base := map[string]any{"general.architecture": "test"}
+	var base convert.KV = map[string]any{"general.architecture": "test"}
 	maps.Copy(base, kv)
 
 	if err := ggml.WriteGGUF(f, base, ti); err != nil {

server/sched.go

@@ -21,6 +21,7 @@ import (
 	"github.com/ollama/ollama/logutil"
 	"github.com/ollama/ollama/ml"
 	"github.com/ollama/ollama/types/model"
+	"github.com/ollama/ollama/x/imagegen"
 )
 
 type LlmRequest struct {
@@ -194,6 +195,14 @@ func (s *Scheduler) processPending(ctx context.Context) {
 						slog.Debug("updating default concurrency", "OLLAMA_MAX_LOADED_MODELS", maxRunners, "gpu_count", len(gpus))
 					}
 
+					// Check for image generation model before attempting GGML load
+					if slices.Contains(pending.model.Config.Capabilities, "image") {
+						if s.loadImageGen(pending) {
+							break
+						}
+						continue
+					}
+
 					// Load model for fitting
 					logutil.Trace("loading model metadata", "model", pending.model.ModelPath)
 					ggml, err := llm.LoadModel(pending.model.ModelPath, 1024)
@@ -543,6 +552,48 @@ iGPUScan:
 	return false
 }
 
+// loadImageGen loads an image generation model.
+func (s *Scheduler) loadImageGen(req *LlmRequest) bool {
+	// Use model name for imagegen (it resolves manifests by name, not file path)
+	modelName := req.model.ShortName
+	server, err := imagegen.NewServer(modelName)
+	if err != nil {
+		req.errCh <- err
+		return true
+	}
+
+	sessionDuration := envconfig.KeepAlive()
+	if req.sessionDuration != nil {
+		sessionDuration = req.sessionDuration.Duration
+	}
+
+	runner := &runnerRef{
+		model:           req.model,
+		modelPath:       req.model.ModelPath,
+		llama:           server,
+		Options:         &req.opts,
+		loading:         false,
+		sessionDuration: sessionDuration,
+		refCount:        1,
+	}
+
+	s.loadedMu.Lock()
+	s.loaded[req.model.ModelPath] = runner
+	s.loadedMu.Unlock()
+
+	// Set up expiration timer
+	runner.refMu.Lock()
+	if sessionDuration > 0 {
+		runner.expireTimer = time.AfterFunc(sessionDuration, func() {
+			s.expiredCh <- runner
+		})
+	}
+	runner.refMu.Unlock()
+
+	req.useLoadedRunner(runner, s.finishedReqCh)
+	return true
+}
+
 func (s *Scheduler) updateFreeSpace(allGpus []ml.DeviceInfo) {
 	if len(allGpus) == 0 {
 		return

server/sched_test.go

@@ -6,6 +6,7 @@ import (
 	"errors"
 	"log/slog"
 	"os"
+	"slices"
 	"testing"
 	"time"
 
@@ -16,6 +17,7 @@ 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) {
@@ -804,3 +806,61 @@ func (s *mockLlm) GetPort() int                                       { return -
 func (s *mockLlm) GetDeviceInfos(ctx context.Context) []ml.DeviceInfo { return nil }
 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.
+func TestImageGenRunnerCanBeEvicted(t *testing.T) {
+	ctx, done := context.WithTimeout(t.Context(), 500*time.Millisecond)
+	defer done()
+
+	s := InitScheduler(ctx)
+	s.getGpuFn = getGpuFn
+	s.getSystemInfoFn = getSystemInfoFn
+
+	// Simulate an image gen runner already loaded
+	imageGenRunner := &runnerRef{
+		model:           &Model{Name: "z-image", ModelPath: "/fake/image/model"},
+		modelPath:       "/fake/image/model",
+		llama:           &mockLlm{vramSize: 21 * format.GigaByte, vramByGPU: map[ml.DeviceID]uint64{}},
+		sessionDuration: 5 * time.Millisecond,
+		refCount:        0, // idle
+	}
+
+	s.loadedMu.Lock()
+	s.loaded["/fake/image/model"] = imageGenRunner
+	s.loadedMu.Unlock()
+
+	// Verify the image gen runner is loaded
+	s.loadedMu.Lock()
+	require.Len(t, s.loaded, 1)
+	s.loadedMu.Unlock()
+
+	// findRunnerToUnload should find the idle image gen runner
+	runner := s.findRunnerToUnload()
+	require.NotNil(t, runner)
+	require.Equal(t, "/fake/image/model", runner.modelPath)
+}

template/template.go

@@ -381,6 +381,28 @@ func (t templateTools) String() string {
 	return string(bts)
 }
 
+// templateArgs is a map type with JSON string output for templates.
+type templateArgs map[string]any
+
+func (t templateArgs) String() string {
+	if t == nil {
+		return "{}"
+	}
+	bts, _ := json.Marshal(t)
+	return string(bts)
+}
+
+// templateProperties is a map type with JSON string output for templates.
+type templateProperties map[string]api.ToolProperty
+
+func (t templateProperties) String() string {
+	if t == nil {
+		return "{}"
+	}
+	bts, _ := json.Marshal(t)
+	return string(bts)
+}
+
 // templateTool is a template-compatible representation of api.Tool
 // with Properties as a regular map for template ranging.
 type templateTool struct {
@@ -396,11 +418,11 @@ type templateToolFunction struct {
 }
 
 type templateToolFunctionParameters struct {
-	Type       string                      `json:"type"`
-	Defs       any                         `json:"$defs,omitempty"`
-	Items      any                         `json:"items,omitempty"`
-	Required   []string                    `json:"required,omitempty"`
-	Properties map[string]api.ToolProperty `json:"properties"`
+	Type       string             `json:"type"`
+	Defs       any                `json:"$defs,omitempty"`
+	Items      any                `json:"items,omitempty"`
+	Required   []string           `json:"required,omitempty"`
+	Properties templateProperties `json:"properties"`
 }
 
 // templateToolCall is a template-compatible representation of api.ToolCall
@@ -413,7 +435,7 @@ type templateToolCall struct {
 type templateToolCallFunction struct {
 	Index     int
 	Name      string
-	Arguments map[string]any
+	Arguments templateArgs
 }
 
 // templateMessage is a template-compatible representation of api.Message
@@ -446,7 +468,7 @@ func convertToolsForTemplate(tools api.Tools) templateTools {
 					Defs:       tool.Function.Parameters.Defs,
 					Items:      tool.Function.Parameters.Items,
 					Required:   tool.Function.Parameters.Required,
-					Properties: tool.Function.Parameters.Properties.ToMap(),
+					Properties: templateProperties(tool.Function.Parameters.Properties.ToMap()),
 				},
 			},
 		}
@@ -468,7 +490,7 @@ func convertMessagesForTemplate(messages []*api.Message) []*templateMessage {
 				Function: templateToolCallFunction{
 					Index:     tc.Function.Index,
 					Name:      tc.Function.Name,
-					Arguments: tc.Function.Arguments.ToMap(),
+					Arguments: templateArgs(tc.Function.Arguments.ToMap()),
 				},
 			})
 		}

template/template_test.go

@@ -613,3 +613,159 @@ func TestCollate(t *testing.T) {
 		})
 	}
 }
+
+func TestTemplateArgumentsJSON(t *testing.T) {
+	// Test that {{ .Function.Arguments }} outputs valid JSON, not map[key:value]
+	tmpl := `{{- range .Messages }}{{- range .ToolCalls }}{{ .Function.Arguments }}{{- end }}{{- end }}`
+
+	template, err := Parse(tmpl)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	args := api.NewToolCallFunctionArguments()
+	args.Set("location", "Tokyo")
+	args.Set("unit", "celsius")
+
+	var buf bytes.Buffer
+	err = template.Execute(&buf, Values{
+		Messages: []api.Message{{
+			Role: "assistant",
+			ToolCalls: []api.ToolCall{{
+				Function: api.ToolCallFunction{
+					Name:      "get_weather",
+					Arguments: args,
+				},
+			}},
+		}},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	got := buf.String()
+	// Should be valid JSON, not "map[location:Tokyo unit:celsius]"
+	if strings.HasPrefix(got, "map[") {
+		t.Errorf("Arguments output as Go map format: %s", got)
+	}
+
+	var parsed map[string]any
+	if err := json.Unmarshal([]byte(got), &parsed); err != nil {
+		t.Errorf("Arguments not valid JSON: %s, error: %v", got, err)
+	}
+}
+
+func TestTemplatePropertiesJSON(t *testing.T) {
+	// Test that {{ .Function.Parameters.Properties }} outputs valid JSON
+	// Note: template must reference .Messages to trigger the modern code path that converts Tools
+	tmpl := `{{- range .Messages }}{{- end }}{{- range .Tools }}{{ .Function.Parameters.Properties }}{{- end }}`
+
+	template, err := Parse(tmpl)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	props := api.NewToolPropertiesMap()
+	props.Set("location", api.ToolProperty{Type: api.PropertyType{"string"}, Description: "City name"})
+
+	var buf bytes.Buffer
+	err = template.Execute(&buf, Values{
+		Messages: []api.Message{{Role: "user", Content: "test"}},
+		Tools: api.Tools{{
+			Type: "function",
+			Function: api.ToolFunction{
+				Name:        "get_weather",
+				Description: "Get weather",
+				Parameters: api.ToolFunctionParameters{
+					Type:       "object",
+					Properties: props,
+				},
+			},
+		}},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	got := buf.String()
+	// Should be valid JSON, not "map[location:{...}]"
+	if strings.HasPrefix(got, "map[") {
+		t.Errorf("Properties output as Go map format: %s", got)
+	}
+
+	var parsed map[string]any
+	if err := json.Unmarshal([]byte(got), &parsed); err != nil {
+		t.Errorf("Properties not valid JSON: %s, error: %v", got, err)
+	}
+}
+
+func TestTemplateArgumentsRange(t *testing.T) {
+	// Test that we can range over Arguments in templates
+	tmpl := `{{- range .Messages }}{{- range .ToolCalls }}{{- range $k, $v := .Function.Arguments }}{{ $k }}={{ $v }};{{- end }}{{- end }}{{- end }}`
+
+	template, err := Parse(tmpl)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	args := api.NewToolCallFunctionArguments()
+	args.Set("city", "Tokyo")
+
+	var buf bytes.Buffer
+	err = template.Execute(&buf, Values{
+		Messages: []api.Message{{
+			Role: "assistant",
+			ToolCalls: []api.ToolCall{{
+				Function: api.ToolCallFunction{
+					Name:      "get_weather",
+					Arguments: args,
+				},
+			}},
+		}},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	got := buf.String()
+	if got != "city=Tokyo;" {
+		t.Errorf("Range over Arguments failed, got: %s, want: city=Tokyo;", got)
+	}
+}
+
+func TestTemplatePropertiesRange(t *testing.T) {
+	// Test that we can range over Properties in templates
+	// Note: template must reference .Messages to trigger the modern code path that converts Tools
+	tmpl := `{{- range .Messages }}{{- end }}{{- range .Tools }}{{- range $name, $prop := .Function.Parameters.Properties }}{{ $name }}:{{ $prop.Type }};{{- end }}{{- end }}`
+
+	template, err := Parse(tmpl)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	props := api.NewToolPropertiesMap()
+	props.Set("location", api.ToolProperty{Type: api.PropertyType{"string"}})
+
+	var buf bytes.Buffer
+	err = template.Execute(&buf, Values{
+		Messages: []api.Message{{Role: "user", Content: "test"}},
+		Tools: api.Tools{{
+			Type: "function",
+			Function: api.ToolFunction{
+				Name: "get_weather",
+				Parameters: api.ToolFunctionParameters{
+					Type:       "object",
+					Properties: props,
+				},
+			},
+		}},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	got := buf.String()
+	if got != "location:string;" {
+		t.Errorf("Range over Properties failed, got: %s, want: location:string;", got)
+	}
+}

types/model/capability.go

@@ -3,12 +3,13 @@ package model
 type Capability string
 
 const (
-	CapabilityCompletion = Capability("completion")
-	CapabilityTools      = Capability("tools")
-	CapabilityInsert     = Capability("insert")
-	CapabilityVision     = Capability("vision")
-	CapabilityEmbedding  = Capability("embedding")
-	CapabilityThinking   = Capability("thinking")
+	CapabilityCompletion      = Capability("completion")
+	CapabilityTools           = Capability("tools")
+	CapabilityInsert          = Capability("insert")
+	CapabilityVision          = Capability("vision")
+	CapabilityEmbedding       = Capability("embedding")
+	CapabilityThinking        = Capability("thinking")
+	CapabilityImageGeneration = Capability("image")
 )
 
 func (c Capability) String() string {

x/README.md

@@ -0,0 +1,24 @@
+# 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.  To build:
+
+```
+cmake --preset MLX
+cmake --build --preset MLX --parallel
+cmake --install build --component MLX
+go build -tags mlx .
+```
+
+On linux, use the preset "MLX CUDA 13" or "MLX CUDA 12" to enable CUDA with the default Ollama NVIDIA GPU architectures enabled. 
+
+## Image Generation
+
+Based on the experimental MLX backend, we're working on adding imagegen support.  After running the cmake commands above:
+
+```
+go build -o imagegen ./x/imagegen/cmd/engine
+```

x/agent/approval.go

@@ -4,6 +4,7 @@ package agent
 import (
 	"fmt"
 	"os"
+	"path"
 	"path/filepath"
 	"strings"
 	"sync"
@@ -32,10 +33,29 @@ type ApprovalResult struct {
 // Option labels for the selector (numbered for quick selection)
 var optionLabels = []string{
 	"1. Execute once",
-	"2. Always allow",
+	"2. Allow for this session",
 	"3. Deny",
 }
 
+// toolDisplayNames maps internal tool names to human-readable display names.
+var toolDisplayNames = map[string]string{
+	"bash":       "Bash",
+	"web_search": "Web Search",
+}
+
+// ToolDisplayName returns the human-readable display name for a tool.
+func ToolDisplayName(toolName string) string {
+	if displayName, ok := toolDisplayNames[toolName]; ok {
+		return displayName
+	}
+	// Default: capitalize first letter and replace underscores with spaces
+	name := strings.ReplaceAll(toolName, "_", " ")
+	if len(name) > 0 {
+		return strings.ToUpper(name[:1]) + name[1:]
+	}
+	return toolName
+}
+
 // autoAllowCommands are commands that are always allowed without prompting.
 // These are zero-risk, read-only commands.
 var autoAllowCommands = map[string]bool{
@@ -179,6 +199,7 @@ func FormatDeniedResult(command string, pattern string) string {
 // extractBashPrefix extracts a prefix pattern from a bash command.
 // For commands like "cat tools/tools_test.go | head -200", returns "cat:tools/"
 // For commands without path args, returns empty string.
+// Paths with ".." traversal that escape the base directory return empty string for security.
 func extractBashPrefix(command string) string {
 	// Split command by pipes and get the first part
 	parts := strings.Split(command, "|")
@@ -204,8 +225,8 @@ func extractBashPrefix(command string) string {
 		return ""
 	}
 
-	// Find the first path-like argument (must contain / or start with .)
-	// First pass: look for clear paths (containing / or starting with .)
+	// Find the first path-like argument (must contain / or \ or start with .)
+	// First pass: look for clear paths (containing path separators or starting with .)
 	for _, arg := range fields[1:] {
 		// Skip flags
 		if strings.HasPrefix(arg, "-") {
@@ -215,19 +236,49 @@ func extractBashPrefix(command string) string {
 		if isNumeric(arg) {
 			continue
 		}
-		// Only process if it looks like a path (contains / or starts with .)
-		if !strings.Contains(arg, "/") && !strings.HasPrefix(arg, ".") {
+		// Only process if it looks like a path (contains / or \ or starts with .)
+		if !strings.Contains(arg, "/") && !strings.Contains(arg, "\\") && !strings.HasPrefix(arg, ".") {
 			continue
 		}
-		// If arg ends with /, it's a directory - use it directly
-		if strings.HasSuffix(arg, "/") {
-			return fmt.Sprintf("%s:%s", baseCmd, arg)
+		// Normalize to forward slashes for consistent cross-platform matching
+		arg = strings.ReplaceAll(arg, "\\", "/")
+
+		// Security: reject absolute paths
+		if path.IsAbs(arg) {
+			return "" // Absolute path - don't create prefix
 		}
-		// Get the directory part of a file path
-		dir := filepath.Dir(arg)
+
+		// Normalize the path using stdlib path.Clean (resolves . and ..)
+		cleaned := path.Clean(arg)
+
+		// Security: reject if cleaned path escapes to parent directory
+		if strings.HasPrefix(cleaned, "..") {
+			return "" // Path escapes - don't create prefix
+		}
+
+		// Security: if original had "..", verify cleaned path didn't escape to sibling
+		// e.g., "tools/a/b/../../../etc" -> "etc" (escaped tools/ to sibling)
+		if strings.Contains(arg, "..") {
+			origBase := strings.SplitN(arg, "/", 2)[0]
+			cleanedBase := strings.SplitN(cleaned, "/", 2)[0]
+			if origBase != cleanedBase {
+				return "" // Path escaped to sibling directory
+			}
+		}
+
+		// Check if arg ends with / (explicit directory)
+		isDir := strings.HasSuffix(arg, "/")
+
+		// Get the directory part
+		var dir string
+		if isDir {
+			dir = cleaned
+		} else {
+			dir = path.Dir(cleaned)
+		}
+
 		if dir == "." {
-			// Path is just a directory like "tools" or "src" (no trailing /)
-			return fmt.Sprintf("%s:%s/", baseCmd, arg)
+			return fmt.Sprintf("%s:./", baseCmd)
 		}
 		return fmt.Sprintf("%s:%s/", baseCmd, dir)
 	}
@@ -332,6 +383,8 @@ func AllowlistKey(toolName string, args map[string]any) string {
 }
 
 // IsAllowed checks if a tool/command is allowed (exact match or prefix match).
+// For bash commands, hierarchical path matching is used - if "cat:tools/" is allowed,
+// then "cat:tools/subdir/" is also allowed (subdirectories inherit parent permissions).
 func (a *ApprovalManager) IsAllowed(toolName string, args map[string]any) bool {
 	a.mu.RLock()
 	defer a.mu.RUnlock()
@@ -342,12 +395,20 @@ func (a *ApprovalManager) IsAllowed(toolName string, args map[string]any) bool {
 		return true
 	}
 
-	// For bash commands, check prefix matches
+	// For bash commands, check prefix matches with hierarchical path support
 	if toolName == "bash" {
 		if cmd, ok := args["command"].(string); ok {
 			prefix := extractBashPrefix(cmd)
-			if prefix != "" && a.prefixes[prefix] {
-				return true
+			if prefix != "" {
+				// Check exact prefix match first
+				if a.prefixes[prefix] {
+					return true
+				}
+				// Check hierarchical match: if any stored prefix is a parent of current prefix
+				// e.g., stored "cat:tools/" should match current "cat:tools/subdir/"
+				if a.matchesHierarchicalPrefix(prefix) {
+					return true
+				}
 			}
 		}
 	}
@@ -360,6 +421,40 @@ func (a *ApprovalManager) IsAllowed(toolName string, args map[string]any) bool {
 	return false
 }
 
+// matchesHierarchicalPrefix checks if the given prefix matches any stored prefix hierarchically.
+// For example, if "cat:tools/" is stored, it will match "cat:tools/subdir/" or "cat:tools/a/b/c/".
+func (a *ApprovalManager) matchesHierarchicalPrefix(currentPrefix string) bool {
+	// Split prefix into command and path parts (format: "cmd:path/")
+	colonIdx := strings.Index(currentPrefix, ":")
+	if colonIdx == -1 {
+		return false
+	}
+	currentCmd := currentPrefix[:colonIdx]
+	currentPath := currentPrefix[colonIdx+1:]
+
+	for storedPrefix := range a.prefixes {
+		storedColonIdx := strings.Index(storedPrefix, ":")
+		if storedColonIdx == -1 {
+			continue
+		}
+		storedCmd := storedPrefix[:storedColonIdx]
+		storedPath := storedPrefix[storedColonIdx+1:]
+
+		// Commands must match exactly
+		if currentCmd != storedCmd {
+			continue
+		}
+
+		// Check if current path starts with stored path (hierarchical match)
+		// e.g., "tools/subdir/" starts with "tools/"
+		if strings.HasPrefix(currentPath, storedPath) {
+			return true
+		}
+	}
+
+	return false
+}
+
 // AddToAllowlist adds a tool/command to the session allowlist.
 // For bash commands, it adds the prefix pattern instead of exact command.
 func (a *ApprovalManager) AddToAllowlist(toolName string, args map[string]any) {
@@ -399,16 +494,32 @@ func (a *ApprovalManager) RequestApproval(toolName string, args map[string]any)
 	// This prevents buffered input from causing double-press issues
 	flushStdin(fd)
 
-	// Check if bash command targets paths outside cwd
 	isWarning := false
+	var warningMsg string
+	var allowlistInfo string
 	if toolName == "bash" {
 		if cmd, ok := args["command"].(string); ok {
-			isWarning = isCommandOutsideCwd(cmd)
+			if isCommandOutsideCwd(cmd) {
+				isWarning = true
+				warningMsg = "command targets paths outside project"
+			}
+			if prefix := extractBashPrefix(cmd); prefix != "" {
+				colonIdx := strings.Index(prefix, ":")
+				if colonIdx != -1 {
+					cmdName := prefix[:colonIdx]
+					dirPath := prefix[colonIdx+1:]
+					if dirPath != "./" {
+						allowlistInfo = fmt.Sprintf("%s in %s directory (includes subdirs)", cmdName, dirPath)
+					} else {
+						allowlistInfo = fmt.Sprintf("%s in %s directory", cmdName, dirPath)
+					}
+				}
+			}
 		}
 	}
 
 	// Run interactive selector
-	selected, denyReason, err := runSelector(fd, oldState, toolDisplay, isWarning)
+	selected, denyReason, err := runSelector(fd, oldState, toolDisplay, isWarning, warningMsg, allowlistInfo)
 	if err != nil {
 		term.Restore(fd, oldState)
 		return ApprovalResult{Decision: ApprovalDeny}, err
@@ -433,27 +544,29 @@ func (a *ApprovalManager) RequestApproval(toolName string, args map[string]any)
 // formatToolDisplay creates the display string for a tool call.
 func formatToolDisplay(toolName string, args map[string]any) string {
 	var sb strings.Builder
+	displayName := ToolDisplayName(toolName)
 
 	// For bash, show command directly
 	if toolName == "bash" {
 		if cmd, ok := args["command"].(string); ok {
-			sb.WriteString(fmt.Sprintf("Tool: %s\n", toolName))
+			sb.WriteString(fmt.Sprintf("Tool: %s\n", displayName))
 			sb.WriteString(fmt.Sprintf("Command: %s", cmd))
 			return sb.String()
 		}
 	}
 
-	// For web search, show query
+	// For web search, show query and internet notice
 	if toolName == "web_search" {
 		if query, ok := args["query"].(string); ok {
-			sb.WriteString(fmt.Sprintf("Tool: %s\n", toolName))
-			sb.WriteString(fmt.Sprintf("Query: %s", query))
+			sb.WriteString(fmt.Sprintf("Tool: %s\n", displayName))
+			sb.WriteString(fmt.Sprintf("Query: %s\n", query))
+			sb.WriteString("Uses internet via ollama.com")
 			return sb.String()
 		}
 	}
 
 	// Generic display
-	sb.WriteString(fmt.Sprintf("Tool: %s", toolName))
+	sb.WriteString(fmt.Sprintf("Tool: %s", displayName))
 	if len(args) > 0 {
 		sb.WriteString("\nArguments: ")
 		first := true
@@ -470,24 +583,28 @@ func formatToolDisplay(toolName string, args map[string]any) string {
 
 // selectorState holds the state for the interactive selector
 type selectorState struct {
-	toolDisplay string
-	selected    int
-	totalLines  int
-	termWidth   int
-	termHeight  int
-	boxWidth    int
-	innerWidth  int
-	denyReason  string // deny reason (always visible in box)
-	isWarning   bool   // true if command targets paths outside cwd (red box)
+	toolDisplay    string
+	selected       int
+	totalLines     int
+	termWidth      int
+	termHeight     int
+	boxWidth       int
+	innerWidth     int
+	denyReason     string // deny reason (always visible in box)
+	isWarning      bool   // true if command has warning
+	warningMessage string // dynamic warning message to display
+	allowlistInfo  string // show what will be allowlisted (for "Allow for this session" option)
 }
 
 // runSelector runs the interactive selector and returns the selected index and optional deny reason.
 // If isWarning is true, the box is rendered in red to indicate the command targets paths outside cwd.
-func runSelector(fd int, oldState *term.State, toolDisplay string, isWarning bool) (int, string, error) {
+func runSelector(fd int, oldState *term.State, toolDisplay string, isWarning bool, warningMessage string, allowlistInfo string) (int, string, error) {
 	state := &selectorState{
-		toolDisplay: toolDisplay,
-		selected:    0,
-		isWarning:   isWarning,
+		toolDisplay:    toolDisplay,
+		selected:       0,
+		isWarning:      isWarning,
+		warningMessage: warningMessage,
+		allowlistInfo:  allowlistInfo,
 	}
 
 	// Get terminal size
@@ -647,7 +764,7 @@ func wrapText(text string, maxWidth int) []string {
 
 // getHintLines returns the hint text wrapped to terminal width
 func getHintLines(state *selectorState) []string {
-	hint := "↑/↓ navigate, Enter confirm, 1-3 quick, Ctrl+C cancel"
+	hint := "up/down select, enter confirm, 1-3 quick select, ctrl+c cancel"
 	if state.termWidth >= len(hint)+1 {
 		return []string{hint}
 	}
@@ -657,86 +774,70 @@ func getHintLines(state *selectorState) []string {
 
 // calculateTotalLines calculates how many lines the selector will use
 func calculateTotalLines(state *selectorState) int {
-	toolLines := wrapText(state.toolDisplay, state.innerWidth)
+	toolLines := strings.Split(state.toolDisplay, "\n")
 	hintLines := getHintLines(state)
-	// top border + (warning line if applicable) + tool lines + separator + options + bottom border + hint lines
+	// warning line (if applicable) + tool lines + blank line + options + blank line + hint lines
 	warningLines := 0
 	if state.isWarning {
-		warningLines = 1
+		warningLines = 2 // warning line + blank line after
 	}
-	return 1 + warningLines + len(toolLines) + 1 + len(optionLabels) + 1 + len(hintLines)
+	return warningLines + len(toolLines) + 1 + len(optionLabels) + 1 + len(hintLines)
 }
 
-// renderSelectorBox renders the complete selector box
+// renderSelectorBox renders the selector (minimal, no box)
 func renderSelectorBox(state *selectorState) {
-	toolLines := wrapText(state.toolDisplay, state.innerWidth)
+	toolLines := strings.Split(state.toolDisplay, "\n")
 	hintLines := getHintLines(state)
 
-	// Use red for warning (outside cwd), cyan for normal
-	boxColor := "\033[36m" // cyan
+	// Draw warning line if needed
 	if state.isWarning {
-		boxColor = "\033[91m" // bright red
-	}
-
-	// Draw box top
-	fmt.Fprintf(os.Stderr, "%s┌%s┐\033[0m\033[K\r\n", boxColor, strings.Repeat("─", state.boxWidth-2))
-
-	// Draw warning line if needed (inside the box)
-	if state.isWarning {
-		warning := "!! OUTSIDE PROJECT !!"
-		padding := (state.innerWidth - len(warning)) / 2
-		if padding < 0 {
-			padding = 0
+		if state.warningMessage != "" {
+			fmt.Fprintf(os.Stderr, "\033[1mwarning:\033[0m %s\033[K\r\n", state.warningMessage)
+		} else {
+			fmt.Fprintf(os.Stderr, "\033[1mwarning:\033[0m command targets paths outside project\033[K\r\n")
 		}
-		fmt.Fprintf(os.Stderr, "%s│\033[0m %s%s%s %s│\033[0m\033[K\r\n", boxColor,
-			strings.Repeat(" ", padding), warning, strings.Repeat(" ", state.innerWidth-len(warning)-padding), boxColor)
+		fmt.Fprintf(os.Stderr, "\033[K\r\n") // blank line after warning
 	}
 
-	// Draw tool info
+	// Draw tool info (plain white)
 	for _, line := range toolLines {
-		fmt.Fprintf(os.Stderr, "%s│\033[0m %-*s %s│\033[0m\033[K\r\n", boxColor, state.innerWidth, line, boxColor)
+		fmt.Fprintf(os.Stderr, "%s\033[K\r\n", line)
 	}
 
-	// Draw separator
-	fmt.Fprintf(os.Stderr, "%s├%s┤\033[0m\033[K\r\n", boxColor, strings.Repeat("─", state.boxWidth-2))
+	// Blank line separator
+	fmt.Fprintf(os.Stderr, "\033[K\r\n")
 
-	// Draw options with numbers (Deny option includes reason input)
 	for i, label := range optionLabels {
-		if i == 2 { // Deny option - show with reason input beside it
+		if i == 2 {
 			denyLabel := "3. Deny: "
-			availableWidth := state.innerWidth - 2 - len(denyLabel)
-			if availableWidth < 5 {
-				availableWidth = 5
-			}
 			inputDisplay := state.denyReason
-			if len(inputDisplay) > availableWidth {
-				inputDisplay = inputDisplay[len(inputDisplay)-availableWidth:]
+			if inputDisplay == "" {
+				inputDisplay = "\033[90m(optional reason)\033[0m"
 			}
 			if i == state.selected {
-				fmt.Fprintf(os.Stderr, "%s│\033[0m \033[1;32m> %s\033[0m%-*s %s│\033[0m\033[K\r\n", boxColor, denyLabel, availableWidth, inputDisplay, boxColor)
+				fmt.Fprintf(os.Stderr, "  \033[1m%s\033[0m%s\033[K\r\n", denyLabel, inputDisplay)
 			} else {
-				fmt.Fprintf(os.Stderr, "%s│\033[0m   \033[90m%s\033[0m%-*s %s│\033[0m\033[K\r\n", boxColor, denyLabel, availableWidth, inputDisplay, boxColor)
+				fmt.Fprintf(os.Stderr, "  \033[37m%s\033[0m%s\033[K\r\n", denyLabel, inputDisplay)
 			}
 		} else {
 			displayLabel := label
-			if len(displayLabel) > state.innerWidth-2 {
-				displayLabel = displayLabel[:state.innerWidth-5] + "..."
+			if i == 1 && state.allowlistInfo != "" {
+				displayLabel = fmt.Sprintf("%s  \033[90m%s\033[0m", label, state.allowlistInfo)
 			}
 			if i == state.selected {
-				fmt.Fprintf(os.Stderr, "%s│\033[0m \033[1;32m> %-*s\033[0m %s│\033[0m\033[K\r\n", boxColor, state.innerWidth-2, displayLabel, boxColor)
+				fmt.Fprintf(os.Stderr, "  \033[1m%s\033[0m\033[K\r\n", displayLabel)
 			} else {
-				fmt.Fprintf(os.Stderr, "%s│\033[0m   %-*s %s│\033[0m\033[K\r\n", boxColor, state.innerWidth-2, displayLabel, boxColor)
+				fmt.Fprintf(os.Stderr, "  \033[37m%s\033[0m\033[K\r\n", displayLabel)
 			}
 		}
 	}
 
-	// Draw box bottom
-	fmt.Fprintf(os.Stderr, "%s└%s┘\033[0m\033[K\r\n", boxColor, strings.Repeat("─", state.boxWidth-2))
+	// Blank line before hint
+	fmt.Fprintf(os.Stderr, "\033[K\r\n")
 
-	// Draw hint (may be multiple lines)
+	// Draw hint (dark grey)
 	for i, line := range hintLines {
 		if i == len(hintLines)-1 {
-			// Last line - no newline
 			fmt.Fprintf(os.Stderr, "\033[90m%s\033[0m\033[K", line)
 		} else {
 			fmt.Fprintf(os.Stderr, "\033[90m%s\033[0m\033[K\r\n", line)
@@ -748,50 +849,39 @@ func renderSelectorBox(state *selectorState) {
 func updateSelectorOptions(state *selectorState) {
 	hintLines := getHintLines(state)
 
-	// Use red for warning (outside cwd), cyan for normal
-	boxColor := "\033[36m" // cyan
-	if state.isWarning {
-		boxColor = "\033[91m" // bright red
-	}
-
 	// Move up to the first option line
 	// Cursor is at end of last hint line, need to go up:
-	// (hint lines - 1) + 1 (bottom border) + numOptions
+	// (hint lines - 1) + 1 (blank line) + numOptions
 	linesToMove := len(hintLines) - 1 + 1 + len(optionLabels)
 	fmt.Fprintf(os.Stderr, "\033[%dA\r", linesToMove)
 
-	// Redraw options (Deny option includes reason input)
 	for i, label := range optionLabels {
-		if i == 2 { // Deny option
+		if i == 2 {
 			denyLabel := "3. Deny: "
-			availableWidth := state.innerWidth - 2 - len(denyLabel)
-			if availableWidth < 5 {
-				availableWidth = 5
-			}
 			inputDisplay := state.denyReason
-			if len(inputDisplay) > availableWidth {
-				inputDisplay = inputDisplay[len(inputDisplay)-availableWidth:]
+			if inputDisplay == "" {
+				inputDisplay = "\033[90m(optional reason)\033[0m"
 			}
 			if i == state.selected {
-				fmt.Fprintf(os.Stderr, "%s│\033[0m \033[1;32m> %s\033[0m%-*s %s│\033[0m\033[K\r\n", boxColor, denyLabel, availableWidth, inputDisplay, boxColor)
+				fmt.Fprintf(os.Stderr, "  \033[1m%s\033[0m%s\033[K\r\n", denyLabel, inputDisplay)
 			} else {
-				fmt.Fprintf(os.Stderr, "%s│\033[0m   \033[90m%s\033[0m%-*s %s│\033[0m\033[K\r\n", boxColor, denyLabel, availableWidth, inputDisplay, boxColor)
+				fmt.Fprintf(os.Stderr, "  \033[37m%s\033[0m%s\033[K\r\n", denyLabel, inputDisplay)
 			}
 		} else {
 			displayLabel := label
-			if len(displayLabel) > state.innerWidth-2 {
-				displayLabel = displayLabel[:state.innerWidth-5] + "..."
+			if i == 1 && state.allowlistInfo != "" {
+				displayLabel = fmt.Sprintf("%s  \033[90m%s\033[0m", label, state.allowlistInfo)
 			}
 			if i == state.selected {
-				fmt.Fprintf(os.Stderr, "%s│\033[0m \033[1;32m> %-*s\033[0m %s│\033[0m\033[K\r\n", boxColor, state.innerWidth-2, displayLabel, boxColor)
+				fmt.Fprintf(os.Stderr, "  \033[1m%s\033[0m\033[K\r\n", displayLabel)
 			} else {
-				fmt.Fprintf(os.Stderr, "%s│\033[0m   %-*s %s│\033[0m\033[K\r\n", boxColor, state.innerWidth-2, displayLabel, boxColor)
+				fmt.Fprintf(os.Stderr, "  \033[37m%s\033[0m\033[K\r\n", displayLabel)
 			}
 		}
 	}
 
-	// Redraw bottom and hint
-	fmt.Fprintf(os.Stderr, "%s└%s┘\033[0m\033[K\r\n", boxColor, strings.Repeat("─", state.boxWidth-2))
+	// Blank line + hint
+	fmt.Fprintf(os.Stderr, "\033[K\r\n")
 	for i, line := range hintLines {
 		if i == len(hintLines)-1 {
 			fmt.Fprintf(os.Stderr, "\033[90m%s\033[0m\033[K", line)
@@ -805,36 +895,26 @@ func updateSelectorOptions(state *selectorState) {
 func updateReasonInput(state *selectorState) {
 	hintLines := getHintLines(state)
 
-	// Use red for warning (outside cwd), cyan for normal
-	boxColor := "\033[36m" // cyan
-	if state.isWarning {
-		boxColor = "\033[91m" // bright red
-	}
-
 	// Move up to the Deny line (3rd option, index 2)
 	// Cursor is at end of last hint line, need to go up:
-	// (hint lines - 1) + 1 (bottom border) + 1 (Deny is last option)
+	// (hint lines - 1) + 1 (blank line) + 1 (Deny is last option)
 	linesToMove := len(hintLines) - 1 + 1 + 1
 	fmt.Fprintf(os.Stderr, "\033[%dA\r", linesToMove)
 
 	// Redraw Deny line with reason
 	denyLabel := "3. Deny: "
-	availableWidth := state.innerWidth - 2 - len(denyLabel)
-	if availableWidth < 5 {
-		availableWidth = 5
-	}
 	inputDisplay := state.denyReason
-	if len(inputDisplay) > availableWidth {
-		inputDisplay = inputDisplay[len(inputDisplay)-availableWidth:]
+	if inputDisplay == "" {
+		inputDisplay = "\033[90m(optional reason)\033[0m"
 	}
 	if state.selected == 2 {
-		fmt.Fprintf(os.Stderr, "%s│\033[0m \033[1;32m> %s\033[0m%-*s %s│\033[0m\033[K\r\n", boxColor, denyLabel, availableWidth, inputDisplay, boxColor)
+		fmt.Fprintf(os.Stderr, "  \033[1m%s\033[0m%s\033[K\r\n", denyLabel, inputDisplay)
 	} else {
-		fmt.Fprintf(os.Stderr, "%s│\033[0m   \033[90m%s\033[0m%-*s %s│\033[0m\033[K\r\n", boxColor, denyLabel, availableWidth, inputDisplay, boxColor)
+		fmt.Fprintf(os.Stderr, "  \033[37m%s\033[0m%s\033[K\r\n", denyLabel, inputDisplay)
 	}
 
-	// Redraw bottom and hint
-	fmt.Fprintf(os.Stderr, "%s└%s┘\033[0m\033[K\r\n", boxColor, strings.Repeat("─", state.boxWidth-2))
+	// Blank line + hint
+	fmt.Fprintf(os.Stderr, "\033[K\r\n")
 	for i, line := range hintLines {
 		if i == len(hintLines)-1 {
 			fmt.Fprintf(os.Stderr, "\033[90m%s\033[0m\033[K", line)
@@ -858,11 +938,10 @@ func clearSelectorBox(state *selectorState) {
 // fallbackApproval handles approval when terminal control isn't available.
 func (a *ApprovalManager) fallbackApproval(toolDisplay string) (ApprovalResult, error) {
 	fmt.Fprintln(os.Stderr)
-	fmt.Fprintln(os.Stderr, "━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━")
 	fmt.Fprintln(os.Stderr, toolDisplay)
-	fmt.Fprintln(os.Stderr, "━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━")
-	fmt.Fprintln(os.Stderr, "[1] Execute once  [2] Always allow  [3] Deny")
-	fmt.Fprint(os.Stderr, "Choice: ")
+	fmt.Fprintln(os.Stderr)
+	fmt.Fprintln(os.Stderr, "[1] Execute once  [2] Allow for this session  [3] Deny")
+	fmt.Fprint(os.Stderr, "choice: ")
 
 	var input string
 	fmt.Scanln(&input)
@@ -905,19 +984,16 @@ func (a *ApprovalManager) AllowedTools() []string {
 
 // FormatApprovalResult returns a formatted string showing the approval result.
 func FormatApprovalResult(toolName string, args map[string]any, result ApprovalResult) string {
-	var status string
-	var icon string
+	var label string
+	displayName := ToolDisplayName(toolName)
 
 	switch result.Decision {
 	case ApprovalOnce:
-		status = "Approved"
-		icon = "\033[32m✓\033[0m"
+		label = "Approved"
 	case ApprovalAlways:
-		status = "Always allowed"
-		icon = "\033[32m✓\033[0m"
+		label = "Always allowed"
 	case ApprovalDeny:
-		status = "Denied"
-		icon = "\033[31m✗\033[0m"
+		label = "Denied"
 	}
 
 	// Format based on tool type
@@ -927,7 +1003,7 @@ func FormatApprovalResult(toolName string, args map[string]any, result ApprovalR
 			if len(cmd) > 40 {
 				cmd = cmd[:37] + "..."
 			}
-			return fmt.Sprintf("▶ bash: %s [%s] %s", cmd, status, icon)
+			return fmt.Sprintf("\033[1m%s:\033[0m %s: %s", label, displayName, cmd)
 		}
 	}
 
@@ -937,11 +1013,11 @@ func FormatApprovalResult(toolName string, args map[string]any, result ApprovalR
 			if len(query) > 40 {
 				query = query[:37] + "..."
 			}
-			return fmt.Sprintf("▶ web_search: %s [%s] %s", query, status, icon)
+			return fmt.Sprintf("\033[1m%s:\033[0m %s: %s", label, displayName, query)
 		}
 	}
 
-	return fmt.Sprintf("▶ %s [%s] %s", toolName, status, icon)
+	return fmt.Sprintf("\033[1m%s:\033[0m %s", label, displayName)
 }
 
 // FormatDenyResult returns the tool result message when a tool is denied.
@@ -951,3 +1027,78 @@ func FormatDenyResult(toolName string, reason string) string {
 	}
 	return fmt.Sprintf("User denied execution of %s.", toolName)
 }
+
+// PromptYesNo displays a simple Yes/No prompt and returns the user's choice.
+// Returns true for Yes, false for No.
+func PromptYesNo(question string) (bool, error) {
+	fd := int(os.Stdin.Fd())
+	oldState, err := term.MakeRaw(fd)
+	if err != nil {
+		return false, err
+	}
+	defer term.Restore(fd, oldState)
+
+	selected := 0 // 0 = Yes, 1 = No
+	options := []string{"Yes", "No"}
+
+	// Hide cursor
+	fmt.Fprint(os.Stderr, "\033[?25l")
+	defer fmt.Fprint(os.Stderr, "\033[?25h")
+
+	renderYesNo := func() {
+		// Move to start of line and clear
+		fmt.Fprintf(os.Stderr, "\r\033[K")
+		fmt.Fprintf(os.Stderr, "%s  ", question)
+		for i, opt := range options {
+			if i == selected {
+				fmt.Fprintf(os.Stderr, "\033[1m%s\033[0m  ", opt)
+			} else {
+				fmt.Fprintf(os.Stderr, "\033[37m%s\033[0m  ", opt)
+			}
+		}
+	}
+
+	renderYesNo()
+
+	buf := make([]byte, 3)
+	for {
+		n, err := os.Stdin.Read(buf)
+		if err != nil {
+			return false, err
+		}
+
+		if n == 1 {
+			switch buf[0] {
+			case 'y', 'Y':
+				selected = 0
+				renderYesNo()
+			case 'n', 'N':
+				selected = 1
+				renderYesNo()
+			case '\r', '\n': // Enter
+				fmt.Fprintf(os.Stderr, "\r\033[K") // Clear line
+				return selected == 0, nil
+			case 3: // Ctrl+C
+				fmt.Fprintf(os.Stderr, "\r\033[K")
+				return false, nil
+			case 27: // Escape - could be arrow key
+				// Read more bytes for arrow keys
+				continue
+			}
+		} else if n == 3 && buf[0] == 27 && buf[1] == 91 {
+			// Arrow keys
+			switch buf[2] {
+			case 'D': // Left
+				if selected > 0 {
+					selected--
+				}
+				renderYesNo()
+			case 'C': // Right
+				if selected < len(options)-1 {
+					selected++
+				}
+				renderYesNo()
+			}
+		}
+	}
+}

x/agent/approval_test.go

@@ -151,6 +151,27 @@ func TestExtractBashPrefix(t *testing.T) {
 			command:  "head -n 100",
 			expected: "",
 		},
+		// Path traversal security tests
+		{
+			name:     "path traversal - parent escape",
+			command:  "cat tools/../../etc/passwd",
+			expected: "", // Should NOT create a prefix - path escapes
+		},
+		{
+			name:     "path traversal - deep escape",
+			command:  "cat tools/a/b/../../../etc/passwd",
+			expected: "", // Normalizes to "../etc/passwd" - escapes
+		},
+		{
+			name:     "path traversal - absolute path",
+			command:  "cat /etc/passwd",
+			expected: "", // Absolute paths should not create prefix
+		},
+		{
+			name:     "path with safe dotdot - normalized",
+			command:  "cat tools/subdir/../file.go",
+			expected: "cat:tools/", // Normalizes to tools/file.go - safe, creates prefix
+		},
 	}
 
 	for _, tt := range tests {
@@ -164,6 +185,34 @@ func TestExtractBashPrefix(t *testing.T) {
 	}
 }
 
+func TestApprovalManager_PathTraversalBlocked(t *testing.T) {
+	am := NewApprovalManager()
+
+	// Allow "cat tools/file.go" - creates prefix "cat:tools/"
+	am.AddToAllowlist("bash", map[string]any{"command": "cat tools/file.go"})
+
+	// Path traversal attack: should NOT be allowed
+	if am.IsAllowed("bash", map[string]any{"command": "cat tools/../../etc/passwd"}) {
+		t.Error("SECURITY: path traversal attack should NOT be allowed")
+	}
+
+	// Another traversal variant
+	if am.IsAllowed("bash", map[string]any{"command": "cat tools/../../../etc/shadow"}) {
+		t.Error("SECURITY: deep path traversal should NOT be allowed")
+	}
+
+	// Valid subdirectory access should still work
+	if !am.IsAllowed("bash", map[string]any{"command": "cat tools/subdir/file.go"}) {
+		t.Error("expected cat tools/subdir/file.go to be allowed")
+	}
+
+	// Safe ".." that normalizes to within allowed directory should work
+	// tools/subdir/../other.go normalizes to tools/other.go which is under tools/
+	if !am.IsAllowed("bash", map[string]any{"command": "cat tools/subdir/../other.go"}) {
+		t.Error("expected cat tools/subdir/../other.go to be allowed (normalizes to tools/other.go)")
+	}
+}
+
 func TestApprovalManager_PrefixAllowlist(t *testing.T) {
 	am := NewApprovalManager()
 
@@ -186,6 +235,119 @@ func TestApprovalManager_PrefixAllowlist(t *testing.T) {
 	}
 }
 
+func TestApprovalManager_HierarchicalPrefixAllowlist(t *testing.T) {
+	am := NewApprovalManager()
+
+	// Allow "cat tools/file.go" - this creates prefix "cat:tools/"
+	am.AddToAllowlist("bash", map[string]any{"command": "cat tools/file.go"})
+
+	// Should allow subdirectories (hierarchical matching)
+	if !am.IsAllowed("bash", map[string]any{"command": "cat tools/subdir/file.go"}) {
+		t.Error("expected cat tools/subdir/file.go to be allowed via hierarchical prefix")
+	}
+
+	// Should allow deeply nested subdirectories
+	if !am.IsAllowed("bash", map[string]any{"command": "cat tools/a/b/c/deep.go"}) {
+		t.Error("expected cat tools/a/b/c/deep.go to be allowed via hierarchical prefix")
+	}
+
+	// Should still allow same directory
+	if !am.IsAllowed("bash", map[string]any{"command": "cat tools/another.go"}) {
+		t.Error("expected cat tools/another.go to be allowed")
+	}
+
+	// Should NOT allow different base directory
+	if am.IsAllowed("bash", map[string]any{"command": "cat src/main.go"}) {
+		t.Error("expected cat src/main.go to NOT be allowed")
+	}
+
+	// Should NOT allow different command even in subdirectory
+	if am.IsAllowed("bash", map[string]any{"command": "ls tools/subdir/"}) {
+		t.Error("expected ls tools/subdir/ to NOT be allowed (different command)")
+	}
+
+	// Should NOT allow similar but different directory name
+	if am.IsAllowed("bash", map[string]any{"command": "cat toolsbin/file.go"}) {
+		t.Error("expected cat toolsbin/file.go to NOT be allowed (different directory)")
+	}
+}
+
+func TestApprovalManager_HierarchicalPrefixAllowlist_CrossPlatform(t *testing.T) {
+	am := NewApprovalManager()
+
+	// Allow with forward slashes (Unix-style)
+	am.AddToAllowlist("bash", map[string]any{"command": "cat tools/file.go"})
+
+	// Should work with backslashes too (Windows-style) - normalized internally
+	if !am.IsAllowed("bash", map[string]any{"command": "cat tools\\subdir\\file.go"}) {
+		t.Error("expected cat tools\\subdir\\file.go to be allowed via hierarchical prefix (Windows path)")
+	}
+
+	// Mixed slashes should also work
+	if !am.IsAllowed("bash", map[string]any{"command": "cat tools\\a/b\\c/deep.go"}) {
+		t.Error("expected mixed slash path to be allowed via hierarchical prefix")
+	}
+}
+
+func TestMatchesHierarchicalPrefix(t *testing.T) {
+	am := NewApprovalManager()
+
+	// Add prefix for "cat:tools/"
+	am.prefixes["cat:tools/"] = true
+
+	tests := []struct {
+		name     string
+		prefix   string
+		expected bool
+	}{
+		{
+			name:     "exact match",
+			prefix:   "cat:tools/",
+			expected: true, // exact match also passes HasPrefix - caller handles exact match first
+		},
+		{
+			name:     "subdirectory",
+			prefix:   "cat:tools/subdir/",
+			expected: true,
+		},
+		{
+			name:     "deeply nested",
+			prefix:   "cat:tools/a/b/c/",
+			expected: true,
+		},
+		{
+			name:     "different base directory",
+			prefix:   "cat:src/",
+			expected: false,
+		},
+		{
+			name:     "different command same path",
+			prefix:   "ls:tools/",
+			expected: false,
+		},
+		{
+			name:     "similar directory name",
+			prefix:   "cat:toolsbin/",
+			expected: false,
+		},
+		{
+			name:     "invalid prefix format",
+			prefix:   "cattools",
+			expected: false,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			result := am.matchesHierarchicalPrefix(tt.prefix)
+			if result != tt.expected {
+				t.Errorf("matchesHierarchicalPrefix(%q) = %v, expected %v",
+					tt.prefix, result, tt.expected)
+			}
+		})
+	}
+}
+
 func TestFormatApprovalResult(t *testing.T) {
 	tests := []struct {
 		name     string

x/cmd/run.go

@@ -6,10 +6,12 @@ import (
 	"errors"
 	"fmt"
 	"io"
+	"net/url"
 	"os"
 	"os/signal"
 	"strings"
 	"syscall"
+	"time"
 
 	"github.com/spf13/cobra"
 	"golang.org/x/term"
@@ -22,6 +24,101 @@ import (
 	"github.com/ollama/ollama/x/tools"
 )
 
+// Tool output capping constants
+const (
+	// localModelTokenLimit is the token limit for local models (smaller context).
+	localModelTokenLimit = 4000
+
+	// defaultTokenLimit is the token limit for cloud/remote models.
+	defaultTokenLimit = 10000
+
+	// charsPerToken is a rough estimate of characters per token.
+	// TODO: Estimate tokens more accurately using tokenizer if available
+	charsPerToken = 4
+)
+
+// isLocalModel checks if the model is running locally (not a cloud model).
+// TODO: Improve local/cloud model identification - could check model metadata
+func isLocalModel(modelName string) bool {
+	return !strings.HasSuffix(modelName, "-cloud")
+}
+
+// isLocalServer checks if connecting to a local Ollama server.
+// TODO: Could also check other indicators of local vs cloud server
+func isLocalServer() bool {
+	host := os.Getenv("OLLAMA_HOST")
+	if host == "" {
+		return true // Default is localhost:11434
+	}
+
+	// Parse the URL to check host
+	parsed, err := url.Parse(host)
+	if err != nil {
+		return true // If can't parse, assume local
+	}
+
+	hostname := parsed.Hostname()
+	return hostname == "localhost" || hostname == "127.0.0.1" || strings.Contains(parsed.Host, ":11434")
+}
+
+// truncateToolOutput truncates tool output to prevent context overflow.
+// Uses a smaller limit (4k tokens) for local models, larger (10k) for cloud/remote.
+func truncateToolOutput(output, modelName string) string {
+	var tokenLimit int
+	if isLocalModel(modelName) && isLocalServer() {
+		tokenLimit = localModelTokenLimit
+	} else {
+		tokenLimit = defaultTokenLimit
+	}
+
+	maxChars := tokenLimit * charsPerToken
+	if len(output) > maxChars {
+		return output[:maxChars] + "\n... (output truncated)"
+	}
+	return output
+}
+
+// waitForOllamaSignin shows the signin URL and polls until authentication completes.
+func waitForOllamaSignin(ctx context.Context) error {
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		return err
+	}
+
+	// Get signin URL from initial Whoami call
+	_, err = client.Whoami(ctx)
+	if err != nil {
+		var aErr api.AuthorizationError
+		if errors.As(err, &aErr) && aErr.SigninURL != "" {
+			fmt.Fprintf(os.Stderr, "\n  To sign in, navigate to:\n")
+			fmt.Fprintf(os.Stderr, "      %s\n\n", aErr.SigninURL)
+			fmt.Fprintf(os.Stderr, "  \033[90mwaiting for sign in to complete...\033[0m")
+
+			// Poll until auth succeeds
+			ticker := time.NewTicker(2 * time.Second)
+			defer ticker.Stop()
+
+			for {
+				select {
+				case <-ctx.Done():
+					fmt.Fprintf(os.Stderr, "\n")
+					return ctx.Err()
+				case <-ticker.C:
+					user, whoamiErr := client.Whoami(ctx)
+					if whoamiErr == nil && user != nil && user.Name != "" {
+						fmt.Fprintf(os.Stderr, "\r\033[K\033[A\r\033[K  \033[1msigned in:\033[0m %s\n", user.Name)
+						return nil
+					}
+					// Still waiting, show dot
+					fmt.Fprintf(os.Stderr, ".")
+				}
+			}
+		}
+		return err
+	}
+	return nil
+}
+
 // RunOptions contains options for running an interactive agent session.
 type RunOptions struct {
 	Model        string
@@ -37,6 +134,9 @@ type RunOptions struct {
 	// Agent fields (managed externally for session persistence)
 	Tools    *tools.Registry
 	Approval *agent.ApprovalManager
+
+	// YoloMode skips all tool approval prompts
+	YoloMode bool
 }
 
 // Chat runs an agent chat loop with tool support.
@@ -77,6 +177,7 @@ func Chat(ctx context.Context, opts RunOptions) (*api.Message, error) {
 	var thinkTagOpened bool = false
 	var thinkTagClosed bool = false
 	var pendingToolCalls []api.ToolCall
+	var consecutiveErrors int // Track consecutive 500 errors for retry limit
 
 	role := "assistant"
 	messages := opts.Messages
@@ -159,6 +260,58 @@ func Chat(ctx context.Context, opts RunOptions) (*api.Message, error) {
 				return nil, nil
 			}
 
+			// Check for 401 Unauthorized - prompt user to sign in
+			var authErr api.AuthorizationError
+			if errors.As(err, &authErr) {
+				p.StopAndClear()
+				fmt.Fprintf(os.Stderr, "\033[1mauth required:\033[0m cloud model requires authentication\n")
+				result, promptErr := agent.PromptYesNo("Sign in to Ollama?")
+				if promptErr == nil && result {
+					if signinErr := waitForOllamaSignin(ctx); signinErr == nil {
+						// Retry the chat request
+						fmt.Fprintf(os.Stderr, "\033[90mretrying...\033[0m\n")
+						continue // Retry the loop
+					}
+				}
+				return nil, fmt.Errorf("authentication required - run 'ollama signin' to authenticate")
+			}
+
+			// Check for 500 errors (often tool parsing failures) - inform the model
+			var statusErr api.StatusError
+			if errors.As(err, &statusErr) && statusErr.StatusCode >= 500 {
+				consecutiveErrors++
+				p.StopAndClear()
+
+				if consecutiveErrors >= 3 {
+					fmt.Fprintf(os.Stderr, "\033[1merror:\033[0m too many consecutive errors, giving up\n")
+					return nil, fmt.Errorf("too many consecutive server errors: %s", statusErr.ErrorMessage)
+				}
+
+				fmt.Fprintf(os.Stderr, "\033[1mwarning:\033[0m server error (attempt %d/3): %s\n", consecutiveErrors, statusErr.ErrorMessage)
+
+				// Include both the model's response and the error so it can learn
+				assistantContent := fullResponse.String()
+				if assistantContent == "" {
+					assistantContent = "(empty response)"
+				}
+				errorMsg := fmt.Sprintf("Your previous response caused an error: %s\n\nYour response was:\n%s\n\nPlease try again with a valid response.", statusErr.ErrorMessage, assistantContent)
+				messages = append(messages,
+					api.Message{Role: "user", Content: errorMsg},
+				)
+
+				// Reset state and retry
+				fullResponse.Reset()
+				thinkingContent.Reset()
+				thinkTagOpened = false
+				thinkTagClosed = false
+				pendingToolCalls = nil
+				state = &displayResponseState{}
+				p = progress.NewProgress(os.Stderr)
+				spinner = progress.NewSpinner("")
+				p.Add("", spinner)
+				continue
+			}
+
 			if strings.Contains(err.Error(), "upstream error") {
 				p.StopAndClear()
 				fmt.Println("An error occurred while processing your message. Please try again.")
@@ -168,6 +321,9 @@ func Chat(ctx context.Context, opts RunOptions) (*api.Message, error) {
 			return nil, err
 		}
 
+		// Reset consecutive error counter on success
+		consecutiveErrors = 0
+
 		// If no tool calls, we're done
 		if len(pendingToolCalls) == 0 || toolRegistry == nil {
 			break
@@ -197,8 +353,8 @@ func Chat(ctx context.Context, opts RunOptions) (*api.Message, error) {
 				if cmd, ok := args["command"].(string); ok {
 					// Check if command is denied (dangerous pattern)
 					if denied, pattern := agent.IsDenied(cmd); denied {
-						fmt.Fprintf(os.Stderr, "\033[91m✗ Blocked: %s\033[0m\n", formatToolShort(toolName, args))
-						fmt.Fprintf(os.Stderr, "\033[91m  Matches dangerous pattern: %s\033[0m\n", pattern)
+						fmt.Fprintf(os.Stderr, "\033[1mblocked:\033[0m %s\n", formatToolShort(toolName, args))
+						fmt.Fprintf(os.Stderr, "  matches dangerous pattern: %s\n", pattern)
 						toolResults = append(toolResults, api.Message{
 							Role:       "tool",
 							Content:    agent.FormatDeniedResult(cmd, pattern),
@@ -208,15 +364,21 @@ func Chat(ctx context.Context, opts RunOptions) (*api.Message, error) {
 					}
 
 					// Check if command is auto-allowed (safe command)
-					if agent.IsAutoAllowed(cmd) {
-						fmt.Fprintf(os.Stderr, "\033[90m▶ Auto-allowed: %s\033[0m\n", formatToolShort(toolName, args))
-						skipApproval = true
-					}
+					// TODO(parthsareen): re-enable with tighter scoped allowlist
+					// if agent.IsAutoAllowed(cmd) {
+					// 	fmt.Fprintf(os.Stderr, "\033[1mauto-allowed:\033[0m %s\n", formatToolShort(toolName, args))
+					// 	skipApproval = true
+					// }
 				}
 			}
 
 			// Check approval (uses prefix matching for bash commands)
-			if !skipApproval && !approval.IsAllowed(toolName, args) {
+			// In yolo mode, skip all approval prompts
+			if opts.YoloMode {
+				if !skipApproval {
+					fmt.Fprintf(os.Stderr, "\033[1mrunning:\033[0m %s\n", formatToolShort(toolName, args))
+				}
+			} else if !skipApproval && !approval.IsAllowed(toolName, args) {
 				result, err := approval.RequestApproval(toolName, args)
 				if err != nil {
 					fmt.Fprintf(os.Stderr, "Error requesting approval: %v\n", err)
@@ -244,13 +406,30 @@ func Chat(ctx context.Context, opts RunOptions) (*api.Message, error) {
 				}
 			} else if !skipApproval {
 				// Already allowed - show running indicator
-				fmt.Fprintf(os.Stderr, "\033[90m▶ Running: %s\033[0m\n", formatToolShort(toolName, args))
+				fmt.Fprintf(os.Stderr, "\033[1mrunning:\033[0m %s\n", formatToolShort(toolName, args))
 			}
 
 			// Execute the tool
 			toolResult, err := toolRegistry.Execute(call)
 			if err != nil {
-				fmt.Fprintf(os.Stderr, "\033[31m  Error: %v\033[0m\n", err)
+				// Check if web search needs authentication
+				if errors.Is(err, tools.ErrWebSearchAuthRequired) {
+					// Prompt user to sign in
+					fmt.Fprintf(os.Stderr, "\033[1mauth required:\033[0m web search requires authentication\n")
+					result, promptErr := agent.PromptYesNo("Sign in to Ollama?")
+					if promptErr == nil && result {
+						// Get signin URL and wait for auth completion
+						if signinErr := waitForOllamaSignin(ctx); signinErr == nil {
+							// Retry the web search
+							fmt.Fprintf(os.Stderr, "\033[90mretrying web search...\033[0m\n")
+							toolResult, err = toolRegistry.Execute(call)
+							if err == nil {
+								goto toolSuccess
+							}
+						}
+					}
+				}
+				fmt.Fprintf(os.Stderr, "\033[1merror:\033[0m %v\n", err)
 				toolResults = append(toolResults, api.Message{
 					Role:       "tool",
 					Content:    fmt.Sprintf("Error: %v", err),
@@ -258,6 +437,7 @@ func Chat(ctx context.Context, opts RunOptions) (*api.Message, error) {
 				})
 				continue
 			}
+		toolSuccess:
 
 			// Display tool output (truncated for display)
 			if toolResult != "" {
@@ -269,9 +449,12 @@ func Chat(ctx context.Context, opts RunOptions) (*api.Message, error) {
 				fmt.Fprintf(os.Stderr, "\033[90m  %s\033[0m\n", strings.ReplaceAll(output, "\n", "\n  "))
 			}
 
+			// Truncate output to prevent context overflow
+			toolResultForLLM := truncateToolOutput(toolResult, opts.Model)
+
 			toolResults = append(toolResults, api.Message{
 				Role:       "tool",
-				Content:    toolResult,
+				Content:    toolResultForLLM,
 				ToolCallID: call.ID,
 			})
 		}
@@ -317,17 +500,18 @@ func truncateUTF8(s string, limit int) string {
 
 // formatToolShort returns a short description of a tool call.
 func formatToolShort(toolName string, args map[string]any) string {
+	displayName := agent.ToolDisplayName(toolName)
 	if toolName == "bash" {
 		if cmd, ok := args["command"].(string); ok {
-			return fmt.Sprintf("bash: %s", truncateUTF8(cmd, 50))
+			return fmt.Sprintf("%s: %s", displayName, truncateUTF8(cmd, 50))
 		}
 	}
 	if toolName == "web_search" {
 		if query, ok := args["query"].(string); ok {
-			return fmt.Sprintf("web_search: %s", truncateUTF8(query, 50))
+			return fmt.Sprintf("%s: %s", displayName, truncateUTF8(query, 50))
 		}
 	}
-	return toolName
+	return displayName
 }
 
 // Helper types and functions for display
@@ -449,7 +633,8 @@ func checkModelCapabilities(ctx context.Context, modelName string) (supportsTool
 
 // GenerateInteractive runs an interactive agent session.
 // This is called from cmd.go when --experimental flag is set.
-func GenerateInteractive(cmd *cobra.Command, modelName string, wordWrap bool, options map[string]any, think *api.ThinkValue, hideThinking bool, keepAlive *api.Duration) error {
+// If yoloMode is true, all tool approvals are skipped.
+func GenerateInteractive(cmd *cobra.Command, modelName string, wordWrap bool, options map[string]any, think *api.ThinkValue, hideThinking bool, keepAlive *api.Duration, yoloMode bool) error {
 	scanner, err := readline.New(readline.Prompt{
 		Prompt:         ">>> ",
 		AltPrompt:      "... ",
@@ -466,7 +651,7 @@ func GenerateInteractive(cmd *cobra.Command, modelName string, wordWrap bool, op
 	// Check if model supports tools
 	supportsTools, err := checkModelCapabilities(cmd.Context(), modelName)
 	if err != nil {
-		fmt.Fprintf(os.Stderr, "\033[33mWarning: Could not check model capabilities: %v\033[0m\n", err)
+		fmt.Fprintf(os.Stderr, "\033[1mwarning:\033[0m could not check model capabilities: %v\n", err)
 		supportsTools = false
 	}
 
@@ -474,14 +659,17 @@ func GenerateInteractive(cmd *cobra.Command, modelName string, wordWrap bool, op
 	var toolRegistry *tools.Registry
 	if supportsTools {
 		toolRegistry = tools.DefaultRegistry()
-		fmt.Fprintf(os.Stderr, "Tools available: %s\n", strings.Join(toolRegistry.Names(), ", "))
 
-		// Check for OLLAMA_API_KEY for web search
-		if os.Getenv("OLLAMA_API_KEY") == "" {
-			fmt.Fprintf(os.Stderr, "\033[33mWarning: OLLAMA_API_KEY not set - web search will not work\033[0m\n")
+		if toolRegistry.Has("bash") {
+			fmt.Fprintln(os.Stderr)
+			fmt.Fprintln(os.Stderr, "This experimental version of Ollama has the \033[1mbash\033[0m tool enabled.")
+			fmt.Fprintln(os.Stderr, "Models can read files on your computer, or run commands (after you allow them).")
+			fmt.Fprintln(os.Stderr)
+		}
+
+		if yoloMode {
+			fmt.Fprintf(os.Stderr, "\033[1mwarning:\033[0m yolo mode - all tool approvals will be skipped\n")
 		}
-	} else {
-		fmt.Fprintf(os.Stderr, "\033[33mNote: Model does not support tools - running in chat-only mode\033[0m\n")
 	}
 
 	// Create approval manager for session
@@ -524,6 +712,9 @@ func GenerateInteractive(cmd *cobra.Command, modelName string, wordWrap bool, op
 			fmt.Fprintln(os.Stderr, "  /bye            Exit")
 			fmt.Fprintln(os.Stderr, "  /?, /help       Help for a command")
 			fmt.Fprintln(os.Stderr, "")
+			fmt.Fprintln(os.Stderr, "Keyboard Shortcuts:")
+			fmt.Fprintln(os.Stderr, "  Ctrl+O          Expand last tool output")
+			fmt.Fprintln(os.Stderr, "")
 			continue
 		case strings.HasPrefix(line, "/"):
 			fmt.Printf("Unknown command '%s'. Type /? for help\n", strings.Fields(line)[0])
@@ -546,6 +737,7 @@ func GenerateInteractive(cmd *cobra.Command, modelName string, wordWrap bool, op
 				KeepAlive:    keepAlive,
 				Tools:        toolRegistry,
 				Approval:     approval,
+				YoloMode:     yoloMode,
 			}
 
 			assistant, err := Chat(cmd.Context(), opts)

x/cmd/run_test.go

@@ -0,0 +1,180 @@
+package cmd
+
+import (
+	"testing"
+)
+
+func TestIsLocalModel(t *testing.T) {
+	tests := []struct {
+		name      string
+		modelName string
+		expected  bool
+	}{
+		{
+			name:      "local model without suffix",
+			modelName: "llama3.2",
+			expected:  true,
+		},
+		{
+			name:      "local model with version",
+			modelName: "qwen2.5:7b",
+			expected:  true,
+		},
+		{
+			name:      "cloud model",
+			modelName: "gpt-4-cloud",
+			expected:  false,
+		},
+		{
+			name:      "cloud model with version",
+			modelName: "claude-3-cloud",
+			expected:  false,
+		},
+		{
+			name:      "empty model name",
+			modelName: "",
+			expected:  true,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			result := isLocalModel(tt.modelName)
+			if result != tt.expected {
+				t.Errorf("isLocalModel(%q) = %v, expected %v", tt.modelName, result, tt.expected)
+			}
+		})
+	}
+}
+
+func TestIsLocalServer(t *testing.T) {
+	tests := []struct {
+		name     string
+		host     string
+		expected bool
+	}{
+		{
+			name:     "empty host (default)",
+			host:     "",
+			expected: true,
+		},
+		{
+			name:     "localhost",
+			host:     "http://localhost:11434",
+			expected: true,
+		},
+		{
+			name:     "127.0.0.1",
+			host:     "http://127.0.0.1:11434",
+			expected: true,
+		},
+		{
+			name:     "custom port on localhost",
+			host:     "http://localhost:8080",
+			expected: true, // localhost is always considered local
+		},
+		{
+			name:     "remote host",
+			host:     "http://ollama.example.com:11434",
+			expected: true, // has :11434
+		},
+		{
+			name:     "remote host different port",
+			host:     "http://ollama.example.com:8080",
+			expected: false,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			t.Setenv("OLLAMA_HOST", tt.host)
+			result := isLocalServer()
+			if result != tt.expected {
+				t.Errorf("isLocalServer() with OLLAMA_HOST=%q = %v, expected %v", tt.host, result, tt.expected)
+			}
+		})
+	}
+}
+
+func TestTruncateToolOutput(t *testing.T) {
+	// Create outputs of different sizes
+	localLimitOutput := make([]byte, 20000)   // > 4k tokens (16k chars)
+	defaultLimitOutput := make([]byte, 50000) // > 10k tokens (40k chars)
+	for i := range localLimitOutput {
+		localLimitOutput[i] = 'a'
+	}
+	for i := range defaultLimitOutput {
+		defaultLimitOutput[i] = 'b'
+	}
+
+	tests := []struct {
+		name          string
+		output        string
+		modelName     string
+		host          string
+		shouldTrim    bool
+		expectedLimit int
+	}{
+		{
+			name:          "short output local model",
+			output:        "hello world",
+			modelName:     "llama3.2",
+			host:          "",
+			shouldTrim:    false,
+			expectedLimit: localModelTokenLimit,
+		},
+		{
+			name:          "long output local model - trimmed at 4k",
+			output:        string(localLimitOutput),
+			modelName:     "llama3.2",
+			host:          "",
+			shouldTrim:    true,
+			expectedLimit: localModelTokenLimit,
+		},
+		{
+			name:          "long output cloud model - uses 10k limit",
+			output:        string(localLimitOutput), // 20k chars, under 10k token limit
+			modelName:     "gpt-4-cloud",
+			host:          "",
+			shouldTrim:    false,
+			expectedLimit: defaultTokenLimit,
+		},
+		{
+			name:          "very long output cloud model - trimmed at 10k",
+			output:        string(defaultLimitOutput),
+			modelName:     "gpt-4-cloud",
+			host:          "",
+			shouldTrim:    true,
+			expectedLimit: defaultTokenLimit,
+		},
+		{
+			name:          "long output remote server - uses 10k limit",
+			output:        string(localLimitOutput),
+			modelName:     "llama3.2",
+			host:          "http://remote.example.com:8080",
+			shouldTrim:    false,
+			expectedLimit: defaultTokenLimit,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			t.Setenv("OLLAMA_HOST", tt.host)
+			result := truncateToolOutput(tt.output, tt.modelName)
+
+			if tt.shouldTrim {
+				maxLen := tt.expectedLimit * charsPerToken
+				if len(result) > maxLen+50 { // +50 for the truncation message
+					t.Errorf("expected output to be truncated to ~%d chars, got %d", maxLen, len(result))
+				}
+				if result == tt.output {
+					t.Error("expected output to be truncated but it wasn't")
+				}
+			} else {
+				if result != tt.output {
+					t.Error("expected output to not be truncated")
+				}
+			}
+		})
+	}
+}

x/imagegen/.gitignore

@@ -0,0 +1,38 @@
+# Build directories
+build/
+dist/
+
+# CMake
+CMakeCache.txt
+CMakeFiles/
+cmake_install.cmake
+Makefile
+*.cmake
+
+# IDE
+.idea/
+.vscode/
+*.swp
+*.swo
+*~
+
+# macOS
+.DS_Store
+*.dSYM/
+
+# Go
+*.exe
+*.exe~
+*.dll
+*.so
+*.dylib
+
+# Python
+*.npy
+
+/engine
+weights
+outputs
+
+prompt.txt
+negative.txt

x/imagegen/README.md

@@ -0,0 +1,250 @@
+# Image Generation in Ollama (Experimental)
+
+Generate images from text prompts using local AI models.
+
+## Quick Start
+
+```bash
+# Run with a prompt
+ollama run z-image "a sunset over mountains"
+Generating: step 30/30
+Image saved to: /tmp/ollama-image-1704067200.png
+```
+
+On macOS, the generated image will automatically open in Preview.
+
+## Supported Models
+
+| Model | VRAM Required | Notes |
+|-------|---------------|-------|
+| z-image | ~12GB | Based on Flux architecture |
+
+## CLI Usage
+
+```bash
+# Generate an image
+ollama run z-image "a cat playing piano"
+
+# Check if model is running
+ollama ps
+
+# Stop the model
+ollama stop z-image
+```
+
+## API
+
+### OpenAI-Compatible Endpoint
+
+```bash
+POST /v1/images/generations
+```
+
+**Request:**
+```json
+{
+  "model": "z-image",
+  "prompt": "a sunset over mountains",
+  "size": "1024x1024",
+  "response_format": "b64_json"
+}
+```
+
+**Response:**
+```json
+{
+  "created": 1704067200,
+  "data": [
+    {
+      "b64_json": "iVBORw0KGgo..."
+    }
+  ]
+}
+```
+
+### Example: cURL
+
+```bash
+curl http://localhost:11434/v1/images/generations \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "z-image",
+    "prompt": "a white cat",
+    "size": "1024x1024"
+  }'
+```
+
+### Example: Save to File
+
+```bash
+curl -s http://localhost:11434/v1/images/generations \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "z-image",
+    "prompt": "a white cat",
+    "size": "1024x1024"
+  }' | jq -r '.data[0].b64_json' | base64 -d > image.png
+```
+
+### Streaming Progress
+
+Enable streaming to receive progress updates via SSE:
+
+```bash
+curl http://localhost:11434/v1/images/generations \
+  -H "Content-Type: application/json" \
+  -d '{"model": "z-image", "prompt": "a sunset", "stream": true}'
+```
+
+Events:
+```
+event: progress
+data: {"step": 1, "total": 30}
+
+event: progress
+data: {"step": 2, "total": 30}
+...
+
+event: done
+data: {"created": 1704067200, "data": [{"b64_json": "..."}]}
+```
+
+## Parameters
+
+| Parameter | Type | Default | Description |
+|-----------|------|---------|-------------|
+| model | string | required | Model name |
+| prompt | string | required | Text description of image |
+| size | string | "1024x1024" | Image dimensions (WxH) |
+| n | int | 1 | Number of images (currently only 1 supported) |
+| response_format | string | "b64_json" | "b64_json" or "url" |
+| stream | bool | false | Enable progress streaming |
+
+## Requirements
+
+- macOS with Apple Silicon (M1/M2/M3/M4)
+- CUDA: tested on CUDA 12 Blackwell, more testing coming soon
+- Sufficient VRAM (see model table above)
+- Ollama built with MLX support
+
+## Limitations
+
+- macOS only (uses MLX backend)
+- Single image per request
+- Fixed step count (30 steps)
+- Modelfiles not yet supported (use `ollama create` from model directory)
+
+---
+
+# Tensor Model Storage Format
+
+Tensor models store each tensor as a separate blob with metadata in the manifest. This enables faster downloads (parallel fetching) and deduplication (shared tensors are stored once).
+
+## Manifest Structure
+
+The manifest follows the standard ollama format with tensor-specific layer metadata:
+
+```json
+{
+  "schemaVersion": 2,
+  "mediaType": "application/vnd.docker.distribution.manifest.v2+json",
+  "config": { "digest": "sha256:...", "size": 1234 },
+  "layers": [
+    {
+      "mediaType": "application/vnd.ollama.image.tensor",
+      "digest": "sha256:25b36eed...",
+      "size": 49807448,
+      "name": "text_encoder/model.layers.0.mlp.down_proj.weight",
+      "dtype": "BF16",
+      "shape": [2560, 9728]
+    },
+    {
+      "mediaType": "application/vnd.ollama.image.json",
+      "digest": "sha256:abc123...",
+      "size": 512,
+      "name": "text_encoder/config.json"
+    }
+  ]
+}
+```
+
+Each tensor layer includes:
+- `name`: Path-style tensor name (e.g., `text_encoder/model.layers.0.mlp.down_proj.weight`)
+- `dtype`: Data type (BF16, F32, etc.)
+- `shape`: Tensor dimensions
+
+Config layers use the same path-style naming (e.g., `tokenizer/tokenizer.json`).
+
+## Blob Format
+
+Each tensor blob is a minimal safetensors file:
+
+```
+[8 bytes: header size (uint64 LE)]
+[~80 bytes: JSON header, padded to 8-byte alignment]
+[N bytes: raw tensor data]
+```
+
+Header contains a single tensor named `"data"`:
+
+```json
+{"data":{"dtype":"BF16","shape":[2560,9728],"data_offsets":[0,49807360]}}
+```
+
+## Why Include the Header?
+
+The ~88 byte safetensors header enables MLX's native `mlx_load_safetensors` function, which:
+
+1. **Uses mmap** - Maps file directly into memory, no copies
+2. **Zero-copy to GPU** - MLX reads directly from mapped pages
+3. **No custom code** - Standard MLX API, battle-tested
+
+Without the header, we'd need custom C++ code to create MLX arrays from raw mmap'd data. MLX's public API doesn't expose this - it always copies when creating arrays from external pointers.
+
+The overhead is negligible: 88 bytes per tensor = ~100KB total for a 13GB model (0.0007%).
+
+## Why Per-Tensor Blobs?
+
+**Deduplication**: Blobs are content-addressed by SHA256. If two models share identical tensors (same weights, dtype, shape), they share the same blob file.
+
+Example: Model A and Model B both use the same text encoder. The text encoder's 400 tensors are stored once, referenced by both manifests.
+
+```
+~/.ollama/models/
+  blobs/
+    sha256-25b36eed...  <- shared by both models
+    sha256-abc123...
+  manifests/
+    library/model-a/latest  <- references sha256-25b36eed
+    library/model-b/latest  <- references sha256-25b36eed
+```
+
+## Import Flow
+
+```
+cd ./weights/Z-Image-Turbo
+ollama create z-image
+
+1. Scan component directories (text_encoder/, transformer/, vae/)
+2. For each .safetensors file:
+   - Extract individual tensors
+   - Wrap each in minimal safetensors format (88B header + data)
+   - Write to blob store (SHA256 content-addressed)
+   - Add layer entry to manifest with path-style name
+3. Copy config files (*.json) as config layers
+4. Write manifest
+```
+
+## FP8 Quantization
+
+Z-Image supports FP8 quantization to reduce memory usage by ~50% while maintaining image quality.
+
+### Usage
+
+```bash
+cd ./weights/Z-Image-Turbo
+ollama create z-image-fp8 --quantize fp8
+```
+
+This quantizes weights during import. The resulting model will be ~15GB instead of ~31GB.
+

x/imagegen/api/handler.go

@@ -0,0 +1,231 @@
+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

@@ -0,0 +1,31 @@
+// 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/cache/cache.go

@@ -0,0 +1,156 @@
+//go:build mlx
+
+package cache
+
+import "github.com/ollama/ollama/x/imagegen/mlx"
+
+type Cache interface {
+	Update(k, v *mlx.Array, seqLen int) (*mlx.Array, *mlx.Array)
+	Offset() int
+	Len() int
+	State() []*mlx.Array
+}
+
+type KVCache struct {
+	keys, values *mlx.Array
+	offset       int
+	step         int
+}
+
+func NewKVCache() *KVCache {
+	return &KVCache{step: 256}
+}
+
+func (c *KVCache) Update(k, v *mlx.Array, seqLen int) (*mlx.Array, *mlx.Array) {
+	prev := c.offset
+	shape := k.Shape()
+	B, H, Dk := shape[0], shape[1], shape[3]
+	Dv := v.Shape()[3]
+
+	// Grow buffer if needed
+	if c.keys == nil || (prev+seqLen) > int(c.keys.Shape()[2]) {
+		nSteps := (c.step + seqLen - 1) / c.step
+		newK := mlx.Zeros([]int32{B, H, int32(nSteps * c.step), Dk}, k.Dtype())
+		newV := mlx.Zeros([]int32{B, H, int32(nSteps * c.step), Dv}, v.Dtype())
+
+		if c.keys != nil {
+			if prev%c.step != 0 {
+				c.keys = mlx.Slice(c.keys, []int32{0, 0, 0, 0}, []int32{B, H, int32(prev), Dk})
+				c.values = mlx.Slice(c.values, []int32{0, 0, 0, 0}, []int32{B, H, int32(prev), Dv})
+			}
+			c.keys = mlx.Concatenate([]*mlx.Array{c.keys, newK}, 2)
+			c.values = mlx.Concatenate([]*mlx.Array{c.values, newV}, 2)
+		} else {
+			c.keys, c.values = newK, newV
+		}
+	}
+
+	c.offset += seqLen
+	c.keys = mlx.SliceUpdateInplace(c.keys, k, []int32{0, 0, int32(prev), 0}, []int32{B, H, int32(c.offset), Dk})
+	c.values = mlx.SliceUpdateInplace(c.values, v, []int32{0, 0, int32(prev), 0}, []int32{B, H, int32(c.offset), Dv})
+
+	return mlx.Slice(c.keys, []int32{0, 0, 0, 0}, []int32{B, H, int32(c.offset), Dk}),
+		mlx.Slice(c.values, []int32{0, 0, 0, 0}, []int32{B, H, int32(c.offset), Dv})
+}
+
+func (c *KVCache) State() []*mlx.Array {
+	if c.keys == nil {
+		return nil
+	}
+	return []*mlx.Array{c.keys, c.values}
+}
+
+func (c *KVCache) Offset() int { return c.offset }
+func (c *KVCache) Len() int    { return c.offset }
+
+// RotatingKVCache implements sliding window attention with bounded memory
+type RotatingKVCache struct {
+	keys, values *mlx.Array
+	offset       int
+	maxSize      int
+	step         int
+	idx          int
+}
+
+func NewRotatingKVCache(maxSize int) *RotatingKVCache {
+	return &RotatingKVCache{maxSize: maxSize, step: 256}
+}
+
+func (c *RotatingKVCache) Update(k, v *mlx.Array, seqLen int) (*mlx.Array, *mlx.Array) {
+	if seqLen > 1 {
+		return c.updateConcat(k, v, seqLen)
+	}
+	return c.updateInPlace(k, v)
+}
+
+func (c *RotatingKVCache) updateInPlace(k, v *mlx.Array) (*mlx.Array, *mlx.Array) {
+	shape := k.Shape()
+	B, H, Dk := shape[0], shape[1], shape[3]
+	Dv := v.Shape()[3]
+
+	// Grow buffer if not yet at max
+	if c.keys == nil || (c.idx >= int(c.keys.Shape()[2]) && int(c.keys.Shape()[2]) < c.maxSize) {
+		var cap int
+		if c.keys != nil {
+			cap = int(c.keys.Shape()[2])
+		}
+		newSize := min(c.step, c.maxSize-cap)
+		newK := mlx.Zeros([]int32{B, H, int32(newSize), Dk}, k.Dtype())
+		newV := mlx.Zeros([]int32{B, H, int32(newSize), Dv}, v.Dtype())
+		if c.keys != nil {
+			c.keys = mlx.Concatenate([]*mlx.Array{c.keys, newK}, 2)
+			c.values = mlx.Concatenate([]*mlx.Array{c.values, newV}, 2)
+		} else {
+			c.keys, c.values = newK, newV
+		}
+	}
+
+	// Rotate when hitting max
+	if c.idx >= c.maxSize {
+		c.idx = 0
+	}
+
+	c.keys = mlx.SliceUpdateInplace(c.keys, k, []int32{0, 0, int32(c.idx), 0}, []int32{B, H, int32(c.idx + 1), Dk})
+	c.values = mlx.SliceUpdateInplace(c.values, v, []int32{0, 0, int32(c.idx), 0}, []int32{B, H, int32(c.idx + 1), Dv})
+
+	c.offset++
+	c.idx++
+
+	validLen := int32(min(c.offset, c.maxSize))
+	return mlx.Slice(c.keys, []int32{0, 0, 0, 0}, []int32{B, H, validLen, Dk}),
+		mlx.Slice(c.values, []int32{0, 0, 0, 0}, []int32{B, H, validLen, Dv})
+}
+
+func (c *RotatingKVCache) updateConcat(k, v *mlx.Array, seqLen int) (*mlx.Array, *mlx.Array) {
+	shape := k.Shape()
+	B, H, Dk := shape[0], shape[1], shape[3]
+	Dv := v.Shape()[3]
+
+	if c.keys == nil {
+		c.keys, c.values = k, v
+	} else {
+		c.keys = mlx.Concatenate([]*mlx.Array{c.keys, k}, 2)
+		c.values = mlx.Concatenate([]*mlx.Array{c.values, v}, 2)
+	}
+	c.offset += seqLen
+
+	// Trim to max_size to maintain sliding window
+	cap := int(c.keys.Shape()[2])
+	if trim := cap - c.maxSize; trim > 0 {
+		c.keys = mlx.Slice(c.keys, []int32{0, 0, int32(trim), 0}, []int32{B, H, int32(cap), Dk})
+		c.values = mlx.Slice(c.values, []int32{0, 0, int32(trim), 0}, []int32{B, H, int32(cap), Dv})
+	}
+
+	c.idx = int(c.keys.Shape()[2])
+	return c.keys, c.values
+}
+
+func (c *RotatingKVCache) State() []*mlx.Array {
+	if c.keys == nil {
+		return nil
+	}
+	return []*mlx.Array{c.keys, c.values}
+}
+
+func (c *RotatingKVCache) Offset() int { return c.offset }
+func (c *RotatingKVCache) Len() int    { return min(c.offset, c.maxSize) }

x/imagegen/cache/step.go

@@ -0,0 +1,164 @@
+//go:build mlx
+
+package cache
+
+import "github.com/ollama/ollama/x/imagegen/mlx"
+
+// StepCache caches layer outputs across diffusion denoising steps.
+// Based on DeepCache (CVPR 2024) and Learning-to-Cache (NeurIPS 2024):
+// shallow layers change little between consecutive steps, so we can
+// cache their outputs and skip recomputation on non-refresh steps.
+//
+// Supports both single-stream (Z-Image) and dual-stream (Qwen-Image) architectures:
+//   - Single-stream: use Get/Set for the single output per layer
+//   - Dual-stream: use Get/Set for stream 1 (imgH), Get2/Set2 for stream 2 (txtH)
+//
+// Usage (single-stream):
+//
+//	cache := NewStepCache(15)  // cache first 15 layers
+//	for step := 0; step < numSteps; step++ {
+//	    refresh := cache.ShouldRefresh(step, 3)  // refresh every 3 steps
+//	    for i, layer := range layers {
+//	        if i < 15 && !refresh && cache.Get(i) != nil {
+//	            output = cache.Get(i)  // reuse cached
+//	        } else {
+//	            output = layer.Forward(input)
+//	            if i < 15 && refresh {
+//	                cache.Set(i, output)
+//	            }
+//	        }
+//	    }
+//	}
+//	cache.Free()  // cleanup when done
+//
+// Usage (dual-stream):
+//
+//	cache := NewStepCache(15)
+//	for step := 0; step < numSteps; step++ {
+//	    refresh := cache.ShouldRefresh(step, 3)
+//	    for i, layer := range layers {
+//	        if i < 15 && !refresh && cache.Get(i) != nil {
+//	            imgH, txtH = cache.Get(i), cache.Get2(i)
+//	        } else {
+//	            imgH, txtH = layer.Forward(imgH, txtH, ...)
+//	            if i < 15 && refresh {
+//	                cache.Set(i, imgH)
+//	                cache.Set2(i, txtH)
+//	            }
+//	        }
+//	    }
+//	}
+type StepCache struct {
+	layers   []*mlx.Array // cached layer outputs (stream 1)
+	layers2  []*mlx.Array // cached layer outputs (stream 2, for dual-stream models)
+	constant *mlx.Array   // optional constant (e.g., text embeddings)
+}
+
+// NewStepCache creates a cache for the given number of layers.
+func NewStepCache(numLayers int) *StepCache {
+	return &StepCache{
+		layers:  make([]*mlx.Array, numLayers),
+		layers2: make([]*mlx.Array, numLayers),
+	}
+}
+
+// ShouldRefresh returns true if the cache should be refreshed at this step.
+// Refresh happens on step 0, interval, 2*interval, etc.
+func (c *StepCache) ShouldRefresh(step, interval int) bool {
+	return step%interval == 0
+}
+
+// Get returns the cached output for a layer, or nil if not cached.
+func (c *StepCache) Get(layer int) *mlx.Array {
+	if layer < len(c.layers) {
+		return c.layers[layer]
+	}
+	return nil
+}
+
+// Set stores a layer output (stream 1), freeing any previous value.
+func (c *StepCache) Set(layer int, arr *mlx.Array) {
+	if layer < len(c.layers) {
+		if c.layers[layer] != nil {
+			c.layers[layer].Free()
+		}
+		c.layers[layer] = arr
+	}
+}
+
+// Get2 returns the cached output for a layer (stream 2), or nil if not cached.
+// Used for dual-stream architectures like Qwen-Image.
+func (c *StepCache) Get2(layer int) *mlx.Array {
+	if layer < len(c.layers2) {
+		return c.layers2[layer]
+	}
+	return nil
+}
+
+// Set2 stores a layer output (stream 2), freeing any previous value.
+// Used for dual-stream architectures like Qwen-Image.
+func (c *StepCache) Set2(layer int, arr *mlx.Array) {
+	if layer < len(c.layers2) {
+		if c.layers2[layer] != nil {
+			c.layers2[layer].Free()
+		}
+		c.layers2[layer] = arr
+	}
+}
+
+// GetConstant returns the cached constant value.
+func (c *StepCache) GetConstant() *mlx.Array {
+	return c.constant
+}
+
+// SetConstant stores a constant value, freeing any previous value.
+func (c *StepCache) SetConstant(arr *mlx.Array) {
+	if c.constant != nil {
+		c.constant.Free()
+	}
+	c.constant = arr
+}
+
+// Arrays returns all non-nil cached arrays (for pool.Keep).
+func (c *StepCache) Arrays() []*mlx.Array {
+	var result []*mlx.Array
+	if c.constant != nil {
+		result = append(result, c.constant)
+	}
+	for _, arr := range c.layers {
+		if arr != nil {
+			result = append(result, arr)
+		}
+	}
+	for _, arr := range c.layers2 {
+		if arr != nil {
+			result = append(result, arr)
+		}
+	}
+	return result
+}
+
+// Free releases all cached arrays. Call when generation completes.
+func (c *StepCache) Free() {
+	if c.constant != nil {
+		c.constant.Free()
+		c.constant = nil
+	}
+	for i, arr := range c.layers {
+		if arr != nil {
+			arr.Free()
+			c.layers[i] = nil
+		}
+	}
+	for i, arr := range c.layers2 {
+		if arr != nil {
+			arr.Free()
+			c.layers2[i] = nil
+		}
+	}
+}
+
+// NumLayers returns the number of layers this cache can store.
+func (c *StepCache) NumLayers() int {
+	return len(c.layers)
+}

x/imagegen/cache/teacache.go

@@ -0,0 +1,197 @@
+//go:build mlx
+
+// Package cache provides caching mechanisms for diffusion model inference.
+package cache
+
+import (
+	"github.com/ollama/ollama/x/imagegen/mlx"
+)
+
+// TeaCache implements Timestep Embedding Aware Caching for diffusion models.
+// It caches the transformer output and reuses it when timestep values
+// are similar between consecutive steps.
+//
+// For CFG (classifier-free guidance), it caches pos and neg predictions
+// separately and always computes CFG fresh to avoid error amplification.
+//
+// Reference: "Timestep Embedding Tells: It's Time to Cache for Video Diffusion Model"
+// https://github.com/ali-vilab/TeaCache
+type TeaCache struct {
+	// Cached transformer output from last computed step (non-CFG mode)
+	cachedOutput *mlx.Array
+
+	// Cached CFG outputs (pos and neg separately)
+	cachedPosOutput *mlx.Array
+	cachedNegOutput *mlx.Array
+
+	// Previous timestep value for difference calculation
+	prevTimestep float32
+
+	// Accumulated difference for rescaling
+	accumulatedDiff float32
+
+	// Configuration
+	threshold      float32 // Threshold for recomputation decision
+	rescaleFactor  float32 // Model-specific rescaling factor
+	skipEarlySteps int     // Number of early steps to never cache
+
+	// Statistics
+	cacheHits   int
+	cacheMisses int
+}
+
+// TeaCacheConfig holds configuration for TeaCache.
+type TeaCacheConfig struct {
+	// Threshold for recomputation. Lower = more cache hits, potential quality loss.
+	// Recommended: 0.05-0.15 for image models
+	Threshold float32
+
+	// Rescale factor to adjust timestep embedding differences.
+	// Model-specific, typically 1.0-2.0
+	RescaleFactor float32
+
+	// SkipEarlySteps: number of early steps to always compute (never cache).
+	// Set to 2-3 for CFG mode to preserve structure. 0 = no skipping.
+	SkipEarlySteps int
+}
+
+// DefaultTeaCacheConfig returns default configuration for TeaCache.
+func DefaultTeaCacheConfig() *TeaCacheConfig {
+	return &TeaCacheConfig{
+		Threshold:     0.1,
+		RescaleFactor: 1.0,
+	}
+}
+
+// NewTeaCache creates a new TeaCache instance.
+func NewTeaCache(cfg *TeaCacheConfig) *TeaCache {
+	if cfg == nil {
+		cfg = DefaultTeaCacheConfig()
+	}
+	return &TeaCache{
+		threshold:      cfg.Threshold,
+		rescaleFactor:  cfg.RescaleFactor,
+		skipEarlySteps: cfg.SkipEarlySteps,
+	}
+}
+
+// ShouldCompute determines if we should compute the full forward pass
+// or reuse the cached output based on timestep similarity.
+//
+// Algorithm:
+// 1. First step always computes
+// 2. Subsequent steps compare |currTimestep - prevTimestep| * rescaleFactor
+// 3. If accumulated difference > threshold, compute new output
+// 4. Otherwise, reuse cached output
+func (tc *TeaCache) ShouldCompute(step int, timestep float32) bool {
+	// Always compute early steps (critical for structure)
+	// Check both regular cache and CFG cache
+	hasCachedOutput := tc.cachedOutput != nil || tc.HasCFGCache()
+	if step < tc.skipEarlySteps || step == 0 || !hasCachedOutput {
+		return true
+	}
+
+	// Compute absolute difference between current and previous timestep
+	diff := timestep - tc.prevTimestep
+	if diff < 0 {
+		diff = -diff
+	}
+
+	// Apply rescaling factor
+	scaledDiff := diff * tc.rescaleFactor
+
+	// Accumulate difference (helps track drift over multiple cached steps)
+	tc.accumulatedDiff += scaledDiff
+
+	// Decision based on accumulated difference
+	if tc.accumulatedDiff > tc.threshold {
+		tc.accumulatedDiff = 0 // Reset accumulator
+		return true
+	}
+
+	return false
+}
+
+// UpdateCache stores the computed output for potential reuse (non-CFG mode).
+func (tc *TeaCache) UpdateCache(output *mlx.Array, timestep float32) {
+	// Free previous cached output
+	if tc.cachedOutput != nil {
+		tc.cachedOutput.Free()
+	}
+
+	// Store new cached values
+	tc.cachedOutput = output
+	tc.prevTimestep = timestep
+	tc.cacheMisses++
+}
+
+// UpdateCFGCache stores pos and neg outputs separately for CFG mode.
+// This allows CFG to be computed fresh each step, avoiding error amplification.
+func (tc *TeaCache) UpdateCFGCache(posOutput, negOutput *mlx.Array, timestep float32) {
+	// Free previous cached outputs
+	if tc.cachedPosOutput != nil {
+		tc.cachedPosOutput.Free()
+	}
+	if tc.cachedNegOutput != nil {
+		tc.cachedNegOutput.Free()
+	}
+
+	// Store new cached values
+	tc.cachedPosOutput = posOutput
+	tc.cachedNegOutput = negOutput
+	tc.prevTimestep = timestep
+	tc.cacheMisses++
+}
+
+// GetCached returns the cached output (non-CFG mode).
+func (tc *TeaCache) GetCached() *mlx.Array {
+	tc.cacheHits++
+	return tc.cachedOutput
+}
+
+// GetCFGCached returns cached pos and neg outputs for CFG mode.
+func (tc *TeaCache) GetCFGCached() (pos, neg *mlx.Array) {
+	tc.cacheHits++
+	return tc.cachedPosOutput, tc.cachedNegOutput
+}
+
+// HasCFGCache returns true if CFG cache is available.
+func (tc *TeaCache) HasCFGCache() bool {
+	return tc.cachedPosOutput != nil && tc.cachedNegOutput != nil
+}
+
+// Arrays returns all arrays that should be kept alive.
+func (tc *TeaCache) Arrays() []*mlx.Array {
+	var arrays []*mlx.Array
+	if tc.cachedOutput != nil {
+		arrays = append(arrays, tc.cachedOutput)
+	}
+	if tc.cachedPosOutput != nil {
+		arrays = append(arrays, tc.cachedPosOutput)
+	}
+	if tc.cachedNegOutput != nil {
+		arrays = append(arrays, tc.cachedNegOutput)
+	}
+	return arrays
+}
+
+// Stats returns cache hit/miss statistics.
+func (tc *TeaCache) Stats() (hits, misses int) {
+	return tc.cacheHits, tc.cacheMisses
+}
+
+// Free releases all cached arrays.
+func (tc *TeaCache) Free() {
+	if tc.cachedOutput != nil {
+		tc.cachedOutput.Free()
+		tc.cachedOutput = nil
+	}
+	if tc.cachedPosOutput != nil {
+		tc.cachedPosOutput.Free()
+		tc.cachedPosOutput = nil
+	}
+	if tc.cachedNegOutput != nil {
+		tc.cachedNegOutput.Free()
+		tc.cachedNegOutput = nil
+	}
+}

x/imagegen/cli.go

@@ -0,0 +1,541 @@
+// cli.go provides CLI commands for image generation models.
+//
+// TODO (jmorganca): Integrate these commands into cmd/cmd.go when stable.
+// Currently these are separate to keep experimental code isolated.
+
+package imagegen
+
+import (
+	"encoding/base64"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"io"
+	"os"
+	"strconv"
+	"strings"
+	"time"
+
+	"github.com/spf13/cobra"
+
+	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/envconfig"
+	"github.com/ollama/ollama/progress"
+	"github.com/ollama/ollama/readline"
+)
+
+// ImageGenOptions holds options for image generation.
+// These can be set via environment variables or interactive commands.
+type ImageGenOptions struct {
+	Width          int
+	Height         int
+	Steps          int
+	Seed           int
+	NegativePrompt string
+}
+
+// DefaultOptions returns the default image generation options.
+func DefaultOptions() ImageGenOptions {
+	return ImageGenOptions{
+		Width:  1024,
+		Height: 1024,
+		Steps:  9,
+		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("seed", 0, "Random seed (0 for random)")
+	cmd.Flags().String("negative", "", "Negative prompt")
+	cmd.Flags().MarkHidden("width")
+	cmd.Flags().MarkHidden("height")
+	cmd.Flags().MarkHidden("steps")
+	cmd.Flags().MarkHidden("seed")
+	cmd.Flags().MarkHidden("negative")
+}
+
+// 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
+func RunCLI(cmd *cobra.Command, name string, prompt string, interactive bool, keepAlive *api.Duration) error {
+	// Verify it's a valid image gen model
+	if ResolveModelName(name) == "" {
+		return fmt.Errorf("unknown image generation model: %s", name)
+	}
+
+	// Get options from flags (with env var defaults)
+	opts := DefaultOptions()
+	if cmd != nil && cmd.Flags() != nil {
+		if v, err := cmd.Flags().GetInt("width"); err == nil && v > 0 {
+			opts.Width = v
+		}
+		if v, err := cmd.Flags().GetInt("height"); err == nil && v > 0 {
+			opts.Height = v
+		}
+		if v, err := cmd.Flags().GetInt("steps"); err == nil && v > 0 {
+			opts.Steps = v
+		}
+		if v, err := cmd.Flags().GetInt("seed"); err == nil && v != 0 {
+			opts.Seed = v
+		}
+		if v, err := cmd.Flags().GetString("negative"); err == nil && v != "" {
+			opts.NegativePrompt = v
+		}
+	}
+
+	if interactive {
+		return runInteractive(cmd, name, keepAlive, opts)
+	}
+
+	// One-shot generation
+	return generateImageWithOptions(cmd, name, prompt, keepAlive, opts)
+}
+
+// generateImageWithOptions generates an image with the given options.
+func generateImageWithOptions(cmd *cobra.Command, modelName, prompt string, keepAlive *api.Duration, opts ImageGenOptions) error {
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		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,
+		},
+	}
+	if keepAlive != nil {
+		req.KeepAlive = keepAlive
+	}
+
+	// Show loading spinner until generation starts
+	p := progress.NewProgress(os.Stderr)
+	spinner := progress.NewSpinner("")
+	p.Add("", spinner)
+
+	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 {
+				spinner.Stop()
+				stepBar = progress.NewStepBar("Generating", total)
+				p.Add("", stepBar)
+			}
+			if stepBar != nil {
+				stepBar.Set(step)
+			}
+			return nil
+		}
+
+		// Handle final response with base64 image data
+		if resp.Done && strings.HasPrefix(content, "IMAGE_BASE64:") {
+			imageBase64 = content[13:]
+		}
+
+		return nil
+	})
+
+	p.Stop()
+	if err != nil {
+		return err
+	}
+
+	if imageBase64 != "" {
+		// Decode base64 and save to CWD
+		imageData, err := base64.StdEncoding.DecodeString(imageBase64)
+		if err != nil {
+			return fmt.Errorf("failed to decode image: %w", err)
+		}
+
+		// Create filename from prompt
+		safeName := sanitizeFilename(prompt)
+		if len(safeName) > 50 {
+			safeName = safeName[:50]
+		}
+		timestamp := time.Now().Format("20060102-150405")
+		filename := fmt.Sprintf("%s-%s.png", safeName, timestamp)
+
+		if err := os.WriteFile(filename, imageData, 0o644); err != nil {
+			return fmt.Errorf("failed to save image: %w", err)
+		}
+
+		displayImageInTerminal(filename)
+		fmt.Printf("Image saved to: %s\n", filename)
+	}
+
+	return nil
+}
+
+// runInteractive runs an interactive REPL for image generation.
+func runInteractive(cmd *cobra.Command, modelName string, keepAlive *api.Duration, opts ImageGenOptions) error {
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		return err
+	}
+
+	scanner, err := readline.New(readline.Prompt{
+		Prompt:      ">>> ",
+		Placeholder: "Describe an image to generate (/help for commands)",
+	})
+	if err != nil {
+		return err
+	}
+
+	if envconfig.NoHistory() {
+		scanner.HistoryDisable()
+	}
+
+	for {
+		line, err := scanner.Readline()
+		switch {
+		case errors.Is(err, io.EOF):
+			fmt.Println()
+			return nil
+		case errors.Is(err, readline.ErrInterrupt):
+			if line == "" {
+				fmt.Println("\nUse Ctrl + d or /bye to exit.")
+			}
+			continue
+		case err != nil:
+			return err
+		}
+
+		line = strings.TrimSpace(line)
+		if line == "" {
+			continue
+		}
+
+		// Handle commands
+		switch {
+		case strings.HasPrefix(line, "/bye"):
+			return nil
+		case strings.HasPrefix(line, "/?"), strings.HasPrefix(line, "/help"):
+			printInteractiveHelp(opts)
+			continue
+		case strings.HasPrefix(line, "/set "):
+			if err := handleSetCommand(line[5:], &opts); err != nil {
+				fmt.Fprintf(os.Stderr, "Error: %v\n", err)
+			}
+			continue
+		case strings.HasPrefix(line, "/show"):
+			printCurrentSettings(opts)
+			continue
+		case strings.HasPrefix(line, "/"):
+			fmt.Fprintf(os.Stderr, "Unknown command: %s (try /help)\n", line)
+			continue
+		}
+
+		// Generate image with current options
+		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,
+			},
+		}
+		if keepAlive != nil {
+			req.KeepAlive = keepAlive
+		}
+
+		// Show loading spinner until generation starts
+		p := progress.NewProgress(os.Stderr)
+		spinner := progress.NewSpinner("")
+		p.Add("", spinner)
+
+		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 {
+					spinner.Stop()
+					stepBar = progress.NewStepBar("Generating", total)
+					p.Add("", stepBar)
+				}
+				if stepBar != nil {
+					stepBar.Set(step)
+				}
+				return nil
+			}
+
+			// Handle final response with base64 image data
+			if resp.Done && strings.HasPrefix(content, "IMAGE_BASE64:") {
+				imageBase64 = content[13:]
+			}
+
+			return nil
+		})
+
+		p.Stop()
+		if err != nil {
+			fmt.Fprintf(os.Stderr, "Error: %v\n", err)
+			continue
+		}
+
+		// Save image to current directory with descriptive name
+		if imageBase64 != "" {
+			// Decode base64 image data
+			imageData, err := base64.StdEncoding.DecodeString(imageBase64)
+			if err != nil {
+				fmt.Fprintf(os.Stderr, "Error decoding image: %v\n", err)
+				continue
+			}
+
+			// Create filename from prompt (sanitized)
+			safeName := sanitizeFilename(line)
+			if len(safeName) > 50 {
+				safeName = safeName[:50]
+			}
+			timestamp := time.Now().Format("20060102-150405")
+			filename := fmt.Sprintf("%s-%s.png", safeName, timestamp)
+
+			if err := os.WriteFile(filename, imageData, 0o644); err != nil {
+				fmt.Fprintf(os.Stderr, "Error saving image: %v\n", err)
+				continue
+			}
+
+			displayImageInTerminal(filename)
+			fmt.Printf("Image saved to: %s\n", filename)
+		}
+
+		fmt.Println()
+	}
+}
+
+// sanitizeFilename removes characters that aren't safe for filenames.
+func sanitizeFilename(s string) string {
+	s = strings.ToLower(s)
+	s = strings.ReplaceAll(s, " ", "-")
+	// Remove any character that's not alphanumeric or hyphen
+	var result strings.Builder
+	for _, r := range s {
+		if (r >= 'a' && r <= 'z') || (r >= '0' && r <= '9') || r == '-' {
+			result.WriteRune(r)
+		}
+	}
+	return result.String()
+}
+
+// printInteractiveHelp prints help for interactive mode commands.
+func printInteractiveHelp(opts ImageGenOptions) {
+	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 negative <s>  Set negative prompt")
+	fmt.Fprintln(os.Stderr, "  /show              Show current settings")
+	fmt.Fprintln(os.Stderr, "  /bye               Exit")
+	fmt.Fprintln(os.Stderr)
+	fmt.Fprintln(os.Stderr, "Or type a prompt to generate an image.")
+	fmt.Fprintln(os.Stderr)
+}
+
+// printCurrentSettings prints the current image generation settings.
+func printCurrentSettings(opts ImageGenOptions) {
+	fmt.Fprintf(os.Stderr, "Current settings:\n")
+	fmt.Fprintf(os.Stderr, "  width:    %d\n", opts.Width)
+	fmt.Fprintf(os.Stderr, "  height:   %d\n", opts.Height)
+	fmt.Fprintf(os.Stderr, "  steps:    %d\n", opts.Steps)
+	fmt.Fprintf(os.Stderr, "  seed:     %d (0=random)\n", opts.Seed)
+	if opts.NegativePrompt != "" {
+		fmt.Fprintf(os.Stderr, "  negative: %s\n", opts.NegativePrompt)
+	}
+	fmt.Fprintln(os.Stderr)
+}
+
+// handleSetCommand handles /set commands to change options.
+func handleSetCommand(args string, opts *ImageGenOptions) error {
+	parts := strings.SplitN(args, " ", 2)
+	if len(parts) < 2 {
+		return fmt.Errorf("usage: /set <option> <value>")
+	}
+
+	key := strings.ToLower(parts[0])
+	value := strings.TrimSpace(parts[1])
+
+	switch key {
+	case "width", "w":
+		v, err := strconv.Atoi(value)
+		if err != nil || v <= 0 {
+			return fmt.Errorf("width must be a positive integer")
+		}
+		opts.Width = v
+		fmt.Fprintf(os.Stderr, "Set width to %d\n", v)
+	case "height", "h":
+		v, err := strconv.Atoi(value)
+		if err != nil || v <= 0 {
+			return fmt.Errorf("height must be a positive integer")
+		}
+		opts.Height = v
+		fmt.Fprintf(os.Stderr, "Set height to %d\n", v)
+	case "steps", "s":
+		v, err := strconv.Atoi(value)
+		if err != nil || v <= 0 {
+			return fmt.Errorf("steps must be a positive integer")
+		}
+		opts.Steps = v
+		fmt.Fprintf(os.Stderr, "Set steps to %d\n", v)
+	case "seed":
+		v, err := strconv.Atoi(value)
+		if err != nil {
+			return fmt.Errorf("seed must be an integer")
+		}
+		opts.Seed = v
+		fmt.Fprintf(os.Stderr, "Set seed to %d\n", v)
+	case "negative", "neg", "n":
+		opts.NegativePrompt = value
+		if value == "" {
+			fmt.Fprintln(os.Stderr, "Cleared negative prompt")
+		} else {
+			fmt.Fprintf(os.Stderr, "Set negative prompt to: %s\n", value)
+		}
+	default:
+		return fmt.Errorf("unknown option: %s (try /help)", key)
+	}
+	return nil
+}
+
+// displayImageInTerminal attempts to render an image inline in the terminal.
+// Supports iTerm2, Kitty, WezTerm, Ghostty, and other terminals with inline image support.
+// Returns true if the image was displayed, false otherwise.
+func displayImageInTerminal(imagePath string) bool {
+	// Check if terminal supports inline images
+	termProgram := os.Getenv("TERM_PROGRAM")
+	kittyWindowID := os.Getenv("KITTY_WINDOW_ID")
+	weztermPane := os.Getenv("WEZTERM_PANE")
+	ghostty := os.Getenv("GHOSTTY_RESOURCES_DIR")
+
+	// Read the image file
+	data, err := os.ReadFile(imagePath)
+	if err != nil {
+		return false
+	}
+
+	encoded := base64.StdEncoding.EncodeToString(data)
+
+	switch {
+	case termProgram == "iTerm.app" || termProgram == "WezTerm" || weztermPane != "":
+		// iTerm2/WezTerm inline image protocol
+		// ESC ] 1337 ; File = [arguments] : base64 BEL
+		fmt.Printf("\033]1337;File=inline=1;preserveAspectRatio=1:%s\a\n", encoded)
+		return true
+
+	case kittyWindowID != "" || ghostty != "" || termProgram == "ghostty":
+		// Kitty graphics protocol (also used by Ghostty)
+		// Send in chunks for large images
+		const chunkSize = 4096
+		for i := 0; i < len(encoded); i += chunkSize {
+			end := i + chunkSize
+			if end > len(encoded) {
+				end = len(encoded)
+			}
+			chunk := encoded[i:end]
+
+			if i == 0 {
+				// First chunk: a=T (transmit), f=100 (PNG), m=1 (more chunks follow) or m=0 (last chunk)
+				more := 1
+				if end >= len(encoded) {
+					more = 0
+				}
+				fmt.Printf("\033_Ga=T,f=100,m=%d;%s\033\\", more, chunk)
+			} else if end >= len(encoded) {
+				// Last chunk
+				fmt.Printf("\033_Gm=0;%s\033\\", chunk)
+			} else {
+				// Middle chunk
+				fmt.Printf("\033_Gm=1;%s\033\\", chunk)
+			}
+		}
+		fmt.Println()
+		return true
+
+	default:
+		return false
+	}
+}

x/imagegen/client/create.go

@@ -0,0 +1,190 @@
+// 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/client/quantize.go

@@ -0,0 +1,120 @@
+//go:build mlx
+
+package client
+
+import (
+	"fmt"
+	"io"
+	"os"
+	"path/filepath"
+
+	"github.com/ollama/ollama/x/imagegen/mlx"
+)
+
+// quantizeTensor loads a tensor from safetensors format, quantizes it to affine int8,
+// and returns safetensors data for the quantized weights, scales, and biases.
+// 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) {
+	tmpDir := ensureTempDir()
+
+	// Read safetensors data to a temp file (LoadSafetensorsNative needs a path)
+	tmpFile, err := os.CreateTemp(tmpDir, "quant-input-*.safetensors")
+	if err != nil {
+		return nil, nil, nil, nil, nil, nil, fmt.Errorf("failed to create temp file: %w", err)
+	}
+	tmpPath := tmpFile.Name()
+	defer os.Remove(tmpPath)
+
+	if _, err := io.Copy(tmpFile, r); err != nil {
+		tmpFile.Close()
+		return nil, nil, nil, nil, nil, nil, fmt.Errorf("failed to write temp file: %w", err)
+	}
+	tmpFile.Close()
+
+	// Load the tensor using MLX's native loader
+	st, err := mlx.LoadSafetensorsNative(tmpPath)
+	if err != nil {
+		return nil, nil, nil, nil, nil, nil, fmt.Errorf("failed to load safetensors: %w", err)
+	}
+	defer st.Free()
+
+	// Get the tensor (it's stored as "data" in our minimal safetensors format)
+	arr := st.Get("data")
+	if arr == nil {
+		return nil, nil, nil, nil, nil, nil, fmt.Errorf("tensor 'data' not found in safetensors")
+	}
+
+	// Convert to BFloat16 if needed (quantize expects float type)
+	if arr.Dtype() != mlx.DtypeBFloat16 && arr.Dtype() != mlx.DtypeFloat32 && arr.Dtype() != mlx.DtypeFloat16 {
+		arr = mlx.AsType(arr, mlx.DtypeBFloat16)
+		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")
+
+	// Eval and make contiguous for data access
+	qweight = mlx.Contiguous(qweight)
+	scales = mlx.Contiguous(scales)
+	if qbiases != nil {
+		qbiases = mlx.Contiguous(qbiases)
+		mlx.Eval(qweight, scales, qbiases)
+	} else {
+		mlx.Eval(qweight, scales)
+	}
+
+	// Get shapes
+	qweightShape = qweight.Shape()
+	scalesShape = scales.Shape()
+
+	// Save quantized weight using MLX's native safetensors (correctly handles uint32 dtype)
+	qweightPath := filepath.Join(tmpDir, "qweight.safetensors")
+	defer os.Remove(qweightPath)
+	if err := mlx.SaveSafetensors(qweightPath, map[string]*mlx.Array{"data": qweight}); err != nil {
+		return nil, nil, nil, nil, nil, nil, fmt.Errorf("failed to save quantized weight: %w", err)
+	}
+	qweightData, err = os.ReadFile(qweightPath)
+	if err != nil {
+		return nil, nil, nil, nil, nil, nil, fmt.Errorf("failed to read quantized weight: %w", err)
+	}
+
+	// Save scales using MLX's native safetensors
+	scalesPath := filepath.Join(tmpDir, "scales.safetensors")
+	defer os.Remove(scalesPath)
+	if err := mlx.SaveSafetensors(scalesPath, map[string]*mlx.Array{"data": scales}); err != nil {
+		return nil, nil, nil, nil, nil, nil, fmt.Errorf("failed to save scales: %w", err)
+	}
+	scalesData, err = os.ReadFile(scalesPath)
+	if err != nil {
+		return nil, nil, nil, nil, nil, nil, fmt.Errorf("failed to read scales: %w", err)
+	}
+
+	// Affine mode returns qbiases for zero-point offset
+	if qbiases != nil {
+		qbiasShape = qbiases.Shape()
+		qbiasPath := filepath.Join(tmpDir, "qbias.safetensors")
+		defer os.Remove(qbiasPath)
+		if err := mlx.SaveSafetensors(qbiasPath, map[string]*mlx.Array{"data": qbiases}); err != nil {
+			return nil, nil, nil, nil, nil, nil, fmt.Errorf("failed to save qbiases: %w", err)
+		}
+		qbiasData, err = os.ReadFile(qbiasPath)
+		if err != nil {
+			return nil, nil, nil, nil, nil, nil, fmt.Errorf("failed to read qbiases: %w", err)
+		}
+	}
+
+	return qweightData, scalesData, qbiasData, qweightShape, scalesShape, qbiasShape, nil
+}
+
+// QuantizeSupported returns true if quantization is supported (MLX build)
+func QuantizeSupported() bool {
+	return true
+}
+
+// ensureTempDir creates the temp directory for quantization if it doesn't exist
+func ensureTempDir() string {
+	tmpDir := filepath.Join(os.TempDir(), "ollama-quantize")
+	os.MkdirAll(tmpDir, 0755)
+	return tmpDir
+}

x/imagegen/client/quantize_stub.go

@@ -0,0 +1,18 @@
+//go:build !mlx
+
+package client
+
+import (
+	"fmt"
+	"io"
+)
+
+// 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) {
+	return nil, nil, nil, nil, nil, nil, fmt.Errorf("quantization requires MLX support (build with mlx tag)")
+}
+
+// QuantizeSupported returns false when MLX is not available
+func QuantizeSupported() bool {
+	return false
+}

x/imagegen/cmd/engine/README.md

@@ -0,0 +1,35 @@
+# MLX Engine
+
+Experimental MLX backend for running models on Apple Silicon and CUDA.
+
+## Build
+
+```bash
+go build -tags mlx -o engine ./x/imagegen/cmd/engine
+```
+
+## Text Generation
+
+```bash
+./engine -model /path/to/model -prompt "Hello" -max-tokens 100
+```
+
+Options:
+
+- `-temperature` - sampling temperature (default 0.7)
+- `-top-p` - nucleus sampling (default 0.9)
+- `-top-k` - top-k sampling (default 40)
+
+Supports: Llama, Gemma3, GPT-OSS
+
+## Image Generation
+
+```bash
+./engine -zimage -model /path/to/z-image -prompt "a cat" -output cat.png
+```
+
+Options:
+
+- `-width`, `-height` - image dimensions (default 1024x1024)
+- `-steps` - denoising steps (default 9)
+- `-seed` - random seed (default 42)

x/imagegen/cmd/engine/generate.go

@@ -0,0 +1,359 @@
+//go:build mlx
+
+package main
+
+import (
+	"context"
+	"fmt"
+	"time"
+	"unicode/utf8"
+
+	"github.com/ollama/ollama/x/imagegen/cache"
+	"github.com/ollama/ollama/x/imagegen/mlx"
+	"github.com/ollama/ollama/x/imagegen/tokenizer"
+)
+
+// Dedicated stream for generation (like mlx-lm's generation_stream)
+var generationStream *mlx.Stream
+
+// utf8Streamer buffers decoded text and emits only complete UTF-8 characters.
+// This handles cases where tokenizers output partial multi-byte sequences.
+type utf8Streamer struct {
+	buffer []byte
+}
+
+// Write adds decoded text to the buffer and returns complete UTF-8 characters.
+func (s *utf8Streamer) Write(text string) string {
+	s.buffer = append(s.buffer, text...)
+
+	// Find the last position that ends with a complete UTF-8 character
+	validLen := 0
+	for i := 0; i < len(s.buffer); {
+		r, size := utf8.DecodeRune(s.buffer[i:])
+		if r == utf8.RuneError && size == 1 {
+			// Invalid or incomplete UTF-8 sequence at this position
+			// Check if it could be a valid start of a multi-byte sequence
+			if len(s.buffer)-i < 4 {
+				// Might be incomplete, keep it in buffer
+				break
+			}
+			// Definitely invalid, skip this byte
+			i++
+			validLen = i
+		} else {
+			i += size
+			validLen = i
+		}
+	}
+
+	if validLen == 0 {
+		return ""
+	}
+
+	result := string(s.buffer[:validLen])
+	s.buffer = s.buffer[validLen:]
+	return result
+}
+
+// Flush returns any remaining buffered bytes (may be incomplete UTF-8).
+func (s *utf8Streamer) Flush() string {
+	if len(s.buffer) == 0 {
+		return ""
+	}
+	result := string(s.buffer)
+	s.buffer = nil
+	return result
+}
+
+func init() {
+	generationStream = mlx.NewStream()
+}
+
+// withStream runs fn with the generation stream as default
+func withStream(fn func()) {
+	orig := mlx.GetDefaultStream()
+	mlx.SetDefaultStream(generationStream)
+	fn()
+	mlx.SetDefaultStream(orig)
+}
+
+type Model interface {
+	Tokenizer() *tokenizer.Tokenizer
+	VocabSize() int32
+	NewCache(maxSeqLen int32) []cache.Cache
+	Forward(input *mlx.Array, caches []cache.Cache) *mlx.Array
+}
+
+// ChatModel is an optional interface for models that support chat formatting
+type ChatModel interface {
+	FormatPrompt(prompt string) string
+}
+
+// MultimodalModel is for models that support image input
+type MultimodalModel interface {
+	Model
+	FormatPromptWithImage(prompt string) string
+	ExpandImageTokens(tokens []int32) []int32
+	ForwardWithImage(tokens *mlx.Array, image *mlx.Array, caches []cache.Cache) *mlx.Array
+	ImageSize() int32 // Returns expected image size for preprocessing
+}
+
+// ImageLoader loads and preprocesses an image for multimodal models
+// Returns nil if path is empty
+type ImageLoader func(path string, imageSize int32) (*mlx.Array, error)
+
+type input struct {
+	Prompt       string
+	Image        *mlx.Array // Optional preprocessed image for multimodal models
+	MaxTokens    int
+	Temperature  float32
+	TopP         float32
+	TopK         int
+	WiredLimitGB int // Metal wired memory limit in GB (default 32)
+}
+
+type output struct {
+	Text          string
+	Done          bool
+	PrefillTokSec float64
+	GenTokSec     float64
+}
+
+// Decoder wraps model + cache for autoregressive generation.
+type Decoder struct {
+	model         Model
+	caches        []cache.Cache
+	vocabSize     int32
+	temp          float32
+	topK          int
+	topP          float32
+	token         *mlx.Array   // Current token (kept across pools)
+	oldCacheState []*mlx.Array // Preallocated slice for old cache state
+	image         *mlx.Array   // Optional image for multimodal prefill
+}
+
+func NewDecoder(m Model, temp float32, topK int, topP float32) *Decoder {
+	caches := m.NewCache(0)
+	return &Decoder{
+		model:         m,
+		caches:        caches,
+		vocabSize:     m.VocabSize(),
+		temp:          temp,
+		topK:          topK,
+		topP:          topP,
+		oldCacheState: make([]*mlx.Array, 0, len(caches)*2),
+	}
+}
+
+// SetImage sets the image for multimodal prefill (call before prefill)
+func (d *Decoder) SetImage(img *mlx.Array) {
+	d.image = img
+}
+
+func (d *Decoder) prefill(inputIDs []int32) int {
+	processed := 0
+
+	// Track old cache state to free after each chunk
+	var oldCacheState []*mlx.Array
+
+	// For multimodal models with an image, we need to process all tokens together
+	// in the first forward pass so the image embeddings can be inserted properly.
+	// Skip chunking for multimodal prefill.
+	isMultimodal := d.image != nil
+
+	// Process all-but-1 tokens in chunks, eval cache state for memory management
+	// Skip chunking for multimodal - process everything in the final step
+	if !isMultimodal {
+		for len(inputIDs) > 1 {
+			chunkSize := min(2048, len(inputIDs)-1)
+			if chunkSize <= 0 {
+				break
+			}
+			chunk := inputIDs[:chunkSize]
+
+			// Save old cache state before forward
+			oldCacheState = oldCacheState[:0]
+			for _, c := range d.caches {
+				oldCacheState = append(oldCacheState, c.State()...)
+			}
+
+			var cacheState []*mlx.Array
+			withStream(func() {
+				x := mlx.NewArrayInt32(chunk, []int32{1, int32(len(chunk))})
+				d.model.Forward(x, d.caches)
+				for _, c := range d.caches {
+					cacheState = append(cacheState, c.State()...)
+				}
+			})
+			mlx.Eval(cacheState...)
+
+			// Free old cache state
+			for _, arr := range oldCacheState {
+				if arr != nil {
+					arr.Free()
+				}
+			}
+
+			inputIDs = inputIDs[chunkSize:]
+			processed += chunkSize
+		}
+	}
+
+	// Save old cache state before final step
+	oldCacheState = oldCacheState[:0]
+	for _, c := range d.caches {
+		oldCacheState = append(oldCacheState, c.State()...)
+	}
+
+	// Final token + sampling (or all tokens for multimodal)
+	withStream(func() {
+		x := mlx.NewArrayInt32(inputIDs, []int32{1, int32(len(inputIDs))})
+		mlx.Eval(x) // Materialize before any other evals
+
+		var logits *mlx.Array
+		// Use ForwardWithImage if we have an image and model supports it
+		if d.image != nil {
+			if mm, ok := d.model.(MultimodalModel); ok {
+				logits = mm.ForwardWithImage(x, d.image, d.caches)
+				d.image = nil // Only use image for first forward
+			} else {
+				logits = d.model.Forward(x, d.caches)
+			}
+		} else {
+			logits = d.model.Forward(x, d.caches)
+		}
+		d.token = sample(logits, d.temp, d.topK, d.topP, d.vocabSize)
+	})
+	// Keep cache state (token auto-kept by AsyncEval)
+	for _, c := range d.caches {
+		mlx.Keep(c.State()...)
+	}
+	mlx.AsyncEval(d.token)
+
+	// Free old cache state from before final step
+	for _, arr := range oldCacheState {
+		if arr != nil {
+			arr.Free()
+		}
+	}
+
+	mlx.ClearCache()
+
+	return processed + len(inputIDs)
+}
+
+func (d *Decoder) step() int32 {
+	prevToken := d.token
+
+	// Save old cache state (reuse preallocated slice)
+	d.oldCacheState = d.oldCacheState[:0]
+	for _, c := range d.caches {
+		d.oldCacheState = append(d.oldCacheState, c.State()...)
+	}
+
+	withStream(func() {
+		logits := d.model.Forward(mlx.Reshape(prevToken, 1, 1), d.caches)
+		d.token = sample(logits, d.temp, d.topK, d.topP, d.vocabSize)
+	})
+	// Keep token and new cache state so they survive cleanup
+	mlx.Keep(d.token)
+	for _, c := range d.caches {
+		mlx.Keep(c.State()...)
+	}
+	mlx.AsyncEval(d.token)
+
+	// Sync on previous token (GPU already working on next step)
+	val := prevToken.ItemInt32()
+
+	// Free old token and old cache state
+	prevToken.Free()
+	for _, arr := range d.oldCacheState {
+		arr.Free()
+	}
+	return val
+}
+
+func generate(ctx context.Context, m Model, in input, cb func(output)) error {
+	mlx.EnableCompile()
+	wiredLimit := in.WiredLimitGB
+	if wiredLimit <= 0 {
+		wiredLimit = 32 // default 32GB
+	}
+	mlx.MetalSetWiredLimit(uint64(wiredLimit) << 30)
+
+	temp := in.Temperature
+	if temp < 0 {
+		temp = 0.7
+	}
+
+	tok := m.Tokenizer()
+	dec := NewDecoder(m, temp, in.TopK, in.TopP)
+
+	// Apply chat template - use image template if we have an image
+	prompt := in.Prompt
+	var tokens []int32
+	if mm, ok := m.(MultimodalModel); ok && in.Image != nil {
+		prompt = mm.FormatPromptWithImage(prompt)
+		tokens = tok.Encode(prompt, true)
+		tokens = mm.ExpandImageTokens(tokens) // Expand <start_of_image> to 256 image tokens
+		dec.SetImage(in.Image)
+	} else if cm, ok := m.(ChatModel); ok {
+		prompt = cm.FormatPrompt(prompt)
+		tokens = tok.Encode(prompt, true)
+	} else {
+		tokens = tok.Encode(prompt, true)
+	}
+
+	prefillStart := time.Now()
+	prefillTokens := dec.prefill(tokens)
+	// Prefill measurement should include time to first token (like mlx-lm)
+	// Step() waits for prefill to complete and returns first token
+	firstToken := dec.step()
+	prefillTokSec := float64(prefillTokens) / time.Since(prefillStart).Seconds()
+
+	genStart := time.Now()
+	maxTokens := max(in.MaxTokens, 100)
+	var genTokens int
+
+	// UTF-8 streamer to handle partial multi-byte characters
+	streamer := &utf8Streamer{}
+
+	// Handle first token
+	genTokens++
+	if tok.IsEOS(firstToken) {
+		cb(output{Done: true, PrefillTokSec: prefillTokSec, GenTokSec: 0})
+		return nil
+	}
+	if text := streamer.Write(tok.Decode([]int32{firstToken})); text != "" {
+		cb(output{Text: text})
+	}
+
+	for n := 1; n < maxTokens; n++ {
+		if ctx.Err() != nil {
+			return ctx.Err()
+		}
+		token := dec.step()
+		genTokens++
+
+		if tok.IsEOS(token) {
+			break
+		}
+		if text := streamer.Write(tok.Decode([]int32{token})); text != "" {
+			cb(output{Text: text})
+		}
+
+		if n%256 == 0 {
+			mlx.ClearCache()
+		}
+	}
+
+	// Flush any remaining buffered bytes
+	if text := streamer.Flush(); text != "" {
+		cb(output{Text: text})
+	}
+
+	fmt.Printf("\nPeak memory: %.2fGB\n", float64(mlx.MetalGetPeakMemory())/(1<<30))
+	cb(output{Done: true, PrefillTokSec: prefillTokSec,
+		GenTokSec: float64(genTokens) / time.Since(genStart).Seconds()})
+	return nil
+}

x/imagegen/cmd/engine/image.go

@@ -0,0 +1,89 @@
+//go:build mlx
+
+package main
+
+import (
+	"fmt"
+	"image"
+	"image/png"
+	"os"
+	"path/filepath"
+
+	"github.com/ollama/ollama/x/imagegen/mlx"
+)
+
+// saveImageArray saves an MLX array as a PNG image.
+// Expected format: [B, C, H, W] with values in [0, 1] range and C=3 (RGB).
+func saveImageArray(arr *mlx.Array, path string) error {
+	img, err := arrayToImage(arr)
+	if err != nil {
+		return err
+	}
+	return savePNG(img, path)
+}
+
+func savePNG(img *image.RGBA, path string) error {
+	if filepath.Ext(path) != ".png" {
+		path = path + ".png"
+	}
+	f, err := os.Create(path)
+	if err != nil {
+		return err
+	}
+	defer f.Close()
+	return png.Encode(f, img)
+}
+
+func arrayToImage(arr *mlx.Array) (*image.RGBA, error) {
+	shape := arr.Shape()
+	if len(shape) != 4 {
+		return nil, fmt.Errorf("expected 4D array [B, C, H, W], got %v", shape)
+	}
+
+	// Transform to [H, W, C] for image conversion
+	img := mlx.Squeeze(arr, 0)
+	arr.Free()
+	img = mlx.Transpose(img, 1, 2, 0)
+	img = mlx.Contiguous(img)
+	mlx.Eval(img)
+
+	imgShape := img.Shape()
+	H := int(imgShape[0])
+	W := int(imgShape[1])
+	C := int(imgShape[2])
+
+	if C != 3 {
+		img.Free()
+		return nil, fmt.Errorf("expected 3 channels (RGB), got %d", C)
+	}
+
+	// Copy to CPU and free GPU memory
+	data := img.Data()
+	img.Free()
+
+	// Write directly to Pix slice (faster than SetRGBA)
+	goImg := image.NewRGBA(image.Rect(0, 0, W, H))
+	pix := goImg.Pix
+	for y := 0; y < H; y++ {
+		for x := 0; x < W; x++ {
+			srcIdx := (y*W + x) * C
+			dstIdx := (y*W + x) * 4
+			pix[dstIdx+0] = uint8(clampF(data[srcIdx+0]*255+0.5, 0, 255))
+			pix[dstIdx+1] = uint8(clampF(data[srcIdx+1]*255+0.5, 0, 255))
+			pix[dstIdx+2] = uint8(clampF(data[srcIdx+2]*255+0.5, 0, 255))
+			pix[dstIdx+3] = 255
+		}
+	}
+
+	return goImg, nil
+}
+
+func clampF(v, min, max float32) float32 {
+	if v < min {
+		return min
+	}
+	if v > max {
+		return max
+	}
+	return v
+}

x/imagegen/cmd/engine/main.go

@@ -0,0 +1,293 @@
+//go:build mlx
+
+package main
+
+import (
+	"context"
+	"encoding/json"
+	"flag"
+	"fmt"
+	"log"
+	"os"
+	"path/filepath"
+	"runtime/pprof"
+
+	"github.com/ollama/ollama/x/imagegen/mlx"
+	"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"
+	"github.com/ollama/ollama/x/imagegen/models/qwen_image"
+	"github.com/ollama/ollama/x/imagegen/models/qwen_image_edit"
+	"github.com/ollama/ollama/x/imagegen/models/zimage"
+	"github.com/ollama/ollama/x/imagegen/safetensors"
+)
+
+// stringSlice is a flag type that accumulates multiple values
+type stringSlice []string
+
+func (s *stringSlice) String() string {
+	return fmt.Sprintf("%v", *s)
+}
+
+func (s *stringSlice) Set(value string) error {
+	*s = append(*s, value)
+	return nil
+}
+
+func main() {
+	modelPath := flag.String("model", "", "Model directory")
+	prompt := flag.String("prompt", "Hello", "Prompt")
+
+	// Text generation params
+	maxTokens := flag.Int("max-tokens", 100, "Max tokens")
+	temperature := flag.Float64("temperature", 0.7, "Temperature")
+	topP := flag.Float64("top-p", 0.9, "Top-p sampling")
+	topK := flag.Int("top-k", 40, "Top-k sampling")
+	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")
+	seed := flag.Int64("seed", 42, "Random seed")
+	out := flag.String("output", "output.png", "Output path")
+
+	// Utility flags
+	listTensors := flag.Bool("list", false, "List tensors only")
+	cpuProfile := flag.String("cpuprofile", "", "Write CPU profile to file")
+	gpuCapture := flag.String("gpu-capture", "", "Capture GPU trace to .gputrace file (run with MTL_CAPTURE_ENABLED=1)")
+	layerCache := flag.Bool("layer-cache", false, "Enable layer caching for faster diffusion (Z-Image, Qwen-Image). Not compatible with CFG/negative prompts.")
+	wiredLimitGB := flag.Int("wired-limit", 32, "Metal wired memory limit in GB")
+
+	// Legacy mode flags
+	zimageFlag := flag.Bool("zimage", false, "Z-Image 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
+	flag.Var(&inputImages, "input-image", "Input image for image editing (can be specified multiple times)")
+	negativePrompt := flag.String("negative-prompt", "", "Negative prompt for CFG (empty = no CFG, matching Python)")
+	cfgScale := flag.Float64("cfg-scale", 4.0, "CFG scale for image editing")
+	teaCache := flag.Bool("teacache", false, "Enable TeaCache for faster inference")
+	teaCacheThreshold := flag.Float64("teacache-threshold", 0.1, "TeaCache threshold (lower = more aggressive caching)")
+	fusedQKV := flag.Bool("fused-qkv", false, "Enable fused QKV projection for faster attention")
+
+	flag.Parse()
+
+	if *modelPath == "" {
+		flag.Usage()
+		return
+	}
+
+	// CPU profiling
+	if *cpuProfile != "" {
+		f, err := os.Create(*cpuProfile)
+		if err != nil {
+			log.Fatal(err)
+		}
+		defer f.Close()
+		if err := pprof.StartCPUProfile(f); err != nil {
+			log.Fatal(err)
+		}
+		defer pprof.StopCPUProfile()
+	}
+
+	var err error
+
+	// Handle legacy mode flags that aren't unified yet
+	switch {
+	case *zimageFlag:
+		m := &zimage.Model{}
+		if loadErr := m.Load(*modelPath); loadErr != nil {
+			log.Fatal(loadErr)
+		}
+		var img *mlx.Array
+		img, err = m.GenerateFromConfig(context.Background(), &zimage.GenerateConfig{
+			Prompt:            *prompt,
+			NegativePrompt:    *negativePrompt,
+			CFGScale:          float32(*cfgScale),
+			Width:             int32(*width),
+			Height:            int32(*height),
+			Steps:             *steps,
+			Seed:              *seed,
+			CapturePath:       *gpuCapture,
+			TeaCache:          *teaCache,
+			TeaCacheThreshold: float32(*teaCacheThreshold),
+			FusedQKV:          *fusedQKV,
+		})
+		if err == nil {
+			err = saveImageArray(img, *out)
+		}
+	case *qwenImage:
+		m, loadErr := qwen_image.LoadPersistent(*modelPath)
+		if loadErr != nil {
+			log.Fatal(loadErr)
+		}
+		var img *mlx.Array
+		img, err = m.GenerateFromConfig(&qwen_image.GenerateConfig{
+			Prompt:         *prompt,
+			NegativePrompt: *negativePrompt,
+			CFGScale:       float32(*cfgScale),
+			Width:          int32(*width),
+			Height:         int32(*height),
+			Steps:          *steps,
+			Seed:           *seed,
+			LayerCache:     *layerCache,
+		})
+		if err == nil {
+			err = saveImageArray(img, *out)
+		}
+	case *qwenImageEdit:
+		if len(inputImages) == 0 {
+			log.Fatal("qwen-image-edit requires at least one -input-image")
+		}
+
+		m, loadErr := qwen_image_edit.LoadPersistent(*modelPath)
+		if loadErr != nil {
+			log.Fatal(loadErr)
+		}
+		// For image editing, use 0 for dimensions to auto-detect from input image
+		// unless explicitly overridden from defaults
+		editWidth := int32(0)
+		editHeight := int32(0)
+		if *width != 1024 {
+			editWidth = int32(*width)
+		}
+		if *height != 1024 {
+			editHeight = int32(*height)
+		}
+
+		cfg := &qwen_image_edit.GenerateConfig{
+			Prompt:         *prompt,
+			NegativePrompt: *negativePrompt,
+			CFGScale:       float32(*cfgScale),
+			Width:          editWidth,
+			Height:         editHeight,
+			Steps:          *steps,
+			Seed:           *seed,
+		}
+
+		var img *mlx.Array
+		img, err = m.EditFromConfig(inputImages, cfg)
+		if err == nil {
+			err = saveImageArray(img, *out)
+		}
+	case *listTensors:
+		err = listModelTensors(*modelPath)
+	default:
+		// llm path
+		m, err := load(*modelPath)
+		if err != nil {
+			log.Fatal(err)
+		}
+
+		// Load image if provided and model supports it
+		var image *mlx.Array
+		if *imagePath != "" {
+			if mm, ok := m.(interface{ ImageSize() int32 }); ok {
+				image, err = gemma3.ProcessImage(*imagePath, mm.ImageSize())
+				if err != nil {
+					log.Fatal("load image:", err)
+				}
+			} else {
+				log.Fatal("model does not support image input")
+			}
+		}
+
+		err = generate(context.Background(), m, input{
+			Prompt:       *prompt,
+			Image:        image,
+			MaxTokens:    *maxTokens,
+			Temperature:  float32(*temperature),
+			TopP:         float32(*topP),
+			TopK:         *topK,
+			WiredLimitGB: *wiredLimitGB,
+		}, func(out output) {
+			if out.Text != "" {
+				fmt.Print(out.Text)
+			}
+			if out.Done {
+				fmt.Printf("\n\n[prefill: %.1f tok/s, gen: %.1f tok/s]\n", out.PrefillTokSec, out.GenTokSec)
+			}
+		})
+	}
+
+	if err != nil {
+		log.Fatal(err)
+	}
+}
+
+func listModelTensors(modelPath string) error {
+	weights, err := safetensors.LoadModelWeights(modelPath)
+	if err != nil {
+		return err
+	}
+	for _, name := range weights.ListTensors() {
+		info, _ := weights.GetTensorInfo(name)
+		fmt.Printf("%s: %v (%s)\n", name, info.Shape, info.Dtype)
+	}
+	return nil
+}
+
+// loadModel builds and evaluates a model using the common load pattern.
+// Release safetensors BEFORE eval - lazy arrays have captured their data,
+// and this reduces peak memory by ~6GB (matches mlx-lm behavior).
+func loadModel[T Model](build func() T, cleanup func()) T {
+	m := build()
+	weights := mlx.Collect(m)
+	cleanup()
+	mlx.Eval(weights...)
+	return m
+}
+
+func load(modelPath string) (Model, error) {
+	kind, err := detectModelKind(modelPath)
+	if err != nil {
+		return nil, fmt.Errorf("detect model kind: %w", err)
+	}
+
+	switch kind {
+	case "gpt_oss":
+		return gpt_oss.Load(modelPath)
+	case "gemma3":
+		return gemma3.Load(modelPath)
+	case "gemma3_text":
+		return gemma3.LoadText(modelPath)
+	default:
+		return llama.Load(modelPath)
+	}
+}
+
+func detectModelKind(modelPath string) (string, error) {
+	indexPath := filepath.Join(modelPath, "model_index.json")
+	if _, err := os.Stat(indexPath); err == nil {
+		data, err := os.ReadFile(indexPath)
+		if err != nil {
+			return "zimage", nil
+		}
+		var index struct {
+			ClassName string `json:"_class_name"`
+		}
+		if err := json.Unmarshal(data, &index); err == nil {
+			switch index.ClassName {
+			case "FluxPipeline", "ZImagePipeline":
+				return "zimage", nil
+			}
+		}
+		return "zimage", nil
+	}
+
+	configPath := filepath.Join(modelPath, "config.json")
+	data, err := os.ReadFile(configPath)
+	if err != nil {
+		return "", fmt.Errorf("no config.json or model_index.json found: %w", err)
+	}
+
+	var cfg struct {
+		ModelType string `json:"model_type"`
+	}
+	if err := json.Unmarshal(data, &cfg); err != nil {
+		return "", fmt.Errorf("parse config.json: %w", err)
+	}
+
+	return cfg.ModelType, nil
+}

x/imagegen/cmd/engine/sample.go

@@ -0,0 +1,49 @@
+//go:build mlx
+
+package main
+
+import "github.com/ollama/ollama/x/imagegen/mlx"
+
+// sampleTopK samples from top-k logits using global random state
+func sampleTopK(scaledLogits *mlx.Array, k int) *mlx.Array {
+	neg := mlx.Neg(scaledLogits)
+	indices := mlx.Argpartition(neg, k-1, -1)
+	topKIdx := mlx.Slice(indices, []int32{0}, []int32{int32(k)})
+	values := mlx.TakeAlongAxis(scaledLogits, topKIdx, -1)
+	sampled := mlx.RandomCategorical(values, -1, 1)
+	return mlx.Take(topKIdx, sampled, -1)
+}
+
+// sampleTopP samples using nucleus sampling with global random state
+func sampleTopP(scaledLogits *mlx.Array, p float32, vocabSize int32) *mlx.Array {
+	sorted := mlx.Argsort(mlx.Neg(scaledLogits), -1)
+	sortedLogits := mlx.TakeAlongAxis(scaledLogits, sorted, -1)
+	probs := mlx.Softmax(sortedLogits, -1)
+	cumProbs := mlx.Cumsum(probs, -1)
+	mask := mlx.LessScalar(cumProbs, p)
+	negInf := mlx.FullDtype(float32(-1e9), scaledLogits.Dtype(), vocabSize)
+	masked := mlx.Where(mask, sortedLogits, negInf)
+	sampled := mlx.RandomCategorical(masked, -1, 1)
+	return mlx.Take(sorted, sampled, -1)
+}
+
+// sample samples from logits at the last position
+func sample(logits *mlx.Array, temp float32, topK int, topP float32, vocab int32) *mlx.Array {
+	// Get last position logits: [1, L, vocab] -> [vocab]
+	shape := logits.Shape()
+	seqLen := shape[1]
+	lastLogits := mlx.Slice(logits, []int32{0, seqLen - 1, 0}, []int32{1, seqLen, vocab})
+	lastLogits = mlx.Reshape(lastLogits, vocab)
+
+	if temp == 0 {
+		return mlx.Argmax(lastLogits, -1, false)
+	}
+	scaled := mlx.DivScalar(lastLogits, temp)
+	if topK > 0 && topK < int(vocab) {
+		return sampleTopK(scaled, topK)
+	}
+	if topP > 0 && topP < 1.0 {
+		return sampleTopP(scaled, topP, vocab)
+	}
+	return mlx.RandomCategorical(scaled, -1, 1)
+}

x/imagegen/create.go

@@ -0,0 +1,213 @@
+package imagegen
+
+import (
+	"bytes"
+	"encoding/json"
+	"fmt"
+	"io"
+	"os"
+	"path/filepath"
+	"strings"
+
+	"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.
+// 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.
+// 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 {
+	var layers []LayerInfo
+	var configLayer LayerInfo
+	var totalParams int64 // Count parameters from original tensor shapes
+
+	// Components to process - extract individual tensors from each
+	components := []string{"text_encoder", "transformer", "vae"}
+
+	for _, component := range components {
+		componentDir := filepath.Join(modelDir, component)
+		if _, err := os.Stat(componentDir); os.IsNotExist(err) {
+			continue
+		}
+
+		// Find all safetensors files in this component
+		entries, err := os.ReadDir(componentDir)
+		if err != nil {
+			return fmt.Errorf("failed to read %s: %w", component, err)
+		}
+
+		for _, entry := range entries {
+			if !strings.HasSuffix(entry.Name(), ".safetensors") {
+				continue
+			}
+
+			stPath := filepath.Join(componentDir, 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 == "fp8" && component != "vae" {
+				quantizeMsg = ", quantizing to fp8"
+			}
+			fn(fmt.Sprintf("importing %s/%s (%d tensors%s)", component, 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)
+				}
+
+				// Count parameters from original tensor shape
+				if len(td.Shape) > 0 {
+					numElements := int64(1)
+					for _, dim := range td.Shape {
+						numElements *= int64(dim)
+					}
+					totalParams += numElements
+				}
+
+				// Store as minimal safetensors format (88 bytes header overhead)
+				// This enables native mmap loading via mlx_load_safetensors
+				// Use path-style name: "component/tensor_name"
+				fullName := component + "/" + tensorName
+
+				// Determine if this tensor should be quantized
+				doQuantize := quantize == "fp8" && ShouldQuantize(tensorName, component)
+
+				// createTensorLayer returns multiple layers if quantizing (weight + scales)
+				newLayers, err := createTensorLayer(td.SafetensorsReader(), fullName, td.Dtype, td.Shape, doQuantize)
+				if err != nil {
+					extractor.Close()
+					return fmt.Errorf("failed to create layer for %s: %w", fullName, err)
+				}
+				layers = append(layers, newLayers...)
+			}
+
+			extractor.Close()
+		}
+	}
+
+	// Import config files
+	configFiles := []string{
+		"model_index.json",
+		"text_encoder/config.json",
+		"text_encoder/generation_config.json",
+		"transformer/config.json",
+		"vae/config.json",
+		"scheduler/scheduler_config.json",
+		"tokenizer/tokenizer.json",
+		"tokenizer/tokenizer_config.json",
+		"tokenizer/vocab.json",
+	}
+
+	for _, cfgPath := range configFiles {
+		fullPath := filepath.Join(modelDir, cfgPath)
+		if _, err := os.Stat(fullPath); os.IsNotExist(err) {
+			continue
+		}
+
+		fn(fmt.Sprintf("importing config %s", cfgPath))
+
+		var r io.Reader
+
+		// For model_index.json, normalize to Ollama format and add metadata
+		if cfgPath == "model_index.json" {
+			data, err := os.ReadFile(fullPath)
+			if err != nil {
+				return fmt.Errorf("failed to read %s: %w", cfgPath, err)
+			}
+
+			var cfg map[string]any
+			if err := json.Unmarshal(data, &cfg); err != nil {
+				return fmt.Errorf("failed to parse %s: %w", cfgPath, err)
+			}
+
+			// Rename _class_name to architecture, remove diffusers-specific fields
+			if className, ok := cfg["_class_name"]; ok {
+				cfg["architecture"] = className
+				delete(cfg, "_class_name")
+			}
+			delete(cfg, "_diffusers_version")
+
+			// Add parameter count (counted from tensor shapes during import)
+			cfg["parameter_count"] = totalParams
+
+			// Add quantization info
+			if quantize == "fp8" {
+				cfg["quantization"] = "FP8"
+			} else {
+				cfg["quantization"] = "BF16"
+			}
+
+			data, err = json.MarshalIndent(cfg, "", "    ")
+			if err != nil {
+				return fmt.Errorf("failed to marshal %s: %w", cfgPath, err)
+			}
+			r = bytes.NewReader(data)
+		} else {
+			f, err := os.Open(fullPath)
+			if err != nil {
+				return fmt.Errorf("failed to open %s: %w", cfgPath, err)
+			}
+			defer f.Close()
+			r = f
+		}
+
+		layer, err := createLayer(r, "application/vnd.ollama.image.json", cfgPath)
+		if err != nil {
+			return fmt.Errorf("failed to create layer for %s: %w", cfgPath, err)
+		}
+
+		// Use model_index.json as the config layer
+		if cfgPath == "model_index.json" {
+			configLayer = layer
+		}
+
+		layers = append(layers, layer)
+	}
+
+	if configLayer.Digest == "" {
+		return fmt.Errorf("model_index.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/imagegen/image.go

@@ -0,0 +1,110 @@
+//go:build mlx
+
+package imagegen
+
+import (
+	"bytes"
+	"encoding/base64"
+	"fmt"
+	"image"
+	"image/png"
+	"os"
+	"path/filepath"
+
+	"github.com/ollama/ollama/x/imagegen/mlx"
+)
+
+// SaveImage saves an MLX array as a PNG image file.
+// Expected format: [B, C, H, W] with values in [0, 1] range and C=3 (RGB).
+func SaveImage(arr *mlx.Array, path string) error {
+	img, err := ArrayToImage(arr)
+	if err != nil {
+		return err
+	}
+
+	if filepath.Ext(path) != ".png" {
+		path = path + ".png"
+	}
+
+	f, err := os.Create(path)
+	if err != nil {
+		return err
+	}
+	defer f.Close()
+
+	return png.Encode(f, img)
+}
+
+// EncodeImageBase64 encodes an MLX array as a base64-encoded PNG.
+// Expected format: [B, C, H, W] with values in [0, 1] range and C=3 (RGB).
+func EncodeImageBase64(arr *mlx.Array) (string, error) {
+	img, err := ArrayToImage(arr)
+	if err != nil {
+		return "", err
+	}
+
+	var buf bytes.Buffer
+	if err := png.Encode(&buf, img); err != nil {
+		return "", err
+	}
+
+	return base64.StdEncoding.EncodeToString(buf.Bytes()), nil
+}
+
+// ArrayToImage converts an MLX array to a Go image.RGBA.
+// Expected format: [B, C, H, W] with values in [0, 1] range and C=3 (RGB).
+func ArrayToImage(arr *mlx.Array) (*image.RGBA, error) {
+	shape := arr.Shape()
+	if len(shape) != 4 {
+		return nil, fmt.Errorf("expected 4D array [B, C, H, W], got %v", shape)
+	}
+
+	// Transform to [H, W, C] for image conversion
+	// Free intermediate arrays to avoid memory leak
+	squeezed := mlx.Squeeze(arr, 0)
+	transposed := mlx.Transpose(squeezed, 1, 2, 0)
+	squeezed.Free()
+	img := mlx.Contiguous(transposed)
+	transposed.Free()
+	mlx.Eval(img)
+
+	imgShape := img.Shape()
+	H := int(imgShape[0])
+	W := int(imgShape[1])
+	C := int(imgShape[2])
+
+	if C != 3 {
+		img.Free()
+		return nil, fmt.Errorf("expected 3 channels (RGB), got %d", C)
+	}
+
+	// Copy to CPU and free GPU memory
+	data := img.Data()
+	img.Free()
+
+	// Write directly to Pix slice (faster than SetRGBA)
+	goImg := image.NewRGBA(image.Rect(0, 0, W, H))
+	pix := goImg.Pix
+	for y := 0; y < H; y++ {
+		for x := 0; x < W; x++ {
+			srcIdx := (y*W + x) * C
+			dstIdx := (y*W + x) * 4
+			pix[dstIdx+0] = uint8(clampF(data[srcIdx+0]*255+0.5, 0, 255))
+			pix[dstIdx+1] = uint8(clampF(data[srcIdx+1]*255+0.5, 0, 255))
+			pix[dstIdx+2] = uint8(clampF(data[srcIdx+2]*255+0.5, 0, 255))
+			pix[dstIdx+3] = 255
+		}
+	}
+
+	return goImg, nil
+}
+
+func clampF(v, min, max float32) float32 {
+	if v < min {
+		return min
+	}
+	if v > max {
+		return max
+	}
+	return v
+}

x/imagegen/manifest.go

@@ -0,0 +1,177 @@
+package imagegen
+
+import (
+	"encoding/json"
+	"fmt"
+	"io"
+	"os"
+	"path/filepath"
+	"runtime"
+	"strings"
+)
+
+// 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"` // Path-style name: "component/tensor" or "path/to/config.json"
+}
+
+// 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"`
+}
+
+// ModelManifest holds a parsed manifest with helper methods.
+type ModelManifest struct {
+	Manifest *Manifest
+	BlobDir  string
+}
+
+// DefaultBlobDir returns the default blob storage directory.
+func DefaultBlobDir() string {
+	home, err := os.UserHomeDir()
+	if err != nil {
+		home = "."
+	}
+	switch runtime.GOOS {
+	case "darwin":
+		return filepath.Join(home, ".ollama", "models", "blobs")
+	case "linux":
+		return filepath.Join(home, ".ollama", "models", "blobs")
+	case "windows":
+		return filepath.Join(home, ".ollama", "models", "blobs")
+	default:
+		return filepath.Join(home, ".ollama", "models", "blobs")
+	}
+}
+
+// DefaultManifestDir returns the default manifest storage directory.
+func DefaultManifestDir() string {
+	home, err := os.UserHomeDir()
+	if err != nil {
+		home = "."
+	}
+	return filepath.Join(home, ".ollama", "models", "manifests")
+}
+
+// LoadManifest loads a manifest for the given model name.
+// Model name format: "modelname" or "modelname:tag" or "host/namespace/name:tag"
+func LoadManifest(modelName string) (*ModelManifest, error) {
+	manifestPath := resolveManifestPath(modelName)
+
+	data, err := os.ReadFile(manifestPath)
+	if err != nil {
+		return nil, fmt.Errorf("read manifest: %w", err)
+	}
+
+	var manifest Manifest
+	if err := json.Unmarshal(data, &manifest); err != nil {
+		return nil, fmt.Errorf("parse manifest: %w", err)
+	}
+
+	return &ModelManifest{
+		Manifest: &manifest,
+		BlobDir:  DefaultBlobDir(),
+	}, nil
+}
+
+// resolveManifestPath converts a model name to a manifest file path.
+func resolveManifestPath(modelName string) string {
+	// Parse model name into components
+	// Default: registry.ollama.ai/library/<name>/<tag>
+	host := "registry.ollama.ai"
+	namespace := "library"
+	name := modelName
+	tag := "latest"
+
+	// Handle explicit tag
+	if idx := strings.LastIndex(name, ":"); idx != -1 {
+		tag = name[idx+1:]
+		name = name[:idx]
+	}
+
+	// Handle full path like "host/namespace/name"
+	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)
+}
+
+// BlobPath returns the full path to a blob given its digest.
+func (m *ModelManifest) BlobPath(digest string) string {
+	// Convert "sha256:abc123" to "sha256-abc123"
+	blobName := strings.Replace(digest, ":", "-", 1)
+	return filepath.Join(m.BlobDir, blobName)
+}
+
+// GetTensorLayers returns all tensor layers for a given component.
+// Component should be "text_encoder", "transformer", or "vae".
+// Tensor names are path-style: "component/tensor_name" (e.g., "text_encoder/model.embed_tokens.weight").
+func (m *ModelManifest) GetTensorLayers(component string) []ManifestLayer {
+	prefix := component + "/"
+	var layers []ManifestLayer
+	for _, layer := range m.Manifest.Layers {
+		if layer.MediaType == "application/vnd.ollama.image.tensor" && strings.HasPrefix(layer.Name, prefix) {
+			layers = append(layers, layer)
+		}
+	}
+	return layers
+}
+
+// GetConfigLayer returns the config layer for a given path.
+func (m *ModelManifest) GetConfigLayer(configPath string) *ManifestLayer {
+	for _, layer := range m.Manifest.Layers {
+		if layer.MediaType == "application/vnd.ollama.image.json" && layer.Name == configPath {
+			return &layer
+		}
+	}
+	return nil
+}
+
+// ReadConfig reads and returns the content of a config file.
+func (m *ModelManifest) ReadConfig(configPath string) ([]byte, error) {
+	layer := m.GetConfigLayer(configPath)
+	if layer == nil {
+		return nil, fmt.Errorf("config %q not found in manifest", configPath)
+	}
+
+	blobPath := m.BlobPath(layer.Digest)
+	return os.ReadFile(blobPath)
+}
+
+// ReadConfigJSON reads and unmarshals a config file.
+func (m *ModelManifest) ReadConfigJSON(configPath string, v any) error {
+	data, err := m.ReadConfig(configPath)
+	if err != nil {
+		return err
+	}
+	return json.Unmarshal(data, v)
+}
+
+// OpenBlob opens a blob for reading.
+func (m *ModelManifest) OpenBlob(digest string) (io.ReadCloser, error) {
+	return os.Open(m.BlobPath(digest))
+}
+
+// HasTensorLayers returns true if the manifest has any tensor layers.
+func (m *ModelManifest) HasTensorLayers() bool {
+	for _, layer := range m.Manifest.Layers {
+		if layer.MediaType == "application/vnd.ollama.image.tensor" {
+			return true
+		}
+	}
+	return false
+}

x/imagegen/memory.go

@@ -0,0 +1,102 @@
+// Package imagegen provides experimental image generation capabilities for Ollama.
+//
+// This package is in x/ because the tensor model storage format is under development.
+// The goal is to integrate these capabilities into the main Ollama packages once
+// the format is stable.
+//
+// TODO (jmorganca): Integrate into main packages when stable:
+//   - CLI commands → cmd/
+//   - API endpoints → api/
+//   - Model creation → server/
+package imagegen
+
+import (
+	"encoding/json"
+	"fmt"
+	"runtime"
+)
+
+// GB is a convenience constant for gigabytes.
+const GB = 1024 * 1024 * 1024
+
+// SupportedBackends lists the backends that support image generation.
+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
+}
+
+// CheckPlatformSupport validates that image generation is supported on the current platform.
+// Returns nil if supported, or an error describing why it's not supported.
+func CheckPlatformSupport() error {
+	switch runtime.GOOS {
+	case "darwin":
+		// macOS: Metal is supported via MLX
+		if runtime.GOARCH != "arm64" {
+			return fmt.Errorf("image generation on macOS requires Apple Silicon (arm64), got %s", runtime.GOARCH)
+		}
+		return nil
+	case "linux", "windows":
+		// Linux/Windows: CUDA support (requires mlx or cuda build)
+		// The actual backend availability is checked at runtime
+		return nil
+	default:
+		return fmt.Errorf("image generation is not supported on %s", runtime.GOOS)
+	}
+}
+
+// CheckMemoryRequirements validates that there's enough memory for image generation.
+// Returns nil if memory is sufficient, or an error if not.
+func CheckMemoryRequirements(modelName string, availableMemory uint64) error {
+	required := EstimateVRAM(modelName)
+	if availableMemory < required {
+		return fmt.Errorf("insufficient memory for image generation: need %d GB, have %d GB",
+			required/GB, availableMemory/GB)
+	}
+	return nil
+}
+
+// ResolveModelName checks if a model name is a known image generation model.
+// Returns the normalized model name if found, empty string otherwise.
+func ResolveModelName(modelName string) string {
+	manifest, err := LoadManifest(modelName)
+	if err == nil && manifest.HasTensorLayers() {
+		return modelName
+	}
+	return ""
+}
+
+// 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 {
+		return estimate
+	}
+	return 21 * GB
+}
+
+// HasTensorLayers checks if the given model has tensor layers.
+func HasTensorLayers(modelName string) bool {
+	return ResolveModelName(modelName) != ""
+}

x/imagegen/memory_test.go

@@ -0,0 +1,110 @@
+package imagegen
+
+import (
+	"runtime"
+	"testing"
+)
+
+func TestCheckPlatformSupport(t *testing.T) {
+	err := CheckPlatformSupport()
+
+	switch runtime.GOOS {
+	case "darwin":
+		if runtime.GOARCH == "arm64" {
+			if err != nil {
+				t.Errorf("Expected nil error on darwin/arm64, got: %v", err)
+			}
+		} else {
+			if err == nil {
+				t.Error("Expected error on darwin/non-arm64")
+			}
+		}
+	case "linux", "windows":
+		if err != nil {
+			t.Errorf("Expected nil error on %s, got: %v", runtime.GOOS, err)
+		}
+	default:
+		if err == nil {
+			t.Errorf("Expected error on unsupported platform %s", runtime.GOOS)
+		}
+	}
+}
+
+func TestCheckMemoryRequirements(t *testing.T) {
+	tests := []struct {
+		name            string
+		availableMemory uint64
+		wantErr         bool
+	}{
+		{
+			name:            "sufficient memory",
+			availableMemory: 32 * GB,
+			wantErr:         false,
+		},
+		{
+			name:            "exactly enough memory",
+			availableMemory: 21 * GB,
+			wantErr:         false,
+		},
+		{
+			name:            "insufficient memory",
+			availableMemory: 16 * GB,
+			wantErr:         true,
+		},
+		{
+			name:            "zero memory",
+			availableMemory: 0,
+			wantErr:         true,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			// Use a non-existent model name which will default to 21GB estimate
+			err := CheckMemoryRequirements("nonexistent-model", tt.availableMemory)
+			if (err != nil) != tt.wantErr {
+				t.Errorf("CheckMemoryRequirements() error = %v, wantErr %v", err, tt.wantErr)
+			}
+		})
+	}
+}
+
+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,
+	}
+
+	for name, expectedVRAM := range expected {
+		if actual, ok := modelVRAMEstimates[name]; !ok {
+			t.Errorf("Missing VRAM estimate for %s", name)
+		} else if actual != expectedVRAM {
+			t.Errorf("VRAM estimate for %s = %d GB, want %d GB", name, actual/GB, expectedVRAM/GB)
+		}
+	}
+}
+
+func TestEstimateVRAMDefault(t *testing.T) {
+	// Non-existent model should return default 21GB
+	vram := EstimateVRAM("nonexistent-model-that-does-not-exist")
+	if vram != 21*GB {
+		t.Errorf("EstimateVRAM() = %d GB, want 21 GB", vram/GB)
+	}
+}
+
+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")
+	if result != "" {
+		t.Errorf("ResolveModelName() = %q, want empty string", result)
+	}
+}

x/imagegen/mlx/README.md

@@ -0,0 +1,46 @@
+# MLX Memory Management
+
+| This package will get consolidated with `x/ml/backend/mlx` in the future.
+
+## Automatic Tracking
+
+All arrays are automatically tracked when created. On `Eval()`, non-kept arrays are freed.
+
+### API
+
+```go
+result := mlx.Matmul(x, w) // arrays automatically tracked
+mlx.Eval(result)           // free non-kept, eval result (auto-kept)
+```
+
+### Key Functions
+
+- `mlx.Eval(outputs...)` - free non-kept arrays, then evaluate (outputs auto-kept)
+- `mlx.AsyncEval(outputs...)` - async version of Eval (outputs auto-kept)
+- `mlx.Keep(arrays...)` - mark arrays to survive cleanup (for weights, caches)
+- `array.Free()` - mark array for cleanup on next Eval
+
+### Loop Pattern
+
+```go
+for step := 0; step < maxTokens; step++ {
+    logits := model.Forward(token, caches)
+    oldToken := token
+    token = sample(logits)
+
+    // Keep cache state across iterations
+    for _, c := range caches {
+        mlx.Keep(c.State()...)
+    }
+
+    oldToken.Free()       // mark for cleanup
+    mlx.AsyncEval(token)  // frees old, evals new
+}
+```
+
+### Notes
+
+- `Eval()` and `AsyncEval()` auto-keep their outputs
+- `Free()` marks for cleanup - actual free happens during next Eval
+- Use `Keep()` for weights and cache state that must survive multiple Eval cycles
+- Arrays created inside compiled closures are managed by MLX, not tracked

x/imagegen/mlx/compile.go

@@ -0,0 +1,173 @@
+//go:build mlx
+
+package mlx
+
+/*
+#include "mlx/c/mlx.h"
+#include <stdlib.h>
+
+// Forward declaration for Go callback
+extern int goClosureCallback(mlx_vector_array* res, mlx_vector_array input, void* payload);
+
+// Destructor for payload (Go handle)
+extern void goClosureDestructor(void* payload);
+*/
+import "C"
+import (
+	"runtime/cgo"
+	"sync"
+	"unsafe"
+)
+
+// inClosureCallback is set to true during closure callback execution.
+var inClosureCallback bool
+var closureCallbackMu sync.Mutex
+
+// InClosureCallback returns true if we're currently executing inside a closure callback.
+func InClosureCallback() bool {
+	closureCallbackMu.Lock()
+	defer closureCallbackMu.Unlock()
+	return inClosureCallback
+}
+
+// CompiledFunc is a compiled MLX function that can be called efficiently.
+// All intermediate arrays during execution stay inside MLX - only inputs
+// and outputs cross the Go boundary.
+type CompiledFunc struct {
+	closure  C.mlx_closure
+	compiled C.mlx_closure
+}
+
+// ClosureFunc is the signature for functions that can be compiled.
+// It takes a slice of input arrays and returns a slice of output arrays.
+type ClosureFunc func(inputs []*Array) []*Array
+
+// Compile compiles a Go function into an optimized MLX closure.
+// The function is traced once during compilation, then subsequent calls
+// run the optimized graph without creating Go intermediate arrays.
+//
+// Example:
+//
+//	compiled := mlx.Compile(func(inputs []*mlx.Array) []*mlx.Array {
+//	    a, b := inputs[0], inputs[1]
+//	    c := mlx.Add(a, b)
+//	    d := mlx.Mul(c, c)
+//	    return []*mlx.Array{d}
+//	})
+//	defer compiled.Free()
+//
+//	result := compiled.Call(x, y)[0]
+func Compile(fn ClosureFunc) *CompiledFunc {
+	return CompileShapeless(fn, false)
+}
+
+// CompileShapeless compiles with optional shapeless mode.
+// If shapeless=true, the function works for any input shape after tracing.
+func CompileShapeless(fn ClosureFunc, shapeless bool) *CompiledFunc {
+	// Create a cgo.Handle to prevent the Go function from being GC'd
+	handle := cgo.NewHandle(fn)
+
+	// Create the closure from the Go callback
+	closure := C.mlx_closure_new_func_payload(
+		(*[0]byte)(C.goClosureCallback),
+		unsafe.Pointer(handle),
+		(*[0]byte)(C.goClosureDestructor),
+	)
+
+	// Compile the closure
+	compiled := C.mlx_closure_new()
+	C.mlx_compile(&compiled, closure, C.bool(shapeless))
+
+	return &CompiledFunc{
+		closure:  closure,
+		compiled: compiled,
+	}
+}
+
+// Call invokes the compiled function with the given inputs.
+func (cf *CompiledFunc) Call(inputs ...*Array) []*Array {
+	// Pack inputs into vector
+	inputVec := C.mlx_vector_array_new()
+	for _, arr := range inputs {
+		C.mlx_vector_array_append_value(inputVec, arr.c)
+	}
+
+	// Apply compiled closure
+	outputVec := C.mlx_vector_array_new()
+	C.mlx_closure_apply(&outputVec, cf.compiled, inputVec)
+	C.mlx_vector_array_free(inputVec)
+
+	// Unpack outputs
+	numOutputs := int(C.mlx_vector_array_size(outputVec))
+	outputs := make([]*Array, numOutputs)
+	for i := 0; i < numOutputs; i++ {
+		var arr C.mlx_array
+		C.mlx_vector_array_get(&arr, outputVec, C.size_t(i))
+		outputs[i] = newArray(arr)
+	}
+	C.mlx_vector_array_free(outputVec)
+
+	return outputs
+}
+
+// CallEval invokes the compiled function and evaluates the results.
+func (cf *CompiledFunc) CallEval(inputs ...*Array) []*Array {
+	outputs := cf.Call(inputs...)
+	Eval(outputs...)
+	return outputs
+}
+
+// Free releases the compiled function resources.
+func (cf *CompiledFunc) Free() {
+	C.mlx_closure_free(cf.compiled)
+	C.mlx_closure_free(cf.closure)
+}
+
+// borrowArray wraps a C array WITHOUT setting up GC cleanup.
+// Use this for arrays we don't own (e.g., borrowed references in callbacks).
+func borrowArray(array C.mlx_array) *Array {
+	return &Array{c: array}
+}
+
+//export goClosureCallback
+func goClosureCallback(res *C.mlx_vector_array, input C.mlx_vector_array, payload unsafe.Pointer) C.int {
+	// Set flag to disable AddCleanup during callback
+	closureCallbackMu.Lock()
+	inClosureCallback = true
+	closureCallbackMu.Unlock()
+	defer func() {
+		closureCallbackMu.Lock()
+		inClosureCallback = false
+		closureCallbackMu.Unlock()
+	}()
+
+	// Recover the Go function from the handle
+	handle := cgo.Handle(payload)
+	fn := handle.Value().(ClosureFunc)
+
+	// Convert input vector to Go slice - use borrowArray since MLX owns these
+	numInputs := int(C.mlx_vector_array_size(input))
+	inputs := make([]*Array, numInputs)
+	for i := 0; i < numInputs; i++ {
+		var arr C.mlx_array
+		C.mlx_vector_array_get(&arr, input, C.size_t(i))
+		inputs[i] = borrowArray(arr) // Don't set up cleanup - MLX owns these
+	}
+
+	// Call the Go function
+	outputs := fn(inputs)
+
+	// Build output vector
+	*res = C.mlx_vector_array_new()
+	for _, arr := range outputs {
+		C.mlx_vector_array_append_value(*res, arr.c)
+	}
+
+	return 0
+}
+
+//export goClosureDestructor
+func goClosureDestructor(payload unsafe.Pointer) {
+	handle := cgo.Handle(payload)
+	handle.Delete()
+}

x/imagegen/models/gemma3/gemma3.go

@@ -0,0 +1,614 @@
+//go:build mlx
+
+package gemma3
+
+import (
+	"encoding/json"
+	"fmt"
+	"math"
+	"os"
+	"path/filepath"
+
+	"github.com/ollama/ollama/x/imagegen/cache"
+	"github.com/ollama/ollama/x/imagegen/mlx"
+	"github.com/ollama/ollama/x/imagegen/nn"
+	"github.com/ollama/ollama/x/imagegen/safetensors"
+	"github.com/ollama/ollama/x/imagegen/tokenizer"
+)
+
+// TextConfig holds configuration for the text model
+type TextConfig struct {
+	HiddenSize            int32   `json:"hidden_size"`
+	NumHiddenLayers       int32   `json:"num_hidden_layers"`
+	IntermediateSize      int32   `json:"intermediate_size"`
+	NumAttentionHeads     int32   `json:"num_attention_heads"`
+	NumKeyValueHeads      int32   `json:"num_key_value_heads"`
+	HeadDim               int32   `json:"head_dim"`
+	VocabSize             int32   `json:"vocab_size"`
+	RMSNormEps            float32 `json:"rms_norm_eps"`
+	RopeTheta             float32 `json:"rope_theta"`
+	RopeLocalBaseFreq     float32 `json:"rope_local_base_freq"`
+	MaxPositionEmbeddings int32   `json:"max_position_embeddings"`
+	SlidingWindow         int32   `json:"sliding_window"`
+	SlidingWindowPattern  int32   `json:"sliding_window_pattern"`
+
+	// Computed fields
+	Scale float32 `json:"-"`
+}
+
+// TextModel is the Gemma 3 text-only model
+type TextModel struct {
+	EmbedTokens *nn.Embedding   `weight:"model.embed_tokens"`
+	Layers      []*DecoderLayer `weight:"model.layers"`
+	Norm        *nn.RMSNorm     `weight:"model.norm"`
+	Output      *nn.Linear      `weight:"-"` // Tied to EmbedTokens, set manually
+
+	// Precomputed (1 + weight) for Gemma-style RMSNorm to avoid allocation per forward
+	NormScaled *mlx.Array `weight:"-"`
+
+	tok *tokenizer.Tokenizer
+	*TextConfig
+}
+
+// DecoderLayer is a single transformer block
+type DecoderLayer struct {
+	InputNorm    *nn.RMSNorm `weight:"input_layernorm"`
+	Attention    *Attention
+	PostAttnNorm *nn.RMSNorm `weight:"post_attention_layernorm"`
+	PreFFNorm    *nn.RMSNorm `weight:"pre_feedforward_layernorm"`
+	MLP          *MLP
+	PostFFNorm   *nn.RMSNorm `weight:"post_feedforward_layernorm"`
+
+	// Precomputed (1 + weight) for Gemma-style RMSNorm
+	InputNormScaled    *mlx.Array `weight:"-"`
+	PostAttnNormScaled *mlx.Array `weight:"-"`
+	PreFFNormScaled    *mlx.Array `weight:"-"`
+	PostFFNormScaled   *mlx.Array `weight:"-"`
+
+	// Whether this layer uses sliding window attention
+	IsSliding bool
+	LayerIdx  int32
+}
+
+// Attention implements Gemma 3 attention with Q/K normalization
+type Attention struct {
+	QProj *nn.Linear  `weight:"self_attn.q_proj"`
+	KProj *nn.Linear  `weight:"self_attn.k_proj"`
+	VProj *nn.Linear  `weight:"self_attn.v_proj"`
+	OProj *nn.Linear  `weight:"self_attn.o_proj"`
+	QNorm *nn.RMSNorm `weight:"self_attn.q_norm"`
+	KNorm *nn.RMSNorm `weight:"self_attn.k_norm"`
+
+	// Precomputed (1 + weight) for Gemma-style RMSNorm
+	QNormScaled *mlx.Array `weight:"-"`
+	KNormScaled *mlx.Array `weight:"-"`
+}
+
+// MLP is the feed-forward network with GELU activation
+type MLP struct {
+	GateProj *nn.Linear `weight:"mlp.gate_proj"`
+	UpProj   *nn.Linear `weight:"mlp.up_proj"`
+	DownProj *nn.Linear `weight:"mlp.down_proj"`
+}
+
+// LoadText loads the text-only Gemma 3 model
+func LoadText(modelPath string) (*TextModel, error) {
+	data, err := os.ReadFile(filepath.Join(modelPath, "config.json"))
+	if err != nil {
+		return nil, fmt.Errorf("load config: %w", err)
+	}
+	var cfg TextConfig
+	if err := json.Unmarshal(data, &cfg); err != nil {
+		return nil, fmt.Errorf("parse config: %w", err)
+	}
+
+	// Compute scale
+	cfg.Scale = float32(1.0 / math.Sqrt(float64(cfg.HeadDim)))
+
+	// Set defaults if not specified
+	if cfg.RopeTheta == 0 {
+		cfg.RopeTheta = 1000000
+	}
+	if cfg.RopeLocalBaseFreq == 0 {
+		cfg.RopeLocalBaseFreq = 10000
+	}
+	if cfg.RMSNormEps == 0 {
+		cfg.RMSNormEps = 1e-6
+	}
+
+	weights, err := safetensors.LoadModelWeights(modelPath)
+	if err != nil {
+		return nil, fmt.Errorf("load weights: %w", err)
+	}
+
+	tok, err := tokenizer.Load(filepath.Join(modelPath, "tokenizer.json"))
+	if err != nil {
+		return nil, fmt.Errorf("load tokenizer: %w", err)
+	}
+
+	m := &TextModel{
+		Layers:     make([]*DecoderLayer, cfg.NumHiddenLayers),
+		TextConfig: &cfg,
+		tok:        tok,
+	}
+
+	// Initialize layer metadata
+	for i := range m.Layers {
+		m.Layers[i] = &DecoderLayer{
+			LayerIdx:  int32(i),
+			IsSliding: isLayerSliding(int32(i), cfg.SlidingWindowPattern),
+		}
+	}
+
+	if err := safetensors.LoadModule(m, weights, ""); err != nil {
+		return nil, err
+	}
+
+	// Tied embeddings for output
+	m.Output = nn.NewLinear(m.EmbedTokens.Weight, nil)
+
+	mlx.Eval(mlx.Collect(m)...)
+	weights.ReleaseAll()
+
+	// Precompute (1 + weight) for Gemma-style RMSNorm to avoid per-forward allocation
+	precomputeGemmaScaledWeights(m)
+
+	return m, nil
+}
+
+// precomputeGemmaScaledWeights computes (1 + weight) for all RMSNorm layers
+// This avoids creating temporary arrays on every forward pass
+func precomputeGemmaScaledWeights(m *TextModel) {
+	m.NormScaled = mlx.AddScalar(m.Norm.Weight, 1.0)
+
+	for _, layer := range m.Layers {
+		layer.InputNormScaled = mlx.AddScalar(layer.InputNorm.Weight, 1.0)
+		layer.PostAttnNormScaled = mlx.AddScalar(layer.PostAttnNorm.Weight, 1.0)
+		layer.PreFFNormScaled = mlx.AddScalar(layer.PreFFNorm.Weight, 1.0)
+		layer.PostFFNormScaled = mlx.AddScalar(layer.PostFFNorm.Weight, 1.0)
+
+		layer.Attention.QNormScaled = mlx.AddScalar(layer.Attention.QNorm.Weight, 1.0)
+		layer.Attention.KNormScaled = mlx.AddScalar(layer.Attention.KNorm.Weight, 1.0)
+	}
+
+	// Eval all the precomputed weights
+	var scaled []*mlx.Array
+	scaled = append(scaled, m.NormScaled)
+	for _, layer := range m.Layers {
+		scaled = append(scaled, layer.InputNormScaled, layer.PostAttnNormScaled,
+			layer.PreFFNormScaled, layer.PostFFNormScaled,
+			layer.Attention.QNormScaled, layer.Attention.KNormScaled)
+	}
+	mlx.Eval(scaled...)
+}
+
+// isLayerSliding determines if a layer uses sliding window attention
+// Pattern N means: layers 0 to N-1 sliding, N full, N+1 to 2N-1 sliding, 2N full, etc.
+func isLayerSliding(layerIdx, pattern int32) bool {
+	if pattern <= 0 {
+		return false // No sliding window
+	}
+	// Layer is full attention if (layerIdx + 1) % pattern == 0
+	return (layerIdx+1)%pattern != 0
+}
+
+// Forward runs the text model forward pass
+func (m *TextModel) Forward(tokens *mlx.Array, caches []cache.Cache) *mlx.Array {
+	B, L := tokens.Shape()[0], tokens.Shape()[1]
+
+	// Get embeddings and scale by sqrt(hidden_size)
+	h := m.EmbedTokens.Forward(tokens)
+	h = mlx.MulScalar(h, float32(math.Sqrt(float64(m.HiddenSize))))
+
+	for i, layer := range m.Layers {
+		h = layer.Forward(h, caches[i], B, L, m.TextConfig)
+	}
+
+	// Final norm and output projection
+	return m.Output.Forward(mlx.RMSNorm(h, m.NormScaled, m.RMSNormEps))
+}
+
+// Forward runs a decoder layer
+func (l *DecoderLayer) Forward(x *mlx.Array, c cache.Cache, B, L int32, cfg *TextConfig) *mlx.Array {
+	// Pre-attention norm (use precomputed scaled weight)
+	normed := mlx.RMSNorm(x, l.InputNormScaled, cfg.RMSNormEps)
+
+	// Attention
+	attnOut := l.Attention.Forward(normed, c, B, L, l.IsSliding, cfg)
+
+	// Post-attention norm and residual
+	attnOut = mlx.RMSNorm(attnOut, l.PostAttnNormScaled, cfg.RMSNormEps)
+	h := mlx.Add(x, attnOut)
+
+	// Pre-FFN norm
+	normed = mlx.RMSNorm(h, l.PreFFNormScaled, cfg.RMSNormEps)
+
+	// MLP
+	mlpOut := l.MLP.Forward(normed)
+
+	// Post-FFN norm and residual
+	mlpOut = mlx.RMSNorm(mlpOut, l.PostFFNormScaled, cfg.RMSNormEps)
+	return mlx.Add(h, mlpOut)
+}
+
+// Forward runs attention with Q/K normalization
+func (a *Attention) Forward(x *mlx.Array, c cache.Cache, B, L int32, isSliding bool, cfg *TextConfig) *mlx.Array {
+	q := a.QProj.Forward(x)
+	k := a.KProj.Forward(x)
+	v := a.VProj.Forward(x)
+
+	// Reshape to [B, num_heads, L, head_dim]
+	q = mlx.AsStrided(q, []int32{B, cfg.NumAttentionHeads, L, cfg.HeadDim},
+		[]int64{int64(L * cfg.NumAttentionHeads * cfg.HeadDim), int64(cfg.HeadDim), int64(cfg.NumAttentionHeads * cfg.HeadDim), 1}, 0)
+	k = mlx.AsStrided(k, []int32{B, cfg.NumKeyValueHeads, L, cfg.HeadDim},
+		[]int64{int64(L * cfg.NumKeyValueHeads * cfg.HeadDim), int64(cfg.HeadDim), int64(cfg.NumKeyValueHeads * cfg.HeadDim), 1}, 0)
+	v = mlx.AsStrided(v, []int32{B, cfg.NumKeyValueHeads, L, cfg.HeadDim},
+		[]int64{int64(L * cfg.NumKeyValueHeads * cfg.HeadDim), int64(cfg.HeadDim), int64(cfg.NumKeyValueHeads * cfg.HeadDim), 1}, 0)
+
+	// Q/K normalization after reshaping (use precomputed scaled weight)
+	q = mlx.RMSNorm(q, a.QNormScaled, cfg.RMSNormEps)
+	k = mlx.RMSNorm(k, a.KNormScaled, cfg.RMSNormEps)
+
+	// Apply RoPE with appropriate theta
+	ropeTheta := cfg.RopeTheta
+	if isSliding {
+		ropeTheta = cfg.RopeLocalBaseFreq
+	}
+	q = mlx.RoPE(q, int(cfg.HeadDim), false, ropeTheta, 1.0, c.Offset())
+	k = mlx.RoPE(k, int(cfg.HeadDim), false, ropeTheta, 1.0, c.Offset())
+
+	// Update cache
+	k, v = c.Update(k, v, int(L))
+
+	// Repeat K/V for GQA if needed
+	repeatFactor := cfg.NumAttentionHeads / cfg.NumKeyValueHeads
+	if repeatFactor > 1 {
+		k = nn.RepeatKV(k, repeatFactor)
+		v = nn.RepeatKV(v, repeatFactor)
+	}
+
+	// Attention
+	out := mlx.ScaledDotProductAttention(q, k, v, cfg.Scale, L > 1)
+	out = mlx.Reshape(mlx.Transpose(out, 0, 2, 1, 3), B, L, cfg.NumAttentionHeads*cfg.HeadDim)
+	return a.OProj.Forward(out)
+}
+
+// compiledGeluApprox is a singleton compiled GELU function shared across all layers
+var compiledGeluApprox *mlx.CompiledFunc
+
+// getCompiledGeluApprox returns the compiled GELU function, creating it once if needed
+func getCompiledGeluApprox() *mlx.CompiledFunc {
+	if compiledGeluApprox == nil {
+		compiledGeluApprox = mlx.CompileShapeless(func(inputs []*mlx.Array) []*mlx.Array {
+			return []*mlx.Array{geluApproxImpl(inputs[0])}
+		}, true)
+	}
+	return compiledGeluApprox
+}
+
+// Forward runs the MLP with GELU approximation (tanh variant)
+func (m *MLP) Forward(x *mlx.Array) *mlx.Array {
+	gate := getCompiledGeluApprox().Call(m.GateProj.Forward(x))[0]
+	return m.DownProj.Forward(mlx.Mul(gate, m.UpProj.Forward(x)))
+}
+
+// geluApproxImpl computes GELU using the tanh approximation (gelu_pytorch_tanh):
+// 0.5 * x * (1 + tanh(sqrt(2/pi) * (x + 0.044715 * x^3)))
+func geluApproxImpl(x *mlx.Array) *mlx.Array {
+	// Constants
+	const sqrt2OverPi = 0.7978845608028654 // sqrt(2/pi)
+	const coeff = 0.044715
+
+	// x^3
+	x3 := mlx.Mul(mlx.Mul(x, x), x)
+	// x + 0.044715 * x^3
+	inner := mlx.Add(x, mlx.MulScalar(x3, coeff))
+	// sqrt(2/pi) * (x + 0.044715 * x^3)
+	scaled := mlx.MulScalar(inner, sqrt2OverPi)
+	// tanh(...)
+	tanh := mlx.Tanh(scaled)
+	// 1 + tanh(...)
+	onePlusTanh := mlx.AddScalar(tanh, 1.0)
+	// 0.5 * x * (1 + tanh(...))
+	return mlx.Mul(mlx.MulScalar(x, 0.5), onePlusTanh)
+}
+
+// gemmaRMSNorm applies Gemma-style RMS normalization: x * rsqrt(mean(x^2) + eps) * (1 + weight)
+// Uses mlx.RMSNorm fast kernel with pre-computed (1 + weight)
+func gemmaRMSNorm(x, weight *mlx.Array, eps float32) *mlx.Array {
+	// Gemma uses (1 + weight) instead of weight
+	scaledWeight := mlx.AddScalar(weight, 1.0)
+	return mlx.RMSNorm(x, scaledWeight, eps)
+}
+
+// Interface methods
+func (m *TextModel) NumLayers() int          { return len(m.Layers) }
+func (m *TextModel) MaxContextLength() int32 { return m.MaxPositionEmbeddings }
+func (m *TextModel) VocabSize() int32        { return m.TextConfig.VocabSize }
+
+// Tokenizer returns the tokenizer wrapped to add BOS and apply chat template
+func (m *TextModel) Tokenizer() *tokenizer.Tokenizer {
+	return m.tok
+}
+
+// FormatPrompt applies the Gemma 3 chat template to a prompt
+func (m *TextModel) FormatPrompt(prompt string) string {
+	// Gemma 3 chat format: <start_of_turn>user\n{prompt}<end_of_turn>\n<start_of_turn>model\n
+	return fmt.Sprintf("<start_of_turn>user\n%s<end_of_turn>\n<start_of_turn>model\n", prompt)
+}
+
+func (m *TextModel) NewCache(maxSeqLen int32) []cache.Cache {
+	caches := make([]cache.Cache, len(m.Layers))
+	for i := range caches {
+		if m.Layers[i].IsSliding {
+			// Use rotating cache for sliding window layers
+			caches[i] = cache.NewRotatingKVCache(int(m.SlidingWindow))
+		} else {
+			// Use regular cache for global attention layers
+			caches[i] = cache.NewKVCache()
+		}
+	}
+	return caches
+}
+
+// Config holds config for the full multimodal model
+type Config struct {
+	TextConfig   TextConfig   `json:"text_config"`
+	VisionConfig VisionConfig `json:"vision_config"`
+
+	// Image token config (from config.json)
+	BOITokenIndex     int32 `json:"boi_token_index"`     // <start_of_image> = 255999
+	EOITokenIndex     int32 `json:"eoi_token_index"`     // <end_of_image> = 256000
+	ImageTokenIndex   int32 `json:"image_token_index"`   // <image_soft_token> = 262144
+	MMTokensPerImage  int32 `json:"mm_tokens_per_image"` // 256
+}
+
+// Model is the full Gemma 3 multimodal model
+type Model struct {
+	VisionTower *VisionTower         `weight:"vision_tower"`
+	Projector   *MultiModalProjector `weight:"multi_modal_projector"`
+	TextModel   *TextModel           `weight:"language_model"`
+	Config      *Config
+	tok         *tokenizer.Tokenizer
+}
+
+// Load loads the full multimodal Gemma 3 model
+func Load(modelPath string) (*Model, error) {
+	data, err := os.ReadFile(filepath.Join(modelPath, "config.json"))
+	if err != nil {
+		return nil, fmt.Errorf("load config: %w", err)
+	}
+
+	var cfg Config
+	if err := json.Unmarshal(data, &cfg); err != nil {
+		return nil, fmt.Errorf("parse config: %w", err)
+	}
+
+	// Set defaults for text config (multimodal config often has incomplete text_config)
+	// These defaults match transformers.Gemma3TextConfig defaults
+	tc := &cfg.TextConfig
+	if tc.HeadDim == 0 {
+		tc.HeadDim = 256 // Gemma 3 uses head_dim=256
+	}
+	if tc.NumAttentionHeads == 0 {
+		// Gemma 3 4B uses 8 attention heads (cannot infer from hidden_size/head_dim)
+		tc.NumAttentionHeads = 8
+	}
+	if tc.NumKeyValueHeads == 0 {
+		// Gemma 3 4B uses 4 KV heads (GQA with 2:1 ratio)
+		tc.NumKeyValueHeads = 4
+	}
+	if tc.VocabSize == 0 {
+		tc.VocabSize = 262208 // Gemma 3 vocab size (not 262144!)
+	}
+	if tc.RopeTheta == 0 {
+		tc.RopeTheta = 1000000
+	}
+	if tc.RopeLocalBaseFreq == 0 {
+		tc.RopeLocalBaseFreq = 10000
+	}
+	if tc.RMSNormEps == 0 {
+		tc.RMSNormEps = 1e-6
+	}
+	if tc.SlidingWindowPattern == 0 {
+		tc.SlidingWindowPattern = 6
+	}
+	if tc.MaxPositionEmbeddings == 0 {
+		tc.MaxPositionEmbeddings = 131072 // Gemma 3 4B default
+	}
+
+	// Compute text model scale
+	tc.Scale = float32(1.0 / math.Sqrt(float64(tc.HeadDim)))
+
+	// Set defaults for image token config
+	if cfg.BOITokenIndex == 0 {
+		cfg.BOITokenIndex = 255999 // <start_of_image>
+	}
+	if cfg.EOITokenIndex == 0 {
+		cfg.EOITokenIndex = 256000 // <end_of_image>
+	}
+	if cfg.ImageTokenIndex == 0 {
+		cfg.ImageTokenIndex = 262144 // <image_soft_token>
+	}
+	if cfg.MMTokensPerImage == 0 {
+		cfg.MMTokensPerImage = 256
+	}
+
+	weights, err := safetensors.LoadModelWeights(modelPath)
+	if err != nil {
+		return nil, fmt.Errorf("load weights: %w", err)
+	}
+
+	tok, err := tokenizer.Load(filepath.Join(modelPath, "tokenizer.json"))
+	if err != nil {
+		return nil, fmt.Errorf("load tokenizer: %w", err)
+	}
+
+	m := &Model{
+		VisionTower: &VisionTower{
+			Embeddings: &VisionEmbeddings{},
+			Encoder:    make([]*VisionEncoderLayer, cfg.VisionConfig.NumHiddenLayers),
+			Config:     &cfg.VisionConfig,
+		},
+		Projector: &MultiModalProjector{},
+		TextModel: &TextModel{
+			Layers:     make([]*DecoderLayer, cfg.TextConfig.NumHiddenLayers),
+			TextConfig: &cfg.TextConfig,
+		},
+		Config: &cfg,
+		tok:    tok,
+	}
+
+	// Initialize text layer metadata
+	for i := range m.TextModel.Layers {
+		m.TextModel.Layers[i] = &DecoderLayer{
+			LayerIdx:  int32(i),
+			IsSliding: isLayerSliding(int32(i), cfg.TextConfig.SlidingWindowPattern),
+		}
+	}
+
+	// Initialize vision encoder layers
+	for i := range m.VisionTower.Encoder {
+		m.VisionTower.Encoder[i] = &VisionEncoderLayer{}
+	}
+
+	if err := safetensors.LoadModule(m, weights, ""); err != nil {
+		return nil, err
+	}
+
+	// Tied embeddings for text output
+	m.TextModel.Output = nn.NewLinear(m.TextModel.EmbedTokens.Weight, nil)
+	m.TextModel.tok = tok
+
+	mlx.Eval(mlx.Collect(m)...)
+	weights.ReleaseAll()
+
+	// Precompute (1 + weight) for Gemma-style RMSNorm
+	precomputeGemmaScaledWeights(m.TextModel)
+
+	// Precompute projector's scaled weight
+	m.Projector.SoftEmbNormScaled = mlx.AddScalar(m.Projector.SoftEmbNorm.Weight, 1.0)
+	mlx.Eval(m.Projector.SoftEmbNormScaled)
+
+	return m, nil
+}
+
+// Forward runs the text-only forward pass
+func (m *Model) Forward(tokens *mlx.Array, caches []cache.Cache) *mlx.Array {
+	return m.TextModel.Forward(tokens, caches)
+}
+
+// ForwardWithImage runs the multimodal forward pass
+// tokens: [B, L] input token IDs (with image placeholder tokens)
+// image: [B, H, W, C] preprocessed image tensor
+func (m *Model) ForwardWithImage(tokens *mlx.Array, image *mlx.Array, caches []cache.Cache) *mlx.Array {
+	B, L := tokens.Shape()[0], tokens.Shape()[1]
+	cfg := m.Config.TextConfig
+
+	// Find image token position FIRST before any eval that might free tokens
+	imageStartPos := int32(-1)
+	if image != nil && B == 1 {
+		tokenData := tokens.DataInt32() // This evals tokens
+		for i, t := range tokenData {
+			if t == m.Config.ImageTokenIndex {
+				imageStartPos = int32(i)
+				break
+			}
+		}
+	}
+
+	// Get text embeddings and scale
+	h := m.TextModel.EmbedTokens.Forward(tokens)
+	h = mlx.MulScalar(h, float32(math.Sqrt(float64(cfg.HiddenSize))))
+
+	// Process image if provided
+	if image != nil && imageStartPos >= 0 {
+		// Vision tower: [B, H, W, C] -> [B, num_patches, vision_hidden]
+		visionFeatures := m.VisionTower.Forward(image)
+
+		// Project to text space: [B, num_patches, vision_hidden] -> [B, 256, text_hidden]
+		imageEmbeds := m.Projector.Forward(visionFeatures, cfg.RMSNormEps)
+
+		// Eval h and imageEmbeds together so neither gets freed
+		mlx.Eval(h, imageEmbeds)
+
+		// Cast imageEmbeds to match text embeddings dtype (bf16)
+		if imageEmbeds.Dtype() != h.Dtype() {
+			imageEmbeds = mlx.AsType(imageEmbeds, h.Dtype())
+			mlx.Eval(imageEmbeds)
+		}
+
+		// Insert image embeddings at the known position
+		h = m.insertImageEmbeddingsAt(h, imageEmbeds, imageStartPos)
+	}
+
+	// Run through text model layers
+	for i, layer := range m.TextModel.Layers {
+		h = layer.Forward(h, caches[i], B, L, m.TextModel.TextConfig)
+	}
+
+	// Final norm and output projection
+	return m.TextModel.Output.Forward(mlx.RMSNorm(h, m.TextModel.NormScaled, cfg.RMSNormEps))
+}
+
+// insertImageEmbeddingsAt replaces image placeholder tokens with actual image embeddings
+// at a known position (to avoid re-scanning tokens after eval)
+// textEmbeds: [B, L, hidden_size] text embeddings
+// imageEmbeds: [B, 256, hidden_size] image embeddings from projector
+// startPos: starting position of image tokens in the sequence
+func (m *Model) insertImageEmbeddingsAt(textEmbeds, imageEmbeds *mlx.Array, startPos int32) *mlx.Array {
+	numImageTokens := imageEmbeds.Shape()[1]
+	L := textEmbeds.Shape()[1]
+
+	// Split text embeddings: [0:startPos] + imageEmbeds + [startPos+256:L]
+	afterStart := startPos + numImageTokens
+
+	// Slice before image tokens: textEmbeds[:, 0:startPos, :]
+	before := mlx.SliceAxis(textEmbeds, 1, 0, startPos)
+
+	// Slice after image tokens: textEmbeds[:, startPos+256:L, :]
+	after := mlx.SliceAxis(textEmbeds, 1, afterStart, L)
+
+	// Concatenate: before + imageEmbeds + after along axis 1
+	return mlx.Concatenate([]*mlx.Array{before, imageEmbeds, after}, 1)
+}
+
+// Interface methods for Model
+func (m *Model) NumLayers() int                         { return len(m.TextModel.Layers) }
+func (m *Model) MaxContextLength() int32                { return m.Config.TextConfig.MaxPositionEmbeddings }
+func (m *Model) VocabSize() int32                       { return m.Config.TextConfig.VocabSize }
+func (m *Model) Tokenizer() *tokenizer.Tokenizer     { return m.tok }
+func (m *Model) NewCache(maxSeqLen int32) []cache.Cache { return m.TextModel.NewCache(maxSeqLen) }
+func (m *Model) ImageSize() int32                       { return m.Config.VisionConfig.ImageSize }
+
+// FormatPrompt applies the Gemma 3 multimodal chat template
+func (m *Model) FormatPrompt(prompt string) string {
+	return fmt.Sprintf("<start_of_turn>user\n%s<end_of_turn>\n<start_of_turn>model\n", prompt)
+}
+
+// FormatPromptWithImage applies the Gemma 3 multimodal chat template with image
+func (m *Model) FormatPromptWithImage(prompt string) string {
+	return fmt.Sprintf("<start_of_turn>user\n<start_of_image>%s<end_of_turn>\n<start_of_turn>model\n", prompt)
+}
+
+// ExpandImageTokens expands <start_of_image> into 256 image placeholder tokens
+// Input tokens containing boi_token (255999) are expanded to:
+// boi_token + 256 * image_token + eoi_token
+func (m *Model) ExpandImageTokens(tokens []int32) []int32 {
+	result := make([]int32, 0, len(tokens)+int(m.Config.MMTokensPerImage)+1)
+
+	for _, t := range tokens {
+		if t == m.Config.BOITokenIndex {
+			// Expand: boi + 256 * image_token + eoi
+			result = append(result, m.Config.BOITokenIndex)
+			for i := int32(0); i < m.Config.MMTokensPerImage; i++ {
+				result = append(result, m.Config.ImageTokenIndex)
+			}
+			result = append(result, m.Config.EOITokenIndex)
+		} else {
+			result = append(result, t)
+		}
+	}
+
+	return result
+}

x/imagegen/models/gemma3/image.go

@@ -0,0 +1,58 @@
+//go:build mlx
+
+package gemma3
+
+import (
+	"fmt"
+	"image"
+	_ "image/jpeg"
+	_ "image/png"
+	"os"
+
+	"github.com/ollama/ollama/x/imagegen/mlx"
+	"golang.org/x/image/draw"
+)
+
+// ProcessImage loads and preprocesses an image for the vision tower
+// Returns [1, H, W, C] tensor in NHWC format normalized for SigLIP
+func ProcessImage(path string, imageSize int32) (*mlx.Array, error) {
+	f, err := os.Open(path)
+	if err != nil {
+		return nil, fmt.Errorf("open image: %w", err)
+	}
+	defer f.Close()
+
+	img, _, err := image.Decode(f)
+	if err != nil {
+		return nil, fmt.Errorf("decode image: %w", err)
+	}
+
+	return ProcessImageData(img, imageSize)
+}
+
+// ProcessImageData preprocesses an image.Image for the vision tower
+func ProcessImageData(img image.Image, imageSize int32) (*mlx.Array, error) {
+	// Resize to target size using bilinear interpolation
+	resized := image.NewRGBA(image.Rect(0, 0, int(imageSize), int(imageSize)))
+	draw.BiLinear.Scale(resized, resized.Bounds(), img, img.Bounds(), draw.Over, nil)
+
+	// Convert to float32 array [H, W, C] and normalize
+	// SigLIP normalization: (pixel / 255.0 - 0.5) / 0.5 = pixel / 127.5 - 1.0
+	data := make([]float32, imageSize*imageSize*3)
+	idx := 0
+	for y := int32(0); y < imageSize; y++ {
+		for x := int32(0); x < imageSize; x++ {
+			r, g, b, _ := resized.At(int(x), int(y)).RGBA()
+			// RGBA returns 16-bit values, convert to 8-bit
+			data[idx] = float32(r>>8)/127.5 - 1.0
+			data[idx+1] = float32(g>>8)/127.5 - 1.0
+			data[idx+2] = float32(b>>8)/127.5 - 1.0
+			idx += 3
+		}
+	}
+
+	// Create MLX array [1, H, W, C] for NHWC layout
+	arr := mlx.NewArrayFloat32(data, []int32{1, imageSize, imageSize, 3})
+	mlx.Eval(arr) // Materialize to prevent use-after-free
+	return arr, nil
+}

x/imagegen/models/gemma3/projector.go

@@ -0,0 +1,50 @@
+//go:build mlx
+
+package gemma3
+
+import (
+	"github.com/ollama/ollama/x/imagegen/mlx"
+	"github.com/ollama/ollama/x/imagegen/nn"
+)
+
+// MultiModalProjector projects vision features to text embedding space
+type MultiModalProjector struct {
+	// mm_input_projection_weight: [vision_hidden, text_hidden]
+	InputProjection *mlx.Array  `weight:"mm_input_projection_weight"`
+	SoftEmbNorm     *nn.RMSNorm `weight:"mm_soft_emb_norm"`
+
+	// Precomputed (1 + weight) for Gemma-style RMSNorm
+	SoftEmbNormScaled *mlx.Array `weight:"-"`
+}
+
+// Forward projects vision features to text space
+// Input: [B, num_patches, vision_hidden] (e.g., [1, 4096, 1152])
+// Output: [B, num_image_tokens, text_hidden] (e.g., [1, 256, 2560])
+func (p *MultiModalProjector) Forward(visionFeatures *mlx.Array, eps float32) *mlx.Array {
+	// Average pool 4x4: [B, 4096, 1152] -> [B, 256, 1152]
+	// 4096 patches = 64x64 grid, pool to 16x16 = 256 tokens
+	B := visionFeatures.Shape()[0]
+	visionHidden := visionFeatures.Shape()[2]
+
+	// Reshape to [B, 64, 64, hidden]
+	gridSize := int32(64) // sqrt(4096)
+	pooledSize := int32(16) // 64/4
+	h := mlx.Reshape(visionFeatures, B, gridSize, gridSize, visionHidden)
+
+	// Reshape to [B, 16, 4, 16, 4, hidden] for 4x4 pooling
+	h = mlx.Reshape(h, B, pooledSize, 4, pooledSize, 4, visionHidden)
+
+	// Average over pooling dimensions (axes 2 and 4)
+	h = mlx.Mean(h, 4, false)
+	h = mlx.Mean(h, 2, false)
+
+	// h is now [B, 16, 16, hidden], reshape to [B, 256, hidden]
+	numTokens := pooledSize * pooledSize
+	h = mlx.Reshape(h, B, numTokens, visionHidden)
+
+	// Apply Gemma-style RMS norm (use precomputed 1 + weight)
+	h = mlx.RMSNorm(h, p.SoftEmbNormScaled, eps)
+
+	// Project to text space: [B, 256, vision_hidden] @ [vision_hidden, text_hidden]
+	return mlx.Linear(h, p.InputProjection)
+}

x/imagegen/models/gemma3/vision.go

@@ -0,0 +1,138 @@
+//go:build mlx
+
+package gemma3
+
+import (
+	"math"
+
+	"github.com/ollama/ollama/x/imagegen/mlx"
+	"github.com/ollama/ollama/x/imagegen/nn"
+)
+
+// VisionConfig holds configuration for the SigLIP vision tower
+type VisionConfig struct {
+	HiddenSize        int32 `json:"hidden_size"`
+	ImageSize         int32 `json:"image_size"`
+	IntermediateSize  int32 `json:"intermediate_size"`
+	NumAttentionHeads int32 `json:"num_attention_heads"`
+	NumHiddenLayers   int32 `json:"num_hidden_layers"`
+	PatchSize         int32 `json:"patch_size"`
+}
+
+// VisionTower is the SigLIP vision encoder
+type VisionTower struct {
+	Embeddings    *VisionEmbeddings     `weight:"vision_model.embeddings"`
+	Encoder       []*VisionEncoderLayer `weight:"vision_model.encoder.layers"`
+	PostLayerNorm *nn.LayerNorm         `weight:"vision_model.post_layernorm"`
+	Config        *VisionConfig
+}
+
+// VisionEmbeddings handles patch and position embeddings
+type VisionEmbeddings struct {
+	// PatchWeight: [O, C, kH, kW] from PyTorch, transposed to [O, kH, kW, C] for MLX
+	PatchWeight *mlx.Array    `weight:"patch_embedding.weight"`
+	PatchBias   *mlx.Array    `weight:"patch_embedding.bias"`
+	PosEmbed    *nn.Embedding `weight:"position_embedding"`
+}
+
+// VisionEncoderLayer is a single transformer encoder layer
+type VisionEncoderLayer struct {
+	LayerNorm1 *nn.LayerNorm     `weight:"layer_norm1"`
+	Attention  *VisionAttention  `weight:"self_attn"`
+	LayerNorm2 *nn.LayerNorm     `weight:"layer_norm2"`
+	MLP        *VisionMLP        `weight:"mlp"`
+}
+
+// VisionAttention implements multi-head self-attention
+type VisionAttention struct {
+	QProj   *nn.Linear `weight:"q_proj"`
+	KProj   *nn.Linear `weight:"k_proj"`
+	VProj   *nn.Linear `weight:"v_proj"`
+	OutProj *nn.Linear `weight:"out_proj"`
+}
+
+// VisionMLP is the feed-forward network
+type VisionMLP struct {
+	FC1 *nn.Linear `weight:"fc1"`
+	FC2 *nn.Linear `weight:"fc2"`
+}
+
+// Forward runs the vision tower on preprocessed images
+// Input: [B, H, W, C] normalized image tensor (NHWC layout for MLX)
+// Output: [B, num_patches, hidden_size]
+func (v *VisionTower) Forward(x *mlx.Array) *mlx.Array {
+	// Patch embedding conv: input [B, H, W, C], weight [O, kH, kW, C] -> [B, grid, grid, O]
+	// Weight comes as [O, C, kH, kW] from PyTorch, transpose to [O, kH, kW, C]
+	weight := mlx.Transpose(v.Embeddings.PatchWeight, 0, 2, 3, 1)
+	h := mlx.Conv2d(x, weight, v.Config.PatchSize, 0) // stride=patch_size, no padding
+
+	// Add bias: [O] -> [1, 1, 1, O] for broadcasting
+	bias := mlx.Reshape(v.Embeddings.PatchBias, 1, 1, 1, v.Embeddings.PatchBias.Shape()[0])
+	h = mlx.Add(h, bias)
+
+	// h is [B, grid, grid, hidden], flatten to [B, num_patches, hidden]
+	B := h.Shape()[0]
+	gridH, gridW := h.Shape()[1], h.Shape()[2]
+	hidden := h.Shape()[3]
+	numPatches := gridH * gridW
+	h = mlx.Reshape(h, B, numPatches, hidden)
+
+	// Add position embeddings
+	posIds := mlx.ArangeInt(0, numPatches, 1, mlx.DtypeInt32)
+	posEmbed := v.Embeddings.PosEmbed.Forward(posIds)
+	h = mlx.Add(h, posEmbed)
+
+	// Encoder layers
+	headDim := float32(v.Config.HiddenSize / v.Config.NumAttentionHeads)
+	scale := float32(1.0 / math.Sqrt(float64(headDim)))
+	for _, layer := range v.Encoder {
+		h = layer.Forward(h, v.Config, scale)
+	}
+
+	// Final layer norm
+	h = v.PostLayerNorm.Forward(h)
+
+	return h
+}
+
+// Forward runs a vision encoder layer
+func (l *VisionEncoderLayer) Forward(x *mlx.Array, cfg *VisionConfig, scale float32) *mlx.Array {
+	// Pre-norm attention
+	h := l.LayerNorm1.Forward(x)
+	h = l.Attention.Forward(h, cfg, scale)
+	x = mlx.Add(x, h)
+
+	// Pre-norm MLP
+	h = l.LayerNorm2.Forward(x)
+	h = l.MLP.Forward(h)
+	return mlx.Add(x, h)
+}
+
+// Forward runs multi-head self-attention
+func (a *VisionAttention) Forward(x *mlx.Array, cfg *VisionConfig, scale float32) *mlx.Array {
+	B, L := x.Shape()[0], x.Shape()[1]
+	headDim := cfg.HiddenSize / cfg.NumAttentionHeads
+
+	q := a.QProj.Forward(x)
+	k := a.KProj.Forward(x)
+	v := a.VProj.Forward(x)
+
+	// Reshape to [B, num_heads, L, head_dim]
+	q = mlx.Transpose(mlx.Reshape(q, B, L, cfg.NumAttentionHeads, headDim), 0, 2, 1, 3)
+	k = mlx.Transpose(mlx.Reshape(k, B, L, cfg.NumAttentionHeads, headDim), 0, 2, 1, 3)
+	v = mlx.Transpose(mlx.Reshape(v, B, L, cfg.NumAttentionHeads, headDim), 0, 2, 1, 3)
+
+	// Scaled dot-product attention (no causal mask for vision)
+	out := mlx.ScaledDotProductAttention(q, k, v, scale, false)
+
+	// Reshape back: [B, num_heads, L, head_dim] -> [B, L, hidden]
+	out = mlx.Reshape(mlx.Transpose(out, 0, 2, 1, 3), B, L, cfg.HiddenSize)
+
+	return a.OutProj.Forward(out)
+}
+
+// Forward runs the MLP with GELU activation
+func (m *VisionMLP) Forward(x *mlx.Array) *mlx.Array {
+	h := mlx.GELU(m.FC1.Forward(x))
+	return m.FC2.Forward(h)
+}

x/imagegen/models/gpt_oss/gpt_oss.go

@@ -0,0 +1,487 @@
+//go:build mlx
+
+package gpt_oss
+
+import (
+	"encoding/json"
+	"fmt"
+	"math"
+	"os"
+	"path/filepath"
+
+	"github.com/ollama/ollama/x/imagegen/cache"
+	"github.com/ollama/ollama/x/imagegen/mlx"
+	"github.com/ollama/ollama/x/imagegen/nn"
+	"github.com/ollama/ollama/x/imagegen/safetensors"
+	"github.com/ollama/ollama/x/imagegen/tokenizer"
+)
+
+// RopeScaling holds YaRN or other RoPE scaling configuration
+type RopeScaling struct {
+	RopeType                      string  `json:"rope_type"`
+	Factor                        float32 `json:"factor"`
+	OriginalMaxPositionEmbeddings int32   `json:"original_max_position_embeddings"`
+	BetaFast                      float32 `json:"beta_fast"`
+	BetaSlow                      float32 `json:"beta_slow"`
+}
+
+type Config struct {
+	HiddenSize        int32        `json:"hidden_size"`
+	NumHiddenLayers   int32        `json:"num_hidden_layers"`
+	IntermediateSize  int32        `json:"intermediate_size"`
+	NumAttentionHeads int32        `json:"num_attention_heads"`
+	NumKeyValueHeads  int32        `json:"num_key_value_heads"`
+	VocabSize         int32        `json:"vocab_size"`
+	RMSNormEps        float32      `json:"rms_norm_eps"`
+	RopeTheta         float32      `json:"rope_theta"`
+	HeadDim           int32        `json:"head_dim"`
+	SlidingWindow     int32        `json:"sliding_window"`
+	NumLocalExperts   int32        `json:"num_local_experts"`
+	NumExpertsPerTok  int32        `json:"num_experts_per_tok"`
+	LayerTypes        []string     `json:"layer_types"`
+	SwiGLULimit       float32      `json:"swiglu_limit"`
+	RopeScaling       *RopeScaling `json:"rope_scaling"`
+	Scale             float32      `json:"-"` // computed: 1/sqrt(HeadDim)
+}
+
+type Attention struct {
+	QProj      *nn.Linear `weight:"self_attn.q_proj"`
+	KProj      *nn.Linear `weight:"self_attn.k_proj"`
+	VProj      *nn.Linear `weight:"self_attn.v_proj"`
+	OProj      *nn.Linear `weight:"self_attn.o_proj"`
+	Sinks      *mlx.Array `weight:"self_attn.sinks,optional"`
+	YarnFreqs  *mlx.Array // computed
+	YarnMscale float32
+}
+
+// swiGLU applies the GPT-OSS custom SwiGLU activation.
+// Formula: (gate * sigmoid(alpha * gate)) * (up + 1)
+// with clipping: gate to [None, limit], up to [-limit, limit]
+func swiGLU(gate, up *mlx.Array, alpha, limit float32) *mlx.Array {
+	// Clip gate to [None, limit]
+	gateClipped := mlx.ClipScalar(gate, 0, limit, false, true)
+
+	// Clip up to [-limit, limit]
+	upClipped := mlx.ClipScalar(up, -limit, limit, true, true)
+
+	// glu_scaled = alpha * gate_clipped
+	gluScaled := mlx.MulScalar(gateClipped, alpha)
+
+	// sig = sigmoid(glu_scaled)
+	sig := mlx.Sigmoid(gluScaled)
+
+	// out_glu = gate_clipped * sig
+	outGlu := mlx.Mul(gateClipped, sig)
+
+	// result = out_glu * (up_clipped + 1)
+	return mlx.Mul(outGlu, mlx.AddScalar(upClipped, 1.0))
+}
+
+// compiledSwiGLU is a singleton compiled SwiGLU function shared across all layers
+var compiledSwiGLU *mlx.CompiledFunc
+
+// getCompiledSwiGLU returns the compiled SwiGLU function, creating it once if needed
+func getCompiledSwiGLU() *mlx.CompiledFunc {
+	if compiledSwiGLU == nil {
+		const alpha float32 = 1.702
+		const limit float32 = 7.0
+		compiledSwiGLU = mlx.CompileShapeless(func(inputs []*mlx.Array) []*mlx.Array {
+			return []*mlx.Array{swiGLU(inputs[0], inputs[1], alpha, limit)}
+		}, true) // shapeless=true so it works for any input size
+	}
+	return compiledSwiGLU
+}
+
+// ComputeYarnFreqs computes YaRN-modified RoPE frequencies
+// Based on mlx-lm's YarnRoPE implementation
+func ComputeYarnFreqs(dims int32, base, scalingFactor float32, origMaxPos int32, betaFast, betaSlow float32) (*mlx.Array, float32) {
+	// yarn_find_correction_dim
+	yarnFindCorrectionDim := func(numRotations float64) float64 {
+		return float64(dims) * math.Log(float64(origMaxPos)/(numRotations*2*math.Pi)) / (2 * math.Log(float64(base)))
+	}
+
+	// yarn_find_correction_range
+	low := int(math.Floor(yarnFindCorrectionDim(float64(betaFast))))
+	high := int(math.Ceil(yarnFindCorrectionDim(float64(betaSlow))))
+	if low < 0 {
+		low = 0
+	}
+	if high > int(dims)-1 {
+		high = int(dims) - 1
+	}
+
+	// yarn_get_mscale
+	yarnGetMscale := func(scale, mscale float64) float64 {
+		if scale <= 1 {
+			return 1.0
+		}
+		return 0.1*mscale*math.Log(scale) + 1.0
+	}
+	mscale := float32(yarnGetMscale(float64(scalingFactor), 1.0) / yarnGetMscale(float64(scalingFactor), 0.0))
+
+	// Compute frequencies
+	// freq_extra = base ** (arange(0, dims, 2) / dims)
+	// freq_inter = scaling_factor * freq_extra
+	halfDims := dims / 2
+	freqData := make([]float32, halfDims)
+	for i := int32(0); i < halfDims; i++ {
+		exp := float64(2*i) / float64(dims)
+		freqExtra := math.Pow(float64(base), exp)
+		freqInter := float64(scalingFactor) * freqExtra
+
+		// linear ramp mask
+		var freqMask float64
+		if low == high {
+			freqMask = 0.0
+		} else {
+			t := (float64(i) - float64(low)) / float64(high-low)
+			if t < 0 {
+				t = 0
+			}
+			if t > 1 {
+				t = 1
+			}
+			freqMask = 1.0 - t
+		}
+
+		// Combined frequency: (inter * extra) / (inter * mask + extra * (1 - mask))
+		freqData[i] = float32((freqInter * freqExtra) / (freqInter*freqMask + freqExtra*(1-freqMask)))
+	}
+
+	return mlx.NewArray(freqData, []int32{halfDims}), mscale
+}
+
+// initYarn initializes YaRN RoPE if configured
+func (a *Attention) initYarn(cfg *Config) {
+	a.YarnMscale = 1.0
+	if cfg.RopeScaling != nil && cfg.RopeScaling.RopeType == "yarn" {
+		a.YarnFreqs, a.YarnMscale = ComputeYarnFreqs(
+			cfg.HeadDim,
+			cfg.RopeTheta,
+			cfg.RopeScaling.Factor,
+			cfg.RopeScaling.OriginalMaxPositionEmbeddings,
+			cfg.RopeScaling.BetaFast,
+			cfg.RopeScaling.BetaSlow,
+		)
+	}
+}
+
+func (a *Attention) Forward(x *mlx.Array, c cache.Cache, B, L int32, mask *mlx.Array, maskMode string, cfg *Config) *mlx.Array {
+	q := a.QProj.Forward(x)
+	k := a.KProj.Forward(x)
+	v := a.VProj.Forward(x)
+
+	// Reshape via AsStrided: [B, L, n_heads * head_dim] -> [B, n_heads, L, head_dim]
+	q = mlx.AsStrided(q, []int32{B, cfg.NumAttentionHeads, L, cfg.HeadDim},
+		[]int64{int64(L * cfg.NumAttentionHeads * cfg.HeadDim), int64(cfg.HeadDim), int64(cfg.NumAttentionHeads * cfg.HeadDim), 1}, 0)
+	k = mlx.AsStrided(k, []int32{B, cfg.NumKeyValueHeads, L, cfg.HeadDim},
+		[]int64{int64(L * cfg.NumKeyValueHeads * cfg.HeadDim), int64(cfg.HeadDim), int64(cfg.NumKeyValueHeads * cfg.HeadDim), 1}, 0)
+	v = mlx.AsStrided(v, []int32{B, cfg.NumKeyValueHeads, L, cfg.HeadDim},
+		[]int64{int64(L * cfg.NumKeyValueHeads * cfg.HeadDim), int64(cfg.HeadDim), int64(cfg.NumKeyValueHeads * cfg.HeadDim), 1}, 0)
+
+	offset := 0
+	if c != nil {
+		offset = c.Offset()
+	}
+	if a.YarnFreqs != nil {
+		if a.YarnMscale != 1.0 {
+			q = mlx.MulScalar(q, a.YarnMscale)
+		}
+		q = mlx.RoPEWithFreqs(q, a.YarnFreqs, int(cfg.HeadDim), false, 1.0, offset)
+		k = mlx.RoPEWithFreqs(k, a.YarnFreqs, int(cfg.HeadDim), false, 1.0, offset)
+	} else {
+		q = mlx.RoPE(q, int(cfg.HeadDim), false, cfg.RopeTheta, 1.0, offset)
+		k = mlx.RoPE(k, int(cfg.HeadDim), false, cfg.RopeTheta, 1.0, offset)
+	}
+
+	if c != nil {
+		k, v = c.Update(k, v, int(L))
+	}
+
+	out := mlx.ScaledDotProductAttentionWithSinks(q, k, v, cfg.Scale, maskMode, mask, a.Sinks)
+	out = mlx.Reshape(mlx.Transpose(out, 0, 2, 1, 3), B, L, cfg.NumAttentionHeads*cfg.HeadDim)
+	return a.OProj.Forward(out)
+}
+
+// CreateSlidingWindowMask creates a causal mask with sliding window
+// Mirrors mlx-lm's create_causal_mask with window_size
+func CreateSlidingWindowMask(seqLen, queryStart, keyStart, keyLen, windowSize int) *mlx.Array {
+	// Build mask aligned to actual cache length (may be rotated)
+	// rinds covers existing keys: [keyStart, keyStart+keyLen)
+	// linds covers new queries: [queryStart, queryStart+seqLen)
+	rinds := mlx.Arange(float32(keyStart), float32(keyStart+keyLen), 1)     // [keyLen]
+	linds := mlx.Arange(float32(queryStart), float32(queryStart+seqLen), 1) // [seqLen]
+
+	linds = mlx.ExpandDims(linds, 1) // [seqLen, 1]
+	rinds = mlx.ExpandDims(rinds, 0) // [1, keyLen]
+
+	causalMask := mlx.GreaterEqual(linds, rinds) // [seqLen, keyLen]
+	windowLimit := mlx.AddScalar(rinds, float32(windowSize))
+	windowMask := mlx.LessArray(linds, windowLimit) // [seqLen, keyLen]
+
+	return mlx.LogicalAnd(causalMask, windowMask)
+}
+
+// MoE represents the Mixture of Experts SwiGLU layer with quantized experts.
+type MoE struct {
+	Router     *nn.Linear `weight:"mlp.router"`
+	TopK       int32
+	HiddenSize int32
+	GroupSize  int
+	Bits       int
+	// Expert weights (loaded manually via sanitizeExpertWeights)
+	GateBlocks, GateScales, GateBias *mlx.Array
+	UpBlocks, UpScales, UpBias       *mlx.Array
+	DownBlocks, DownScales, DownBias *mlx.Array
+}
+
+func (moe *MoE) Forward(x *mlx.Array, B, L int32) *mlx.Array {
+	logits := moe.Router.Forward(x)
+	neg := mlx.Neg(logits)
+	part := mlx.Argpartition(neg, int(moe.TopK)-1, -1)
+	topKIdx := mlx.Slice(part, []int32{0, 0, 0}, []int32{B, L, moe.TopK})
+	topKVal := mlx.TakeAlongAxis(logits, topKIdx, -1)
+	weights := mlx.Softmax(topKVal, -1)
+
+	xFlat := mlx.Reshape(x, B*L, 1, 1, moe.HiddenSize)
+	idxFlat := mlx.Reshape(topKIdx, B*L, moe.TopK)
+
+	doSort := B*L >= 64
+	var invOrder *mlx.Array
+	sorted := false
+	n := B * L * moe.TopK
+
+	if doSort {
+		idxAll := mlx.Flatten(idxFlat)
+		order := mlx.Argsort(idxAll, 0)
+		invOrder = mlx.Argsort(order, 0)
+		xFlat = mlx.ExpandDims(mlx.Take(mlx.Squeeze(xFlat, 1), mlx.FloorDivideScalar(order, moe.TopK), 0), 1)
+		idxFlat = mlx.Reshape(mlx.Take(idxAll, order, 0), n, 1)
+		sorted = true
+	}
+
+	gate := mlx.GatherQMM(xFlat, moe.GateBlocks, moe.GateScales, nil, nil, idxFlat, true, moe.GroupSize, moe.Bits, "mxfp4", sorted)
+	up := mlx.GatherQMM(xFlat, moe.UpBlocks, moe.UpScales, nil, nil, idxFlat, true, moe.GroupSize, moe.Bits, "mxfp4", sorted)
+
+	if moe.GateBias != nil {
+		gate = mlx.Add(gate, mlx.ExpandDims(mlx.Take(moe.GateBias, idxFlat, 0), 2))
+	}
+	if moe.UpBias != nil {
+		up = mlx.Add(up, mlx.ExpandDims(mlx.Take(moe.UpBias, idxFlat, 0), 2))
+	}
+
+	hidden := getCompiledSwiGLU().Call(gate, up)[0]
+
+	down := mlx.GatherQMM(hidden, moe.DownBlocks, moe.DownScales, nil, nil, idxFlat, true, moe.GroupSize, moe.Bits, "mxfp4", sorted)
+	if moe.DownBias != nil {
+		down = mlx.Add(down, mlx.ExpandDims(mlx.Take(moe.DownBias, idxFlat, 0), 2))
+	}
+
+	if doSort {
+		down = mlx.Reshape(mlx.Take(mlx.Squeeze(mlx.Squeeze(down, 2), 1), invOrder, 0), B*L, moe.TopK, moe.HiddenSize)
+	} else {
+		down = mlx.Squeeze(down, 2)
+	}
+
+	ewFlat := mlx.Reshape(weights, B*L, moe.TopK, 1)
+	return mlx.Reshape(mlx.Sum(mlx.Mul(down, ewFlat), 1, false), B, L, moe.HiddenSize)
+}
+
+type Block struct {
+	Attention    *Attention
+	MLP          *MoE
+	InputNorm    *nn.RMSNorm `weight:"input_layernorm"`
+	PostAttnNorm *nn.RMSNorm `weight:"post_attention_layernorm"`
+	LayerType    string      // "sliding_attention" or "full_attention"
+}
+
+func (b *Block) Forward(x *mlx.Array, c cache.Cache, B, L int32, mask *mlx.Array, maskMode string, cfg *Config) *mlx.Array {
+	h := mlx.Add(x, b.Attention.Forward(b.InputNorm.Forward(x, cfg.RMSNormEps), c, B, L, mask, maskMode, cfg))
+	return mlx.Add(h, b.MLP.Forward(b.PostAttnNorm.Forward(h, cfg.RMSNormEps), B, L))
+}
+
+type Model struct {
+	EmbedTokens *nn.Embedding `weight:"model.embed_tokens"`
+	Layers      []*Block      `weight:"-"` // loaded manually due to MoE sanitization
+	Norm        *nn.RMSNorm   `weight:"model.norm"`
+	LMHead      *nn.Linear    `weight:"lm_head"`
+
+	tok *tokenizer.Tokenizer
+	*Config
+}
+
+func (m *Model) Tokenizer() *tokenizer.Tokenizer { return m.tok }
+func (m *Model) NumLayers() int                     { return len(m.Layers) }
+func (m *Model) VocabSize() int32                   { return m.Config.VocabSize }
+
+func (m *Model) NewCache(int32) []cache.Cache {
+	caches := make([]cache.Cache, len(m.Layers))
+	for i, layer := range m.Layers {
+		if layer.LayerType == "sliding_attention" && m.SlidingWindow > 0 {
+			caches[i] = cache.NewRotatingKVCache(int(m.SlidingWindow))
+		} else {
+			caches[i] = cache.NewKVCache()
+		}
+	}
+	return caches
+}
+
+func (m *Model) Forward(tokens *mlx.Array, caches []cache.Cache) *mlx.Array {
+	B, L := tokens.Shape()[0], tokens.Shape()[1]
+	x := m.EmbedTokens.Forward(tokens)
+
+	// Find representative cache indices for sliding window attention
+	var swaIdx int = -1
+	for i, layer := range m.Layers {
+		if layer.LayerType == "sliding_attention" {
+			swaIdx = i
+			break
+		}
+	}
+
+	// Create masks once at model level
+	var fullMask, swaMask *mlx.Array
+	var fullMaskMode, swaMaskMode string
+
+	if L > 1 {
+		fullMaskMode = "causal"
+		if swaIdx >= 0 && m.SlidingWindow > 0 && caches != nil {
+			c := caches[swaIdx]
+			offset := c.Offset()
+			windowSize := int(m.SlidingWindow)
+			cacheLen := min(int(L), windowSize)
+			if offset > 0 {
+				cacheLen = min(c.Len()+int(L), windowSize)
+			}
+			if int(L) > windowSize {
+				swaMask = CreateSlidingWindowMask(int(L), offset, offset+int(L)-cacheLen, cacheLen, windowSize)
+			} else {
+				swaMaskMode = "causal"
+			}
+		} else {
+			swaMaskMode = "causal"
+		}
+	}
+
+	for i, layer := range m.Layers {
+		var c cache.Cache
+		if caches != nil {
+			c = caches[i]
+		}
+		mask, maskMode := fullMask, fullMaskMode
+		if layer.LayerType == "sliding_attention" {
+			mask, maskMode = swaMask, swaMaskMode
+		}
+		x = layer.Forward(x, c, B, L, mask, maskMode, m.Config)
+	}
+
+	return m.LMHead.Forward(m.Norm.Forward(x, m.RMSNormEps))
+}
+
+// sanitizeExpertWeights splits merged gate_up weights into separate gate/up arrays.
+// MXFP4 quantized weights require contiguous memory - strided views give wrong results.
+func sanitizeExpertWeights(weights *safetensors.ModelWeights, prefix string) (moe *MoE) {
+	gateUpBlocks, _ := weights.GetTensor(prefix + ".mlp.experts.gate_up_proj_blocks")
+	gateUpScales, _ := weights.GetTensor(prefix + ".mlp.experts.gate_up_proj_scales")
+	gateUpBias, _ := weights.GetTensor(prefix + ".mlp.experts.gate_up_proj_bias")
+	downBlocks, _ := weights.GetTensor(prefix + ".mlp.experts.down_proj_blocks")
+	downScales, _ := weights.GetTensor(prefix + ".mlp.experts.down_proj_scales")
+	downBias, _ := weights.GetTensor(prefix + ".mlp.experts.down_proj_bias")
+
+	moe = &MoE{GroupSize: 32, Bits: 4, DownScales: downScales, DownBias: downBias}
+
+	if gateUpBlocks != nil {
+		gub := mlx.FlattenRange(mlx.View(gateUpBlocks, int(mlx.DtypeUint32)), -2, -1)
+		s := gub.Shape()
+		moe.GateBlocks = mlx.Contiguous(mlx.SliceStride(gub, []int32{0, 0, 0}, []int32{s[0], s[1], s[2]}, []int32{1, 2, 1}))
+		moe.UpBlocks = mlx.Contiguous(mlx.SliceStride(gub, []int32{0, 1, 0}, []int32{s[0], s[1], s[2]}, []int32{1, 2, 1}))
+	}
+	if gateUpScales != nil {
+		s := gateUpScales.Shape()
+		moe.GateScales = mlx.Contiguous(mlx.SliceStride(gateUpScales, []int32{0, 0, 0}, []int32{s[0], s[1], s[2]}, []int32{1, 2, 1}))
+		moe.UpScales = mlx.Contiguous(mlx.SliceStride(gateUpScales, []int32{0, 1, 0}, []int32{s[0], s[1], s[2]}, []int32{1, 2, 1}))
+	}
+	if gateUpBias != nil {
+		s := gateUpBias.Shape()
+		moe.GateBias = mlx.Contiguous(mlx.SliceStride(gateUpBias, []int32{0, 0}, []int32{s[0], s[1]}, []int32{1, 2}))
+		moe.UpBias = mlx.Contiguous(mlx.SliceStride(gateUpBias, []int32{0, 1}, []int32{s[0], s[1]}, []int32{1, 2}))
+	}
+	if downBlocks != nil {
+		moe.DownBlocks = mlx.FlattenRange(mlx.View(downBlocks, int(mlx.DtypeUint32)), -2, -1)
+	}
+	return moe
+}
+
+func Load(modelPath string) (*Model, error) {
+	data, err := os.ReadFile(filepath.Join(modelPath, "config.json"))
+	if err != nil {
+		return nil, fmt.Errorf("load config: %w", err)
+	}
+	var cfg Config
+	if err := json.Unmarshal(data, &cfg); err != nil {
+		return nil, fmt.Errorf("parse config: %w", err)
+	}
+	cfg.Scale = float32(1.0 / math.Sqrt(float64(cfg.HeadDim)))
+
+	weights, err := safetensors.LoadModelWeights(modelPath)
+	if err != nil {
+		return nil, fmt.Errorf("load weights: %w", err)
+	}
+
+	tok, err := tokenizer.Load(filepath.Join(modelPath, "tokenizer.json"))
+	if err != nil {
+		return nil, fmt.Errorf("load tokenizer: %w", err)
+	}
+
+	m := &Model{
+		Layers: make([]*Block, cfg.NumHiddenLayers),
+		Config: &cfg,
+		tok:    tok,
+	}
+
+	// Load simple weights via struct tags
+	if err := safetensors.LoadModule(m, weights, ""); err != nil {
+		return nil, err
+	}
+
+	// Load layers with custom MoE handling
+	for i := int32(0); i < cfg.NumHiddenLayers; i++ {
+		prefix := fmt.Sprintf("model.layers.%d", i)
+		layer := &Block{}
+		if err := safetensors.LoadModule(layer, weights, prefix); err != nil {
+			return nil, fmt.Errorf("layer %d: %w", i, err)
+		}
+
+		// Initialize attention YaRN
+		layer.Attention.initYarn(&cfg)
+
+		// Load MoE with weight sanitization
+		moe := sanitizeExpertWeights(weights, prefix)
+		moe.Router = layer.MLP.Router // Router was loaded by LoadModule
+		moe.TopK = cfg.NumExpertsPerTok
+		moe.HiddenSize = cfg.HiddenSize
+		layer.MLP = moe
+
+		// Set layer type
+		layer.LayerType = "full_attention"
+		if int(i) < len(cfg.LayerTypes) {
+			layer.LayerType = cfg.LayerTypes[i]
+		}
+
+		m.Layers[i] = layer
+	}
+
+	// Release safetensors BEFORE eval - lazy arrays have captured data,
+	// this reduces peak memory by freeing mmap during materialization
+	weights.ReleaseAll()
+	mlx.Eval(mlx.Collect(m)...)
+
+	return m, nil
+}
+
+func (m *Model) MaxContextLength() int32 {
+	if m.RopeScaling != nil && m.RopeScaling.OriginalMaxPositionEmbeddings > 0 {
+		return m.RopeScaling.OriginalMaxPositionEmbeddings
+	}
+	return 131072
+}

x/imagegen/models/llama/llama.go

@@ -0,0 +1,152 @@
+//go:build mlx
+
+package llama
+
+import (
+	"encoding/json"
+	"fmt"
+	"math"
+	"os"
+	"path/filepath"
+
+	"github.com/ollama/ollama/x/imagegen/cache"
+	"github.com/ollama/ollama/x/imagegen/mlx"
+	"github.com/ollama/ollama/x/imagegen/nn"
+	"github.com/ollama/ollama/x/imagegen/safetensors"
+	"github.com/ollama/ollama/x/imagegen/tokenizer"
+)
+
+type Config struct {
+	HiddenSize            int32   `json:"hidden_size"`
+	NumHiddenLayers       int32   `json:"num_hidden_layers"`
+	IntermediateSize      int32   `json:"intermediate_size"`
+	NumAttentionHeads     int32   `json:"num_attention_heads"`
+	NumKeyValueHeads      int32   `json:"num_key_value_heads"`
+	VocabSize             int32   `json:"vocab_size"`
+	RMSNormEps            float32 `json:"rms_norm_eps"`
+	RopeTheta             float32 `json:"rope_theta"`
+	MaxPositionEmbeddings int32   `json:"max_position_embeddings"`
+	HeadDim               int32   `json:"-"`
+	Scale                 float32 `json:"-"`
+}
+
+type Model struct {
+	EmbedTokens *nn.Embedding `weight:"model.embed_tokens"`
+	Layers      []*Layer      `weight:"model.layers"`
+	Norm        *nn.RMSNorm   `weight:"model.norm"`
+	Output      *nn.Linear    `weight:"lm_head,optional"`
+
+	tok *tokenizer.Tokenizer
+	*Config
+}
+
+type Layer struct {
+	Attention     *Attention
+	MLP           *MLP
+	AttentionNorm *nn.RMSNorm `weight:"input_layernorm"`
+	MLPNorm       *nn.RMSNorm `weight:"post_attention_layernorm"`
+}
+
+type Attention struct {
+	QProj *nn.Linear `weight:"self_attn.q_proj"`
+	KProj *nn.Linear `weight:"self_attn.k_proj"`
+	VProj *nn.Linear `weight:"self_attn.v_proj"`
+	OProj *nn.Linear `weight:"self_attn.o_proj"`
+}
+
+type MLP struct {
+	GateProj *nn.Linear `weight:"mlp.gate_proj"`
+	UpProj   *nn.Linear `weight:"mlp.up_proj"`
+	DownProj *nn.Linear `weight:"mlp.down_proj"`
+}
+
+func Load(modelPath string) (*Model, error) {
+	data, err := os.ReadFile(filepath.Join(modelPath, "config.json"))
+	if err != nil {
+		return nil, fmt.Errorf("load config: %w", err)
+	}
+	var cfg Config
+	if err := json.Unmarshal(data, &cfg); err != nil {
+		return nil, fmt.Errorf("parse config: %w", err)
+	}
+	cfg.HeadDim = cfg.HiddenSize / cfg.NumAttentionHeads
+	cfg.Scale = float32(1.0 / math.Sqrt(float64(cfg.HeadDim)))
+
+	weights, err := safetensors.LoadModelWeights(modelPath)
+	if err != nil {
+		return nil, fmt.Errorf("load weights: %w", err)
+	}
+
+	tok, err := tokenizer.Load(filepath.Join(modelPath, "tokenizer.json"))
+	if err != nil {
+		return nil, fmt.Errorf("load tokenizer: %w", err)
+	}
+
+	m := &Model{
+		Layers: make([]*Layer, cfg.NumHiddenLayers),
+		Config: &cfg,
+		tok:    tok,
+	}
+	if err := safetensors.LoadModule(m, weights, ""); err != nil {
+		return nil, err
+	}
+	m.Output = nn.NewLinear(m.EmbedTokens.Weight, nil)
+
+	mlx.Eval(mlx.Collect(m)...)
+	weights.ReleaseAll()
+
+	return m, nil
+}
+
+func (m *Model) Forward(tokens *mlx.Array, caches []cache.Cache) *mlx.Array {
+	B, L := tokens.Shape()[0], tokens.Shape()[1]
+	h := m.EmbedTokens.Forward(tokens)
+	for i, layer := range m.Layers {
+		h = layer.Forward(h, caches[i], B, L, m.Config)
+	}
+	return m.Output.Forward(m.Norm.Forward(h, m.RMSNormEps))
+}
+
+func (l *Layer) Forward(x *mlx.Array, c cache.Cache, B, L int32, cfg *Config) *mlx.Array {
+	h := mlx.Add(x, l.Attention.Forward(l.AttentionNorm.Forward(x, cfg.RMSNormEps), c, B, L, cfg))
+	return mlx.Add(h, l.MLP.Forward(l.MLPNorm.Forward(h, cfg.RMSNormEps)))
+}
+
+func (a *Attention) Forward(x *mlx.Array, c cache.Cache, B, L int32, cfg *Config) *mlx.Array {
+	q := a.QProj.Forward(x)
+	k := a.KProj.Forward(x)
+	v := a.VProj.Forward(x)
+
+	q = mlx.AsStrided(q, []int32{B, cfg.NumAttentionHeads, L, cfg.HeadDim},
+		[]int64{int64(L * cfg.NumAttentionHeads * cfg.HeadDim), int64(cfg.HeadDim), int64(cfg.NumAttentionHeads * cfg.HeadDim), 1}, 0)
+	k = mlx.AsStrided(k, []int32{B, cfg.NumKeyValueHeads, L, cfg.HeadDim},
+		[]int64{int64(L * cfg.NumKeyValueHeads * cfg.HeadDim), int64(cfg.HeadDim), int64(cfg.NumKeyValueHeads * cfg.HeadDim), 1}, 0)
+	v = mlx.AsStrided(v, []int32{B, cfg.NumKeyValueHeads, L, cfg.HeadDim},
+		[]int64{int64(L * cfg.NumKeyValueHeads * cfg.HeadDim), int64(cfg.HeadDim), int64(cfg.NumKeyValueHeads * cfg.HeadDim), 1}, 0)
+
+	q = mlx.RoPE(q, int(cfg.HeadDim), false, cfg.RopeTheta, 1.0, c.Offset())
+	k = mlx.RoPE(k, int(cfg.HeadDim), false, cfg.RopeTheta, 1.0, c.Offset())
+
+	k, v = c.Update(k, v, int(L))
+	out := mlx.ScaledDotProductAttention(q, k, v, cfg.Scale, L > 1)
+	out = mlx.Reshape(mlx.Transpose(out, 0, 2, 1, 3), B, L, cfg.NumAttentionHeads*cfg.HeadDim)
+	return a.OProj.Forward(out)
+}
+
+func (m *MLP) Forward(x *mlx.Array) *mlx.Array {
+	return m.DownProj.Forward(mlx.Mul(mlx.SiLU(m.GateProj.Forward(x)), m.UpProj.Forward(x)))
+}
+
+// Interface methods
+func (m *Model) NumLayers() int                     { return len(m.Layers) }
+func (m *Model) MaxContextLength() int32            { return m.MaxPositionEmbeddings }
+func (m *Model) VocabSize() int32                   { return m.Config.VocabSize }
+func (m *Model) Tokenizer() *tokenizer.Tokenizer { return m.tok }
+
+func (m *Model) NewCache(maxSeqLen int32) []cache.Cache {
+	caches := make([]cache.Cache, len(m.Layers))
+	for i := range caches {
+		caches[i] = cache.NewKVCache()
+	}
+	return caches
+}

x/imagegen/models/qwen_image/pipeline_test.go

@@ -0,0 +1,66 @@
+//go:build mlx
+
+package qwen_image
+
+import (
+	"os"
+	"testing"
+
+	"github.com/ollama/ollama/x/imagegen/mlx"
+)
+
+// TestPipelineOutput runs the full pipeline (integration test).
+// Skips if model weights not found. Requires ~50GB VRAM.
+func TestPipelineOutput(t *testing.T) {
+	modelPath := "../../../weights/Qwen-Image-2512"
+	if _, err := os.Stat(modelPath); os.IsNotExist(err) {
+		t.Skip("Skipping: model weights not found at " + modelPath)
+	}
+
+	// Load model
+	pm, err := LoadPersistent(modelPath)
+	if err != nil {
+		t.Skipf("Skipping: failed to load model: %v", err)
+	}
+
+	// Run 2-step pipeline (minimum for stable scheduler)
+	cfg := &GenerateConfig{
+		Prompt: "a cat",
+		Width:  256,
+		Height: 256,
+		Steps:  2,
+		Seed:   42,
+	}
+
+	output, err := pm.GenerateFromConfig(cfg)
+	if err != nil {
+		t.Fatalf("Pipeline failed: %v", err)
+	}
+	mlx.Eval(output)
+
+	// Verify output shape [1, C, H, W]
+	shape := output.Shape()
+	if len(shape) != 4 {
+		t.Errorf("Expected 4D output, got %v", shape)
+	}
+	if shape[0] != 1 || shape[1] != 3 || shape[2] != cfg.Height || shape[3] != cfg.Width {
+		t.Errorf("Shape mismatch: got %v, expected [1, 3, %d, %d]", shape, cfg.Height, cfg.Width)
+	}
+
+	// Verify values in expected range [0, 1]
+	data := output.Data()
+	minVal, maxVal := float32(1.0), float32(0.0)
+	for _, v := range data {
+		if v < minVal {
+			minVal = v
+		}
+		if v > maxVal {
+			maxVal = v
+		}
+	}
+	t.Logf("Output range: [%.4f, %.4f]", minVal, maxVal)
+
+	if minVal < -0.1 || maxVal > 1.1 {
+		t.Errorf("Output values out of range: [%.4f, %.4f]", minVal, maxVal)
+	}
+}