cmd: set codex env vars on launch and handle zstd request bodies

The Codex runner was not setting OPENAI_BASE_URL or OPENAI_API_KEY, this prevents Codex from sending requests to api.openai.com instead of the local Ollama server. This mirrors the approach used by the Claude runner.

Codex v0.98.0 sends zstd-compressed request bodies to the /v1/responses endpoint. Add decompression support in ResponsesMiddleware with an 8MB max decompressed size limit to prevent resource exhaustion.

.github/workflows/test-install.yaml

@@ -0,0 +1,22 @@
+name: test-install
+
+on:
+  pull_request:
+    paths:
+      - 'scripts/install.sh'
+      - '.github/workflows/test-install.yaml'
+
+jobs:
+  test:
+    strategy:
+      matrix:
+        os: [ubuntu-latest, macos-latest]
+    runs-on: ${{ matrix.os }}
+    steps:
+      - uses: actions/checkout@v4
+      - name: Run install script
+        run: sh ./scripts/install.sh
+        env:
+          OLLAMA_NO_START: 1 # do not start app
+      - name: Verify ollama is available
+        run: ollama --version

CMakeLists.txt

@@ -182,7 +182,7 @@ option(MLX_ENGINE "Enable MLX backend" OFF)
 
 if(MLX_ENGINE)
     message(STATUS "Setting up MLX (this takes a while...)")
-    add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/x/ml/backend/mlx)
+    add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/x/imagegen/mlx)
 
     # Find CUDA toolkit if MLX is built with CUDA support
     find_package(CUDAToolkit)
@@ -216,4 +216,4 @@ if(MLX_ENGINE)
                 COMPONENT MLX)
         endif()
     endif()
-endif()
+endif()

anthropic/anthropic.go

@@ -811,7 +811,9 @@ type CountTokensResponse struct {
 	InputTokens int `json:"input_tokens"`
 }
 
-// estimateTokens returns a rough estimate of tokens (len/4)
+// estimateTokens returns a rough estimate of tokens (len/4).
+// TODO: Replace with actual tokenization via Tokenize API for accuracy.
+// Current len/4 heuristic is a rough approximation (~4 chars/token average).
 func estimateTokens(req CountTokensRequest) int {
 	var totalLen int
 

anthropic/anthropic_test.go

@@ -321,8 +321,6 @@ func TestFromMessagesRequest_WithThinking(t *testing.T) {
 	}
 }
 
-// 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",
@@ -842,10 +840,6 @@ func TestStreamConverter_MultipleToolCallsWithMixedValidity(t *testing.T) {
 	}
 }
 
-// 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
@@ -899,8 +893,6 @@ func TestContentBlockJSON_EmptyFieldsPresent(t *testing.T) {
 	}
 }
 
-// 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", 0)
@@ -969,3 +961,105 @@ func TestStreamConverter_ContentBlockStartIncludesEmptyFields(t *testing.T) {
 		}
 	})
 }
+
+func TestEstimateTokens_SimpleMessage(t *testing.T) {
+	req := CountTokensRequest{
+		Model: "test-model",
+		Messages: []MessageParam{
+			{Role: "user", Content: "Hello, world!"},
+		},
+	}
+
+	tokens := estimateTokens(req)
+
+	// "user" (4) + "Hello, world!" (13) = 17 chars / 4 = 4 tokens
+	if tokens < 1 {
+		t.Errorf("expected at least 1 token, got %d", tokens)
+	}
+	// Sanity check: shouldn't be wildly off
+	if tokens > 10 {
+		t.Errorf("expected fewer than 10 tokens for short message, got %d", tokens)
+	}
+}
+
+func TestEstimateTokens_WithSystemPrompt(t *testing.T) {
+	req := CountTokensRequest{
+		Model:  "test-model",
+		System: "You are a helpful assistant.",
+		Messages: []MessageParam{
+			{Role: "user", Content: "Hello"},
+		},
+	}
+
+	tokens := estimateTokens(req)
+
+	// System prompt adds to count
+	if tokens < 5 {
+		t.Errorf("expected at least 5 tokens with system prompt, got %d", tokens)
+	}
+}
+
+func TestEstimateTokens_WithTools(t *testing.T) {
+	req := CountTokensRequest{
+		Model: "test-model",
+		Messages: []MessageParam{
+			{Role: "user", Content: "What's the weather?"},
+		},
+		Tools: []Tool{
+			{
+				Name:        "get_weather",
+				Description: "Get the current weather for a location",
+				InputSchema: json.RawMessage(`{"type":"object","properties":{"location":{"type":"string"}}}`),
+			},
+		},
+	}
+
+	tokens := estimateTokens(req)
+
+	// Tools add significant content
+	if tokens < 10 {
+		t.Errorf("expected at least 10 tokens with tools, got %d", tokens)
+	}
+}
+
+func TestEstimateTokens_WithThinking(t *testing.T) {
+	req := CountTokensRequest{
+		Model: "test-model",
+		Messages: []MessageParam{
+			{Role: "user", Content: "Hello"},
+			{
+				Role: "assistant",
+				Content: []any{
+					map[string]any{
+						"type":     "thinking",
+						"thinking": "Let me think about this carefully...",
+					},
+					map[string]any{
+						"type": "text",
+						"text": "Here is my response.",
+					},
+				},
+			},
+		},
+	}
+
+	tokens := estimateTokens(req)
+
+	// Thinking content should be counted
+	if tokens < 10 {
+		t.Errorf("expected at least 10 tokens with thinking content, got %d", tokens)
+	}
+}
+
+func TestEstimateTokens_EmptyContent(t *testing.T) {
+	req := CountTokensRequest{
+		Model:    "test-model",
+		Messages: []MessageParam{},
+	}
+
+	tokens := estimateTokens(req)
+
+	if tokens != 0 {
+		t.Errorf("expected 0 tokens for empty content, got %d", tokens)
+	}
+}

api/client.go

@@ -478,3 +478,13 @@ type AliasRequest struct {
 func (c *Client) SetAliasExperimental(ctx context.Context, req *AliasRequest) error {
 	return c.do(ctx, http.MethodPost, "/api/experimental/aliases", req, nil)
 }
+
+// AliasDeleteRequest is the request body for deleting a model alias.
+type AliasDeleteRequest struct {
+	Alias string `json:"alias"`
+}
+
+// DeleteAliasExperimental deletes a model alias via the experimental aliases API.
+func (c *Client) DeleteAliasExperimental(ctx context.Context, req *AliasDeleteRequest) error {
+	return c.do(ctx, http.MethodDelete, "/api/experimental/aliases", req, nil)
+}

cmd/cmd.go

@@ -367,14 +367,25 @@ func loadOrUnloadModel(cmd *cobra.Command, opts *runOptions) error {
 		return err
 	} else if info.RemoteHost != "" {
 		// Cloud model, no need to load/unload
+
+		isCloud := strings.HasPrefix(info.RemoteHost, "https://ollama.com")
+
+		// Check if user is signed in for ollama.com cloud models
+		if isCloud {
+			if _, err := client.Whoami(cmd.Context()); err != nil {
+				return err
+			}
+		}
+
 		if opts.ShowConnect {
 			p.StopAndClear()
-			if strings.HasPrefix(info.RemoteHost, "https://ollama.com") {
+			if isCloud {
 				fmt.Fprintf(os.Stderr, "Connecting to '%s' on 'ollama.com' ⚡\n", info.RemoteModel)
 			} else {
 				fmt.Fprintf(os.Stderr, "Connecting to '%s' on '%s'\n", info.RemoteModel, info.RemoteHost)
 			}
 		}
+
 		return nil
 	}
 
@@ -1752,7 +1763,7 @@ func checkServerHeartbeat(cmd *cobra.Command, _ []string) error {
 			return err
 		}
 		if err := startApp(cmd.Context(), client); err != nil {
-			return fmt.Errorf("ollama server not responding - %w", err)
+			return err
 		}
 	}
 	return nil

cmd/cmd_test.go

@@ -3,6 +3,7 @@ package cmd
 import (
 	"bytes"
 	"encoding/json"
+	"errors"
 	"fmt"
 	"io"
 	"net/http"
@@ -1659,3 +1660,103 @@ func TestRunOptions_Copy_Independence(t *testing.T) {
 		t.Error("Copy Think should not be affected by original modification")
 	}
 }
+
+func TestLoadOrUnloadModel_CloudModelAuth(t *testing.T) {
+	tests := []struct {
+		name          string
+		remoteHost    string
+		whoamiStatus  int
+		whoamiResp    any
+		expectedError string
+	}{
+		{
+			name:         "ollama.com cloud model - user signed in",
+			remoteHost:   "https://ollama.com",
+			whoamiStatus: http.StatusOK,
+			whoamiResp:   api.UserResponse{Name: "testuser"},
+		},
+		{
+			name:         "ollama.com cloud model - user not signed in",
+			remoteHost:   "https://ollama.com",
+			whoamiStatus: http.StatusUnauthorized,
+			whoamiResp: map[string]string{
+				"error":      "unauthorized",
+				"signin_url": "https://ollama.com/signin",
+			},
+			expectedError: "unauthorized",
+		},
+		{
+			name:         "non-ollama.com remote - no auth check",
+			remoteHost:   "https://other-remote.com",
+			whoamiStatus: http.StatusUnauthorized, // should not be called
+			whoamiResp:   nil,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			whoamiCalled := false
+			mockServer := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+				switch r.URL.Path {
+				case "/api/show":
+					w.Header().Set("Content-Type", "application/json")
+					if err := json.NewEncoder(w).Encode(api.ShowResponse{
+						RemoteHost:  tt.remoteHost,
+						RemoteModel: "test-model",
+					}); err != nil {
+						http.Error(w, err.Error(), http.StatusInternalServerError)
+					}
+				case "/api/me":
+					whoamiCalled = true
+					w.Header().Set("Content-Type", "application/json")
+					w.WriteHeader(tt.whoamiStatus)
+					if tt.whoamiResp != nil {
+						if err := json.NewEncoder(w).Encode(tt.whoamiResp); err != nil {
+							http.Error(w, err.Error(), http.StatusInternalServerError)
+						}
+					}
+				default:
+					http.NotFound(w, r)
+				}
+			}))
+			defer mockServer.Close()
+
+			t.Setenv("OLLAMA_HOST", mockServer.URL)
+
+			cmd := &cobra.Command{}
+			cmd.SetContext(t.Context())
+
+			opts := &runOptions{
+				Model:       "test-cloud-model",
+				ShowConnect: false,
+			}
+
+			err := loadOrUnloadModel(cmd, opts)
+
+			if strings.HasPrefix(tt.remoteHost, "https://ollama.com") {
+				if !whoamiCalled {
+					t.Error("expected whoami to be called for ollama.com cloud model")
+				}
+			} else {
+				if whoamiCalled {
+					t.Error("whoami should not be called for non-ollama.com remote")
+				}
+			}
+
+			if tt.expectedError != "" {
+				if err == nil {
+					t.Errorf("expected error containing %q, got nil", tt.expectedError)
+				} else {
+					var authErr api.AuthorizationError
+					if !errors.As(err, &authErr) {
+						t.Errorf("expected AuthorizationError, got %T: %v", err, err)
+					}
+				}
+			} else {
+				if err != nil {
+					t.Errorf("expected no error, got %v", err)
+				}
+			}
+		})
+	}
+}

cmd/config/claude.go

@@ -58,27 +58,64 @@ func (c *Claude) Run(model string, args []string) error {
 	cmd.Stdin = os.Stdin
 	cmd.Stdout = os.Stdout
 	cmd.Stderr = os.Stderr
-	cmd.Env = append(os.Environ(),
+
+	env := append(os.Environ(),
 		"ANTHROPIC_BASE_URL="+envconfig.Host().String(),
 		"ANTHROPIC_API_KEY=",
 		"ANTHROPIC_AUTH_TOKEN=ollama",
 	)
+
+	env = append(env, c.modelEnvVars(model)...)
+
+	cmd.Env = env
 	return cmd.Run()
 }
 
-// ConfigureAliases sets up Primary and Fast model aliases for Claude Code.
-func (c *Claude) ConfigureAliases(ctx context.Context, primaryModel string, existing map[string]string, force bool) (map[string]string, bool, error) {
+// modelEnvVars returns Claude Code env vars that route all model tiers through Ollama.
+func (c *Claude) modelEnvVars(model string) []string {
+	primary := model
+	fast := model
+	if cfg, err := loadIntegration("claude"); err == nil && cfg.Aliases != nil {
+		if p := cfg.Aliases["primary"]; p != "" {
+			primary = p
+		}
+		if f := cfg.Aliases["fast"]; f != "" {
+			fast = f
+		}
+	}
+	return []string{
+		"ANTHROPIC_DEFAULT_OPUS_MODEL=" + primary,
+		"ANTHROPIC_DEFAULT_SONNET_MODEL=" + primary,
+		"ANTHROPIC_DEFAULT_HAIKU_MODEL=" + fast,
+		"CLAUDE_CODE_SUBAGENT_MODEL=" + primary,
+	}
+}
+
+// ConfigureAliases sets up model aliases for Claude Code.
+// model: the model to use (if empty, user will be prompted to select)
+// aliases: existing alias configuration to preserve/update
+// Cloud-only: subagent routing (fast model) is gated to cloud models only until
+// there is a better strategy for prompt caching on local models.
+func (c *Claude) ConfigureAliases(ctx context.Context, model string, existingAliases map[string]string, force bool) (map[string]string, bool, error) {
 	aliases := make(map[string]string)
-	for k, v := range existing {
+	for k, v := range existingAliases {
 		aliases[k] = v
 	}
 
-	if primaryModel != "" {
-		aliases["primary"] = primaryModel
+	if model != "" {
+		aliases["primary"] = model
 	}
 
-	if !force && aliases["primary"] != "" && aliases["fast"] != "" {
-		return aliases, false, nil
+	if !force && aliases["primary"] != "" {
+		client, _ := api.ClientFromEnvironment()
+		if isCloudModel(ctx, client, aliases["primary"]) {
+			if isCloudModel(ctx, client, aliases["fast"]) {
+				return aliases, false, nil
+			}
+		} else {
+			delete(aliases, "fast")
+			return aliases, false, nil
+		}
 	}
 
 	items, existingModels, cloudModels, client, err := listModels(ctx)
@@ -86,14 +123,11 @@ func (c *Claude) ConfigureAliases(ctx context.Context, primaryModel string, exis
 		return nil, false, err
 	}
 
-	fmt.Fprintf(os.Stderr, "\n%sModel Configuration%s\n", ansiBold, ansiReset)
-	fmt.Fprintf(os.Stderr, "%sClaude Code uses multiple models for various tasks%s\n\n", ansiGray, ansiReset)
-
-	fmt.Fprintf(os.Stderr, "%sPrimary%s\n", ansiBold, ansiReset)
-	fmt.Fprintf(os.Stderr, "%sHandles complex reasoning: planning, code generation, debugging.%s\n\n", ansiGray, ansiReset)
+	fmt.Fprintf(os.Stderr, "\n%sModel Configuration%s\n\n", ansiBold, ansiReset)
 
 	if aliases["primary"] == "" || force {
-		primary, err := selectPrompt("Select Primary model:", items)
+		primary, err := selectPrompt("Select model:", items)
+		fmt.Fprintf(os.Stderr, "\033[3A\033[J")
 		if err != nil {
 			return nil, false, err
 		}
@@ -104,38 +138,37 @@ func (c *Claude) ConfigureAliases(ctx context.Context, primaryModel string, exis
 			return nil, false, err
 		}
 		aliases["primary"] = primary
-	} else {
-		fmt.Fprintf(os.Stderr, "  %s\n\n", aliases["primary"])
 	}
 
-	fmt.Fprintf(os.Stderr, "%sFast%s\n", ansiBold, ansiReset)
-	fmt.Fprintf(os.Stderr, "%sHandles quick operations: file searches, simple edits, status checks.%s\n", ansiGray, ansiReset)
-	fmt.Fprintf(os.Stderr, "%sSmaller models work well and respond faster.%s\n\n", ansiGray, ansiReset)
-
-	if aliases["fast"] == "" || force {
-		fast, err := selectPrompt("Select Fast model:", items)
-		if err != nil {
-			return nil, false, err
+	if isCloudModel(ctx, client, aliases["primary"]) {
+		if aliases["fast"] == "" || !isCloudModel(ctx, client, aliases["fast"]) {
+			aliases["fast"] = aliases["primary"]
 		}
-		if err := pullIfNeeded(ctx, client, existingModels, fast); err != nil {
-			return nil, false, err
-		}
-		if err := ensureAuth(ctx, client, cloudModels, []string{fast}); err != nil {
-			return nil, false, err
-		}
-		aliases["fast"] = fast
+	} else {
+		delete(aliases, "fast")
 	}
 
 	return aliases, true, nil
 }
 
 // SetAliases syncs the configured aliases to the Ollama server using prefix matching.
+// Cloud-only: for local models (fast is empty), we delete any existing aliases to
+// prevent stale routing to a previous cloud model.
 func (c *Claude) SetAliases(ctx context.Context, aliases map[string]string) error {
 	client, err := api.ClientFromEnvironment()
 	if err != nil {
 		return err
 	}
 
+	prefixes := []string{"claude-sonnet-", "claude-haiku-"}
+
+	if aliases["fast"] == "" {
+		for _, prefix := range prefixes {
+			_ = client.DeleteAliasExperimental(ctx, &api.AliasDeleteRequest{Alias: prefix})
+		}
+		return nil
+	}
+
 	prefixAliases := map[string]string{
 		"claude-sonnet-": aliases["primary"],
 		"claude-haiku-":  aliases["fast"],

cmd/config/claude_test.go

@@ -5,6 +5,7 @@ import (
 	"path/filepath"
 	"runtime"
 	"slices"
+	"strings"
 	"testing"
 )
 
@@ -103,3 +104,95 @@ func TestClaudeArgs(t *testing.T) {
 		})
 	}
 }
+
+func TestClaudeModelEnvVars(t *testing.T) {
+	c := &Claude{}
+
+	envMap := func(envs []string) map[string]string {
+		m := make(map[string]string)
+		for _, e := range envs {
+			k, v, _ := strings.Cut(e, "=")
+			m[k] = v
+		}
+		return m
+	}
+
+	t.Run("falls back to model param when no aliases saved", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+
+		got := envMap(c.modelEnvVars("llama3.2"))
+		if got["ANTHROPIC_DEFAULT_OPUS_MODEL"] != "llama3.2" {
+			t.Errorf("OPUS = %q, want llama3.2", got["ANTHROPIC_DEFAULT_OPUS_MODEL"])
+		}
+		if got["ANTHROPIC_DEFAULT_SONNET_MODEL"] != "llama3.2" {
+			t.Errorf("SONNET = %q, want llama3.2", got["ANTHROPIC_DEFAULT_SONNET_MODEL"])
+		}
+		if got["ANTHROPIC_DEFAULT_HAIKU_MODEL"] != "llama3.2" {
+			t.Errorf("HAIKU = %q, want llama3.2", got["ANTHROPIC_DEFAULT_HAIKU_MODEL"])
+		}
+		if got["CLAUDE_CODE_SUBAGENT_MODEL"] != "llama3.2" {
+			t.Errorf("SUBAGENT = %q, want llama3.2", got["CLAUDE_CODE_SUBAGENT_MODEL"])
+		}
+	})
+
+	t.Run("uses primary alias for opus sonnet and subagent", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+
+		saveIntegration("claude", []string{"qwen3:8b"})
+		saveAliases("claude", map[string]string{"primary": "qwen3:8b"})
+
+		got := envMap(c.modelEnvVars("qwen3:8b"))
+		if got["ANTHROPIC_DEFAULT_OPUS_MODEL"] != "qwen3:8b" {
+			t.Errorf("OPUS = %q, want qwen3:8b", got["ANTHROPIC_DEFAULT_OPUS_MODEL"])
+		}
+		if got["ANTHROPIC_DEFAULT_SONNET_MODEL"] != "qwen3:8b" {
+			t.Errorf("SONNET = %q, want qwen3:8b", got["ANTHROPIC_DEFAULT_SONNET_MODEL"])
+		}
+		if got["ANTHROPIC_DEFAULT_HAIKU_MODEL"] != "qwen3:8b" {
+			t.Errorf("HAIKU = %q, want qwen3:8b (no fast alias)", got["ANTHROPIC_DEFAULT_HAIKU_MODEL"])
+		}
+		if got["CLAUDE_CODE_SUBAGENT_MODEL"] != "qwen3:8b" {
+			t.Errorf("SUBAGENT = %q, want qwen3:8b", got["CLAUDE_CODE_SUBAGENT_MODEL"])
+		}
+	})
+
+	t.Run("uses fast alias for haiku", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+
+		saveIntegration("claude", []string{"llama3.2:70b"})
+		saveAliases("claude", map[string]string{
+			"primary": "llama3.2:70b",
+			"fast":    "llama3.2:8b",
+		})
+
+		got := envMap(c.modelEnvVars("llama3.2:70b"))
+		if got["ANTHROPIC_DEFAULT_OPUS_MODEL"] != "llama3.2:70b" {
+			t.Errorf("OPUS = %q, want llama3.2:70b", got["ANTHROPIC_DEFAULT_OPUS_MODEL"])
+		}
+		if got["ANTHROPIC_DEFAULT_SONNET_MODEL"] != "llama3.2:70b" {
+			t.Errorf("SONNET = %q, want llama3.2:70b", got["ANTHROPIC_DEFAULT_SONNET_MODEL"])
+		}
+		if got["ANTHROPIC_DEFAULT_HAIKU_MODEL"] != "llama3.2:8b" {
+			t.Errorf("HAIKU = %q, want llama3.2:8b", got["ANTHROPIC_DEFAULT_HAIKU_MODEL"])
+		}
+		if got["CLAUDE_CODE_SUBAGENT_MODEL"] != "llama3.2:70b" {
+			t.Errorf("SUBAGENT = %q, want llama3.2:70b", got["CLAUDE_CODE_SUBAGENT_MODEL"])
+		}
+	})
+
+	t.Run("alias primary overrides model param", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+
+		saveIntegration("claude", []string{"saved-model"})
+		saveAliases("claude", map[string]string{"primary": "saved-model"})
+
+		got := envMap(c.modelEnvVars("different-model"))
+		if got["ANTHROPIC_DEFAULT_OPUS_MODEL"] != "saved-model" {
+			t.Errorf("OPUS = %q, want saved-model", got["ANTHROPIC_DEFAULT_OPUS_MODEL"])
+		}
+	})
+}

cmd/config/codex.go

@@ -6,6 +6,7 @@ import (
 	"os/exec"
 	"strings"
 
+	"github.com/ollama/ollama/envconfig"
 	"golang.org/x/mod/semver"
 )
 
@@ -32,6 +33,10 @@ func (c *Codex) Run(model string, args []string) error {
 	cmd.Stdin = os.Stdin
 	cmd.Stdout = os.Stdout
 	cmd.Stderr = os.Stderr
+	cmd.Env = append(os.Environ(),
+		"OPENAI_BASE_URL="+envconfig.Host().String()+"/v1/",
+		"OPENAI_API_KEY=ollama",
+	)
 	return cmd.Run()
 }
 

cmd/config/config.go

@@ -6,7 +6,6 @@ import (
 	"encoding/json"
 	"errors"
 	"fmt"
-	"log/slog"
 	"os"
 	"path/filepath"
 	"strings"
@@ -54,7 +53,6 @@ func migrateConfig() (bool, error) {
 
 	var js json.RawMessage
 	if err := json.Unmarshal(oldData, &js); err != nil {
-		slog.Warn("legacy config has invalid JSON, skipping migration", "path", oldPath, "error", err)
 		return false, nil
 	}
 
@@ -73,7 +71,6 @@ func migrateConfig() (bool, error) {
 	_ = os.Remove(oldPath)
 	_ = os.Remove(filepath.Dir(oldPath)) // clean up empty directory
 
-	slog.Info("migrated config", "from", oldPath, "to", newPath)
 	return true, nil
 }
 
@@ -179,12 +176,8 @@ func saveAliases(appName string, aliases map[string]string) error {
 		existing = &integration{}
 	}
 
-	if existing.Aliases == nil {
-		existing.Aliases = make(map[string]string)
-	}
-	for k, v := range aliases {
-		existing.Aliases[k] = v
-	}
+	// Replace aliases entirely (not merge) so deletions are persisted
+	existing.Aliases = aliases
 
 	cfg.Integrations[key] = existing
 	return save(cfg)

cmd/config/config_cloud_test.go

@@ -0,0 +1,677 @@
+package config
+
+import (
+	"context"
+	"errors"
+	"os"
+	"path/filepath"
+	"testing"
+)
+
+func TestSetAliases_CloudModel(t *testing.T) {
+	// Test the SetAliases logic by checking the alias map behavior
+	aliases := map[string]string{
+		"primary": "kimi-k2.5:cloud",
+		"fast":    "kimi-k2.5:cloud",
+	}
+
+	// Verify fast is set (cloud model behavior)
+	if aliases["fast"] == "" {
+		t.Error("cloud model should have fast alias set")
+	}
+	if aliases["fast"] != aliases["primary"] {
+		t.Errorf("fast should equal primary for auto-set, got fast=%q primary=%q", aliases["fast"], aliases["primary"])
+	}
+}
+
+func TestSetAliases_LocalModel(t *testing.T) {
+	aliases := map[string]string{
+		"primary": "llama3.2:latest",
+	}
+	// Simulate local model behavior: fast should be empty
+	delete(aliases, "fast")
+
+	if aliases["fast"] != "" {
+		t.Error("local model should have empty fast alias")
+	}
+}
+
+func TestSaveAliases_ReplacesNotMerges(t *testing.T) {
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	// First save with both primary and fast
+	initial := map[string]string{
+		"primary": "cloud-model",
+		"fast":    "cloud-model",
+	}
+	if err := saveAliases("claude", initial); err != nil {
+		t.Fatalf("failed to save initial aliases: %v", err)
+	}
+
+	// Verify both are saved
+	loaded, err := loadIntegration("claude")
+	if err != nil {
+		t.Fatalf("failed to load: %v", err)
+	}
+	if loaded.Aliases["fast"] != "cloud-model" {
+		t.Errorf("expected fast=cloud-model, got %q", loaded.Aliases["fast"])
+	}
+
+	// Now save without fast (simulating switch to local model)
+	updated := map[string]string{
+		"primary": "local-model",
+		// fast intentionally missing
+	}
+	if err := saveAliases("claude", updated); err != nil {
+		t.Fatalf("failed to save updated aliases: %v", err)
+	}
+
+	// Verify fast is GONE (not merged/preserved)
+	loaded, err = loadIntegration("claude")
+	if err != nil {
+		t.Fatalf("failed to load after update: %v", err)
+	}
+	if loaded.Aliases["fast"] != "" {
+		t.Errorf("fast should be removed after saving without it, got %q", loaded.Aliases["fast"])
+	}
+	if loaded.Aliases["primary"] != "local-model" {
+		t.Errorf("primary should be updated to local-model, got %q", loaded.Aliases["primary"])
+	}
+}
+
+func TestSaveAliases_PreservesModels(t *testing.T) {
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	// First save integration with models
+	if err := saveIntegration("claude", []string{"model1", "model2"}); err != nil {
+		t.Fatalf("failed to save integration: %v", err)
+	}
+
+	// Then update aliases
+	aliases := map[string]string{"primary": "new-model"}
+	if err := saveAliases("claude", aliases); err != nil {
+		t.Fatalf("failed to save aliases: %v", err)
+	}
+
+	// Verify models are preserved
+	loaded, err := loadIntegration("claude")
+	if err != nil {
+		t.Fatalf("failed to load: %v", err)
+	}
+	if len(loaded.Models) != 2 || loaded.Models[0] != "model1" {
+		t.Errorf("models should be preserved, got %v", loaded.Models)
+	}
+}
+
+// TestSaveAliases_EmptyMap clears all aliases
+func TestSaveAliases_EmptyMap(t *testing.T) {
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	// Save with aliases
+	if err := saveAliases("claude", map[string]string{"primary": "model", "fast": "model"}); err != nil {
+		t.Fatalf("failed to save: %v", err)
+	}
+
+	// Save empty map
+	if err := saveAliases("claude", map[string]string{}); err != nil {
+		t.Fatalf("failed to save empty: %v", err)
+	}
+
+	loaded, err := loadIntegration("claude")
+	if err != nil {
+		t.Fatalf("failed to load: %v", err)
+	}
+	if len(loaded.Aliases) != 0 {
+		t.Errorf("aliases should be empty, got %v", loaded.Aliases)
+	}
+}
+
+// TestSaveAliases_NilMap handles nil gracefully
+func TestSaveAliases_NilMap(t *testing.T) {
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	// Save with aliases first
+	if err := saveAliases("claude", map[string]string{"primary": "model"}); err != nil {
+		t.Fatalf("failed to save: %v", err)
+	}
+
+	// Save nil map - should clear aliases
+	if err := saveAliases("claude", nil); err != nil {
+		t.Fatalf("failed to save nil: %v", err)
+	}
+
+	loaded, err := loadIntegration("claude")
+	if err != nil {
+		t.Fatalf("failed to load: %v", err)
+	}
+	if len(loaded.Aliases) > 0 {
+		t.Errorf("aliases should be nil or empty, got %v", loaded.Aliases)
+	}
+}
+
+// TestSaveAliases_EmptyAppName returns error
+func TestSaveAliases_EmptyAppName(t *testing.T) {
+	err := saveAliases("", map[string]string{"primary": "model"})
+	if err == nil {
+		t.Error("expected error for empty app name")
+	}
+}
+
+func TestSaveAliases_CaseInsensitive(t *testing.T) {
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	if err := saveAliases("Claude", map[string]string{"primary": "model1"}); err != nil {
+		t.Fatalf("failed to save: %v", err)
+	}
+
+	// Load with different case
+	loaded, err := loadIntegration("claude")
+	if err != nil {
+		t.Fatalf("failed to load: %v", err)
+	}
+	if loaded.Aliases["primary"] != "model1" {
+		t.Errorf("expected primary=model1, got %q", loaded.Aliases["primary"])
+	}
+
+	// Update with different case
+	if err := saveAliases("CLAUDE", map[string]string{"primary": "model2"}); err != nil {
+		t.Fatalf("failed to update: %v", err)
+	}
+
+	loaded, err = loadIntegration("claude")
+	if err != nil {
+		t.Fatalf("failed to load after update: %v", err)
+	}
+	if loaded.Aliases["primary"] != "model2" {
+		t.Errorf("expected primary=model2, got %q", loaded.Aliases["primary"])
+	}
+}
+
+// TestSaveAliases_CreatesIntegration creates integration if it doesn't exist
+func TestSaveAliases_CreatesIntegration(t *testing.T) {
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	// Save aliases for non-existent integration
+	if err := saveAliases("newintegration", map[string]string{"primary": "model"}); err != nil {
+		t.Fatalf("failed to save: %v", err)
+	}
+
+	loaded, err := loadIntegration("newintegration")
+	if err != nil {
+		t.Fatalf("failed to load: %v", err)
+	}
+	if loaded.Aliases["primary"] != "model" {
+		t.Errorf("expected primary=model, got %q", loaded.Aliases["primary"])
+	}
+}
+
+func TestConfigureAliases_AliasMap(t *testing.T) {
+	t.Run("cloud model auto-sets fast to primary", func(t *testing.T) {
+		aliases := make(map[string]string)
+		aliases["primary"] = "cloud-model"
+
+		// Simulate cloud model behavior
+		isCloud := true
+		if isCloud {
+			if aliases["fast"] == "" {
+				aliases["fast"] = aliases["primary"]
+			}
+		}
+
+		if aliases["fast"] != "cloud-model" {
+			t.Errorf("expected fast=cloud-model, got %q", aliases["fast"])
+		}
+	})
+
+	t.Run("cloud model preserves custom fast", func(t *testing.T) {
+		aliases := map[string]string{
+			"primary": "cloud-model",
+			"fast":    "custom-fast-model",
+		}
+
+		// Simulate cloud model behavior - should preserve existing fast
+		isCloud := true
+		if isCloud {
+			if aliases["fast"] == "" {
+				aliases["fast"] = aliases["primary"]
+			}
+		}
+
+		if aliases["fast"] != "custom-fast-model" {
+			t.Errorf("expected fast=custom-fast-model (preserved), got %q", aliases["fast"])
+		}
+	})
+
+	t.Run("local model clears fast", func(t *testing.T) {
+		aliases := map[string]string{
+			"primary": "local-model",
+			"fast":    "should-be-cleared",
+		}
+
+		// Simulate local model behavior
+		isCloud := false
+		if !isCloud {
+			delete(aliases, "fast")
+		}
+
+		if aliases["fast"] != "" {
+			t.Errorf("expected fast to be cleared, got %q", aliases["fast"])
+		}
+	})
+
+	t.Run("switching cloud to local clears fast", func(t *testing.T) {
+		// Start with cloud config
+		aliases := map[string]string{
+			"primary": "cloud-model",
+			"fast":    "cloud-model",
+		}
+
+		// Switch to local
+		aliases["primary"] = "local-model"
+		isCloud := false
+		if !isCloud {
+			delete(aliases, "fast")
+		}
+
+		if aliases["fast"] != "" {
+			t.Errorf("fast should be cleared when switching to local, got %q", aliases["fast"])
+		}
+		if aliases["primary"] != "local-model" {
+			t.Errorf("primary should be updated, got %q", aliases["primary"])
+		}
+	})
+
+	t.Run("switching local to cloud sets fast", func(t *testing.T) {
+		// Start with local config (no fast)
+		aliases := map[string]string{
+			"primary": "local-model",
+		}
+
+		// Switch to cloud
+		aliases["primary"] = "cloud-model"
+		isCloud := true
+		if isCloud {
+			if aliases["fast"] == "" {
+				aliases["fast"] = aliases["primary"]
+			}
+		}
+
+		if aliases["fast"] != "cloud-model" {
+			t.Errorf("fast should be set when switching to cloud, got %q", aliases["fast"])
+		}
+	})
+}
+
+func TestSetAliases_PrefixMapping(t *testing.T) {
+	// This tests the expected mapping without needing a real client
+	aliases := map[string]string{
+		"primary": "my-cloud-model",
+		"fast":    "my-fast-model",
+	}
+
+	expectedMappings := map[string]string{
+		"claude-sonnet-": aliases["primary"],
+		"claude-haiku-":  aliases["fast"],
+	}
+
+	if expectedMappings["claude-sonnet-"] != "my-cloud-model" {
+		t.Errorf("claude-sonnet- should map to primary")
+	}
+	if expectedMappings["claude-haiku-"] != "my-fast-model" {
+		t.Errorf("claude-haiku- should map to fast")
+	}
+}
+
+func TestSetAliases_LocalDeletesPrefixes(t *testing.T) {
+	aliases := map[string]string{
+		"primary": "local-model",
+		// fast is empty/missing - indicates local model
+	}
+
+	prefixesToDelete := []string{"claude-sonnet-", "claude-haiku-"}
+
+	// Verify the logic: when fast is empty, we should delete
+	if aliases["fast"] != "" {
+		t.Error("fast should be empty for local model")
+	}
+
+	// Verify we have the right prefixes to delete
+	if len(prefixesToDelete) != 2 {
+		t.Errorf("expected 2 prefixes to delete, got %d", len(prefixesToDelete))
+	}
+}
+
+// TestAtomicUpdate_ServerFailsConfigNotSaved simulates atomic update behavior
+func TestAtomicUpdate_ServerFailsConfigNotSaved(t *testing.T) {
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	// Simulate: server fails, config should NOT be saved
+	serverErr := errors.New("server unavailable")
+
+	if serverErr == nil {
+		t.Error("config should NOT be saved when server fails")
+	}
+}
+
+// TestAtomicUpdate_ServerSucceedsConfigSaved simulates successful atomic update
+func TestAtomicUpdate_ServerSucceedsConfigSaved(t *testing.T) {
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	// Simulate: server succeeds, config should be saved
+	var serverErr error
+	if serverErr != nil {
+		t.Fatal("server should succeed")
+	}
+
+	if err := saveAliases("claude", map[string]string{"primary": "model"}); err != nil {
+		t.Fatalf("saveAliases failed: %v", err)
+	}
+
+	// Verify it was actually saved
+	loaded, err := loadIntegration("claude")
+	if err != nil {
+		t.Fatalf("failed to load: %v", err)
+	}
+	if loaded.Aliases["primary"] != "model" {
+		t.Errorf("expected primary=model, got %q", loaded.Aliases["primary"])
+	}
+}
+
+func TestConfigFile_PreservesUnknownFields(t *testing.T) {
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	// Write config with extra fields
+	configPath := filepath.Join(tmpDir, ".ollama", "config.json")
+	os.MkdirAll(filepath.Dir(configPath), 0o755)
+
+	// Note: Our config struct only has Integrations, so top-level unknown fields
+	// won't be preserved by our current implementation. This test documents that.
+	initialConfig := `{
+  "integrations": {
+    "claude": {
+      "models": ["model1"],
+      "aliases": {"primary": "model1"},
+      "unknownField": "should be lost"
+    }
+  },
+  "topLevelUnknown": "will be lost"
+}`
+	os.WriteFile(configPath, []byte(initialConfig), 0o644)
+
+	// Update aliases
+	if err := saveAliases("claude", map[string]string{"primary": "model2"}); err != nil {
+		t.Fatalf("failed to save: %v", err)
+	}
+
+	// Read raw file to check
+	data, _ := os.ReadFile(configPath)
+	content := string(data)
+
+	// models should be preserved
+	if !contains(content, "model1") {
+		t.Error("models should be preserved")
+	}
+
+	// primary should be updated
+	if !contains(content, "model2") {
+		t.Error("primary should be updated to model2")
+	}
+}
+
+func contains(s, substr string) bool {
+	return len(s) >= len(substr) && (s == substr || len(s) > 0 && containsHelper(s, substr))
+}
+
+func containsHelper(s, substr string) bool {
+	for i := 0; i <= len(s)-len(substr); i++ {
+		if s[i:i+len(substr)] == substr {
+			return true
+		}
+	}
+	return false
+}
+
+func TestClaudeImplementsAliasConfigurer(t *testing.T) {
+	c := &Claude{}
+	var _ AliasConfigurer = c // Compile-time check
+}
+
+func TestModelNameEdgeCases(t *testing.T) {
+	testCases := []struct {
+		name  string
+		model string
+	}{
+		{"simple", "llama3.2"},
+		{"with tag", "llama3.2:latest"},
+		{"with cloud tag", "kimi-k2.5:cloud"},
+		{"with namespace", "library/llama3.2"},
+		{"with dots", "glm-4.7-flash"},
+		{"with numbers", "qwen3:8b"},
+	}
+
+	for _, tc := range testCases {
+		t.Run(tc.name, func(t *testing.T) {
+			tmpDir := t.TempDir()
+			setTestHome(t, tmpDir)
+
+			aliases := map[string]string{"primary": tc.model}
+			if err := saveAliases("claude", aliases); err != nil {
+				t.Fatalf("failed to save model %q: %v", tc.model, err)
+			}
+
+			loaded, err := loadIntegration("claude")
+			if err != nil {
+				t.Fatalf("failed to load: %v", err)
+			}
+			if loaded.Aliases["primary"] != tc.model {
+				t.Errorf("expected primary=%q, got %q", tc.model, loaded.Aliases["primary"])
+			}
+		})
+	}
+}
+
+func TestSwitchingScenarios(t *testing.T) {
+	t.Run("cloud to local removes fast", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+
+		// Initial cloud config
+		if err := saveAliases("claude", map[string]string{
+			"primary": "cloud-model",
+			"fast":    "cloud-model",
+		}); err != nil {
+			t.Fatal(err)
+		}
+
+		// Switch to local (no fast)
+		if err := saveAliases("claude", map[string]string{
+			"primary": "local-model",
+		}); err != nil {
+			t.Fatal(err)
+		}
+
+		loaded, _ := loadIntegration("claude")
+		if loaded.Aliases["fast"] != "" {
+			t.Errorf("fast should be removed, got %q", loaded.Aliases["fast"])
+		}
+		if loaded.Aliases["primary"] != "local-model" {
+			t.Errorf("primary should be local-model, got %q", loaded.Aliases["primary"])
+		}
+	})
+
+	t.Run("local to cloud adds fast", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+
+		// Initial local config
+		if err := saveAliases("claude", map[string]string{
+			"primary": "local-model",
+		}); err != nil {
+			t.Fatal(err)
+		}
+
+		// Switch to cloud (with fast)
+		if err := saveAliases("claude", map[string]string{
+			"primary": "cloud-model",
+			"fast":    "cloud-model",
+		}); err != nil {
+			t.Fatal(err)
+		}
+
+		loaded, _ := loadIntegration("claude")
+		if loaded.Aliases["fast"] != "cloud-model" {
+			t.Errorf("fast should be cloud-model, got %q", loaded.Aliases["fast"])
+		}
+	})
+
+	t.Run("cloud to different cloud updates both", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+
+		// Initial cloud config
+		if err := saveAliases("claude", map[string]string{
+			"primary": "cloud-model-1",
+			"fast":    "cloud-model-1",
+		}); err != nil {
+			t.Fatal(err)
+		}
+
+		// Switch to different cloud
+		if err := saveAliases("claude", map[string]string{
+			"primary": "cloud-model-2",
+			"fast":    "cloud-model-2",
+		}); err != nil {
+			t.Fatal(err)
+		}
+
+		loaded, _ := loadIntegration("claude")
+		if loaded.Aliases["primary"] != "cloud-model-2" {
+			t.Errorf("primary should be cloud-model-2, got %q", loaded.Aliases["primary"])
+		}
+		if loaded.Aliases["fast"] != "cloud-model-2" {
+			t.Errorf("fast should be cloud-model-2, got %q", loaded.Aliases["fast"])
+		}
+	})
+}
+
+func TestToolCapabilityFiltering(t *testing.T) {
+	t.Run("all models checked for tool capability", func(t *testing.T) {
+		// Both cloud and local models are checked for tool capability via Show API
+		// Only models with "tools" in capabilities are included
+		m := modelInfo{Name: "tool-model", Remote: false, ToolCapable: true}
+		if !m.ToolCapable {
+			t.Error("tool capable model should be marked as such")
+		}
+	})
+
+	t.Run("modelInfo includes ToolCapable field", func(t *testing.T) {
+		m := modelInfo{Name: "test", Remote: true, ToolCapable: true}
+		if !m.ToolCapable {
+			t.Error("ToolCapable field should be accessible")
+		}
+	})
+}
+
+func TestIsCloudModel_RequiresClient(t *testing.T) {
+	t.Run("nil client always returns false", func(t *testing.T) {
+		// isCloudModel now only uses Show API, no suffix detection
+		if isCloudModel(context.Background(), nil, "model:cloud") {
+			t.Error("nil client should return false regardless of suffix")
+		}
+		if isCloudModel(context.Background(), nil, "local-model") {
+			t.Error("nil client should return false")
+		}
+	})
+}
+
+func TestModelsAndAliasesMustStayInSync(t *testing.T) {
+	t.Run("saveAliases followed by saveIntegration keeps them in sync", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+
+		// Save aliases with one model
+		if err := saveAliases("claude", map[string]string{"primary": "model-a"}); err != nil {
+			t.Fatal(err)
+		}
+
+		// Save integration with same model (this is the pattern we use)
+		if err := saveIntegration("claude", []string{"model-a"}); err != nil {
+			t.Fatal(err)
+		}
+
+		loaded, _ := loadIntegration("claude")
+		if loaded.Aliases["primary"] != loaded.Models[0] {
+			t.Errorf("aliases.primary (%q) != models[0] (%q)", loaded.Aliases["primary"], loaded.Models[0])
+		}
+	})
+
+	t.Run("out of sync config is detectable", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+
+		// Simulate out-of-sync state (like manual edit or bug)
+		if err := saveIntegration("claude", []string{"old-model"}); err != nil {
+			t.Fatal(err)
+		}
+		if err := saveAliases("claude", map[string]string{"primary": "new-model"}); err != nil {
+			t.Fatal(err)
+		}
+
+		loaded, _ := loadIntegration("claude")
+
+		// They should be different (this is the bug state)
+		if loaded.Models[0] == loaded.Aliases["primary"] {
+			t.Error("expected out-of-sync state for this test")
+		}
+
+		// The fix: when updating aliases, also update models
+		if err := saveIntegration("claude", []string{loaded.Aliases["primary"]}); err != nil {
+			t.Fatal(err)
+		}
+
+		loaded, _ = loadIntegration("claude")
+		if loaded.Models[0] != loaded.Aliases["primary"] {
+			t.Errorf("after fix: models[0] (%q) should equal aliases.primary (%q)",
+				loaded.Models[0], loaded.Aliases["primary"])
+		}
+	})
+
+	t.Run("updating primary alias updates models too", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+
+		// Initial state
+		if err := saveIntegration("claude", []string{"initial-model"}); err != nil {
+			t.Fatal(err)
+		}
+		if err := saveAliases("claude", map[string]string{"primary": "initial-model"}); err != nil {
+			t.Fatal(err)
+		}
+
+		// Update aliases AND models together
+		newAliases := map[string]string{"primary": "updated-model"}
+		if err := saveAliases("claude", newAliases); err != nil {
+			t.Fatal(err)
+		}
+		if err := saveIntegration("claude", []string{newAliases["primary"]}); err != nil {
+			t.Fatal(err)
+		}
+
+		loaded, _ := loadIntegration("claude")
+		if loaded.Models[0] != "updated-model" {
+			t.Errorf("models[0] should be updated-model, got %q", loaded.Models[0])
+		}
+		if loaded.Aliases["primary"] != "updated-model" {
+			t.Errorf("aliases.primary should be updated-model, got %q", loaded.Aliases["primary"])
+		}
+	})
+}

cmd/config/droid.go

@@ -1,6 +1,7 @@
 package config
 
 import (
+	"context"
 	"encoding/json"
 	"fmt"
 	"os"
@@ -8,6 +9,7 @@ import (
 	"path/filepath"
 	"slices"
 
+	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/envconfig"
 )
 
@@ -112,9 +114,17 @@ func (d *Droid) Edit(models []string) error {
 	}
 
 	// Build new Ollama model entries with sequential indices (0, 1, 2, ...)
+	client, _ := api.ClientFromEnvironment()
+
 	var newModels []any
 	var defaultModelID string
 	for i, model := range models {
+		maxOutput := 64000
+		if isCloudModel(context.Background(), client, model) {
+			if l, ok := lookupCloudModelLimit(model); ok {
+				maxOutput = l.Output
+			}
+		}
 		modelID := fmt.Sprintf("custom:%s-%d", model, i)
 		newModels = append(newModels, modelEntry{
 			Model:           model,
@@ -122,7 +132,7 @@ func (d *Droid) Edit(models []string) error {
 			BaseURL:         envconfig.Host().String() + "/v1",
 			APIKey:          "ollama",
 			Provider:        "generic-chat-completion-api",
-			MaxOutputTokens: 64000,
+			MaxOutputTokens: maxOutput,
 			SupportsImages:  false,
 			ID:              modelID,
 			Index:           i,

cmd/config/droid_test.go

@@ -1251,6 +1251,55 @@ func TestDroidEdit_LargeNumberOfModels(t *testing.T) {
 	}
 }
 
+func TestDroidEdit_LocalModelDefaultMaxOutput(t *testing.T) {
+	d := &Droid{}
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	settingsDir := filepath.Join(tmpDir, ".factory")
+	settingsPath := filepath.Join(settingsDir, "settings.json")
+
+	if err := d.Edit([]string{"llama3.2"}); err != nil {
+		t.Fatal(err)
+	}
+
+	data, _ := os.ReadFile(settingsPath)
+	var settings map[string]any
+	json.Unmarshal(data, &settings)
+
+	models := settings["customModels"].([]any)
+	entry := models[0].(map[string]any)
+	if entry["maxOutputTokens"] != float64(64000) {
+		t.Errorf("local model maxOutputTokens = %v, want 64000", entry["maxOutputTokens"])
+	}
+}
+
+func TestDroidEdit_CloudModelLimitsUsed(t *testing.T) {
+	// Verify that every cloud model in cloudModelLimits has a valid output
+	// value that would be used for maxOutputTokens when isCloudModel returns true.
+	// :cloud suffix stripping must also work since that's how users specify them.
+	for name, expected := range cloudModelLimits {
+		t.Run(name, func(t *testing.T) {
+			l, ok := lookupCloudModelLimit(name)
+			if !ok {
+				t.Fatalf("lookupCloudModelLimit(%q) returned false", name)
+			}
+			if l.Output != expected.Output {
+				t.Errorf("output = %d, want %d", l.Output, expected.Output)
+			}
+			// Also verify :cloud suffix lookup
+			cloudName := name + ":cloud"
+			l2, ok := lookupCloudModelLimit(cloudName)
+			if !ok {
+				t.Fatalf("lookupCloudModelLimit(%q) returned false", cloudName)
+			}
+			if l2.Output != expected.Output {
+				t.Errorf(":cloud output = %d, want %d", l2.Output, expected.Output)
+			}
+		})
+	}
+}
+
 func TestDroidEdit_ArraysWithMixedTypes(t *testing.T) {
 	d := &Droid{}
 	tmpDir := t.TempDir()

cmd/config/integrations.go

@@ -207,7 +207,10 @@ func listModels(ctx context.Context) ([]selectItem, map[string]bool, map[string]
 
 	var existing []modelInfo
 	for _, m := range models.Models {
-		existing = append(existing, modelInfo{Name: m.Name, Remote: m.RemoteModel != ""})
+		existing = append(existing, modelInfo{
+			Name:   m.Name,
+			Remote: m.RemoteModel != "",
+		})
 	}
 
 	items, _, existingModels, cloudModels := buildModelList(existing, nil, "")
@@ -286,52 +289,38 @@ func ensureAuth(ctx context.Context, client *api.Client, cloudModels map[string]
 	}
 }
 
-func ensureAliases(ctx context.Context, r Runner, name string, primaryModel string, existing map[string]string, force bool) (bool, error) {
-	ac, ok := r.(AliasConfigurer)
-	if !ok {
-		return false, nil
-	}
-
-	aliases, updated, err := ac.ConfigureAliases(ctx, primaryModel, existing, force)
-	if err != nil {
-		return false, err
-	}
-	if !updated {
-		return false, nil
-	}
-
-	if err := saveAliases(name, aliases); err != nil {
-		return false, err
-	}
-
-	if err := ac.SetAliases(ctx, aliases); err != nil {
-		fmt.Fprintf(os.Stderr, "%sWarning: Could not sync aliases to server: %v%s\n", ansiGray, err, ansiReset)
-		fmt.Fprintf(os.Stderr, "%sAliases saved locally. Server sync will retry on next launch.%s\n\n", ansiGray, ansiReset)
-	}
-
-	return true, nil
-}
-
 func runIntegration(name, modelName string, args []string) error {
 	r, ok := integrations[name]
 	if !ok {
 		return fmt.Errorf("unknown integration: %s", name)
 	}
 
-	if _, ok := r.(AliasConfigurer); ok {
-		if config, err := loadIntegration(name); err == nil && config.Aliases != nil {
-			primary, fast := config.Aliases["primary"], config.Aliases["fast"]
-			if primary != "" && fast != "" {
-				fmt.Fprintf(os.Stderr, "\nLaunching %s with Primary: %s, Fast: %s...\n", r, primary, fast)
-				return r.Run(modelName, args)
-			}
-		}
-	}
-
 	fmt.Fprintf(os.Stderr, "\nLaunching %s with %s...\n", r, modelName)
 	return r.Run(modelName, args)
 }
 
+// syncAliases syncs aliases to server and saves locally for an AliasConfigurer.
+func syncAliases(ctx context.Context, client *api.Client, ac AliasConfigurer, name, model string, existing map[string]string) error {
+	aliases := make(map[string]string)
+	for k, v := range existing {
+		aliases[k] = v
+	}
+	aliases["primary"] = model
+
+	if isCloudModel(ctx, client, model) {
+		if aliases["fast"] == "" || !isCloudModel(ctx, client, aliases["fast"]) {
+			aliases["fast"] = model
+		}
+	} else {
+		delete(aliases, "fast")
+	}
+
+	if err := ac.SetAliases(ctx, aliases); err != nil {
+		return err
+	}
+	return saveAliases(name, aliases)
+}
+
 // LaunchCmd returns the cobra command for launching integrations.
 func LaunchCmd(checkServerHeartbeat func(cmd *cobra.Command, args []string) error) *cobra.Command {
 	var modelFlag string
@@ -399,50 +388,88 @@ Examples:
 				return fmt.Errorf("unknown integration: %s", name)
 			}
 
-			if !configFlag && modelFlag == "" {
-				if config, err := loadIntegration(name); err == nil && len(config.Models) > 0 {
-					if _, err := ensureAliases(cmd.Context(), r, name, config.Models[0], config.Aliases, false); errors.Is(err, errCancelled) {
-						return nil
-					} else if err != nil {
-						return err
-					}
-					return runIntegration(name, config.Models[0], passArgs)
-				}
-			}
-
+			// Handle AliasConfigurer integrations (claude, codex)
 			if ac, ok := r.(AliasConfigurer); ok {
-				var existingAliases map[string]string
-				if existing, err := loadIntegration(name); err == nil {
-					existingAliases = existing.Aliases
-				}
-				aliases, updated, err := ac.ConfigureAliases(cmd.Context(), "", existingAliases, configFlag)
-				if errors.Is(err, errCancelled) {
-					return nil
-				}
+				client, err := api.ClientFromEnvironment()
 				if err != nil {
 					return err
 				}
-				if updated {
-					if err := saveAliases(name, aliases); err != nil {
+
+				// Validate --model flag if provided
+				if modelFlag != "" {
+					if _, err := client.Show(cmd.Context(), &api.ShowRequest{Name: modelFlag}); err != nil {
+						return fmt.Errorf("model %q not found", modelFlag)
+					}
+				}
+
+				var model string
+				var existingAliases map[string]string
+
+				// Load saved config
+				if cfg, err := loadIntegration(name); err == nil {
+					existingAliases = cfg.Aliases
+					if len(cfg.Models) > 0 {
+						model = cfg.Models[0]
+						// AliasConfigurer integrations use single model; sanitize if multiple
+						if len(cfg.Models) > 1 {
+							_ = saveIntegration(name, []string{model})
+						}
+					}
+				}
+
+				// --model flag overrides saved model
+				if modelFlag != "" {
+					model = modelFlag
+				}
+
+				// Validate saved model still exists
+				if model != "" && modelFlag == "" {
+					if _, err := client.Show(cmd.Context(), &api.ShowRequest{Name: model}); err != nil {
+						fmt.Fprintf(os.Stderr, "%sConfigured model %q not found%s\n\n", ansiGray, model, ansiReset)
+						model = ""
+					}
+				}
+
+				// If no valid model or --config flag, show picker
+				if model == "" || configFlag {
+					aliases, _, err := ac.ConfigureAliases(cmd.Context(), model, existingAliases, configFlag)
+					if errors.Is(err, errCancelled) {
+						return nil
+					}
+					if err != nil {
 						return err
 					}
-					if err := ac.SetAliases(cmd.Context(), aliases); err != nil {
-						fmt.Fprintf(os.Stderr, "%sWarning: Could not sync aliases to server: %v%s\n", ansiGray, err, ansiReset)
-					}
-					fmt.Fprintf(os.Stderr, "\n%sConfiguration Complete%s\n", ansiBold, ansiReset)
-					fmt.Fprintf(os.Stderr, "Primary: %s\n", aliases["primary"])
-					fmt.Fprintf(os.Stderr, "Fast: %s\n\n", aliases["fast"])
+					model = aliases["primary"]
+					existingAliases = aliases
 				}
-				if err := saveIntegration(name, []string{aliases["primary"]}); err != nil {
+
+				// Sync aliases and save
+				if err := syncAliases(cmd.Context(), client, ac, name, model, existingAliases); err != nil {
+					fmt.Fprintf(os.Stderr, "%sWarning: Could not sync aliases: %v%s\n", ansiGray, err, ansiReset)
+				}
+				if err := saveIntegration(name, []string{model}); err != nil {
 					return fmt.Errorf("failed to save: %w", err)
 				}
+
+				// Launch (unless --config without confirmation)
 				if configFlag {
 					if launch, _ := confirmPrompt(fmt.Sprintf("Launch %s now?", r)); launch {
-						return runIntegration(name, aliases["primary"], passArgs)
+						return runIntegration(name, model, passArgs)
 					}
 					return nil
 				}
-				return runIntegration(name, aliases["primary"], passArgs)
+				return runIntegration(name, model, passArgs)
+			}
+
+			// Validate --model flag for non-AliasConfigurer integrations
+			if modelFlag != "" {
+				client, err := api.ClientFromEnvironment()
+				if err != nil {
+					return err
+				}
+				if _, err := client.Show(cmd.Context(), &api.ShowRequest{Name: modelFlag}); err != nil {
+					return fmt.Errorf("model %q not found", modelFlag)
+				}
 			}
 
 			var models []string
@@ -455,6 +482,8 @@ Examples:
 						}
 					}
 				}
+			} else if saved, err := loadIntegration(name); err == nil && len(saved.Models) > 0 && !configFlag {
+				return runIntegration(name, saved.Models[0], passArgs)
 			} else {
 				var err error
 				models, err = selectModels(cmd.Context(), name, "")
@@ -517,8 +546,9 @@ Examples:
 }
 
 type modelInfo struct {
-	Name   string
-	Remote bool
+	Name        string
+	Remote      bool
+	ToolCapable bool
 }
 
 // buildModelList merges existing models with recommendations, sorts them, and returns
@@ -555,7 +585,7 @@ func buildModelList(existing []modelInfo, preChecked []string, current string) (
 			continue
 		}
 		items = append(items, rec)
-		if isCloudModel(rec.Name) {
+		if strings.HasSuffix(rec.Name, ":cloud") {
 			cloudModels[rec.Name] = true
 		}
 	}
@@ -615,8 +645,16 @@ func buildModelList(existing []modelInfo, preChecked []string, current string) (
 	return items, preChecked, existingModels, cloudModels
 }
 
-func isCloudModel(name string) bool {
-	return strings.HasSuffix(name, ":cloud")
+// isCloudModel checks if a model is a cloud model using the Show API.
+func isCloudModel(ctx context.Context, client *api.Client, name string) bool {
+	if client == nil {
+		return false
+	}
+	resp, err := client.Show(ctx, &api.ShowRequest{Name: name})
+	if err != nil {
+		return false
+	}
+	return resp.RemoteModel != ""
 }
 
 func pullModel(ctx context.Context, client *api.Client, model string) error {

cmd/config/integrations_test.go

@@ -1,6 +1,7 @@
 package config
 
 import (
+	"context"
 	"fmt"
 	"slices"
 	"strings"
@@ -297,24 +298,15 @@ func TestParseArgs(t *testing.T) {
 }
 
 func TestIsCloudModel(t *testing.T) {
-	tests := []struct {
-		name string
-		want bool
-	}{
-		{"glm-4.7:cloud", true},
-		{"kimi-k2.5:cloud", true},
-		{"glm-4.7-flash", false},
-		{"glm-4.7-flash:latest", false},
-		{"cloud-model", false},
-		{"model:cloudish", false},
-	}
-	for _, tt := range tests {
-		t.Run(tt.name, func(t *testing.T) {
-			if got := isCloudModel(tt.name); got != tt.want {
-				t.Errorf("isCloudModel(%q) = %v, want %v", tt.name, got, tt.want)
+	// isCloudModel now only uses Show API, so nil client always returns false
+	t.Run("nil client returns false", func(t *testing.T) {
+		models := []string{"glm-4.7:cloud", "kimi-k2.5:cloud", "local-model"}
+		for _, model := range models {
+			if isCloudModel(context.Background(), nil, model) {
+				t.Errorf("isCloudModel(%q) with nil client should return false", model)
 			}
-		})
-	}
+		}
+	})
 }
 
 func names(items []selectItem) []string {
@@ -510,6 +502,28 @@ func TestBuildModelList_ReturnsExistingAndCloudMaps(t *testing.T) {
 	}
 }
 
+func TestEditorIntegration_SavedConfigSkipsSelection(t *testing.T) {
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	// Save a config for opencode so it looks like a previous launch
+	if err := saveIntegration("opencode", []string{"llama3.2"}); err != nil {
+		t.Fatal(err)
+	}
+
+	// Verify loadIntegration returns the saved models
+	saved, err := loadIntegration("opencode")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(saved.Models) == 0 {
+		t.Fatal("expected saved models")
+	}
+	if saved.Models[0] != "llama3.2" {
+		t.Errorf("expected llama3.2, got %s", saved.Models[0])
+	}
+}
+
 func TestAliasConfigurerInterface(t *testing.T) {
 	t.Run("claude implements AliasConfigurer", func(t *testing.T) {
 		claude := &Claude{}

cmd/config/openclaw.go

@@ -17,8 +17,6 @@ type Openclaw struct{}
 
 func (c *Openclaw) String() string { return "OpenClaw" }
 
-const ansiGreen = "\033[32m"
-
 func (c *Openclaw) Run(model string, args []string) error {
 	bin := "openclaw"
 	if _, err := exec.LookPath(bin); err != nil {

cmd/config/opencode.go

@@ -1,6 +1,7 @@
 package config
 
 import (
+	"context"
 	"encoding/json"
 	"fmt"
 	"maps"
@@ -10,12 +11,53 @@ import (
 	"slices"
 	"strings"
 
+	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/envconfig"
 )
 
 // OpenCode implements Runner and Editor for OpenCode integration
 type OpenCode struct{}
 
+// cloudModelLimit holds context and output token limits for a cloud model.
+type cloudModelLimit struct {
+	Context int
+	Output  int
+}
+
+// cloudModelLimits maps cloud model base names to their token limits.
+// TODO(parthsareen): grab context/output limits from model info instead of hardcoding
+var cloudModelLimits = map[string]cloudModelLimit{
+	"cogito-2.1:671b":     {Context: 163_840, Output: 65_536},
+	"deepseek-v3.1:671b":  {Context: 163_840, Output: 163_840},
+	"deepseek-v3.2":       {Context: 163_840, Output: 65_536},
+	"glm-4.6":             {Context: 202_752, Output: 131_072},
+	"glm-4.7":             {Context: 202_752, Output: 131_072},
+	"gpt-oss:120b":        {Context: 131_072, Output: 131_072},
+	"gpt-oss:20b":         {Context: 131_072, Output: 131_072},
+	"kimi-k2:1t":          {Context: 262_144, Output: 262_144},
+	"kimi-k2.5":           {Context: 262_144, Output: 262_144},
+	"kimi-k2-thinking":    {Context: 262_144, Output: 262_144},
+	"nemotron-3-nano:30b": {Context: 1_048_576, Output: 131_072},
+	"qwen3-coder:480b":    {Context: 262_144, Output: 65_536},
+	"qwen3-coder-next":    {Context: 262_144, Output: 32_768},
+	"qwen3-next:80b":      {Context: 262_144, Output: 32_768},
+}
+
+// lookupCloudModelLimit returns the token limits for a cloud model.
+// It tries the exact name first, then strips the ":cloud" suffix.
+func lookupCloudModelLimit(name string) (cloudModelLimit, bool) {
+	if l, ok := cloudModelLimits[name]; ok {
+		return l, true
+	}
+	base := strings.TrimSuffix(name, ":cloud")
+	if base != name {
+		if l, ok := cloudModelLimits[base]; ok {
+			return l, true
+		}
+	}
+	return cloudModelLimit{}, false
+}
+
 func (o *OpenCode) String() string { return "OpenCode" }
 
 func (o *OpenCode) Run(model string, args []string) error {
@@ -113,6 +155,8 @@ func (o *OpenCode) Edit(modelList []string) error {
 		}
 	}
 
+	client, _ := api.ClientFromEnvironment()
+
 	for _, model := range modelList {
 		if existing, ok := models[model].(map[string]any); ok {
 			// migrate existing models without _launch marker
@@ -122,12 +166,29 @@ func (o *OpenCode) Edit(modelList []string) error {
 					existing["name"] = strings.TrimSuffix(name, " [Ollama]")
 				}
 			}
+			if isCloudModel(context.Background(), client, model) {
+				if l, ok := lookupCloudModelLimit(model); ok {
+					existing["limit"] = map[string]any{
+						"context": l.Context,
+						"output":  l.Output,
+					}
+				}
+			}
 			continue
 		}
-		models[model] = map[string]any{
+		entry := map[string]any{
 			"name":    model,
 			"_launch": true,
 		}
+		if isCloudModel(context.Background(), client, model) {
+			if l, ok := lookupCloudModelLimit(model); ok {
+				entry["limit"] = map[string]any{
+					"context": l.Context,
+					"output":  l.Output,
+				}
+			}
+		}
+		models[model] = entry
 	}
 
 	ollama["models"] = models

cmd/config/opencode_test.go

@@ -2,6 +2,7 @@ package config
 
 import (
 	"encoding/json"
+	"fmt"
 	"os"
 	"path/filepath"
 	"testing"
@@ -495,6 +496,166 @@ func TestOpenCodeEdit_SpecialCharsInModelName(t *testing.T) {
 	}
 }
 
+func readOpenCodeModel(t *testing.T, configPath, model string) map[string]any {
+	t.Helper()
+	data, err := os.ReadFile(configPath)
+	if err != nil {
+		t.Fatal(err)
+	}
+	var cfg map[string]any
+	json.Unmarshal(data, &cfg)
+	provider := cfg["provider"].(map[string]any)
+	ollama := provider["ollama"].(map[string]any)
+	models := ollama["models"].(map[string]any)
+	entry, ok := models[model].(map[string]any)
+	if !ok {
+		t.Fatalf("model %s not found in config", model)
+	}
+	return entry
+}
+
+func TestOpenCodeEdit_LocalModelNoLimit(t *testing.T) {
+	o := &OpenCode{}
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	configPath := filepath.Join(tmpDir, ".config", "opencode", "opencode.json")
+
+	if err := o.Edit([]string{"llama3.2"}); err != nil {
+		t.Fatal(err)
+	}
+
+	entry := readOpenCodeModel(t, configPath, "llama3.2")
+	if entry["limit"] != nil {
+		t.Errorf("local model should not have limit set, got %v", entry["limit"])
+	}
+}
+
+func TestOpenCodeEdit_PreservesUserLimit(t *testing.T) {
+	o := &OpenCode{}
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	configDir := filepath.Join(tmpDir, ".config", "opencode")
+	configPath := filepath.Join(configDir, "opencode.json")
+
+	// Set up a model with a user-configured limit
+	os.MkdirAll(configDir, 0o755)
+	os.WriteFile(configPath, []byte(`{
+		"provider": {
+			"ollama": {
+				"models": {
+					"llama3.2": {
+						"name": "llama3.2",
+						"_launch": true,
+						"limit": {"context": 8192, "output": 4096}
+					}
+				}
+			}
+		}
+	}`), 0o644)
+
+	// Re-edit should preserve the user's limit (not delete it)
+	if err := o.Edit([]string{"llama3.2"}); err != nil {
+		t.Fatal(err)
+	}
+
+	entry := readOpenCodeModel(t, configPath, "llama3.2")
+	limit, ok := entry["limit"].(map[string]any)
+	if !ok {
+		t.Fatal("user-configured limit was removed")
+	}
+	if limit["context"] != float64(8192) {
+		t.Errorf("context limit changed: got %v, want 8192", limit["context"])
+	}
+	if limit["output"] != float64(4096) {
+		t.Errorf("output limit changed: got %v, want 4096", limit["output"])
+	}
+}
+
+func TestOpenCodeEdit_CloudModelLimitStructure(t *testing.T) {
+	// Verify that when a cloud model entry has limits set (as Edit would do),
+	// the structure matches what opencode expects and re-edit preserves them.
+	o := &OpenCode{}
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	configDir := filepath.Join(tmpDir, ".config", "opencode")
+	configPath := filepath.Join(configDir, "opencode.json")
+
+	expected := cloudModelLimits["glm-4.7"]
+
+	// Simulate a cloud model that already has the limit set by a previous Edit
+	os.MkdirAll(configDir, 0o755)
+	os.WriteFile(configPath, []byte(fmt.Sprintf(`{
+		"provider": {
+			"ollama": {
+				"models": {
+					"glm-4.7:cloud": {
+						"name": "glm-4.7:cloud",
+						"_launch": true,
+						"limit": {"context": %d, "output": %d}
+					}
+				}
+			}
+		}
+	}`, expected.Context, expected.Output)), 0o644)
+
+	// Re-edit should preserve the cloud model limit
+	if err := o.Edit([]string{"glm-4.7:cloud"}); err != nil {
+		t.Fatal(err)
+	}
+
+	entry := readOpenCodeModel(t, configPath, "glm-4.7:cloud")
+	limit, ok := entry["limit"].(map[string]any)
+	if !ok {
+		t.Fatal("cloud model limit was removed on re-edit")
+	}
+	if limit["context"] != float64(expected.Context) {
+		t.Errorf("context = %v, want %d", limit["context"], expected.Context)
+	}
+	if limit["output"] != float64(expected.Output) {
+		t.Errorf("output = %v, want %d", limit["output"], expected.Output)
+	}
+}
+
+func TestLookupCloudModelLimit(t *testing.T) {
+	tests := []struct {
+		name        string
+		wantOK      bool
+		wantContext int
+		wantOutput  int
+	}{
+		{"glm-4.7", true, 202_752, 131_072},
+		{"glm-4.7:cloud", true, 202_752, 131_072},
+		{"kimi-k2.5", true, 262_144, 262_144},
+		{"kimi-k2.5:cloud", true, 262_144, 262_144},
+		{"deepseek-v3.2", true, 163_840, 65_536},
+		{"deepseek-v3.2:cloud", true, 163_840, 65_536},
+		{"qwen3-coder:480b", true, 262_144, 65_536},
+		{"qwen3-coder-next:cloud", true, 262_144, 32_768},
+		{"llama3.2", false, 0, 0},
+		{"unknown-model:cloud", false, 0, 0},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			l, ok := lookupCloudModelLimit(tt.name)
+			if ok != tt.wantOK {
+				t.Errorf("lookupCloudModelLimit(%q) ok = %v, want %v", tt.name, ok, tt.wantOK)
+			}
+			if ok {
+				if l.Context != tt.wantContext {
+					t.Errorf("context = %d, want %d", l.Context, tt.wantContext)
+				}
+				if l.Output != tt.wantOutput {
+					t.Errorf("output = %d, want %d", l.Output, tt.wantOutput)
+				}
+			}
+		})
+	}
+}
+
 func TestOpenCodeModels_NoConfig(t *testing.T) {
 	o := &OpenCode{}
 	tmpDir := t.TempDir()

cmd/config/selector.go

@@ -17,6 +17,7 @@ const (
 	ansiBold       = "\033[1m"
 	ansiReset      = "\033[0m"
 	ansiGray       = "\033[37m"
+	ansiGreen      = "\033[32m"
 	ansiClearDown  = "\033[J"
 )
 
@@ -65,10 +66,6 @@ func (s *selectState) handleInput(event inputEvent, char byte) (done bool, resul
 		if len(filtered) > 0 && s.selected < len(filtered) {
 			return true, filtered[s.selected].Name, nil
 		}
-		// No matches but user typed something - return filter for pull prompt
-		if len(filtered) == 0 && s.filter != "" {
-			return true, s.filter, nil
-		}
 	case eventEscape:
 		return true, "", errCancelled
 	case eventBackspace:
@@ -287,11 +284,7 @@ func renderSelect(w io.Writer, prompt string, s *selectState) int {
 	lineCount := 1
 
 	if len(filtered) == 0 {
-		if s.filter != "" {
-			fmt.Fprintf(w, "  %s→ Download model: '%s'? Press Enter%s\r\n", ansiGray, s.filter, ansiReset)
-		} else {
-			fmt.Fprintf(w, "  %s(no matches)%s\r\n", ansiGray, ansiReset)
-		}
+		fmt.Fprintf(w, "  %s(no matches)%s\r\n", ansiGray, ansiReset)
 		lineCount++
 	} else {
 		displayCount := min(len(filtered), maxDisplayedItems)

cmd/config/selector_test.go

@@ -87,12 +87,12 @@ func TestSelectState(t *testing.T) {
 		}
 	})
 
-	t.Run("Enter_EmptyFilteredList_ReturnsFilter", func(t *testing.T) {
+	t.Run("Enter_EmptyFilteredList_DoesNothing", func(t *testing.T) {
 		s := newSelectState(items)
 		s.filter = "nonexistent"
 		done, result, err := s.handleInput(eventEnter, 0)
-		if !done || result != "nonexistent" || err != nil {
-			t.Errorf("expected (true, 'nonexistent', nil), got (%v, %v, %v)", done, result, err)
+		if done || result != "" || err != nil {
+			t.Errorf("expected (false, '', nil), got (%v, %v, %v)", done, result, err)
 		}
 	})
 
@@ -576,15 +576,15 @@ func TestRenderSelect(t *testing.T) {
 		}
 	})
 
-	t.Run("EmptyFilteredList_ShowsPullPrompt", func(t *testing.T) {
+	t.Run("EmptyFilteredList_ShowsNoMatches", func(t *testing.T) {
 		s := newSelectState(items)
 		s.filter = "xyz"
 		var buf bytes.Buffer
 		renderSelect(&buf, "Select:", s)
 
 		output := buf.String()
-		if !strings.Contains(output, "Download model: 'xyz'?") {
-			t.Errorf("expected 'Download model: xyz?' message, got: %s", output)
+		if !strings.Contains(output, "no matches") {
+			t.Errorf("expected 'no matches' message, got: %s", output)
 		}
 	})
 

cmd/start_darwin.go

@@ -10,19 +10,21 @@ import (
 	"github.com/ollama/ollama/api"
 )
 
+var errNotRunning = errors.New("could not connect to ollama server, run 'ollama serve' to start it")
+
 func startApp(ctx context.Context, client *api.Client) error {
 	exe, err := os.Executable()
 	if err != nil {
-		return err
+		return errNotRunning
 	}
 	link, err := os.Readlink(exe)
 	if err != nil {
-		return err
+		return errNotRunning
 	}
 	r := regexp.MustCompile(`^.*/Ollama\s?\d*.app`)
 	m := r.FindStringSubmatch(link)
 	if len(m) != 1 {
-		return errors.New("could not find ollama app")
+		return errNotRunning
 	}
 	if err := exec.Command("/usr/bin/open", "-j", "-a", m[0], "--args", "--fast-startup").Run(); err != nil {
 		return err

go.mod

@@ -24,6 +24,7 @@ require (
 	github.com/d4l3k/go-bfloat16 v0.0.0-20211005043715-690c3bdd05f1
 	github.com/dlclark/regexp2 v1.11.4
 	github.com/emirpasic/gods/v2 v2.0.0-alpha
+	github.com/klauspost/compress v1.18.3
 	github.com/mattn/go-runewidth v0.0.14
 	github.com/nlpodyssey/gopickle v0.3.0
 	github.com/pdevine/tensor v0.0.0-20240510204454-f88f4562727c

go.sum

@@ -106,7 +106,6 @@ github.com/golang/protobuf v1.5.0/go.mod h1:FsONVRAS9T7sI+LIUmWTfcYkHO4aIWwzhcaS
 github.com/golang/protobuf v1.5.2/go.mod h1:XVQd3VNwM+JqD3oG2Ue2ip4fOMUkwXdXDdiuN0vRsmY=
 github.com/golang/protobuf v1.5.4 h1:i7eJL8qZTpSEXOPTxNKhASYpMn+8e5Q6AdndVa1dWek=
 github.com/golang/protobuf v1.5.4/go.mod h1:lnTiLA8Wa4RWRcIUkrtSVa5nRhsEGBg48fD6rSs7xps=
-github.com/golang/snappy v0.0.3 h1:fHPg5GQYlCeLIPB9BZqMVR5nR9A+IM5zcgeTdjMYmLA=
 github.com/golang/snappy v0.0.3/go.mod h1:/XxbfmMg8lxefKM7IXC3fBNl/7bRcc72aCRzEWrmP2Q=
 github.com/google/flatbuffers v2.0.0+incompatible/go.mod h1:1AeVuKshWv4vARoZatz6mlQ0JxURH0Kv5+zNeJKJCa8=
 github.com/google/flatbuffers v24.3.25+incompatible h1:CX395cjN9Kke9mmalRoL3d81AtFUxJM+yDthflgJGkI=
@@ -134,8 +133,9 @@ github.com/jung-kurt/gofpdf v1.0.0/go.mod h1:7Id9E/uU8ce6rXgefFLlgrJj/GYY22cpxn+
 github.com/jung-kurt/gofpdf v1.0.3-0.20190309125859-24315acbbda5/go.mod h1:7Id9E/uU8ce6rXgefFLlgrJj/GYY22cpxn+r32jIOes=
 github.com/kisielk/errcheck v1.5.0/go.mod h1:pFxgyoBC7bSaBwPgfKdkLd5X25qrDl4LWUI2bnpBCr8=
 github.com/kisielk/gotool v1.0.0/go.mod h1:XhKaO+MFFWcvkIS/tQcRk01m1F5IRFswLeQ+oQHNcck=
-github.com/klauspost/compress v1.13.1 h1:wXr2uRxZTJXHLly6qhJabee5JqIhTRoLBhDOA74hDEQ=
 github.com/klauspost/compress v1.13.1/go.mod h1:8dP1Hq4DHOhN9w426knH3Rhby4rFm6D8eO+e+Dq5Gzg=
+github.com/klauspost/compress v1.18.3 h1:9PJRvfbmTabkOX8moIpXPbMMbYN60bWImDDU7L+/6zw=
+github.com/klauspost/compress v1.18.3/go.mod h1:R0h/fSBs8DE4ENlcrlib3PsXS61voFxhIs2DeRhCvJ4=
 github.com/klauspost/cpuid/v2 v2.0.9/go.mod h1:FInQzS24/EEf25PyTYn52gqo7WaD8xa0213Md/qVLRg=
 github.com/klauspost/cpuid/v2 v2.2.7 h1:ZWSB3igEs+d0qvnxR/ZBzXVmxkgt8DdzP6m9pfuVLDM=
 github.com/klauspost/cpuid/v2 v2.2.7/go.mod h1:Lcz8mBdAVJIBVzewtcLocK12l3Y+JytZYpaMropDUws=

integration/basic_test.go

@@ -144,3 +144,47 @@ func TestUnicodeModelDir(t *testing.T) {
 	}
 	ChatTestHelper(ctx, t, req, blueSkyExpected)
 }
+
+// TestNumPredict verifies that when num_predict is set, the model generates
+// exactly that many tokens. It uses logprobs to count the actual tokens output.
+func TestNumPredict(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute)
+	defer cancel()
+
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
+
+	if err := PullIfMissing(ctx, client, "qwen3:0.6b"); err != nil {
+		t.Fatalf("failed to pull model: %v", err)
+	}
+
+	req := api.GenerateRequest{
+		Model:    "qwen3:0.6b",
+		Prompt:   "Write a long story.",
+		Stream:   &stream,
+		Logprobs: true,
+		Options: map[string]any{
+			"num_predict": 10,
+			"temperature": 0,
+			"seed":        123,
+		},
+	}
+
+	logprobCount := 0
+	var finalResponse api.GenerateResponse
+	err := client.Generate(ctx, &req, func(resp api.GenerateResponse) error {
+		logprobCount += len(resp.Logprobs)
+		if resp.Done {
+			finalResponse = resp
+		}
+		return nil
+	})
+	if err != nil {
+		t.Fatalf("generate failed: %v", err)
+	}
+
+	if logprobCount != 10 {
+		t.Errorf("expected 10 tokens (logprobs), got %d (EvalCount=%d, DoneReason=%s)",
+			logprobCount, finalResponse.EvalCount, finalResponse.DoneReason)
+	}
+}

llm/server.go

@@ -34,6 +34,7 @@ import (
 	"github.com/ollama/ollama/logutil"
 	"github.com/ollama/ollama/ml"
 	"github.com/ollama/ollama/model"
+	"github.com/ollama/ollama/tokenizer"
 )
 
 type filteredEnv []string
@@ -116,7 +117,7 @@ type llamaServer struct {
 type ollamaServer struct {
 	llmServer
 
-	textProcessor model.TextProcessor // textProcessor handles text encoding/decoding
+	tokenizer tokenizer.Tokenizer // tokenizer handles text encoding/decoding
 }
 
 // LoadModel will load a model from disk. The model must be in the GGML format.
@@ -142,11 +143,11 @@ func LoadModel(model string, maxArraySize int) (*ggml.GGML, error) {
 // NewLlamaServer will run a server for the given GPUs
 func NewLlamaServer(systemInfo ml.SystemInfo, gpus []ml.DeviceInfo, modelPath string, f *ggml.GGML, adapters, projectors []string, opts api.Options, numParallel int) (LlamaServer, error) {
 	var llamaModel *llama.Model
-	var textProcessor model.TextProcessor
+	var tok tokenizer.Tokenizer
 	var err error
 	if envconfig.NewEngine() || f.KV().OllamaEngineRequired() {
 		if len(projectors) == 0 {
-			textProcessor, err = model.NewTextProcessor(modelPath)
+			tok, err = model.NewTextProcessor(modelPath)
 		} else {
 			err = errors.New("split vision models aren't supported")
 		}
@@ -155,7 +156,7 @@ func NewLlamaServer(systemInfo ml.SystemInfo, gpus []ml.DeviceInfo, modelPath st
 			slog.Debug("model not yet supported by Ollama engine, switching to compatibility mode", "model", modelPath, "error", err)
 		}
 	}
-	if textProcessor == nil {
+	if tok == nil {
 		llamaModel, err = llama.LoadModelFromFile(modelPath, llama.ModelParams{VocabOnly: true})
 		if err != nil {
 			return nil, err
@@ -211,7 +212,7 @@ func NewLlamaServer(systemInfo ml.SystemInfo, gpus []ml.DeviceInfo, modelPath st
 
 	kvct := strings.ToLower(envconfig.KvCacheType())
 
-	if textProcessor == nil {
+	if tok == nil {
 		flashAttention := ml.FlashAttentionAuto
 		if faUserSet {
 			if fa {
@@ -261,7 +262,7 @@ func NewLlamaServer(systemInfo ml.SystemInfo, gpus []ml.DeviceInfo, modelPath st
 	gpuLibs := ml.LibraryPaths(gpus)
 	status := NewStatusWriter(os.Stderr)
 	cmd, port, err := StartRunner(
-		textProcessor != nil,
+		tok != nil,
 		modelPath,
 		gpuLibs,
 		status,
@@ -310,8 +311,8 @@ func NewLlamaServer(systemInfo ml.SystemInfo, gpus []ml.DeviceInfo, modelPath st
 		}
 	}()
 
-	if textProcessor != nil {
-		return &ollamaServer{llmServer: s, textProcessor: textProcessor}, nil
+	if tok != nil {
+		return &ollamaServer{llmServer: s, tokenizer: tok}, nil
 	} else {
 		return &llamaServer{llmServer: s, ggml: f}, nil
 	}
@@ -1774,7 +1775,7 @@ func (s *llamaServer) Tokenize(ctx context.Context, content string) ([]int, erro
 }
 
 func (s *ollamaServer) Tokenize(ctx context.Context, content string) ([]int, error) {
-	tokens, err := s.textProcessor.Encode(content, false)
+	tokens, err := s.tokenizer.Encode(content, false)
 	if err != nil {
 		return nil, err
 	}
@@ -1809,7 +1810,7 @@ func (s *ollamaServer) Detokenize(ctx context.Context, tokens []int) (string, er
 		toks[i] = int32(t)
 	}
 
-	content, err := s.textProcessor.Decode(toks)
+	content, err := s.tokenizer.Decode(toks)
 	if err != nil {
 		return "", err
 	}

middleware/openai.go

@@ -11,6 +11,7 @@ import (
 	"time"
 
 	"github.com/gin-gonic/gin"
+	"github.com/klauspost/compress/zstd"
 
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/openai"
@@ -496,6 +497,17 @@ func (w *ResponsesWriter) Write(data []byte) (int, error) {
 
 func ResponsesMiddleware() gin.HandlerFunc {
 	return func(c *gin.Context) {
+		if c.GetHeader("Content-Encoding") == "zstd" {
+			reader, err := zstd.NewReader(c.Request.Body, zstd.WithDecoderMaxMemory(8<<20))
+			if err != nil {
+				c.AbortWithStatusJSON(http.StatusBadRequest, openai.NewError(http.StatusBadRequest, "failed to decompress zstd body"))
+				return
+			}
+			defer reader.Close()
+			c.Request.Body = io.NopCloser(reader)
+			c.Request.Header.Del("Content-Encoding")
+		}
+
 		var req openai.ResponsesRequest
 		if err := c.ShouldBindJSON(&req); err != nil {
 			c.AbortWithStatusJSON(http.StatusBadRequest, openai.NewError(http.StatusBadRequest, err.Error()))

middleware/openai_test.go

@@ -14,6 +14,7 @@ import (
 
 	"github.com/gin-gonic/gin"
 	"github.com/google/go-cmp/cmp"
+	"github.com/klauspost/compress/zstd"
 
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/openai"
@@ -1238,3 +1239,102 @@ func TestImageEditsMiddleware(t *testing.T) {
 		})
 	}
 }
+
+func zstdCompress(t *testing.T, data []byte) []byte {
+	t.Helper()
+	var buf bytes.Buffer
+	w, err := zstd.NewWriter(&buf)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, err := w.Write(data); err != nil {
+		t.Fatal(err)
+	}
+	if err := w.Close(); err != nil {
+		t.Fatal(err)
+	}
+	return buf.Bytes()
+}
+
+func TestResponsesMiddlewareZstd(t *testing.T) {
+	tests := []struct {
+		name        string
+		body        string
+		useZstd     bool
+		oversized   bool
+		wantCode    int
+		wantModel   string
+		wantMessage string
+	}{
+		{
+			name:        "plain JSON",
+			body:        `{"model": "test-model", "input": "Hello"}`,
+			wantCode:    http.StatusOK,
+			wantModel:   "test-model",
+			wantMessage: "Hello",
+		},
+		{
+			name:        "zstd compressed",
+			body:        `{"model": "test-model", "input": "Hello"}`,
+			useZstd:     true,
+			wantCode:    http.StatusOK,
+			wantModel:   "test-model",
+			wantMessage: "Hello",
+		},
+		{
+			name:      "zstd over max decompressed size",
+			oversized: true,
+			useZstd:   true,
+			wantCode:  http.StatusBadRequest,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			var capturedRequest *api.ChatRequest
+
+			gin.SetMode(gin.TestMode)
+			router := gin.New()
+			router.Use(ResponsesMiddleware(), captureRequestMiddleware(&capturedRequest))
+			router.Handle(http.MethodPost, "/v1/responses", func(c *gin.Context) {
+				c.Status(http.StatusOK)
+			})
+
+			var bodyReader io.Reader
+			if tt.oversized {
+				bodyReader = bytes.NewReader(zstdCompress(t, bytes.Repeat([]byte("A"), 9<<20)))
+			} else if tt.useZstd {
+				bodyReader = bytes.NewReader(zstdCompress(t, []byte(tt.body)))
+			} else {
+				bodyReader = strings.NewReader(tt.body)
+			}
+
+			req, _ := http.NewRequest(http.MethodPost, "/v1/responses", bodyReader)
+			req.Header.Set("Content-Type", "application/json")
+			if tt.useZstd || tt.oversized {
+				req.Header.Set("Content-Encoding", "zstd")
+			}
+
+			resp := httptest.NewRecorder()
+			router.ServeHTTP(resp, req)
+
+			if resp.Code != tt.wantCode {
+				t.Fatalf("expected status %d, got %d: %s", tt.wantCode, resp.Code, resp.Body.String())
+			}
+
+			if tt.wantCode != http.StatusOK {
+				return
+			}
+
+			if capturedRequest == nil {
+				t.Fatal("expected captured request, got nil")
+			}
+			if capturedRequest.Model != tt.wantModel {
+				t.Fatalf("expected model %q, got %q", tt.wantModel, capturedRequest.Model)
+			}
+			if len(capturedRequest.Messages) != 1 || capturedRequest.Messages[0].Content != tt.wantMessage {
+				t.Fatalf("expected single user message %q, got %+v", tt.wantMessage, capturedRequest.Messages)
+			}
+		})
+	}
+}

ml/backend.go

@@ -175,6 +175,7 @@ type Tensor interface {
 	SILU(ctx Context, up ...Tensor) Tensor
 	RELU(ctx Context, up ...Tensor) Tensor
 	Sigmoid(ctx Context) Tensor
+	SigmoidOut(ctx Context) Tensor
 
 	// AlphaLimitSILU is a variant of SILU that clamps the input to the range [-limit, limit]
 	SILUAlphaLimit(ctx Context, up Tensor, alpha, limit float32) Tensor

ml/backend/ggml/ggml.go

@@ -1468,6 +1468,13 @@ func (t *Tensor) Sigmoid(ctx ml.Context) ml.Tensor {
 	}
 }
 
+func (t *Tensor) SigmoidOut(ctx ml.Context) ml.Tensor {
+	return &Tensor{
+		b: t.b,
+		t: C.ggml_sigmoid(ctx.(*Context).ctx, t.t),
+	}
+}
+
 func (t *Tensor) View(ctx ml.Context, offset int, shape ...int) ml.Tensor {
 	switch len(shape) {
 	case 1:

model/bytepairencoding.go

@@ -1,272 +0,0 @@
-package model
-
-import (
-	"cmp"
-	"iter"
-	"slices"
-	"strings"
-
-	"github.com/dlclark/regexp2"
-	heap "github.com/emirpasic/gods/v2/trees/binaryheap"
-	"github.com/ollama/ollama/logutil"
-)
-
-type BytePairEncoding struct {
-	vocab   *Vocabulary
-	regexps []*regexp2.Regexp
-}
-
-var _ TextProcessor = (*BytePairEncoding)(nil)
-
-func NewBytePairEncoding(vocab *Vocabulary, pretokenizers ...string) BytePairEncoding {
-	if len(pretokenizers) == 0 {
-		// set default byte-level pretokenizer if none provided, e.g.
-		// https://github.com/huggingface/tokenizers/blob/main/tokenizers/src/pre_tokenizers/byte_level.rs#L44
-		pretokenizers = []string{`'s|'t|'re|'ve|'m|'ll|'d| ?\p{L}+| ?\p{N}+| ?[^\s\p{L}\p{N}]+|\s+(?!\S)|\s+`}
-	}
-
-	return BytePairEncoding{
-		vocab: vocab,
-		regexps: slices.Collect(func(yield func(*regexp2.Regexp) bool) {
-			for _, p := range pretokenizers {
-				if !yield(regexp2.MustCompile(p, regexp2.RE2)) {
-					return
-				}
-			}
-		}),
-	}
-}
-
-func (bpe BytePairEncoding) Vocabulary() *Vocabulary {
-	return bpe.vocab
-}
-
-func (bpe BytePairEncoding) Is(id int32, special Special) bool {
-	return bpe.vocab.Is(id, special)
-}
-
-func (bpe *BytePairEncoding) split(s string) iter.Seq[string] {
-	parts := []string{s}
-	for _, re := range bpe.regexps {
-		parts = slices.Collect(func(yield func(string) bool) {
-			for _, part := range parts {
-				r := []rune(part)
-				var offset int
-				for m, _ := re.FindRunesMatch(r); m != nil; m, _ = re.FindNextMatch(m) {
-					if offset-m.Index != 0 {
-						if !yield(string(r[:m.Index])) {
-							return
-						}
-					}
-
-					if !yield(m.String()) {
-						return
-					}
-
-					offset = m.Index + m.Length
-				}
-
-				if offset < len(r) {
-					if !yield(string(r[offset:])) {
-						return
-					}
-				}
-			}
-		})
-	}
-
-	return slices.Values(parts)
-}
-
-// fragment is a string fragment and their corresponding token IDs
-type fragment struct {
-	value string
-	ids   []int32
-}
-
-// pair is a pair of runes and its rank
-type pair struct {
-	a, b  int
-	rank  int
-	value string
-}
-
-type merge struct {
-	p, n  int
-	runes []rune
-}
-
-func (bpe BytePairEncoding) Encode(s string, addSpecial bool) ([]int32, error) {
-	fragments := []fragment{{value: s}}
-	for _, special := range bpe.vocab.SpecialVocabulary() {
-		// TODO: process special tokens concurrently
-		id := bpe.vocab.Encode(special)
-		for i := 0; i < len(fragments); i++ {
-			frag := fragments[i]
-			if len(frag.ids) > 0 {
-				continue
-			}
-
-			var middle []fragment
-			switch i := strings.Index(frag.value, special); {
-			case i < 0:
-				middle = append(middle, frag)
-			case i > 0:
-				middle = append(middle, fragment{value: frag.value[:i]})
-				fallthrough
-			default:
-				middle = append(middle, fragment{value: special, ids: []int32{id}})
-				if rest := frag.value[i+len(special):]; rest != "" {
-					middle = append(middle, fragment{value: rest})
-				}
-			}
-
-			fragments = append(fragments[:i], append(middle, fragments[i+1:]...)...)
-		}
-	}
-
-	var ids []int32
-	for _, frag := range fragments {
-		if len(frag.ids) > 0 {
-			ids = append(ids, frag.ids...)
-			continue
-		}
-
-		for split := range bpe.split(frag.value) {
-			// TODO: process splits concurrently
-			var sb strings.Builder
-			for _, b := range []byte(split) {
-				r := rune(b)
-				switch {
-				case r == 0x00ad:
-					r = 0x0143
-				case r <= 0x0020:
-					r = r + 0x0100
-				case r >= 0x007f && r <= 0x00a0:
-					r = r + 0x00a2
-				}
-
-				sb.WriteRune(r)
-			}
-
-			// short circuit if the fragment is in the vocabulary
-			if id := bpe.vocab.Encode(sb.String()); id >= 0 {
-				ids = append(ids, id)
-				continue
-			}
-
-			runes := []rune(sb.String())
-			merges := make([]merge, len(runes))
-			for r := range runes {
-				merges[r] = merge{
-					p:     r - 1,
-					n:     r + 1,
-					runes: []rune{runes[r]},
-				}
-			}
-
-			pairwise := func(a, b int) *pair {
-				if a < 0 || b >= len(runes) {
-					return nil
-				}
-
-				left, right := string(merges[a].runes), string(merges[b].runes)
-				rank := bpe.vocab.Merge(left, right)
-				if rank < 0 {
-					return nil
-				}
-
-				return &pair{
-					a:     a,
-					b:     b,
-					rank:  rank,
-					value: left + right,
-				}
-			}
-
-			pairs := heap.NewWith(func(i, j *pair) int {
-				return cmp.Compare(i.rank, j.rank)
-			})
-
-			for i := range len(runes) - 1 {
-				if pair := pairwise(i, i+1); pair != nil {
-					pairs.Push(pair)
-				}
-			}
-
-			for !pairs.Empty() {
-				pair, _ := pairs.Pop()
-
-				left, right := merges[pair.a], merges[pair.b]
-				if len(left.runes) == 0 || len(right.runes) == 0 ||
-					string(left.runes)+string(right.runes) != pair.value {
-					continue
-				}
-
-				if id := bpe.vocab.Encode(pair.value); id < 0 {
-					continue
-				}
-
-				merges[pair.a].runes = append(left.runes, right.runes...)
-				merges[pair.b].runes = nil
-
-				merges[pair.a].n = right.n
-				if right.n < len(merges) {
-					merges[right.n].p = pair.a
-				}
-
-				if pair := pairwise(merges[pair.a].p, pair.a); pair != nil {
-					pairs.Push(pair)
-				}
-
-				if pair := pairwise(pair.a, merges[pair.a].n); pair != nil {
-					pairs.Push(pair)
-				}
-			}
-
-			for _, merge := range merges {
-				if len(merge.runes) > 0 {
-					// TODO: handle the edge case where the rune isn't in the vocabulary
-					if id := bpe.vocab.Encode(string(merge.runes)); id >= 0 {
-						ids = append(ids, id)
-					}
-				}
-			}
-		}
-	}
-
-	if addSpecial {
-		ids = bpe.vocab.addSpecials(ids)
-	}
-
-	logutil.Trace("encoded", "string", s, "ids", ids)
-	return ids, nil
-}
-
-func (bpe BytePairEncoding) Decode(ids []int32) (string, error) {
-	var sb strings.Builder
-	for _, id := range ids {
-		for _, r := range bpe.vocab.Decode(id) {
-			switch {
-			case r == 0x0100:
-				// this produces 0x00 aka NULL
-				continue
-			case r == 0x0143:
-				r = 0x00ad
-			case r > 0x0100 && r <= 0x0120:
-				r = r - 0x0100
-			case r > 0x0120 && r <= 0x0142:
-				r = r - 0x00a2
-			}
-
-			// NOTE: not using WriteRune here because it writes the UTF-8
-			// encoding of the rune which is _not_ what we want
-			if err := sb.WriteByte(byte(r)); err != nil {
-				return "", err
-			}
-		}
-	}
-
-	logutil.Trace("decoded", "string", sb.String(), "from", ids)
-	return sb.String(), nil
-}

model/model.go

@@ -23,6 +23,7 @@ import (
 	_ "github.com/ollama/ollama/ml/backend"
 	"github.com/ollama/ollama/ml/nn/pooling"
 	"github.com/ollama/ollama/model/input"
+	"github.com/ollama/ollama/tokenizer"
 )
 
 var (
@@ -133,7 +134,7 @@ func New(modelPath string, params ml.BackendParams) (Model, error) {
 	return m, nil
 }
 
-func NewTextProcessor(s string) (TextProcessor, error) {
+func NewTextProcessor(s string) (tokenizer.Tokenizer, error) {
 	r, err := os.Open(s)
 	if err != nil {
 		return nil, err
@@ -150,7 +151,7 @@ func NewTextProcessor(s string) (TextProcessor, error) {
 		return nil, err
 	}
 
-	tp, ok := m.(TextProcessor)
+	tp, ok := m.(tokenizer.Tokenizer)
 	if !ok {
 		return nil, ErrUnsupportedTokenizer
 	}

model/models/bert/embed.go

@@ -10,11 +10,12 @@ import (
 	"github.com/ollama/ollama/ml/nn/pooling"
 	"github.com/ollama/ollama/model"
 	"github.com/ollama/ollama/model/input"
+	"github.com/ollama/ollama/tokenizer"
 )
 
 type Model struct {
 	model.Base
-	model.TextProcessor
+	tokenizer.Tokenizer
 
 	TokenEmbedding     *nn.Embedding `gguf:"token_embd"`
 	TypeEmbedding      *nn.Embedding `gguf:"token_types"`
@@ -129,7 +130,7 @@ func (o Options) headDim() int {
 }
 
 func New(c fs.Config) (model.Model, error) {
-	vocab := &model.Vocabulary{
+	vocab := &tokenizer.Vocabulary{
 		Values: c.Strings("tokenizer.ggml.tokens"),
 		Scores: c.Floats("tokenizer.ggml.scores"),
 		Types:  c.Ints("tokenizer.ggml.token_type"),
@@ -153,17 +154,17 @@ func New(c fs.Config) (model.Model, error) {
 		},
 	}
 
-	var processor model.TextProcessor
+	var t tokenizer.Tokenizer
 	switch c.String("tokenizer.ggml.model", "bert") {
 	case "bert":
-		processor = model.NewWordPiece(vocab, true)
+		t = tokenizer.NewWordPiece(vocab, true)
 	default:
 		return nil, model.ErrUnsupportedTokenizer
 	}
 
 	return &Model{
-		TextProcessor: processor,
-		Layers:        make([]EncoderLayer, c.Uint("block_count")),
+		Tokenizer: t,
+		Layers:    make([]EncoderLayer, c.Uint("block_count")),
 		Options: Options{
 			hiddenSize:  int(c.Uint("embedding_length")),
 			numHeads:    int(c.Uint("attention.head_count")),

model/models/deepseek2/model.go

@@ -13,6 +13,7 @@ import (
 	"github.com/ollama/ollama/ml/nn/rope"
 	"github.com/ollama/ollama/model"
 	"github.com/ollama/ollama/model/input"
+	"github.com/ollama/ollama/tokenizer"
 )
 
 type Options struct {
@@ -222,7 +223,7 @@ func (t *Layer) Forward(ctx ml.Context, hiddenStates, positions, outputs ml.Tens
 
 type Model struct {
 	model.Base
-	model.BytePairEncoding
+	tokenizer.Tokenizer
 
 	TokenEmbedding *nn.Embedding `gguf:"token_embd"`
 	Layers         []Layer       `gguf:"blk"`
@@ -277,8 +278,8 @@ func New(c fs.Config) (model.Model, error) {
 	}
 
 	m := Model{
-		BytePairEncoding: model.NewBytePairEncoding(
-			&model.Vocabulary{
+		Tokenizer: tokenizer.NewBytePairEncoding(
+			&tokenizer.Vocabulary{
 				Values: c.Strings("tokenizer.ggml.tokens"),
 				Types:  c.Ints("tokenizer.ggml.token_type"),
 				Merges: c.Strings("tokenizer.ggml.merges"),

model/models/deepseekocr/model.go

@@ -10,11 +10,12 @@ import (
 	"github.com/ollama/ollama/ml/nn"
 	"github.com/ollama/ollama/model"
 	"github.com/ollama/ollama/model/input"
+	"github.com/ollama/ollama/tokenizer"
 )
 
 type Model struct {
 	model.Base
-	model.TextProcessor
+	tokenizer.Tokenizer
 
 	Sam    *samModel    `gguf:"s"`
 	Vision *visionModel `gguf:"v"`
@@ -134,8 +135,8 @@ func init() {
 		}
 
 		m := Model{
-			TextProcessor: model.NewBytePairEncoding(
-				&model.Vocabulary{
+			Tokenizer: tokenizer.NewBytePairEncoding(
+				&tokenizer.Vocabulary{
 					Values: c.Strings("tokenizer.ggml.tokens"),
 					Types:  c.Ints("tokenizer.ggml.token_type"),
 					Merges: c.Strings("tokenizer.ggml.merges"),

model/models/gemma2/model.go

@@ -10,6 +10,7 @@ import (
 	"github.com/ollama/ollama/ml/nn/rope"
 	"github.com/ollama/ollama/model"
 	"github.com/ollama/ollama/model/input"
+	"github.com/ollama/ollama/tokenizer"
 )
 
 type Options struct {
@@ -27,7 +28,7 @@ func (o Options) applyRotaryPositionEmbeddings(ctx ml.Context, states, positions
 
 type Model struct {
 	model.Base
-	model.SentencePiece
+	tokenizer.Tokenizer
 
 	TokenEmbedding *nn.Embedding `gguf:"token_embd"`
 	Layers         []Layer       `gguf:"blk"`
@@ -43,8 +44,8 @@ const (
 
 func New(c fs.Config) (model.Model, error) {
 	m := Model{
-		SentencePiece: model.NewSentencePiece(
-			&model.Vocabulary{
+		Tokenizer: tokenizer.NewSentencePiece(
+			&tokenizer.Vocabulary{
 				Values: c.Strings("tokenizer.ggml.tokens"),
 				Scores: c.Floats("tokenizer.ggml.scores"),
 				Types:  c.Ints("tokenizer.ggml.token_type"),

model/models/gemma3/embed.go

@@ -7,11 +7,12 @@ import (
 	"github.com/ollama/ollama/ml/nn/pooling"
 	"github.com/ollama/ollama/model"
 	"github.com/ollama/ollama/model/input"
+	"github.com/ollama/ollama/tokenizer"
 )
 
 type embedModel struct {
 	model.Base
-	model.SentencePiece
+	tokenizer.Tokenizer
 
 	*TextModel
 	poolingType pooling.Type
@@ -31,8 +32,8 @@ func (m *embedModel) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, erro
 
 func newEmbedModel(c fs.Config) (model.Model, error) {
 	m := &embedModel{
-		SentencePiece: model.NewSentencePiece(
-			&model.Vocabulary{
+		Tokenizer: tokenizer.NewSentencePiece(
+			&tokenizer.Vocabulary{
 				Values: c.Strings("tokenizer.ggml.tokens"),
 				Scores: c.Floats("tokenizer.ggml.scores"),
 				Types:  c.Ints("tokenizer.ggml.token_type"),

model/models/gemma3/model.go

@@ -12,11 +12,12 @@ import (
 	"github.com/ollama/ollama/ml/nn"
 	"github.com/ollama/ollama/model"
 	"github.com/ollama/ollama/model/input"
+	"github.com/ollama/ollama/tokenizer"
 )
 
 type Model struct {
 	model.Base
-	model.TextProcessor
+	tokenizer.Tokenizer
 
 	*VisionModel `gguf:"v"`
 	*TextModel
@@ -54,7 +55,7 @@ func (p *MultiModalProjector) Forward(ctx ml.Context, visionOutputs ml.Tensor, i
 }
 
 func New(c fs.Config) (model.Model, error) {
-	vocabulary := model.Vocabulary{
+	vocabulary := tokenizer.Vocabulary{
 		Values: c.Strings("tokenizer.ggml.tokens"),
 		Scores: c.Floats("tokenizer.ggml.scores"),
 		Types:  c.Ints("tokenizer.ggml.token_type"),
@@ -70,19 +71,19 @@ func New(c fs.Config) (model.Model, error) {
 		),
 	}
 
-	var processor model.TextProcessor
+	var t tokenizer.Tokenizer
 	switch c.String("tokenizer.ggml.model") {
 	case "gpt2":
-		processor = model.NewBytePairEncoding(&vocabulary)
+		t = tokenizer.NewBytePairEncoding(&vocabulary)
 	default:
 		// Previous uploads of Gemma 3 on Ollama did not have token 106
 		// (i.e. "<end_of_turn>") so we need to add in case it's not already present
 		vocabulary.EOS = append(vocabulary.EOS, int32(c.Uint("tokenizer.ggml.eot_token_id", 106)))
-		processor = model.NewSentencePiece(&vocabulary)
+		t = tokenizer.NewSentencePiece(&vocabulary)
 	}
 
 	m := Model{
-		TextProcessor:  processor,
+		Tokenizer:      t,
 		ImageProcessor: newImageProcessor(c),
 		VisionModel:    newVisionModel(c),
 		TextModel:      newTextModel(c),

model/models/gemma3n/model.go

@@ -6,11 +6,12 @@ import (
 	"github.com/ollama/ollama/ml"
 	"github.com/ollama/ollama/model"
 	"github.com/ollama/ollama/model/input"
+	"github.com/ollama/ollama/tokenizer"
 )
 
 type Model struct {
 	model.Base
-	model.SentencePiece
+	tokenizer.Tokenizer
 
 	*TextModel
 }
@@ -23,8 +24,8 @@ func (m *Model) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
 func New(c fs.Config) (model.Model, error) {
 	m := Model{
 		TextModel: newTextModel(c),
-		SentencePiece: model.NewSentencePiece(
-			&model.Vocabulary{
+		Tokenizer: tokenizer.NewSentencePiece(
+			&tokenizer.Vocabulary{
 				Values: c.Strings("tokenizer.ggml.tokens"),
 				Scores: c.Floats("tokenizer.ggml.scores"),
 				Types:  c.Ints("tokenizer.ggml.token_type"),

model/models/glm4moelite/model.go

@@ -10,6 +10,7 @@ import (
 	"github.com/ollama/ollama/ml/nn"
 	"github.com/ollama/ollama/model"
 	"github.com/ollama/ollama/model/input"
+	"github.com/ollama/ollama/tokenizer"
 )
 
 var ErrOldModelFormat = errors.New("this model uses a weight format that is no longer supported; please re-download it")
@@ -198,7 +199,7 @@ func (t *Layer) Forward(ctx ml.Context, hiddenStates, positions, outputs ml.Tens
 
 type Model struct {
 	model.Base
-	model.BytePairEncoding
+	tokenizer.Tokenizer
 
 	TokenEmbedding *nn.Embedding `gguf:"token_embd"`
 	Layers         []Layer       `gguf:"blk"`
@@ -236,8 +237,8 @@ func New(c fs.Config) (model.Model, error) {
 	}
 
 	m := Model{
-		BytePairEncoding: model.NewBytePairEncoding(
-			&model.Vocabulary{
+		Tokenizer: tokenizer.NewBytePairEncoding(
+			&tokenizer.Vocabulary{
 				Values: c.Strings("tokenizer.ggml.tokens"),
 				Types:  c.Ints("tokenizer.ggml.token_type"),
 				Merges: c.Strings("tokenizer.ggml.merges"),

model/models/glmocr/model.go

@@ -11,11 +11,12 @@ import (
 	"github.com/ollama/ollama/ml"
 	"github.com/ollama/ollama/model"
 	"github.com/ollama/ollama/model/input"
+	"github.com/ollama/ollama/tokenizer"
 )
 
 type Model struct {
 	model.Base
-	model.BytePairEncoding
+	tokenizer.Tokenizer
 
 	*TextModel
 	*VisionModel     `gguf:"v"`
@@ -37,8 +38,8 @@ func New(c fs.Config) (model.Model, error) {
 	allEOS := append([]int32{eosTokenID}, eosTokenIDs...)
 
 	m := &Model{
-		BytePairEncoding: model.NewBytePairEncoding(
-			&model.Vocabulary{
+		Tokenizer: tokenizer.NewBytePairEncoding(
+			&tokenizer.Vocabulary{
 				Values: c.Strings("tokenizer.ggml.tokens"),
 				Types:  c.Ints("tokenizer.ggml.token_type"),
 				Merges: c.Strings("tokenizer.ggml.merges"),

model/models/gptoss/model.go

@@ -12,11 +12,12 @@ import (
 	"github.com/ollama/ollama/ml/nn/rope"
 	"github.com/ollama/ollama/model"
 	"github.com/ollama/ollama/model/input"
+	"github.com/ollama/ollama/tokenizer"
 )
 
 type Transformer struct {
 	model.Base
-	model.BytePairEncoding
+	tokenizer.Tokenizer
 
 	TokenEmbedding    *nn.Embedding      `gguf:"token_embd"`
 	TransformerBlocks []TransformerBlock `gguf:"blk"`
@@ -196,8 +197,8 @@ func (mlp *MLPBlock) Forward(ctx ml.Context, hiddenStates ml.Tensor, opts *Optio
 func New(c fs.Config) (model.Model, error) {
 	m := Transformer{
 		TransformerBlocks: make([]TransformerBlock, c.Uint("block_count")),
-		BytePairEncoding: model.NewBytePairEncoding(
-			&model.Vocabulary{
+		Tokenizer: tokenizer.NewBytePairEncoding(
+			&tokenizer.Vocabulary{
 				Values: c.Strings("tokenizer.ggml.tokens"),
 				Types:  c.Ints("tokenizer.ggml.token_type"),
 				Merges: c.Strings("tokenizer.ggml.merges"),

model/models/lfm2/model.go

@@ -10,6 +10,7 @@ import (
 	"github.com/ollama/ollama/ml/nn/rope"
 	"github.com/ollama/ollama/model"
 	"github.com/ollama/ollama/model/input"
+	"github.com/ollama/ollama/tokenizer"
 )
 
 type Options struct {
@@ -59,7 +60,7 @@ func (o Options) applyRotaryPositionEmbeddings(ctx ml.Context, states, positions
 
 type Model struct {
 	model.Base
-	model.TextProcessor
+	tokenizer.Tokenizer
 
 	TokenEmbedding *nn.Embedding `gguf:"token_embd"`
 	Layers         []Layer       `gguf:"blk"`
@@ -78,7 +79,7 @@ func New(c fs.Config) (model.Model, error) {
 		return nil, model.ErrUnsupportedTokenizer
 	}
 
-	vocabulary := model.Vocabulary{
+	vocabulary := tokenizer.Vocabulary{
 		Values: c.Strings("tokenizer.ggml.tokens"),
 		Scores: c.Floats("tokenizer.ggml.scores"),
 		Types:  c.Ints("tokenizer.ggml.token_type"),
@@ -104,8 +105,8 @@ func New(c fs.Config) (model.Model, error) {
 	}
 
 	m := Model{
-		TextProcessor: model.NewBytePairEncoding(&vocabulary, pretokenizers...),
-		Layers:        make([]Layer, c.Uint("block_count")),
+		Tokenizer: tokenizer.NewBytePairEncoding(&vocabulary, pretokenizers...),
+		Layers:    make([]Layer, c.Uint("block_count")),
 		Options: Options{
 			hiddenSize:            int(c.Uint("embedding_length")),
 			headDim:               int(c.Uint("attention.key_length")),

model/models/llama/model.go

@@ -11,6 +11,7 @@ import (
 	"github.com/ollama/ollama/ml/nn/rope"
 	"github.com/ollama/ollama/model"
 	"github.com/ollama/ollama/model/input"
+	"github.com/ollama/ollama/tokenizer"
 )
 
 type Options struct {
@@ -25,7 +26,7 @@ func (o Options) applyRotaryPositionEmbeddings(ctx ml.Context, states, positions
 
 type Model struct {
 	model.Base
-	model.TextProcessor
+	tokenizer.Tokenizer
 
 	TokenEmbedding *nn.Embedding `gguf:"token_embd"`
 	Layers         []Layer       `gguf:"blk"`
@@ -41,8 +42,8 @@ func New(c fs.Config) (model.Model, error) {
 		return nil, model.ErrUnsupportedModel
 	}
 
-	var processor model.TextProcessor
-	vocabulary := model.Vocabulary{
+	var processor tokenizer.Tokenizer
+	vocabulary := tokenizer.Vocabulary{
 		Values: c.Strings("tokenizer.ggml.tokens"),
 		Scores: c.Floats("tokenizer.ggml.scores"),
 		Types:  c.Ints("tokenizer.ggml.token_type"),
@@ -80,16 +81,16 @@ func New(c fs.Config) (model.Model, error) {
 				"(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
 			}
 		}
-		processor = model.NewBytePairEncoding(&vocabulary, pretokenizers...)
+		processor = tokenizer.NewBytePairEncoding(&vocabulary, pretokenizers...)
 	case "llama":
-		processor = model.NewSentencePiece(&vocabulary)
+		processor = tokenizer.NewSentencePiece(&vocabulary)
 	default:
 		return nil, model.ErrUnsupportedTokenizer
 	}
 
 	m := Model{
-		TextProcessor: processor,
-		Layers:        make([]Layer, c.Uint("block_count")),
+		Tokenizer: processor,
+		Layers:    make([]Layer, c.Uint("block_count")),
 		Options: Options{
 			hiddenSize: int(c.Uint("embedding_length")),
 			numHeads:   int(c.Uint("attention.head_count")),

model/models/llama4/model.go

@@ -11,11 +11,12 @@ import (
 	"github.com/ollama/ollama/ml/nn"
 	"github.com/ollama/ollama/model"
 	"github.com/ollama/ollama/model/input"
+	"github.com/ollama/ollama/tokenizer"
 )
 
 type Model struct {
 	model.Base
-	model.BytePairEncoding
+	tokenizer.Tokenizer
 	ImageProcessor
 
 	*VisionModel `gguf:"v"`
@@ -33,8 +34,8 @@ func (p *Projector) Forward(ctx ml.Context, visionOutputs ml.Tensor) ml.Tensor {
 
 func New(c fs.Config) (model.Model, error) {
 	m := Model{
-		BytePairEncoding: model.NewBytePairEncoding(
-			&model.Vocabulary{
+		Tokenizer: tokenizer.NewBytePairEncoding(
+			&tokenizer.Vocabulary{
 				Values: c.Strings("tokenizer.ggml.tokens"),
 				Types:  c.Ints("tokenizer.ggml.token_type"),
 				Merges: c.Strings("tokenizer.ggml.merges"),

model/models/mistral3/model.go

@@ -11,11 +11,12 @@ import (
 	"github.com/ollama/ollama/ml/nn"
 	"github.com/ollama/ollama/model"
 	"github.com/ollama/ollama/model/input"
+	"github.com/ollama/ollama/tokenizer"
 )
 
 type Model struct {
 	model.Base
-	model.BytePairEncoding
+	tokenizer.Tokenizer
 
 	*TextModel
 	*VisionModel         `gguf:"v"`
@@ -28,12 +29,12 @@ type Model struct {
 var _ model.MultimodalProcessor = (*Model)(nil)
 
 // Implement TextProcessor interface
-var _ model.TextProcessor = (*Model)(nil)
+var _ tokenizer.Tokenizer = (*Model)(nil)
 
 func New(c fs.Config) (model.Model, error) {
 	m := &Model{
-		BytePairEncoding: model.NewBytePairEncoding(
-			&model.Vocabulary{
+		Tokenizer: tokenizer.NewBytePairEncoding(
+			&tokenizer.Vocabulary{
 				Values: c.Strings("tokenizer.ggml.tokens"),
 				Types:  c.Ints("tokenizer.ggml.token_type"),
 				Merges: c.Strings("tokenizer.ggml.merges"),

model/models/mllama/model.go

@@ -11,11 +11,12 @@ import (
 	"github.com/ollama/ollama/ml/nn"
 	"github.com/ollama/ollama/model"
 	"github.com/ollama/ollama/model/input"
+	"github.com/ollama/ollama/tokenizer"
 )
 
 type Model struct {
 	model.Base
-	model.BytePairEncoding
+	tokenizer.Tokenizer
 
 	*VisionModel `gguf:"v"`
 	*TextModel
@@ -32,8 +33,8 @@ const (
 
 func New(c fs.Config) (model.Model, error) {
 	m := Model{
-		BytePairEncoding: model.NewBytePairEncoding(
-			&model.Vocabulary{
+		Tokenizer: tokenizer.NewBytePairEncoding(
+			&tokenizer.Vocabulary{
 				Values: c.Strings("tokenizer.ggml.tokens"),
 				Types:  c.Ints("tokenizer.ggml.token_type"),
 				Merges: c.Strings("tokenizer.ggml.merges"),

model/models/nomicbert/model.go

@@ -11,11 +11,12 @@ import (
 	"github.com/ollama/ollama/ml/nn/rope"
 	"github.com/ollama/ollama/model"
 	"github.com/ollama/ollama/model/input"
+	"github.com/ollama/ollama/tokenizer"
 )
 
 type Model struct {
 	model.Base
-	model.TextProcessor
+	tokenizer.Tokenizer
 
 	TokenEmbedding     *nn.Embedding `gguf:"token_embd"`
 	TypeEmbedding      *nn.Embedding `gguf:"token_types"`
@@ -178,29 +179,6 @@ func New(c fs.Config) (model.Model, error) {
 	numHeads := int(c.Uint("attention.head_count"))
 	headDim := hiddenSize / numHeads
 
-	processor := model.NewWordPiece(
-		&model.Vocabulary{
-			Values: c.Strings("tokenizer.ggml.tokens"),
-			Scores: c.Floats("tokenizer.ggml.scores"),
-			Types:  c.Ints("tokenizer.ggml.token_type"),
-			AddBOS: c.Bool("tokenizer.ggml.add_bos_token", true),
-			BOS: []int32{
-				int32(cmp.Or(
-					c.Uint("tokenizer.ggml.cls_token_id"),
-					c.Uint("tokenizer.ggml.bos_token_id"),
-				)),
-			},
-			AddEOS: c.Bool("tokenizer.ggml.add_eos_token", true),
-			EOS: []int32{
-				int32(cmp.Or(
-					c.Uint("tokenizer.ggml.separator_token_id"),
-					c.Uint("tokenizer.ggml.eos_token_id"),
-				)),
-			},
-		},
-		false,
-	)
-
 	blockCount := int(c.Uint("block_count"))
 	moeEveryNLayers := int(c.Uint("moe_every_n_layers", 0))
 	layers := make([]EncoderLayer, blockCount)
@@ -219,8 +197,29 @@ func New(c fs.Config) (model.Model, error) {
 	}
 
 	return &Model{
-		TextProcessor: processor,
-		Layers:        layers,
+		Tokenizer: tokenizer.NewWordPiece(
+			&tokenizer.Vocabulary{
+				Values: c.Strings("tokenizer.ggml.tokens"),
+				Scores: c.Floats("tokenizer.ggml.scores"),
+				Types:  c.Ints("tokenizer.ggml.token_type"),
+				AddBOS: c.Bool("tokenizer.ggml.add_bos_token", true),
+				BOS: []int32{
+					int32(cmp.Or(
+						c.Uint("tokenizer.ggml.cls_token_id"),
+						c.Uint("tokenizer.ggml.bos_token_id"),
+					)),
+				},
+				AddEOS: c.Bool("tokenizer.ggml.add_eos_token", true),
+				EOS: []int32{
+					int32(cmp.Or(
+						c.Uint("tokenizer.ggml.separator_token_id"),
+						c.Uint("tokenizer.ggml.eos_token_id"),
+					)),
+				},
+			},
+			false,
+		),
+		Layers: layers,
 		Options: Options{
 			hiddenSize:      hiddenSize,
 			numHeads:        numHeads,

model/models/olmo3/model.go

@@ -11,6 +11,7 @@ import (
 	"github.com/ollama/ollama/ml/nn/rope"
 	"github.com/ollama/ollama/model"
 	"github.com/ollama/ollama/model/input"
+	"github.com/ollama/ollama/tokenizer"
 )
 
 const (
@@ -33,7 +34,7 @@ type Options struct {
 
 type Model struct {
 	model.Base
-	model.TextProcessor
+	tokenizer.Tokenizer
 
 	TokenEmbedding *nn.Embedding `gguf:"token_embd"`
 	Layers         []Layer       `gguf:"blk"`
@@ -44,28 +45,24 @@ type Model struct {
 }
 
 func New(c fs.Config) (model.Model, error) {
-	vocabulary := model.Vocabulary{
-		Values: c.Strings("tokenizer.ggml.tokens"),
-		Scores: c.Floats("tokenizer.ggml.scores"),
-		Types:  c.Ints("tokenizer.ggml.token_type"),
-		Merges: c.Strings("tokenizer.ggml.merges"),
-		AddBOS: c.Bool("tokenizer.ggml.add_bos_token", false),
-		BOS:    []int32{int32(c.Uint("tokenizer.ggml.bos_token_id"))},
-		AddEOS: c.Bool("tokenizer.ggml.add_eos_token", false),
-		EOS: append(
-			[]int32{int32(c.Uint("tokenizer.ggml.eos_token_id"))},
-			c.Ints("tokenizer.ggml.eos_token_ids")...,
-		),
-	}
-
-	processor := model.NewBytePairEncoding(
-		&vocabulary,
-		"(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
-	)
-
 	m := Model{
-		TextProcessor: processor,
-		Layers:        make([]Layer, c.Uint("block_count")),
+		Tokenizer: tokenizer.NewBytePairEncoding(
+			&tokenizer.Vocabulary{
+				Values: c.Strings("tokenizer.ggml.tokens"),
+				Scores: c.Floats("tokenizer.ggml.scores"),
+				Types:  c.Ints("tokenizer.ggml.token_type"),
+				Merges: c.Strings("tokenizer.ggml.merges"),
+				AddBOS: c.Bool("tokenizer.ggml.add_bos_token", false),
+				BOS:    []int32{int32(c.Uint("tokenizer.ggml.bos_token_id"))},
+				AddEOS: c.Bool("tokenizer.ggml.add_eos_token", false),
+				EOS: append(
+					[]int32{int32(c.Uint("tokenizer.ggml.eos_token_id"))},
+					c.Ints("tokenizer.ggml.eos_token_ids")...,
+				),
+			},
+			"(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
+		),
+		Layers: make([]Layer, c.Uint("block_count")),
 		Options: Options{
 			hiddenSize:            int(c.Uint("embedding_length")),
 			numHeads:              int(c.Uint("attention.head_count")),

model/models/qwen2/model.go

@@ -13,6 +13,7 @@ import (
 	"github.com/ollama/ollama/ml/nn/rope"
 	"github.com/ollama/ollama/model"
 	"github.com/ollama/ollama/model/input"
+	"github.com/ollama/ollama/tokenizer"
 )
 
 type Options struct {
@@ -92,7 +93,7 @@ func (d DecoderLayer) Forward(ctx ml.Context, hiddenStates, positions, outputs m
 
 type Model struct {
 	model.Base
-	model.BytePairEncoding
+	tokenizer.Tokenizer
 
 	TokenEmbedding *nn.Embedding  `gguf:"token_embd"`
 	Layers         []DecoderLayer `gguf:"blk"`
@@ -139,8 +140,8 @@ func New(c fs.Config) (model.Model, error) {
 	}
 	m := Model{
 		Layers: make([]DecoderLayer, c.Uint("block_count")),
-		BytePairEncoding: model.NewBytePairEncoding(
-			&model.Vocabulary{
+		Tokenizer: tokenizer.NewBytePairEncoding(
+			&tokenizer.Vocabulary{
 				Values: c.Strings("tokenizer.ggml.tokens"),
 				Types:  c.Ints("tokenizer.ggml.token_type"),
 				Merges: c.Strings("tokenizer.ggml.merges"),

model/models/qwen25vl/model.go

@@ -10,11 +10,12 @@ import (
 	"github.com/ollama/ollama/ml"
 	"github.com/ollama/ollama/model"
 	"github.com/ollama/ollama/model/input"
+	"github.com/ollama/ollama/tokenizer"
 )
 
 type Model struct {
 	model.Base
-	model.BytePairEncoding
+	tokenizer.Tokenizer
 
 	*TextModel
 	*VisionModel `gguf:"v"`
@@ -27,8 +28,8 @@ var _ model.MultimodalProcessor = (*Model)(nil)
 
 func New(c fs.Config) (model.Model, error) {
 	m := &Model{
-		BytePairEncoding: model.NewBytePairEncoding(
-			&model.Vocabulary{
+		Tokenizer: tokenizer.NewBytePairEncoding(
+			&tokenizer.Vocabulary{
 				Values: c.Strings("tokenizer.ggml.tokens"),
 				Types:  c.Ints("tokenizer.ggml.token_type"),
 				Merges: c.Strings("tokenizer.ggml.merges"),

model/models/qwen3/embed.go

@@ -7,11 +7,12 @@ import (
 	"github.com/ollama/ollama/ml/nn/pooling"
 	"github.com/ollama/ollama/model"
 	"github.com/ollama/ollama/model/input"
+	"github.com/ollama/ollama/tokenizer"
 )
 
 type embedModel struct {
 	model.Base
-	model.BytePairEncoding
+	tokenizer.Tokenizer
 
 	*Model
 	poolingType pooling.Type
@@ -34,8 +35,8 @@ func newEmbed(c fs.Config) (model.Model, error) {
 		layers[i].MLP = &dense{}
 	}
 	m := embedModel{
-		BytePairEncoding: model.NewBytePairEncoding(
-			&model.Vocabulary{
+		Tokenizer: tokenizer.NewBytePairEncoding(
+			&tokenizer.Vocabulary{
 				Values: c.Strings("tokenizer.ggml.tokens"),
 				Types:  c.Ints("tokenizer.ggml.token_type"),
 				Merges: c.Strings("tokenizer.ggml.merges"),

model/models/qwen3/model.go

@@ -12,6 +12,7 @@ import (
 	"github.com/ollama/ollama/ml/nn/rope"
 	"github.com/ollama/ollama/model"
 	"github.com/ollama/ollama/model/input"
+	"github.com/ollama/ollama/tokenizer"
 )
 
 type Options struct {
@@ -159,7 +160,7 @@ func (d *Layer) Forward(ctx ml.Context, hiddenStates, positions, outputs ml.Tens
 
 type Model struct {
 	model.Base
-	model.BytePairEncoding
+	tokenizer.Tokenizer
 
 	TokenEmbedding *nn.Embedding `gguf:"token_embd"`
 	OutputNorm     *nn.RMSNorm   `gguf:"output_norm"`
@@ -218,8 +219,8 @@ func New(c fs.Config) (model.Model, error) {
 	}
 
 	m := Model{
-		BytePairEncoding: model.NewBytePairEncoding(
-			&model.Vocabulary{
+		Tokenizer: tokenizer.NewBytePairEncoding(
+			&tokenizer.Vocabulary{
 				Values: c.Strings("tokenizer.ggml.tokens"),
 				Types:  c.Ints("tokenizer.ggml.token_type"),
 				Merges: c.Strings("tokenizer.ggml.merges"),

model/models/qwen3next/deltanet.go

@@ -406,8 +406,10 @@ func (gdn *GatedDeltaNet) deltaNetChunked(
 	gDiff := gCumsum.Neg(ctx).Add(ctx, gLast)
 	gDiffExp := gDiff.Exp(ctx)
 
-	// key_gdiff = k * exp(g_diff)
-	keyGDiff := k.Mul(ctx, gDiffExp)
+	// Reshapes g_diff_exp to [1, chunkSize, nChunks, ...]
+	gDiffExpReshaped := gDiffExp.Reshape(ctx, 1, chunkSize, nChunks, numVHeads*nSeqs)
+	keyGDiff := k.Mul(ctx, gDiffExpReshaped)
+	keyGDiffT := keyGDiff.Permute(ctx, 1, 0, 2, 3).Contiguous(ctx)
 
 	// Process chunks and update state
 	var coreAttnOut ml.Tensor
@@ -444,12 +446,9 @@ func (gdn *GatedDeltaNet) deltaNetChunked(
 			coreAttnOut = coreAttnOut.Concat(ctx, coreAttnOutChunk, 1)
 		}
 
-		// Update state for next chunk using pre-computed values
+		// Update state for next chunk
 		gExpLastChunk := gLastExp.Slice(ctx, 2, chunk, chunk+1, 1)
-		kGDiffChunk := keyGDiff.Slice(ctx, 2, chunk, chunk+1, 1)
-
-		// kgdmulvnew = key_gdiff^T @ v_new
-		kGDiffChunkT := kGDiffChunk.Permute(ctx, 1, 0, 2, 3).Contiguous(ctx)
+		kGDiffChunkT := keyGDiffT.Slice(ctx, 2, chunk, chunk+1, 1)
 		kgdMulVNew := vNewT.Mulmat(ctx, kGDiffChunkT)
 
 		// state = state * g_last + kgdmulvnew

model/models/qwen3next/model.go

@@ -11,6 +11,7 @@ import (
 	"github.com/ollama/ollama/ml/nn/rope"
 	"github.com/ollama/ollama/model"
 	"github.com/ollama/ollama/model/input"
+	"github.com/ollama/ollama/tokenizer"
 )
 
 // Options contains model configuration
@@ -135,7 +136,7 @@ func (mlp *sparse) Forward(ctx ml.Context, hiddenStates ml.Tensor, opts *Options
 		// Apply shared expert gating
 		if mlp.SharedGateInp != nil {
 			sharedGateVal := mlp.SharedGateInp.Forward(ctx, hiddenStates2D)
-			sharedGateVal = sharedGateVal.Sigmoid(ctx)
+			sharedGateVal = sharedGateVal.SigmoidOut(ctx)
 			// Broadcast gate to match dimensions
 			sharedGateVal = sharedGateVal.Repeat(ctx, 0, sharedOut.Dim(0))
 			sharedOut = sharedOut.Mul(ctx, sharedGateVal)
@@ -207,7 +208,7 @@ func (l *Layer) Forward(ctx ml.Context, layer int, hiddenStates, positions, outp
 // Model is the main Qwen3-Next model
 type Model struct {
 	model.Base
-	model.BytePairEncoding
+	tokenizer.Tokenizer
 
 	TokenEmbedding *nn.Embedding `gguf:"token_embd"`
 	OutputNorm     *nn.RMSNorm   `gguf:"output_norm"`
@@ -353,8 +354,8 @@ func New(c fs.Config) (model.Model, error) {
 	}
 
 	m := Model{
-		BytePairEncoding: model.NewBytePairEncoding(
-			&model.Vocabulary{
+		Tokenizer: tokenizer.NewBytePairEncoding(
+			&tokenizer.Vocabulary{
 				Values: c.Strings("tokenizer.ggml.tokens"),
 				Types:  c.Ints("tokenizer.ggml.token_type"),
 				Merges: c.Strings("tokenizer.ggml.merges"),

model/models/qwen3vl/model.go

@@ -10,11 +10,12 @@ import (
 	"github.com/ollama/ollama/ml"
 	"github.com/ollama/ollama/model"
 	"github.com/ollama/ollama/model/input"
+	"github.com/ollama/ollama/tokenizer"
 )
 
 type Model struct {
 	model.Base
-	model.TextProcessor
+	tokenizer.Tokenizer
 
 	*TextModel
 	*VisionModel `gguf:"v"`
@@ -172,8 +173,8 @@ func (m *Model) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
 
 func New(c fs.Config) (model.Model, error) {
 	m := Model{
-		TextProcessor: model.NewBytePairEncoding(
-			&model.Vocabulary{
+		Tokenizer: tokenizer.NewBytePairEncoding(
+			&tokenizer.Vocabulary{
 				Values: c.Strings("tokenizer.ggml.tokens"),
 				Types:  c.Ints("tokenizer.ggml.token_type"),
 				Merges: c.Strings("tokenizer.ggml.merges"),

model/wordpiece_test.go

@@ -1,53 +0,0 @@
-package model
-
-import (
-	"slices"
-	"testing"
-
-	"github.com/google/go-cmp/cmp"
-)
-
-func TestWordPiece(t *testing.T) {
-	wpm := NewWordPiece(
-		&Vocabulary{
-			Values: []string{"[UNK]", "[CLS]", "[SEP]", "▁hello", "▁world", "s", "▁!", "▁@", "▁#"},
-			AddBOS: true,
-			AddEOS: true,
-			BOS:    []int32{1},
-			EOS:    []int32{2},
-		},
-		true, // lowercase
-	)
-
-	ids, err := wpm.Encode("Hello world!", true)
-	if err != nil {
-		t.Fatal(err)
-	}
-
-	if diff := cmp.Diff([]int32{1, 3, 4, 6, 2}, ids); diff != "" {
-		t.Errorf("unexpected ids (-want +got):\n%s", diff)
-	}
-
-	words, err := wpm.Decode(ids)
-	if err != nil {
-		t.Fatal(err)
-	}
-
-	if diff := cmp.Diff("[CLS] hello world! [SEP]", words); diff != "" {
-		t.Errorf("unexpected words (-want +got):\n%s", diff)
-	}
-}
-
-func TestWordPieceWords(t *testing.T) {
-	var wpm WordPiece
-
-	basic := slices.Collect(wpm.words("Hey friend!     How are you?!?"))
-	if diff := cmp.Diff([]string{"Hey", "friend", "!", "How", "are", "you", "?", "!", "?"}, basic); diff != "" {
-		t.Errorf("unexpected words (-want +got):\n%s", diff)
-	}
-
-	chinese := slices.Collect(wpm.words("野口里佳 Noguchi Rika"))
-	if diff := cmp.Diff([]string{"野", "口", "里", "佳", "Noguchi", "Rika"}, chinese); diff != "" {
-		t.Errorf("unexpected words (-want +got):\n%s", diff)
-	}
-}

runner/ollamarunner/runner.go

@@ -37,6 +37,7 @@ import (
 	"github.com/ollama/ollama/model/input"
 	"github.com/ollama/ollama/runner/common"
 	"github.com/ollama/ollama/sample"
+	"github.com/ollama/ollama/tokenizer"
 
 	_ "github.com/ollama/ollama/model/models"
 )
@@ -210,9 +211,9 @@ func (s *Server) NewSequence(prompt string, images []llm.ImageData, params NewSe
 }
 
 // calculateLogprobs converts raw logits to log probabilities and finds top K tokens
-func calculateLogprobs(logits []float32, selectedToken int32, topK int, textProcessor model.TextProcessor) []llm.Logprob {
+func calculateLogprobs(logits []float32, selectedToken int32, topK int, tok tokenizer.Tokenizer) []llm.Logprob {
 	decoder := func(tokenID int) string {
-		text, _ := textProcessor.Decode([]int32{int32(tokenID)})
+		text, _ := tok.Decode([]int32{int32(tokenID)})
 		return text
 	}
 	return common.CalculateLogprobs(logits, int(selectedToken), topK, decoder)
@@ -242,7 +243,7 @@ func (s *Server) inputs(prompt string, images []llm.ImageData) ([]*input.Input,
 
 	for i, part := range parts {
 		// text - tokenize
-		tokens, err := s.model.(model.TextProcessor).Encode(part, i == 0)
+		tokens, err := s.model.(tokenizer.Tokenizer).Encode(part, i == 0)
 		if err != nil {
 			return nil, nil, nil, err
 		}
@@ -514,13 +515,6 @@ func (s *Server) forwardBatch(pendingBatch batchState) (nextBatch batchState, er
 			continue
 		}
 
-		// if past the num predict limit
-		if seq.numPredict > 0 && seq.numPredicted >= seq.numPredict {
-			s.removeSequence(seqIdx, llm.DoneReasonLength)
-			nextBatch.seqs[seqIdx] = nil
-			continue
-		}
-
 		if !s.cache.enabled {
 			seq.inputs = append(seq.cache.Inputs, seq.inputs...)
 			seq.cache.Inputs = []*input.Input{}
@@ -709,7 +703,6 @@ func (s *Server) computeBatch(activeBatch batchState) {
 			continue
 		}
 
-		seq.numPredicted++
 		nextToken := &input.Input{Token: 0} // placeholder we'll fill in after Compute/Floats
 		seq.inputs = []*input.Input{nextToken}
 		nextBatchTokens[i] = nextToken
@@ -740,8 +733,14 @@ func (s *Server) computeBatch(activeBatch batchState) {
 		if seq == nil || nextBatchTokens[i] == nil {
 			continue
 		}
+		// If the sequence was replaced while this batch was computing, discard results.
+		if activeBatch.seqs[i] != seq {
+			logutil.Trace("computeBatch: sequence replaced, discarding its results", "batchID", activeBatch.id, "seqIdx", i)
+			continue
+		}
 
 		seq.lastUpdatedAt = t
+		seq.numPredicted++
 		if seq.numPredicted == 1 {
 			seq.processingDuration = seq.lastUpdatedAt.Sub(seq.startedAt)
 			seq.startedAt = seq.lastUpdatedAt
@@ -766,7 +765,7 @@ func (s *Server) computeBatch(activeBatch batchState) {
 		nextBatchTokens[i].Token = token
 
 		// if it's an end of sequence token, break
-		if s.model.(model.TextProcessor).Is(token, model.SpecialEOS) {
+		if s.model.(tokenizer.Tokenizer).Is(token, tokenizer.SpecialEOS) {
 			// TODO (jmorganca): we should send this back
 			// as it's important for the /api/generate context
 			// seq.responses <- piece
@@ -775,18 +774,25 @@ func (s *Server) computeBatch(activeBatch batchState) {
 			continue
 		}
 
-		piece, err := s.model.(model.TextProcessor).Decode([]int32{token})
+		piece, err := s.model.(tokenizer.Tokenizer).Decode([]int32{token})
 		if err != nil {
 			panic("failed to decode token")
 		}
 
 		// Calculate logprobs if requested (after EOS check to avoid logprobs for EOS tokens)
 		if seq.logprobs {
-			logprobs := calculateLogprobs(logits, token, seq.topLogprobs, s.model.(model.TextProcessor))
+			logprobs := calculateLogprobs(logits, token, seq.topLogprobs, s.model.(tokenizer.Tokenizer))
 			seq.pendingLogprobs = append(seq.pendingLogprobs, logprobs...)
 		}
 
 		seq.pendingResponses = append(seq.pendingResponses, piece)
+
+		// if past the num predict limit
+		if seq.numPredict > 0 && seq.numPredicted >= seq.numPredict {
+			s.removeSequence(i, llm.DoneReasonLength)
+			continue
+		}
+
 		sequence := strings.Join(seq.pendingResponses, "")
 
 		if ok, stop := common.FindStop(sequence, seq.stop); ok {
@@ -873,7 +879,7 @@ func (s *Server) completion(w http.ResponseWriter, r *http.Request) {
 	var grammar *sample.GrammarSampler
 	var err error
 	if req.Grammar != "" {
-		grammar, err = sample.NewGrammarSampler(s.model.(model.TextProcessor), req.Grammar)
+		grammar, err = sample.NewGrammarSampler(s.model.(tokenizer.Tokenizer), req.Grammar)
 		if err != nil {
 			http.Error(w, "failed to load model vocabulary required for format", http.StatusInternalServerError)
 			return

runner/runner.go

@@ -3,7 +3,7 @@ package runner
 import (
 	"github.com/ollama/ollama/runner/llamarunner"
 	"github.com/ollama/ollama/runner/ollamarunner"
-	"github.com/ollama/ollama/x/mlxrunner"
+	"github.com/ollama/ollama/x/imagegen"
 )
 
 func Execute(args []string) error {
@@ -11,22 +11,13 @@ func Execute(args []string) error {
 		args = args[1:]
 	}
 
-	var newRunner bool
-	var mlxRunner bool
-	if len(args) > 0 && args[0] == "--ollama-engine" {
-		args = args[1:]
-		newRunner = true
-	}
-	if len(args) > 0 && args[0] == "--mlx-engine" {
-		args = args[1:]
-		mlxRunner = true
-	}
-
-	if mlxRunner {
-		return mlxrunner.Execute(args)
-	} else if newRunner {
-		return ollamarunner.Execute(args)
-	} else {
-		return llamarunner.Execute(args)
+	if len(args) > 0 {
+		switch args[0] {
+		case "--ollama-engine":
+			return ollamarunner.Execute(args[1:])
+		case "--imagegen-engine":
+			return imagegen.Execute(args[1:])
+		}
 	}
+	return llamarunner.Execute(args)
 }

sample/samplers.go

@@ -7,7 +7,7 @@ import (
 	"slices"
 
 	"github.com/ollama/ollama/llama"
-	"github.com/ollama/ollama/model"
+	"github.com/ollama/ollama/tokenizer"
 )
 
 // token represents information about a single token during sampling
@@ -168,15 +168,15 @@ type GrammarSampler struct {
 	grammar *llama.Grammar
 }
 
-func NewGrammarSampler(model model.TextProcessor, grammarStr string) (*GrammarSampler, error) {
-	vocabIds := make([]uint32, len(model.Vocabulary().Values))
-	pieces := make([]string, len(model.Vocabulary().Values))
-	for i := range model.Vocabulary().Values {
-		pieces[i], _ = model.Decode([]int32{int32(i)})
+func NewGrammarSampler(tok tokenizer.Tokenizer, grammarStr string) (*GrammarSampler, error) {
+	vocabIds := make([]uint32, len(tok.Vocabulary().Values))
+	pieces := make([]string, len(tok.Vocabulary().Values))
+	for i := range tok.Vocabulary().Values {
+		pieces[i], _ = tok.Decode([]int32{int32(i)})
 		vocabIds[i] = uint32(i)
 	}
 
-	grammar := llama.NewGrammar(grammarStr, vocabIds, pieces, model.Vocabulary().EOS)
+	grammar := llama.NewGrammar(grammarStr, vocabIds, pieces, tok.Vocabulary().EOS)
 	if grammar == nil {
 		return nil, errors.New("sample: failed to initialize grammar")
 	}

sample/samplers_test.go

@@ -8,7 +8,7 @@ import (
 	"path/filepath"
 	"testing"
 
-	"github.com/ollama/ollama/model"
+	"github.com/ollama/ollama/tokenizer"
 )
 
 func TestWeighted(t *testing.T) {
@@ -60,10 +60,10 @@ func TestWeighted(t *testing.T) {
 	}
 }
 
-func modelHelper(t testing.TB) model.BytePairEncoding {
+func modelHelper(t testing.TB) tokenizer.Tokenizer {
 	t.Helper()
 
-	f, err := os.Open(filepath.Join("..", "model", "testdata", "llama3.2", "encoder.json"))
+	f, err := os.Open(filepath.FromSlash("../tokenizer/testdata/llama3.2/encoder.json"))
 	if err != nil {
 		t.Fatal(err)
 	}
@@ -81,8 +81,8 @@ func modelHelper(t testing.TB) model.BytePairEncoding {
 
 	merges := make([]string, 0, 1)
 	// Only need vocab for Grammar Test
-	return model.NewBytePairEncoding(
-		&model.Vocabulary{
+	return tokenizer.NewBytePairEncoding(
+		&tokenizer.Vocabulary{
 			Values: tokens,
 			Types:  make([]int32, len(vocab)),
 			Merges: merges,

scripts/install.sh

@@ -1,5 +1,5 @@
 #!/bin/sh
-# This script installs Ollama on Linux.
+# This script installs Ollama on Linux and macOS.
 # It detects the current operating system architecture and installs the appropriate version of Ollama.
 
 set -eu
@@ -27,8 +27,7 @@ require() {
     echo $MISSING
 }
 
-[ "$(uname -s)" = "Linux" ] || error 'This script is intended to run on Linux only.'
-
+OS="$(uname -s)"
 ARCH=$(uname -m)
 case "$ARCH" in
     x86_64) ARCH="amd64" ;;
@@ -36,6 +35,65 @@ case "$ARCH" in
     *) error "Unsupported architecture: $ARCH" ;;
 esac
 
+###########################################
+# macOS
+###########################################
+
+if [ "$OS" = "Darwin" ]; then
+    NEEDS=$(require curl unzip)
+    if [ -n "$NEEDS" ]; then
+        status "ERROR: The following tools are required but missing:"
+        for NEED in $NEEDS; do
+            echo "  - $NEED"
+        done
+        exit 1
+    fi
+
+    if [ -n "${OLLAMA_VERSION:-}" ]; then
+        DOWNLOAD_URL="https://github.com/ollama/ollama/releases/download/${OLLAMA_VERSION}/Ollama-darwin.zip"
+    else
+        DOWNLOAD_URL="https://github.com/ollama/ollama/releases/latest/download/Ollama-darwin.zip"
+    fi
+
+    if pgrep -x Ollama >/dev/null 2>&1; then
+        status "Stopping running Ollama instance..."
+        pkill -x Ollama 2>/dev/null || true
+        sleep 2
+    fi
+
+    if [ -d "/Applications/Ollama.app" ]; then
+        status "Removing existing Ollama installation..."
+        rm -rf "/Applications/Ollama.app"
+    fi
+
+    status "Downloading Ollama for macOS..."
+    curl --fail --show-error --location --progress-bar \
+        -o "$TEMP_DIR/Ollama-darwin.zip" "$DOWNLOAD_URL"
+
+    status "Installing Ollama to /Applications..."
+    unzip -q "$TEMP_DIR/Ollama-darwin.zip" -d "$TEMP_DIR"
+    mv "$TEMP_DIR/Ollama.app" "/Applications/"
+
+    status "Adding 'ollama' command to PATH (may require password)..."
+    mkdir -p "/usr/local/bin" 2>/dev/null || sudo mkdir -p "/usr/local/bin"
+    ln -sf "/Applications/Ollama.app/Contents/Resources/ollama" "/usr/local/bin/ollama" 2>/dev/null || \
+        sudo ln -sf "/Applications/Ollama.app/Contents/Resources/ollama" "/usr/local/bin/ollama"
+
+    if [ -z "${OLLAMA_NO_START:-}" ]; then
+        status "Starting Ollama..."
+        open -a Ollama --args hidden
+    fi
+
+    status "Install complete. You can now run 'ollama'."
+    exit 0
+fi
+
+###########################################
+# Linux
+###########################################
+
+[ "$OS" = "Linux" ] || error 'This script is intended to run on Linux and macOS only.'
+
 IS_WSL2=false
 
 KERN=$(uname -r)

server/aliases.go

@@ -16,8 +16,8 @@ import (
 )
 
 const (
-	routerConfigFilename = "server.json"
-	routerConfigVersion  = 1
+	serverConfigFilename = "server.json"
+	serverConfigVersion  = 1
 )
 
 var errAliasCycle = errors.New("alias cycle detected")
@@ -28,20 +28,20 @@ type aliasEntry struct {
 	PrefixMatching bool   `json:"prefix_matching,omitempty"`
 }
 
-type routerConfig struct {
+type serverConfig struct {
 	Version int          `json:"version"`
 	Aliases []aliasEntry `json:"aliases"`
 }
 
-type aliasStore struct {
+type store struct {
 	mu            sync.RWMutex
 	path          string
 	entries       map[string]aliasEntry // normalized alias -> entry (exact matches)
 	prefixEntries []aliasEntry          // prefix matches, sorted longest-first
 }
 
-func newAliasStore(path string) (*aliasStore, error) {
-	store := &aliasStore{
+func createStore(path string) (*store, error) {
+	store := &store{
 		path:    path,
 		entries: make(map[string]aliasEntry),
 	}
@@ -51,7 +51,7 @@ func newAliasStore(path string) (*aliasStore, error) {
 	return store, nil
 }
 
-func (s *aliasStore) load() error {
+func (s *store) load() error {
 	data, err := os.ReadFile(s.path)
 	if err != nil {
 		if errors.Is(err, os.ErrNotExist) {
@@ -60,12 +60,12 @@ func (s *aliasStore) load() error {
 		return err
 	}
 
-	var cfg routerConfig
+	var cfg serverConfig
 	if err := json.Unmarshal(data, &cfg); err != nil {
 		return err
 	}
 
-	if cfg.Version != 0 && cfg.Version != routerConfigVersion {
+	if cfg.Version != 0 && cfg.Version != serverConfigVersion {
 		return fmt.Errorf("unsupported router config version %d", cfg.Version)
 	}
 
@@ -109,7 +109,7 @@ func (s *aliasStore) load() error {
 	return nil
 }
 
-func (s *aliasStore) saveLocked() error {
+func (s *store) saveLocked() error {
 	dir := filepath.Dir(s.path)
 	if err := os.MkdirAll(dir, 0o755); err != nil {
 		return err
@@ -126,8 +126,8 @@ func (s *aliasStore) saveLocked() error {
 		return strings.Compare(entries[i].Alias, entries[j].Alias) < 0
 	})
 
-	cfg := routerConfig{
-		Version: routerConfigVersion,
+	cfg := serverConfig{
+		Version: serverConfigVersion,
 		Aliases: entries,
 	}
 
@@ -157,7 +157,7 @@ func (s *aliasStore) saveLocked() error {
 	return os.Rename(f.Name(), s.path)
 }
 
-func (s *aliasStore) ResolveName(name model.Name) (model.Name, bool, error) {
+func (s *store) ResolveName(name model.Name) (model.Name, bool, error) {
 	// If a local model exists, do not allow alias shadowing (highest priority).
 	exists, err := localModelExists(name)
 	if err != nil {
@@ -226,7 +226,7 @@ func (s *aliasStore) ResolveName(name model.Name) (model.Name, bool, error) {
 	}
 }
 
-func (s *aliasStore) Set(alias, target model.Name, prefixMatching bool) error {
+func (s *store) Set(alias, target model.Name, prefixMatching bool) error {
 	targetKey := normalizeAliasKey(target)
 
 	s.mu.Lock()
@@ -283,7 +283,7 @@ func (s *aliasStore) Set(alias, target model.Name, prefixMatching bool) error {
 	return s.saveLocked()
 }
 
-func (s *aliasStore) Delete(alias model.Name) (bool, error) {
+func (s *store) Delete(alias model.Name) (bool, error) {
 	aliasKey := normalizeAliasKey(alias)
 
 	s.mu.Lock()
@@ -309,7 +309,7 @@ func (s *aliasStore) Delete(alias model.Name) (bool, error) {
 
 // DeleteByString deletes an alias by its raw string value, useful for prefix
 // aliases that may not be valid model names.
-func (s *aliasStore) DeleteByString(alias string) (bool, error) {
+func (s *store) DeleteByString(alias string) (bool, error) {
 	alias = strings.TrimSpace(alias)
 	aliasLower := strings.ToLower(alias)
 
@@ -333,7 +333,7 @@ func (s *aliasStore) DeleteByString(alias string) (bool, error) {
 	return false, nil
 }
 
-func (s *aliasStore) List() []aliasEntry {
+func (s *store) List() []aliasEntry {
 	s.mu.RLock()
 	defer s.mu.RUnlock()
 
@@ -353,7 +353,7 @@ func normalizeAliasKey(name model.Name) string {
 	return strings.ToLower(displayAliasName(name))
 }
 
-func (s *aliasStore) sortPrefixEntriesLocked() {
+func (s *store) sortPrefixEntriesLocked() {
 	sort.Slice(s.prefixEntries, func(i, j int) bool {
 		// Sort by length descending (longest prefix first)
 		return len(s.prefixEntries[i].Alias) > len(s.prefixEntries[j].Alias)
@@ -392,23 +392,23 @@ func localModelExists(name model.Name) (bool, error) {
 	return false, nil
 }
 
-func routerConfigPath() string {
+func serverConfigPath() string {
 	home, err := os.UserHomeDir()
 	if err != nil {
-		return filepath.Join(".ollama", routerConfigFilename)
+		return filepath.Join(".ollama", serverConfigFilename)
 	}
-	return filepath.Join(home, ".ollama", routerConfigFilename)
+	return filepath.Join(home, ".ollama", serverConfigFilename)
 }
 
-func (s *Server) aliasStore() (*aliasStore, error) {
+func (s *Server) aliasStore() (*store, error) {
 	s.aliasesOnce.Do(func() {
-		s.aliases, s.aliasesErr = newAliasStore(routerConfigPath())
+		s.aliases, s.aliasesErr = createStore(serverConfigPath())
 	})
 
 	return s.aliases, s.aliasesErr
 }
 
-func (s *Server) resolveModelAliasName(name model.Name) (model.Name, bool, error) {
+func (s *Server) resolveAlias(name model.Name) (model.Name, bool, error) {
 	store, err := s.aliasStore()
 	if err != nil {
 		return name, false, err

server/prompt.go

@@ -27,14 +27,12 @@ func chatPrompt(ctx context.Context, m *Model, tokenize tokenizeFunc, opts *api.
 	// Clip images are represented as 768 tokens, each an embedding
 	imageNumTokens := 768
 
-	n := len(msgs) - 1
-	// in reverse, find all messages that fit into context window
-	for i := n; i >= 0; i-- {
-		// always include the last message
-		if i == n {
-			continue
-		}
+	lastMsgIdx := len(msgs) - 1
+	currMsgIdx := 0
 
+	// Start with all messages and remove from the front until it fits in context
+	for i := 0; i <= lastMsgIdx; i++ {
+		// Collect system messages from the portion we're about to skip
 		system = make([]api.Message, 0)
 		for j := range i {
 			if msgs[j].Role == "system" {
@@ -54,20 +52,26 @@ func chatPrompt(ctx context.Context, m *Model, tokenize tokenizeFunc, opts *api.
 
 		ctxLen := len(s)
 		if m.ProjectorPaths != nil {
-			for _, m := range msgs[i:] {
-				ctxLen += imageNumTokens * len(m.Images)
+			for _, msg := range msgs[i:] {
+				ctxLen += imageNumTokens * len(msg.Images)
 			}
 		}
 
-		if truncate && ctxLen > opts.NumCtx {
-			slog.Debug("truncating input messages which exceed context length", "truncated", len(msgs[i:]))
+		if !truncate || ctxLen <= opts.NumCtx {
+			currMsgIdx = i
+			break
+		}
+
+		// Must always include at least the last message
+		if i == lastMsgIdx {
+			currMsgIdx = lastMsgIdx
 			break
-		} else {
-			n = i
 		}
 	}
 
-	currMsgIdx := n
+	if currMsgIdx > 0 {
+		slog.Debug("truncating input messages which exceed context length", "truncated", len(msgs[currMsgIdx:]))
+	}
 
 	for cnt, msg := range msgs[currMsgIdx:] {
 		if slices.Contains(m.Config.ModelFamilies, "mllama") && len(msg.Images) > 1 {

server/prompt_test.go

@@ -2,6 +2,7 @@ package server
 
 import (
 	"bytes"
+	"context"
 	"testing"
 
 	"github.com/google/go-cmp/cmp"
@@ -264,3 +265,68 @@ func TestChatPrompt(t *testing.T) {
 		})
 	}
 }
+
+func TestChatPromptTokenizeCalls(t *testing.T) {
+	tmpl, err := template.Parse(`
+{{- if .System }}{{ .System }} {{ end }}
+{{- if .Prompt }}{{ .Prompt }} {{ end }}
+{{- if .Response }}{{ .Response }} {{ end }}`)
+	if err != nil {
+		t.Fatal(err)
+	}
+	model := Model{Template: tmpl}
+
+	cases := []struct {
+		name         string
+		limit        int
+		msgs         []api.Message
+		maxTokenizes int
+	}{
+		{
+			name:  "all messages fit",
+			limit: 2048,
+			msgs: []api.Message{
+				{Role: "user", Content: "message 1"},
+				{Role: "assistant", Content: "response 1"},
+				{Role: "user", Content: "message 2"},
+				{Role: "assistant", Content: "response 2"},
+				{Role: "user", Content: "message 3"},
+			},
+			maxTokenizes: 1,
+		},
+		{
+			name:  "truncate to last message",
+			limit: 5,
+			msgs: []api.Message{
+				{Role: "user", Content: "message 1"},
+				{Role: "assistant", Content: "response 1"},
+				{Role: "user", Content: "message 2"},
+				{Role: "assistant", Content: "response 2"},
+				{Role: "user", Content: "message 3"},
+			},
+			maxTokenizes: 5,
+		},
+	}
+
+	for _, tt := range cases {
+		t.Run(tt.name, func(t *testing.T) {
+			tokenizeCount := 0
+			countingTokenize := func(ctx context.Context, s string) ([]int, error) {
+				tokenizeCount++
+				tokens, err := mockRunner{}.Tokenize(ctx, s)
+				return tokens, err
+			}
+
+			opts := api.Options{Runner: api.Runner{NumCtx: tt.limit}}
+			think := false
+			_, _, err := chatPrompt(t.Context(), &model, countingTokenize, &opts, tt.msgs, nil, &api.ThinkValue{Value: think}, true)
+			if err != nil {
+				t.Fatal(err)
+			}
+
+			if tokenizeCount > tt.maxTokenizes {
+				t.Errorf("tokenize called %d times, expected at most %d", tokenizeCount, tt.maxTokenizes)
+			}
+		})
+	}
+}

server/routes.go

@@ -52,7 +52,7 @@ import (
 	"github.com/ollama/ollama/types/errtypes"
 	"github.com/ollama/ollama/types/model"
 	"github.com/ollama/ollama/version"
-	"github.com/ollama/ollama/x/imagegen"
+	imagegenmanifest "github.com/ollama/ollama/x/imagegen/manifest"
 	xserver "github.com/ollama/ollama/x/server"
 )
 
@@ -83,7 +83,7 @@ type Server struct {
 	sched         *Scheduler
 	defaultNumCtx int
 	aliasesOnce   sync.Once
-	aliases       *aliasStore
+	aliases       *store
 	aliasesErr    error
 }
 
@@ -195,7 +195,7 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 		return
 	}
 
-	resolvedName, _, err := s.resolveModelAliasName(name)
+	resolvedName, _, err := s.resolveAlias(name)
 	if err != nil {
 		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
 		return
@@ -1106,7 +1106,7 @@ func GetModelInfo(req api.ShowRequest) (*api.ShowResponse, error) {
 
 	// For image generation models, populate details from imagegen package
 	if slices.Contains(m.Capabilities(), model.CapabilityImage) {
-		if info, err := imagegen.GetModelInfo(name.String()); err == nil {
+		if info, err := imagegenmanifest.GetModelInfo(name.String()); err == nil {
 			modelDetails.Family = info.Architecture
 			modelDetails.ParameterSize = format.HumanNumber(uint64(info.ParameterCount))
 			modelDetails.QuantizationLevel = info.Quantization
@@ -1964,7 +1964,7 @@ func (s *Server) ChatHandler(c *gin.Context) {
 		return
 	}
 
-	resolvedName, _, err := s.resolveModelAliasName(name)
+	resolvedName, _, err := s.resolveAlias(name)
 	if err != nil {
 		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
 		return

server/routes_aliases_test.go

@@ -113,7 +113,7 @@ func TestAliasResolvesForChatRemote(t *testing.T) {
 
 func TestPrefixAliasBasicMatching(t *testing.T) {
 	tmpDir := t.TempDir()
-	store, err := newAliasStore(filepath.Join(tmpDir, "server.json"))
+	store, err := createStore(filepath.Join(tmpDir, "server.json"))
 	if err != nil {
 		t.Fatal(err)
 	}
@@ -174,7 +174,7 @@ func TestPrefixAliasBasicMatching(t *testing.T) {
 
 func TestPrefixAliasLongestMatchWins(t *testing.T) {
 	tmpDir := t.TempDir()
-	store, err := newAliasStore(filepath.Join(tmpDir, "server.json"))
+	store, err := createStore(filepath.Join(tmpDir, "server.json"))
 	if err != nil {
 		t.Fatal(err)
 	}
@@ -219,7 +219,7 @@ func TestPrefixAliasLongestMatchWins(t *testing.T) {
 
 func TestPrefixAliasChain(t *testing.T) {
 	tmpDir := t.TempDir()
-	store, err := newAliasStore(filepath.Join(tmpDir, "server.json"))
+	store, err := createStore(filepath.Join(tmpDir, "server.json"))
 	if err != nil {
 		t.Fatal(err)
 	}
@@ -325,7 +325,7 @@ func TestPrefixAliasCRUD(t *testing.T) {
 
 func TestPrefixAliasCaseInsensitive(t *testing.T) {
 	tmpDir := t.TempDir()
-	store, err := newAliasStore(filepath.Join(tmpDir, "server.json"))
+	store, err := createStore(filepath.Join(tmpDir, "server.json"))
 	if err != nil {
 		t.Fatal(err)
 	}

server/sched.go

@@ -21,7 +21,7 @@ import (
 	"github.com/ollama/ollama/logutil"
 	"github.com/ollama/ollama/ml"
 	"github.com/ollama/ollama/types/model"
-	"github.com/ollama/ollama/x/mlxrunner"
+	"github.com/ollama/ollama/x/imagegen"
 )
 
 type LlmRequest struct {
@@ -417,9 +417,9 @@ func (s *Scheduler) load(req *LlmRequest, f *ggml.GGML, systemInfo ml.SystemInfo
 		numParallel = 1
 	}
 
-	// `mllama`, `qwen3vl`, and `qwen3vlmoe` are snowflakes and uses an encoder cache which cannot be used with num_parallel > 1
+	// Some architectures are not safe with num_parallel > 1.
 	// ref: https://github.com/ollama/ollama/issues/4165
-	if slices.Contains([]string{"mllama", "qwen3vl", "qwen3vlmoe"}, req.model.Config.ModelFamily) && numParallel != 1 {
+	if slices.Contains([]string{"mllama", "qwen3vl", "qwen3vlmoe", "qwen3next", "lfm2", "lfm2moe"}, req.model.Config.ModelFamily) && numParallel != 1 {
 		numParallel = 1
 		slog.Warn("model architecture does not currently support parallel requests", "architecture", req.model.Config.ModelFamily)
 	}
@@ -567,16 +567,16 @@ iGPUScan:
 // This supports both LLM (completion) and image generation models.
 func (s *Scheduler) loadMLX(req *LlmRequest) bool {
 	// Determine mode based on capabilities
-	var mode mlxrunner.ModelMode
+	var mode imagegen.ModelMode
 	if slices.Contains(req.model.Config.Capabilities, "image") {
-		mode = mlxrunner.ModeImageGen
+		mode = imagegen.ModeImageGen
 	} else {
-		mode = mlxrunner.ModeLLM
+		mode = imagegen.ModeLLM
 	}
 
 	// Use model name for MLX (it resolves manifests by name, not file path)
 	modelName := req.model.ShortName
-	server, err := mlxrunner.NewServer(modelName, mode)
+	server, err := imagegen.NewServer(modelName, mode)
 	if err != nil {
 		req.errCh <- err
 		return true

tokenizer/bytepairencoding.go

@@ -1,4 +1,4 @@
-package model
+package tokenizer
 
 import (
 	"cmp"
@@ -18,19 +18,19 @@ type BytePairEncoding struct {
 	regexps []*regexp2.Regexp
 }
 
-var _ TextProcessor = (*BytePairEncoding)(nil)
+var _ Tokenizer = (*BytePairEncoding)(nil)
 
-func NewBytePairEncoding(vocab *Vocabulary, pretokenizers ...string) BytePairEncoding {
-	if len(pretokenizers) == 0 {
+func NewBytePairEncoding(vocab *Vocabulary, pretokenizer ...string) BytePairEncoding {
+	if len(pretokenizer) == 0 {
 		// set default byte-level pretokenizer if none provided, e.g.
-		// https://github.com/huggingface/tokenizers/blob/main/tokenizers/src/pre_tokenizers/byte_level.rs#L44
-		pretokenizers = []string{`'s|'t|'re|'ve|'m|'ll|'d| ?\p{L}+| ?\p{N}+| ?[^\s\p{L}\p{N}]+|\s+(?!\S)|\s+`}
+		// https://github.com/huggingface/tokenizer/blob/main/tokenizer/src/pre_tokenizer/byte_level.rs#L44
+		pretokenizer = []string{`'s|'t|'re|'ve|'m|'ll|'d| ?\p{L}+| ?\p{N}+| ?[^\s\p{L}\p{N}]+|\s+(?!\S)|\s+`}
 	}
 
 	return BytePairEncoding{
 		vocab: vocab,
 		regexps: slices.Collect(func(yield func(*regexp2.Regexp) bool) {
-			for _, p := range pretokenizers {
+			for _, p := range pretokenizer {
 				if !yield(regexp2.MustCompile(p, regexp2.RE2)) {
 					return
 				}

tokenizer/bytepairencoding_test.go

@@ -1,4 +1,4 @@
-package model
+package tokenizer
 
 import (
 	"bufio"
@@ -17,7 +17,7 @@ import (
 func llama(t testing.TB) BytePairEncoding {
 	t.Helper()
 
-	f, err := os.Open(filepath.Join("testdata", "llama3.2", "encoder.json"))
+	f, err := os.Open(filepath.FromSlash("testdata/llama3.2/encoder.json"))
 	if err != nil {
 		t.Fatal(err)
 	}
@@ -43,7 +43,7 @@ func llama(t testing.TB) BytePairEncoding {
 		}
 	}
 
-	f, err = os.Open(filepath.Join("testdata", "llama3.2", "vocab.bpe"))
+	f, err = os.Open(filepath.FromSlash("testdata/llama3.2/vocab.bpe"))
 	if err != nil {
 		t.Fatal(err)
 	}

tokenizer/sentencepiece.go

@@ -1,4 +1,4 @@
-package model
+package tokenizer
 
 import (
 	"container/heap"
@@ -17,7 +17,7 @@ type SentencePiece struct {
 	vocab       *Vocabulary
 }
 
-var _ TextProcessor = (*SentencePiece)(nil)
+var _ Tokenizer = (*SentencePiece)(nil)
 
 func (spm SentencePiece) Vocabulary() *Vocabulary {
 	return spm.vocab
@@ -224,7 +224,7 @@ func (spm SentencePiece) Decode(ids []int32) (string, error) {
 		data := spm.vocab.Decode(id)
 		data = strings.ReplaceAll(data, spmWhitespaceSep, " ")
 
-		// For tokenizers that use byte tokens like "<0xEA>"
+		// For tokenizer that use byte tokens like "<0xEA>"
 		// convert them to the partial unicode character
 		// so they are buffered correctly by the runner instead
 		// of being sent back to the api as "<0xEA>"

tokenizer/sentencepiece_test.go

@@ -1,4 +1,4 @@
-package model
+package tokenizer
 
 import (
 	"log/slog"
@@ -15,7 +15,7 @@ import (
 func loadSentencePieceVocab(t *testing.T) SentencePiece {
 	t.Helper()
 
-	bts, err := os.ReadFile(filepath.Join("testdata", "gemma2", "tokenizer.model"))
+	bts, err := os.ReadFile(filepath.FromSlash("testdata/gemma2/tokenizer.model"))
 	if err != nil {
 		t.Fatal(err)
 	}

tokenizer/tokenizer.go

@@ -1,4 +1,4 @@
-package model
+package tokenizer
 
 const (
 	TOKEN_TYPE_NORMAL = iota + 1
@@ -9,7 +9,7 @@ const (
 	TOKEN_TYPE_BYTE
 )
 
-type TextProcessor interface {
+type Tokenizer interface {
 	Encode(s string, addSpecial bool) ([]int32, error)
 	Decode([]int32) (string, error)
 	Is(int32, Special) bool

tokenizer/vocabulary.go

@@ -1,4 +1,4 @@
-package model
+package tokenizer
 
 import (
 	"log/slog"

tokenizer/vocabulary_test.go

@@ -1,4 +1,4 @@
-package model
+package tokenizer
 
 import (
 	"testing"

tokenizer/wordpiece.go

@@ -1,4 +1,4 @@
-package model
+package tokenizer
 
 import (
 	"fmt"
@@ -32,7 +32,7 @@ var wordPieceReplacer = strings.NewReplacer(
 	" 're", "'re",
 )
 
-// Decode implements TextProcessor.
+// Decode implements Tokenizer.
 func (wpm WordPiece) Decode(ids []int32) (string, error) {
 	var sb strings.Builder
 	for i, id := range ids {
@@ -96,7 +96,7 @@ func (wpm WordPiece) words(s string) iter.Seq[string] {
 	}
 }
 
-// Encode implements TextProcessor.
+// Encode implements Tokenizer.
 func (wpm WordPiece) Encode(s string, addSpecial bool) ([]int32, error) {
 	var ids []int32
 
@@ -151,17 +151,17 @@ func (wpm WordPiece) Encode(s string, addSpecial bool) ([]int32, error) {
 	return ids, nil
 }
 
-// Is implements TextProcessor.
+// Is implements Tokenizer.
 func (wpm WordPiece) Is(id int32, special Special) bool {
 	return wpm.vocab.Is(id, special)
 }
 
-// Vocabulary implements TextProcessor.
+// Vocabulary implements Tokenizer.
 func (wpm WordPiece) Vocabulary() *Vocabulary {
 	return wpm.vocab
 }
 
-var _ TextProcessor = (*WordPiece)(nil)
+var _ Tokenizer = (*WordPiece)(nil)
 
 func NewWordPiece(vocab *Vocabulary, lowercase bool) WordPiece {
 	return WordPiece{

tokenizer/wordpiece_test.go

@@ -1,4 +1,4 @@
-package model
+package tokenizer
 
 import (
 	"slices"

x/imagegen/imagegen.go

@@ -1,6 +1,6 @@
 //go:build mlx
 
-package mlxrunner
+package imagegen
 
 import (
 	"context"
@@ -11,7 +11,7 @@ import (
 	"sync"
 	"time"
 
-	"github.com/ollama/ollama/x/imagegen"
+	"github.com/ollama/ollama/x/imagegen/manifest"
 	"github.com/ollama/ollama/x/imagegen/mlx"
 	"github.com/ollama/ollama/x/imagegen/models/flux2"
 	"github.com/ollama/ollama/x/imagegen/models/zimage"
@@ -28,8 +28,8 @@ var imageGenMu sync.Mutex
 func (s *server) loadImageModel() error {
 	// Check memory requirements before loading
 	var requiredMemory uint64
-	if manifest, err := imagegen.LoadManifest(s.modelName); err == nil {
-		requiredMemory = uint64(manifest.TotalTensorSize())
+	if modelManifest, err := manifest.LoadManifest(s.modelName); err == nil {
+		requiredMemory = uint64(modelManifest.TotalTensorSize())
 	}
 	availableMemory := mlx.GetMemoryLimit()
 	if availableMemory > 0 && requiredMemory > 0 && availableMemory < requiredMemory {
@@ -38,7 +38,7 @@ func (s *server) loadImageModel() error {
 	}
 
 	// Detect model type and load appropriate model
-	modelType := imagegen.DetectModelType(s.modelName)
+	modelType := DetectModelType(s.modelName)
 	slog.Info("detected image model type", "type", modelType)
 
 	var model ImageModel
@@ -108,7 +108,7 @@ func (s *server) handleImageCompletion(w http.ResponseWriter, r *http.Request, r
 	}
 
 	// Encode image as base64 PNG
-	imageData, err := imagegen.EncodeImageBase64(img)
+	imageData, err := EncodeImageBase64(img)
 	if err != nil {
 		resp := Response{Content: fmt.Sprintf("error encoding: %v", err), Done: true}
 		data, _ := json.Marshal(resp)

x/imagegen/llm.go

@@ -1,6 +1,6 @@
 //go:build mlx
 
-package mlxrunner
+package imagegen
 
 import (
 	"encoding/json"
@@ -12,8 +12,8 @@ import (
 	"sync"
 	"time"
 
-	"github.com/ollama/ollama/x/imagegen"
 	"github.com/ollama/ollama/x/imagegen/cache"
+	"github.com/ollama/ollama/x/imagegen/manifest"
 	"github.com/ollama/ollama/x/imagegen/mlx"
 	"github.com/ollama/ollama/x/imagegen/models/glm4_moe_lite"
 	"github.com/ollama/ollama/x/imagegen/tokenizer"
@@ -197,13 +197,13 @@ func sample(logits *mlx.Array, temp float32, vocabSize int32) *mlx.Array {
 // loadLLMModel loads a safetensors LLM model and its tokenizer from manifest storage.
 func (s *server) loadLLMModel() error {
 	// Load the manifest to get model information
-	manifest, err := imagegen.LoadManifest(s.modelName)
+	modelManifest, err := manifest.LoadManifest(s.modelName)
 	if err != nil {
 		return fmt.Errorf("failed to load manifest: %w", err)
 	}
 
 	// Detect model architecture from config.json
-	configData, err := manifest.ReadConfig("config.json")
+	configData, err := modelManifest.ReadConfig("config.json")
 	if err != nil {
 		return fmt.Errorf("failed to read config.json: %w", err)
 	}
@@ -232,7 +232,7 @@ func (s *server) loadLLMModel() error {
 
 	switch {
 	case strings.Contains(archLower, "glm4moelite"):
-		m, err := glm4_moe_lite.LoadFromManifest(manifest)
+		m, err := glm4_moe_lite.LoadFromManifest(modelManifest)
 		if err != nil {
 			return fmt.Errorf("failed to load glm4-moe-lite model: %w", err)
 		}

x/imagegen/manifest/manifest.go

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

x/imagegen/manifest/manifest_test.go

@@ -1,4 +1,4 @@
-package imagegen
+package manifest
 
 import (
 	"path/filepath"

x/imagegen/manifest/weights.go

@@ -1,6 +1,6 @@
 //go:build mlx
 
-package imagegen
+package manifest
 
 import (
 	"fmt"
@@ -15,9 +15,9 @@ import (
 type ManifestWeights struct {
 	manifest    *ModelManifest
 	component   string
-	tensors     map[string]ManifestLayer      // name -> layer
-	cache       map[string]*mlx.Array         // name -> loaded array
-	nativeCache []*mlx.SafetensorsFile        // keep native handles alive
+	tensors     map[string]ManifestLayer // name -> layer
+	cache       map[string]*mlx.Array    // name -> loaded array
+	nativeCache []*mlx.SafetensorsFile   // keep native handles alive
 }
 
 // LoadWeightsFromManifest creates a weight loader from manifest storage.

x/imagegen/memory.go

@@ -14,6 +14,8 @@ import (
 	"encoding/json"
 	"fmt"
 	"runtime"
+
+	"github.com/ollama/ollama/x/imagegen/manifest"
 )
 
 // SupportedBackends lists the backends that support image generation.
@@ -41,8 +43,8 @@ func CheckPlatformSupport() error {
 // 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() {
+	modelManifest, err := manifest.LoadManifest(modelName)
+	if err == nil && modelManifest.HasTensorLayers() {
 		return modelName
 	}
 	return ""
@@ -52,12 +54,12 @@ func ResolveModelName(modelName string) string {
 // Checks both "architecture" (Ollama format) and "_class_name" (diffusers format).
 // Returns empty string if detection fails.
 func DetectModelType(modelName string) string {
-	manifest, err := LoadManifest(modelName)
+	modelManifest, err := manifest.LoadManifest(modelName)
 	if err != nil {
 		return ""
 	}
 
-	data, err := manifest.ReadConfig("model_index.json")
+	data, err := modelManifest.ReadConfig("model_index.json")
 	if err != nil {
 		return ""
 	}

x/imagegen/models/flux2/flux2.go

@@ -12,7 +12,7 @@ import (
 	"math"
 	"time"
 
-	"github.com/ollama/ollama/x/imagegen"
+	"github.com/ollama/ollama/x/imagegen/manifest"
 	"github.com/ollama/ollama/x/imagegen/mlx"
 	"github.com/ollama/ollama/x/imagegen/models/qwen3"
 	"github.com/ollama/ollama/x/imagegen/tokenizer"
@@ -61,7 +61,7 @@ func (m *Model) Load(modelName string) error {
 	m.ModelName = modelName
 
 	// Load manifest
-	manifest, err := imagegen.LoadManifest(modelName)
+	manifest, err := manifest.LoadManifest(modelName)
 	if err != nil {
 		return fmt.Errorf("load manifest: %w", err)
 	}

x/imagegen/models/flux2/transformer.go

@@ -6,7 +6,7 @@ import (
 	"fmt"
 	"math"
 
-	"github.com/ollama/ollama/x/imagegen"
+	"github.com/ollama/ollama/x/imagegen/manifest"
 	"github.com/ollama/ollama/x/imagegen/mlx"
 	"github.com/ollama/ollama/x/imagegen/nn"
 	"github.com/ollama/ollama/x/imagegen/safetensors"
@@ -14,19 +14,19 @@ import (
 
 // TransformerConfig holds Flux2 transformer configuration
 type TransformerConfig struct {
-	AttentionHeadDim         int32   `json:"attention_head_dim"`          // 128
-	AxesDimsRoPE             []int32 `json:"axes_dims_rope"`              // [32, 32, 32, 32]
-	Eps                      float32 `json:"eps"`                         // 1e-6
-	GuidanceEmbeds           bool    `json:"guidance_embeds"`             // false for Klein
-	InChannels               int32   `json:"in_channels"`                 // 128
-	JointAttentionDim        int32   `json:"joint_attention_dim"`         // 7680
-	MLPRatio                 float32 `json:"mlp_ratio"`                   // 3.0
-	NumAttentionHeads        int32   `json:"num_attention_heads"`         // 24
-	NumLayers                int32   `json:"num_layers"`                  // 5
-	NumSingleLayers          int32   `json:"num_single_layers"`           // 20
-	PatchSize                int32   `json:"patch_size"`                  // 1
-	RopeTheta                int32   `json:"rope_theta"`                  // 2000
-	TimestepGuidanceChannels int32   `json:"timestep_guidance_channels"`  // 256
+	AttentionHeadDim         int32   `json:"attention_head_dim"`         // 128
+	AxesDimsRoPE             []int32 `json:"axes_dims_rope"`             // [32, 32, 32, 32]
+	Eps                      float32 `json:"eps"`                        // 1e-6
+	GuidanceEmbeds           bool    `json:"guidance_embeds"`            // false for Klein
+	InChannels               int32   `json:"in_channels"`                // 128
+	JointAttentionDim        int32   `json:"joint_attention_dim"`        // 7680
+	MLPRatio                 float32 `json:"mlp_ratio"`                  // 3.0
+	NumAttentionHeads        int32   `json:"num_attention_heads"`        // 24
+	NumLayers                int32   `json:"num_layers"`                 // 5
+	NumSingleLayers          int32   `json:"num_single_layers"`          // 20
+	PatchSize                int32   `json:"patch_size"`                 // 1
+	RopeTheta                int32   `json:"rope_theta"`                 // 2000
+	TimestepGuidanceChannels int32   `json:"timestep_guidance_channels"` // 256
 }
 
 // Computed dimensions
@@ -392,12 +392,12 @@ type Flux2Transformer2DModel struct {
 }
 
 // Load loads the Flux2 transformer from ollama blob storage.
-func (m *Flux2Transformer2DModel) Load(manifest *imagegen.ModelManifest) error {
+func (m *Flux2Transformer2DModel) Load(modelManifest *manifest.ModelManifest) error {
 	fmt.Print("  Loading transformer... ")
 
 	// Load config from blob
 	var cfg TransformerConfig
-	if err := manifest.ReadConfigJSON("transformer/config.json", &cfg); err != nil {
+	if err := modelManifest.ReadConfigJSON("transformer/config.json", &cfg); err != nil {
 		return fmt.Errorf("config: %w", err)
 	}
 	m.TransformerConfig = &cfg
@@ -412,7 +412,7 @@ func (m *Flux2Transformer2DModel) Load(manifest *imagegen.ModelManifest) error {
 	}
 
 	// Load weights from tensor blobs
-	weights, err := imagegen.LoadWeightsFromManifest(manifest, "transformer")
+	weights, err := manifest.LoadWeightsFromManifest(modelManifest, "transformer")
 	if err != nil {
 		return fmt.Errorf("weights: %w", err)
 	}

x/imagegen/models/flux2/vae.go

@@ -6,7 +6,7 @@ import (
 	"fmt"
 	"math"
 
-	"github.com/ollama/ollama/x/imagegen"
+	"github.com/ollama/ollama/x/imagegen/manifest"
 	"github.com/ollama/ollama/x/imagegen/mlx"
 	"github.com/ollama/ollama/x/imagegen/nn"
 	"github.com/ollama/ollama/x/imagegen/safetensors"
@@ -15,21 +15,21 @@ import (
 
 // VAEConfig holds AutoencoderKLFlux2 configuration
 type VAEConfig struct {
-	ActFn             string  `json:"act_fn"`              // "silu"
-	BatchNormEps      float32 `json:"batch_norm_eps"`      // 0.0001
-	BatchNormMomentum float32 `json:"batch_norm_momentum"` // 0.1
-	BlockOutChannels  []int32 `json:"block_out_channels"`  // [128, 256, 512, 512]
-	ForceUpcast       bool    `json:"force_upcast"`        // true
-	InChannels        int32   `json:"in_channels"`         // 3
-	LatentChannels    int32   `json:"latent_channels"`     // 32
-	LayersPerBlock    int32   `json:"layers_per_block"`    // 2
+	ActFn             string  `json:"act_fn"`                  // "silu"
+	BatchNormEps      float32 `json:"batch_norm_eps"`          // 0.0001
+	BatchNormMomentum float32 `json:"batch_norm_momentum"`     // 0.1
+	BlockOutChannels  []int32 `json:"block_out_channels"`      // [128, 256, 512, 512]
+	ForceUpcast       bool    `json:"force_upcast"`            // true
+	InChannels        int32   `json:"in_channels"`             // 3
+	LatentChannels    int32   `json:"latent_channels"`         // 32
+	LayersPerBlock    int32   `json:"layers_per_block"`        // 2
 	MidBlockAddAttn   bool    `json:"mid_block_add_attention"` // true
-	NormNumGroups     int32   `json:"norm_num_groups"`     // 32
-	OutChannels       int32   `json:"out_channels"`        // 3
-	PatchSize         []int32 `json:"patch_size"`          // [2, 2]
-	SampleSize        int32   `json:"sample_size"`         // 1024
-	UsePostQuantConv  bool    `json:"use_post_quant_conv"` // true
-	UseQuantConv      bool    `json:"use_quant_conv"`      // true
+	NormNumGroups     int32   `json:"norm_num_groups"`         // 32
+	OutChannels       int32   `json:"out_channels"`            // 3
+	PatchSize         []int32 `json:"patch_size"`              // [2, 2]
+	SampleSize        int32   `json:"sample_size"`             // 1024
+	UsePostQuantConv  bool    `json:"use_post_quant_conv"`     // true
+	UseQuantConv      bool    `json:"use_quant_conv"`          // true
 }
 
 // BatchNorm2D implements 2D batch normalization with running statistics
@@ -356,18 +356,18 @@ func (db *DownEncoderBlock2D) Forward(x *mlx.Array) *mlx.Array {
 }
 
 // Load loads the Flux2 VAE from ollama blob storage.
-func (m *AutoencoderKLFlux2) Load(manifest *imagegen.ModelManifest) error {
+func (m *AutoencoderKLFlux2) Load(modelManifest *manifest.ModelManifest) error {
 	fmt.Print("  Loading VAE... ")
 
 	// Load config from blob
 	var cfg VAEConfig
-	if err := manifest.ReadConfigJSON("vae/config.json", &cfg); err != nil {
+	if err := modelManifest.ReadConfigJSON("vae/config.json", &cfg); err != nil {
 		return fmt.Errorf("config: %w", err)
 	}
 	m.Config = &cfg
 
 	// Load weights from tensor blobs
-	weights, err := imagegen.LoadWeightsFromManifest(manifest, "vae")
+	weights, err := manifest.LoadWeightsFromManifest(modelManifest, "vae")
 	if err != nil {
 		return fmt.Errorf("weights: %w", err)
 	}

x/imagegen/models/glm4_moe_lite/glm4_moe_lite.go

@@ -9,8 +9,8 @@ import (
 	"fmt"
 	"math"
 
-	"github.com/ollama/ollama/x/imagegen"
 	"github.com/ollama/ollama/x/imagegen/cache"
+	"github.com/ollama/ollama/x/imagegen/manifest"
 	"github.com/ollama/ollama/x/imagegen/mlx"
 	"github.com/ollama/ollama/x/imagegen/nn"
 	"github.com/ollama/ollama/x/imagegen/safetensors"
@@ -38,11 +38,11 @@ type Config struct {
 	AttentionBias         bool    `json:"attention_bias"`
 
 	// MLA (Multi-head Latent Attention) parameters
-	QLoraRank      int32 `json:"q_lora_rank"`
-	KVLoraRank     int32 `json:"kv_lora_rank"`
-	QKRopeHeadDim  int32 `json:"qk_rope_head_dim"`
-	QKNopeHeadDim  int32 `json:"qk_nope_head_dim"`
-	VHeadDim       int32 `json:"v_head_dim"`
+	QLoraRank     int32 `json:"q_lora_rank"`
+	KVLoraRank    int32 `json:"kv_lora_rank"`
+	QKRopeHeadDim int32 `json:"qk_rope_head_dim"`
+	QKNopeHeadDim int32 `json:"qk_nope_head_dim"`
+	VHeadDim      int32 `json:"v_head_dim"`
 
 	// MoE parameters
 	NRoutedExperts      int32   `json:"n_routed_experts"`
@@ -82,7 +82,7 @@ type MLAAttention struct {
 	// Absorbed MLA projections (derived from kv_b_proj)
 	// EmbedQ: projects q_nope to latent space [num_heads, kv_lora_rank, qk_nope_head_dim]
 	// UnembedOut: projects attention output from latent space [num_heads, v_head_dim, kv_lora_rank]
-	EmbedQ    *nn.MultiLinear `weight:"-"`
+	EmbedQ     *nn.MultiLinear `weight:"-"`
 	UnembedOut *nn.MultiLinear `weight:"-"`
 
 	// Output projection
@@ -194,8 +194,8 @@ func (m *DenseMLP) Forward(x *mlx.Array) *mlx.Array {
 
 // MoEGate implements the expert gating mechanism
 type MoEGate struct {
-	Gate                   nn.LinearLayer `weight:"mlp.gate"`
-	EScoreCorrectionBias   *mlx.Array     `weight:"mlp.gate.e_score_correction_bias,optional"`
+	Gate                 nn.LinearLayer `weight:"mlp.gate"`
+	EScoreCorrectionBias *mlx.Array     `weight:"mlp.gate.e_score_correction_bias,optional"`
 }
 
 // Forward computes expert selection indices and scores
@@ -617,9 +617,9 @@ func sanitizeExpertWeights(weights safetensors.WeightSource, prefix string, numE
 }
 
 // LoadFromManifest loads a GLM4-MoE-Lite model from a manifest (Ollama blob storage).
-func LoadFromManifest(manifest *imagegen.ModelManifest) (*Model, error) {
+func LoadFromManifest(modelManifest *manifest.ModelManifest) (*Model, error) {
 	// Read config from manifest
-	configData, err := manifest.ReadConfig("config.json")
+	configData, err := modelManifest.ReadConfig("config.json")
 	if err != nil {
 		return nil, fmt.Errorf("load config: %w", err)
 	}
@@ -634,7 +634,7 @@ func LoadFromManifest(manifest *imagegen.ModelManifest) (*Model, error) {
 	cfg.Scale = computeScale(&cfg)
 
 	// Load weights from manifest blobs
-	weights, err := imagegen.LoadWeightsFromManifest(manifest, "")
+	weights, err := manifest.LoadWeightsFromManifest(modelManifest, "")
 	if err != nil {
 		return nil, fmt.Errorf("load weights: %w", err)
 	}
@@ -653,7 +653,7 @@ func LoadFromManifest(manifest *imagegen.ModelManifest) (*Model, error) {
 	}
 
 	// Load tokenizer from manifest with config files for EOS token detection
-	tokData, err := manifest.ReadConfig("tokenizer.json")
+	tokData, err := modelManifest.ReadConfig("tokenizer.json")
 	if err != nil {
 		return nil, fmt.Errorf("load tokenizer config: %w", err)
 	}
@@ -664,12 +664,12 @@ func LoadFromManifest(manifest *imagegen.ModelManifest) (*Model, error) {
 	}
 
 	// Try to load generation_config.json if available (preferred source for EOS)
-	if genConfigData, err := manifest.ReadConfig("generation_config.json"); err == nil {
+	if genConfigData, err := modelManifest.ReadConfig("generation_config.json"); err == nil {
 		tokConfig.GenerationConfigJSON = genConfigData
 	}
 
 	// Try to load tokenizer_config.json if available
-	if tokConfigData, err := manifest.ReadConfig("tokenizer_config.json"); err == nil {
+	if tokConfigData, err := modelManifest.ReadConfig("tokenizer_config.json"); err == nil {
 		tokConfig.TokenizerConfigJSON = tokConfigData
 	}
 

x/imagegen/models/qwen3/text_encoder.go

@@ -7,7 +7,7 @@ import (
 	"fmt"
 	"math"
 
-	"github.com/ollama/ollama/x/imagegen"
+	"github.com/ollama/ollama/x/imagegen/manifest"
 	"github.com/ollama/ollama/x/imagegen/mlx"
 	"github.com/ollama/ollama/x/imagegen/nn"
 	"github.com/ollama/ollama/x/imagegen/safetensors"
@@ -181,19 +181,19 @@ type TextEncoder struct {
 }
 
 // Load loads the Qwen3 text encoder from ollama blob storage.
-func (m *TextEncoder) Load(manifest *imagegen.ModelManifest, configPath string) error {
+func (m *TextEncoder) Load(modelManifest *manifest.ModelManifest, configPath string) error {
 	fmt.Print("  Loading text encoder... ")
 
 	// Load config from blob
 	var cfg Config
-	if err := manifest.ReadConfigJSON(configPath, &cfg); err != nil {
+	if err := modelManifest.ReadConfigJSON(configPath, &cfg); err != nil {
 		return fmt.Errorf("config: %w", err)
 	}
 	m.Config = &cfg
 	m.Layers = make([]*Block, cfg.NumHiddenLayers)
 
 	// Load weights from tensor blobs
-	weights, err := imagegen.LoadWeightsFromManifest(manifest, "text_encoder")
+	weights, err := manifest.LoadWeightsFromManifest(modelManifest, "text_encoder")
 	if err != nil {
 		return fmt.Errorf("weights: %w", err)
 	}

x/imagegen/models/zimage/transformer.go

@@ -7,8 +7,8 @@ import (
 	"fmt"
 	"math"
 
-	"github.com/ollama/ollama/x/imagegen"
 	"github.com/ollama/ollama/x/imagegen/cache"
+	"github.com/ollama/ollama/x/imagegen/manifest"
 	"github.com/ollama/ollama/x/imagegen/mlx"
 	"github.com/ollama/ollama/x/imagegen/nn"
 	"github.com/ollama/ollama/x/imagegen/safetensors"
@@ -38,7 +38,7 @@ type TransformerConfig struct {
 type TimestepEmbedder struct {
 	Linear1       nn.LinearLayer `weight:"mlp.0"`
 	Linear2       nn.LinearLayer `weight:"mlp.2"`
-	FreqEmbedSize int32      // 256 (computed)
+	FreqEmbedSize int32          // 256 (computed)
 }
 
 // Forward computes timestep embeddings -> [B, 256]
@@ -85,9 +85,9 @@ func (xe *XEmbedder) Forward(x *mlx.Array) *mlx.Array {
 
 // CapEmbedder projects caption features to model dimension
 type CapEmbedder struct {
-	Norm     *nn.RMSNorm `weight:"0"`
-	Linear   nn.LinearLayer  `weight:"1"`
-	PadToken *mlx.Array  // loaded separately at root level
+	Norm     *nn.RMSNorm    `weight:"0"`
+	Linear   nn.LinearLayer `weight:"1"`
+	PadToken *mlx.Array     // loaded separately at root level
 }
 
 // Forward projects caption embeddings: [B, L, cap_feat_dim] -> [B, L, dim]
@@ -103,10 +103,9 @@ type FeedForward struct {
 	W1     nn.LinearLayer `weight:"w1"` // gate projection
 	W2     nn.LinearLayer `weight:"w2"` // down projection
 	W3     nn.LinearLayer `weight:"w3"` // up projection
-	OutDim int32      // computed from W2
+	OutDim int32          // computed from W2
 }
 
-
 // Forward applies SwiGLU: silu(W1(x)) * W3(x), then W2
 func (ff *FeedForward) Forward(x *mlx.Array) *mlx.Array {
 	shape := x.Shape()
@@ -132,11 +131,11 @@ type Attention struct {
 	ToK   nn.LinearLayer `weight:"to_k"`
 	ToV   nn.LinearLayer `weight:"to_v"`
 	ToOut nn.LinearLayer `weight:"to_out.0"`
-	NormQ *mlx.Array `weight:"norm_q.weight"` // [head_dim] for per-head RMSNorm
-	NormK *mlx.Array `weight:"norm_k.weight"`
+	NormQ *mlx.Array     `weight:"norm_q.weight"` // [head_dim] for per-head RMSNorm
+	NormK *mlx.Array     `weight:"norm_k.weight"`
 	// Fused QKV (computed at init time for efficiency, not loaded from weights)
 	ToQKV nn.LinearLayer `weight:"-"` // Fused Q+K+V projection (created by FuseQKV)
-	Fused bool       `weight:"-"` // Whether to use fused QKV path
+	Fused bool           `weight:"-"` // Whether to use fused QKV path
 	// Computed fields (not loaded from weights)
 	NHeads  int32   `weight:"-"`
 	HeadDim int32   `weight:"-"`
@@ -288,13 +287,13 @@ func applyRoPE3D(x *mlx.Array, cos, sin *mlx.Array) *mlx.Array {
 
 // TransformerBlock is a single transformer block with optional AdaLN modulation
 type TransformerBlock struct {
-	Attention      *Attention   `weight:"attention"`
-	FeedForward    *FeedForward `weight:"feed_forward"`
-	AttentionNorm1 *nn.RMSNorm  `weight:"attention_norm1"`
-	AttentionNorm2 *nn.RMSNorm  `weight:"attention_norm2"`
-	FFNNorm1       *nn.RMSNorm  `weight:"ffn_norm1"`
-	FFNNorm2       *nn.RMSNorm  `weight:"ffn_norm2"`
-	AdaLN          nn.LinearLayer   `weight:"adaLN_modulation.0,optional"` // only if modulation
+	Attention      *Attention     `weight:"attention"`
+	FeedForward    *FeedForward   `weight:"feed_forward"`
+	AttentionNorm1 *nn.RMSNorm    `weight:"attention_norm1"`
+	AttentionNorm2 *nn.RMSNorm    `weight:"attention_norm2"`
+	FFNNorm1       *nn.RMSNorm    `weight:"ffn_norm1"`
+	FFNNorm2       *nn.RMSNorm    `weight:"ffn_norm2"`
+	AdaLN          nn.LinearLayer `weight:"adaLN_modulation.0,optional"` // only if modulation
 	// Computed fields
 	HasModulation bool
 	Dim           int32
@@ -350,7 +349,7 @@ func (tb *TransformerBlock) Forward(x *mlx.Array, adaln *mlx.Array, cos, sin *ml
 type FinalLayer struct {
 	AdaLN  nn.LinearLayer `weight:"adaLN_modulation.1"` // [256] -> [dim]
 	Output nn.LinearLayer `weight:"linear"`             // [dim] -> [out_channels]
-	OutDim int32      // computed from Output
+	OutDim int32          // computed from Output
 }
 
 // Forward computes final output
@@ -401,12 +400,12 @@ type Transformer struct {
 }
 
 // Load loads the Z-Image transformer from ollama blob storage.
-func (m *Transformer) Load(manifest *imagegen.ModelManifest) error {
+func (m *Transformer) Load(modelManifest *manifest.ModelManifest) error {
 	fmt.Print("  Loading transformer... ")
 
 	// Load config from blob
 	var cfg TransformerConfig
-	if err := manifest.ReadConfigJSON("transformer/config.json", &cfg); err != nil {
+	if err := modelManifest.ReadConfigJSON("transformer/config.json", &cfg); err != nil {
 		return fmt.Errorf("config: %w", err)
 	}
 	if len(cfg.AllPatchSize) > 0 {
@@ -417,7 +416,7 @@ func (m *Transformer) Load(manifest *imagegen.ModelManifest) error {
 	m.ContextRefiners = make([]*TransformerBlock, cfg.NRefinerLayers)
 	m.Layers = make([]*TransformerBlock, cfg.NLayers)
 
-	weights, err := imagegen.LoadWeightsFromManifest(manifest, "transformer")
+	weights, err := manifest.LoadWeightsFromManifest(modelManifest, "transformer")
 	if err != nil {
 		return fmt.Errorf("weights: %w", err)
 	}

x/imagegen/models/zimage/vae.go

@@ -6,7 +6,7 @@ import (
 	"fmt"
 	"math"
 
-	"github.com/ollama/ollama/x/imagegen"
+	"github.com/ollama/ollama/x/imagegen/manifest"
 	"github.com/ollama/ollama/x/imagegen/mlx"
 	"github.com/ollama/ollama/x/imagegen/safetensors"
 	"github.com/ollama/ollama/x/imagegen/vae"
@@ -562,7 +562,7 @@ func (ub *UpDecoderBlock2D) Forward(x *mlx.Array) *mlx.Array {
 	if ub.Upsample != nil {
 		// Stage 1: Upsample2x (nearest neighbor)
 		{
-					prev := x
+			prev := x
 			x = Upsample2x(x)
 			prev.Free()
 			mlx.Eval(x)
@@ -570,7 +570,7 @@ func (ub *UpDecoderBlock2D) Forward(x *mlx.Array) *mlx.Array {
 
 		// Stage 2: Upsample conv
 		{
-					prev := x
+			prev := x
 			x = ub.Upsample.Forward(x)
 			prev.Free()
 			mlx.Eval(x)
@@ -643,16 +643,16 @@ type VAEDecoder struct {
 }
 
 // Load loads the VAE decoder from ollama blob storage.
-func (m *VAEDecoder) Load(manifest *imagegen.ModelManifest) error {
+func (m *VAEDecoder) Load(modelManifest *manifest.ModelManifest) error {
 	// Load config from blob
 	var cfg VAEConfig
-	if err := manifest.ReadConfigJSON("vae/config.json", &cfg); err != nil {
+	if err := modelManifest.ReadConfigJSON("vae/config.json", &cfg); err != nil {
 		return fmt.Errorf("config: %w", err)
 	}
 	m.Config = &cfg
 
 	// Load weights from tensor blobs
-	weights, err := imagegen.LoadWeightsFromManifest(manifest, "vae")
+	weights, err := manifest.LoadWeightsFromManifest(modelManifest, "vae")
 	if err != nil {
 		return fmt.Errorf("weights: %w", err)
 	}

x/imagegen/models/zimage/zimage.go

@@ -8,8 +8,8 @@ import (
 	"fmt"
 	"time"
 
-	"github.com/ollama/ollama/x/imagegen"
 	"github.com/ollama/ollama/x/imagegen/cache"
+	"github.com/ollama/ollama/x/imagegen/manifest"
 	"github.com/ollama/ollama/x/imagegen/mlx"
 	"github.com/ollama/ollama/x/imagegen/tokenizer"
 	"github.com/ollama/ollama/x/imagegen/vae"
@@ -18,14 +18,14 @@ import (
 // GenerateConfig holds all options for image generation.
 type GenerateConfig struct {
 	Prompt         string
-	NegativePrompt string                // Empty = no CFG
-	CFGScale       float32               // Only used if NegativePrompt is set (default: 4.0)
-	Width          int32                 // Image width (default: 1024)
-	Height         int32                 // Image height (default: 1024)
-	Steps          int                   // Denoising steps (default: 9 for turbo)
-	Seed           int64                 // Random seed
+	NegativePrompt string                     // Empty = no CFG
+	CFGScale       float32                    // Only used if NegativePrompt is set (default: 4.0)
+	Width          int32                      // Image width (default: 1024)
+	Height         int32                      // Image height (default: 1024)
+	Steps          int                        // Denoising steps (default: 9 for turbo)
+	Seed           int64                      // Random seed
 	Progress       func(step, totalSteps int) // Optional progress callback
-	CapturePath    string                // GPU capture path (debug)
+	CapturePath    string                     // GPU capture path (debug)
 
 	// TeaCache options (timestep embedding aware caching)
 	TeaCache          bool    // TeaCache is always enabled for faster inference
@@ -58,7 +58,7 @@ func (m *Model) Load(modelName string) error {
 	m.ModelName = modelName
 
 	// Load manifest
-	manifest, err := imagegen.LoadManifest(modelName)
+	manifest, err := manifest.LoadManifest(modelName)
 	if err != nil {
 		return fmt.Errorf("load manifest: %w", err)
 	}

x/imagegen/runner.go

@@ -1,7 +1,7 @@
 //go:build mlx
 
-// Package mlxrunner provides a unified MLX runner for both LLM and image generation models.
-package mlxrunner
+// Package imagegen provides a unified MLX runner for both LLM and image generation models.
+package imagegen
 
 import (
 	"context"
@@ -16,7 +16,6 @@ import (
 	"time"
 
 	"github.com/ollama/ollama/envconfig"
-	"github.com/ollama/ollama/x/imagegen"
 	"github.com/ollama/ollama/x/imagegen/mlx"
 )
 
@@ -98,7 +97,7 @@ func Execute(args []string) error {
 // detectModelMode determines whether a model is an LLM or image generation model.
 func detectModelMode(modelName string) ModelMode {
 	// Check for image generation model by looking at model_index.json
-	modelType := imagegen.DetectModelType(modelName)
+	modelType := DetectModelType(modelName)
 	if modelType != "" {
 		// Known image generation model types
 		switch modelType {

x/imagegen/runner_stub.go

@@ -1,6 +1,6 @@
 //go:build !mlx
 
-package mlxrunner
+package imagegen
 
 import "errors"
 

x/imagegen/server.go

@@ -1,4 +1,4 @@
-package mlxrunner
+package imagegen
 
 import (
 	"bufio"
@@ -23,7 +23,7 @@ import (
 
 	"github.com/ollama/ollama/llm"
 	"github.com/ollama/ollama/ml"
-	"github.com/ollama/ollama/x/imagegen"
+	"github.com/ollama/ollama/x/imagegen/manifest"
 )
 
 // Server wraps an MLX runner subprocess to implement llm.LlamaServer.
@@ -46,7 +46,7 @@ type Server struct {
 // NewServer spawns a new MLX runner subprocess and waits until it's ready.
 func NewServer(modelName string, mode ModelMode) (*Server, error) {
 	// Validate platform support before attempting to start
-	if err := imagegen.CheckPlatformSupport(); err != nil {
+	if err := CheckPlatformSupport(); err != nil {
 		return nil, err
 	}
 
@@ -71,8 +71,8 @@ func NewServer(modelName string, mode ModelMode) (*Server, error) {
 		exe = eval
 	}
 
-	// Spawn subprocess: ollama runner --mlx-engine --model <path> --port <port>
-	cmd := exec.Command(exe, "runner", "--mlx-engine", "--model", modelName, "--port", strconv.Itoa(port))
+	// Spawn subprocess: ollama runner --imagegen-engine --model <path> --port <port>
+	cmd := exec.Command(exe, "runner", "--imagegen-engine", "--model", modelName, "--port", strconv.Itoa(port))
 	cmd.Env = os.Environ()
 
 	// On Linux, set LD_LIBRARY_PATH to include MLX library directories
@@ -107,8 +107,8 @@ func NewServer(modelName string, mode ModelMode) (*Server, error) {
 
 	// Estimate VRAM based on tensor size from manifest
 	var vramSize uint64
-	if manifest, err := imagegen.LoadManifest(modelName); err == nil {
-		vramSize = uint64(manifest.TotalTensorSize())
+	if modelManifest, err := manifest.LoadManifest(modelName); err == nil {
+		vramSize = uint64(modelManifest.TotalTensorSize())
 	} else {
 		// Fallback: default to 8GB if manifest can't be loaded
 		vramSize = 8 * 1024 * 1024 * 1024

x/imagegen/types.go

@@ -1,9 +1,9 @@
-// Package mlxrunner provides a unified MLX runner for both LLM and image generation models.
+// Package imagegen provides a unified MLX runner for both LLM and image generation models.
 //
 // This package handles safetensors models created with `ollama create --experimental`,
 // supporting both text generation (LLM) and image generation (diffusion) models
 // through a single unified interface.
-package mlxrunner
+package imagegen
 
 // Request is the request format for completion requests.
 type Request struct {