launch: use bundled OpenClaw ollama web search (#15757)

---------

Co-authored-by: Parth Sareen <parth.sareen@ollama.com>

app/ui/app/src/hooks/useChats.ts

@@ -381,7 +381,7 @@ export const useSendMessage = (chatId: string) => {
                     role: "assistant",
                     content: "",
                     thinking: "",
-                    model: effectiveModel,
+                    model: effectiveModel.model,
                   }),
                 );
                 lastMessage = newMessages[newMessages.length - 1];
@@ -433,7 +433,7 @@ export const useSendMessage = (chatId: string) => {
                     role: "assistant",
                     content: "",
                     thinking: "",
-                    model: effectiveModel,
+                    model: effectiveModel.model,
                   }),
                 );
                 lastMessage = newMessages[newMessages.length - 1];
@@ -520,7 +520,7 @@ export const useSendMessage = (chatId: string) => {
                     thinkingTimeStart:
                       lastMessage.thinkingTimeStart || event.thinkingTimeStart,
                     thinkingTimeEnd: event.thinkingTimeEnd,
-                    model: selectedModel,
+                    model: selectedModel.model,
                   });
                   newMessages[newMessages.length - 1] = updatedMessage;
                 } else {
@@ -533,7 +533,7 @@ export const useSendMessage = (chatId: string) => {
                       tool_calls: event.toolCalls,
                       thinkingTimeStart: event.thinkingTimeStart,
                       thinkingTimeEnd: event.thinkingTimeEnd,
-                      model: selectedModel,
+                      model: selectedModel.model,
                     }),
                   );
                 }
@@ -699,7 +699,7 @@ export const useSendMessage = (chatId: string) => {
             queryClient.setQueryData(["chat", newId], {
               chat: new Chat({
                 id: newId,
-                model: effectiveModel,
+                model: effectiveModel.model,
                 messages: [
                   new Message({
                     role: "user",

cmd/cmd.go

@@ -1975,8 +1975,61 @@ func launchInteractiveModel(cmd *cobra.Command, modelName string) error {
 		Options:     map[string]any{},
 		ShowConnect: true,
 	}
-	// loadOrUnloadModel is cloud-safe here: remote/cloud models skip local preload
-	// and only validate auth/connectivity before interactive chat starts.
+
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		return err
+	}
+
+	requestedCloud := modelref.HasExplicitCloudSource(modelName)
+
+	info, err := func() (*api.ShowResponse, error) {
+		showReq := &api.ShowRequest{Name: modelName}
+		info, err := client.Show(cmd.Context(), showReq)
+		var se api.StatusError
+		if errors.As(err, &se) && se.StatusCode == http.StatusNotFound {
+			if requestedCloud {
+				return nil, err
+			}
+			if err := PullHandler(cmd, []string{modelName}); err != nil {
+				return nil, err
+			}
+			return client.Show(cmd.Context(), &api.ShowRequest{Name: modelName})
+		}
+		return info, err
+	}()
+	if err != nil {
+		if handleCloudAuthorizationError(err) {
+			return nil
+		}
+		return err
+	}
+
+	ensureCloudStub(cmd.Context(), client, modelName)
+
+	opts.Think, err = inferThinkingOption(&info.Capabilities, &opts, false)
+	if err != nil {
+		return err
+	}
+
+	audioCapable := slices.Contains(info.Capabilities, model.CapabilityAudio)
+	opts.MultiModal = slices.Contains(info.Capabilities, model.CapabilityVision) || audioCapable
+
+	// TODO: remove the projector info and vision info checks below,
+	// these are left in for backwards compatibility with older servers
+	// that don't have the capabilities field in the model info
+	if len(info.ProjectorInfo) != 0 {
+		opts.MultiModal = true
+	}
+	for k := range info.ModelInfo {
+		if strings.Contains(k, ".vision.") {
+			opts.MultiModal = true
+			break
+		}
+	}
+
+	applyShowResponseToRunOptions(&opts, info)
+
 	if err := loadOrUnloadModel(cmd, &opts); err != nil {
 		return fmt.Errorf("error loading model: %w", err)
 	}

cmd/launch/integrations_test.go

@@ -301,7 +301,7 @@ func TestParseArgs(t *testing.T) {
 func TestIsCloudModel(t *testing.T) {
 	// 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-5.1:cloud", "kimi-k2.5:cloud", "local-model"}
+		models := []string{"glm-5.1:cloud", "kimi-k2.6:cloud", "local-model"}
 		for _, model := range models {
 			if isCloudModel(context.Background(), nil, model) {
 				t.Errorf("isCloudModel(%q) with nil client should return false", model)
@@ -321,7 +321,7 @@ func names(items []ModelItem) []string {
 func TestBuildModelList_NoExistingModels(t *testing.T) {
 	items, _, _, _ := buildModelList(nil, nil, "")
 
-	want := []string{"kimi-k2.5:cloud", "qwen3.5:cloud", "glm-5.1:cloud", "minimax-m2.7:cloud", "gemma4", "qwen3.5"}
+	want := []string{"kimi-k2.6:cloud", "qwen3.5:cloud", "glm-5.1:cloud", "minimax-m2.7:cloud", "gemma4", "qwen3.5"}
 	if diff := cmp.Diff(want, names(items)); diff != "" {
 		t.Errorf("with no existing models, items should be recommended in order (-want +got):\n%s", diff)
 	}
@@ -350,7 +350,7 @@ func TestBuildModelList_OnlyLocalModels_CloudRecsStillFirst(t *testing.T) {
 
 	// Cloud recs always come first among recommended, regardless of installed inventory.
 	// Cloud disablement is handled upstream in loadSelectableModels via filterCloudItems.
-	want := []string{"kimi-k2.5:cloud", "qwen3.5:cloud", "glm-5.1:cloud", "minimax-m2.7:cloud", "gemma4", "qwen3.5", "llama3.2", "qwen2.5"}
+	want := []string{"kimi-k2.6:cloud", "qwen3.5:cloud", "glm-5.1:cloud", "minimax-m2.7:cloud", "gemma4", "qwen3.5", "llama3.2", "qwen2.5"}
 	if diff := cmp.Diff(want, got); diff != "" {
 		t.Errorf("cloud recs pinned first even when no cloud models installed (-want +got):\n%s", diff)
 	}
@@ -366,7 +366,7 @@ func TestBuildModelList_BothCloudAndLocal_RegularSort(t *testing.T) {
 	got := names(items)
 
 	// All recs pinned at top (cloud before local in mixed case), then non-recs
-	want := []string{"kimi-k2.5:cloud", "qwen3.5:cloud", "glm-5.1:cloud", "minimax-m2.7:cloud", "gemma4", "qwen3.5", "llama3.2"}
+	want := []string{"kimi-k2.6:cloud", "qwen3.5:cloud", "glm-5.1:cloud", "minimax-m2.7:cloud", "gemma4", "qwen3.5", "llama3.2"}
 	if diff := cmp.Diff(want, got); diff != "" {
 		t.Errorf("recs pinned at top, cloud recs first in mixed case (-want +got):\n%s", diff)
 	}
@@ -437,7 +437,7 @@ func TestBuildModelList_ExistingRecommendedMarked(t *testing.T) {
 			if !strings.HasSuffix(item.Description, "(not downloaded)") {
 				t.Errorf("non-installed recommended %q should have '(not downloaded)' suffix, got %q", item.Name, item.Description)
 			}
-		case "minimax-m2.7:cloud", "kimi-k2.5:cloud", "qwen3.5:cloud":
+		case "minimax-m2.7:cloud", "kimi-k2.6:cloud", "qwen3.5:cloud":
 			if strings.HasSuffix(item.Description, "(not downloaded)") {
 				t.Errorf("cloud model %q should not have '(not downloaded)' suffix, got %q", item.Name, item.Description)
 			}
@@ -455,9 +455,9 @@ func TestBuildModelList_ExistingCloudModelsNotPushedToBottom(t *testing.T) {
 	got := names(items)
 
 	// gemma4 and glm-5.1:cloud are installed so they sort normally;
-	// kimi-k2.5:cloud, qwen3.5:cloud, and qwen3.5 are not installed so they go to the bottom
+	// qwen3.5:cloud and qwen3.5 are not installed so they go to the bottom
 	// All recs: cloud first in mixed case, then local, in rec order within each
-	want := []string{"kimi-k2.5:cloud", "qwen3.5:cloud", "glm-5.1:cloud", "minimax-m2.7:cloud", "gemma4", "qwen3.5"}
+	want := []string{"kimi-k2.6:cloud", "qwen3.5:cloud", "glm-5.1:cloud", "minimax-m2.7:cloud", "gemma4", "qwen3.5"}
 	if diff := cmp.Diff(want, got); diff != "" {
 		t.Errorf("all recs, cloud first in mixed case (-want +got):\n%s", diff)
 	}
@@ -466,23 +466,23 @@ func TestBuildModelList_ExistingCloudModelsNotPushedToBottom(t *testing.T) {
 func TestBuildModelList_HasRecommendedCloudModel_OnlyNonInstalledAtBottom(t *testing.T) {
 	existing := []modelInfo{
 		{Name: "llama3.2:latest", Remote: false},
-		{Name: "kimi-k2.5:cloud", Remote: true},
+		{Name: "kimi-k2.6:cloud", Remote: true},
 	}
 
 	items, _, _, _ := buildModelList(existing, nil, "")
 	got := names(items)
 
-	// kimi-k2.5:cloud is installed so it sorts normally;
+	// kimi-k2.6:cloud is installed so it sorts normally;
 	// the rest of the recommendations are not installed so they go to the bottom
 	// All recs pinned at top (cloud first in mixed case), then non-recs
-	want := []string{"kimi-k2.5:cloud", "qwen3.5:cloud", "glm-5.1:cloud", "minimax-m2.7:cloud", "gemma4", "qwen3.5", "llama3.2"}
+	want := []string{"kimi-k2.6:cloud", "qwen3.5:cloud", "glm-5.1:cloud", "minimax-m2.7:cloud", "gemma4", "qwen3.5", "llama3.2"}
 	if diff := cmp.Diff(want, got); diff != "" {
 		t.Errorf("recs pinned at top, cloud first in mixed case (-want +got):\n%s", diff)
 	}
 
 	for _, item := range items {
 		isCloud := strings.HasSuffix(item.Name, ":cloud")
-		isInstalled := slices.Contains([]string{"kimi-k2.5:cloud", "llama3.2"}, item.Name)
+		isInstalled := slices.Contains([]string{"kimi-k2.6:cloud", "llama3.2"}, item.Name)
 		if isInstalled || isCloud {
 			if strings.HasSuffix(item.Description, "(not downloaded)") {
 				t.Errorf("installed or cloud model %q should not have '(not downloaded)' suffix, got %q", item.Name, item.Description)
@@ -549,8 +549,8 @@ func TestBuildModelList_ReturnsExistingAndCloudMaps(t *testing.T) {
 	if !cloudModels["glm-5.1:cloud"] {
 		t.Error("glm-5.1:cloud should be in cloudModels")
 	}
-	if !cloudModels["kimi-k2.5:cloud"] {
-		t.Error("kimi-k2.5:cloud should be in cloudModels (recommended cloud)")
+	if !cloudModels["kimi-k2.6:cloud"] {
+		t.Error("kimi-k2.6:cloud should be in cloudModels (recommended cloud)")
 	}
 	if !cloudModels["qwen3.5:cloud"] {
 		t.Error("qwen3.5:cloud should be in cloudModels (recommended cloud)")
@@ -570,7 +570,7 @@ func TestBuildModelList_RecommendedFieldSet(t *testing.T) {
 
 	for _, item := range items {
 		switch item.Name {
-		case "gemma4", "qwen3.5", "glm-5.1:cloud", "kimi-k2.5:cloud", "qwen3.5:cloud":
+		case "gemma4", "qwen3.5", "glm-5.1:cloud", "kimi-k2.6:cloud", "qwen3.5:cloud":
 			if !item.Recommended {
 				t.Errorf("%q should have Recommended=true", item.Name)
 			}
@@ -628,7 +628,7 @@ func TestBuildModelList_RecsAboveNonRecs(t *testing.T) {
 	lastRecIdx := -1
 	firstNonRecIdx := len(got)
 	for i, name := range got {
-		isRec := name == "gemma4" || name == "qwen3.5" || name == "minimax-m2.7:cloud" || name == "glm-5.1:cloud" || name == "kimi-k2.5:cloud" || name == "qwen3.5:cloud"
+		isRec := name == "gemma4" || name == "qwen3.5" || name == "minimax-m2.7:cloud" || name == "glm-5.1:cloud" || name == "kimi-k2.6:cloud" || name == "qwen3.5:cloud"
 		if isRec && i > lastRecIdx {
 			lastRecIdx = i
 		}

cmd/launch/models.go

@@ -21,7 +21,7 @@ import (
 )
 
 var recommendedModels = []ModelItem{
-	{Name: "kimi-k2.5:cloud", Description: "Multimodal reasoning with subagents", Recommended: true},
+	{Name: "kimi-k2.6:cloud", Description: "State-of-the-art coding, long-horizon execution, and multimodal agent swarm capability", Recommended: true},
 	{Name: "qwen3.5:cloud", Description: "Reasoning, coding, and agentic tool use with vision", Recommended: true},
 	{Name: "glm-5.1:cloud", Description: "Reasoning and code generation", Recommended: true},
 	{Name: "minimax-m2.7:cloud", Description: "Fast, efficient coding and real-world productivity", Recommended: true},
@@ -56,6 +56,7 @@ var cloudModelLimits = map[string]cloudModelLimit{
 	"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-k2.6":           {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},

cmd/launch/openclaw.go

@@ -14,8 +14,6 @@ import (
 	"strings"
 	"time"
 
-	"golang.org/x/mod/semver"
-
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/cmd/internal/fileutil"
 	"github.com/ollama/ollama/envconfig"
@@ -98,9 +96,7 @@ func (c *Openclaw) Run(model string, args []string) error {
 		patchDeviceScopes()
 	}
 
-	if ensureWebSearchPlugin() {
-		registerWebSearchPlugin()
-	}
+	configureOllamaWebSearch()
 
 	// When extra args are passed through, run exactly what the user asked for
 	// after setup and skip the built-in gateway+TUI convenience flow.
@@ -738,89 +734,13 @@ func clearSessionModelOverride(primary string) {
 	_ = os.WriteFile(path, out, 0o600)
 }
 
-const (
-	webSearchNpmPackage = "@ollama/openclaw-web-search"
-	webSearchMinVersion = "0.2.1"
-)
-
-// ensureWebSearchPlugin installs the openclaw-web-search extension into the
-// user-level extensions directory (~/.openclaw/extensions/) if it isn't already
-// present, or re-installs if the installed version is older than webSearchMinVersion.
-// Returns true if the extension is available.
-func ensureWebSearchPlugin() bool {
-	home, err := os.UserHomeDir()
-	if err != nil {
-		return false
-	}
-
-	pluginDir := filepath.Join(home, ".openclaw", "extensions", "openclaw-web-search")
-	if webSearchPluginUpToDate(pluginDir) {
-		return true
-	}
-
-	npmBin, err := exec.LookPath("npm")
-	if err != nil {
-		return false
-	}
-
-	if err := os.MkdirAll(pluginDir, 0o755); err != nil {
-		return false
-	}
-
-	// Download the tarball via `npm pack`, extract it flat into the plugin dir.
-	pack := exec.Command(npmBin, "pack", webSearchNpmPackage, "--pack-destination", pluginDir)
-	out, err := pack.Output()
-	if err != nil {
-		fmt.Fprintf(os.Stderr, "%s  Warning: could not download web search plugin: %v%s\n", ansiYellow, err, ansiReset)
-		return false
-	}
-
-	tgzName := strings.TrimSpace(string(out))
-	tgzPath := filepath.Join(pluginDir, tgzName)
-	defer os.Remove(tgzPath)
-
-	tar := exec.Command("tar", "xzf", tgzPath, "--strip-components=1", "-C", pluginDir)
-	if err := tar.Run(); err != nil {
-		fmt.Fprintf(os.Stderr, "%s  Warning: could not extract web search plugin: %v%s\n", ansiYellow, err, ansiReset)
-		return false
-	}
-
-	fmt.Fprintf(os.Stderr, "%s  ✓ Installed Ollama web search %s\n", ansiGreen, ansiReset)
-	return true
-}
-
-// webSearchPluginUpToDate returns true if the plugin is installed and its
-// package.json version is >= webSearchMinVersion.
-func webSearchPluginUpToDate(pluginDir string) bool {
-	data, err := os.ReadFile(filepath.Join(pluginDir, "package.json"))
-	if err != nil {
-		return false
-	}
-	var pkg struct {
-		Version string `json:"version"`
-	}
-	if json.Unmarshal(data, &pkg) != nil || pkg.Version == "" {
-		return false
-	}
-	return !versionLessThan(pkg.Version, webSearchMinVersion)
-}
-
-// versionLessThan compares two semver version strings (major.minor.patch).
-// Inputs may omit the "v" prefix; it is added automatically for semver.Compare.
-func versionLessThan(a, b string) bool {
-	if !strings.HasPrefix(a, "v") {
-		a = "v" + a
-	}
-	if !strings.HasPrefix(b, "v") {
-		b = "v" + b
-	}
-	return semver.Compare(a, b) < 0
-}
-
-// registerWebSearchPlugin adds plugins.entries.openclaw-web-search to the OpenClaw
-// config so the gateway activates it on next start. Best-effort; silently returns
-// on any error.
-func registerWebSearchPlugin() {
+// configureOllamaWebSearch keeps launch-managed OpenClaw installs on the
+// bundled Ollama web_search provider. Older launch builds installed an
+// external openclaw-web-search plugin that added custom ollama_web_search and
+// ollama_web_fetch tools. Current OpenClaw versions ship Ollama web_search as
+// the bundled "ollama" plugin instead, so we migrate stale config and ensure
+// fresh installs select the bundled provider.
+func configureOllamaWebSearch() {
 	home, err := os.UserHomeDir()
 	if err != nil {
 		return
@@ -835,6 +755,8 @@ func registerWebSearchPlugin() {
 		return
 	}
 
+	stalePluginConfigured := false
+
 	plugins, _ := config["plugins"].(map[string]any)
 	if plugins == nil {
 		plugins = make(map[string]any)
@@ -843,68 +765,100 @@ func registerWebSearchPlugin() {
 	if entries == nil {
 		entries = make(map[string]any)
 	}
-	entries["openclaw-web-search"] = map[string]any{"enabled": true}
-	plugins["entries"] = entries
-
-	// Pin trust so the gateway doesn't warn about untracked plugins.
-	allow, _ := plugins["allow"].([]any)
-	hasAllow := false
-	for _, v := range allow {
-		if s, ok := v.(string); ok && s == "openclaw-web-search" {
-			hasAllow = true
-			break
-		}
-	}
-	if !hasAllow {
-		allow = append(allow, "openclaw-web-search")
-	}
-	plugins["allow"] = allow
-
-	// Record install provenance so the loader can verify the plugin origin.
-	installs, _ := plugins["installs"].(map[string]any)
-	if installs == nil {
-		installs = make(map[string]any)
-	}
-	pluginDir := filepath.Join(home, ".openclaw", "extensions", "openclaw-web-search")
-	installs["openclaw-web-search"] = map[string]any{
-		"source":      "npm",
-		"spec":        webSearchNpmPackage,
-		"installPath": pluginDir,
-	}
-	plugins["installs"] = installs
-
-	config["plugins"] = plugins
-
-	// Add plugin tools to tools.alsoAllow so they survive the coding profile's
-	// policy pipeline (which has an explicit allow list of core tools only).
 	tools, _ := config["tools"].(map[string]any)
 	if tools == nil {
 		tools = make(map[string]any)
 	}
-
-	alsoAllow, _ := tools["alsoAllow"].([]any)
-	needed := []string{"ollama_web_search", "ollama_web_fetch"}
-	have := make(map[string]bool, len(alsoAllow))
-	for _, v := range alsoAllow {
-		if s, ok := v.(string); ok {
-			have[s] = true
-		}
-	}
-	for _, name := range needed {
-		if !have[name] {
-			alsoAllow = append(alsoAllow, name)
-		}
-	}
-	tools["alsoAllow"] = alsoAllow
-
-	// Disable built-in web search/fetch since our plugin replaces them.
 	web, _ := tools["web"].(map[string]any)
 	if web == nil {
 		web = make(map[string]any)
 	}
-	web["search"] = map[string]any{"enabled": false}
-	web["fetch"] = map[string]any{"enabled": false}
+	search, _ := web["search"].(map[string]any)
+	if search == nil {
+		search = make(map[string]any)
+	}
+	fetch, _ := web["fetch"].(map[string]any)
+	if fetch == nil {
+		fetch = make(map[string]any)
+	}
+
+	alsoAllow, _ := tools["alsoAllow"].([]any)
+	var filteredAlsoAllow []any
+	for _, v := range alsoAllow {
+		s, ok := v.(string)
+		if !ok {
+			filteredAlsoAllow = append(filteredAlsoAllow, v)
+			continue
+		}
+		if s == "ollama_web_search" || s == "ollama_web_fetch" {
+			stalePluginConfigured = true
+			continue
+		}
+		filteredAlsoAllow = append(filteredAlsoAllow, v)
+	}
+	if len(filteredAlsoAllow) > 0 {
+		tools["alsoAllow"] = filteredAlsoAllow
+	} else {
+		delete(tools, "alsoAllow")
+	}
+
+	if _, ok := entries["openclaw-web-search"]; ok {
+		delete(entries, "openclaw-web-search")
+		stalePluginConfigured = true
+	}
+	ollamaEntry, _ := entries["ollama"].(map[string]any)
+	if ollamaEntry == nil {
+		ollamaEntry = make(map[string]any)
+	}
+	ollamaEntry["enabled"] = true
+	entries["ollama"] = ollamaEntry
+	plugins["entries"] = entries
+
+	if allow, ok := plugins["allow"].([]any); ok {
+		var nextAllow []any
+		hasOllama := false
+		for _, v := range allow {
+			s, ok := v.(string)
+			if ok && s == "openclaw-web-search" {
+				stalePluginConfigured = true
+				continue
+			}
+			if ok && s == "ollama" {
+				hasOllama = true
+			}
+			nextAllow = append(nextAllow, v)
+		}
+		if !hasOllama {
+			nextAllow = append(nextAllow, "ollama")
+		}
+		plugins["allow"] = nextAllow
+	}
+
+	if installs, ok := plugins["installs"].(map[string]any); ok {
+		if _, exists := installs["openclaw-web-search"]; exists {
+			delete(installs, "openclaw-web-search")
+			stalePluginConfigured = true
+		}
+		if len(installs) > 0 {
+			plugins["installs"] = installs
+		} else {
+			delete(plugins, "installs")
+		}
+	}
+
+	if stalePluginConfigured || search["provider"] == nil {
+		search["provider"] = "ollama"
+	}
+	if stalePluginConfigured {
+		fetch["enabled"] = true
+	}
+	search["enabled"] = true
+	web["search"] = search
+	if len(fetch) > 0 {
+		web["fetch"] = fetch
+	}
 	tools["web"] = web
+	config["plugins"] = plugins
 	config["tools"] = tools
 
 	out, err := json.MarshalIndent(config, "", "  ")

cmd/launch/openclaw_test.go

@@ -2242,95 +2242,7 @@ func TestIntegrationOnboarded(t *testing.T) {
 	})
 }
 
-func TestVersionLessThan(t *testing.T) {
-	tests := []struct {
-		a, b string
-		want bool
-	}{
-		{"0.1.7", "0.2.1", true},
-		{"0.2.0", "0.2.1", true},
-		{"0.2.1", "0.2.1", false},
-		{"0.2.2", "0.2.1", false},
-		{"1.0.0", "0.2.1", false},
-		{"0.2.1", "1.0.0", true},
-		{"v0.1.7", "0.2.1", true},
-		{"0.2.1", "v0.2.1", false},
-	}
-	for _, tt := range tests {
-		t.Run(tt.a+"_vs_"+tt.b, func(t *testing.T) {
-			if got := versionLessThan(tt.a, tt.b); got != tt.want {
-				t.Errorf("versionLessThan(%q, %q) = %v, want %v", tt.a, tt.b, got, tt.want)
-			}
-		})
-	}
-}
-
-func TestWebSearchPluginUpToDate(t *testing.T) {
-	t.Run("missing directory", func(t *testing.T) {
-		if webSearchPluginUpToDate(filepath.Join(t.TempDir(), "nonexistent")) {
-			t.Error("expected false for missing directory")
-		}
-	})
-
-	t.Run("missing package.json", func(t *testing.T) {
-		dir := t.TempDir()
-		if webSearchPluginUpToDate(dir) {
-			t.Error("expected false for missing package.json")
-		}
-	})
-
-	t.Run("old version", func(t *testing.T) {
-		dir := t.TempDir()
-		if err := os.WriteFile(filepath.Join(dir, "package.json"), []byte(`{"version":"0.1.7"}`), 0o644); err != nil {
-			t.Fatal(err)
-		}
-		if webSearchPluginUpToDate(dir) {
-			t.Error("expected false for old version 0.1.7")
-		}
-	})
-
-	t.Run("exact minimum version", func(t *testing.T) {
-		dir := t.TempDir()
-		if err := os.WriteFile(filepath.Join(dir, "package.json"), []byte(`{"version":"0.2.1"}`), 0o644); err != nil {
-			t.Fatal(err)
-		}
-		if !webSearchPluginUpToDate(dir) {
-			t.Error("expected true for exact minimum version 0.2.1")
-		}
-	})
-
-	t.Run("newer version", func(t *testing.T) {
-		dir := t.TempDir()
-		if err := os.WriteFile(filepath.Join(dir, "package.json"), []byte(`{"version":"1.0.0"}`), 0o644); err != nil {
-			t.Fatal(err)
-		}
-		if !webSearchPluginUpToDate(dir) {
-			t.Error("expected true for newer version 1.0.0")
-		}
-	})
-
-	t.Run("invalid json", func(t *testing.T) {
-		dir := t.TempDir()
-		if err := os.WriteFile(filepath.Join(dir, "package.json"), []byte(`not json`), 0o644); err != nil {
-			t.Fatal(err)
-		}
-		if webSearchPluginUpToDate(dir) {
-			t.Error("expected false for invalid json")
-		}
-	})
-
-	t.Run("empty version", func(t *testing.T) {
-		dir := t.TempDir()
-		if err := os.WriteFile(filepath.Join(dir, "package.json"), []byte(`{"version":""}`), 0o644); err != nil {
-			t.Fatal(err)
-		}
-		if webSearchPluginUpToDate(dir) {
-			t.Error("expected false for empty version")
-		}
-	})
-}
-
-func TestRegisterWebSearchPlugin(t *testing.T) {
+func TestConfigureOllamaWebSearch(t *testing.T) {
 	home := t.TempDir()
 	setTestHome(t, home)
 
@@ -2345,7 +2257,7 @@ func TestRegisterWebSearchPlugin(t *testing.T) {
 			t.Fatal(err)
 		}
 
-		registerWebSearchPlugin()
+		configureOllamaWebSearch()
 
 		data, err := os.ReadFile(configPath)
 		if err != nil {
@@ -2361,40 +2273,30 @@ func TestRegisterWebSearchPlugin(t *testing.T) {
 			t.Fatal("plugins section missing")
 		}
 
-		// Check entries
 		entries, _ := plugins["entries"].(map[string]any)
-		entry, _ := entries["openclaw-web-search"].(map[string]any)
+		entry, _ := entries["ollama"].(map[string]any)
 		if enabled, _ := entry["enabled"].(bool); !enabled {
-			t.Error("expected entries.openclaw-web-search.enabled = true")
+			t.Error("expected entries.ollama.enabled = true")
+		}
+		if _, ok := entries["openclaw-web-search"]; ok {
+			t.Error("expected stale openclaw-web-search entry to be absent")
 		}
 
-		// Check allow list
-		allow, _ := plugins["allow"].([]any)
-		found := false
-		for _, v := range allow {
-			if s, ok := v.(string); ok && s == "openclaw-web-search" {
-				found = true
-			}
+		if _, ok := plugins["allow"]; ok {
+			t.Error("did not expect plugins.allow to be created when no allowlist exists")
 		}
-		if !found {
-			t.Error("expected plugins.allow to contain openclaw-web-search")
+		if _, ok := plugins["installs"]; ok {
+			t.Error("did not expect plugins.installs to be created")
 		}
 
-		// Check install provenance
-		installs, _ := plugins["installs"].(map[string]any)
-		record, _ := installs["openclaw-web-search"].(map[string]any)
-		if record == nil {
-			t.Fatal("expected plugins.installs.openclaw-web-search")
+		tools, _ := config["tools"].(map[string]any)
+		web, _ := tools["web"].(map[string]any)
+		search, _ := web["search"].(map[string]any)
+		if got, _ := search["provider"].(string); got != "ollama" {
+			t.Errorf("search provider = %q, want %q", got, "ollama")
 		}
-		if source, _ := record["source"].(string); source != "npm" {
-			t.Errorf("install source = %q, want %q", source, "npm")
-		}
-		if spec, _ := record["spec"].(string); spec != webSearchNpmPackage {
-			t.Errorf("install spec = %q, want %q", spec, webSearchNpmPackage)
-		}
-		expectedPath := filepath.Join(home, ".openclaw", "extensions", "openclaw-web-search")
-		if installPath, _ := record["installPath"].(string); installPath != expectedPath {
-			t.Errorf("installPath = %q, want %q", installPath, expectedPath)
+		if enabled, _ := search["enabled"].(bool); !enabled {
+			t.Error("expected tools.web.search.enabled = true")
 		}
 	})
 
@@ -2403,8 +2305,8 @@ func TestRegisterWebSearchPlugin(t *testing.T) {
 			t.Fatal(err)
 		}
 
-		registerWebSearchPlugin()
-		registerWebSearchPlugin()
+		configureOllamaWebSearch()
+		configureOllamaWebSearch()
 
 		data, err := os.ReadFile(configPath)
 		if err != nil {
@@ -2416,30 +2318,39 @@ func TestRegisterWebSearchPlugin(t *testing.T) {
 		}
 
 		plugins, _ := config["plugins"].(map[string]any)
-		allow, _ := plugins["allow"].([]any)
-		count := 0
-		for _, v := range allow {
-			if s, ok := v.(string); ok && s == "openclaw-web-search" {
-				count++
-			}
+		entries, _ := plugins["entries"].(map[string]any)
+		if len(entries) != 1 {
+			t.Fatalf("expected only bundled ollama entry, got %v", entries)
 		}
-		if count != 1 {
-			t.Errorf("expected exactly 1 openclaw-web-search in allow, got %d", count)
+		if _, ok := entries["ollama"]; !ok {
+			t.Fatalf("expected entries.ollama to exist, got %v", entries)
 		}
 	})
 
-	t.Run("preserves existing config", func(t *testing.T) {
+	t.Run("migrates stale plugin config and preserves unrelated settings", func(t *testing.T) {
 		initial := map[string]any{
 			"plugins": map[string]any{
-				"allow": []any{"some-other-plugin"},
+				"allow": []any{"some-other-plugin", "openclaw-web-search"},
 				"entries": map[string]any{
-					"some-other-plugin": map[string]any{"enabled": true},
+					"some-other-plugin":   map[string]any{"enabled": true},
+					"openclaw-web-search": map[string]any{"enabled": true},
 				},
 				"installs": map[string]any{
 					"some-other-plugin": map[string]any{
 						"source":      "npm",
 						"installPath": "/some/path",
 					},
+					"openclaw-web-search": map[string]any{
+						"source":      "npm",
+						"installPath": "/old/path",
+					},
+				},
+			},
+			"tools": map[string]any{
+				"alsoAllow": []any{"ollama_web_search", "ollama_web_fetch", "browser"},
+				"web": map[string]any{
+					"search": map[string]any{"enabled": false},
+					"fetch":  map[string]any{"enabled": false},
 				},
 			},
 			"customField": "preserved",
@@ -2449,7 +2360,7 @@ func TestRegisterWebSearchPlugin(t *testing.T) {
 			t.Fatal(err)
 		}
 
-		registerWebSearchPlugin()
+		configureOllamaWebSearch()
 
 		out, err := os.ReadFile(configPath)
 		if err != nil {
@@ -2469,28 +2380,61 @@ func TestRegisterWebSearchPlugin(t *testing.T) {
 		if entries["some-other-plugin"] == nil {
 			t.Error("existing plugin entry was lost")
 		}
+		if entries["openclaw-web-search"] != nil {
+			t.Error("stale openclaw-web-search entry should be removed")
+		}
+		if ollamaEntry, _ := entries["ollama"].(map[string]any); ollamaEntry == nil {
+			t.Fatal("expected bundled ollama entry to be enabled")
+		}
 
 		installs, _ := plugins["installs"].(map[string]any)
 		if installs["some-other-plugin"] == nil {
 			t.Error("existing install record was lost")
 		}
+		if installs["openclaw-web-search"] != nil {
+			t.Error("stale openclaw-web-search install record should be removed")
+		}
 
 		allow, _ := plugins["allow"].([]any)
-		hasOther, hasWebSearch := false, false
+		hasOther, hasStalePlugin, hasOllama := false, false, false
 		for _, v := range allow {
 			s, _ := v.(string)
 			if s == "some-other-plugin" {
 				hasOther = true
 			}
 			if s == "openclaw-web-search" {
-				hasWebSearch = true
+				hasStalePlugin = true
+			}
+			if s == "ollama" {
+				hasOllama = true
 			}
 		}
 		if !hasOther {
 			t.Error("existing allow entry was lost")
 		}
-		if !hasWebSearch {
-			t.Error("openclaw-web-search not added to allow")
+		if hasStalePlugin {
+			t.Error("stale openclaw-web-search allow entry should be removed")
+		}
+		if !hasOllama {
+			t.Error("expected plugins.allow to contain bundled ollama plugin")
+		}
+
+		tools, _ := config["tools"].(map[string]any)
+		alsoAllow, _ := tools["alsoAllow"].([]any)
+		if len(alsoAllow) != 1 || alsoAllow[0] != "browser" {
+			t.Errorf("expected stale custom web tools to be removed, got %v", alsoAllow)
+		}
+		web, _ := tools["web"].(map[string]any)
+		search, _ := web["search"].(map[string]any)
+		fetch, _ := web["fetch"].(map[string]any)
+		if got, _ := search["provider"].(string); got != "ollama" {
+			t.Errorf("search provider = %q, want %q", got, "ollama")
+		}
+		if enabled, _ := search["enabled"].(bool); !enabled {
+			t.Error("expected migrated tools.web.search.enabled = true")
+		}
+		if enabled, _ := fetch["enabled"].(bool); !enabled {
+			t.Error("expected migrated tools.web.fetch.enabled = true")
 		}
 	})
 }

docs/capabilities/structured-outputs.mdx

@@ -2,6 +2,10 @@
 title: Structured Outputs
 ---
 
+<Note>
+  Ollama's Cloud currently does not support structured outputs.
+</Note>
+
 Structured outputs let you enforce a JSON schema on model responses so you can reliably extract structured data, describe images, or keep every reply consistent.
 
 ## Generating structured JSON

docs/integrations/openclaw.mdx

@@ -15,7 +15,7 @@ Ollama handles everything automatically:
 1. **Install** — If OpenClaw isn't installed, Ollama prompts to install it via npm
 2. **Security** — On the first launch, a security notice explains the risks of tool access
 3. **Model** — Pick a model from the selector (local or cloud)
-4. **Onboarding** — Ollama configures the provider, installs the gateway daemon, sets your model as the primary, and installs the web search and fetch plugin
+4. **Onboarding** — Ollama configures the provider, installs the gateway daemon, sets your model as the primary, and enables OpenClaw's bundled Ollama web search
 5. **Gateway** — Starts in the background and opens the OpenClaw TUI
 
 <Note>OpenClaw requires a larger context window. It is recommended to use a context window of at least 64k tokens if using local models. See [Context length](/context-length) for more information.</Note>
@@ -24,19 +24,19 @@ Ollama handles everything automatically:
 
 ## Web search and fetch
 
-OpenClaw ships with a web search and fetch plugin that gives local or cloud models the ability to search the web and extract readable page content.
+OpenClaw ships with a bundled Ollama `web_search` provider that lets local or cloud-backed Ollama setups search the web through the configured Ollama host.
 
 ```bash
 ollama launch openclaw
 ```
 
-Web search and fetch is enabled automatically when launching OpenClaw through Ollama. To install the plugin directly:
+Ollama web search is enabled automatically when launching OpenClaw through Ollama. To configure it manually:
 
 ```bash
-openclaw plugins install @ollama/openclaw-web-search
+openclaw configure --section web
 ```
 
-<Note>Web search for local models requires `ollama signin`.</Note>
+<Note>Ollama web search for local models requires `ollama signin`.</Note>
 
 ## Configure without launching
 
@@ -93,4 +93,3 @@ Link WhatsApp, Telegram, Slack, Discord, or iMessage to chat with your local mod
 ```bash
 openclaw gateway stop
 ```
-

integration/api_test.go

@@ -406,10 +406,6 @@ func TestAPIShowModel(t *testing.T) {
 }
 
 func TestAPIGenerateLogprobs(t *testing.T) {
-	if testModel != "" {
-		// Logprobs requires runner support (e.g. llama.cpp has it, MLX does not).
-		t.Skip("logprobs not supported by all runners")
-	}
 	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute)
 	defer cancel()
 
@@ -523,10 +519,6 @@ func TestAPIGenerateLogprobs(t *testing.T) {
 }
 
 func TestAPIChatLogprobs(t *testing.T) {
-	if testModel != "" {
-		// Logprobs requires runner support (e.g. llama.cpp has it, MLX does not).
-		t.Skip("logprobs not supported by all runners")
-	}
 	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute)
 	defer cancel()
 

x/mlxrunner/client.go

@@ -151,22 +151,11 @@ func (c *Client) WaitUntilRunning(ctx context.Context) error {
 	}
 }
 
-// completionRequest is a properly-tagged version of llm.CompletionRequest for JSON serialization.
-type completionRequest struct {
-	Prompt  string          `json:"prompt"`
-	Options *completionOpts `json:"options,omitempty"`
-}
-
-type completionOpts struct {
-	Temperature      float32 `json:"temperature,omitempty"`
-	TopP             float32 `json:"top_p,omitempty"`
-	MinP             float32 `json:"min_p,omitempty"`
-	TopK             int     `json:"top_k,omitempty"`
-	RepeatLastN      int     `json:"repeat_last_n,omitempty"`
-	RepeatPenalty    float32 `json:"repeat_penalty,omitempty"`
-	PresencePenalty  float32 `json:"presence_penalty,omitempty"`
-	FrequencyPenalty float32 `json:"frequency_penalty,omitempty"`
-	NumPredict       int     `json:"num_predict,omitempty"`
+type CompletionRequest struct {
+	Prompt      string
+	Options     api.Options
+	Logprobs    bool
+	TopLogprobs int
 }
 
 type CompletionResponse struct {
@@ -179,6 +168,8 @@ type CompletionResponse struct {
 	EvalCount          int
 	EvalDuration       time.Duration
 
+	Logprobs []llm.Logprob
+
 	Error *api.StatusError
 }
 
@@ -203,21 +194,13 @@ func (c *Client) Close() error {
 
 // Completion implements llm.LlamaServer.
 func (c *Client) Completion(ctx context.Context, req llm.CompletionRequest, fn func(llm.CompletionResponse)) error {
-	creq := completionRequest{
-		Prompt: req.Prompt,
+	creq := CompletionRequest{
+		Prompt:      req.Prompt,
+		Logprobs:    req.Logprobs,
+		TopLogprobs: req.TopLogprobs,
 	}
 	if req.Options != nil {
-		creq.Options = &completionOpts{
-			Temperature:      req.Options.Temperature,
-			TopP:             req.Options.TopP,
-			MinP:             req.Options.MinP,
-			TopK:             req.Options.TopK,
-			RepeatLastN:      req.Options.RepeatLastN,
-			RepeatPenalty:    req.Options.RepeatPenalty,
-			PresencePenalty:  req.Options.PresencePenalty,
-			FrequencyPenalty: req.Options.FrequencyPenalty,
-			NumPredict:       req.Options.NumPredict,
-		}
+		creq.Options = *req.Options
 	}
 
 	body, err := json.Marshal(creq)
@@ -243,7 +226,7 @@ func (c *Client) Completion(ctx context.Context, req llm.CompletionRequest, fn f
 
 	if resp.StatusCode != http.StatusOK {
 		respBody, _ := io.ReadAll(resp.Body)
-		return fmt.Errorf("%s", strings.TrimSpace(string(respBody)))
+		return api.StatusError{StatusCode: resp.StatusCode, ErrorMessage: strings.TrimSpace(string(respBody))}
 	}
 
 	scanner := bufio.NewScanner(resp.Body)
@@ -266,6 +249,7 @@ func (c *Client) Completion(ctx context.Context, req llm.CompletionRequest, fn f
 			PromptEvalDuration: raw.PromptEvalDuration,
 			EvalCount:          raw.EvalCount,
 			EvalDuration:       raw.EvalDuration,
+			Logprobs:           raw.Logprobs,
 		}
 
 		fn(cresp)

x/mlxrunner/mlx/array.go

@@ -10,6 +10,8 @@ import (
 	"reflect"
 	"sort"
 	"strings"
+	"sync"
+	"sync/atomic"
 	"unsafe"
 
 	"github.com/ollama/ollama/logutil"
@@ -18,20 +20,28 @@ import (
 type Array struct {
 	ctx    C.mlx_array
 	name   string
-	pinned int
+	pinned atomic.Int32
 }
 
-var arrays []*Array
+var (
+	arrays   []*Array
+	arraysMu sync.Mutex
+)
 
 // constructor utilities
 
 func New(name string) *Array {
 	t := &Array{name: name}
+
 	if tracing {
 		traceScratch = append(traceScratch, t)
 	} else {
+		arraysMu.Lock()
+		defer arraysMu.Unlock()
+
 		arrays = append(arrays, t)
 	}
+
 	return t
 }
 
@@ -131,7 +141,7 @@ func (t *Array) Clone() *Array {
 func Pin(s ...*Array) {
 	for _, t := range s {
 		if t != nil {
-			t.pinned++
+			t.pinned.Add(1)
 		}
 	}
 }
@@ -140,8 +150,7 @@ func Pin(s ...*Array) {
 func Unpin(s ...*Array) {
 	for _, t := range s {
 		if t != nil {
-			t.pinned--
-			if t.pinned < 0 {
+			if t.pinned.Add(-1) < 0 {
 				panic(fmt.Sprintf("mlx.Unpin: negative pin count on array %q", t.name))
 			}
 		}
@@ -151,9 +160,11 @@ func Unpin(s ...*Array) {
 // Sweep releases all unpinned arrays, primarily intermediate tensors. MLX will truly
 // free them when there are no other references, including dependencies in the graph.
 func Sweep() {
+	arraysMu.Lock()
+	defer arraysMu.Unlock()
 	n := 0
 	for _, t := range arrays {
-		if t.pinned > 0 && t.Valid() {
+		if t.pinned.Load() > 0 && t.Valid() {
 			arrays[n] = t
 			n++
 		} else if t.Valid() {
@@ -180,7 +191,7 @@ func (t *Array) String() string {
 func (t *Array) LogValue() slog.Value {
 	attrs := []slog.Attr{
 		slog.String("name", t.name),
-		slog.Int("pinned", t.pinned),
+		slog.Int("pinned", int(t.pinned.Load())),
 	}
 	if t.Valid() {
 		attrs = append(attrs,
@@ -194,19 +205,19 @@ func (t *Array) LogValue() slog.Value {
 
 // shape utilities
 
-func (t Array) Size() int {
+func (t *Array) Size() int {
 	return int(C.mlx_array_size(t.ctx))
 }
 
-func (t Array) NumBytes() int {
+func (t *Array) NumBytes() int {
 	return int(C.mlx_array_nbytes(t.ctx))
 }
 
-func (t Array) NumDims() int {
+func (t *Array) NumDims() int {
 	return int(C.mlx_array_ndim(t.ctx))
 }
 
-func (t Array) Dims() []int {
+func (t *Array) Dims() []int {
 	dims := make([]int, t.NumDims())
 	for i := range dims {
 		dims[i] = t.Dim(i)
@@ -215,29 +226,32 @@ func (t Array) Dims() []int {
 	return dims
 }
 
-func (t Array) Dim(dim int) int {
+func (t *Array) Dim(dim int) int {
 	return int(C.mlx_array_dim(t.ctx, C.int(dim)))
 }
 
-func (t Array) DType() DType {
+func (t *Array) DType() DType {
 	return DType(C.mlx_array_dtype(t.ctx))
 }
 
 // data utilities
 
-func (t Array) Int() int {
+func (t *Array) Int() int {
 	var item C.int64_t
 	C.mlx_array_item_int64(&item, t.ctx)
 	return int(item)
 }
 
-func (t Array) Float() float64 {
+func (t *Array) Float() float64 {
 	var item C.double
 	C.mlx_array_item_float64(&item, t.ctx)
 	return float64(item)
 }
 
-func (t Array) Ints() []int {
+func (t *Array) Ints() []int {
+	if dt := t.DType(); dt != DTypeInt32 {
+		panic(fmt.Sprintf("mlx: Ints requires DTypeInt32, got %v", dt))
+	}
 	ints := make([]int, t.Size())
 	for i, f := range unsafe.Slice(C.mlx_array_data_int32(t.ctx), len(ints)) {
 		ints[i] = int(f)
@@ -245,7 +259,10 @@ func (t Array) Ints() []int {
 	return ints
 }
 
-func (t Array) Floats() []float32 {
+func (t *Array) Floats() []float32 {
+	if dt := t.DType(); dt != DTypeFloat32 {
+		panic(fmt.Sprintf("mlx: Floats requires DTypeFloat32, got %v", dt))
+	}
 	floats := make([]float32, t.Size())
 	for i, f := range unsafe.Slice(C.mlx_array_data_float32(t.ctx), len(floats)) {
 		floats[i] = float32(f)
@@ -253,7 +270,7 @@ func (t Array) Floats() []float32 {
 	return floats
 }
 
-func (t Array) Save(name string) error {
+func (t *Array) Save(name string) error {
 	cName := C.CString(name)
 	defer C.free(unsafe.Pointer(cName))
 	C.mlx_save(cName, t.ctx)
@@ -262,6 +279,8 @@ func (t Array) Save(name string) error {
 
 // LogArrays logs all live arrays, sorted by size
 func LogArrays() {
+	arraysMu.Lock()
+	defer arraysMu.Unlock()
 	sort.Slice(arrays, func(i, j int) bool {
 		return arrays[i].NumBytes() > arrays[j].NumBytes()
 	})
@@ -270,7 +289,7 @@ func LogArrays() {
 	for _, t := range arrays {
 		nb := t.NumBytes()
 		total += nb
-		logutil.Trace(fmt.Sprintf("tensor %-60s %5s %5s pinned=%d %v", t.name, t.DType(), PrettyBytes(nb), t.pinned, t.Dims()))
+		logutil.Trace(fmt.Sprintf("tensor %-60s %5s %5s pinned=%d %v", t.name, t.DType(), PrettyBytes(nb), t.pinned.Load(), t.Dims()))
 	}
 	logutil.Trace(fmt.Sprintf("tensors total: %d, size: %s, active: %s", len(arrays), PrettyBytes(total), PrettyBytes(ActiveMemory())))
 }

x/mlxrunner/mlx/compile.go

@@ -150,7 +150,7 @@ func closureCallback(res *C.mlx_vector_array, input C.mlx_vector_array, payload
 	traceScratch = nil
 	defer func() {
 		for _, a := range traceScratch {
-			if a.pinned > 0 {
+			if a.pinned.Load() > 0 {
 				panic("mlx: traced array was pinned during compilation")
 			}
 			if a.Valid() {

x/mlxrunner/mlx/fast.go

@@ -24,8 +24,8 @@ func ScaledDotProductAttention(query, key, value, mask *Array, scale float32) *A
 }
 
 type LayerNorm struct {
-	Weight Array `weight:"weight"`
-	Bias   Array `weight:"bias"`
+	Weight *Array `weight:"weight"`
+	Bias   *Array `weight:"bias"`
 }
 
 func (r *LayerNorm) Forward(x *Array, eps float32) *Array {
@@ -35,10 +35,10 @@ func (r *LayerNorm) Forward(x *Array, eps float32) *Array {
 }
 
 type RMSNorm struct {
-	Weight Array `weight:"weight"`
+	Weight *Array `weight:"weight"`
 }
 
-func (r RMSNorm) Forward(x *Array, eps float32) *Array {
+func (r *RMSNorm) Forward(x *Array, eps float32) *Array {
 	out := New("FAST_RMSNORM")
 	C.mlx_fast_rms_norm(&out.ctx, x.ctx, r.Weight.ctx, C.float(eps), DefaultStream().ctx)
 	return out

x/mlxrunner/mlx/nn.go

@@ -1,12 +1,12 @@
 package mlx
 
 type Linear struct {
-	Weight Array `weight:"weight"`
-	Bias   Array `weight:"bias"`
+	Weight *Array `weight:"weight"`
+	Bias   *Array `weight:"bias"`
 }
 
 // Forward computes the linear transformation: x @ Weight.T + Bias
-func (m Linear) Forward(x *Array) *Array {
+func (m *Linear) Forward(x *Array) *Array {
 	w := m.Weight.Transpose(1, 0)
 	if m.Bias.Valid() {
 		return m.Bias.Addmm(x, w, 1.0, 1.0)
@@ -15,14 +15,14 @@ func (m Linear) Forward(x *Array) *Array {
 	return x.Matmul(w)
 }
 
-func (m Linear) Gather(x, lhs, rhs *Array, sorted bool) *Array {
+func (m *Linear) Gather(x, lhs, rhs *Array, sorted bool) *Array {
 	w := m.Weight.Transpose(0, 2, 1)
 	// TODO: bias
 	return x.GatherMM(w, lhs, rhs, sorted)
 }
 
 type Embedding struct {
-	Weight Array `weight:"weight"`
+	Weight *Array `weight:"weight"`
 }
 
 func (e *Embedding) Forward(indices *Array) *Array {

x/mlxrunner/mlx/ops.go

@@ -139,6 +139,12 @@ func (t *Array) Less(other *Array) *Array {
 	return out
 }
 
+func (t *Array) MaxAxis(axis int, keepDims bool) *Array {
+	out := New("MAX_AXIS")
+	C.mlx_max_axis(&out.ctx, t.ctx, C.int(axis), C.bool(keepDims), DefaultStream().ctx)
+	return out
+}
+
 func (t *Array) Matmul(other *Array) *Array {
 	out := New("MATMUL")
 	C.mlx_matmul(&out.ctx, t.ctx, other.ctx, DefaultStream().ctx)

x/mlxrunner/pipeline.go

@@ -6,36 +6,59 @@ import (
 	"errors"
 	"fmt"
 	"log/slog"
-	"net/http"
+	"sort"
 	"time"
 
-	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/llm"
 	"github.com/ollama/ollama/logutil"
 	"github.com/ollama/ollama/x/mlxrunner/mlx"
+	sampler "github.com/ollama/ollama/x/mlxrunner/sample"
+	"github.com/ollama/ollama/x/tokenizer"
 )
 
 func prefillChunkSize() int {
 	return 2 << 10
 }
 
-func (r *Runner) TextGenerationPipeline(request Request) error {
+// Prepare tokenizes the prompt and validates it against the model's
+// context length. It is safe to call from any goroutine. On success it
+// populates request.Tokens and adjusts request.Options.NumPredict.
+func (r *Runner) Prepare(request *Request) error {
 	if r.Model == nil {
 		return errors.New("model not loaded")
 	}
 
+	tokens := r.Tokenizer.Encode(request.Prompt, r.Tokenizer.AddBOS())
+	if len(tokens) == 0 {
+		return errors.New("empty prompt")
+	}
+
+	if len(tokens) >= r.contextLength {
+		return fmt.Errorf("input length (%d tokens) exceeds the model's maximum context length (%d tokens)", len(tokens), r.contextLength)
+	}
+
+	// Cap generation to stay within the model's context length
+	maxGenerate := r.contextLength - len(tokens)
+	if request.Options.NumPredict <= 0 {
+		request.Options.NumPredict = maxGenerate
+	} else {
+		request.Options.NumPredict = min(request.Options.NumPredict, maxGenerate)
+	}
+
+	request.Tokens = tokens
+	return nil
+}
+
+func (r *Runner) TextGenerationPipeline(ctx context.Context, request Request) error {
 	mlx.ResetPeakMemory()
-	ctx := request.Ctx
-	var (
-		sample     *mlx.Array
-		nextSample *mlx.Array
-	)
+	var sample, nextSample sampler.Result
 
 	defer func() {
 		if request.Sampler != nil {
 			request.Sampler.Free()
 		}
-		mlx.Unpin(sample)
-		mlx.Unpin(nextSample)
+		mlx.Unpin(sample.Arrays()...)
+		mlx.Unpin(nextSample.Arrays()...)
 		mlx.Sweep()
 		mlx.ClearCache()
 
@@ -46,26 +69,7 @@ func (r *Runner) TextGenerationPipeline(request Request) error {
 		slog.Info("peak memory", "size", mlx.PrettyBytes(mlx.PeakMemory()))
 	}()
 
-	inputs := r.Tokenizer.Encode(request.Prompt, r.Tokenizer.AddBOS())
-	if len(inputs) == 0 {
-		return errors.New("empty prompt")
-	}
-
-	if len(inputs) >= r.contextLength {
-		return api.StatusError{
-			StatusCode:   http.StatusBadRequest,
-			ErrorMessage: fmt.Sprintf("input length (%d tokens) exceeds the model's maximum context length (%d tokens)", len(inputs), r.contextLength),
-		}
-	}
-
-	// Cap generation to stay within the model's context length
-	maxGenerate := r.contextLength - len(inputs)
-	if request.Options.MaxTokens <= 0 {
-		request.Options.MaxTokens = maxGenerate
-	} else {
-		request.Options.MaxTokens = min(request.Options.MaxTokens, maxGenerate)
-	}
-
+	inputs := request.Tokens
 	request.Sampler.ResetHistory(inputs)
 
 	session := r.cache.begin(r.Model, inputs)
@@ -135,40 +139,38 @@ func (r *Runner) TextGenerationPipeline(request Request) error {
 		mlx.ClearCache()
 	}
 
-	step := func(token *mlx.Array) *mlx.Array {
+	step := func(token *mlx.Array) sampler.Result {
 		fwd := r.Model.Forward(token.ExpandDims(0), caches)
 		logits := r.Model.Unembed(fwd)
 		logits = logits.Slice(mlx.Slice(), mlx.Slice(logits.Dim(1)-1), mlx.Slice()).Squeeze(1)
 
 		sample := request.Sampler.Sample(logits)
-
-		mlx.Pin(sample)
+		mlx.Pin(sample.Arrays()...)
 		mlx.Sweep()
-		mlx.AsyncEval(sample)
-
+		mlx.AsyncEval(sample.Arrays()...)
 		return sample
 	}
 
 	sample = step(mlx.FromValues(tokens[processed:], total-processed))
 
-	var b bytes.Buffer
+	dec := decoder{tokenizer: r.Tokenizer}
 
-	final := CompletionResponse{Done: true, PromptEvalCount: len(inputs), EvalCount: request.Options.MaxTokens, DoneReason: 1}
-	for i := range request.Options.MaxTokens {
+	final := CompletionResponse{Done: true, PromptEvalCount: len(inputs), EvalCount: request.Options.NumPredict, DoneReason: 1}
+	for i := range request.Options.NumPredict {
 		if err := ctx.Err(); err != nil {
 			return err
 		}
 
-		request.Sampler.AppendToken(sample)
-		nextSample = step(sample)
+		request.Sampler.AppendToken(sample.Token)
+		nextSample = step(sample.Token)
 
 		if i == 0 {
-			mlx.Eval(sample)
+			mlx.Eval(sample.Arrays()...)
 			final.PromptEvalDuration = time.Since(now)
 			now = time.Now()
 		}
 
-		output := int32(sample.Int())
+		output := int32(sample.Token.Int())
 		session.outputs = append(session.outputs, output)
 
 		if r.Tokenizer.IsEOS(output) {
@@ -177,17 +179,16 @@ func (r *Runner) TextGenerationPipeline(request Request) error {
 			break
 		}
 
-		select {
-		case <-ctx.Done():
-			return ctx.Err()
-		case request.Responses <- CompletionResponse{
-			Content: r.Decode(output, &b),
-		}:
+		if resp, ok := dec.decode(sample); ok {
+			select {
+			case <-ctx.Done():
+				return ctx.Err()
+			case request.Responses <- resp:
+			}
 		}
 
-		mlx.Unpin(sample)
-		sample = nextSample
-		nextSample = nil
+		mlx.Unpin(sample.Arrays()...)
+		sample, nextSample = nextSample, sampler.Result{}
 
 		if i%256 == 0 {
 			mlx.ClearCache()
@@ -203,13 +204,57 @@ func (r *Runner) TextGenerationPipeline(request Request) error {
 	}
 }
 
-func (r Runner) Decode(sample int32, b *bytes.Buffer) string {
-	token := r.Tokenizer.Decode([]int32{sample})
+// decoder serializes sampled tokens into response chunks, holding bytes
+// whose UTF-8 sequence hasn't completed yet and the logprobs that belong
+// with those bytes so Content and Logprobs stay aligned when a chunk does
+// flush.
+type decoder struct {
+	tokenizer *tokenizer.Tokenizer
+	buf       bytes.Buffer
+	logprobs  []llm.Logprob
+}
 
-	if _, err := b.WriteString(token); err != nil {
-		slog.Error("Failed to write token to buffer", "error", err)
-		return ""
+func (d *decoder) decode(res sampler.Result) (CompletionResponse, bool) {
+	output := int32(res.Token.Int())
+	d.buf.WriteString(d.tokenizer.Decode([]int32{output}))
+	d.logprobs = append(d.logprobs, buildLogprob(res, d.tokenizer.Decode)...)
+
+	content := flushValidUTF8Prefix(&d.buf)
+	if content == "" {
+		return CompletionResponse{}, false
+	}
+	resp := CompletionResponse{Content: content, Logprobs: d.logprobs}
+	d.logprobs = nil
+	return resp, true
+}
+
+func buildLogprob(sample sampler.Result, decode func([]int32) string) []llm.Logprob {
+	if sample.Logprob == nil {
+		return nil
+	}
+	tok := func(id int32) string { return decode([]int32{id}) }
+
+	out := llm.Logprob{
+		TokenLogprob: llm.TokenLogprob{
+			Token:   tok(int32(sample.Token.Int())),
+			Logprob: float64(sample.Logprob.Floats()[0]),
+		},
 	}
 
-	return flushValidUTF8Prefix(b)
+	if sample.TopTokens != nil {
+		ids := sample.TopTokens.Ints()
+		vals := sample.TopLogprobs.Floats()
+		pairs := make([]llm.TokenLogprob, len(ids))
+		for i, id := range ids {
+			pairs[i] = llm.TokenLogprob{
+				Token:   tok(int32(id)),
+				Logprob: float64(vals[i]),
+			}
+		}
+		sort.Slice(pairs, func(i, j int) bool {
+			return pairs[i].Logprob > pairs[j].Logprob
+		})
+		out.TopLogprobs = pairs
+	}
+	return []llm.Logprob{out}
 }

x/mlxrunner/runner.go

@@ -18,34 +18,20 @@ import (
 	"github.com/ollama/ollama/x/tokenizer"
 )
 
+// Request is a short-lived struct that carries a completion request through
+// a channel from the HTTP handler to the runner goroutine. The ctx field
+// must travel with the request so that cancellation propagates across the
+// channel boundary.
 type Request struct {
-	TextCompletionsRequest
+	CompletionRequest
 	Responses chan CompletionResponse
-	Pipeline  func(Request) error
-
-	Ctx context.Context
+	Pipeline  func(context.Context, Request) error
 
+	Ctx     context.Context //nolint:containedctx
+	Tokens  []int32
 	Sampler *sample.Sampler
 }
 
-type TextCompletionsRequest struct {
-	Prompt  string `json:"prompt"`
-	Options struct {
-		Temperature      float32 `json:"temperature"`
-		TopP             float32 `json:"top_p"`
-		MinP             float32 `json:"min_p"`
-		TopK             int     `json:"top_k"`
-		RepeatLastN      int     `json:"repeat_last_n"`
-		RepeatPenalty    float32 `json:"repeat_penalty"`
-		PresencePenalty  float32 `json:"presence_penalty"`
-		FrequencyPenalty float32 `json:"frequency_penalty"`
-		MaxTokens        int     `json:"max_tokens"`
-
-		// Deprecated: use MaxTokens instead
-		NumPredict int `json:"num_predict"`
-	} `json:"options"`
-}
-
 type Runner struct {
 	Model         base.Model
 	Tokenizer     *tokenizer.Tokenizer
@@ -149,7 +135,7 @@ func (r *Runner) Run(host, port string, mux http.Handler) error {
 			case <-ctx.Done():
 				return nil
 			case request := <-r.Requests:
-				if err := request.Pipeline(request); err != nil {
+				if err := request.Pipeline(request.Ctx, request); err != nil {
 					slog.Info("Request terminated", "error", err)
 					var statusErr api.StatusError
 					if !errors.As(err, &statusErr) {

x/mlxrunner/sample/logprob_test.go

@@ -0,0 +1,249 @@
+//go:build mlx
+
+package sample
+
+import (
+	"math"
+	"sort"
+	"testing"
+
+	"github.com/ollama/ollama/x/mlxrunner/mlx"
+)
+
+// logprobEntry is the (token id, logprob) pair returned by the sampler's
+// top-K extraction, used after the test-side descending sort.
+type logprobEntry struct {
+	id      int
+	logprob float64
+}
+
+// runSampleLogprobs drives Sample on a fresh Sampler configured for logprobs
+// and returns the greedily-sampled token id, its logprob, and the top-K
+// entries sorted descending by logprob. Logits must be a [vocab]-shaped
+// slice; the helper reshapes it to [1, vocab] before calling the sampler.
+func runSampleLogprobs(t *testing.T, logits []float32, topK int) (int, float64, []logprobEntry) {
+	t.Helper()
+
+	s := New(Options{Logprobs: true, TopLogprobs: topK})
+	defer func() {
+		s.Free()
+		mlx.Sweep()
+	}()
+
+	tensor := mlx.FromValues(logits, 1, len(logits))
+	res := s.Sample(tensor)
+
+	mlx.Pin(res.Arrays()...)
+	defer mlx.Unpin(res.Arrays()...)
+	mlx.Sweep()
+	mlx.Eval(res.Arrays()...)
+
+	selected := res.Token.Int()
+	selLP := float64(res.Logprob.Floats()[0])
+
+	var top []logprobEntry
+	if topK > 0 && res.TopTokens != nil {
+		ids := res.TopTokens.Ints()
+		vals := res.TopLogprobs.Floats()
+		top = make([]logprobEntry, len(ids))
+		for i, id := range ids {
+			top[i] = logprobEntry{id: id, logprob: float64(vals[i])}
+		}
+		sort.Slice(top, func(i, j int) bool { return top[i].logprob > top[j].logprob })
+	}
+	return selected, selLP, top
+}
+
+func TestSampleLogprobsBasic(t *testing.T) {
+	tests := []struct {
+		name           string
+		logits         []float32
+		topK           int
+		wantSelectedID int
+		wantTopLen     int
+	}{
+		{
+			name:           "single token without top logprobs",
+			logits:         []float32{1.0, 0.5, 0.3, 0.1},
+			topK:           0,
+			wantSelectedID: 0,
+			wantTopLen:     0,
+		},
+		{
+			name:           "single token with top logprobs",
+			logits:         []float32{1.0, 0.5, 0.3, 0.1},
+			topK:           3,
+			wantSelectedID: 0,
+			wantTopLen:     3,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			selected, _, top := runSampleLogprobs(t, tt.logits, tt.topK)
+			if selected != tt.wantSelectedID {
+				t.Errorf("selected = %d, want %d", selected, tt.wantSelectedID)
+			}
+			if len(top) != tt.wantTopLen {
+				t.Errorf("top-K length = %d, want %d", len(top), tt.wantTopLen)
+			}
+		})
+	}
+}
+
+func TestSampleLogprobsNumericalStability(t *testing.T) {
+	logits := []float32{1000.0, 999.0, 998.0}
+	_, selLP, top := runSampleLogprobs(t, logits, 3)
+
+	if math.IsInf(selLP, 0) || math.IsNaN(selLP) {
+		t.Errorf("selected logprob is not finite: %f", selLP)
+	}
+	for i, e := range top {
+		if math.IsInf(e.logprob, 0) || math.IsNaN(e.logprob) {
+			t.Errorf("top[%d] logprob is not finite: %f", i, e.logprob)
+		}
+	}
+	for i := 1; i < len(top); i++ {
+		if top[i].logprob > top[i-1].logprob {
+			t.Errorf("top logprobs not descending: %f > %f", top[i].logprob, top[i-1].logprob)
+		}
+	}
+}
+
+func TestSampleLogprobsProbabilityCorrectness(t *testing.T) {
+	tests := []struct {
+		name   string
+		logits []float32
+	}{
+		{"uniform", []float32{1.0, 1.0, 1.0, 1.0}},
+		{"different", []float32{2.0, 1.0, 0.5, 0.1}},
+		{"negative", []float32{-1.0, -2.0, -3.0, -4.0}},
+		{"mixed", []float32{5.0, -5.0, 0.0, 2.5}},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			selected, selLP, top := runSampleLogprobs(t, tt.logits, len(tt.logits))
+
+			if selLP > 0 {
+				t.Errorf("selected logprob should be <= 0, got %f", selLP)
+			}
+			for i, e := range top {
+				if e.logprob > 0 {
+					t.Errorf("top[%d] logprob should be <= 0, got %f", i, e.logprob)
+				}
+			}
+
+			if tt.name == "uniform" {
+				want := 1.0 / float64(len(tt.logits))
+				got := math.Exp(selLP)
+				if math.Abs(got-want) > 1e-6 {
+					t.Errorf("uniform logits: selected prob = %f, want %f", got, want)
+				}
+			}
+
+			for i := 1; i < len(top); i++ {
+				if top[i].logprob > top[i-1].logprob {
+					t.Errorf("top logprobs not descending at %d: %f > %f",
+						i, top[i].logprob, top[i-1].logprob)
+				}
+			}
+
+			found := false
+			for _, e := range top {
+				if e.id == selected {
+					found = true
+					if math.Abs(e.logprob-selLP) > 1e-6 {
+						t.Errorf("selected logprob mismatch: selLP=%f top=%f", selLP, e.logprob)
+					}
+					break
+				}
+			}
+			if !found {
+				t.Errorf("selected token %d not present in top-K", selected)
+			}
+		})
+	}
+}
+
+func TestSampleLogprobsSoftmaxCorrectness(t *testing.T) {
+	tests := []struct {
+		name   string
+		logits []float32
+	}{
+		{"small vocabulary", []float32{1.0, 2.0, 3.0}},
+		{"large differences", []float32{10.0, 0.0, -10.0}},
+		{"all equal", []float32{5.0, 5.0, 5.0, 5.0, 5.0}},
+		{"very large values", []float32{500.0, 499.0, 498.0}},
+		{"very small values", []float32{-500.0, -499.0, -498.0}},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			_, _, top := runSampleLogprobs(t, tt.logits, len(tt.logits))
+			if len(top) != len(tt.logits) {
+				t.Fatalf("top-K length = %d, want %d", len(top), len(tt.logits))
+			}
+
+			var sum float64
+			for _, e := range top {
+				p := math.Exp(e.logprob)
+				if p < 0 || p > 1 {
+					t.Errorf("token %d: probability %f out of [0,1]", e.id, p)
+				}
+				sum += p
+			}
+
+			if math.Abs(sum-1.0) > 1e-5 {
+				t.Errorf("probabilities sum = %f, want 1.0", sum)
+			}
+		})
+	}
+}
+
+func TestSampleLogprobsSelectedTokenCorrectness(t *testing.T) {
+	logits := []float32{3.0, 1.0, 2.0, 0.5}
+
+	maxIdx := 0
+	for i, v := range logits[1:] {
+		if v > logits[maxIdx] {
+			maxIdx = i + 1
+		}
+	}
+
+	selected, selLP, top := runSampleLogprobs(t, logits, len(logits))
+
+	if selected != maxIdx {
+		t.Errorf("selected = %d, want argmax %d", selected, maxIdx)
+	}
+
+	if top[0].id != maxIdx {
+		t.Errorf("top[0].id = %d, want argmax %d", top[0].id, maxIdx)
+	}
+	if math.Abs(top[0].logprob-selLP) > 1e-6 {
+		t.Errorf("top[0].logprob = %f, want selected %f", top[0].logprob, selLP)
+	}
+}
+
+func TestSampleLogprobsTopKOrdering(t *testing.T) {
+	// Logits chosen so argmax order differs from index order.
+	logits := []float32{2.0, 5.0, 1.0, 4.0, 3.0}
+	wantOrder := []int{1, 3, 4, 0, 2}
+
+	_, _, top := runSampleLogprobs(t, logits, len(logits))
+
+	if len(top) != len(wantOrder) {
+		t.Fatalf("top-K length = %d, want %d", len(top), len(wantOrder))
+	}
+	for i, e := range top {
+		if e.id != wantOrder[i] {
+			t.Errorf("top[%d].id = %d, want %d", i, e.id, wantOrder[i])
+		}
+	}
+	for i := 1; i < len(top); i++ {
+		if top[i].logprob > top[i-1].logprob {
+			t.Errorf("top[%d].logprob (%f) > top[%d].logprob (%f)",
+				i, top[i].logprob, i-1, top[i-1].logprob)
+		}
+	}
+}

x/mlxrunner/sample/sample.go

@@ -8,7 +8,7 @@ import (
 
 type Transform func(*Sampler, *mlx.Array) *mlx.Array
 
-type Sampler struct {
+type Options struct {
 	Temperature      float32
 	TopP             float32
 	MinP             float32
@@ -18,45 +18,66 @@ type Sampler struct {
 	PresencePenalty  float32
 	FrequencyPenalty float32
 
+	// Logprobs causes Sample to populate Result.Logprob with the selected
+	// token's log-probability. TopLogprobs (when > 0) adds top-K pairs.
+	Logprobs    bool
+	TopLogprobs int
+}
+
+type Sampler struct {
+	Options
+
 	history    *mlx.Array
 	historyLen int
 	transforms []Transform
 }
 
-func New(temp, top_p, min_p float32, top_k, repeatLastN int, repeatPenalty, presencePenalty, frequencyPenalty float32) *Sampler {
-	if repeatPenalty <= 0 {
-		repeatPenalty = 1
+// Result bundles the outputs of one decode step. The logprob tensors are
+// populated only when the sampler is configured to report them.
+type Result struct {
+	Token       *mlx.Array // sampled token id, shape [B]
+	Logprob     *mlx.Array // sampled-token logprob, shape [B,1]; nil unless Logprobs
+	TopTokens   *mlx.Array // top-K token ids, shape [B,K]; nil unless TopLogprobs > 0
+	TopLogprobs *mlx.Array // top-K logprobs, shape [B,K]; nil unless TopLogprobs > 0
+}
+
+// Arrays returns the tensor fields as a slice so callers can drive the mlx
+// lifecycle verbs (Pin, Unpin, Eval, AsyncEval) over the whole group. Unset
+// fields stay nil; the mlx helpers skip them.
+func (r Result) Arrays() []*mlx.Array {
+	return []*mlx.Array{r.Token, r.Logprob, r.TopTokens, r.TopLogprobs}
+}
+
+func New(opts Options) *Sampler {
+	if opts.RepeatPenalty <= 0 {
+		opts.RepeatPenalty = 1
 	}
 
-	s := &Sampler{
-		Temperature:      temp,
-		TopP:             top_p,
-		MinP:             min_p,
-		TopK:             top_k,
-		RepeatLastN:      repeatLastN,
-		RepeatPenalty:    repeatPenalty,
-		PresencePenalty:  presencePenalty,
-		FrequencyPenalty: frequencyPenalty,
-	}
+	s := &Sampler{Options: opts}
 
 	var transforms []Transform
 	if s.usesHistory() {
 		transforms = append(transforms, penalty)
 	}
 
-	if top_p > 0 && top_p < 1 {
-		transforms = append(transforms, topP)
-	}
-
-	if min_p != 0 {
-		transforms = append(transforms, minP)
-	}
-
-	if top_k > 0 {
+	hasTopP := opts.TopP > 0 && opts.TopP < 1
+	hasTopK := opts.TopK > 0
+	switch {
+	case hasTopP:
+		// topKTopP always does a full descending sort for the top-P
+		// cumulative mask and opportunistically masks top-K during the
+		// same pass when it is also configured.
+		transforms = append(transforms, topKTopP)
+	case hasTopK:
+		// Argpartition (partial sort) is cheaper than a full sort.
 		transforms = append(transforms, topK)
 	}
 
-	if temp == 0 {
+	if opts.MinP != 0 {
+		transforms = append(transforms, minP)
+	}
+
+	if opts.Temperature == 0 {
 		transforms = append(transforms, greedy)
 	} else {
 		transforms = append(transforms, temperature)
@@ -123,76 +144,121 @@ func (s *Sampler) Free() {
 	s.setHistory(nil, 0)
 }
 
-func (s *Sampler) Sample(logits *mlx.Array) *mlx.Array {
+// Sample runs the configured transform chain on the raw per-token logits
+// and returns the sampled token id plus, when configured, the reported
+// log-probability tensors for the selected token and the top-K tokens.
+func (s *Sampler) Sample(logits *mlx.Array) Result {
+	scores := logits
 	for _, transform := range s.transforms {
-		logits = transform(s, logits)
+		scores = transform(s, scores)
 	}
-	return logits
-}
+	res := Result{Token: scores}
 
-func greedy(_ *Sampler, logits *mlx.Array) *mlx.Array {
-	return logits.Argmax(-1, false)
-}
-
-func temperature(s *Sampler, logits *mlx.Array) *mlx.Array {
-	return mlx.DivScalar(logits, s.Temperature).Categorical(-1)
-}
-
-func topP(s *Sampler, logits *mlx.Array) *mlx.Array {
-	if s.TopP <= 0 || s.TopP >= 1 {
-		return logits
+	if s.Logprobs {
+		// Compute log_softmax in fp32 and subtract the max before
+		// logsumexp so the final subtraction stays on small values.
+		// Otherwise it cancels two large numbers and loses precision.
+		lp := logits.AsType(mlx.DTypeFloat32)
+		lp = lp.Subtract(lp.MaxAxis(-1, true))
+		lp = lp.Subtract(lp.Logsumexp(true))
+		res.Logprob = lp.TakeAlongAxis(res.Token.ExpandDims(-1), -1)
+		if k := s.TopLogprobs; k > 0 {
+			if vocab := lp.Dim(lp.NumDims() - 1); k > vocab {
+				k = vocab
+			}
+			// Argpartition on the negated values places the K largest
+			// (unsorted) in positions [0:K].
+			idx := lp.Negative().ArgpartitionAxis(k-1, -1).Slice(mlx.Slice(), mlx.Slice(0, k))
+			res.TopTokens = idx.AsType(mlx.DTypeInt32)
+			res.TopLogprobs = lp.TakeAlongAxis(idx, -1)
+		}
 	}
+	return res
+}
 
-	order := logits.Negative().ArgsortAxis(-1)
-	sortedLogits := logits.TakeAlongAxis(order, -1)
-	sortedProbs := mlx.SoftmaxAxis(sortedLogits, -1, true)
-	prevCumProbs := sortedProbs.Cumsum(-1, false, true).Subtract(sortedProbs)
+func greedy(_ *Sampler, scores *mlx.Array) *mlx.Array {
+	return scores.Argmax(-1, false)
+}
+
+func temperature(s *Sampler, scores *mlx.Array) *mlx.Array {
+	return mlx.DivScalar(scores, s.Temperature).Categorical(-1)
+}
+
+// topKTopP applies top-P in a descending sort pass and, when top-K is also
+// configured, masks any surviving value below the K-th largest in the same
+// pass. Callers dispatch here whenever top-P is enabled — the top-K-only
+// case uses a cheaper partial sort via the topK transform.
+func topKTopP(s *Sampler, scores *mlx.Array) *mlx.Array {
+	vocab := scores.Dim(scores.NumDims() - 1)
+	applyTopK := s.TopK > 0 && s.TopK < vocab
+
+	order := scores.Negative().ArgsortAxis(-1)
+	sorted := scores.TakeAlongAxis(order, -1)
+	negInf := mlx.FromValue(float32(math.Inf(-1)))
+
+	// Top-P: in descending order, keep tokens whose exclusive cumulative
+	// probability is still below s.TopP.
+	probs := mlx.SoftmaxAxis(sorted, -1, true)
+	prevCumProbs := probs.Cumsum(-1, false, true).Subtract(probs)
 	keep := prevCumProbs.Less(mlx.FromValue(s.TopP))
-	filtered := mlx.Where(keep, sortedLogits, mlx.FromValue(float32(math.Inf(-1))))
-	return logits.PutAlongAxis(order, filtered, -1)
-}
+	sorted = mlx.Where(keep, sorted, negInf)
 
-func minP(s *Sampler, logits *mlx.Array) *mlx.Array {
-	if s.MinP <= 0 || s.MinP > 1 {
-		return logits
+	out := scores.PutAlongAxis(order, sorted, -1)
+
+	// Top-K: sorted is already in descending order, so positions [K, V)
+	// are the ones to drop. Scatter -inf through their original-layout
+	// indices (order[K:]). Positional (not value-based) so exactly K
+	// tokens survive — ties at the K-th logit get broken by the sort
+	// order rather than promoted through the filter.
+	if applyTopK {
+		dropOrder := order.Slice(mlx.Slice(), mlx.Slice(s.TopK, mlx.End))
+		out = out.PutAlongAxis(dropOrder, negInf, -1)
 	}
 
-	maxLogits := logits.TakeAlongAxis(logits.Argmax(-1, true), -1)
-	minLogits := mlx.AddScalar(maxLogits, float32(math.Log(float64(s.MinP))))
+	return out
+}
+
+func minP(s *Sampler, scores *mlx.Array) *mlx.Array {
+	if s.MinP <= 0 || s.MinP > 1 {
+		return scores
+	}
+
+	maxScore := scores.MaxAxis(-1, true)
+	threshold := mlx.AddScalar(maxScore, float32(math.Log(float64(s.MinP))))
 
 	return mlx.Where(
-		logits.Less(minLogits),
+		scores.Less(threshold),
 		mlx.FromValue(float32(math.Inf(-1))),
-		logits,
+		scores,
 	)
 }
 
-func topK(s *Sampler, logits *mlx.Array) *mlx.Array {
+func topK(s *Sampler, scores *mlx.Array) *mlx.Array {
 	if s.TopK <= 0 {
-		return logits
+		return scores
 	}
 
-	vocab := logits.Dim(logits.NumDims() - 1)
+	vocab := scores.Dim(scores.NumDims() - 1)
 	if s.TopK >= vocab {
-		return logits
+		return scores
 	}
 
-	mask := logits.Negative().ArgpartitionAxis(s.TopK-1, -1).Slice(mlx.Slice(), mlx.Slice(s.TopK, mlx.End))
-	return logits.PutAlongAxis(mask, mlx.FromValue(float32(math.Inf(-1))), -1)
+	mask := scores.Negative().ArgpartitionAxis(s.TopK-1, -1).Slice(mlx.Slice(), mlx.Slice(s.TopK, mlx.End))
+	return scores.PutAlongAxis(mask, mlx.FromValue(float32(math.Inf(-1))), -1)
 }
 
-func penalty(s *Sampler, logits *mlx.Array) *mlx.Array {
+func penalty(s *Sampler, scores *mlx.Array) *mlx.Array {
 	if s.historyLen == 0 {
-		return logits
+		return scores
 	}
 
 	tokenIndices := s.history
-	if logits.NumDims() > 1 {
+	if scores.NumDims() > 1 {
 		tokenIndices = tokenIndices.ExpandDims(0)
 	}
 
 	if s.RepeatPenalty != 1 || s.PresencePenalty != 0 {
-		adjusted := logits.TakeAlongAxis(tokenIndices, -1)
+		adjusted := scores.TakeAlongAxis(tokenIndices, -1)
 		if s.RepeatPenalty != 1 {
 			factor := mlx.Where(
 				adjusted.Less(mlx.FromValue(float32(0))),
@@ -204,12 +270,12 @@ func penalty(s *Sampler, logits *mlx.Array) *mlx.Array {
 		if s.PresencePenalty != 0 {
 			adjusted = mlx.AddScalar(adjusted, -s.PresencePenalty)
 		}
-		logits = logits.PutAlongAxis(tokenIndices, adjusted, -1)
+		scores = scores.PutAlongAxis(tokenIndices, adjusted, -1)
 	}
 
 	if s.FrequencyPenalty != 0 {
-		logits = logits.ScatterAddAxis(tokenIndices, mlx.FromValue(-s.FrequencyPenalty), -1)
+		scores = scores.ScatterAddAxis(tokenIndices, mlx.FromValue(-s.FrequencyPenalty), -1)
 	}
 
-	return logits
+	return scores
 }

x/mlxrunner/sample/sample_test.go

@@ -10,8 +10,7 @@ import (
 )
 
 func TestPresencePenaltyUsesAppendedTokenImmediately(t *testing.T) {
-	// RepeatLastN = 1, PresencePenalty = 6
-	s := New(0, 0, 0, 0, 1, 1, 6, 0)
+	s := New(Options{RepeatLastN: 1, PresencePenalty: 6})
 	defer func() {
 		s.Free()
 		mlx.Sweep()
@@ -21,7 +20,7 @@ func TestPresencePenaltyUsesAppendedTokenImmediately(t *testing.T) {
 	s.AppendToken(mlx.NewArrayInt32([]int32{1}, []int32{1}))
 
 	logits := mlx.FromValues([]float32{0, 5, 4}, 3)
-	got := s.Sample(logits)
+	got := s.Sample(logits).Token
 	mlx.Eval(got)
 
 	// logits will be [0, -1, 4] after the penalty
@@ -33,7 +32,7 @@ func TestPresencePenaltyUsesAppendedTokenImmediately(t *testing.T) {
 }
 
 func TestRepeatPenaltyUsesHistoryWithoutPresencePenalty(t *testing.T) {
-	s := New(0, 0, 0, 0, 1, 2, 0, 0)
+	s := New(Options{RepeatLastN: 1, RepeatPenalty: 2})
 	defer func() {
 		s.Free()
 		mlx.Sweep()
@@ -42,7 +41,7 @@ func TestRepeatPenaltyUsesHistoryWithoutPresencePenalty(t *testing.T) {
 	s.ResetHistory([]int32{1})
 
 	logits := mlx.FromValues([]float32{0, 5, 4}, 3)
-	got := s.Sample(logits)
+	got := s.Sample(logits).Token
 	mlx.Eval(got)
 
 	// token 1 is repeated and positive, so 5 / 2 falls below token 2.
@@ -53,7 +52,7 @@ func TestRepeatPenaltyUsesHistoryWithoutPresencePenalty(t *testing.T) {
 }
 
 func TestFrequencyPenaltyUsesTokenCounts(t *testing.T) {
-	s := New(0, 0, 0, 0, 4, 1, 0, 2)
+	s := New(Options{RepeatLastN: 4, FrequencyPenalty: 2})
 	defer func() {
 		s.Free()
 		mlx.Sweep()
@@ -62,7 +61,7 @@ func TestFrequencyPenaltyUsesTokenCounts(t *testing.T) {
 	s.ResetHistory([]int32{1, 1})
 
 	logits := mlx.FromValues([]float32{0, 5, 4}, 3)
-	got := s.Sample(logits)
+	got := s.Sample(logits).Token
 	mlx.Eval(got)
 
 	// token 1 appears twice, so 5 - (2 * 2) falls below token 2.
@@ -73,7 +72,7 @@ func TestFrequencyPenaltyUsesTokenCounts(t *testing.T) {
 }
 
 func TestMinPMasksTokensBelowThreshold(t *testing.T) {
-	s := New(0, 0, 0.5, 0, 0, 1, 0, 0)
+	s := New(Options{MinP: 0.5})
 	defer func() {
 		s.Free()
 		mlx.Sweep()

x/mlxrunner/server.go

@@ -2,7 +2,6 @@ package mlxrunner
 
 import (
 	"bytes"
-	"cmp"
 	"context"
 	"encoding/json"
 	"flag"
@@ -87,25 +86,30 @@ func Execute(args []string) error {
 	mux.HandleFunc("POST /v1/completions", func(w http.ResponseWriter, r *http.Request) {
 		request := Request{Responses: make(chan CompletionResponse)}
 
-		if err := json.NewDecoder(r.Body).Decode(&request.TextCompletionsRequest); err != nil {
+		if err := json.NewDecoder(r.Body).Decode(&request.CompletionRequest); err != nil {
 			slog.Error("Failed to decode request", "error", err)
 			http.Error(w, "Bad Request", http.StatusBadRequest)
 			return
 		}
 
-		request.Options.MaxTokens = cmp.Or(request.Options.MaxTokens, request.Options.NumPredict)
-
 		request.Pipeline = runner.TextGenerationPipeline
-		request.Sampler = sample.New(
-			request.Options.Temperature,
-			request.Options.TopP,
-			request.Options.MinP,
-			request.Options.TopK,
-			request.Options.RepeatLastN,
-			request.Options.RepeatPenalty,
-			request.Options.PresencePenalty,
-			request.Options.FrequencyPenalty,
-		)
+		request.Sampler = sample.New(sample.Options{
+			Temperature:      request.Options.Temperature,
+			TopP:             request.Options.TopP,
+			MinP:             request.Options.MinP,
+			TopK:             request.Options.TopK,
+			RepeatLastN:      request.Options.RepeatLastN,
+			RepeatPenalty:    request.Options.RepeatPenalty,
+			PresencePenalty:  request.Options.PresencePenalty,
+			FrequencyPenalty: request.Options.FrequencyPenalty,
+			Logprobs:         request.Logprobs,
+			TopLogprobs:      request.TopLogprobs,
+		})
+
+		if err := runner.Prepare(&request); err != nil {
+			http.Error(w, err.Error(), http.StatusBadRequest)
+			return
+		}
 
 		var cancel context.CancelFunc
 		request.Ctx, cancel = context.WithCancel(r.Context())

x/models/gemma4/gemma4_moe_test.go

@@ -144,6 +144,8 @@ func TestRouterForwardMatchesLegacy(t *testing.T) {
 
 	gotScores, gotInds := r.Forward(x, cfg)
 	wantScores, wantInds := legacyRouterForward(r, x, cfg)
+	gotInds = gotInds.AsType(mlx.DTypeInt32)
+	wantInds = wantInds.AsType(mlx.DTypeInt32)
 	mlx.Eval(gotScores, gotInds, wantScores, wantInds)
 
 	if got, want := gotInds.Ints(), wantInds.Ints(); !intSlicesEqual(got, want) {

x/models/nn/nn_test.go

@@ -169,8 +169,8 @@ func TestQuantizedLinearMXFP4MatchesDequantizedWeight(t *testing.T) {
 	dequantizedWeight := mlx.Dequantize(ql.Weight, ql.Scales, ql.QBiases, 32, 4, "mxfp4")
 	mlx.Eval(dequantizedWeight)
 
-	qOut := ql.Forward(input)
-	dOut := NewLinear(dequantizedWeight, nil).Forward(input)
+	qOut := ql.Forward(input).AsType(mlx.DTypeFloat32)
+	dOut := NewLinear(dequantizedWeight, nil).Forward(input).AsType(mlx.DTypeFloat32)
 	mlx.Eval(qOut, dOut)
 
 	got := qOut.Floats()