add min vram for glm-4.7-flash

Show recommended models (glm-4.7-flash, glm-4.7:cloud, kimi-k2.5:cloud)
in the model selector. Non-installed models are labeled "recommended,
install?" and sorted to the bottom. Selecting a non-installed model
prompts for confirmation before pulling. Strip :latest suffix from
display names. Update docs to match new recommendations.

cmd/config/integrations.go

@@ -13,6 +13,7 @@ import (
 	"time"
 
 	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/progress"
 	"github.com/spf13/cobra"
 )
 
@@ -49,6 +50,14 @@ var integrations = map[string]Runner{
 	"openclaw": &Openclaw{},
 }
 
+// recommendedModels are shown when the user has no models or as suggestions.
+// Order matters: local models first, then cloud models.
+var recommendedModels = []selectItem{
+	{Name: "glm-4.7-flash", Description: "Recommended (requires ~25GB VRAM)"},
+	{Name: "glm-4.7:cloud", Description: "recommended"},
+	{Name: "kimi-k2.5:cloud", Description: "recommended"},
+}
+
 // integrationAliases are hidden from the interactive selector but work as CLI arguments.
 var integrationAliases = map[string]bool{
 	"clawdbot": true,
@@ -94,62 +103,25 @@ func selectModels(ctx context.Context, name, current string) ([]string, error) {
 		return nil, err
 	}
 
-	if len(models.Models) == 0 {
-		return nil, fmt.Errorf("no models available, run 'ollama pull <model>' first")
-	}
-
-	var items []selectItem
-	cloudModels := make(map[string]bool)
+	var existing []modelInfo
 	for _, m := range models.Models {
-		if m.RemoteModel != "" {
-			cloudModels[m.Name] = true
-		}
-		items = append(items, selectItem{Name: m.Name})
+		existing = append(existing, modelInfo{Name: m.Name, Remote: m.RemoteModel != ""})
 	}
 
-	if len(items) == 0 {
-		return nil, fmt.Errorf("no local models available, run 'ollama pull <model>' first")
-	}
-
-	// Get previously configured models (saved config takes precedence)
 	var preChecked []string
 	if saved, err := loadIntegration(name); err == nil {
 		preChecked = saved.Models
 	} else if editor, ok := r.(Editor); ok {
 		preChecked = editor.Models()
 	}
-	checked := make(map[string]bool, len(preChecked))
-	for _, n := range preChecked {
-		checked[n] = true
-	}
 
-	// Resolve current to full name (e.g., "llama3.2" -> "llama3.2:latest")
-	for _, item := range items {
-		if item.Name == current || strings.HasPrefix(item.Name, current+":") {
-			current = item.Name
-			break
-		}
-	}
+	items, preChecked, existingModels, cloudModels := buildModelList(existing, preChecked, current)
 
-	// If current model is configured, move to front of preChecked
-	if checked[current] {
-		preChecked = append([]string{current}, slices.DeleteFunc(preChecked, func(m string) bool { return m == current })...)
+	if len(items) == 0 {
+		return nil, fmt.Errorf("no models available")
 	}
 
-	// Sort: checked first, then alphabetical
-	slices.SortFunc(items, func(a, b selectItem) int {
-		ac, bc := checked[a.Name], checked[b.Name]
-		if ac != bc {
-			if ac {
-				return -1
-			}
-			return 1
-		}
-		return strings.Compare(strings.ToLower(a.Name), strings.ToLower(b.Name))
-	})
-
 	var selected []string
-	// only editors support multi-model selection
 	if _, ok := r.(Editor); ok {
 		selected, err = multiSelectPrompt(fmt.Sprintf("Select models for %s:", r), items, preChecked)
 		if err != nil {
@@ -163,7 +135,27 @@ func selectModels(ctx context.Context, name, current string) ([]string, error) {
 		selected = []string{model}
 	}
 
-	// if any model in selected is a cloud model, ensure signed in
+	var toPull []string
+	for _, m := range selected {
+		if !existingModels[m] {
+			toPull = append(toPull, m)
+		}
+	}
+	if len(toPull) > 0 {
+		msg := fmt.Sprintf("Download %s?", strings.Join(toPull, ", "))
+		if ok, err := confirmPrompt(msg); err != nil {
+			return nil, err
+		} else if !ok {
+			return nil, errCancelled
+		}
+		for _, m := range toPull {
+			fmt.Fprintf(os.Stderr, "\n")
+			if err := pullModel(ctx, client, m); err != nil {
+				return nil, fmt.Errorf("failed to pull %s: %w", m, err)
+			}
+		}
+	}
+
 	var selectedCloudModels []string
 	for _, m := range selected {
 		if cloudModels[m] {
@@ -286,7 +278,6 @@ Examples:
 				return fmt.Errorf("unknown integration: %s", name)
 			}
 
-			// If launching without --model, use saved config if available
 			if !configFlag && modelFlag == "" {
 				if config, err := loadIntegration(name); err == nil && len(config.Models) > 0 {
 					return runIntegration(name, config.Models[0])
@@ -295,7 +286,6 @@ Examples:
 
 			var models []string
 			if modelFlag != "" {
-				// When --model is specified, merge with existing models (new model becomes default)
 				models = []string{modelFlag}
 				if existing, err := loadIntegration(name); err == nil && len(existing.Models) > 0 {
 					for _, m := range existing.Models {
@@ -364,3 +354,150 @@ Examples:
 	cmd.Flags().BoolVar(&configFlag, "config", false, "Configure without launching")
 	return cmd
 }
+
+type modelInfo struct {
+	Name   string
+	Remote bool
+}
+
+// buildModelList merges existing models with recommendations, sorts them, and returns
+// the ordered items along with maps of existing and cloud model names.
+func buildModelList(existing []modelInfo, preChecked []string, current string) (items []selectItem, orderedChecked []string, existingModels, cloudModels map[string]bool) {
+	existingModels = make(map[string]bool)
+	cloudModels = make(map[string]bool)
+	recommended := make(map[string]bool)
+	var hasLocalModel, hasCloudModel bool
+
+	for _, rec := range recommendedModels {
+		recommended[rec.Name] = true
+	}
+
+	for _, m := range existing {
+		existingModels[m.Name] = true
+		if m.Remote {
+			cloudModels[m.Name] = true
+			hasCloudModel = true
+		} else {
+			hasLocalModel = true
+		}
+		displayName := strings.TrimSuffix(m.Name, ":latest")
+		existingModels[displayName] = true
+		item := selectItem{Name: displayName}
+		if recommended[displayName] {
+			item.Description = "recommended"
+		}
+		items = append(items, item)
+	}
+
+	for _, rec := range recommendedModels {
+		if existingModels[rec.Name] || existingModels[rec.Name+":latest"] {
+			continue
+		}
+		items = append(items, rec)
+		if isCloudModel(rec.Name) {
+			cloudModels[rec.Name] = true
+		}
+	}
+
+	checked := make(map[string]bool, len(preChecked))
+	for _, n := range preChecked {
+		checked[n] = true
+	}
+
+	// Resolve current to full name (e.g., "llama3.2" -> "llama3.2:latest")
+	for _, item := range items {
+		if item.Name == current || strings.HasPrefix(item.Name, current+":") {
+			current = item.Name
+			break
+		}
+	}
+
+	if checked[current] {
+		preChecked = append([]string{current}, slices.DeleteFunc(preChecked, func(m string) bool { return m == current })...)
+	}
+
+	// Non-existing models get "install?" suffix and are pushed to the bottom.
+	// When user has no models, preserve recommended order.
+	notInstalled := make(map[string]bool)
+	for i := range items {
+		if !existingModels[items[i].Name] {
+			notInstalled[items[i].Name] = true
+			items[i].Description = "recommended, install?"
+		}
+	}
+
+	if hasLocalModel || hasCloudModel {
+		slices.SortStableFunc(items, func(a, b selectItem) int {
+			ac, bc := checked[a.Name], checked[b.Name]
+			aNew, bNew := notInstalled[a.Name], notInstalled[b.Name]
+
+			if ac != bc {
+				if ac {
+					return -1
+				}
+				return 1
+			}
+			if !ac && !bc && aNew != bNew {
+				if aNew {
+					return 1
+				}
+				return -1
+			}
+			return strings.Compare(strings.ToLower(a.Name), strings.ToLower(b.Name))
+		})
+	}
+
+	return items, preChecked, existingModels, cloudModels
+}
+
+func isCloudModel(name string) bool {
+	return strings.HasSuffix(name, ":cloud")
+}
+
+func pullModel(ctx context.Context, client *api.Client, model string) error {
+	p := progress.NewProgress(os.Stderr)
+	defer p.Stop()
+
+	bars := make(map[string]*progress.Bar)
+	var status string
+	var spinner *progress.Spinner
+
+	fn := func(resp api.ProgressResponse) error {
+		if resp.Digest != "" {
+			if resp.Completed == 0 {
+				return nil
+			}
+
+			if spinner != nil {
+				spinner.Stop()
+			}
+
+			bar, ok := bars[resp.Digest]
+			if !ok {
+				name, isDigest := strings.CutPrefix(resp.Digest, "sha256:")
+				name = strings.TrimSpace(name)
+				if isDigest {
+					name = name[:min(12, len(name))]
+				}
+				bar = progress.NewBar(fmt.Sprintf("pulling %s:", name), resp.Total, resp.Completed)
+				bars[resp.Digest] = bar
+				p.Add(resp.Digest, bar)
+			}
+
+			bar.Set(resp.Completed)
+		} else if status != resp.Status {
+			if spinner != nil {
+				spinner.Stop()
+			}
+
+			status = resp.Status
+			spinner = progress.NewSpinner(status)
+			p.Add(status, spinner)
+		}
+
+		return nil
+	}
+
+	request := api.PullRequest{Name: model}
+	return client.Pull(ctx, &request, fn)
+}

cmd/config/integrations_test.go

@@ -5,6 +5,7 @@ import (
 	"strings"
 	"testing"
 
+	"github.com/google/go-cmp/cmp"
 	"github.com/spf13/cobra"
 )
 
@@ -174,15 +175,226 @@ func TestLaunchCmd_NilHeartbeat(t *testing.T) {
 func TestAllIntegrations_HaveRequiredMethods(t *testing.T) {
 	for name, r := range integrations {
 		t.Run(name, func(t *testing.T) {
-			// Test String() doesn't panic and returns non-empty
 			displayName := r.String()
 			if displayName == "" {
 				t.Error("String() should not return empty")
 			}
-
-			// Test Run() exists (we can't call it without actually running the command)
-			// Just verify the method is available
 			var _ func(string) error = r.Run
 		})
 	}
 }
+
+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)
+			}
+		})
+	}
+}
+
+func names(items []selectItem) []string {
+	var out []string
+	for _, item := range items {
+		out = append(out, item.Name)
+	}
+	return out
+}
+
+func TestBuildModelList_NoExistingModels(t *testing.T) {
+	items, _, _, _ := buildModelList(nil, nil, "")
+
+	want := []string{"glm-4.7-flash", "glm-4.7:cloud", "kimi-k2.5:cloud"}
+	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)
+	}
+
+	for _, item := range items {
+		if item.Description != "recommended, install?" {
+			t.Errorf("item %q should have description 'install?', got %q", item.Name, item.Description)
+		}
+	}
+}
+
+func TestBuildModelList_OnlyLocalModels_CloudRecsAtBottom(t *testing.T) {
+	existing := []modelInfo{
+		{Name: "llama3.2:latest", Remote: false},
+		{Name: "qwen2.5:latest", Remote: false},
+	}
+
+	items, _, _, _ := buildModelList(existing, nil, "")
+	got := names(items)
+
+	want := []string{"llama3.2", "qwen2.5", "glm-4.7-flash", "glm-4.7:cloud", "kimi-k2.5:cloud"}
+	if diff := cmp.Diff(want, got); diff != "" {
+		t.Errorf("cloud recs should be at bottom (-want +got):\n%s", diff)
+	}
+}
+
+func TestBuildModelList_BothCloudAndLocal_RegularSort(t *testing.T) {
+	existing := []modelInfo{
+		{Name: "llama3.2:latest", Remote: false},
+		{Name: "glm-4.7:cloud", Remote: true},
+	}
+
+	items, _, _, _ := buildModelList(existing, nil, "")
+	got := names(items)
+
+	want := []string{"glm-4.7:cloud", "llama3.2", "glm-4.7-flash", "kimi-k2.5:cloud"}
+	if diff := cmp.Diff(want, got); diff != "" {
+		t.Errorf("mixed models should be alphabetical (-want +got):\n%s", diff)
+	}
+}
+
+func TestBuildModelList_PreCheckedFirst(t *testing.T) {
+	existing := []modelInfo{
+		{Name: "llama3.2:latest", Remote: false},
+		{Name: "glm-4.7:cloud", Remote: true},
+	}
+
+	items, _, _, _ := buildModelList(existing, []string{"llama3.2"}, "")
+	got := names(items)
+
+	if got[0] != "llama3.2" {
+		t.Errorf("pre-checked model should be first, got %v", got)
+	}
+}
+
+func TestBuildModelList_ExistingRecommendedMarked(t *testing.T) {
+	existing := []modelInfo{
+		{Name: "glm-4.7-flash", Remote: false},
+		{Name: "glm-4.7:cloud", Remote: true},
+	}
+
+	items, _, _, _ := buildModelList(existing, nil, "")
+
+	for _, item := range items {
+		switch item.Name {
+		case "glm-4.7-flash", "glm-4.7:cloud":
+			if item.Description != "recommended" {
+				t.Errorf("installed recommended %q should have description 'recommended', got %q", item.Name, item.Description)
+			}
+		case "kimi-k2.5:cloud":
+			if item.Description != "recommended, install?" {
+				t.Errorf("non-installed recommended %q should have description 'install?', got %q", item.Name, item.Description)
+			}
+		}
+	}
+}
+
+func TestBuildModelList_ExistingCloudModelsNotPushedToBottom(t *testing.T) {
+	existing := []modelInfo{
+		{Name: "glm-4.7-flash", Remote: false},
+		{Name: "glm-4.7:cloud", Remote: true},
+	}
+
+	items, _, _, _ := buildModelList(existing, nil, "")
+	got := names(items)
+
+	// glm-4.7-flash and glm-4.7:cloud are installed so they sort normally;
+	// kimi-k2.5:cloud and qwen3:0.6b are not installed so they go to the bottom
+	want := []string{"glm-4.7-flash", "glm-4.7:cloud", "kimi-k2.5:cloud"}
+	if diff := cmp.Diff(want, got); diff != "" {
+		t.Errorf("existing cloud models should sort normally (-want +got):\n%s", diff)
+	}
+}
+
+func TestBuildModelList_HasRecommendedCloudModel_OnlyNonInstalledAtBottom(t *testing.T) {
+	existing := []modelInfo{
+		{Name: "llama3.2:latest", Remote: false},
+		{Name: "kimi-k2.5:cloud", Remote: true},
+	}
+
+	items, _, _, _ := buildModelList(existing, nil, "")
+	got := names(items)
+
+	// kimi-k2.5:cloud is installed so it sorts normally;
+	// the rest of the recommendations are not installed so they go to the bottom
+	want := []string{"kimi-k2.5:cloud", "llama3.2", "glm-4.7-flash", "glm-4.7:cloud"}
+	if diff := cmp.Diff(want, got); diff != "" {
+		t.Errorf("only non-installed models should be at bottom (-want +got):\n%s", diff)
+	}
+
+	// Non-installed models should have "recommended, install?" description
+	for _, item := range items {
+		if !slices.Contains([]string{"kimi-k2.5:cloud", "llama3.2"}, item.Name) {
+			if item.Description != "recommended, install?" {
+				t.Errorf("non-installed %q should have description 'install?', got %q", item.Name, item.Description)
+			}
+		}
+	}
+}
+
+func TestBuildModelList_LatestTagStripped(t *testing.T) {
+	existing := []modelInfo{
+		{Name: "glm-4.7-flash:latest", Remote: false},
+		{Name: "llama3.2:latest", Remote: false},
+	}
+
+	items, _, existingModels, _ := buildModelList(existing, nil, "")
+	got := names(items)
+
+	// :latest should be stripped from display names
+	for _, name := range got {
+		if strings.HasSuffix(name, ":latest") {
+			t.Errorf("name %q should not have :latest suffix", name)
+		}
+	}
+
+	// glm-4.7-flash should not be duplicated (existing :latest matches the recommendation)
+	count := 0
+	for _, name := range got {
+		if name == "glm-4.7-flash" {
+			count++
+		}
+	}
+	if count != 1 {
+		t.Errorf("glm-4.7-flash should appear exactly once, got %d in %v", count, got)
+	}
+
+	// Stripped name should be in existingModels so it won't be pulled
+	if !existingModels["glm-4.7-flash"] {
+		t.Error("glm-4.7-flash should be in existingModels")
+	}
+}
+
+func TestBuildModelList_ReturnsExistingAndCloudMaps(t *testing.T) {
+	existing := []modelInfo{
+		{Name: "llama3.2:latest", Remote: false},
+		{Name: "glm-4.7:cloud", Remote: true},
+	}
+
+	_, _, existingModels, cloudModels := buildModelList(existing, nil, "")
+
+	if !existingModels["llama3.2"] {
+		t.Error("llama3.2 should be in existingModels")
+	}
+	if !existingModels["glm-4.7:cloud"] {
+		t.Error("glm-4.7:cloud should be in existingModels")
+	}
+	if existingModels["glm-4.7-flash"] {
+		t.Error("glm-4.7-flash should not be in existingModels (it's a recommendation)")
+	}
+
+	if !cloudModels["glm-4.7:cloud"] {
+		t.Error("glm-4.7: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["llama3.2"] {
+		t.Error("llama3.2 should not be in cloudModels")
+	}
+}

cmd/config/selector.go

@@ -353,10 +353,15 @@ func renderMultiSelect(w io.Writer, prompt string, s *multiSelectState) int {
 				suffix = " " + ansiGray + "(default)" + ansiReset
 			}
 
+			desc := ""
+			if item.Description != "" {
+				desc = " " + ansiGray + "- " + item.Description + ansiReset
+			}
+
 			if idx == s.highlighted && !s.focusOnButton {
-				fmt.Fprintf(w, "  %s%s %s %s%s%s\r\n", ansiBold, prefix, checkbox, item.Name, ansiReset, suffix)
+				fmt.Fprintf(w, "  %s%s %s %s%s%s%s\r\n", ansiBold, prefix, checkbox, item.Name, ansiReset, desc, suffix)
 			} else {
-				fmt.Fprintf(w, "  %s %s %s%s\r\n", prefix, checkbox, item.Name, suffix)
+				fmt.Fprintf(w, "  %s %s %s%s%s\r\n", prefix, checkbox, item.Name, desc, suffix)
 			}
 			lineCount++
 		}

envconfig/config.go

@@ -201,7 +201,7 @@ var (
 	// Enable the new Ollama engine
 	NewEngine = Bool("OLLAMA_NEW_ENGINE")
 	// ContextLength sets the default context length
-	ContextLength = Uint("OLLAMA_CONTEXT_LENGTH", 4096)
+	ContextLength = Uint("OLLAMA_CONTEXT_LENGTH", 0)
 	// Auth enables authentication between the Ollama client and server
 	UseAuth = Bool("OLLAMA_AUTH")
 	// Enable Vulkan backend
@@ -290,7 +290,7 @@ func AsMap() map[string]EnvVar {
 		"OLLAMA_ORIGINS":           {"OLLAMA_ORIGINS", AllowedOrigins(), "A comma separated list of allowed origins"},
 		"OLLAMA_SCHED_SPREAD":      {"OLLAMA_SCHED_SPREAD", SchedSpread(), "Always schedule model across all GPUs"},
 		"OLLAMA_MULTIUSER_CACHE":   {"OLLAMA_MULTIUSER_CACHE", MultiUserCache(), "Optimize prompt caching for multi-user scenarios"},
-		"OLLAMA_CONTEXT_LENGTH":    {"OLLAMA_CONTEXT_LENGTH", ContextLength(), "Context length to use unless otherwise specified (default: 4096)"},
+		"OLLAMA_CONTEXT_LENGTH":    {"OLLAMA_CONTEXT_LENGTH", ContextLength(), "Context length to use unless otherwise specified (default: 4k/32k/256k based on VRAM)"},
 		"OLLAMA_NEW_ENGINE":        {"OLLAMA_NEW_ENGINE", NewEngine(), "Enable the new Ollama engine"},
 		"OLLAMA_REMOTES":           {"OLLAMA_REMOTES", Remotes(), "Allowed hosts for remote models (default \"ollama.com\")"},
 

envconfig/config_test.go

@@ -282,7 +282,7 @@ func TestVar(t *testing.T) {
 
 func TestContextLength(t *testing.T) {
 	cases := map[string]uint{
-		"":     4096,
+		"":     0,
 		"2048": 2048,
 	}
 

llm/server.go

@@ -80,6 +80,7 @@ type LlamaServer interface {
 	GetPort() int
 	GetDeviceInfos(ctx context.Context) []ml.DeviceInfo
 	HasExited() bool
+	ContextLength() int
 }
 
 // llmServer is an instance of a runner hosting a single model
@@ -1200,7 +1201,8 @@ func (s *llmServer) initModel(ctx context.Context, req LoadRequest, operation Lo
 
 	resp, err := http.DefaultClient.Do(r)
 	if err != nil {
-		return nil, fmt.Errorf("do load request: %w", err)
+		slog.Error("do load request", "error", err)
+		return nil, errors.New("model failed to load, this may be due to resource limitations or an internal error, check ollama server logs for details")
 	}
 	defer resp.Body.Close()
 
@@ -1901,6 +1903,10 @@ func (s *llmServer) VRAMByGPU(id ml.DeviceID) uint64 {
 	return 0
 }
 
+func (s *llmServer) ContextLength() int {
+	return s.options.NumCtx
+}
+
 func (s *ollamaServer) GetDeviceInfos(ctx context.Context) []ml.DeviceInfo {
 	devices, err := ml.GetDevicesFromRunner(ctx, s)
 	if err != nil {

server/routes.go

@@ -75,16 +75,12 @@ func experimentEnabled(name string) bool {
 
 var useClient2 = experimentEnabled("client2")
 
-// Low VRAM mode is based on the sum of total VRAM (not free) and triggers
-// reduced context length on some models
-var lowVRAMThreshold uint64 = 20 * format.GibiByte
-
 var mode string = gin.DebugMode
 
 type Server struct {
-	addr    net.Addr
-	sched   *Scheduler
-	lowVRAM bool
+	addr          net.Addr
+	sched         *Scheduler
+	defaultNumCtx int
 }
 
 func init() {
@@ -107,8 +103,12 @@ var (
 	errBadTemplate = errors.New("template error")
 )
 
-func modelOptions(model *Model, requestOpts map[string]any) (api.Options, error) {
+func (s *Server) modelOptions(model *Model, requestOpts map[string]any) (api.Options, error) {
 	opts := api.DefaultOptions()
+	if opts.NumCtx == 0 {
+		opts.NumCtx = s.defaultNumCtx
+	}
+
 	if err := opts.FromMap(model.Options); err != nil {
 		return api.Options{}, err
 	}
@@ -140,20 +140,11 @@ func (s *Server) scheduleRunner(ctx context.Context, name string, caps []model.C
 		return nil, nil, nil, fmt.Errorf("%s %w", name, err)
 	}
 
-	opts, err := modelOptions(model, requestOpts)
+	opts, err := s.modelOptions(model, requestOpts)
 	if err != nil {
 		return nil, nil, nil, err
 	}
 
-	// This model is much more capable with a larger context, so set that
-	// unless it would penalize performance too much
-	if !s.lowVRAM && slices.Contains([]string{
-		"gptoss", "gpt-oss",
-		"qwen3vl", "qwen3vlmoe",
-	}, model.Config.ModelFamily) {
-		opts.NumCtx = max(opts.NumCtx, 8192)
-	}
-
 	runnerCh, errCh := s.sched.GetRunner(ctx, model, opts, keepAlive)
 	var runner *runnerRef
 	select {
@@ -1720,10 +1711,18 @@ func Serve(ln net.Listener) error {
 	for _, gpu := range gpus {
 		totalVRAM += gpu.TotalMemory - envconfig.GpuOverhead()
 	}
-	if totalVRAM < lowVRAMThreshold {
-		s.lowVRAM = true
-		slog.Info("entering low vram mode", "total vram", format.HumanBytes2(totalVRAM), "threshold", format.HumanBytes2(lowVRAMThreshold))
+
+	// Set default context based on VRAM tier
+	// Use slightly lower thresholds (47/23 GiB vs. 48/24 GiB) to account for small differences in the exact value
+	switch {
+	case totalVRAM >= 47*format.GibiByte:
+		s.defaultNumCtx = 262144
+	case totalVRAM >= 23*format.GibiByte:
+		s.defaultNumCtx = 32768
+	default:
+		s.defaultNumCtx = 4096
 	}
+	slog.Info("vram-based default context", "total_vram", format.HumanBytes2(totalVRAM), "default_num_ctx", s.defaultNumCtx)
 
 	err = srvr.Serve(ln)
 	// If server is closed from the signal handler, wait for the ctx to be done
@@ -1897,8 +1896,8 @@ func (s *Server) PsHandler(c *gin.Context) {
 			Details:   modelDetails,
 			ExpiresAt: v.expiresAt,
 		}
-		if v.Options != nil {
-			mr.ContextLength = v.Options.NumCtx
+		if v.llama != nil {
+			mr.ContextLength = v.llama.ContextLength()
 		}
 		// The scheduler waits to set expiresAt, so if a model is loading it's
 		// possible that it will be set to the unix epoch. For those cases, just

server/routes_debug_test.go

@@ -15,6 +15,7 @@ import (
 )
 
 func TestGenerateDebugRenderOnly(t *testing.T) {
+	t.Setenv("OLLAMA_CONTEXT_LENGTH", "4096")
 	gin.SetMode(gin.TestMode)
 
 	mock := mockRunner{
@@ -208,6 +209,7 @@ func TestGenerateDebugRenderOnly(t *testing.T) {
 }
 
 func TestChatDebugRenderOnly(t *testing.T) {
+	t.Setenv("OLLAMA_CONTEXT_LENGTH", "4096")
 	gin.SetMode(gin.TestMode)
 
 	mock := mockRunner{

server/routes_generate_renderer_test.go

@@ -20,6 +20,7 @@ import (
 // TestGenerateWithBuiltinRenderer tests that api/generate uses built-in renderers
 // when in chat-like flow (messages present, no suffix, no template)
 func TestGenerateWithBuiltinRenderer(t *testing.T) {
+	t.Setenv("OLLAMA_CONTEXT_LENGTH", "4096")
 	gin.SetMode(gin.TestMode)
 
 	mock := mockRunner{
@@ -204,6 +205,7 @@ func TestGenerateWithBuiltinRenderer(t *testing.T) {
 
 // TestGenerateWithDebugRenderOnly tests that debug_render_only works with built-in renderers
 func TestGenerateWithDebugRenderOnly(t *testing.T) {
+	t.Setenv("OLLAMA_CONTEXT_LENGTH", "4096")
 	gin.SetMode(gin.TestMode)
 
 	mock := mockRunner{

server/routes_generate_test.go

@@ -162,6 +162,7 @@ func TestGenerateChatRemote(t *testing.T) {
 }
 
 func TestGenerateChat(t *testing.T) {
+	t.Setenv("OLLAMA_CONTEXT_LENGTH", "4096")
 	gin.SetMode(gin.TestMode)
 
 	mock := mockRunner{
@@ -878,6 +879,7 @@ func TestGenerateChat(t *testing.T) {
 }
 
 func TestGenerate(t *testing.T) {
+	t.Setenv("OLLAMA_CONTEXT_LENGTH", "4096")
 	gin.SetMode(gin.TestMode)
 
 	mock := mockRunner{
@@ -2355,6 +2357,7 @@ func TestGenerateWithImages(t *testing.T) {
 // TestImageGenerateStreamFalse tests that image generation respects stream=false
 // and returns a single JSON response instead of streaming ndjson.
 func TestImageGenerateStreamFalse(t *testing.T) {
+	t.Setenv("OLLAMA_CONTEXT_LENGTH", "4096")
 	gin.SetMode(gin.TestMode)
 
 	p := t.TempDir()

server/routes_options_test.go

@@ -0,0 +1,127 @@
+package server
+
+import (
+	"testing"
+)
+
+func TestModelOptionsNumCtxPriority(t *testing.T) {
+	tests := []struct {
+		name           string
+		envContextLen  string // empty means not set (uses 0 sentinel)
+		defaultNumCtx  int    // VRAM-based default
+		modelNumCtx    int    // 0 means not set in model
+		requestNumCtx  int    // 0 means not set in request
+		expectedNumCtx int
+	}{
+		{
+			name:           "vram default when nothing else set",
+			envContextLen:  "",
+			defaultNumCtx:  32768,
+			modelNumCtx:    0,
+			requestNumCtx:  0,
+			expectedNumCtx: 32768,
+		},
+		{
+			name:           "env var overrides vram default",
+			envContextLen:  "8192",
+			defaultNumCtx:  32768,
+			modelNumCtx:    0,
+			requestNumCtx:  0,
+			expectedNumCtx: 8192,
+		},
+		{
+			name:           "model overrides vram default",
+			envContextLen:  "",
+			defaultNumCtx:  32768,
+			modelNumCtx:    16384,
+			requestNumCtx:  0,
+			expectedNumCtx: 16384,
+		},
+		{
+			name:           "model overrides env var",
+			envContextLen:  "8192",
+			defaultNumCtx:  32768,
+			modelNumCtx:    16384,
+			requestNumCtx:  0,
+			expectedNumCtx: 16384,
+		},
+		{
+			name:           "request overrides everything",
+			envContextLen:  "8192",
+			defaultNumCtx:  32768,
+			modelNumCtx:    16384,
+			requestNumCtx:  4096,
+			expectedNumCtx: 4096,
+		},
+		{
+			name:           "request overrides vram default",
+			envContextLen:  "",
+			defaultNumCtx:  32768,
+			modelNumCtx:    0,
+			requestNumCtx:  4096,
+			expectedNumCtx: 4096,
+		},
+		{
+			name:           "request overrides model",
+			envContextLen:  "",
+			defaultNumCtx:  32768,
+			modelNumCtx:    16384,
+			requestNumCtx:  4096,
+			expectedNumCtx: 4096,
+		},
+		{
+			name:           "low vram tier default",
+			envContextLen:  "",
+			defaultNumCtx:  4096,
+			modelNumCtx:    0,
+			requestNumCtx:  0,
+			expectedNumCtx: 4096,
+		},
+		{
+			name:           "high vram tier default",
+			envContextLen:  "",
+			defaultNumCtx:  262144,
+			modelNumCtx:    0,
+			requestNumCtx:  0,
+			expectedNumCtx: 262144,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			// Set or clear environment variable
+			if tt.envContextLen != "" {
+				t.Setenv("OLLAMA_CONTEXT_LENGTH", tt.envContextLen)
+			}
+
+			// Create server with VRAM-based default
+			s := &Server{
+				defaultNumCtx: tt.defaultNumCtx,
+			}
+
+			// Create model options (use float64 as FromMap expects JSON-style numbers)
+			var modelOpts map[string]any
+			if tt.modelNumCtx != 0 {
+				modelOpts = map[string]any{"num_ctx": float64(tt.modelNumCtx)}
+			}
+			model := &Model{
+				Options: modelOpts,
+			}
+
+			// Create request options (use float64 as FromMap expects JSON-style numbers)
+			var requestOpts map[string]any
+			if tt.requestNumCtx != 0 {
+				requestOpts = map[string]any{"num_ctx": float64(tt.requestNumCtx)}
+			}
+
+			opts, err := s.modelOptions(model, requestOpts)
+			if err != nil {
+				t.Fatalf("modelOptions failed: %v", err)
+			}
+
+			if opts.NumCtx != tt.expectedNumCtx {
+				t.Errorf("NumCtx = %d, want %d", opts.NumCtx, tt.expectedNumCtx)
+			}
+		})
+	}
+}

server/sched_test.go

@@ -804,6 +804,7 @@ func (s *mockLlm) GetPort() int                                       { return -
 func (s *mockLlm) GetDeviceInfos(ctx context.Context) []ml.DeviceInfo { return nil }
 func (s *mockLlm) HasExited() bool                                    { return false }
 func (s *mockLlm) GetActiveDeviceIDs() []ml.DeviceID                  { return nil }
+func (s *mockLlm) ContextLength() int                                 { return 0 }
 
 // TestImageGenRunnerCanBeEvicted verifies that an image generation model
 // loaded in the scheduler can be evicted when idle.

x/imagegen/server.go

@@ -347,6 +347,11 @@ func (s *Server) VRAMByGPU(id ml.DeviceID) uint64 {
 	return s.vramSize
 }
 
+// Context length is not applicable for image generation.
+func (s *Server) ContextLength() int {
+	return 0
+}
+
 func (s *Server) Embedding(ctx context.Context, input string) ([]float32, int, error) {
 	return nil, 0, errors.New("not supported")
 }