server: use tiered VRAM-based default context length

Replace binary low VRAM mode with tiered VRAM thresholds that set
default context lengths for all models:

- < 24 GiB VRAM: 4,096 context
- 24-48 GiB VRAM: 32,768 context
- >= 48 GiB VRAM: 262,144 context

cmd/cmd.go

@@ -1419,10 +1419,10 @@ func thinkingOutputClosingText(plainText bool) string {
 	return readline.ColorGrey + readline.ColorBold + text + readline.ColorDefault
 }
 
-func chat(cmd *cobra.Command, opts runOptions) (*api.Message, *api.Metrics, error) {
+func chat(cmd *cobra.Command, opts runOptions) (*api.Message, error) {
 	client, err := api.ClientFromEnvironment()
 	if err != nil {
-		return nil, nil, err
+		return nil, err
 	}
 
 	p := progress.NewProgress(os.Stderr)
@@ -1515,7 +1515,7 @@ func chat(cmd *cobra.Command, opts runOptions) (*api.Message, *api.Metrics, erro
 
 	if err := client.Chat(cancelCtx, req, fn); err != nil {
 		if errors.Is(err, context.Canceled) {
-			return nil, nil, nil
+			return nil, nil
 		}
 
 		// this error should ideally be wrapped properly by the client
@@ -1523,9 +1523,9 @@ func chat(cmd *cobra.Command, opts runOptions) (*api.Message, *api.Metrics, erro
 			p.StopAndClear()
 			fmt.Println("An error occurred while processing your message. Please try again.")
 			fmt.Println()
-			return nil, nil, nil
+			return nil, nil
 		}
-		return nil, nil, err
+		return nil, err
 	}
 
 	if len(opts.Messages) > 0 {
@@ -1535,14 +1535,14 @@ func chat(cmd *cobra.Command, opts runOptions) (*api.Message, *api.Metrics, erro
 
 	verbose, err := cmd.Flags().GetBool("verbose")
 	if err != nil {
-		return nil, nil, err
+		return nil, err
 	}
 
 	if verbose {
 		latest.Summary()
 	}
 
-	return &api.Message{Role: role, Thinking: thinkingContent.String(), Content: fullResponse.String()}, &latest.Metrics, nil
+	return &api.Message{Role: role, Thinking: thinkingContent.String(), Content: fullResponse.String()}, nil
 }
 
 func generate(cmd *cobra.Command, opts runOptions) error {

cmd/interactive.go

@@ -30,9 +30,6 @@ const (
 )
 
 func generateInteractive(cmd *cobra.Command, opts runOptions) error {
-	var sessionPromptTokens int64
-	var sessionCompletionTokens int64
-
 	usage := func() {
 		fmt.Fprintln(os.Stderr, "Available Commands:")
 		fmt.Fprintln(os.Stderr, "  /set            Set session variables")
@@ -40,7 +37,6 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 		fmt.Fprintln(os.Stderr, "  /load <model>   Load a session or model")
 		fmt.Fprintln(os.Stderr, "  /save <model>   Save your current session")
 		fmt.Fprintln(os.Stderr, "  /clear          Clear session context")
-		fmt.Fprintln(os.Stderr, "  /usage          Show session token usage")
 		fmt.Fprintln(os.Stderr, "  /bye            Exit")
 		fmt.Fprintln(os.Stderr, "  /?, /help       Help for a command")
 		fmt.Fprintln(os.Stderr, "  /? shortcuts    Help for keyboard shortcuts")
@@ -449,9 +445,6 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 			} else {
 				usageShow()
 			}
-		case strings.HasPrefix(line, "/usage"):
-			fmt.Printf("prompt tokens:     %d\n", sessionPromptTokens)
-			fmt.Printf("completion tokens: %d\n", sessionCompletionTokens)
 		case strings.HasPrefix(line, "/help"), strings.HasPrefix(line, "/?"):
 			args := strings.Fields(line)
 			if len(args) > 1 {
@@ -506,7 +499,7 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 
 			opts.Messages = append(opts.Messages, newMessage)
 
-			assistant, metrics, err := chat(cmd, opts)
+			assistant, err := chat(cmd, opts)
 			if err != nil {
 				if strings.Contains(err.Error(), "does not support thinking") ||
 					strings.Contains(err.Error(), "invalid think value") {
@@ -516,10 +509,6 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 				}
 				return err
 			}
-			if metrics != nil {
-				sessionPromptTokens += int64(metrics.PromptEvalCount)
-				sessionCompletionTokens += int64(metrics.EvalCount)
-			}
 			if assistant != nil {
 				opts.Messages = append(opts.Messages, *assistant)
 			}

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

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
@@ -1901,6 +1902,10 @@ 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")
 }