mlxrunner: Report actual memory usage from runner

The MLX runner previously reported a static VRAM estimate that was
computed at load time and consisted only of the weights. This is
strictly less than the actual memory usage, as it does not include
the KV cache or compute graph.

app/cmd/app/app.go

@@ -35,7 +35,6 @@ import (
 var (
 	wv           = &Webview{}
 	uiServerPort int
-	appStore     *store.Store
 )
 
 var debug = strings.EqualFold(os.Getenv("OLLAMA_DEBUG"), "true") || os.Getenv("OLLAMA_DEBUG") == "1"
@@ -209,7 +208,6 @@ func main() {
 	uiServerPort = port
 
 	st := &store.Store{}
-	appStore = st
 
 	// Enable CORS in development mode
 	if devMode {
@@ -296,15 +294,8 @@ func main() {
 
 	// Check for pending updates on startup (show tray notification if update is ready)
 	if updater.IsUpdatePending() {
-		// On Windows, the tray is initialized in osRun(). Calling UpdateAvailable
-		// before that would dereference a nil tray callback.
-		// TODO: refactor so the update check runs after platform init on all platforms.
-		if runtime.GOOS == "windows" {
-			slog.Debug("update pending on startup, deferring tray notification until tray initialization")
-		} else {
-			slog.Debug("update pending on startup, showing tray notification")
-			UpdateAvailable("")
-		}
+		slog.Debug("update pending on startup, showing tray notification")
+		UpdateAvailable("")
 	}
 
 	hasCompletedFirstRun, err := st.HasCompletedFirstRun()
@@ -369,7 +360,8 @@ func startHiddenTasks() {
 			slog.Info("deferring pending update for fast startup")
 		} else {
 			// Check if auto-update is enabled before automatically upgrading
-			settings, err := appStore.Settings()
+			st := &store.Store{}
+			settings, err := st.Settings()
 			if err != nil {
 				slog.Warn("failed to load settings for upgrade check", "error", err)
 			} else if !settings.AutoUpdateEnabled {

app/cmd/app/app_windows.go

@@ -154,10 +154,6 @@ func handleURLSchemeRequest(urlScheme string) {
 }
 
 func UpdateAvailable(ver string) error {
-	if app.t == nil {
-		slog.Debug("tray not yet initialized, skipping update notification")
-		return nil
-	}
 	return app.t.UpdateAvailable(ver)
 }
 
@@ -169,14 +165,6 @@ func osRun(shutdown func(), hasCompletedFirstRun, startHidden bool) {
 		log.Fatalf("Failed to start: %s", err)
 	}
 
-	// Check for pending updates now that the tray is initialized.
-	// The platform-independent check in app.go fires before osRun,
-	// when app.t is still nil, so we must re-check here.
-	if updater.IsUpdatePending() {
-		slog.Debug("update pending on startup, showing tray notification")
-		UpdateAvailable("")
-	}
-
 	signals := make(chan os.Signal, 1)
 	signal.Notify(signals, syscall.SIGINT, syscall.SIGTERM)
 

app/updater/updater.go

@@ -289,7 +289,6 @@ func (u *Updater) TriggerImmediateCheck() {
 
 func (u *Updater) StartBackgroundUpdaterChecker(ctx context.Context, cb func(string) error) {
 	u.checkNow = make(chan struct{}, 1)
-	u.checkNow <- struct{}{} // Trigger first check after initial delay
 	go func() {
 		// Don't blast an update message immediately after startup
 		time.Sleep(UpdateCheckInitialDelay)
@@ -334,7 +333,7 @@ func (u *Updater) StartBackgroundUpdaterChecker(ctx context.Context, cb func(str
 				continue
 			}
 
-			// Download successful - show tray notification
+			// Download successful - show tray notification (regardless of toggle state)
 			err = cb(resp.UpdateVersion)
 			if err != nil {
 				slog.Warn("failed to register update available with tray", "error", err)

app/updater/updater_test.go

@@ -351,13 +351,10 @@ func TestTriggerImmediateCheck(t *testing.T) {
 
 	updater.StartBackgroundUpdaterChecker(ctx, cb)
 
-	// Wait for the initial check that fires after the initial delay
-	select {
-	case <-checkDone:
-	case <-time.After(2 * time.Second):
-		t.Fatal("initial check did not happen")
-	}
+	// Wait for goroutine to start and pass initial delay
+	time.Sleep(10 * time.Millisecond)
 
+	// With 1 hour interval, no check should have happened yet
 	initialCount := checkCount.Load()
 
 	// Trigger immediate check

llm/server.go

@@ -74,8 +74,7 @@ type LlamaServer interface {
 	Tokenize(ctx context.Context, content string) ([]int, error)
 	Detokenize(ctx context.Context, tokens []int) (string, error)
 	Close() error
-	VRAMSize() uint64 // Total VRAM across all GPUs
-	TotalSize() uint64
+	MemorySize() (total, vram uint64)
 	VRAMByGPU(id ml.DeviceID) uint64
 	Pid() int
 	GetPort() int
@@ -685,8 +684,9 @@ func (s *llamaServer) Load(ctx context.Context, systemInfo ml.SystemInfo, system
 	// Windows CUDA should not use mmap for best performance
 	// Linux  with a model larger than free space, mmap leads to thrashing
 	// For CPU loads we want the memory to be allocated, not FS cache
+	totalSize, _ := s.MemorySize()
 	if (runtime.GOOS == "windows" && len(gpus) > 0 && gpus[0].Library == "CUDA" && s.options.UseMMap == nil) ||
-		(runtime.GOOS == "linux" && systemInfo.FreeMemory < s.TotalSize() && s.options.UseMMap == nil) ||
+		(runtime.GOOS == "linux" && systemInfo.FreeMemory < totalSize && s.options.UseMMap == nil) ||
 		(len(gpus) == 0 && s.options.UseMMap == nil) ||
 		(len(gpus) > 0 && gpus[0].Library == "Vulkan" && s.options.UseMMap == nil) ||
 		(s.options.UseMMap != nil && !*s.options.UseMMap) {
@@ -1848,17 +1848,17 @@ func (s *llamaServer) GetDeviceInfos(ctx context.Context) []ml.DeviceInfo {
 	return nil
 }
 
-func (s *llmServer) VRAMSize() uint64 {
+func (s *llmServer) MemorySize() (total, vram uint64) {
 	if s.mem == nil {
-		return 0
+		return 0, 0
 	}
 
-	var mem uint64
-
 	for _, g := range s.mem.GPUs {
-		mem += g.Size()
+		vram += g.Size()
 	}
 
+	total = s.mem.InputWeights + s.mem.CPU.Size() + vram
+
 	// Some elements are always on CPU. However, if we have allocated all layers
 	// on the GPU then include the CPU components as well, to represent complete offloading.
 	noCPULayers := true
@@ -1869,25 +1869,11 @@ func (s *llmServer) VRAMSize() uint64 {
 		}
 	}
 	if noCPULayers {
-		mem += s.mem.InputWeights
-		mem += s.mem.CPU.Graph
+		vram += s.mem.InputWeights
+		vram += s.mem.CPU.Graph
 	}
 
-	return mem
-}
-
-func (s *llmServer) TotalSize() uint64 {
-	if s.mem == nil {
-		return 0
-	}
-
-	mem := s.mem.InputWeights
-	mem += s.mem.CPU.Size()
-	for _, g := range s.mem.GPUs {
-		mem += g.Size()
-	}
-
-	return mem
+	return total, vram
 }
 
 func (s *llmServer) VRAMByGPU(id ml.DeviceID) uint64 {

model/parsers/qwen3.go

@@ -204,24 +204,6 @@ func (p *Qwen3Parser) eat() ([]qwen3Event, bool) {
 			p.maybeThinkingOpenAtBOL = false
 		}
 
-		thinkingCloseIdx := strings.Index(acc, qwen3ThinkingCloseTag)
-		toolOpenIdx := strings.Index(acc, qwen3ToolOpenTag)
-
-		// If a tool call starts before </think>, treat that as the end of thinking
-		// for parsing purposes and continue in tool-call mode.
-		if toolOpenIdx != -1 && (thinkingCloseIdx == -1 || toolOpenIdx < thinkingCloseIdx) {
-			before, after := p.splitAtTag(qwen3ToolOpenTag, true)
-			if len(before) > 0 {
-				events = append(events, qwen3EventThinkingContent{content: before})
-			}
-			if after == "" {
-				p.state = qwen3ParserStateToolStartedEatingWhitespace
-			} else {
-				p.state = qwen3ParserStateCollectingToolContent
-			}
-			return events, true
-		}
-
 		if strings.Contains(acc, qwen3ThinkingCloseTag) {
 			thinking, remaining := p.splitAtTag(qwen3ThinkingCloseTag, true)
 			if len(thinking) > 0 {
@@ -233,7 +215,7 @@ func (p *Qwen3Parser) eat() ([]qwen3Event, bool) {
 				p.state = qwen3ParserStateCollectingContent
 			}
 			return events, true
-		} else if overlapLen := max(overlap(acc, qwen3ThinkingCloseTag), overlap(acc, qwen3ToolOpenTag)); overlapLen > 0 {
+		} else if overlapLen := overlap(acc, qwen3ThinkingCloseTag); overlapLen > 0 {
 			beforePartialTag := acc[:len(acc)-overlapLen]
 			trailingWsLen := trailingWhitespaceLen(beforePartialTag)
 			ambiguousStart := len(beforePartialTag) - trailingWsLen

model/parsers/qwen3_test.go

@@ -146,68 +146,6 @@ func TestQwen3ParserToolCall(t *testing.T) {
 	}
 }
 
-func TestQwen3ParserThinkingWithToolCallBeforeThinkingClose(t *testing.T) {
-	parser := &Qwen3Parser{hasThinkingSupport: true, defaultThinking: true}
-	parser.Init(nil, nil, &api.ThinkValue{Value: true})
-
-	input := "Let me think<tool_call>{\"name\":\"get_weather\",\"arguments\":{\"location\":\"San Francisco\",\"unit\":\"celsius\"}}</tool_call>"
-	content, thinking, calls, err := parser.Add(input, true)
-	if err != nil {
-		t.Fatalf("parse failed: %v", err)
-	}
-
-	if content != "" {
-		t.Fatalf("expected empty content, got %q", content)
-	}
-	if thinking != "Let me think" {
-		t.Fatalf("expected thinking %q, got %q", "Let me think", thinking)
-	}
-	if len(calls) != 1 {
-		t.Fatalf("expected 1 tool call, got %d", len(calls))
-	}
-	if calls[0].Function.Name != "get_weather" {
-		t.Fatalf("expected tool name %q, got %q", "get_weather", calls[0].Function.Name)
-	}
-}
-
-func TestQwen3ParserThinkingWithSplitToolOpenTag(t *testing.T) {
-	parser := &Qwen3Parser{hasThinkingSupport: true, defaultThinking: true}
-	parser.Init(nil, nil, &api.ThinkValue{Value: true})
-
-	content, thinking, calls, err := parser.Add("Let me think<tool_ca", false)
-	if err != nil {
-		t.Fatalf("parse failed on first chunk: %v", err)
-	}
-	if content != "" || thinking != "Let me think" || len(calls) != 0 {
-		t.Fatalf(
-			"expected content=%q thinking=%q calls=%d, got content=%q thinking=%q calls=%d",
-			"",
-			"Let me think",
-			0,
-			content,
-			thinking,
-			len(calls),
-		)
-	}
-
-	content, thinking, calls, err = parser.Add("ll>{\"name\":\"get_weather\",\"arguments\":{\"location\":\"SF\"}}</tool_call>", true)
-	if err != nil {
-		t.Fatalf("parse failed on second chunk: %v", err)
-	}
-	if content != "" {
-		t.Fatalf("expected empty content, got %q", content)
-	}
-	if thinking != "" {
-		t.Fatalf("expected no additional thinking on second chunk, got %q", thinking)
-	}
-	if len(calls) != 1 {
-		t.Fatalf("expected 1 tool call, got %d", len(calls))
-	}
-	if calls[0].Function.Name != "get_weather" {
-		t.Fatalf("expected tool name %q, got %q", "get_weather", calls[0].Function.Name)
-	}
-}
-
 func TestQwen35ParserRespectsNoThink(t *testing.T) {
 	parser := ParserForName("qwen3.5")
 	if parser == nil {

model/parsers/qwen3vl.go

@@ -180,22 +180,7 @@ func (p *Qwen3VLParser) eat() ([]qwenEvent, bool) {
 			return events, false
 		}
 	case CollectingThinkingContent:
-		acc := p.buffer.String()
-		thinkingCloseIdx := strings.Index(acc, thinkingCloseTag)
-		toolOpenIdx := strings.Index(acc, toolOpenTag)
-
-		// If a tool call starts before </think>, treat that as the end of thinking
-		// for parsing purposes and continue in tool-call mode.
-		if toolOpenIdx != -1 && (thinkingCloseIdx == -1 || toolOpenIdx < thinkingCloseIdx) {
-			before, _ := splitAtTag(&p.buffer, toolOpenTag, false)
-			if len(before) > 0 {
-				events = append(events, qwenEventThinkingContent{content: before})
-			}
-			p.state = CollectingToolContent
-			return events, true
-		}
-
-		if strings.Contains(acc, thinkingCloseTag) {
+		if strings.Contains(p.buffer.String(), thinkingCloseTag) {
 			thinking, remaining := splitAtTag(&p.buffer, thinkingCloseTag, true)
 			if len(thinking) > 0 {
 				events = append(events, qwenEventThinkingContent{content: thinking})
@@ -206,13 +191,13 @@ func (p *Qwen3VLParser) eat() ([]qwenEvent, bool) {
 				p.state = CollectingContent
 			}
 			return events, true
-		} else if overlapLen := max(overlap(acc, thinkingCloseTag), overlap(acc, toolOpenTag)); overlapLen > 0 {
-			beforePartialTag := acc[:len(acc)-overlapLen]
+		} else if overlapLen := overlap(p.buffer.String(), thinkingCloseTag); overlapLen > 0 {
+			beforePartialTag := p.buffer.String()[:len(p.buffer.String())-overlapLen]
 			trailingWhitespaceLen := trailingWhitespaceLen(beforePartialTag)
 			ambiguousStart := len(beforePartialTag) - trailingWhitespaceLen
 
-			unambiguous := acc[:ambiguousStart]
-			ambiguous := acc[ambiguousStart:]
+			unambiguous := p.buffer.String()[:ambiguousStart]
+			ambiguous := p.buffer.String()[ambiguousStart:]
 			p.buffer.Reset()
 			p.buffer.WriteString(ambiguous)
 			if len(unambiguous) > 0 {
@@ -220,11 +205,11 @@ func (p *Qwen3VLParser) eat() ([]qwenEvent, bool) {
 			}
 			return events, false
 		} else {
-			whitespaceLen := trailingWhitespaceLen(acc)
-			ambiguousStart := len(acc) - whitespaceLen
+			whitespaceLen := trailingWhitespaceLen(p.buffer.String())
+			ambiguousStart := len(p.buffer.String()) - whitespaceLen
 
-			unambiguous := acc[:ambiguousStart]
-			ambiguous := acc[ambiguousStart:]
+			unambiguous := p.buffer.String()[:ambiguousStart]
+			ambiguous := p.buffer.String()[ambiguousStart:]
 			p.buffer.Reset()
 			p.buffer.WriteString(ambiguous)
 			if len(unambiguous) > 0 {

model/parsers/qwen3vl_thinking_test.go

@@ -98,12 +98,8 @@ func TestQwen3VLThinkingParserStreaming(t *testing.T) {
 			desc: "nested thinking and tool call (outside thinking, inside tool call)",
 			steps: []step{
 				{
-					input: "I'm thinking<tool_call>I'm nested tool call</tool_call></think>",
-					wantEvents: []qwenEvent{
-						qwenEventThinkingContent{content: "I'm thinking"},
-						qwenEventRawToolCall{raw: "I'm nested tool call"},
-						qwenEventContent{content: "</think>"},
-					},
+					input:      "I'm thinking<tool_call>I'm nested tool call</tool_call></think>",
+					wantEvents: []qwenEvent{qwenEventThinkingContent{content: "I'm thinking<tool_call>I'm nested tool call</tool_call>"}},
 				},
 			},
 		},
@@ -113,7 +109,8 @@ func TestQwen3VLThinkingParserStreaming(t *testing.T) {
 				{
 					input: "<tool_call>I'm nested tool call<think>I'm thinking</think></tool_call>",
 					wantEvents: []qwenEvent{
-						qwenEventRawToolCall{raw: "I'm nested tool call<think>I'm thinking</think>"},
+						qwenEventThinkingContent{content: "<tool_call>I'm nested tool call<think>I'm thinking"},
+						qwenEventContent{content: "</tool_call>"},
 					},
 				},
 			},
@@ -124,8 +121,8 @@ func TestQwen3VLThinkingParserStreaming(t *testing.T) {
 				{
 					input: "I'm thinking<tool_call>I'm NOT a nested tool call</think></tool_call><tool_call>I'm nested tool call 2<think></tool_call></think>",
 					wantEvents: []qwenEvent{
-						qwenEventThinkingContent{content: "I'm thinking"},
-						qwenEventRawToolCall{raw: "I'm NOT a nested tool call</think>"},
+						qwenEventThinkingContent{content: "I'm thinking<tool_call>I'm NOT a nested tool call"},
+						qwenEventContent{content: "</tool_call>"},
 						qwenEventRawToolCall{raw: "I'm nested tool call 2<think>"},
 						qwenEventContent{content: "</think>"},
 					},

server/routes.go

@@ -1951,6 +1951,9 @@ func (s *Server) PsHandler(c *gin.Context) {
 		}
 		if v.llama != nil {
 			mr.ContextLength = v.llama.ContextLength()
+			total, vram := v.llama.MemorySize()
+			mr.Size = int64(total)
+			mr.SizeVRAM = int64(vram)
 		}
 		// 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/sched.go

@@ -536,6 +536,7 @@ iGPUScan:
 		}
 	}
 
+	totalSize, vramSize := llama.MemorySize()
 	runner := &runnerRef{
 		model:           req.model,
 		modelPath:       req.model.ModelPath,
@@ -545,8 +546,8 @@ iGPUScan:
 		sessionDuration: sessionDuration,
 		gpus:            gpuIDs,
 		discreteGPUs:    discreteGPUs,
-		vramSize:        llama.VRAMSize(),
-		totalSize:       llama.TotalSize(),
+		totalSize:       totalSize,
+		vramSize:        vramSize,
 		loading:         true,
 		pid:             llama.Pid(),
 	}
@@ -619,6 +620,7 @@ func (s *Scheduler) loadMLX(req *LlmRequest) bool {
 		sessionDuration = req.sessionDuration.Duration
 	}
 
+	totalSize, vramSize := server.MemorySize()
 	runner := &runnerRef{
 		model:           req.model,
 		modelPath:       req.model.ModelPath,
@@ -628,8 +630,8 @@ func (s *Scheduler) loadMLX(req *LlmRequest) bool {
 		loading:         false,
 		isImagegen:      isImagegen,
 		sessionDuration: sessionDuration,
-		totalSize:       server.TotalSize(),
-		vramSize:        server.VRAMSize(),
+		totalSize:       totalSize,
+		vramSize:        vramSize,
 	}
 
 	s.loadedMu.Lock()

server/sched_test.go

@@ -861,8 +861,7 @@ func (s *mockLlm) Close() error {
 	s.closeCalled = true
 	return s.closeResp
 }
-func (s *mockLlm) VRAMSize() uint64                                   { return s.vramSize }
-func (s *mockLlm) TotalSize() uint64                                  { return s.totalSize }
+func (s *mockLlm) MemorySize() (uint64, uint64)                       { return s.totalSize, s.vramSize }
 func (s *mockLlm) VRAMByGPU(id ml.DeviceID) uint64                    { return s.vramByGPU[id] }
 func (s *mockLlm) Pid() int                                           { return -1 }
 func (s *mockLlm) GetPort() int                                       { return -1 }

x/imagegen/server.go

@@ -374,14 +374,9 @@ func (s *Server) Close() error {
 	return nil
 }
 
-// VRAMSize returns the estimated VRAM usage.
-func (s *Server) VRAMSize() uint64 {
-	return s.vramSize
-}
-
-// TotalSize returns the total memory usage.
-func (s *Server) TotalSize() uint64 {
-	return s.vramSize
+// MemorySize returns the total and VRAM memory usage.
+func (s *Server) MemorySize() (total, vram uint64) {
+	return s.vramSize, s.vramSize
 }
 
 // VRAMByGPU returns VRAM usage for a specific GPU.

x/mlxrunner/cache.go

@@ -78,11 +78,6 @@ func (c *kvCache) findRemaining(tokens []int32) []int32 {
 		prefix++
 	}
 
-	if prefix == len(tokens) && prefix > 0 {
-		// Leave one token to run through the model so we can sample a response.
-		prefix--
-	}
-
 	if prefix < len(c.tokens) {
 		trim := len(c.tokens) - prefix
 		for _, kv := range c.caches {

x/mlxrunner/client.go

@@ -24,14 +24,13 @@ import (
 	"github.com/ollama/ollama/llm"
 	"github.com/ollama/ollama/ml"
 	"github.com/ollama/ollama/x/imagegen"
-	"github.com/ollama/ollama/x/imagegen/manifest"
 )
 
 // Client wraps an MLX runner subprocess to implement llm.LlamaServer for LLM models.
 type Client struct {
 	port        int
 	modelName   string
-	vramSize    uint64
+	memory      uint
 	done        chan error
 	client      *http.Client
 	lastErr     string
@@ -98,18 +97,9 @@ func NewClient(modelName string) (*Client, error) {
 		slog.Debug("mlx subprocess library path", "LD_LIBRARY_PATH", pathEnvVal)
 	}
 
-	// Estimate VRAM based on tensor size from manifest
-	var vramSize uint64
-	if modelManifest, err := manifest.LoadManifest(modelName); err == nil {
-		vramSize = uint64(modelManifest.TotalTensorSize())
-	} else {
-		vramSize = 8 * 1024 * 1024 * 1024
-	}
-
 	c := &Client{
 		port:      port,
 		modelName: modelName,
-		vramSize:  vramSize,
 		done:      make(chan error, 1),
 		client:    &http.Client{Timeout: 10 * time.Minute},
 		cmd:       cmd,
@@ -347,9 +337,15 @@ func (c *Client) Pid() int {
 	return -1
 }
 
+type statusResponse struct {
+	Status   int
+	Progress int
+	Memory   uint
+}
+
 // Ping implements llm.LlamaServer.
 func (c *Client) Ping(ctx context.Context) error {
-	reqURL := fmt.Sprintf("http://127.0.0.1:%d/health", c.port)
+	reqURL := fmt.Sprintf("http://127.0.0.1:%d/v1/status", c.port)
 	req, err := http.NewRequestWithContext(ctx, "GET", reqURL, nil)
 	if err != nil {
 		return err
@@ -362,6 +358,12 @@ func (c *Client) Ping(ctx context.Context) error {
 	if resp.StatusCode != http.StatusOK {
 		return fmt.Errorf("health check failed: %d", resp.StatusCode)
 	}
+
+	var status statusResponse
+	if err := json.NewDecoder(resp.Body).Decode(&status); err != nil {
+		return err
+	}
+	c.memory = status.Memory
 	return nil
 }
 
@@ -388,19 +390,24 @@ func (c *Client) Tokenize(ctx context.Context, content string) ([]int, error) {
 	return tokens, nil
 }
 
-// TotalSize implements llm.LlamaServer.
-func (c *Client) TotalSize() uint64 {
-	return c.vramSize
+func (c *Client) currentMemory() uint64 {
+	ctx, cancel := context.WithTimeout(context.Background(), time.Second)
+	defer cancel()
+	if err := c.Ping(ctx); err != nil {
+		slog.Warn("failed to get current memory", "error", err)
+	}
+	return uint64(c.memory)
+}
+
+// MemorySize implements llm.LlamaServer.
+func (c *Client) MemorySize() (total, vram uint64) {
+	mem := c.currentMemory()
+	return mem, mem
 }
 
 // VRAMByGPU implements llm.LlamaServer.
 func (c *Client) VRAMByGPU(id ml.DeviceID) uint64 {
-	return c.vramSize
-}
-
-// VRAMSize implements llm.LlamaServer.
-func (c *Client) VRAMSize() uint64 {
-	return c.vramSize
+	return c.currentMemory()
 }
 
 // WaitUntilRunning implements llm.LlamaServer.

x/mlxrunner/server.go

@@ -50,9 +50,10 @@ func Execute(args []string) error {
 
 	mux := http.NewServeMux()
 	mux.HandleFunc("GET /v1/status", func(w http.ResponseWriter, r *http.Request) {
-		if err := json.NewEncoder(w).Encode(map[string]any{
-			"status":   0,
-			"progress": 100,
+		if err := json.NewEncoder(w).Encode(statusResponse{
+			Status:   0,
+			Progress: 100,
+			Memory:   uint(mlx.ActiveMemory() + mlx.CacheMemory()),
 		}); err != nil {
 			slog.Error("Failed to encode response", "error", err)
 			http.Error(w, "Internal Server Error", http.StatusInternalServerError)