diff --git a/internal/cmd/ml/cmd_serve.go b/internal/cmd/ml/cmd_serve.go
index 740eba7c..9547f601 100644
--- a/internal/cmd/ml/cmd_serve.go
+++ b/internal/cmd/ml/cmd_serve.go
@@ -79,9 +79,12 @@ type usageInfo struct {
 }
 
 func runServe(cmd *cli.Command, args []string) error {
-	// Create a backend — use HTTP backend pointing to configured API URL.
-	// On macOS with MLX build tag, this will use the native MLX backend instead.
-	backend := ml.NewHTTPBackend(apiURL, modelName)
+	// Try native MLX backend first (macOS arm64 with mlx tag + model-path set),
+	// fall back to HTTP proxy backend.
+	backend, err := createServeBackend()
+	if err != nil {
+		return err
+	}
 
 	mux := http.NewServeMux()
 
diff --git a/internal/cmd/ml/serve_backend_default.go b/internal/cmd/ml/serve_backend_default.go
new file mode 100644
index 00000000..69a1d942
--- /dev/null
+++ b/internal/cmd/ml/serve_backend_default.go
@@ -0,0 +1,9 @@
+//go:build !(darwin && arm64 && mlx)
+
+package ml
+
+import "forge.lthn.ai/core/cli/pkg/ml"
+
+func createServeBackend() (ml.Backend, error) {
+	return ml.NewHTTPBackend(apiURL, modelName), nil
+}
diff --git a/internal/cmd/ml/serve_backend_mlx.go b/internal/cmd/ml/serve_backend_mlx.go
new file mode 100644
index 00000000..f48d1aab
--- /dev/null
+++ b/internal/cmd/ml/serve_backend_mlx.go
@@ -0,0 +1,22 @@
+//go:build darwin && arm64 && mlx
+
+package ml
+
+import (
+	"fmt"
+	"log/slog"
+
+	"forge.lthn.ai/core/cli/pkg/ml"
+)
+
+func createServeBackend() (ml.Backend, error) {
+	if serveModelPath != "" {
+		slog.Info("ml serve: loading native MLX backend", "path", serveModelPath)
+		b, err := ml.NewMLXBackend(serveModelPath)
+		if err != nil {
+			return nil, fmt.Errorf("mlx backend: %w", err)
+		}
+		return b, nil
+	}
+	return ml.NewHTTPBackend(apiURL, modelName), nil
+}