| 1 | # Apple MLX Support (`apple_mlx`) |
| 2 | |
| 3 | Local package collecting experimental Dart wrappers around `lib_mlx`, |
| 4 | `speech_mlx`, and `mlx_audio`. |
| 5 | |
| 6 | > **Development status:** Native LLM inference is disabled in LocalMind. The |
| 7 | > current `lib_mlx` 0.1.1 dependency ships a mock model core that returns stub |
| 8 | > responses; it does not execute downloaded MLX weights. Its bridge targets |
| 9 | > iOS only. The speech dependencies target macOS Apple Silicon and are not |
| 10 | > integrated into the LocalMind mobile app. |
| 11 | |
| 12 | ## Capabilities |
| 13 | |
| 14 | ### 1. LLM Text Generation (`AppleMlxLlmManager` & `lib_mlx`) |
| 15 | - **Prototype only**: Exercises the native lifecycle and local OpenAI-compatible |
| 16 | HTTP/SSE contract against the upstream mock core. |
| 17 | - **Not user-facing**: `AppleMlxCapabilities.nativeLlmInferenceAvailable` stays |
| 18 | `false` until a real, device-validated backend replaces the mock. |
| 19 | |
| 20 | ### 2. Speech-To-Text (`AppleMlxSpeechManager` & `speech_mlx` / `mlx_audio`) |
| 21 | - **macOS only**: These dependencies do not provide iOS speech inference. |
| 22 | - **Parakeet-TDT & Whisper**: High-accuracy local speech recognition. |
| 23 | - **Isolate Offloading**: Dedicated worker isolates keep audio processing and resampling off the UI isolate. |
| 24 | |
| 25 | ### 3. Text-To-Speech (`PocketTTS`) |
| 26 | - **Neural Voice Synthesis**: Real-time voice generation with incremental PCM streaming. |
| 27 | - **Voices**: Built-in multi-voice support (`alba`, `marius`, `javert`, `jean`). |
| 28 | |
| 29 | ### 4. Conversational Turn Detection (`SmartTurn`) |
| 30 | - **Smart Turn v3.2**: Sub-millisecond endpoint classifier to determine whether a speaker finished speaking during a voice conversation. |
| 31 | |
| 32 | --- |
| 33 | |
| 34 | ## Quickstart |
| 35 | |
| 36 | ### LLM Text Generation |
| 37 | |
| 38 | The following example is for bridge development and contract testing only. It |
| 39 | does not perform real inference with `lib_mlx` 0.1.1. |
| 40 | |
| 41 | ```dart |
| 42 | import 'package:apple_mlx/apple_mlx.dart'; |
| 43 | |
| 44 | Future<void> main() async { |
| 45 | final llm = AppleMlxLlmManager(); |
| 46 | |
| 47 | // 1. Load MLX model and launch local server |
| 48 | final serverInfo = await llm.loadAndStartServer( |
| 49 | modelPath: '/path/to/mlx-community/gemma-4-e2b-it-4bit', |
| 50 | modelId: AppleMlxModelIds.gemma4E2b, |
| 51 | thinkingEnabled: true, |
| 52 | ); |
| 53 | print('MLX Server running on ${serverInfo.baseUrl}'); |
| 54 | |
| 55 | // 2. Stream generation |
| 56 | final stream = llm.generateStream( |
| 57 | messages: [ |
| 58 | {'role': 'user', 'content': 'Explain quantum computing in simple terms.'}, |
| 59 | ], |
| 60 | temperature: 0.7, |
| 61 | ); |
| 62 | |
| 63 | await for (final delta in stream) { |
| 64 | if (delta.reasoningContent != null) { |
| 65 | print('[Thinking] ${delta.reasoningContent}'); |
| 66 | } |
| 67 | if (delta.content != null) { |
| 68 | print(delta.content); |
| 69 | } |
| 70 | } |
| 71 | |
| 72 | // 3. Clean up |
| 73 | await llm.dispose(); |
| 74 | } |
| 75 | ``` |
| 76 | |
| 77 | ### STT and TTS Audio |
| 78 | |
| 79 | ```dart |
| 80 | import 'package:apple_mlx/apple_mlx.dart'; |
| 81 | |
| 82 | Future<void> main() async { |
| 83 | final speech = AppleMlxSpeechManager(); |
| 84 | await speech.initialize(); |
| 85 | |
| 86 | // Transcribe audio |
| 87 | final text = await speech.transcribeWavFile('input.wav'); |
| 88 | print('Transcript: $text'); |
| 89 | |
| 90 | // Synthesize voice |
| 91 | await speech.synthesizeToWav( |
| 92 | text: 'Hello from Apple MLX on LocalMind!', |
| 93 | outputPath: 'output.wav', |
| 94 | voice: 'alba', |
| 95 | ); |
| 96 | |
| 97 | await speech.dispose(); |
| 98 | } |
| 99 | ``` |