mlx_localllm
English
A Flutter plugin for high-performance localized LLM inference on macOS using Apple's MLX framework. It provides zero-latency interaction by loading models directly into the application process.
Features
- 🚀 Native Acceleration: Built on Apple's MLX for optimal performance on Apple Silicon.
- 📦 In-Process Inference: No external sidecars or servers (like Ollama) required.
- 🌐 Robust Downloader: Built-in support for Hugging Face and mirrors with resilient chunk-based downloading.
- 🛠️ Full Lifecycle: From model discovery and download to stateful conversational inference.
Platform Support
| Platform | Support | Architecture | Min OS |
|---|---|---|---|
| macOS | ✅ | Apple Silicon (ARM64) / Intel (x86_64 Mock) | 14.0+ |
| iOS | ✅ | ARM64 | 17.0+ |
| Android | ❌ | - | - |
| Windows | ❌ | - | - |
| Linux | ❌ | - | - |
| Web | ❌ | - | - |
Requirements
- macOS: 14.0 or higher.
- iOS: 17.0 or higher.
- Hardware: Apple Silicon (M1, M2, M3, etc.) for inference.
- Flutter: 3.0.0 or higher.
Installation
-
Add the dependency to your
pubspec.yaml:flutter pub add mlx_localllm -
Platform Setup: The plugin uses Swift Package Manager (SPM) natively (requires Flutter 3.24+).
- macOS/iOS: Dependencies are automatically resolved by the Flutter toolchain.
- Architecture:
- Apple Silicon (ARM64): Fully supported with native MLX acceleration.
- Intel (x86_64): Supported via Mocking. You can compile and link on Intel Macs, but inference calls will return an "Unsupported architecture" error. This allows for seamless development across different Mac architectures.
Usage
Check Support
bool supported = await MlxLocalllm().isSupported();
Download Model
// Track progress via modelEvents stream
MlxLocalllm().modelEvents.listen((event) {
if (event['event'] == 'progress') {
print('Download progress: ${event['progress'] * 100}%');
} else if (event['event'] == 'complete') {
print('Download complete at ${event['path']}');
}
});
await MlxLocalllm().downloadModel('mlx-community/Qwen2.5-0.5B-Instruct-4bit');
API Reference
MlxLocalllm Singleton
| Method | Description | Returns |
|---|---|---|
isSupported() |
Checks if hardware supports MLX (requires Apple Silicon GPU). | Future<bool> |
downloadModel(repoId) |
Starts background download from HF. Monitor via modelEvents. |
Future<bool> |
loadModel(path) |
Loads model into memory from absolute path or repoId. | Future<bool> |
unloadModel() |
Releases model and frees system memory. | Future<bool> |
generate(prompt, config) |
Generates complete response (blocking). | Future<String> |
generateStream(prompt, config) |
Yields text chunks in real-time. | Stream<String> |
checkModelExists(repoId) |
Checks if model folder exists locally. | Future<bool> |
deleteModel(repoId) |
Deletes model from disk. | Future<bool> |
setCustomStoragePath(path) |
Sets/Resets model storage root folder. | Future<void> |
GenerateConfig Parameters
Detailed control over model output:
| Parameter | Type | Default | Description |
|---|---|---|---|
temperature |
double |
0.0 |
Higher = creative, Lower = deterministic. |
maxTokens |
int |
1024 |
Maximum length of generated response. |
topP |
double |
0.95 |
Nucleus sampling probability threshold. |
presence_penalty |
double |
0.0 |
Positive values penalize tokens already present. |
stopSequences |
List |
[] |
Tokens that stop generation (e.g. ["\nUser:"]). |
extraBody |
String |
null |
A JSON string for advanced backend options. |
Advanced Configuration (extraBody)
Pass advanced parameters through extraBody as a JSON string:
GenerateConfig(
extraBody: jsonEncode({
"top_k": 50,
"repetition_penalty": 1.2,
"chat_template_kwargs": {
"enable_thinking": true // Required for reasoning models like Qwen3.5
}
})
)
Thinking Mode (Reasoning) Guide
Thinking models (like Qwen/Qwen3.5-7B-Instruct) use a special template logic to externalize their reasoning process.
- Enable Reasoning: Include
"enable_thinking": trueinchat_template_kwargs. - Handle Output: The model will output text enclosed in
<think>...</think>tags before the final answer. - Hide Reasoning: Set
"enable_thinking": false. Depending on the model's template, it may output no tags or empty tags.
Error Handling
The plugin throws MlxEngineException for native errors. Always wrap calls in try-catch:
try {
await MlxLocalllm().generate(prompt: "...");
} on MlxEngineException catch (e) {
print("Code: ${e.code}, Message: ${e.message}");
}
简体中文
基于 Apple MLX 框架的 macOS/iOS 高性能本地大模型推理 Flutter 插件。通过将模型直接加载到应用进程中，提供零延迟的交互体验。
特性
- 🚀 原生加速: 基于 Apple MLX 针对 Apple Silicon 芯片深度优化。
- 📦 进程内推理: 无需安装 Ollama 等外部服务，零延迟通信。
- 🌐 鲁棒下载器: 内置分块下载机制，支持 Hugging Face 及其镜像站。
- 🛠️ 多平台支持: 同时支持 macOS 和 iOS 平台。
- 💻 架构兼容: 支持 x86 架构 Mock 编译，确保 Intel Mac 开发者也能正常运行项目。
平台支持
| 平台 | 支持 | 架构 | 最低系统版本 |
|---|---|---|---|
| macOS | ✅ | Apple Silicon (ARM64) / Intel (x86_64 Mock) | 14.0+ |
| iOS | ✅ | ARM64 | 17.0+ |
| Android | ❌ | - | - |
| Windows | ❌ | - | - |
| Linux | ❌ | - | - |
| Web | ❌ | - | - |
系统要求
- macOS: 14.0 及以上。
- iOS: 17.0 及以上。
- 硬件: Apple Silicon 系列芯片 (M1, M2, M3 等) 仅在这些芯片上支持推理。
- Flutter: 3.0.0 及以上。
安装指南
-
在
pubspec.yaml中添加依赖：flutter pub add mlx_localllm -
平台配置:
- macOS/iOS: 插件使用 Swift Package Manager (SPM) 管理 MLX 依赖。Flutter (3.24+) 会自动处理这些依赖。
- 架构要求:
- Apple Silicon (ARM64): 原生支持，具备 MLX 加速。
- Intel (x86_64): 支持 Mock 编译。您可以在 Intel Mac 上正常编译和链接项目，但推理调用将返回“不支持的架构”错误。这极大地方便了跨架构的协同开发。
API 参考
MlxLocalllm 单例方法
| 方法 | 描述 | 返回值 |
|---|---|---|
isSupported() |
检查硬件是否支持 MLX (需要 Apple Silicon GPU)。 | Future<bool> |
downloadModel(repoId) |
从 HuggingFace 开始后台下载。需通过 modelEvents 监控进度。 |
Future<bool> |
loadModel(path) |
从绝对路径或 repoId 将模型加载到内存。 | Future<bool> |
unloadModel() |
从内存中卸载模型，释放系统资源。 | Future<bool> |
generate(prompt, config) |
生成完整文本响应（阻塞式）。 | Future<String> |
generateStream(prompt, config) |
实时返回生成的文本片段。 | Stream<String> |
checkModelExists(repoId) |
检查模型文件夹是否已在本地存在。 | Future<bool> |
deleteModel(repoId) |
从磁盘删除模型。 | Future<bool> |
setCustomStoragePath(path) |
设置或重置模型的根存储目录。 | Future<void> |
GenerateConfig 配置参数
用于精确控制模型输出行为：
| 参数 | 类型 | 默认值 | 描述 |
|---|---|---|---|
temperature |
double |
0.0 |
越高越具创造性，越低越确定（Deterministic）。 |
maxTokens |
int |
1024 |
响应生成的最大 Token 数量。 |
topP |
double |
0.95 |
核采样（Nucleus sampling）概率阈值。 |
presence_penalty |
double |
0.0 |
正值会惩罚已出现的 Token，减少重复。 |
stopSequences |
List |
[] |
停止生成的序列（如 ["\nUser:"]）。 |
extraBody |
String |
null |
传递给原生层的高级 JSON 配置字符串。 |
高级配置 (extraBody)
通过 extraBody 以 JSON 字符串形式传递底层高级参数：
GenerateConfig(
extraBody: jsonEncode({
"top_k": 50,
"repetition_penalty": 1.2,
"chat_template_kwargs": {
"enable_thinking": true // 针对 Qwen3.5 等推理模型必填
}
})
)
推理模式 (Thinking Mode) 指南
推理模型（如 Qwen/Qwen3.5-7B-Instruct）使用特殊的模板逻辑来外部化其思维链（CoT）。
- 启动推理: 在
chat_template_kwargs中包含"enable_thinking": true。 - 输出处理: 模型会在最终回答前输出被
<think>...</think>标签包裹的思维过程。 - 关闭推理: 设置
"enable_thinking": false。根据模型模板不同，可能不输出标签或输出空标签。
错误处理
插件在发生原生错误时会抛出 MlxEngineException。建议始终使用 try-catch 包裹调用：
try {
await MlxLocalllm().generate(prompt: "...");
} on MlxEngineException catch (e) {
print("错误码: ${e.code}, 消息: ${e.message}");
}
许可证
MIT License.