MLX-MemOS

MLX-MemOS is a high-performance LLM serving and RAG (Retrieval-Augmented Generation) infrastructure toolkit optimized for Apple Silicon (macOS). Built on top of Apple's MLX framework, it provides a seamless experience for running large language models (like Qwen3) and embedding/reranking models locally with OpenAI-compatible APIs.

🚀 Key Features

• Apple Silicon Optimized: Leverages MLX for efficient inference on Mac devices (M1/M2/M3/M4).
• OpenAI Compatible: Provides a drop-in replacement for OpenAI's Chat Completions API.
• RAG Ready: Includes a dedicated server for Embeddings (bge-m3) and Reranking (bge-reranker-v2-m3).
• Model Management: Ready-to-use scripts for converting and managing Qwen3 models (0.6B, 4B, 8B, 14B).
• Production Friendly: Includes startup/shutdown scripts, PID management, and logging.
• Benchmarking: Built-in tools to stress test and verify model performance.

📋 Prerequisites

• macOS 13.0+ (Ventura or later recommended)
• Python 3.10+
• Apple Silicon (M-series chip)

🛠️ Installation

1. Clone the repository

   git clone https://github.com/yourusername/MLX-MemOS.git
   cd MLX-MemOS

2. Create and activate a virtual environment

   python3 -m venv .venv
   source .venv/bin/activate

3. Install dependencies

   pip install -r requirements.txt

🏗️ Model Preparation

41→ 42→MLX-MemOS expects models to be placed in the models/ directory. 43→ 44→### ⚠️ Important: Restore Large Models 45→ 46→Due to GitHub's file size limits, some large model files (over 2GB) are split into chunks. You must run the following command after cloning to restore them: 47→ 48→bash 49→./script/manage_large_files.sh merge 50→ 51→ 52→This will reassemble files like pytorch_model.bin and model.safetensors from their split parts. Specifically, it handles:

• models/bge-m3/pytorch_model.bin
• models/bge-reranker-v2-m3/model.safetensors
• models/Qwen3-8B-MLX/model.safetensors
• models/Qwen3-4B-MLX/model.safetensors
• models/Qwen3-14B-MLX/model-00001-of-00002.safetensors
• models/Qwen3-14B-MLX/model-00002-of-00002.safetensors 53→ 54→### Download & Convert Models 55→ 56→We also provide scripts to help you convert Hugging Face models to MLX format.

# Example: Convert Qwen3-14B
./script/convert_qwen3_14b.sh

# Example: Convert Qwen3-8B
./script/convert_qwen3_8b.sh

Ensure you have sufficient disk space and memory for the conversion process.

🚦 Usage

1. Start the LLM Server (Chat Completions)

This starts an OpenAI-compatible server hosting the LLM (default: Qwen3-14B-MLX).

./script/start_mlx_server.sh start

• Port: 8080
• Endpoint: http://127.0.0.1:8080/v1/chat/completions
• Logs: logs/mlx_server.log

To stop or restart:

./script/start_mlx_server.sh stop
./script/start_mlx_server.sh restart
./script/start_mlx_server.sh status

2. Start the Embedding & Rerank Server

This starts a separate server for text embeddings and document reranking.

./script/start_embedding_server.sh start

• Port: 8081
• Endpoints:
  • Embeddings: http://127.0.0.1:8081/v1/embeddings
  • Rerank: http://127.0.0.1:8081/v1/rerank
• Logs: logs/embedding_server.log

3. Verification

Verify that the servers are running correctly:

# Verify LLM Server
./script/verify_mlx_curl.sh
# OR using Python script
python script/verify_mlx_server.py

# Verify Embedding/Rerank Server
./script/verify_embedding_server.sh

📊 Benchmarking

You can benchmark the performance of the LLM server using the included Python script:

python script/benchmark_mlx.py

This script will simulate concurrent requests and report token generation speeds (TPS) and latency metrics.

📂 Project Structure

MLX-MemOS/
├── models/                 # Model checkpoints (MLX format)
├── script/                 # Operation scripts
│   ├── start_mlx_server.sh         # Manage LLM server
│   ├── start_embedding_server.sh   # Manage Embedding server
│   ├── convert_*.sh                # Model conversion scripts
│   ├── verify_*.sh                 # Verification scripts
│   └── benchmark_mlx.py            # Performance testing
├── logs/                   # Server logs
├── requirements.txt        # Python dependencies
└── README.md               # Project documentation

📜 License

This project is licensed under the MIT License - see the LICENSE file for details.