mlx-engine - Apple MLX LLM Engine for LM Studio

mlx-engine

MLX engine for LM Studio

Built with

• mlx-lm - Apple MLX inference engine (MIT)
• Outlines - Structured output for LLMs (Apache 2.0)
• mlx-vlm - Vision model inferencing for MLX (MIT)

How to use in LM Studio

LM Studio 0.3.4 and newer for Mac ships pre-bundled with mlx-engine. Download LM Studio from here

Standalone Demo

Prerequisites

• macOS 14.0 (Sonoma) or greater.
• python3.11
  • The requirements.txt file is compiled specifically for python3.11. python3.11 is the python version bundled within the LM Studio MLX runtime
  • brew install python@3.11 is a quick way to add python3.11 to your path that doesn't break your default python setup

Install Steps

To run a demo of model load and inference:

1. Clone the repository

git clone https://github.com/lmstudio-ai/mlx-engine.git
cd mlx-engine

2. Create a virtual environment (optional)

 python3.11 -m venv .venv
 source .venv/bin/activate

3. Install the required dependency packages

pip install -U -r requirements.txt

Text Model Demo

Download models with the lms CLI tool. The lms CLI documentation can be found here: https://lmstudio.ai/docs/cli Run the demo.py script with an MLX text generation model:

lms get mlx-community/Meta-Llama-3.1-8B-Instruct-4bit
python demo.py --model mlx-community/Meta-Llama-3.1-8B-Instruct-4bit 

mlx-community/Meta-Llama-3.1-8B-Instruct-4bit - 4.53 GB

This command will use a default prompt. For a different prompt, add a custom --prompt argument like:

lms get mlx-community/Mistral-Small-Instruct-2409-4bit
python demo.py --model mlx-community/Mistral-Small-Instruct-2409-4bit --prompt "How long will it take for an apple to fall from a 10m tree?"

mlx-community/Mistral-Small-Instruct-2409-4bit - 12.52 GB

Vision Model Demo

Run the demo.py script with an MLX vision model:

lms get mlx-community/pixtral-12b-4bit
python demo.py --model mlx-community/pixtral-12b-4bit --prompt "Compare these images" --images demo-data/chameleon.webp demo-data/toucan.jpeg

Currently supported vision models include:

• Llama-3.2-Vision
  • lms get mlx-community/Llama-3.2-11B-Vision-Instruct-4bit
• Pixtral
  • lms get mlx-community/pixtral-12b-4bit
• Qwen2-VL
  • lms get mlx-community/Qwen2-VL-7B-Instruct-4bit
• Llava-v1.6
  • lms get mlx-community/llava-v1.6-mistral-7b-4bit

Speculative Decoding Demo

Run the demo.py script with an MLX text generation model and a compatible --draft-model

lms get mlx-community/Qwen2.5-7B-Instruct-4bit
lms get lmstudio-community/Qwen2.5-0.5B-Instruct-MLX-8bit
python demo.py \
    --model mlx-community/Qwen2.5-7B-Instruct-4bit \
    --draft-model lmstudio-community/Qwen2.5-0.5B-Instruct-MLX-8bit \
    --prompt "<|im_start|>system
You are Qwen, created by Alibaba Cloud. You are a helpful assistant.<|im_end|>
<|im_start|>user
Write a quick sort algorithm in C++<|im_end|>
<|im_start|>assistant
"

Testing

To run tests, run the following command from the root of this repo:

python -m unittest discover tests

To test specific vision models:

python -m unittest tests/test_vision_models.py -k pixtral