MLX-VLM is a Python package for inference and fine-tuning Vision Language Models on macOS using MLX. It supports multi-modal models and a CLI, plus optional UI components.
Collecting history — the radar snapshots this repo daily. The trend line appears after 3 days of data (1 so far).
What it is
MLX-VLM is a package for inference and fine-tuning of Vision Language Models (VLMs) and Omni Models (VLMs with audio and video support) on your Mac using MLX.
How it works
The repository provides a Python package (mlx-vlm) with a CLI for generation and a server for local inference, supports speculative decoding and various drafter-based acceleration methods, and offers model-specific documentation and utilities. It includes an agent-skills bundle for integration with coding agents and a chat UI via Gradio as an optional UI.
Getting started
Installation:
pip install -U mlx-vlm
Optional UI:
pip install -U 'mlx-vlm[ui]'
CLI usage examples are provided in the README for generating text, images, and audio, as well as multi-modal prompts. Example snippets include how to run generation commands and how to start the server with thinking mode options:
mlx_vlm.generate --model mlx-community/Qwen2-VL-2B-Instruct-4bit --max-tokens 100 --prompt "Hello"
Recent releases
Latest releases (latest 5): v0.6.10 (2026-08-04): includes MLX-VLM agent skills enhancements and prefill adjustments. v0.6.9 (2026-08-03): adds Kimi K3, fixes UTF-8 detokenizer issue. v0.6.8 (2026-07-27): vendors standard models and sanitizer fixes. v0.6.7 (2026-07-23): adds Laguna S support and Gemma/Inkling updates. v0.6.6 (2026-07-20): vendors text models and warm-match fixes.
Traction
Stars: 5296 (as of the provided data).
Behind the repo
Not provided beyond the repository linkage and usage details in the README.
Caveats
License: MIT. Creation date: 2024-04-16. Last push: 2026-08-04. Open issues: 145. Language: Python. Notes:
- The Gradio chat UI requires an extra dependency
mlx-vlm[ui]. - The README provides extensive usage sections, including speculative decoding (DFlash, Gemma 4 MTP, EAGLE-3, MiniMax M3), server options, and conversion/quantization workflows.
- Commands must be quoted exactly as written when using shells that expand brackets. E.g., installing UI extras:
pip install -U 'mlx-vlm[ui]'.