blaizzy/mlx-vlm is a Python package that supports inference and fine-tuning of vision language models (Vision Language Models, VLM) on Mac using MLX. It also covers Omni models that support audio and video, and the usage paths are provided as a CLI, a Gradio chat UI, a Python API, and a FastAPI server. The latest release shown on GitHub is v0.7.9.
The model processing and fine-tuning paths are broad. It supports Hugging Face model conversion and quantization, and distributed inference is also possible. For fine-tuning you can choose between LoRA, QLoRA, DoRA, and full fine-tuning, and you can use SFT, ORPO, DPO, and GRPO objectives. The license is MIT, the implementation language is Python, and the GitHub star count is shown as 5640.
What can be confirmed is limited to the feature list. There are no performance comparison figures, so per-model accuracy or processing speed is unknown, and the v0.7.9 changelog is not recorded either, so it is not possible to know what this release fixed. The fact that the runtime environment is limited to Mac should also be taken into account. The only recorded score left is momentum 0.7, and the novelty, technical, and quality scores are not present. You can narrow the candidate down by feature scope alone, and performance remains to be verified separately.