MLX-Knife 2.0
A text-first command-line tool for managing and running MLX models on Apple Silicon, covering listing, cloning, converting, quantizing and serving, with curated support for select vision and audio models.
MLX-Knife works with any standard MLX text model, such as Llama, Mistral, Qwen, Phi or Gemma, through list, clone, run, convert, quantize and serve commands, and adds curated support for a verified set of vision and audio models, including Whisper, Pixtral, Gemma-3, Qwen2-VL and VibeVoice. It explicitly does not try to wrap every model type that its underlying mlx-vlm and mlx-audio libraries can load, rejecting unverified model types at conversion time rather than silently producing a broken workspace.
It supports native audio transcription through Whisper, vision analysis that automatically extracts EXIF metadata such as GPS and camera details, cloning and converting models without repeated round-trips to Hugging Face, resumable downloads, automatic batching to avoid Metal out-of-memory crashes, and a beta feature for piping models together over Unix pipes without temporary files. It makes no background network calls or telemetry beyond explicit Hugging Face interactions.
A recent bugfix release refuses to load a checkpoint that names its own Python file as its model file, closing a code-execution vulnerability. It is written in Python and licensed under Apache-2.0.