mlx-tune
An Unsloth-compatible fine-tuning library, for large language, vision, speech and OCR models, that runs natively on Apple Silicon through Apple's MLX framework.
mlx-tune brings the Unsloth fine-tuning experience to Mac users by wrapping Apple's native MLX framework in an Unsloth-compatible API, so a training script written with FastLanguageModel can be prototyped locally on a Mac and later run unchanged on a CUDA cluster. It supports SFT, DPO and GRPO training methods for language models, plus fine-tuning for vision-language, text-to-speech, speech-to-text, embedding and OCR models.
It takes advantage of a Mac's unified memory, up to 512 GB on a Mac Studio, and can export trained models to the Hugging Face format or to GGUF for use with Ollama or llama.cpp. It was previously named unsloth-mlx and is an independent, unofficial project, not affiliated with Unsloth. It is written in Python under the Apache-2.0 license.