mlx-tune

Overview of mlx-tune

What it is

mlx-tune is a tool for fine-tuning large language models (LLMs), vision-language models (VLMs), text-to-speech (TTS), speech-to-text (STT), embedding, and OCR models on Apple Silicon Macs. It wraps Apple's MLX framework in an API compatible with Unsloth, allowing users to write training scripts locally on a Mac and then run the same script on CUDA clusters. The project is not an official Unsloth project and does not aim to replace Unsloth; its goal is code portability between local Mac prototyping and cloud GPU production.

Key features

• Fine-tune LLMs, VLMs, TTS, STT, embedding, and OCR models natively on MLX.
• Unsloth-compatible API: FastLanguageModel, SFTTrainer, and other classes work with the same code as Unsloth.
• Supports multiple training methods: SFT, DPO, ORPO, GRPO, KTO, SimPO, and JEPA-based objectives (LeJEPA, I-JEPA, V-JEPA 2, LLM-JEPA).
• Export models to HuggingFace format or GGUF for use with Ollama/llama.cpp.
• Supports MoE fine-tuning, continual pretraining, and audio fine-tuning (5 TTS models, 7 STT models).
• Includes JEPA family: from-scratch ViT SSL (LeJEPA), pretrained I-JEPA, V-JEPA 2 video world model, and LLM-JEPA for LLM fine-tuning.

Use cases

• Prototype and experiment with fine-tuning on a local Mac before committing to cloud GPU costs.
• Develop and iterate training pipelines with fast local feedback loops.
• Scale up to cloud NVIDIA GPUs with original Unsloth for production training, using the same script.
• Fine-tune audio models (TTS, STT) and OCR models locally.

Getting started

Install mlx-tune using uv (recommended): bash uv pip install mlx-tune After installation, import FastLanguageModel and SFTTrainer from mlx_tune instead of unsloth. See the repository README for detailed quick-start examples.

License and pricing

The project is licensed under Apache-2.0 and is open source.

Based on the project's public README and repository metadata.