mlx-tune

Why we included this project

mlx-tune is a local fine-tuning tool for Apple Silicon Macs, built on MLX rather than CUDA, so a MacBook or Mac Studio with unified memory can train models without needing a cloud GPU. It covers the usual LLM methods (SFT, DPO, GRPO, KTO, SimPO) and branches into vision, speech-to-text, text-to-speech, embedding, OCR, and continual pretraining. The main appeal for people already using Unsloth is that the API is compatible: changing the import is often all it takes to run an existing training script locally, then move to CUDA later for bigger production runs. Trained models export to HuggingFace, GGUF, and merged weights, which feeds them straight into Ollama or llama.cpp. If your fine-tuning work normally lives in the cloud, this gives you a lighter way to prototype small- and mid-size models on the machine in front of you.