// the find
ARahim3/mlx-tune
Fine-tune LLMs on your Mac with Apple Silicon. SFT, DPO, GRPO, Vision, TTS, STT, Embedding, and OCR fine-tuning — natively on MLX. Unsloth-compatible API.
mlx-tune wraps Apple's MLX framework in an Unsloth-compatible API so you can run the same fine-tuning script locally on a Mac that you'd run on a CUDA cluster with Unsloth. It covers an unusually wide surface: SFT, DPO, GRPO, ORPO, KTO, SimPO, VLM, TTS, STT, OCR, embeddings, MoE, and continual pretraining — all natively on Apple Silicon. Aimed at developers who prototype on a MacBook before paying for cloud GPU time.
The Unsloth API compatibility is the core value prop and it's executed well — the `from mlx_tune import FastLanguageModel, SFTTrainer` swap is genuinely a one-line change. The v0.5.0 GRPO optimisation (KV cache reuse across rollouts, one prompt forward per group) is a real algorithmic improvement, not just a marketing claim — 10x on GRPO is plausible given how naively most implementations do this. The DeepSeek-OCR dependency mess is documented honestly in 'Known Limitations' with an exact working pinset, which is more than most projects bother with. The breadth of audio support (Parakeet TDT with CTC/RNN-T/TDT losses, auto vocab extension for new languages) goes well beyond what you'd expect from a one-person Mac-focused project.
Audio training is batch_size=1 and RL training is single-sample — these aren't footnotes, they're training-speed walls that will hit you on any serious dataset. The project is essentially a wrapper on top of mlx-lm, mlx-vlm, and mlx-audio; when those libraries break or diverge (as the DeepSeek-OCR/transformers 5.0 situation already shows), mlx-tune inherits the breakage with no control over the fix timeline. The test suite exists but leans on mocked or trivially small models — there's no evidence that a real fine-tuned adapter actually improves task performance, and 'E2E tests' in a wrapper library usually just exercise the wiring, not the training math. GGUF export from quantized bases is broken by design (mlx-lm limitation), which is the exact workflow most people reaching for this library will want: QLoRA on Mac, deploy via Ollama.