What it does

mlx-tune wraps Apple’s MLX framework in an Unsloth-compatible API. You import FastLanguageModel and SFTTrainer from mlx_tune instead of unsloth and trl, and the rest of your script stays identical. It supports SFT, DPO, GRPO, vision-language models, TTS, STT, embeddings, and even OCR fine-tuning — all on Apple Silicon, using unified memory up to 512 GB.

The interesting bit

The author calls this the “Context Switch” problem: Unsloth depends on Triton, which doesn’t run on Macs. Rather than reimplementing Unsloth or claiming speed parity, mlx-tune optimizes for code portability. Prototype on an M4 Mac, then move the exact same script to a CUDA cluster for production training. The v0.5.0 release adds real optimizations — GRPO roughly 10× faster via KV-cache reuse, gradient checkpointing wired into every trainer, and ORPO now fitting in 48 GB at 4096 context — but the core pitch remains workflow, not benchmarks.

Key highlights

• Unsloth-compatible API: swap imports, keep your training scripts
• Broad modality coverage: LLMs, vision (Gemma 4, Qwen3.5, LLaVA), audio (5 TTS and 7 STT models), embeddings, OCR, and MoE architectures
• Exports to HuggingFace format and GGUF for Ollama/llama.cpp
• convert() utility for HF → MLX model conversion
• Continual pretraining support with decoupled learning rates

Caveats

• GGUF export has limitations with 4-bit base models; the README points to workarounds
• Merging into a 4-bit base with save_method="merged_4bit" can round away weak LoRA deltas — 16-bit merge is preferred
• Apple Silicon only; no Intel Mac or cross-platform support

Verdict

Grab this if you own a recent Mac and want to iterate on fine-tuning pipelines without burning cloud GPU credits. Skip it if you’re already on CUDA full-time — Unsloth itself is still the gold standard there, and mlx-tune doesn’t pretend otherwise.