[P] mlx-tune – Fine-tune LLMs on Apple Silicon with MLX (SFT, DPO, GRPO, VLM)

Developer A-Rahim has launched mlx-tune, a significant new open-source library that brings sophisticated large language model fine-tuning capabilities natively to Apple Silicon Macs. Built on top of Apple's MLX framework (mlx-lm and mlx-vlm), the tool allows developers to run full training workflows—including Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and newer methods like GRPO and KTO—directly on a Mac with 8GB or more of unified memory. Its API is designed to mirror popular CUDA-based frameworks like Unsloth and TRL, meaning a training script written for NVIDIA GPUs can run on a Mac by changing just a single import line. This dramatically lowers the barrier for local prototyping and experimentation.

The library is specifically positioned as a tool for local development and prototyping before scaling workloads to expensive cloud GPU instances. It supports parameter-efficient techniques like LoRA and QLoRA, includes chat templates for 15 major model families, and can export trained models to the GGUF format for use with tools like llama.cpp. While not intended to replace high-performance CUDA training for large-scale jobs, mlx-tune fills a crucial gap by enabling data scientists and ML engineers to iterate quickly on their fine-tuning recipes using the hardware they already own, potentially saving significant cloud costs during the development phase.