MLX Tune

MLX Tune is a Python library for adapting AI models on an Apple Silicon Mac. Formerly unsloth-mlx, it is an independent project and is not affiliated with Unsloth. Its Unsloth-compatible API lets you prototype training locally with minimal changes to existing scripts, then move to NVIDIA hardware with Unsloth. Training runs through Apple's MLX framework and doesn't require a cloud GPU.

It covers instruction tuning, preference learning and reasoning training through methods such as SFT, DPO and GRPO. LoRA and quantized model training reduce the memory needed to adapt models, while continual pretraining lets you train on raw text for a particular domain or language. It also supports mixture-of-experts models, including Mixtral and Qwen3.5-MoE.

Beyond text, you can fine-tune Gemma 4 and Qwen3.5 for vision tasks, speech models such as Whisper and Qwen3-TTS, and OCR models including DeepSeek-OCR and GLM-OCR. Embedding training supports BERT, ModernBERT and Qwen3-Embedding for semantic search. JEPA support includes training LeJEPA from scratch and adapting pretrained image and video models.

The library requires macOS 13 or later and at least 8 GB of unified memory, with 16 GB recommended. It's open source under Apache 2.0. Exports include HuggingFace format, merged weights and GGUF for Ollama or llama.cpp; direct GGUF export from a quantized base model requires a non-quantized base or dequantization.