MLX LM LoRA native model training View on GitHub
Open source · Apple Silicon

Shape the model.
Keep the control.

A native MLX toolkit for fine-tuning language models locally, with adapters, preference optimization, and quantization-aware training built into one calm, capable workflow.

Start training Advanced notebooks
Runs on the machine in front of you.
mlx_lm_lora / train
current surface Fine-tune with intent.
01model = Qwen3-8B
02train_mode = sft
03train_type = lora
04quantization = qat · 4-bit
adapter rank16
devicemps / native
12+training algorithms
4 to 16bitQAT precision range
1local command surface
MLXnative on Apple Silicon
01 / Capability map

Everything you need to
move a model forward.

Keep the loop close to the work. MLX-LM-LoRA puts efficient adaptation, preference learning, and evaluation in the same local toolkit.

01

One local command

Move from a base model to a focused training run without leaving your terminal or your machine.

mlx_lm_lora.train●
02

Preference-native

DPO, CPO, ORPO, GRPO, online DPO, and more. Train toward the behavior you actually want.

03

Quantized by design

Use QLoRA and QAT to keep memory practical while preserving a direct path to better models.

04

Bring your own signal

Register custom reward functions and shape GRPO around the signal that matters to your domain.

02 / Method library

Choose the pressure.
Keep the surface simple.

Supervised fine-tuning

Teach the model a sharper way to speak.

Use prompt and completion pairs or chat messages to turn a capable base model into one that understands your task, tone, and format.

Read the method notes
run recipeSFT
mlx_lm_lora.train \ --model <model> \ --train-mode sft \ --data <dataset>
ready for your data
03 / Quick start

Make the first run
feel close.

Install the package, point it at a model and dataset, then keep iterating where your weights live: locally.

Advanced notebook runsStep-by-step examples for fine-tuning, preference learning, and RL. MLX Community datasetsDatasets formatted for direct use with MLX-LM-LoRA.
Read the docs PyPI package
terminal / zsh
pip install -U mlx-lm-lora

mlx_lm_lora.train \
  --model <model> \
  --train \
  --data <dataset> \
  --iters 600
local · private · yours
The local edge

Better models begin
with closer loops.

Build with MLX-LM-LoRA