MLX-LoRA-Studio

Welcome to MLX LoRA Studio

A native Mac app for LLM fine-tuning on Apple Silicon — fully on-device, fully open source.

MLX LoRA Studio turns fine-tuning into a normal Mac workflow: pick a model, choose a dataset, select an algorithm, watch live training metrics, generate synthetic data, and publish adapters to Hugging Face — without leaving the window, and without your data ever leaving your Mac.

It is a graphical front-end to the mlx-lm-lora Python training pipeline, vendored at vendor/mlx-lm-lora/, so what runs in the GUI is exactly what you can run from the CLI.

Requirements

Features at a glance

🧠 Training

📊 Live observability

🧪 Synthetic data

🚀 Publish

🛠 Engineering safeguards

Algorithms (detailed reference)

Grouped into three families — supervised, preference, and reinforcement/online. Each page covers the loss, the math, the dataset shape, when to use it, the settings table, and failure modes.

Supervised

Preference

Reinforcement / online

Adaptation methods & foundations

Orthogonal to the loss: which tensors are trainable, how gradients become updates, and how the forward pass is quantised.

Adaptation methods

Quantization

Optimizers

App sections

Pages mirroring the app’s sidebar:

See also