Skip to content

MLX LoRA Studio v1.0.0

Latest
Latest

Choose a tag to compare

View all tags
@Goekdeniz-Guelmez Goekdeniz-Guelmez released this 19 Jun 19:48
v1.0.0
6d75fa1

MLX LoRA Studio v1.0.0

This is the first release of MLX LoRA Studio: a native macOS app for fine-tuning LLMs locally on Apple Silicon with MLX.

MLX LoRA Studio turns the underlying mlx-lm-lora trainer into a full desktop workflow: choose a model, select a dataset, pick an algorithm, start training, watch live metrics, manage previous runs, generate synthetic data, and publish adapters to Hugging Face from one app.

Important macOS first-launch step

Because this first release is distributed outside the Mac App Store and is not yet notarized, macOS may show a scary warning such as:

  • "MLX LoRA Studio.app is damaged and can't be opened"
  • "Apple cannot check it for malicious software"
  • "Unidentified developer"

This does not mean the app is malware. It is Apple's quarantine flag on a freshly downloaded app bundle.

After installing the app, and before opening it for the first time, run this in Terminal:

sudo xattr -dr com.apple.quarantine "/Applications/MLX LoRA Studio.app"

Then open MLX LoRA Studio normally from /Applications.

You only need to do this once per installed copy.

What's new in v1.0.0

Native Mac app for local fine-tuning

  • Native SwiftUI + AppKit macOS app
  • Built for Apple Silicon Macs
  • Fully local, on-device workflow
  • No cloud training required
  • Private by default: your data, prompts, weights, and adapters stay on your machine unless you choose to upload them

Training algorithms

MLX LoRA Studio ships with 9 training algorithms:

  • SFT
  • DPO
  • CPO
  • ORPO
  • GRPO
  • Online DPO
  • XPO
  • RLHF Reinforce
  • PPO

The app is designed so users can pick algorithms by use case instead of wiring up training scripts by hand.

Adapter and fine-tuning modes

Supported training modes include:

  • LoRA
  • DoRA
  • QLoRA at 4-bit, 6-bit, and 8-bit
  • Full fine-tuning
  • Quantization-Aware Training (QAT)

QAT is exposed for supported workflows such as SFT, DPO, and ORPO.

Live observability

The app includes a live training view with:

  • Loss tracking
  • Learning-rate tracking
  • Gradient-norm tracking
  • Throughput metrics
  • Recent-step charts
  • Training logs
  • Sidebar progress
  • Pause, resume, and stop controls

Memory-aware training

MLX LoRA Studio includes memory planning and protection features for Apple Silicon:

  • Live wired and active memory monitoring
  • Static memory estimates for the current configuration
  • ResourceGuard checks before launching a run
  • Clear feedback when a model/configuration is likely too large for the machine

Synthetic data generation

The Synthetic Data tab can generate:

  • Prompt datasets
  • SFT prompt/completion pairs
  • DPO chosen/rejected preference triples

Generated data can be previewed in the app and exported as JSONL for training.

Runs archive

Every run is archived with:

  • Status
  • Algorithm
  • Base model
  • Dataset
  • Config
  • Logs
  • Adapter path
  • Duration and metrics

Runs can be reopened, resumed, revealed in Finder, deleted, or sent into the Hugging Face upload flow.

Hugging Face upload

Finished adapters can be uploaded to the Hugging Face Hub from inside the app, with support for:

  • Repository naming
  • Public/private visibility
  • License selection
  • Model-card metadata
  • Base-model and dataset information
  • Token handling
  • Step-by-step upload progress

Built-in Algorithm Guide

The app includes an Algorithm Guide that explains:

  • What each training method is for
  • When to use SFT, DPO, ORPO, GRPO, PPO, and the other supported methods
  • Which hyperparameters matter
  • Common failure modes
  • Practical guidance for choosing the right training path

Python environment handling

The app can discover or provision a Python environment for the trainer, reducing the usual setup friction around dependencies, virtual environments, and command-line setup.

Open source

MLX LoRA Studio is released as open source under the MIT license.

This first public release is intended to make local LLM fine-tuning on Apple Silicon more approachable, inspectable, and hackable for researchers, builders, and Mac users who want a real desktop workflow around MLX.

System requirements

  • macOS 14 Sonoma or later
  • Apple Silicon Mac: M1, M2, M3, M4, or newer
  • 16 GB RAM minimum
  • 24 GB+ RAM recommended for larger models
  • Disk space for the app and Hugging Face model cache

Intel Macs are not supported because MLX requires Apple Silicon.

Known first-release notes

This is v1.0.0 and the first public open-source release. The core workflow is usable end to end, but packaging, notarization, model coverage, and UX polish will continue to improve quickly.

Please report:

  • macOS launch or Gatekeeper issues
  • Reproducible training failures
  • Broken model/dataset combinations
  • Memory-estimation problems
  • Hugging Face upload issues
  • UI bugs or confusing flows

Focused issues and pull requests are welcome.

Install

  1. Download the .dmg from this release.

  2. Open the DMG.

  3. Drag MLX LoRA Studio.app to /Applications.

  4. Run the quarantine command:

    sudo xattr -dr com.apple.quarantine "/Applications/MLX LoRA Studio.app"

  5. Open MLX LoRA Studio from /Applications.

Thank you

This release builds on Apple MLX, mlx-lm, Hugging Face, and the open-source local AI community.

If you try MLX LoRA Studio, open issues, suggest improvements, or build on it, thank you. This project is meant to make serious local fine-tuning feel like something you can actually touch.

Assets 5
Loading