FAQ¶

Is this installer affiliated with Steven Atkin or the upstream project?¶

No. This is an independent, unaffiliated packaging effort by Jay German. Steven Atkin created MLX Training Studio and all credit for the application belongs to him and contributors. This installer simply automates the clone-build-install workflow that upstream does not provide.

Why not publish a pre-built binary?¶

Two reasons:

1. Upstream choice. The upstream project publishes source only. Redistributing a binary of someone else's app raises licensing and maintenance questions even when the license (Apache-2.0) technically permits it.

2. Apple Silicon code signing. Distributing a non-sandboxed macOS app outside the App Store requires an Apple Developer notarization. Building locally means Gatekeeper sees the app as locally built, which requires only a one-time quarantine attribute removal rather than a notarization identity.

See Architecture — why source-only build for the full explanation.

Can I run this on an Intel Mac?¶

No. The upstream application is built on Apple's MLX framework, which targets Apple Silicon (arm64) exclusively. mlx-training-studio doctor will report a hard failure on Intel hardware:

[ERR ] Architecture 'x86_64' detected. Apple Silicon (arm64) is required.
[ERR ] Intel Macs are not supported by the upstream application.

There is no workaround — this is an upstream constraint.

How do I run training without the GUI?¶

Use mlx-lm-lora directly from the command line. Install it in a Python 3.12+ virtual environment:

python3.12 -m venv ~/mlx-venv
source ~/mlx-venv/bin/activate
pip install mlx-lm-lora

# Fine-tune a model
python -m mlx_lm.lora \
  --model Qwen/Qwen2.5-0.5B-Instruct \
  --data ~/datasets/my-data \
  --train \
  --batch-size 4 \
  --num-epochs 3

See the mlx-lm documentation for the full CLI reference.

Is this the same as "MLX Studio" or "LM Studio"?¶

No — these are three different applications:

Will this installer track upstream forever?¶

By default, yes — mlx-training-studio update fetches and resets to origin/main, so you always get the latest upstream commit.

To lock to a specific commit:

MLX_TS_REF=a3f9c21b84d7e05c1f2843b6d9a0e1f3c5d2e4b8 mlx-training-studio update

Run mlx-training-studio status to see the currently installed commit.

How do I check which version I have installed?¶

mlx-training-studio status

This prints the manifest JSON, which includes:

• installer_version — the version of this installer that performed the install.
• upstream_commit — the full commit SHA of the upstream source that was built.
• installed_at — the UTC timestamp of the last install or update.

Can I install to a location other than /Applications?¶

Yes. Set MLX_TS_INSTALL_DIR:

MLX_TS_INSTALL_DIR="${HOME}/Applications" mlx-training-studio install

Or choose option 2 (~/Applications) at the interactive prompt during mlx-training-studio install.

Does the installer need root / sudo?¶

The installer itself does not run as root. cp and mv to /Applications on macOS does not require root for bundles (it is writable by the owner on a standard macOS install).

The following steps do require sudo if you need them:

Where do my fine-tuned model adapters live?¶

Adapter weights are saved by mlx-lm-lora to the path specified in your training configuration (typically the output_dir key in the training YAML). The installer does not manage, read, or write adapter files.

By default, mlx-lm-lora saves adapters to an adapters/ directory relative to where it was invoked. Inside the MLX Training Studio app, the working directory is typically ~/Library/Application Support/MLX Training Studio/ or a subdirectory you configured in the training job. Check the training output log for the exact path printed after a completed run.

Is there a way to donate or support the upstream project?¶

This installer has no relationship with the upstream project. If you find MLX Training Studio useful, consider opening issues or contributing pull requests at stevenatkin/mlx-lm-gui.