A still life of a pewter jug, a peeled lemon, an Imari bowl, walnuts and a Mac Studio on a stone ledge.

Melix

Your models, at home on your Macs.

Melix is our inference engine for companies that deploy on Apple Silicon. Register models, serve them, fine-tune them with LoRA, then benchmark and evaluate them, from a native Mac app or the command line. Nothing leaves the machine.

The Melix window on macOS: a local server running a model with a LoRA adapter, its gateway and serving settings.
The Melix workspace on macOS, serving a model with a LoRA adapter on the local machine.

One loop, on one machine.

Melix product page: LoRA training and adapters with MLX, built for Apple Silicon.
Melix, from the product's own site.

Model registry

Import models from disk or from Hugging Face into a registry you control.

Serving

Start, pause, resume and stop local model servers with one command or one click.

LoRA and QLoRA

Train an adapter on your own data, then compare it with the base model side by side.

Benchmarks and evaluation

Run repeatable benchmarks and evaluation suites and keep the results in a format you own.

Native Mac app

A menu bar app and workspace for all of the above, backed by the same command line.

Melix is open source under the Apache 2.0 license.

Every commission begins with a conversation.

Tell us what the work is, what data you keep and where it has to run. We reply within two business days.

Start a projectbiz@random-walk.co.jp