Unsloth

Making model adaptation practical

Unsloth develops software for adapting existing models and running them locally. The company, started by brothers Daniel and Michael Han, works on the training workflow around models rather than presenting every supported checkpoint as its own research release. Dataset preparation, training configuration and model export are central parts of that workflow. [1][2]

Unsloth Studio launched in beta in March 2026. Its local interface brought training, inference and dataset tools into one application. The launch account also marks the limits of availability: Mac inference support did not mean MLX training had already shipped, and planned multi-GPU work was distinct from the beta's delivered features. [2]

The repository applies Apache-2.0 to the core training library, tests and scripts, while Studio and its command-line component use AGPL-3.0. Those are materially different terms for people modifying or hosting the software. Supported models and datasets also retain their own conditions; a training-tool license does not replace them. [3]