Mellum2.1 brings coding agent skills to self-hosted AI

Mellum2.1 adds repository-level coding-agent capabilities, letting the 12B mixture-of-experts model explore codebases, edit files, and check its changes.

The update keeps Mellum2's 2.5B active-parameter architecture and Apache 2.0 license, while post-training shifts heavily toward reinforcement learning in millions of sandboxed environments.

The project reports that multi-token prediction makes single requests about 1.6 times faster, and Mellum2.1 is now available on Hugging Face.