JetBrains open-sources Mellum2.1, a 12B MoE coding agent model trained with RL at scale
lmoroney · x · 2026-10-09
JetBrains released Mellum2.1, an open-source (Apache 2.0) coding agent model. It keeps the 12B mixture-of-experts architecture with 2.5B active parameters per token, but nearly all the work went into post-training.
- RL was upgraded from a short final stage to the core of training, using in-house infrastructure that ran millions of sandboxes across thousands of real environments
- The model can now explore a repo, edit files, and verify its own changes
- New RL tasks span math, competitive programming, science, tool use, and software engineering; open datasets were heavily filtered to fix broken tests and remove unverifiable or trivial tasks
It targets fast coding agents and sub-agents running on your own hardware.
Related event: JetBrains Open-Sources Mellum2.1, a 12B MoE Agentic Coding Model(4 posts)→
More from coding & agent
- Satori's next version adds 3D transforms, calc(), min/max/clamp() and more CSS — shuding · 2026-10-09
- Gigacity: a browser city grown entirely from one seed with pure shaders — Promptmethus · 2026-10-09
- Veteran dev: AI-coded it, I never read the code, but it's not vibe coding — judgment still matters — mjuric · 2026-10-09
- shadcn: the most important coding skill in the AI era is reading, not summarizing — shadcn · 2026-10-09
- Philosopher OKF: open-source LLM skill turns any topic into a structured study page — holyshitthatsucks · 2026-10-09
- Codex throttled to 5 tok/s as dev argues local model deployment is the only fix — lxfater · 2026-10-09