JetBrains open-sources Mellum2.1 with scaled RL post-training for agentic coding
huggingface · x · 2026-10-08
JetBrains released Mellum2.1, a major update to the Mellum2 model open-sourced in June. The release comes from significantly scaling RL post-training to enable agentic coding workflows, achieving top agentic coding benchmark scores relative to its speed.
The 12B (MoE, 2.5B active) Thinking model weights are available on Hugging Face in both HF and GGUF formats, with MTP support coming in the following days.
More from coding & agent
- Agent memory, part 4: storage is one INSERT — retrieval at the right moment is what actually breaks — _jaydeepkarale · 2026-10-08
- Manus 2.0 launches with Video Editor, Game Dev tools and multi-agent Cue app — thetripathi58 · 2026-10-08
- OpenAI's agentic software factory detailed — but zero verified cases of fully autonomous long-horizon coding — alex_verem · 2026-10-08
- Codex nearly unusable: 51-task goal runs 46 hours, only half done, says user — manuelkoelman · 2026-10-08
- Make Claude finish overnight tasks: reusable harness with finish line, verifier agent, progress notes — PrajwalTomar_ · 2026-10-08
- COLM 2026 posters: MetaLint easy-to-hard generalization and measuring slop in long-horizon coding agents — dan_fried · 2026-10-08