News · 2026-10-10
JetBrains ships Mellum 2.1: a 12B open-weights model retrained for coding agents
JetBrains released Mellum 2.1 on 8 October 2026 under Apache 2.0: the same 12B mixture-of-experts architecture as Mellum 2 with 2.5B active parameters, post-trained mostly with reinforcement learning across millions of sandboxed runs in thousands of environments so it can explore a repository, edit files, and check its own changes. JetBrains reports the largest gains in agentic coding, improvements across coding, competitive programming, math, tool calling and general knowledge, roughly 1.6 times faster single-request output from multi-token prediction, and almost twice the served tokens of Qwen3.5-9B under heavy load. Weights are on Hugging Face now; GGUF builds for llama.cpp, Ollama and LM Studio are listed as coming soon.
What it means for buyers
Teams that want an agent worker model on their own hardware get a new Apache 2.0 option sized for a single accelerated card, from the vendor that already ships a coding model inside its IDEs; until the GGUF builds land, llama.cpp, Ollama and LM Studio users need the Hugging Face conversion route. Against cloud coding agents it remains a compact model with open-weights privacy as the selling point, so the practical test for buyers is agentic quality per gigabyte against Qwen and Gemma alternatives.
Where to check it
Sources
- Mellum2.1 Gets to Work: A Fast Open Model for Coding Agents (JetBrains AI Blog) — launch post, 8 October 2026, checked 2026-10-10
- JetBrains Mellum 2.1 collection (Hugging Face) — open weights, checked 2026-10-10