THE HOOK
JetBrains spent the summer beating its tiny open model with reinforcement learning, and it came out able to work inside a real repo.
THE GIST
Mellum2.1 is a 12B mixture-of-experts model with 2.5B active parameters, Apache 2.0 licensed and free on Hugging Face. Millions of sandboxed training runs taught it to explore codebases, edit files and check its own changes. JetBrains claims nearly double Qwen3.5-9B's throughput under heavy load.
THE TAKE
This is the right kind of small. Nobody needs a frontier model to find out why a test fails; they need a cheap sub-agent on the box under the desk. The catch is that GGUF builds aren't out yet, so the llama.cpp crowd waits. Fast beats clever for grunt work.
OPINION
The benchmarks in JetBrains' announcement pit Mellum2.1 against Qwen3.5-9B and Gemma 4 E4B, its own weight class, rather than cherry-picked frontier giants, which is refreshingly honest. When your IDE vendor ships an Apache-licensed agent model, the case for piping every keystroke to a cloud API gets noticeably weaker.
- Story type: Breaking news
- Impact: Medium
- Confidence: High (official announcement)
- Audience: Developers
- What changed: JetBrains released Mellum2.1, an RL-trained open 12B MoE model built for coding agents and self-hosted sub-agents.
- What to watch: GGUF builds and the vLLM multi-token prediction head, which decide whether local-model users can actually run it.
- Reading time: 2 minutes
