Mellum2.1
JetBrains’ open 12B MoE (2.5B active) built to work inside coding agents.
Mellum2.1 is an open-weight 12B mixture-of-experts model (2.5B active) from JetBrains. The architecture is unchanged from Mellum2; gains come from post-training, mainly reinforcement learning. It is aimed at sub-agent work such as diagnosing failing tests and drafting fixes, and at private self-hosted deployment. JetBrains reports about 1.6× faster single-request decoding with its multi-token-prediction head on one H200 (vendor-reported).
- Provider
- JetBrains
- Type
- Language model
- Released
- Oct 8, 2026
- Input
- text
- Output
- text
- Open weights
- Yes
- License
- Apache-2.0
Best for
- Coding
- Agents
- Local & on-device
Strengths
- Open weights (Apache-2.0)
- 2.5B active parameters
- MTP head for fast decoding
Tools that use it
- Junie — JetBrains. JetBrains’ coding agent that plans first and debugs with the real IDE debugger.