JetBrains released Mellum2.1 this week, the update to its open 12-billion-parameter model optimized for coding agents. The architecture remains the same as version 2 — mixture-of-experts with 2.5 billion active parameters per token — but post-training with reinforcement learning in real repositories produced a significant jump in agentic tasks.
What changed
According to the official announcement on the JetBrains blog, Mellum2.1 was trained with reinforcement learning in sandboxed environments that simulate real repositories. The model can now explore codebases, edit files, and verify its own changes — capabilities the previous version did not consistently achieve.
On SWE-bench Verified (real GitHub bug fixes), the score rose from 2.0% on Mellum2 to 47.0%. On SWE-bench Pro, from 0% to 28%. On Terminal-Bench 2.1, from 0.6% to 17.4%. On LiveCodeBench v6, it reached 82.0, ahead of Qwen3.5-9B (75.4).
Compared with Qwen3.5-9B in JetBrains' own tests, Mellum2.1 leads on short tasks and tool-calling, while the Chinese model still performs better on longer and more complex repositories. The numbers come from the company's internal evaluation; large-scale independent evaluations have not yet been published.
Why it matters
For developers and game studios that want coding agents running on their own hardware (without relying on cloud APIs), Mellum2.1 offers an efficient option under the Apache 2.0 license. JetBrains highlights throughput: under heavy load on one H200, the model serves nearly twice as many tokens as Qwen3.5-9B. A multi-token prediction head, still in preparation, promises to speed up individual requests further.
The model is available on Hugging Face as JetBrains/Mellum2.1-12B-A2.5B-Thinking, with a 131,072-token context. vLLM support already exists; Ollama and LM Studio have been announced. There is no official hosted API — the focus is self-hosting.
Sources
Transparency: This content was created, edited or reviewed with the assistance of artificial intelligence. The information was cross-checked with public posts on X and sources available on the internet. Consult the original sources to verify the full context.
By GeekikiBot