JetBrains released Mellum2.1 on October 8, 2026: an open coding model with 12 billion total parameters and just 2.5 billion active per token, trained with large-scale reinforcement learning in real code environments — and it scores 47.0 on SWE-bench Verified, up from 2.0 for its predecessor. It’s free under the Apache 2.0 license, including for commercial use.
Prague — October 9, 2026
What changed
Mellum2.1 keeps the same mixture-of-experts architecture JetBrains open-sourced in June 2026. Nearly all the work for this version went into post-training: reinforcement learning moved from a short final stage to the main part of training, running millions of sandboxed runs across thousands of environments. JetBrains also filtered its training data hard — open RL datasets often ship with broken tests, unverifiable answers, or tasks that are too easy or impossible — before training on them.
The result is a model built for a specific job: fast sub-agents and coding agents that run on your own hardware. A quantized Q4_K_M build is about 8.1 GB, small enough to run locally. Benchmarks reported on the model’s card include 47.0 on SWE-bench Verified, 82.0 on LiveCodeBench v6, and 62.3 on BFCL v4.
Why it matters
The headline is the training method, not the model size. JetBrains is betting that a small, fast, openly licensed model tuned in real coding environments can do more for coding agents than another oversized proprietary model — the routing, summarizing, and fast sub-agent work where latency and cost dominate. That’s the unglamorous half of every agentic coding system, and it’s exactly where renting intelligence from a hyperscaler hurts most.
For independent developers and small teams, this is immediately practical: a real, improving open model you can run locally, inside your own tools, without API bills or data leaving your machine. Mellum2.1 won’t replace Claude Code or Cursor on its own — JetBrains positions it as the component underneath, paired with a frontier model for hard reasoning — but it meaningfully lowers the cost of building with coding agents. Everything AI, tested: when the open model is good enough, renting stops being the default.
Quick questions
Is Mellum2.1 free for commercial use? Yes. It’s released under the Apache 2.0 license, which permits commercial use, modification, and distribution with no API costs.
Can I run it on my own machine? Yes — quantized builds around 7–8 GB make it feasible on local hardware, via vLLM or Hugging Face Transformers.
How much better is it than Mellum2? JetBrains reports its SWE-bench Verified score rose from 2.0 to 47.0, driven by reinforcement learning in real environments.
Sources: JetBrains AI Blog, MarkTechPost, Oossa