JetBrains has released Mellum2.1, an Apache 2.0-licensed mixture-of-experts model designed for coding agents. The 12B-parameter model activates only 2.5B parameters per forward pass, keeping inference costs relatively low for an agentic coding workload.
The headline result is the model's improvement on SWE-bench Verified, which jumped from 2.0 to 47.0 after reinforcement learning in real repositories. That gain suggests that RL grounded in realistic coding environments can substantially boost the problem-solving ability of a compact MoE model.
Mellum2.1 is positioned as an open, practical option for teams building coding agents, though the source does not provide direct comparisons with other open models. Its permissive license and small active parameter count make it easy to self-host.