JetBrains has released Mellum2.1, a 12-billion parameter mixture-of-experts open-source model designed specifically for coding agents and fast sub-agents. Available under the Apache 2.0 license on Hugging Face, the model activates 2.5 billion parameters per token and features a 131,072-token context length. The architecture remains unchanged from its predecessor, with performance gains achieved almost entirely through reinforcement learning in real software environments where the model explores repositories, edits files, and verifies its own changes.
In benchmark evaluations, Mellum2.1 achieves an 82.0 score on LiveCodeBench v6 and a 91.5 on HumanEval+, while showing massive improvements in agentic coding tasks such as SWE-bench Verified, which rose from 2.0 to 47.0. The model runs locally via vLLM or SGLang, delivering nearly twice the token throughput of comparable models under heavy load on a single NVIDIA H200 GPU due to its small active parameter footprint.
This release matters because it provides developers with a powerful, self-hostable reasoning model optimized for autonomous software engineering tasks that can be efficiently run on local hardware.