JetBrains releases Mellum 2.1, a 12B MoE coding model with 2.5B active parameters, under Apache 2.0

JetBrains published Mellum 2.1 on Hugging Face, a 12-billion-parameter mixture-of-experts model that activates only 2.5 billion parameters per token. It is released under the permissive Apache 2.0 license and designed specifically for local coding agents. JetBrains reports a 47% score on SWE-Bench, the benchmark that measures whether a model can resolve real GitHub issues.

The MoE design is what makes local use realistic. With only 2.5B parameters active per token, inference speed is closer to a small dense model, while the 12B total gives the model more stored knowledge. That combination lets it run responsively on laptops and workstations, keeping source code on the developer's machine, a requirement for many companies with strict IP or compliance rules. Apache 2.0 licensing removes commercial-use questions that have complicated other open models.

Mellum 2.1 enters a busy week for local coding. Satya Nadella announced that GitHub Copilot now hands off work to local models like MAI-Code-1.1 Flash, a 137B-parameter model; DeepSeek shipped desktop apps for its Harness coding agent; and NVIDIA and Microsoft unveiled RTX Spark PCs built for on-device inference. Mellum sits at the lightweight end of that spectrum. A 47% SWE-Bench score trails frontier cloud models such as Claude Opus 5.5 and GPT-6 Astra by a wide margin, but it is respectable for a model this small. Ai2 separately released new Bolmo byte-level checkpoints alongside a Nature paper.

Caveats: SWE-Bench scores depend heavily on the agent harness, and the source reporting doesn't specify which variant or setup was used. Small models also tend to struggle with long multi-file changes. The best fit is likely as a fast inner-loop model handling routine edits, with harder tasks routed to cloud models, the same cost-per-task routing pattern seen across the industry. Watch for community harness results and integration into JetBrains IDEs.