JetBrains releases Mellum2, an open 12B MoE for code workloads

JetBrains released Mellum2, a 12 billion parameter Mixture-of-Experts model that activates 2.5 billion parameters per token. It is open source under Apache 2.0 and targets text and code, not multimodal use. JetBrains pitches it as a building block for AI systems rather than a Claude or GPT replacement, with the cases they highlight being prompt routing, retrieval pipelines, sub-agent planning, and self-hosted deployments where data has to stay inside the network.

The pitch on numbers is inference speed. JetBrains claims more than 2x faster inference than similarly sized open models at competitive accuracy on code, reasoning, science, and math benchmarks. The active-parameter count is what does the work: at 2.5B active, runtime cost looks closer to a small dense model, while the 12B total gives more knowledge to fall back on. The release includes the model collection on Hugging Face and a technical report on arXiv with the training details.

The framing is worth reading carefully. JetBrains is not claiming a frontier-quality general assistant. They are saying that if you are stitching together a multi-step coding agent, the orchestrator and sub-agents do not need to be the smartest model you can buy, and a small fast MoE can fill those roles cheaper and with steadier latency.

Why it matters

If you build coding agents, the cost of the inner control-loop calls often decides whether a feature is viable. An Apache-licensed model you can self-host, that runs fast and is good enough for routing and code-review work, is the kind of component that quietly changes unit economics. Worth benchmarking against whatever small model your stack uses today.