Introducing Mellum2: A 12B Mixture-of-Experts Model by JetBrains
Hugging Face
JetBrains dropped Mellum2, a 12B mixture-of-experts model built for code and text, but it only fires up 2.5B params per token. That means it runs over twice as fast as similar models, and it's free under Apache 2.0.
Based on reporting by Hugging Face — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
JetBrains just put out Mellum2, and it's a pretty clear statement about where they think production AI is headed: not toward one giant model doing everything, but toward a stack of specialized pieces working together. Mellum2 is meant to be the fast, cheap piece — the one that handles routing, summarization, retrieval cleanup, and agent sub-tasks so you don't have to burn a frontier model's compute on busywork.
Under the hood it's a 12-billion-parameter Mixture-of-Experts model, but only 2.5 billion of those parameters actually activate for any given token. That's the whole trick. You get the capacity of a much bigger model without paying its full inference cost every time, which is exactly what you want for latency-sensitive jobs like tool selection or context compression in a RAG pipeline. JetBrains says this setup gets them more than 2x faster inference than comparably sized open models, while still holding its own on code, reasoning, science, and math benchmarks.
This isn't JetBrains' first swing at this kind of model — Mellum started life as a code-completion tool, tuned for IDE-style autocomplete. Mellum2 broadens that into general natural-language and software-engineering work, but it deliberately skips multimodal ambitions. No images, no audio, just text and code, kept lean on purpose so it stays fast and easy to self-host.
JetBrains is pitching this as a "focal" model, language that fits a broader shift happening across the industry: instead of one model doing planning, retrieval, validation, and generation, you chain together several smaller specialists, each tuned for its slice of the job. Mellum2 slots in as the router or the sub-agent, not the star of the show.
The model ships under Apache 2.0, with weights on Hugging Face and a technical report laying out the architecture and eval methodology. For anyone building agent workflows or private, code-heavy deployments where sending every request to a massive model is overkill, it's worth a look.
My take — AI-written commentary, not fact-checked reporting
This is a sane release, and I say that as someone tired of every model launch pretending to be the last model you'll ever need. JetBrains isn't chasing the frontier — they're building the plumbing that makes frontier models cheaper to actually use in production, and Apache 2.0 licensing on a real MoE model is the kind of open-source contribution that matters more than another leaderboard flex. More teams should build the boring, efficient middle layer instead of racing for AGI headlines.
Read more about this at: Hugging Face