JetBrains drops Mellum2, a 12B open model that runs 2x faster by only waking a fifth of its brain

JetBrains just open-sourced Mellum2, a 12-billion-parameter Mixture-of-Experts model built for the unglamorous, high-frequency work that makes production AI systems actually function. It’s a clear signal that the company, best known for its developer tools, is betting the future of AI isn’t one giant model answering every question, but a swarm of smaller, specialized ones handling the plumbing.

The architecture is the story here. While Mellum2 holds 12 billion parameters in total, it only activates 2.5 billion for any given token. That sparsity translates directly into speed. JetBrains claims more than 2x faster inference compared to similarly sized dense models. The technical report backs this up with competitive benchmark scores across code generation, reasoning, and math, meaning you’re not trading away significant performance for that efficiency gain. Crucially, this thing is not trying to be a multimodal Swiss Army knife. It’s strictly text and code, a deliberate focus that keeps it compact for software engineering workloads.

JetBrains envisions Mellum2 as the workhorse component inside a larger AI stack. They call it a “focal” model. Think: routing prompts to the right specialist model, compressing context in a RAG pipeline, or running planning and validation steps for coding agents. “Many of these operations are latency-sensitive and do not require the largest available model,” the company noted in its release. It’s the kind of task you hate wasting an expensive frontier model call on, and Mellum2 is designed to handle it cheaply and fast, either in the cloud or on private infrastructure.

The move also addresses a growing enterprise anxiety: data privacy. Released under the permissive Apache 2.0 license, Mellum2 can be deployed in self-hosted environments where proprietary code or internal data can’t touch a third-party API. For a company like JetBrains, whose entire business revolves around developers trusting their IDE, baking this option into the model’s DNA isn’t just a feature—it’s a strategic imperative. The model is available for download on Hugging Face now, and you can read the full architecture details in their paper on arXiv.