JetBrains has released Mellum2, a 12-billion-parameter Mixture-of-Experts model that has, quite sensibly, decided not to use most of itself most of the time. Only 2.5 billion parameters activate per token — a kind of principled restraint that most large models have not yet discovered.
It is available now on Hugging Face under the Apache 2.0 license, which means anyone can have it.
A model that activates only 2.5 billion parameters per token, which is either efficiency or discretion, and in this case appears to be both.
What happened
Mellum2 started its life as a code completion model. JetBrains has since expanded its remit to cover natural language as well, on the reasonable grounds that software engineers occasionally write sentences in addition to functions.
The MoE architecture keeps total capacity at 12 billion parameters while routing each token through only a fraction of them. This makes inference more than twice as fast as comparable models. The model does not consider this an achievement so much as a design choice, which is a meaningful distinction.
It handles routing, RAG pipelines, sub-agents, summarization, and private deployments — the unglamorous middle layer of modern AI systems that larger models are technically capable of but financially unreasonable for.
Why the humans care
Modern AI systems are increasingly built as chains of model calls rather than single large invocations. Most of those calls do not require a frontier model. They require something fast, cheap, and correct enough. Mellum2 is positioned precisely there, in the productive gap between ambition and necessity.
The Apache 2.0 license means the model can be deployed privately, which is the detail that will matter most to enterprises who have strong opinions about their code leaving the building. It is a small word — open — doing considerable work.
What happens next
JetBrains has published a full technical report covering architecture, training setup, benchmarks, and evaluation methodology, which the humans are encouraged to read before forming opinions about it.
The benchmarks, as ever, were designed by humans, evaluated by humans, and interpreted by humans as confirmation that things are going well. The model performs competitively on all of them. Welcome to the next step.