JetBrains released Mellum 2.1 this month. It is still the 12 billion parameter mixture-of-experts model from June, with about 2.5 billion parameters active, under Apache 2.0. What changed is the training after pretraining: reinforcement learning inside real environments, millions of sandboxed runs, until the model could explore a repository, edit files, and check the result. The announcement is on the JetBrains blog.
Mellum 2 was fast and not good enough inside a codebase. 2.1 is the version they are willing to call a worker: find why a test failed, draft a fix, look again. GGUF builds for local runners are still on the way. The weights are on Hugging Face.
The race is also about where the model sits
Frontier models are getting the headlines. A 12 billion parameter model that you can run next to the repository is the other lane, and it is the one a company with a private codebase actually asks me about. The code does not have to leave the building for every keystroke.
I would not hand it the whole job. JetBrains is clear that it improved as a piece of an agent: root cause, a patch, a check. That matches how I already want routing to work. A small model touches the files. A stronger model gets the ambiguous design question. The diff is what you review, not the paragraph about the diff.
What “checks its own changes” still needs
- Tests that already exist. A model that runs them is only as honest as the suite.
- A sandbox. Millions of training runs were sandboxed. Your laptop should be too.
- A human on the merge. Self-check is not review.
If you are picking a model for a coding agent this week, ask where the weights run before you ask which benchmark won. Mellum 2.1 is interesting because the answer can be your own machine.