JetBrains shipped Mellum2.1 on October 8 — a 12B MoE coding model with about 2.5B active parameters, Apache 2.0, trained with RL in real repositories. It is on Hugging Face, with GGUF called out as coming. That is the shape of model I actually want to test: small enough to iterate, licensed cleanly, aimed at agents instead of chat cosplay.
Why MoE matters on a desk
Active parameters are what your fans feel. A 12B MoE that only wakes 2.5B is trying to stay snappy for tool loops — propose edit, run test, read failure, try again. That is the coding-agent job. Giant models are fun demos. Tight loops win weeks.
I still want to see latency numbers on consumer GPUs and how it behaves when the repo is messy — half-finished modules, weird vendoring, tests that only fail on Tuesday. Blog benches are optimistic by nature. Real git status is not.
What I will try
Pull the HF weights when I have a free evening, point a local harness at a disposable repo, and see if Mellum2.1 can land a boring refactor without inventing APIs. If GGUF lands clean, that is when it becomes a laptop tool instead of a lab toy. Until then I will treat the announcement as promising and the merge as unproven.
None of this replaces reading the release notes yourself. Models and agent wrappers change weekly. My bar stays the same: license clarity, local option when secrets are involved, and a workflow I can reproduce without a sales call.
Pin the versions you validate. “Latest” is not a release strategy for anything that can write files.
Have a blessed day!