JetBrains Releases Mellum2.1 Coding Model

JetBrains has rolled out Mellum2.1, a coding model built for agentic coding. It keeps the 12B mixture-of-experts design from Mellum2 with 2.5B active parameters, and stays open source under the Apache 2.0 license.

The big change is reinforcement learning, which is now the main part of training instead of a short final step. JetBrains added data for math, coding contests, and software engineering, ran millions of sandboxes during training, and removed low-quality tasks beforehand.

The model can now explore codebases, edit files, and check its own fixes. Multi-token prediction makes it about 1.6x faster per request, and its throughput nearly doubles Qwen3.5-9B under heavy load. It is on Hugging Face for local deployment, so companies can keep code and data in-house.