JetBrains Releases Mellum2.1, an Open Model for Coding Agents

What happened

Per JetBrains' official blog, Mellum2 was fast but couldn't work inside a repository at the level the team wanted. For Mellum2.1, reinforcement learning went from a short final stage to the main part of training: millions of sandboxed runs across thousands of in-house RL environments, with new RL tasks in math, competitive programming, science, tool use and software engineering — and hard filtering of open datasets (broken tests, unverifiable answers, too-easy or impossible tasks were removed).

Official comparisons show the biggest gain over Mellum2 in agentic coding, with across-the-board improvements in coding, competitive programming, math, tool calling and general knowledge; it was benchmarked against Qwen3.5-9B and Gemma 4 E4B in the same class. With the architecture untouched and multi-token prediction (MTP), it's the fastest model in the group under heavy load — serving nearly twice the tokens of Qwen3.5-9B and about 1.6x faster on a single request.

The model is on Hugging Face; GGUF builds for llama.cpp, Ollama and LM Studio, plus the MTP head for speculative decoding in vLLM, are coming soon. JetBrains positions it for developers building coding agents, sub-agents and AI tools on their own infrastructure, keeping code and data under their control.