JetBrains open-sources Mellum2.1: a 12B MoE with 2.5B active parameters, trained with large-scale RL as a self-hostable coding sub-agent

Why it matters

【Key points】

On Oct 8 JetBrains released Mellum2.1, the successor to the Mellum2 model it open-sourced in June, again under Apache 2.0. The architecture is unchanged: a 12B mixture-of-experts model with 2.5B active parameters. Nearly all the work went into post-training, with reinforcement learning becoming the main stage. JetBrains built infrastructure to run thousands of RL environments and launched millions of sandboxes, covering math, competitive programming, science, tool use and software engineering, filtering open data for broken tests and unverifiable answers.

JetBrains says Mellum2.1 can now explore a repository, edit files and check its own changes, with the biggest gains in agentic coding. Compared with Qwen3.5-9B and Gemma 4 E4B, it serves almost twice the tokens of Qwen3.5-9B under heavy load, and multi-token prediction makes single requests about 1.6x faster. All comparisons are JetBrains' own; no independent evaluation yet. Weights are on Hugging Face; GGUF builds for llama.cpp, Ollama and LM Studio and an MTP head for vLLM are coming.

【Verifiable】

Official blog: https://blog.jetbrains.com/ai/2026/10/mellum2-1-gets-to-work-a-fast-open-model-for-coding-agents/

Models: https://huggingface.co/collections/JetBrains/mellum21

【Our read】

1. The pitch is clear: not to out-think big models but to be the cheap worker an agent calls over and over, like finding a failing test's root cause and drafting a fix. It competes on the same line as low-cost small models such as Haiku 5.5, but runs on your own hardware.

2. Same architecture, gains from post-training alone: for small models, RL in real environments pays off more than adding parameters.

3. For companies, a self-hosted coding sub-agent keeps code inside the network. Real performance awaits independent tests and community runs once GGUF builds ship.