# Mellum2.1: JetBrains' open 12B model for coding agents

https://aycode.ai/blog/jetbrains-mellum-2-1/ · 2026-10-10 · News

JetBrains has released Mellum2.1: an open 12B model with 2.5B active parameters, trained with RL to work inside repositories. Apache 2.0.

JetBrains wants a model that can work inside a repository and still run on your own hardware. This October the company announced Mellum2.1 (announcement), the next version of the 12B model it open-sourced in June. The architecture is unchanged; everything after pre-training is new.

## What Mellum2.1 is

- Small and sparse: a mixture of experts with 12B parameters, 2.5B of them active per token, and a context of 131,072 tokens.
- A thinking model: the checkpoint is called Mellum2.1-12B-A2.5B-Thinking.
- Open: Apache 2.0, with the weights on Hugging Face.
- Trained to work in repos: reinforcement learning went from a short final step to the main part of training. For software tasks the model trained in real repositories with a shell and file tools and was rewarded when the tests passed, over millions of sandboxed runs. JetBrains says it now explores a codebase, edits files and checks its own changes.

## JetBrains' numbers

All of these are JetBrains' own figures, measured with one pipeline for every model, the agent tasks in the open-source agent Pi (version 0.73.1):

- SWE-bench Verified: 47.0% (Mellum2: 2.0%, Gemma 4 E4B: 23.0%, Qwen3.5 9B: 50.0%)
- SWE-bench Pro: 28.0% (Qwen3.5 9B: 38.0%)
- Terminal-Bench 2.1: 17.4% (Qwen3.5 9B: 21.7%)
- LiveCodeBench v6: 82.0% (Qwen3.5 9B: 75.4%)

The jump from Mellum2 on agent tasks is huge, but Qwen3.5 9B stays ahead there. Mellum's argument is speed: according to JetBrains it serves almost twice as many tokens as Qwen3.5 9B under heavy load, and multi-token prediction (MTP) makes a single request about 1.6 times faster.

## Running it yourself

The model card shows how to serve it with vLLM, with tool calling through the Hermes parser. The blog post calls builds for llama.cpp, Ollama and LM Studio "coming soon", but JetBrains already has a GGUF repository on Hugging Face: the recommended 4-bit file (Q4_K_M) is 8.1 GB, the smallest (MXFP4_MOE) 7.0 GB, and Ollama starts it with ollama run hf.co/JetBrains/Mellum2.1-12B-A2.5B-Thinking-GGUF:Q4_K_M. The MTP head for faster decoding in vLLM is still to come.

## Our take

A model with 2.5B active parameters at 47% on SWE-bench Verified, measured by its maker, will not replace the big models in a long refactor. Its place is the one JetBrains names: sub-agents, quick fixes, first looks at a failing test, and work where code must not leave the building. That JetBrains compares it with open models of its own class and shows where it loses makes the numbers easier to trust. Under Apache 2.0, it is worth a test.

## In AyCode

In a Codex chat, AyCode can discover models offered by a running Ollama or LM Studio without signing in (docs). That discovery does not establish compatibility: we have not tested Mellum2.1 end to end with Codex in AyCode. Tool calls and longer tasks still need checking on your setup. AyCode also runs Pi, next to Claude Code, Codex, OpenCode and other agents (all agents). Its model comparison sends a prompt to several models and displays their answers side by side.

Sources:
- JetBrains: Mellum2.1 Gets to Work: A Fast Open Model for Coding Agents (https://blog.jetbrains.com/ai/2026/10/mellum2-1-gets-to-work-a-fast-open-model-for-coding-agents/)
- JetBrains: Mellum2.1 model card (https://huggingface.co/JetBrains/Mellum2.1-12B-A2.5B-Thinking)
- JetBrains: Mellum2.1 GGUF builds (https://huggingface.co/JetBrains/Mellum2.1-12B-A2.5B-Thinking-GGUF)