What Is JetBrains Mellum2.1? Open Model for Coding Agents
The short answer
Mellum2.1 is JetBrains’ open model for coding agents: small, fast and meant to run on your own hardware. Released in October 2026 under Apache 2.0, it is a 12B mixture-of-experts model with 2.5B active parameters that JetBrains trained mainly with reinforcement learning so it can work inside a repository — explore the code, edit files and verify its own changes.
| Mellum2.1 | |
|---|---|
| Developer | JetBrains |
| Released | October 2026 (Mellum2 open-sourced June 2026) |
| Architecture | Mixture-of-experts, 12B total / 2.5B active parameters |
| Licence | Apache 2.0 |
| Training focus | Post-training with large-scale RL in sandboxed real environments |
| Speed | ~2x the throughput of Qwen3.5-9B under heavy load; ~1.6x faster single requests with multi-token prediction |
| Where to get it | Hugging Face (GGUF for llama.cpp, Ollama, LM Studio “coming soon”) |
What it is for
JetBrains positions Mellum2.1 for three jobs:
- A worker inside agent systems. It handles parts of an agent’s plan — finding the root cause of a failing test, drafting a fix and checking it — while a larger model plans.
- A general assistant. JetBrains reports gains across coding, competitive programming, maths, tool calling and general knowledge over Mellum2.
- Private, self-hosted deployment. Because it runs locally or on your own infrastructure, code and data stay under your control.
It is not a frontier model and does not try to be. The realistic pattern is a frontier planner such as Claude Sonnet 5.5 or GPT-6.1 Sol directing several cheap local Mellum2.1 sub-agents that do the file reading and test runs. See current API prices for what the planner costs.
What changed from Mellum2
The architecture is identical; the difference is post-training. JetBrains moved reinforcement learning from a short final stage to the main part of training, built in-house infrastructure to run thousands of RL environments, launched millions of sandboxes during training, and filtered open RL datasets for broken tests and unverifiable answers before using them. JetBrains compared it against Mellum2, Qwen3.5-9B and Gemma 4 E4B under one evaluation setup and reports the largest gain in agentic coding.
How to run it
Hardware. 12 billion parameters in 16-bit precision is about 24 GB of weights, so a 24 GB GPU or a large-memory workstation runs the full-precision model; a 4-bit build should need roughly 7–8 GB once GGUF files ship. Only 2.5B parameters are active per token, which is why it is fast once loaded.
Today. Download the weights from the JetBrains collection on Hugging Face and serve them with vLLM or Transformers.
Soon. JetBrains says GGUF builds for llama.cpp, Ollama and LM Studio, plus the multi-token prediction head for speculative decoding in vLLM, are coming. Our guide to running local LLMs on Windows covers those tools.
Should you use it?
Use it if you are building coding agents or sub-agents that must run on your own machines, or want a cheap, permissively licensed worker model to fan out many parallel tasks. Skip it if you need one model to plan and execute hard multi-file changes alone; a frontier model still does that better. For the wider field, see best local LLM to run in 2026.
Last verified: October 13, 2026.