Mellum2.1 is JetBrains' open-weight coding-agent model for engineering teams and enterprises that want fast, private code automation. Released on October 8, 2026, it is a 12-billion-parameter mixture-of-experts model that can explore a repository, edit files, and verify its own changes. It is the successor to Mellum2 and is tuned to act inside a codebase rather than merely autocomplete the next line.
Core Features
- 12B MoE with 2.5B active parameters per token and a 131,072-token context window.
- Apache 2.0 license; self-host on vLLM so source code and data never leave your network.
- Reinforcement learning inside real codebases, using millions of sandboxes across thousands of environments.
- LiveCodeBench v6 score of 82.0, ahead of the comparable Qwen3.5-9B at 75.4.
- Multi-token prediction delivers about 1.6x single-request speed and close to 2x the throughput of Qwen3.5-9B on one H200.
- A single H200 serves more concurrent sub-agents than a 9B rival, which directly lowers per-token cost at scale.
Use Cases
- A sub-agent inside larger agent systems: locate a failing test, draft a patch, and confirm the fix.
- Private or regulated codebases where compliance forbids sending code to a cloud model.
- High-concurrency coding farms where per-token throughput drives the bill more than peak quality.
Pricing
Mellum2.1 is free and open under Apache 2.0; download it from Hugging Face at no cost ($0), with no API fee. The only expense is your own GPU hardware for hosting.
Our Take
Best for teams that need a cheap, fast, on-prem coding worker rather than a giant frontier model; the trade-off is weaker end-to-end scores on Terminal-Bench 2.1 (17.4 versus Qwen3.5-9B's 21.7), so keep it on bounded, well-defined tasks. Compare it with other options in our AI Programming tools category.
