AI news · October 12, 2026
JetBrains releases Mellum2.1, a 12B open model for coding agents
JetBrains' Mellum2.1 is a 12B Apache 2.0 coding model with 2.5B active parameters; download it and compare real repo edits with Qwen3.5-9B before paying for an agent.
- total parameters
- 12B
- active parameters
- 2.5B
- model licence
- Apache 2.0
JetBrains released Mellum2.1 as a 12B mixture-of-experts model with 2.5B active parameters. It was post-trained with reinforcement learning in real coding environments, and JetBrains says it ran millions of sandboxed tasks across thousands of environments. The model can explore a repository, edit files and check its changes. Official support is listed for GGUF with llama.cpp, Ollama and LM Studio, with vLLM support coming later.
JetBrains reports nearly 2x the throughput of Qwen3.5-9B under heavy load and 1.6x faster single requests with multi-token prediction. Those are vendor benchmarks, so they need a local test on your own repositories and hardware. The comparison set is practical: Qwen3.5-9B offers a similar small open coding footprint, while Gemma 4 E4B is smaller but may fit more laptops. The useful question is edit quality per dollar, not raw tokens per second.
What you can do with it
Download the model card from Hugging Face, run a small issue in Ollama or LM Studio, and score compile success, test changes and revert rate against your current model. It is a good fit for private repo work where a 12B download is easier to run than a hosted frontier coding agent.
Our take
Mellum2.1 matters because it targets the full edit-check loop instead of just code completion. The catch is that the strongest speed claims come from JetBrains, and 12B still needs meaningful memory for comfortable local use. A one-hour repo trial will tell you more than the headline benchmark.
Links Try itJetBrains release
Source: JetBrains — Made With Models writes the brief; the reporting is theirs.