JetBrains releases Mellum2.1 Thinking, a 12B coding MoE with 2.5B active, under Apache 2.0
Mellum2.1 Thinking is a 12B mixture-of-experts reasoning model with 2.5B active, retrained mostly with reinforcement learning on real repositories.
JetBrains published Mellum2.1 Thinking, the next version of Mellum2 Thinking. The architecture is unchanged: 12B total and 2.5B active parameters, 28 layers, 64 experts with 8 active, and a 131,072-token context. Almost all the new work went into post-training, where reinforcement learning became the main stage of training.
The card says the model trained inside real repositories with a shell and file-editing tools, rewarded when tests pass, over millions of sandboxed runs. It reports self-measured scores of 82.0 on LiveCodeBench v6, 47.0 on SWE-bench Verified and 17.4 on Terminal-Bench 2.1. JetBrains marks these as self-reported, and the model-index flags them as not verified. In its table Qwen3.5 9B scores higher on SWE-bench Verified.
The licence is Apache 2.0. GGUF builds and an MTP head for speculative decoding are described as coming soon.
The card gives no release date; the date used is the Hugging Face repo creation date.
Honest enough to print the row where it loses.
Sources
Editor's first pass (AI-assisted)
- Real 4: Apache 2.0, ungated, vLLM serving shown; GGUF builds and an MTP head are 'coming soon'.
- Useful 3.5: Repository work with a shell and file editing; the card's own figures are modest next to larger models.
- Notable 3: 173 likes; a known IDE maker's second open model.
- Novel 3: Same architecture as Mellum2; the change is the RL post-training.
Comments