Skip to content- Sources: primary, discussion
- Summary: Mellum2.1 keeps the Mellum2 architecture of 12B total and 2.5B active parameters and moves RL from a short final stage to the main part of training, including rewards for passing tests in real repositories. JetBrains reports SWE-bench Verified at 47.0 versus 2.0 for Mellum2 and 50.0 for Qwen3.5 9B, all self-measured. An included multi-token prediction head speeds decoding 1.4 to 1.8x with vLLM 0.33 or newer.
- Why it matters: It is a first-party account of RL-heavy post-training on a small open-weight agentic model that runs locally with speculative decoding.