Mellum2.1 keeps Mellum2's architecture (12B total, 2.5B active, 64 experts with 8 per token, 131,072-token context) and changes the training. JetBrains says most of the work went into reinforcement learning, including software engineering tasks run inside real repositories with a shell and file tools, rewarded when tests pass.
JetBrains' own numbers, run with the open-source Pi agent harness for every model: 47.0% on SWE-bench Verified (Mellum2 scored 2.0%), 28.0% on SWE-bench Pro and 82.0% on LiveCodeBench v6. Qwen3.5 9B, in the same table, still scores higher on SWE-bench Verified (50.0%) and SWE-bench Pro (38.0%). All figures are self-reported.
The MTP repository adds one extra block that drafts the next token for speculative decoding. JetBrains says that makes decoding 1.4 to 1.8x faster with unchanged output and about 90% of drafted tokens accepted. It needs vLLM 0.33+ or transformers 5.20+; GGUF builds for llama.cpp, Ollama and LM Studio are in a separate repo. The license is Apache 2.0.