JetBrains has released Mellum2.1-12B-A2.5B-Thinking, an open-source reasoning model designed for coding agents. The model features 28 layers and 64 experts, activating 8 experts per token. It supports a context length of 131,072 tokens and utilizes grouped-query attention with 32 query heads and 4 KV heads.
The development team focused on post-training, incorporating reinforcement learning tasks in software engineering, math, and science. The model was trained in millions of sandboxes across thousands of environments. It uses bfloat16 weights and maintains the same architecture as the previous version, ensuring comparable inference speeds.
Performance benchmarks show significant improvements in agentic coding, with SWE-bench Verified scores rising to 47.0 and Terminal-Bench 2.1 reaching 17.4. The model leads in LiveCodeBench v6 with a score of 82.0 and HumanEval+ with 91.5. Safety metrics also improved, as HarmBench scores dropped to 8.5.
JetBrains reports that Mellum2.1 serves nearly twice the tokens of Qwen3.5-9B on an H200 GPU under heavy load. Multi-token prediction provides a 1.6x speed increase for single requests. The model is compatible with vLLM, and GGUF builds are currently in development.
