Mellum2.1 Thinking is a reasoning model from JetBrains, released as the next version of Mellum2 Thinking. The architecture is unchanged: a 12B-parameter mixture-of-experts model with 2.5B active parameters, 28 layers, 64 experts of which 8 are activated per token, and grouped-query attention with 32 query and 4 key/value heads. The context length is 131,072 tokens, and sliding-window attention with a 1,024-token window is used in 3 of every 4 layers. Most of the work in this version went into post-training, mainly reinforcement learning, with new tasks in math, competitive programming, science, tool use and software engineering. For software engineering the model is trained inside real repositories with a shell and file-editing tools, and is rewarded when tests pass. JetBrains positions it for complex agentic tasks such as working in a repository, running commands and calling tools, and for hard non-agentic problems in coding, math and reasoning. The card reports self-reported results from JetBrains' own pipeline in thinking mode against Mellum2 Thinking, Gemma 4 (E4B) and Qwen3.5 (9B): LiveCodeBench v6 82.0 (Mellum2 Thinking 69.4), SWE-bench Verified 47.0 (2.0 for Mellum2 Thinking, 50.0 for Qwen3.5 9B), AIME 25/26 83.3 (Qwen3.5 9B 86.7) and GPQA Diamond 64.6 (Qwen3.5 9B 77.8). The agentic runs use the Pi v0.73.1 harness with a 114K-token context. English is the listed language, weights are in bfloat16, and the model is served with vLLM using the qwen3 reasoning parser, with optional hermes-style tool calling. GGUF builds (BF16, Q8_0, Q6_K, Q4_K_M and MXFP4_MOE) are available in a separate JetBrains repository, while a multi-token prediction head for speculative decoding in vLLM is listed as coming soon. It is released under the Apache 2.0 license.