Mellum model releases

Mellum2.1 Thinking (JetBrains/Mellum2.1-12B-A2.5B-Thinking) is JetBrains' fast, open-weight code model for coding agents and fast sub-agents, released 2026-10-08. It is a sparse Mixture-of-Experts model with 12B total / 2.5B active parameters (28 layers, 64 experts with 8 activated per token, hidden size 2304, grouped-query attention with 32 query / 4 KV heads, a 1,024-token sliding window on three of every four layers, 98,304-token vocabulary, bfloat16), carrying over the Mellum2 architecture unchanged. It is a "thinking" (reasoning-before-answering) model post-trained from the Mellum2-12B-A2.5B base, with reinforcement learning in real repository sandboxes (shell + file-editing tools, rewarded on passing tests) as the main training stage. Text and code only, 131,072-token context. Open weights on Hugging Face under Apache-2.0, with Transformers / vLLM / SGLang serving and GGUF builds for llama.cpp, Ollama and LM Studio; an MTP head for vLLM speculative decoding is promised. Positioned for private self-hosting (no first-party hosted API or pricing). On one NVIDIA H200 under heavy load JetBrains reports it serves nearly 2x the tokens of Qwen3.5-9B, with multi-token prediction making single requests ~1.6x faster. Self-reported benchmarks (thinking mode): LiveCodeBench v6 82.0, HumanEval+ 91.5, MBPP+ 79.4, SWE-bench Verified 47.0, SWE-bench Pro 28.0, Terminal-Bench 2.1 17.4, BFCL v4 62.3, WorkBench 44.6, ToolHop 49.1, AIME 25/26 83.3, GSM-Plus 88.3, IFEval 90.6, GPQA Diamond 64.6, MMLU-Redux 87.8, MixEval-Hard 46.4 - large agentic-coding gains over Mellum2 (SWE-bench Verified 2.0 -> 47.0, SWE-bench Pro 0.0 -> 28.0, Terminal-Bench 2.1 0.6 -> 17.4). Figures vendor-reported.