JetBrains launches Mellum2.1: a 12b moe model for coding agents

JetBrains has released Mellum2.1, a 12B mixture-of-experts model designed for coding agents and sub-agents. The model, available under Apache 2.0 on Hugging Face, retains the same architecture as its predecessor, Mellum2, but incorporates significant improvements through reinforcement learning (RL) in real software environments. Mellum2.1, with a checkpoint named Mellum2.1-12B-A2.5B-Thinking, is a reasoning model that generates its chain of thought before providing answers. It features 28 layers, 64 experts, and a router that activates 8 experts per token. The model supports a 131,072-token context length and uses grouped-query attention with 32 query heads and 4 KV heads. Mellum2.1 outperforms previous models in several benchmarks, including LiveCodeBench v6 (82.0), HumanEval+ (91.5), and BFCL v4 (62.3). Safety metrics also improved, with HarmBench scores dropping from 21.5 to 8.5. JetBrains claims the model is faster than Qwen3.5-9B, handling nearly twice as many tokens on an H200 under heavy load. The release includes technical documentation, model weights, and instructions for deployment, with support for vLLM and tool calling.