JetBrains released Mellum2.1, a coding AI model retaining a 12B MoE architecture with 2.5B active parameters under Apache 2.0, with enhanced agentic coding capabilities and reinforcement learning expanded as the main training stage. Its high-load inference throughput is nearly twice that of Qwen3.5-9B, and MTP improves single-request speed by about 1.6x. The model is now available on Hugging Face, with GGUF versions and vLLM components to follow.