Mellum2 Goes Open Source: A Fast Model for AI Workflows

12B MoE (2.5B active), Apache 2.0, 64 experts. Targets routing, RAG compression, sub-agents, on-premises agentic AI. Not multimodal; text and code only.