MiniMax-M2.5 is MiniMax’s flagship native Agent MoE LLM with 229B total parameters and 10B active parameters. It has a 204K context window and is trained by reinforcement learning on massive real-world environments. It achieves SOTA coding performance and excels at task decomposition, tool calling and office productivity tasks with extremely high cost-performance, ideal for software development, long-chain Agent workflows and complex text processing.