MiniMax-M2.5 is a frontier large language model developed by MiniMax, released in February 2026. Built on a Mixture-of-Experts (MoE) architecture, the model features approximately 230 billion total parameters, with only 10 billion parameters active during any single inference pass. This design is intended to deliver high-performance capabilities in complex reasoning and coding while maintaining high inference efficiency and lower operational costs compared to dense models of similar scale.
The model is specifically optimized for software engineering and agentic workflows, demonstrating high proficiency in multi-file code generation, debugging, and autonomous tool-calling tasks. It was trained using a proprietary reinforcement learning framework called Forge, which emphasizes task decomposition and planning. This training approach enables what the developers describe as a native "architect mindset," where the model outlines structural plans and system designs before executing implementation details.
MiniMax-M2.5 supports a 204,800-token context window, allowing it to analyze and process large codebases or lengthy technical documents in a single prompt. The model is available as open-weights and is also offered through official API services in both Standard and high-throughput Lightning variants.