MiniMax M2.5 is an open-weight large language model developed by MiniMax and released on February 12, 2026, positioned around real-world digital productivity rather than narrow benchmark tuning. Its sparse Mixture-of-Experts architecture carries roughly 229B total parameters while activating only about 10B per token, a design choice that keeps inference economical without sacrificing capability on coding, agentic tool use, web search, and office-document work. The model is text-only, with MiniMax routing vision and audio through separate companion models, and the launch shipped in two variants that share weights but differ in speed: the standard M2.5 and a faster M2.5-Lightning.
Training leaned heavily on reinforcement learning across more than 200,000 simulated digital working environments, and the model was exposed to over ten programming languages spanning full-stack development, which produced an architect-like planning habit where features, structure, and UI are decomposed before code is written. On coding evaluations M2.5 reaches 80.2% on SWE-bench Verified and 76.3% on BrowseComp, while also generating deliverable Word, PowerPoint, and Excel artifacts. Independent provider analysis places M2.5 at an Intelligence Index of 27 with 97 output tokens per second and the lowest blended price among tracked MiniMax models, making it a pragmatic fit for teams that need long-context agentic workflows at competitive cost.
