MiniMax M2.5

MiniMax M2.5 is a large language model built from the ground up for agentic productivity, meaning it is designed to plan multi-step tasks, call tools, and move fluidly between different software environments rather than just produce single-turn answers. The model sits on a Mixture-of-Experts foundation with around 230 billion total parameters but only about 10 billion active per forward pass, a design that keeps inference cost low while preserving frontier-scale capability. It ships in two performance tiers, a standard variant running at roughly 50 tokens per second and a Lightning variant reaching about 100 tokens per second, and supports a very long context window with the underlying architecture extending to one million tokens. The intended sweet spot is real-world digital work: writing and refactoring code, searching the web, operating Word, Excel, and PowerPoint files, and coordinating across agent and human teams, with strong structured-output and tool-calling behavior built in.

The model extends the coding strengths of its predecessor M2.1 into broader office and general productivity territory, using a proprietary reinforcement learning framework called Forge that runs the model across more than 200,000 real-world environments spanning code repositories, browsers, and office applications. The Forge pipeline, which incorporates the CISPO algorithm and is reported to deliver a roughly 40x training speedup, optimizes how the model plans its actions and tokens rather than only chasing raw knowledge, producing measurable efficiency gains over earlier generations. This training focus is visible in the benchmark profile, where M2.5 reaches 80.2% on SWE-Bench Verified, 51.3% on Multi-SWE-Bench, 55.4% on SWE-Bench Pro, and 76.3% on BrowseComp, placing it within a small margin of leading proprietary systems at a fraction of the price. Released as open weights on Hugging Face under a modified MIT license, M2.5 is positioned as a flexible foundation for self-hosted coding agents, office automation, and multi-step workflows, while newer M2.7 updates build directly on the same agentic lineage.