MiniMax M2.5

MiniMax M2.5 is designed for real-world productivity, positioned by its creator as a frontier large language model that extends the coding expertise of the earlier M2.1 generation into general office work. According to the provider, M2.5 was trained across a diverse range of complex real-world digital working environments, reaching fluency in generating and operating Word, Excel, and PowerPoint files, switching context between diverse software environments, and working across different agent and human teams. The model applies reinforcement learning in complex real-world environments spanning hundreds of thousands of machines to achieve efficient inference and optimized task decomposition.

The model demonstrates strong benchmark performance against frontier competitors, scoring 80.2% on SWE-Bench Verified, 51.3% on Multi-SWE-Bench, and 76.3% on BrowseComp, while being more token efficient than previous generations through planning-based optimization of actions and outputs. Its release has been framed as offering comparable performance to leading Western frontier models at a fraction of the cost, positioning it as a competitive option in the productivity AI space. M2.5 is available both in a standard version and an M2.5-Lightning variant, giving users flexibility between full capability and faster inference depending on workload needs.