MiniMax M2.5 Highspeed

MiniMax M2.5 Highspeed is an accelerated variant of a frontier-level AI model engineered specifically for scenarios that demand both intelligence and responsiveness. It inherits the core capabilities of the standard M2.5 family—including the ability to seamlessly manipulate Office documents, coordinate across multiple software tools, and maintain high reasoning precision—while delivering ultra-low latency inference through rigorous engineering optimization. The model is designed to function as a high-velocity engine for interactive applications and automated pipelines, capable of processing complex document streams and high-frequency calls in near real-time without compromising on logical depth or task-planning ability.

The M2.5 family was trained using reinforcement learning deployed across real-world environments involving hundreds of thousands of machines, enabling efficient inference and optimized task decomposition at scale. Source benchmarks show M2.5 achieving an 80.2% score on SWE-Bench Verified, reflecting its capability in software engineering tasks, and operating 37% faster than its predecessor M2.1 while maintaining frontier-level performance. This combination of speed, intelligence, and cost efficiency makes the Highspeed variant particularly well-suited for AI assistant integrations, automated productivity workflows, and applications where both rapid response and robust reasoning are essential.