MiniMax M2.5-highspeed

MiniMax M2.5-highspeed is a throughput-optimized variant of M2.5, running the same Mixture-of-Experts weights and architecture with inference-layer optimizations. It delivers the full M2.5 capability suite at higher output speed for high-volume and multi-turn agentic workflows. With a 204,800-token context window and support for long outputs, it preserves M2.5's agentic tool-use strengths, including 76.3% on BrowseComp and 76.9% on BFCL, while prioritizing inference speed. The model supports native tool calling and extended reasoning. M2.5-highspeed suits real-time agentic applications and high-throughput workflows where speed matters more than maximum context utilization. Recommended sampling settings are temperature 1.0, top_p 0.95, and top_k 40.