MiniMax发布新一代M2.5模型 推动生产级Agent大规模部署

Core Insights - MiniMax has launched its new generation text model, MiniMax M2.5, aimed at addressing cost and performance bottlenecks in AI Agent deployment [2][3] - The model demonstrates superior performance in programming capabilities and tool invocation, achieving an 80.2% score in the SWE-Bench Verified benchmark [2] - M2.5's inference speed and cost control are highlighted, with the lightning version supporting over 100 TPS output speed and competitive pricing for input and output tokens [3] Performance Metrics - M2.5 scored 80.2% in programming capability benchmarks, showcasing its advanced programming skills [2] - The model has improved tool invocation efficiency by 20% compared to its predecessor, achieving better results with fewer iterations in complex tasks [2] - In office applications, M2.5 has shown significant enhancements in financial modeling across Word, PPT, and Excel [2] Cost Efficiency - The M2.5-lightning version offers an output speed of over 100 TPS, approximately double that of mainstream models [3] - Input costs are around $0.3 per million tokens, while output costs are approximately $2.4 per million tokens [3] - The theoretical cost for running four agents continuously for a year is about $10,000, indicating a potential shift in the economic model for agent deployment [3] Deployment and Accessibility - M2.5 was launched on February 12 and is now available for global open-source support and local deployment [4]