MiniMax发布M2.5模型：1美元运行1小时，价格仅为GPT-5的1/20，性能比肩Claude Opus

Core Insights - MiniMax has launched its latest M2.5 series model, significantly reducing inference costs while maintaining industry-leading performance, aiming to address the economic feasibility of complex agent applications [1] - The M2.5 model demonstrates a substantial price advantage, costing only 1/10 to 1/20 of mainstream models like Claude Opus and GPT-5 at a throughput of 50 tokens per second [1][2] - The model has shown strong performance in programming tasks and has achieved first place in the Multi-SWE-Bench multilingual task, with a 37% improvement in task completion speed compared to its predecessor M2.1 [2] Cost Efficiency - M2.5 is designed to eliminate cost constraints for running complex agents, achieving a processing speed of 100 TPS, which is approximately double that of current mainstream models [3] - The model reduces the total token consumption for tasks, averaging 3.52 million tokens per task in SWE-Bench Verified evaluations, down from 3.72 million tokens in M2.1 [3] Programming Capabilities - M2.5 emphasizes system design capabilities in addition to code generation, demonstrating a native specification behavior that allows it to decompose functions and structures from an architect's perspective [4] - The model has been trained in over 10 programming languages and has shown a pass rate of 79.7% on the Droid platform and 76.1% on OpenCode, outperforming previous models [5] Task Handling Efficiency - In search and tool invocation, M2.5 exhibits higher decision maturity, achieving approximately 20% fewer rounds of consumption compared to previous versions while maintaining token efficiency [8] Office Applications - MiniMax has integrated industry-specific knowledge into M2.5's training, resulting in an average win rate of 59.0% in the Cowork Agent evaluation framework against mainstream models, capable of producing industry-standard reports and financial models [10] Technical Foundation - The performance improvements of M2.5 are driven by a large-scale reinforcement learning framework named Forge, which decouples the underlying training engine from the agent [14] - The engineering team has optimized asynchronous scheduling and tree-structured sample merging strategies, achieving approximately 40 times training acceleration [14] Deployment - M2.5 is fully deployed in MiniMax Agent, API, and Coding Plan, with model weights set to be open-sourced on HuggingFace for local deployment [15]