MiniMax M2
Overview
MiniMax M2 is a Mixture of Experts (MoE) language model developed by MiniMax, featuring a total of 230 billion parameters with 10 billion active parameters per forward pass. Released on October 23, 2025, it is designed to excel in coding and agentic tasks while maintaining robust general intelligence. The model's architecture allows for efficient deployment and scaling, making it suitable for a wide range of applications. MiniMax M2's strengths include advanced coding capabilities, such as multi-file editing, compile-run-fix loops, and test-validated repairs. It also demonstrates powerful agentic performance, effectively handling complex, long-chain tool-calling tasks and coordinating various tools like Shell, Browser, Python code interpreter, and MiniMax's proprietary tools. The model's cost-effectiveness and speed are achieved through an efficient design of activated parameters, balancing intelligence, speed, and cost. In benchmark evaluations, MiniMax M2 has shown competitive performance across general reasoning, tool use, and multi-step task execution, positioning it favorably among its peers. In comparison to other models, MiniMax M2 offers a favorable performance-to-cost ratio, operating at approximately 8% of the price of models like Claude Sonnet while delivering comparable or superior performance in various tasks. Its efficient design and robust capabilities make it a compelling choice for developers and organizations seeking a versatile and cost-effective solution for coding and agentic applications.