MiniMax-M1, the world’s first open-source large-scale hybrid architecture inference model

MiniMax, also known as Xiyu Technology, has released the world’s first open-source large-scale hybrid-architecture inference model, MiniMax-M1. The model delivers outstanding performance in complex productivity scenarios, approaching leading international standards, with exceptional cost-effectiveness. M1 supports up to 1 million context input tokens and 80,000 token inference output. Powered by a hybrid architecture and lightning attention mechanism, it significantly improves computational efficiency. Its reinforcement learning algorithm, CISPO, demonstrates excellent convergence performance, with a total training cost of only $537,400.