According to CoinMeta, the Xiaomi team has revealed the architecture that the next generation of MiMo-V3 will adopt, which is designed to address the issue of increasing costs associated with agent. This architecture allows agent to send only short commands each time, yet it can retrieve large amounts of content from web pages, terminals, and tools. The model is divided into two sections: the first section handles input, while the second section performs reasoning and generation, reducing the amount of computation required. Compared to the previous generation, when the input reaches 1 million token, the computational load is reduced to about 1/5, and the memory usage has decreased from 12.09GB to 2.69GB.