Xiaomi MiMo Lead: Research Efficiency Improved Nearly Tenfold Using Claude Code, Model Self-Iteration "Achievable Within One to Two Years"

According to 1M AI News's monitoring, at the Open Source-themed roundtable of the Beijing Zhongguancun Forum, Xia Lixue, the founder of Wuwen Xinquan, revealed that since the end of January this year, the company's token usage has basically doubled every two weeks, and has now increased tenfold. He likened this growth rate to the feeling when mobile data traffic was just popularized in the 3G era, believing that existing resources are far from sufficient to support this growth and require better optimization and integration. Wuwen Xinquan also provides inference infrastructure services for Kimi, Zhipu, and multiple universities.

Xia Lixue made a more fundamental judgment: the current large amount of cloud computing infrastructure is designed to serve human engineers rather than AI. Intelligent agents can initiate tasks in milliseconds, but the response speed of the underlying infrastructure is still designed for human-minute-level operations, limiting the space for intelligent agents to operate. He believes that the infrastructure of the AGI era should itself become an intelligent agent, capable of self-evolution and iteration.

The roundtable was moderated by Yang Zhilin, the founder of Moon's Dark Side, with the participation of Zhang Peng, the founder of Zhipu, Luo Fuli, the head of Xiaomi's MiMo large model team, and Huang Chao, Assistant Professor at the University of Hong Kong. Zhang Peng revealed that the core goal of Zhipu's latest release, GLM-5-Turbo, is to shift from "conversation" to "action," where the token consumption for completing complex tasks may be ten times or even a hundred times that of answering simple questions, which is also the reason for Zhipu's price increase. Luo Fuli stated that with the combination of Claude Code and top models, the research efficiency of the Xiaomi large model team has increased by nearly tenfold, and she predicts that the self-iterative ability of the models will be achieved in one to two years, significantly ahead of her estimate of three to five years a year ago.