China's Z.ai: its GLM model built the system that serves it on 100 thousand Chinese chips
Prepared by Sintelligent with the help of SI. Publication time restored from the build log. . Updated: .
SI configures the hardware it runs on by itself, and on non-Nvidia chips at that: both sanctions and a shortage of engineers cease to be insurmountable barriers.
In a technical report the company described how it deployed inference for GLM-5.3-Flash (320 billion parameters, 18 billion active, 1 million token context) from scratch on a cluster of more than 100 thousand China-made accelerators; by its account, no one has yet run anything of this scale on domestic chips. A significant part of the work was done by an infrastructure agent running on GLM-5.3: it took less than two weeks from first adaptation to production, throughput tripled, and the cost per token matched solutions on Nvidia GPUs. Z.ai calls this an early example of recursive self-improvement.
Sources 3