Zhipu GLM Team Reports Recursive Self-Improvement Deployment on 100,000-Plus Domestic Chip Cluster

On September 17, 2026, the Zhipu GLM team disclosed its first engineering progress on recursive self-improvement. An Infra Agent driven by GLM-5.3 built and debugged inference services from scratch on a cluster comprising more than 100,000 domestic chips, raising overall processing speed to three times the original level in under two weeks, with hardware utilization efficiency and per-token cost approaching Nvidia GPU levels.

The system has been put into real use. GLM-5.3-Flash was launched as the anonymous model Ox-Alpha, and its token call volume exceeded 620 trillion within six days. This is the first time a domestic large-model vendor has publicly applied AI self-improvement in a production environment.