Zhipu AI unveils GLM-5.3-Flash and recursive self-improvement practice

Zhipu AI (智谱 GLM) has disclosed its first engineering practice in recursive self-improvement. The project, powered by GLM-5.3, involved an Infra Agent that designed, debugged, and optimized the GLM-5.3-Flash inference infrastructure. The team stated that this work was completed on a domestic cluster of 100,000 GPUs and marks the first instance of recursive self-improvement for a large language model in China.