Glm team demonstrates first recursive self-improvement in domestic large model

On September 17, 2026, the GLM team from Zhipu AI announced the first engineering implementation of recursive self-improvement (RSI), allowing AI to autonomously enhance its own operational systems. The Infra Agent, driven by GLM-5.3, designed, debugged, and optimized the GLM-5.3-Flash inference infrastructure, achieving reverse improvement of the model's runtime environment. The system was built from scratch on a cluster of over 100,000 domestic chips, achieving a threefold increase in end-to-end throughput within two weeks, with hardware efficiency and single-token cost comparable to mainstream NVIDIA GPUs. The system is now in active use, with the GLM-5.3-Flash model, anonymized as Ox-Alpha, available on OpenCode and OpenRouter, accumulating over 62 trillion tokens in six days. This marks the first time a domestic large model provider has publicly deployed AI self-improvement in production. Zhipu AI's achievement signifies a shift where AI is no longer limited to writing code but can actively participate in the engineering lifecycle of inference systems. The company's success highlights the growing capabilities of domestic AI infrastructure and its potential to compete with international standards.