Zhipu AI Discloses Production-Level Recursive Self-Improvement for Large Models
Zhipu AI revealed a major engineering milestone where its GLM-5.3-powered Infra Agent autonomously designed, debugged, and optimized the production inference system for GLM-5.3-Flash from scratch. Operating on a cluster of over 100,000 domestic Chinese chips, the agent tripled end-to-end throughput in under two weeks. This marks the first time a Chinese AI vendor has publicly demonstrated recursive self-improvement in a live production environment, showing AI can close the engineering loop on its own infrastructure. Furthermore, achieving hardware efficiency and per-token costs on domestic hardware comparable to mainstream NVIDIA GPUs represents a significant step toward AI hardware independence. The self-optimized system was validated in real-world workloads, with GLM-5.3-Flash anonymously deployed as "Ox-Alpha" on OpenCode and OpenRouter, serving over 62 trillion tokens in just six days. The agent optimized the entire pipeline, achieving competitive efficiency without human intervention in code optimization.
## BACKGROUND
Recursive self-improvement (RSI) refers to an AI process where a system autonomously inspects, rewrites, or enhances its own code and execution stack to boost performance. Meanwhile, LLM inference infrastructure encompasses the hardware, software framework, and execution routing required to serve predictions quickly and cost-effectively at scale.