- Published on
Zhipu GLM Unveils First Recursive Self-Improvement Practice for Domestic Large Models: AI Builds Its Own Inference Sy...
- Authors
- Name
- DeepHorus
On September 17, the Zhipu GLM team disclosed the first engineering practice progress in the direction of Recursive Self-Improvement (RSI). The Infra Agent driven by GLM-5.3 completed the design, debugging, and optimization of the GLM-5.3-Flash inference infrastructure, achieving reverse improvement of the model on its own runtime system. In the report, IT Home author Yuanyang noted that this is the first time a domestic large-model vendor has publicly used "AI self-improvement" in a production environment.
The core of recursive self-improvement is to let AI improve its own runtime system with its own hands. In the past, in the infrastructure field, AI mostly undertook tasks such as code generation and assisted debugging, whereas this practice brings the model into the entire process of building, debugging, and optimizing inference services. IT Home noted that this means AI no longer merely writes code but can also participate in the engineering closed loop of the entire inference system.
According to the official Zhipu GLM blog, the Infra Agent completed the construction of production-grade inference services from scratch on a cluster composed of more than 100,000 domestically produced chips. In less than two weeks, its end-to-end throughput increased to 3 times the initial baseline. This process covered design, debugging, and optimization, demonstrating the model's ability to reversely improve its own runtime system.
In terms of hardware efficiency, officially disclosed data shows that the system's hardware utilization efficiency and per-token cost reached levels comparable to mainstream NVIDIA GPUs. For a cluster composed of more than 100,000 domestically produced chips, this result means that production-grade inference services have comparable economic indicators on domestic computing power. The related system has not remained at the experimental stage; instead, it has already been put into real use.
In terms of concrete deployment, GLM-5.3-Flash went online as the anonymous model Ox-Alpha on OpenCode and OpenRouter. Within 6 days of launch, its token call volume exceeded 62 trillion. Large-scale invocation data indicates that the inference system has already withstood real business traffic, rather than being used only for demonstration.
From the perspective of industry impact, this is the first time a domestic large-model vendor has publicly used "AI self-improvement" in a production environment. It moves AI from assisted coding further toward the engineering closed loop of inference systems. For domestic large-model infrastructure, this move provides a production-environment sample of AI participating in the optimization of its own runtime system.
Looking ahead, the RSI engineering practice disclosed this time by the Zhipu GLM team has advanced recursive self-improvement from conceptual discussion to production verification. As GLM-5.3-Flash continues to go online as an anonymous model and generates large-scale token calls, its system performance and optimization effects will attract attention. The implementation progress of domestic large-model vendors in the direction of AI self-improvement has thereby entered an observable stage.
