Zhipu GLM reveals first RSI practice for domestic large models
智谱 GLM 公开国内大模型首个递归自我改进实践：AI 在十万卡国产集群上自建推理系统
This is the first time a domestic Chinese large model company has publicly used AI self-improvement for production environments, showing AI's ability to participate in the entire engineering loop of a reasoning system.
Zhipu GLM team disclosed an engineering practice for Recursive Self-Improvement (RSI) that lets AI improve its own running system. Specifically, the Infra Agent driven by GLM-5.3 completed the design, debugging, and optimization of the GLM-5.3-Flash inference infrastructure, achieving a 3x increase in end-to-end throughput on a 100k-card domestic chip cluster in less than two weeks. The system is now in real use, with the GLM-5.3-Flash model Ox-Alpha deployed on OpenCode and OpenRouter, handling over 6.2 trillion tokens in 6 days.
智谱 GLM 公开国内大模型首个递归自我改进实践：AI 在十万卡国产集群上自建推理系统
IT之家 9 月 17 日消息，今日，智谱 GLM 团队披露递归自我改进（Recursive Self‑Improvement，简称 RSI）方向首个工程化实践进展，也就是让 AI 自己动手改进自己的运行系统。 具体来说，由 GLM-5.3 驱动的 Infra Agent 完成了 GLM-5.3-Flash 推理基础设施的设计、调试与优化，实现模型对自身运行系统的反向改进。 据官方博客，Infra Agent 在超 10 万张国产芯片组成的集群上从零完成生产级推理服务搭建，不到两周将端到端吞吐提升至初始基线的 3 倍，硬件利用效率与单 Token 成本达到主流 NVIDIA GPU 相当水平。这套系统已投入真实使用，GLM-5.3-Flash 以匿名模型 Ox-Alpha 在 OpenCode、OpenRouter 上线，6 天 Token 调用量超过 62 万亿。 IT之家注意到，这是国内大模型厂商首次公开将“AI 自我改进”用于生产环境，意味着 AI 不再只是写代码，还能参与整个推理系统的工程闭环。