Zhipu says a GLM-5.3-powered agent built production inference infra in under two weeks

智东西 · wechat · 2026-09-17

Zhipu's Tang Jie unveiled progress on Recursive Self-Improvement: on a 100k+ domestic-chip cluster, an InfraAgent driven by GLM-5.3 helped build the production inference service for GLM-5.3-Flash in under two weeks, tripling end-to-end throughput. Key methodology: "dense feedback" — decomposing sparse end-to-end metrics into local, cheap, verifiable signals. Cases include fixing a TF32 precision bug in KDA's context-parallel path and a DeepEP GIL bottleneck blocking KVTransfer overlap (gap shrank from >20% to <1%). GLM-5.3-Flash launched anonymously as Ox-Alpha, topping OpenCode/OpenRouter usage with 62T tokens in 6 days. Backed by a $5B round, 60% earmarked for RSI.

Related event: Zhipu Discloses RSI Minimal Loop: GLM-5.3 Agent Built 100K-Card Domestic Inference Stack in Two Weeks(9 posts)→

Original post →

More from AGI Musings

AGI Musings channel →