Zhipu Discloses RSI Loop: GLM-Driven InfraAgent Ships Inference on 100K Domestic AI Accelerators in Two Weeks

Zhipu co-founder and chief scientist Tang Jie posted a long thread on X linking to the technical blog "Toward Recursive Self-Improvement: How GLM Built Its Own Inference Infrastructure," disclosing a minimal recursive self-improvement (RSI) loop: an InfraAgent powered by GLM-5.3 helped build production inference serving for GLM-5.3-Flash on a cluster of 100,000+ domestic AI accelerators, going from first successful run to full production traffic in under two weeks with 3.2x end-to-end throughput over baseline. Zhipu also launched GLM-5.3-FlashX at up to 200 tokens/s—5x faster and 2.5x the price of Flash, with API access open. This is a public demonstration of a model optimizing its own inference infrastructure and a landmark case of a large-scale domestic chip cluster carrying full production traffic.

Confirmed

• Zhipu's official blog states all production inference traffic for GLM-5.3-Flash now runs on 100,000+ domestic AI accelerators, taking under two weeks from first run to production
• End-to-end throughput improved 3.2x over baseline; per-token cost and other metrics were disclosed in Tang's blog (per @APPSO)
• Much of the optimization was done by a GLM-5.3-driven Infra Agent, forming a minimal RSI loop
• Per @赛博禅心, Zhipu launched GLM-5.3-FlashX: up to 200 tokens/s, 5x faster and 2.5x pricier than GLM-5.3-Flash, API open, built on 100K domestic chips
• Per @量子位, the system uses dense feedback as its technical core, seen as a prototype of RSI; GLM-5.3-Flash previously ran anonymously as Ox-Alpha on OpenRouter
• Per Jen Zhu (via @ccerrato147), Zhipu also published its training methodology, arguing the scarce skill is no longer pretraining but scaled agentic RL and serving on constrained hardware
• Per @Elux91 and @petrusenkomax, Zhipu emphasized self-built inference infrastructure over external clouds; Reddit users read it as "shots fired at Dario," and Hacker News debated cost and control of domestic-built serving stacks

Unconfirmed

• The "jab at Dario" reading is community interpretation; Zhipu's posts did not name Anthropic or Dario Amodei
• The specific chip model, supplier, and cluster composition of the "100,000+ domestic AI accelerators" were not disclosed
• The relationship between FlashX and the InfraAgent-built system (whether FlashX runs on it) was not clarified

Why it matters

• First publicly documented domestic case of a model optimizing its own inference infrastructure at production scale, validating RSI beyond papers
• A 100K-card domestic accelerator cluster carrying full production traffic is a real-world test of China's compute supply chain and inference optimization capability, sharpening the self-built-vs-cloud inference debate
• Engineer jietang's account shows a human-agent division of labor shipping full production in two weeks, a reference model for the industry
• FlashX commercializing the gains at 5x speed and 2.5x pricing shows infra optimization directly converting into sellable product differentiation

2026-09-17 ~ 2026-09-18 · 16 related posts

• [source] GLM agent built its own inference infra in two weeks, tripling end-to-end throughput — jietang · 2026-09-17
• Zhipu says a GLM-5.3-powered agent built production inference infra in under two weeks — 智东西 · 2026-09-17
• Inside Zhipu's dense-feedback InfraAgent: an early blueprint for recursive self-improvement — 量子位 · 2026-09-17
• Zhipu says a GLM-5.3 agent optimized its own inference stack, 3.2x throughput on 100k domestic chips — SinclairWang1 · 2026-09-17
• Zhipu's InfraAgent boosted inference throughput 3.2x on 100k domestic chips in two weeks — APPSO · 2026-09-17
• GLM-5.3-powered Infra Agent boosts own inference stack 3x on 100k Chinese accelerators — Dr_Singularity · 2026-09-17
• Zhipu Claims GLM-Powered Infra Agent Delivered 3x Speedup on 100k+ Domestic Accelerators — Dr_Singularity · 2026-09-17
• [source] Zhipu Details How GLM Built Its Own Inference Infrastructure Toward Recursive Self-Improvement — likeastar20 · 2026-09-17
• GLM-5.3-powered Infra Agent cut domestic-chip serving rollout to two weeks with 3.2x throughput — Afinetheorem · 2026-09-17
• Zhipu used GLM-5.3 as an infra agent to ship GLM-5.3-Flash serving in 2 weeks, 3x throughput — ccerrato147 · 2026-09-17
• HN: GLM built its own inference infrastructure — petrusenko_max · 2026-09-17
• GLM's blog on building its own inference infra read as a shot at Dario — Elux91 · 2026-09-18
• GLM-5.3 optimized its own inference stack, tripling throughput in under two weeks — rohanpaul_ai · 2026-09-18
• [source] Zhipu launches GLM-5.3-FlashX at 200 tokens/s, with its inference infra built by LLM agents — 赛博禅心 · 2026-09-18
• Zhipu's GLM-5.3-Flash inference stack built by an agent on 100,000 Chinese accelerators — emmanuelvivier · 2026-09-18

1 near-duplicate retellings: victormustar