Z.ai, the Beijing-based maker of the GLM model family, says a coding agent built on its own GLM-5.3 model designed and deployed the production serving system for its newest release, GLM-5.3-Flash, in under two weeks. The company describes the work in a research post published September 17, and the throughput, chip count, and timeline below are all Z.ai’s own figures about its own system.
The task was not small. Z.ai says it stood up the new service on a cluster of more than 100,000 domestically made AI chips, a scale it says had not been deployed before, while also handling a fresh model architecture, a context window of one million tokens, and both text and image inputs. Kernel support for that hardware was still immature. Z.ai calls the internal tool that handled most of this work an “Infra Agent.”
The company says the result was roughly three times the throughput of its first working version, with hardware efficiency and cost per token it describes as comparable to mainstream Nvidia GPUs. It tested the system anonymously under the name “Ox-Alpha” on two developer platforms, OpenCode and OpenRouter. Within seven days it topped the usage charts on both, and Z.ai says it processed over 62 trillion tokens across that same six-day span. None of these figures come from an outside benchmark; they are the company describing its own launch.
Z.ai attributes the speed to what it calls “dense feedback”: instead of waiting for a full deployment to see whether a change helped, engineers gave the agent kernel-level correctness tests, microbenchmarks, and runtime traces it could check after each edit. Three examples stand out. The agent traced a numerical accuracy bug to a lower-precision math mode that lost accuracy over long contexts, and the fix was merged into the open-source Flash Linear Attention project. It also found that a Python threading lock inside a component called DeepEP was blocking a data transfer thread, pushing a test gap from an acceptable 5 percent to over 20 percent; fixing it cut the gap to under 1 percent. On a separate decode kernel, the agent removed redundant computation for what Z.ai says was a 1.71 times speedup.
Z.ai frames this as evidence its models are edging toward “recursive self-improvement,” the idea of a system that designs and trains its successor without human direction. The company is explicit that it has not reached that point: it says choosing objectives, setting the system’s boundaries, and judging risk remain jobs for its engineers, who reviewed the agent’s riskier proposals before they shipped.
That distinction matters more than the topline throughput number. An agent that finds bugs inside guardrails engineers already defined is a productivity story, not an autonomy story, and Z.ai’s own account keeps a human in the loop at every decision point it describes. The company has not published independent verification of the throughput or token figures, so the appropriate comparison for now is not to a benchmark but to Z.ai’s own prior baseline.
For teams evaluating whether to build inference infrastructure in-house versus buying it, the practical takeaway is narrower than “AI built an AI system”: it is that Z.ai found a way to compress a chip-bring-up cycle that has historically taken frontier labs months into roughly two weeks, on hardware that lacked mature tooling. Anyone weighing a bet on Chinese-made accelerators for large-scale inference now has one company’s own data point to test against their own workloads.
Based on a research post published by Z.ai on September 17, 2026.
