← Back to feed

Z.ai, the Beijing lab formerly known as Zhipu AI, published a research post today describing how GLM-5.3 helped build the inference infrastructure now serving GLM-5.3-Flash. The post is titled “Toward Recursive Self-Improvement: How GLM Built Its Own Inference Infrastructure.” The lab frames what happened as an early, bounded form of RSI — not science fiction, but a documented production deployment.

The core claim: much of the engineering work to bring GLM-5.3-Flash to production was done by an “Infra Agent” powered by GLM-5.3 itself.

100,000 Chinese Chips

All production traffic for GLM-5.3-Flash now runs on a cluster of more than 100,000 Chinese-made AI accelerators. Z.ai says no one had previously deployed a cluster of Chinese-made chips at this scale. No NVIDIA hardware.

The challenge was significant. The chips had limited memory capacity and bandwidth compared to Nvidia’s datacenter parts. The model architecture was new. The context window stretched to 1 million tokens. Multimodal requests added further complexity. Kernel support from the chip ecosystem was incomplete, and documentation was sparse — much of it had to be inferred from first principles.

The Infra Agent analyzed the stack, generated hypotheses, and produced code changes. Human engineers defined objectives, set system boundaries, and reviewed critical changes. The division was explicit.

Technical Stack

The optimizations the agent implemented and validated include intra-node tensor parallelism for linear attention and the LM Head, a technique Z.ai calls ReplaySSM, W8A8 quantization, mixed-precision cache quantization using INT8/FP8/BF16, Layer Split, and an Encode-Prefill-Decode disaggregated serving architecture. Together these brought the cluster to production-grade throughput on hardware that had never run a model of this scale.

From model adaptation to production readiness took under two weeks, according to Z.ai’s account. End-to-end throughput improved roughly 3x from the starting baseline, with per-token cost reaching levels the company describes as comparable to mainstream Nvidia GPUs.

Ox-Alpha

Before the public GLM-5.3-Flash launch, Z.ai ran the model anonymously as Ox-Alpha on OpenCode and OpenRouter. Within a week it became the most-used model on both platforms, processing more than 62 trillion tokens in six days. The infrastructure built by the Infra Agent was already handling that traffic.

Z.ai has confirmed it was behind Ox-Alpha.

The RSI Frame

Z.ai is careful about how far it extends the RSI label. The model did not train itself. It did not modify its own weights. What it did was write, test, and validate the inference code running the system it now depends on to serve users.

But the lab is not shy about where it thinks this leads. From the post: “If this trend continues, given enough compute and enough time, its endpoint is a system that can design and train its own successor entirely autonomously. This is known as Recursive Self-Improvement, or RSI. We are not there yet, but early forms of it are already emerging.”

Z.ai also notes a prior inflection. In October 2025, the lab began using GLM to strengthen its cybersecurity capabilities — reasoning that coding and vulnerability detection overlap. Within less than a year, security partners were using GLM to discover thousands of vulnerabilities in real-world codebases. The company had to design a trusted access program to manage the risk.

The trajectory the lab describes: GLM-4.7 was used for internal coding out of obligation. GLM-5.3 is, by the team’s own account, an indispensable daily coding partner that is “moving steadily toward replacing us.”

What This Actually Is

Agentic coding systems already write and debug software in real codebases. What Z.ai describes is the same capability applied to the infrastructure the model runs on. The feedback loop is tighter — the system the agent improves directly determines the model’s own serving performance — but the underlying mechanism is not categorically new.

What is new: the scale of the Chinese-chip deployment and the explicit RSI framing from a frontier lab. Z.ai does not name the chip manufacturer. The 100,000-chip claim, if accurate, marks a meaningful milestone in China’s push to build AI infrastructure independent of Nvidia export controls.