Forget AGI, Here Comes RSI: Z.ai Says Its GLM Model Built Its Own Inference Infra

For years, the AI argument was about AGI. Z.ai is trying to drag attention to a newer acronym: RSI, short for recursive self-improvement. The company says its GLM-5.3 model didn’t just run on its infrastructure. It helped build that infrastructure too.

In a research post, the Chinese AI company says GLM-5.3 was central to creating a production inference service for GLM-5.3-Flash on a cluster of more than 100,000 AI accelerators made in China. Z.ai says all production inference for that model now runs on the system, and that nobody had previously deployed Chinese-made accelerators at this scale. The company also says it had to work through thin memory and bandwidth, a new model architecture, a one-million-token context window, multimodal requests, immature tooling and incomplete kernel support.

A lot of the heavy lifting was reportedly done by an “Infra Agent” powered by GLM-5.3. Z.ai says the model went from initial adaptation to production readiness in less than two weeks. It also claims end-to-end throughput roughly tripled versus the starting point, while hardware efficiency and per-token cost reached levels comparable to mainstream Nvidia GPUs. Before launch, the model was tested anonymously as Ox-Alpha on OpenCode and OpenRouter, where Z.ai says it became the most-used model on both platforms within a week and processed more than 62 trillion tokens in six days. Z.ai has since confirmed Ox-Alpha was its own model.

The company’s main technical pitch is what it calls “dense feedback.” Instead of giving an agent a fuzzy signal like throughput dropping by 20 percent, Z.ai fed it correctness tests, runtime logs, execution traces, microbenchmarks and end-to-end metrics. Engineers set the goals and guardrails, while the agent handled analysis, hypotheses and code changes. Z.ai gives three examples: a numerical error in the KDA kernel’s context parallelism path that worsened with long contexts and was merged into Flash Linear Attention; a concurrency bottleneck between DeepEP and Mooncake Transfer caused by Python’s GIL; and a decode-kernel optimization that it says produced a 1.71x speedup.

Z.ai is careful not to oversell its own story. It says it has not yet reached RSI, only early forms of it, and that humans should keep responsibility for objectives, boundaries and risk for a long time. Still, the company also says GLM-5.3 has become an indispensable daily coding partner for its team and is “moving steadily toward replacing us.” That’s the part worth watching: not the grand AI mythology, but the very practical moment when a model starts improving the machinery around itself.

The claims are hard to verify independently. The throughput, timeline and cluster-size numbers come from Z.ai, and the post doesn’t cleanly separate what the agent did from what the engineers did. The most concrete outside check is the public code contribution to Flash Linear Attention. Z.ai’s post also lands at a time when the company has been drawing attention for cybersecurity work, with GLM-5.2 described as rivaling Anthropic’s Mythos in that area and security partners reportedly using GLM to find thousands of vulnerabilities in real-world codebases.