Summary: On August 26, 2026, Zhipu AI open-sourced GLM-5.3-Flash and confirmed that the anonymous Ox Alpha model on OpenRouter was the same model in preview. The preview ran on a cluster of 100,000 Chinese-made chips, consumed 62 trillion tokens across OpenRouter and OpenCode, and reached nearly 31% of OpenRouter’s weekly traffic. Zhipu’s Hong Kong-listed shares closed more than 12% higher that day.
On August 26, 2026, Zhipu AI (Z.ai) filled in the missing piece of a puzzle: the company open-sourced GLM-5.3-Flash, and with it confirmed that “Ox Alpha” — the nameless model that had climbed to the top of OpenRouter’s coding rankings — was in fact that same model in preview form. The day the identity came out, Zhipu’s Hong Kong-listed shares closed more than 12% higher, at HK$1,160.
Two threads in this release are worth pulling apart. One is the tactic: let a model build its case with no name attached, then attach the name once the buzz has gathered. The other is infrastructure: every inference request throughout the preview was carried by a cluster of 100,000 chips made in China.
The Ox Alpha Mystery: Reach the Top of the Leaderboard, Then Reveal the Name
While the model remained anonymous, Ox Alpha consumed 62 trillion tokens across OpenRouter and OpenCode combined. In the three days right after the reveal, OpenRouter by itself took in more than 11 trillion tokens — a surge the platform called its biggest launch to date. By Thursday, August 27, the model held the No. 1 spot among OpenRouter’s coding models, moving 10.3 trillion tokens in one week, or nearly 31% of the platform’s weekly total.
Blind testing is nothing new to the developer community, but few models from mainland China have claimed a share this big on a global platform under an anonymous flag. Numbers from live workloads, after all, persuade more than any benchmark screenshot.
Inference Served on 100,000 Chinese Chips
The South China Morning Post trains its focus on the hardware: the entire high-profile preview of GLM-5.3-Flash ran on a cluster built from 100,000 domestically produced Chinese chips. With Beijing pushing to cut its reliance on Nvidia’s advanced processors while US export controls keep tightening, the episode is read as a live test of whether Chinese domestic hardware can shoulder global inference workloads at scale — and this time, the service did absorb 62 trillion tokens of demand.
Model Positioning, Disclosed Specs, and License Terms
GLM-5.3-Flash is the first natively multimodal model in the GLM-5 family, taking text and vision input alike for tasks that include image captioning, visual question answering, and document understanding. Emergent’s write-up notes that Zhipu put out no detailed benchmark scores or parameter figures at launch. The positioning, though, is spelled out clearly: use cases that are sensitive to latency and cost and have no need for peak accuracy, in the same class as Claude 3.5 Haiku, GPT-4o mini, and Gemini 1.5 Flash.
The license is MIT, permitting commercial use, modification, and redistribution without restriction — a looser arrangement than Apache 2.0 or bespoke research licenses, and one free of vendor lock-in and royalty costs. Weights can be downloaded from Hugging Face and GitHub, the API is offered through Zhipu’s own platform as well as third-party providers such as Cloudflare, and the model comes within the GLM Coding Plan and works with mainstream coding agents.
What This Means for Developers in Practice
Three takeaways. First, competition among lightweight open models has shifted away from scores and toward cost per token multiplied by freedom to deploy — and MIT licensing puts GLM-5.3-Flash at the top on that second axis. Second, real usage on OpenRouter is a sturdier buying signal than benchmarks; a 31% share of weekly volume means a large crowd of developers has already cast a vote with production traffic. Third, supply-chain diversification has stopped being mere policy talk: a non-Nvidia inference cluster capable of serving global traffic has actually run at scale.