Ox Alpha Was Z.AI’s GLM-5.3 All Along, and the Weights Are Now Public
TRENDING STORY
By the NewsCenter Desk
Published Sunday, Aug. 23, 2026 · Updated Wednesday, Aug. 26, 1:05 p.m. ET · developing
UPDATE · Wednesday, Aug. 26, 1:05 p.m. ET
Z.AI, the Chinese lab also known as Zhipu, confirmed that Ox Alpha is a new iteration of its GLM series and released the weights the same day, telling Bloomberg News so in response to its queries. The model is GLM-5.3-Flash, published on Hugging Face under an MIT license: 320 billion parameters, 18 billion active, and the first natively multimodal model in the GLM-5 series, per Z.AI’s own model card. Bloomberg reports the company saying it runs on Chinese AI chips. The counter stood at 27.2 trillion tokens Wednesday afternoon.
Ox Alpha, the anonymous model eating the AI charts since Thursday, crossed 20 trillion tokens by Tuesday morning. Nobody will say who owns it.
Five days old, free, listed under a provider called “stealth,” it has spent the week among the busiest models on OpenRouter. Bloomberg reported Sunday that developers keep piling in “even as the identity of its creator remains a mystery.”
One researcher says he settled that.
THE LATEST
‣ Z.AI claimed it, and Bloomberg got the confirmation. Bloomberg News reported Wednesday that the lab confirmed Ox Alpha as a new GLM iteration in response to its questions. Three days earlier this story carried a researcher’s server-side evidence pointing the same direction. Aug. 26
‣ The weights are out, two days before anyone expected them. GLM-5.3-Flash went public on Hugging Face Wednesday under an MIT license, with a BF16 build following hours later. The New York Times had reported the release was coming Friday. Z.AI’s model card claims it beats GLM-5.2 at a tenth the price and approaches Claude Opus 4.8 on coding and agentic benchmarks. Aug. 26
‣ The counter crossed 20 trillion Tuesday morning. OpenRouter’s public counts put Monday at 5.9 trillion tokens, a fifth straight daily record, and Tuesday running ahead of that pace before noon. The listing is still free. Aug. 25
‣ OpenRouter’s paperwork still conflicts. The New Stack reported Monday that OpenRouter “currently publishes conflicting disclosures”: the tension between the listing and the stealth-program terms this story detailed Sunday. The terms page hasn’t changed, and the company hasn’t addressed it. Aug. 24
‣ The viral benchmark didn’t hold up. A widely shared 80% coding score came from 10 tasks; a reproducible community run across 113 tasks landed at 58.4%, per the same New Stack report. Aug. 24
‣ Zhipu stays silent as the week begins. The GLM-5.3 weights it is reportedly expected to publish this week would land on Hugging Face, which, as of Sunday, is reportedly exploring a sale at $13 billion or more. Aug. 24
Everyone else fingerprinted the model. He interrogated the server. In a test published Saturday, a researcher posting as Chetaslua sent one malformed request and got back a Java stack trace naming an internal class on Zhipu AI’s own documented API route. Same error dialect as three confirmed Z.ai models. Thirty of thirty tokenizer probes matching GLM-5.3. He published the command so anyone can rerun it.
That pins down who operates the endpoint, not who trained the weights. Zhipu has said nothing, and a competing analysis reads the tokenizer as Microsoft’s. It is still the only claim here with a receipt attached.
The traffic is its own story. OpenRouter’s public counts show 4.5 trillion tokens Saturday, roughly eight times any previous anonymous debut, and Sunday ran ahead of that pace. At the list price of the Chinese lab it is most often linked to, the giveaway would have cost about $16 million. At top-shelf American rates, more than $110 million.
Somebody is paying for it. The listing says prompts and completions “are retained by the provider and are not used for training.” Retained by a provider no user can identify.
OpenRouter’s stealth-program terms describe the bargain differently. The program exists “for purposes of collecting User Content for use in Stealth Model training and improvement,” free access is the stated consideration for that content, and a section headed “License to User Content for Model Training” grants an irrevocable, perpetual license passed to the anonymous provider.
The same terms promise that OpenRouter “will not disclose the name or origin of Stealth Providers to you.” There is no one to ask.
Who is sending it makes this stranger. Developers can aim their own tools anywhere on the marketplace, and they have been aiming the biggest ones here: Nous Research’s Hermes Agent leads, then Anthropic’s Claude Code at 965 billion tokens. Neither company has any part in the arrangement.
Washington, meanwhile, is weighing a bill to keep Chinese models out of federal agencies. A model with no nationality can’t be checked against any list.
DEEPER COVERAGE
‣ TechCrunch canvasses the guessing game, from Andrew Curran to Reddit. Read · Aug. 23
‣ Wccftech on the 100-trillion-tokens-a-day offer, first pointing to Zhipu, later updated toward Microsoft. Read · Aug. 21
‣ Our story on Stripe buying the tollbooth all of this runs through. Read · Aug. 22
THE NEWSCENTER.IO TAKE · Aug. 23, 2026
Two questions matter here, and neither has an answer. The first is the simplest: what company is this? A researcher has already made the server name itself, and the marketplace forwarding the traffic has a contract with whoever is on the other end. Somebody knows. Nobody is saying.
The second one reaches further. Ten trillion tokens of prompts and working code have gone to a provider no user can identify, and OpenRouter’s own terms say that content is the price of admission. Free was never the deal. Your data was.
Top illustration: NewsCenter.io.
Disclosure: NewsCenter’s newsroom systems run on Claude, made by Anthropic.
