On 20 August a new AI model turned up on OpenRouter, the marketplace where developers shop for models the way the rest of us pick a streaming plan. It was called Ox Alpha. In the box where a maker’s name should sit was a single word: Stealth. No company, no model card, no paper, nothing. It was free, it read text, images and video, and it carried a context window of 1,048,576 tokens. Within days, people were feeding it their code.
By the 23rd it had climbed to the top of OpenRouter’s leaderboard, more than doubling the traffic of DeepSeek according to Bloomberg, and processing more than 11 trillion tokens in its first three days. Thousands of people were pumping prompts, private repositories and half-built products into a machine, and not one of them could tell you whose servers were on the other end.
A model with no return address
The pitch was generous to the point of suspicious. Ox Alpha cost nothing, boasted capacity for a hundred trillion tokens of inference a day, and its anonymous operator said it would not train on what users typed. Coding agents such as OpenCode wired it in with near-unlimited use for a trial week. The catch was the obvious one. Free access to a frontier-grade model, run by a party that will not give its name, is an odd thing to hand your source code to. Plenty of people did it anyway.
Cracked open with a broken request
The detective work was gloriously low-tech. An independent researcher going by Chetaslua sent the model a deliberately malformed request, setting a numeric field to the text “abc,” and the server fell over. The error it coughed up was a Java stack trace naming an internal class, com.wd.paas.api, which maps straight onto the API routes of the Chinese lab Zhipu.
Others piled in. Tokenizer probes matched Zhipu’s GLM family on dozens of tests, across writing systems, emoji, code and SQL. The way Ox Alpha counted video frames lined up with Zhipu’s own multimodal model. Rival theories collapsed under the same tests. Xiaomi’s MiMo happily accepts audio, and Ox Alpha refused it flat. Alibaba’s Qwen produced a different fingerprint entirely. A stranger’s model, and the internet reverse-engineered its birth certificate in about 48 hours.
Surprise, it was China all along
On 26 August, Bloomberg got the confirmation. Z.ai, the international arm of Zhipu, admitted the model was a new entry in its GLM family and released the weights that night. They landed as GLM-5.3-Flash, an MIT-licensed model with a 320-billion-parameter design running 18 billion active. Z.ai’s shares rose sharply on the reveal. The company said the whole stealth trial ran on a cluster of domestically produced Chinese chips, which is its own pointed message to Washington.
Why hide a model this good behind a fake name? Because a nameless listing on a neutral platform buys something a branded launch cannot: billions of tokens of real-world abuse before anyone can pin the result to your reputation. There is form here. Two earlier stealth previews, Hunter Alpha and Healer Alpha, both turned out to be Xiaomi’s MiMo once the curtain dropped.
So the shape of the week was this. The AI industry fell hard for a model with no name, funnelled its secrets into it, argued for days about who built it, and was out-sleuthed by a hobbyist typing “abc” where a number was supposed to go. The machine was clever. The people were the better story.
Did you know: during its anonymous run, Ox Alpha processed 62 trillion tokens across OpenRouter and the OpenCode agent platform, which is a colossal amount of real work to hand a company that would not tell you its name.
Sources
- SCMP: Zhipu AI shares jump as Ox Alpha revealed as GLM-5.3-Flash
- TechCrunch: Z.ai is the lab behind the mysterious Ox Alpha model
- Hugging Face: GLM-5.3-Flash weights
Related on Top Tool Stack: Claude Now Remembers Your Projects Across Chat and Cowork · Google Points Gemini Enterprise at Big Law’s Billable Hour