Ox Alpha arrived as a free, anonymous AI model and quickly became the kind of Silicon Valley puzzle that benchmarks alone could not solve. Developers saw impressive performance; the industry saw a possible new threat to costly frontier systems.

The model appeared on OpenRouter on Thursday as a “stealth model” from an unnamed third-party provider, pitched for coding, sustained agentic work and production workloads. OpenCode offered it free for a week with “near unlimited usage,” while saying the provider could handle 100 trillion tokens a day.¹

Praise fed the frenzy. Stripe chief executive Patrick Collison called Ox Alpha “very impressive,”² but its provenance remained stubbornly unclear. One camp focused on China: observers pointed to similarities with Z.ai’s GLM models and the company’s earlier anonymous test, Pony Alpha. Chinese social-media users also read the name as a clue, connecting “Ox” to the viral film meme Here Comes the Ox.³

That case was never conclusive. Wccftech later highlighted analysis suggesting the tokenizer might instead trace back to Microsoft’s MAI family, while Reddit users argued over whether a model willing to discuss politically sensitive subjects could really be Chinese-made.³ AI analyst Andrew Curran captured the shifting mood: GLM was the leading theory Friday night, but by Saturday “people seem less sure of anything.”⁴

The mystery ended Wednesday. Z.ai, also known as Zhipu, said Ox Alpha had been an anonymous preview of GLM-5.3-Flash, tested through OpenRouter and OpenCode to collect user feedback. It then released the weights on Hugging Face, allowing developers to download, modify and build on the model.³

Z.ai describes the system as multimodal, with a context window of up to 1 million tokens. The reveal vindicated the China theory, not the Microsoft alternative—and transformed a stealth experiment into a public open-model challenge for OpenAI, Anthropic and other premium AI vendors.⁵