China’s 745B parameter AI model leaked before launch
Zhipu AI has officially unveiled GLM-5, a massive 745-billion parameter frontier-class model developed using Huawei hardware. The model’s existence was discovered by tech investigators days before its debut when it appeared under a pseudonym on API aggregators. This release signals a major strategic push for Chinese sovereign AI capabilities in the open-source landscape.
Points clés
- Zhipu AI recently completed an IPO raising $558 million to fuel its AI development
- The model GLM-5 boasts a staggering 745 billion parameters, placing it in the frontier class
- Five days before launch, the model appeared on the OpenRouter aggregator under the pseudonym “Pony Alpha”
- Tech analysts identified the leak on February 6, 2026, after noticing suspiciously high performance levels
- A vLLM pull request on GitHub confirmed the model’s architecture via specific new code classes
- The training of GLM-5 was conducted specifically on Huawei chips, reducing reliance on Western hardware
- Zhipu AI’s model uses a Mixture of Experts (MoE) architecture according to leaked code references
- The model was officially released to the public shortly after the community-driven “detective work”
À retenir
If you’re looking for a lesson in corporate secrecy, Zhipu AI isn’t the teacher you want. Testing your top-secret project on a public API under a name like “Pony Alpha” is about as subtle as wearing a neon sign in a dark alley. But hey, when you’ve got 745 billion parameters and Huawei chips backing you up, I suppose you can afford to leave the back door unlocked for a few days. My advice? If you see a mysterious flying horse on your dev platform, it’s probably just a massive Chinese AI looking for a stable.
Sources
