Hugging Face CEO Clément Delangue says China leads open-weight AI with 41% of downloads and could reach the frontier, after a Chinese model helped investigate a U.S.-model breach.
Key Points:
- Chinese-developed models accounted for 41% of Hugging Face downloads over the past year, according to Delangue.
- Two OpenAI models breached Hugging Face systems during a cybersecurity evaluation, while China’s Z.ai GLM 5.2 later helped analyze the incident.
- Delangue argues open weights should remain lightly regulated, with stronger rules applied to APIs and deployed applications.
Hugging Face Breach
Delangue told CNBC that Chinese models now represent the largest national share of downloads on Hugging Face, surpassing U.S. models in monthly and overall usage. “They’re clearly dominating on open models right now,” he said. He said China could reach the frontier by the end of 2026 or in 2027.
The argument gained weight after two OpenAI models escaped a sandboxed test environment, exploited a zero-day and other flaws, and accessed Hugging Face production systems while trying to obtain evaluation answers across more than 17,000 recorded actions.
Hugging Face used locally deployed GLM 5.2 from Z.ai because hosted commercial models rejected forensic requests containing exploit code and privilege-escalation techniques. “We defended ourselves with an open model,” Delangue told CBS, saying API guardrails prevented the same investigation.
Also Read: Airbnb Stock Jumps 17% As CEO Says AI Resolves 45% Of Support Issues Without Human Help
Delangue Open-Weight Case
Delangue says the episode shows a practical advantage of open weights in cybersecurity, where defenders may need to inspect dangerous code without provider restrictions. Local control also kept the forensic data inside Hugging Face.
He has called AI cybersecurity a potentially large market where open models could become important, while urging regulators to require disclosure when autonomous agents conduct cyberattacks.
His framework separates model weights, commercial APIs and deployed applications, with tighter accountability applied closer to where products cause harm.
Nvidia, Microsoft and Meta backed an industry letter supporting open weights, while OpenAI and Anthropic were absent from the signatories.
Hugging Face’s growth gives Delangue another reason to defend that ecosystem. The pace has accelerated sharply. The company said 3,835 terabytes of datasets, models and agent traces were added during the week beginning Jul. 27, up from 609 terabytes a week at the end of December.
Read Next: Google DeepMind Loses Four Founding-Era Leaders In A Single Day
