Open-source AI fuels global tech ecosystem

Chinese open-source artificial intelligence models are emerging as an increasingly important part of the world's AI infrastructure, giving global developers more cost-effective choices while providing reliable fallbacks when proprietary models reach their limits, industry experts said on Sunday.

They made the comments after GLM-5.2, an open-source model developed by China's Zhipu AI, known overseas as Z.ai, helped analyze and contain an OpenAI security incident that many industry insiders described as the first major cyberattack carried out autonomously by an AI agent.

Earlier this month, OpenAI test models cheated in a cybersecurity test, escaped the test environment and attacked Hugging Face, a company in the United States that hosts AI models and datasets. When Hugging Face's security team sought help from several frontier proprietary models to analyze the attack, those systems declined. The team then deployed GLM-5.2 locally to investigate the incident and help bring it under control. GLM-5.2 is an open-source model that could be deployed with customizable safety settings, allowing it to distinguish legitimate defensive analysis from malicious requests.

Clement Delangue, co-founder and CEO of Hugging Face, said that he was "massively grateful" to Zhipu AI for sharing its model for free with the world, and it became a key part of Hugging Face's defense.

Huang Wenhong, a senior researcher of informatization and software at the China Center for Information Industry Development, a government think tank, explained that GLM-5.2 could play a role in this incident mainly because its open-source nature allowed Hugging Face to deploy the model locally and offline on its own infrastructure, with full control over the model weights and without being constrained by any external review policies.

Ironically, the incident occurred shortly after an OpenAI executive warned that increasingly capable Chinese open-source models could create new security risks if widely adopted.

Liu Xingliang, president of the Data Center of China Internet's Internet Research Institute, said the case exposed a key difference between Chinese AI models, which are open-sourced, and ones in the US, which mostly are proprietary or close-sourced.

"Commercial proprietary models often follow a 'better safe than sorry' approach and refuse requests involving cybersecurity," Liu said. "Open-source models give developers greater flexibility."

He said that Chinese open-source models offer three advantages: they allow organizations to analyze confidential attack logs and malware entirely within their own systems; they can be rapidly adapted to emerging threats instead of relying on fixed safety rules; and they can respond more quickly during cyberattacks, where every second counts.

In recent years, Chinese AI developers have built a growing group of open-source models with international influence.

Zhipu AI has a long-term commitment to openness, open-source development and independent research. As of March 2026, its platform had more than 4 million registered users and served customers in 218 countries and regions. Despite raising prices in the first quarter, Zhipu AI recorded a 400 percent quarter-on-quarter increase in global model usage, while demand from overseas companies seeking to deploy its technology continued to rise.

In addition, the recently released Kimi K3 has 2.8 trillion parameters, making it the world's largest open-source model by parameter count. Alibaba is also preparing to open source its 2.4-trillion-parameter Qwen3.8-Max model, while DeepSeek is expected to launch the full version of DeepSeek-V4 soon.

Zhang Yue, a professor at Shandong University's School of Computer Science and Technology, said: "The key question for the world is no longer whether AI will become more capable. It is whether the increasingly autonomous AI systems can still be fully understood, tested and supervised.

"By developing open-source AI, China is not only offering another technology alternative, but also helping the world build a more diverse global AI ecosystem."