China keeps open-weight models in play as security questions grow

China’s government has not moved to stop domestic companies from releasing open-weight AI models, despite risks that downloadable models could be used for cyberattacks or other harmful purposes. Since DeepSeek’s open-weight reasoning model in January 2025, releases from DeepSeek, Z.ai, Moonshot and Alibaba Cloud’s Qwen have strengthened China’s global AI footprint while challenging closed, higher-priced systems from US companies.

Xi Jinping’s recent support for “open source and openness, cooperation and sharing” does not amount to an unconditional endorsement of open releases. His remarks also emphasized risk awareness, security and controllability, reflecting a long-running policy doctrine that ties development to security, including the 2014 declaration and the 2017 New Generation AI Development Plan.

Possible triggers for tighter controls include cyber capabilities that threaten Chinese systems, misuse in biological or chemical development, or an incident involving advanced models. Beijing may judge that domestic surveillance, political controls, model safeguards and staged mitigation across training and deployment are enough for now. Companies may also choose closed access for business reasons, as Alibaba Cloud did with the most powerful version of its Qwen 3.6 series before later returning to open-weight “Max” releases with Qwen 3.8.