China drafts export controls on frontier AI weights and agents

On 16 October China's commerce and science ministries issued Announcement No. 48: a draft adding 23 AI items to the list of technologies restricted or banned from export, open for comment until 15 November, plus an export registration rule that applies already. According to a Cloud Security Alliance analysis, model weights and know-how are restricted if a model passes 10^26 training FLOPs, 2 trillion parameters, or a global top-10 position on rankings or usage at the time of export, and agents are covered separately, such as tool-using agents with 200 or more integrated tools and at least 90% accuracy. Using a Chinese model through an API, app or SaaS is not treated as an export, but deploying a restricted model locally outside China is, while technology already public as open source or in papers is outside the scope. On Hacker News (8 top-level comments when we read it) one commenter said the rules match earlier guidance, others debated whether Chinese models will stay open, and one feared a US-China deal limiting advanced AI to a few countries; no consensus. Sorted take: models already released as open weights are outside the rules; future releases are the uncertain part.