China's commerce and science ministries publish draft export-control list covering 23 AI technologies, including large models above 10^26 FLOPs, 2 trillion parameters or global top-10 rank

China's commerce and science ministries released a draft update to their catalogue of banned and restricted export technologies. It covers 23 AI areas. Autonomous malware-writing tools aimed at critical infrastructure, the military or government would be banned from export outright, and vulnerability-finding tools face new limits. Controls on large models apply to those trained on China's "important data", and to core technology (weights, code, architecture, methods) once a model reaches 10^26 training FLOPs, 2 trillion parameters or a global top-10 ranking. Open-source releases get automatic licensing on registration, API, app and SaaS access are not controlled, and already-public technology is exempt. It is a draft, and affected exports must be registered on the commerce ministry portal until the rules are final.

If adopted, the rules would give Beijing a licensing lever over the export of weights, code and methods for its largest or top-ranked models, and a ban on autonomous malware tools aimed at critical infrastructure, the military or government. Chinese labs that publish open weights would face a registration step. Foreign firms that depend on those models would face more uncertainty, though API and SaaS access is explicitly not controlled. Impact is limited for now because the draft is not in force and the open-source fast track may keep most releases flowing. Anyone running Chinese models on servers abroad should check the licensing wording once the final text appears.

The headline says China "tightens" controls, but this is an unfinished draft and open-source releases get automatic licensing, so the practical effect on open-weight models is unproven. The Decoder's claims about pushing US firms to domestic models and undermining Europe's sovereignty strategy are its own speculation, with no evidence yet. The article also links no primary document, so every threshold is second-hand until the ministry text is found.

1. 01Publication of the draft text on the Ministry of Commerce or Ministry of Science and Technology sites, and whether its thresholds (10^26 FLOPs, 2 trillion parameters, global top-10 rank) and the 0.1% cleaned-data allowance match The Decoder's account.
2. 02The length of the public comment period and the date the final catalogue is issued, plus any changes to the open-source automatic licensing or the API and SaaS exemption between draft and final.
3. 03Whether Chinese labs that release open weights (for example Qwen, DeepSeek or Kimi) register exports on the commerce ministry portal, delay a release, or withhold weights after the rules take effect.