China tightens export controls on open-weight models and 23 key AI technologies

China's commerce and science ministries have released a draft update to the national list of banned and restricted export technologies, adding 23 key AI technologies spanning large models, AI agents, compute clusters and data processing. Until the final rules take effect, companies must register affected exports through the commerce ministry's portal.
The catalog applies the harshest treatment to AI tools that can autonomously write or alter malware to attack critical infrastructure, the military or government agencies: those face a total export ban. Tools that hunt for software vulnerabilities are not banned outright but will face new export limits. Public rankings determine the top-10 trigger: export officials look at results from major domestic and foreign evaluators or industry platforms, covering overall performance, specific tasks, safety, or worldwide usage numbers.
Large models fall under control in two ways. The first covers models trained on China's "important data" under the Data Security Law and related rules, where services are offered on the basis of that data; lawfully public data is exempt. A carve-out permits tiny amounts of up to 0.1 percent, provided the data was cleaned beforehand, never surfaces in model outputs, and does not substantially boost the model's R&D capacity for core technology in that field.
The second path covers the core technology behind large models: weights, checkpoints, source code, architecture, and training or inference methods. Controls attach once a model reaches 10^26 FLOPs of training compute, has at least 2 trillion parameters, or ranks in the global top 10. For scale, 10^26 FLOPs is roughly the training compute of frontier systems such as GPT-4-class models; 2 trillion parameters is several times the size of most openly released models, so the thresholds target the very top of the field rather than everyday models.
To avoid strangling its domestic open-source scene, the plan introduces a parallel fast track. Publishing restricted AI technology and weights as open source qualifies for automatic licensing. The same applies to Chinese firms running their own data centers abroad, taping out chips overseas, conducting research with foreign partners, or helping set international standards. Under the ministry's rules, eligible exports clear the moment a company files the paperwork. Any AI technology already public through open-source releases or research papers is exempt from the new rules entirely.
Usage rules are split by deployment model. Reaching restricted models through API calls, apps, websites, SaaS tools, or lightweight versions on local devices stays outside these export controls. Installing restricted AI technology on local servers abroad, however, requires a standard export license. Companies can still use the streamlined process when the software runs on their own overseas cloud.
Beijing also pledges to support AI development in the Global South and ease exports for capacity building, technology sharing, and joint rulemaking within the World Artificial Intelligence Cooperation Organization, the bloc China launched to counter Western AI bodies.
The practical effect, as the source frames it, is likely to push US companies that depend on open-weight AI to double down on domestic models rather than Chinese alternatives. Europe sits in a tighter spot: the EU has banked on running open-weight models in domestic clouds to preserve independence, alongside firms that want to host closed top-tier models on local servers. That sovereignty strategy weakens when the strongest open models come from the US or China and can be severed by export controls at short notice. This is a reasonable reading, though the draft leaves real room for workarounds: API access and SaaS use remain outside the controls, so European products can still call Chinese models remotely even if they cannot host the weights.
For developers, the headline mechanic is the open-weight fast track. Automatic licensing on registration means Chinese labs can keep shipping weights publicly, which preserves the supply of downloadable models while the same weights are restricted for export by other means. The exemption for anything already public suggests the rules bite hardest on future frontier checkpoints, private weights transfers, and on-premise installs abroad. A likely trade-off is that foreign teams wanting on-premises Chinese models will need standard licenses, adding legal review and delay, while competitors building on US open weights face no equivalent friction. Whether the top-10 ranking clause creates uncertainty depends on how rankings are read in practice; a model's position can shift with a new evaluation, so the trigger is not fully in a lab's control.
Why it matters: model and platform teams that source weights from Chinese labs should plan for licensing checks on any on-premises deployment outside China, while API-based integration appears less affected. The draft is not final, so the concrete thresholds and the registration portal will determine how much actually changes. European sovereignty plans that assumed unrestricted Chinese open weights now look less certain, and US open-weight providers are the likely near-term beneficiaries in inference.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
China plans to restrict exports of 23 key AI technologies, including model weights, source code, and training methods for systems hitting 10^26 FLOPs or 2 trillion parameters. A fast-track rule will still clear open-source releases as soon as they are registered, but the move will likely push Western companies to double down on US open-weight models over Chinese alternatives. The article China tightens export controls on open-weight models and 23 key AI technologies appeared first on The Decoder .