AivexaNewsSearch
AI news for builders and product teamsChecked every hour

China plans to tighten export controls on open-weight models and 23 key AI technologies

Collected Oct 16, 2026

China's commerce and science ministries have published a draft update to the national catalog of banned and restricted export technologies, aiming it at 23 key AI areas spanning large models, agents, compute clusters and data processing. The draft introduces two distinct pressure points: a full export ban on AI tools that can autonomously write or alter malware to attack critical infrastructure, the military or government agencies, and new export limits — not outright bans — on certain tools for finding software vulnerabilities.

For large models, the draft defines two trigger categories. The first covers models trained on China's "important data" as defined under the Data Security Law and related rules, where the model offers services based on that data. Lawfully public data is exempt. A narrow carve-out permits tiny amounts of up to 0.1 percent, provided the data was cleaned beforehand, never appears in model outputs, and does not substantially boost the model's R&D capacity for core technology in that field.

The second category targets the core technologies behind large models: weights, checkpoints, source code, architecture, and training or inference methods. These controls activate once a system reaches 10^26 FLOPs of training compute, holds at least 2 trillion parameters, or ranks in the global top 10. Export officials judge top-10 status using public rankings from major domestic and foreign evaluators or industry platforms, weighing overall performance, specific tasks, safety, or global usage figures.

For developers, a FLOP is a single floating-point arithmetic operation, and training compute is typically reported as a cumulative total across a training run — so 10^26 FLOPs describes the scale of the run rather than the size of the finished model. Parameter count is a separate axis: a rough proxy for model capacity measured in learned weights. This means a model can trip the threshold through training intensity, sheer size, or external leaderboard placement, and the top-10 criterion is notably elastic because it depends on third-party rankings that shift over time.

The draft also builds a fast-track lane intended to keep China's open-source ecosystem moving. Publishing restricted AI technology and weights as open source qualifies for "automatic licensing," as do 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 are cleared the moment a company files the paperwork. Separately, any AI technology already public through open-source releases or research papers is completely exempt.

The controls are narrower than a blanket ban on foreign use. Accessing restricted models through API calls, apps, websites, SaaS tools, or lightweight on-device versions stays free of these export controls. Installing restricted AI technology on local servers abroad, however, requires a standard export license, though companies can use the streamlined process when the software runs on their own overseas cloud.

Dean W. Ball, head of strategic futures at OpenAI, characterized the framework as establishing "the optionality to shut down future open-weight models at will" rather than a decision to shut them down, and described Beijing's approach as a mirror image of US model-weight export controls already in place. Until final rules take effect, companies must register affected exports through the commerce ministry's portal.

The draft also pledges support for AI development in the Global South and easier exports for capacity building, tech sharing and joint rulemaking within the World Artificial Intelligence Cooperation Organization, a bloc China launched to counter Western AI bodies.

The practical effect on Western AI consumers may be indirect. US companies that lean on open-weight models have an incentive to shift toward domestic alternatives rather than Chinese ones, which suggests American open-weight ecosystems could absorb demand that previously flowed to Chinese releases. Europe faces a sharper problem: an EU strategy of running open-weight models in domestic clouds for sovereignty, alongside hosting closed top-tier models on local servers, becomes fragile when the leading open models originate in the US or China and can be cut off by export controls. That is my reading of the strategic logic, not a claim made by the ministries.

A likely trade-off is visible in the design itself. Fast-tracking open-weight publication protects China's domestic open-source boom while still giving regulators a registration trail and the legal basis to restrict weights later. The already-public exemption limits retroactive disruption, but it also means the control regime bites hardest on frontier-scale work that has not yet shipped — exactly the models most likely to matter commercially and strategically in the next few years. The reliance on public leaderboards to define "top 10" also introduces a mechanism that foreign evaluators and industry platforms could influence indirectly through how they rank systems.

Why it matters: Labs and platform teams that build on Chinese open-weight models should treat supply as contingent, since future frontier releases could require licensing or fall under restriction, while already-published weights stay exempt. In practice, teams may want to track the compute and parameter thresholds when planning model adoption, and European cloud and sovereignty efforts may need contingency plans if top open models become subject to controls on either side. The registration requirement and final rule timeline are worth watching before any architecture or vendor decisions are locked in.

Read at The Decoder

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 plans to tighten export controls on open-weight models and 23 key AI technologies appeared first on The Decoder .