China's Draft AI Export Rules Set Thresholds at 2 Trillion Parameters and 10^26 FLOPs, Exempt API Access

Since China's Ministry of Commerce and Ministry of Science and Technology proposed on Oct. 16 adding 40 AI, biomedicine and solar technologies to the export catalog, the draft's technical annex has drawn closer scrutiny. The AI section accounts for 23 of the 40 entries, and the thresholds are specific.

The thresholds

The core entry, item 266504X, covers a large model's weights, key checkpoints, underlying source code and architecture, plus its training-optimization and inference-acceleration methods. A model falls in scope if it meets any one of three tests: training compute of 10^26 FLOPs, at least 2 trillion parameters, or a global top-10 ranking.

The ranking test is defined in the annex. A model counts as globally leading if mainstream evaluators or large industry platforms place it in the top 10 on overall performance, specialized tasks, call volume, safety or training efficiency. The measurement is taken at the time of export, so a model's status can change with the rankings.

A separate line covers models trained on China's important or core data under its Data Security Law and offered as a service. A model trained with 0.1% or less such data, cleaned and unable to reproduce it in outputs, falls outside that line. The most sensitive version, a core-data model offered as a service, would be banned outright.

The draft does not name individual commercial models, according to the Associated Press.

Agents, chips and malware

Agents get their own entry, 266505X, written as measurable tests. An agent platform is covered if it coordinates teams of 60 agents or uses 200 tools at 90% success. The tool test also asks for 90% accuracy in choosing tools and 90% end-to-end success on tasks that chain three or more of them. A fourth test covers game-playing decision systems that beat 90% of humans on standard benchmarks.

High-performance agents built mainly for critical infrastructure, defense or core government systems would be banned.

The Associated Press reported that the draft also reaches chip-design technologies, including those for AI processing and memory chips. On the banned side, the commerce ministry's spokesperson named AI that autonomously generates or mutates malicious code. AI-assisted vulnerability discovery is the first restricted example the ministry gave.

The other 17 entries are five biomedicine technologies, including cell and gene therapy, and 12 solar technologies covering polysilicon, wafers and cell manufacturing. The announcement says the solar scope can include key equipment that carries the technology.

What is exempt

Technology already in the public domain through open-source release or academic publication is outside the controls, the ministry said. So is API access.

The ministry's Q&A draws the line at where the model runs. Overseas local deployment of a restricted model counts as a technology export. Offering it through API calls, apps, web pages, SaaS products or lightweight on-device deployment does not.

The draft also introduces "automatic licensing," under which certain exports of restricted technology would proceed essentially upon filing. The ministry says this can apply to open-source releases, to deployment on a Chinese company's own overseas cloud, and to other "relatively controllable-risk circumstances." The law firm Baker McKenzie notes that no such mechanism exists under the current framework and that it is expected to be defined in future rules or guidelines.

Registration is already in force

The registration requirement did not wait for the comment period. Since Oct. 16, exporters must register covered transfers through the commerce ministry's technology-trade platform. That includes transfers of technology ownership, overseas licensing, changes that place a Chinese technology owner or controller under a foreign entity, and other transactions the ministry specifies. It also reaches transactions that began before Oct. 16 and are not yet complete.

The filing asks for the exporter, recipient, destination, technology, intended use and expected value. Baker McKenzie stresses that registration is not approval or authorization. It gives regulators visibility while the consultation runs.

Beijing's framing and the skeptics

The commerce ministry says the revision is routine catalog maintenance and is not aimed at any country, region or entity. It also says most AI technologies will remain freely exportable. "We want to ensure that AI technologies remain safe and controllable," a ministry spokesperson said on Oct. 16.

Some analysts read it differently. Ryan Etzcorn, senior director at Ankura China Advisors, called it more than routine maintenance and said Beijing is signaling it intends to protect competitive advantages in industries where Chinese firms have gained ground. Lian Jye Su, chief analyst at Omdia, described it as a response to Western countries imposing sovereignty requirements on AI, robotics and telecommunications.

The stakes for global developers are large. Tech Times, citing its own figures, says Chinese models account for roughly 41% of downloads on Hugging Face and between 30% and 46% of weekly API traffic on OpenRouter. Models already public stay outside the controls, so the practical effect falls on future frontier releases. Whether a given upcoming model crosses a threshold would be decided under tests that include a ranking that moves.

What comes next

Public comment runs through Nov. 15. The ministries have not announced a date for the final catalog. Until it takes effect, registration applies to every listed technology, and the detailed rules for automatic licensing, which determine how much of the open-source ecosystem actually clears the new gate, have not been published.