China's draft export list would cover the weights of its biggest AI models while leaving API access open

Take a team that downloaded Chinese weights last quarter and serves them from its own GPUs. It wants to know on Monday whether anything it already runs is affected, and the answer depends on which report it reads.

PCMag's version turns on one word: existing. Licenses already in place are respected, so weights already released under an open license are protected on that account ^[11]. Neither report addresses a new release of a covered model under a new license. We think the exposure sits in the next release. That is a narrower claim than saying the draft closes off open weights, and it is the one the reports support. PCMag itself frames the conflict with open-weight design as a possibility ^[10].

The ministry's stated aim is safety. A ministry spokesperson said in a statement that the rules aim to "prevent misuse and abuse" and "ensure that AI remains under human control" ^[8]. The ministry added that the plan does not target any country or firm ^[13].

The usage evidence points the same way as the API carve-out, with a limit. PCMag reports that DeepSeek V4.1 Flash and Step 5 Preview have burned through more than 50 trillion tokens, against 5.85 trillion for the latest Claude Opus model ^[14]. That is more than 8.5 times, though it sets two Chinese models against one Claude model ^[15]. A token total is volume. It does not show how many teams depend on a model or whether they reach it by endpoint or by weights, so PCMag's conclusion that the draft should not slow use of these models is a forecast.

For Monday, sort each Chinese model you use on two axes. The first is how you hold it: a hosted endpoint, or weights on your own servers. The second is whether it sits above the draft's size and ranking tests ^[3]. Endpoint use falls under the carve-out the ministry stated ^[6]. Weights of a model below both tests fall outside the model rule as Bloomberg describes it ^[3]. The cell to plan around is weights of a covered model, where the weights rule meets the open-source exemption and the reports differ.

We would leave the other three cells alone and name a hosted fallback for that one. The fallback costs whatever control over data and latency made hosting worth it. And because one test is a world ranking, a model can enter the cell by climbing it, with no change to its weights ^[3].