China drafts export curbs on 23 AI technologies, including model weights

What exactly is being restricted?

China is preparing export restrictions on 23 AI technologies — including model weights, source code, and training methods for systems trained on 10^26 FLOPs or holding more than 2 trillion parameters, according to a report from The Decoder. A separate fast-track provision will still clear open-source releases once they're registered.

The draft framework targets three core artifacts of a frontier AI project. The first is the model weights — the billions or trillions of numerical parameters produced by a training run. Those weights are what make a model work, and they're the single most valuable export a lab produces. The second is the source code used to train and run the model. The third is the training methodology, the recipes and procedures that turn raw data and compute into a working system.

Together, those categories cover most of what any foreign lab would need to reproduce, extend, or audit a Chinese frontier model. The 10^26-FLOP training-compute threshold, paired with a 2-trillion-parameter floor on the released model, defines which systems fall under the gate.

How high is the threshold, really?

10^26 floating-point operations is a very large number in concrete terms. It points roughly at the order of magnitude of compute used to train the current generation of largest AI models. A 2-trillion-parameter floor sits in the same neighborhood. Both gates aim at the frontier tier, not at everyday model releases.

That placement matters. Smaller academic and startup models — typically in the billions of parameters, trained on far less compute — sit well below either line. They stay untouched by the new rules and remain freely exportable.

What does the open-source fast-track do?

The fast-track clause carves out a route for open-weight projects. A developer can register the release and clear it under an accelerated process. In plain terms, you can still publish a Chinese open-weight model — you just have to file first.

The mechanism serves two purposes at once. It keeps Chinese open-weight publishing viable for domestic and international reach. It also creates a registry: Beijing now has a record of every China-origin frontier open-weight system that wants to leave the country. Projects that skip registration fall back under the general export curb.

Why will Western companies likely shift?

The Decoder frames the knock-on effect directly: "Western companies will likely double down on US open-weight models over Chinese alternatives." The logic is mechanical. Friction at export cuts availability. Reduced availability makes the equivalent Western release — same tier, no export gate — more attractive to the same buyer.

The asymmetry is clean. US academic and corporate open-weight projects ship globally without an export license. A Chinese frontier system now needs registration before it can do the same. Buyers who don't want to wait will pick the option that ships today.

What stays outside the framework?

The 2-trillion-parameter floor and the 10^26-FLOP threshold leave most of the open-weight ecosystem untouched. Plenty of useful open-weight models sit in the billions-of-parameters range. Those continue to move freely across borders.

The framework also doesn't reach finished AI products that don't ship weights — chat products, APIs, image generators accessed through a website. The controls sit at the level of the underlying technology, not the consumer-facing service. That's a narrower target than a chip export ban, but a deeper one than controls on finished software.

What are the stakes for the open-weight race?

Open-weight models have become the working currency of global AI development. They let any team run, audit, and fine-tune a frontier-tier system without negotiating a license. Whoever supplies the dominant open-weight stack ends up with structural advantage in adoption and in downstream research.

China's draft rules don't ban the category. They reshape it. Beijing swaps a binary block for a managed pipeline: the frontier tier requires registration, the long tail stays free. Western labs gain ground among buyers who want unrestricted access. Chinese labs keep domestic reach, with state-monitored export from there.

The take-home point: by gating the frontier tier and registering the rest, China has redrawn the map for who can ship top-end AI weights across borders.

What happens next?

Three signals will tell whether the rules land as the draft describes them.

First, watch the final published list of 23 items. Drafts shift. Categories may move in or out between proposal and enactment, and the line between "model weights," "training methods," and adjacent infrastructure definitions can decide whether whole product lines fall inside or outside the curb.

Second, watch the registration mechanism. How fast is "fast-track" in practice? What does the state ask for at filing? A light registration is a minor speed bump. A heavy one is a de facto veto on Chinese open-weight distribution.

Third, watch how Chinese labs respond. They can release smaller open-weight models that stay below the threshold. They can register frontier systems and accept oversight. They can shift weights to overseas entities, a workaround with its own legal questions. Each response shifts where the open-weight market actually clears.

The bottom line: the structural pressure points one way. US open-weight projects — academic and corporate — don't carry an export gate at all. As registration friction lands on the Chinese side, that asymmetry becomes the story of the next twelve months in global AI distribution.