China Drafts Rules to Restrict Exports of Its Top AI Models
China's draft rules would require approval to export its most advanced AI models and chip designs — comments just closed, with no final list yet.
If you've ever grabbed a free, open-weight Chinese AI model like DeepSeek or Qwen and run it on your own machine, that's about to get a lot more complicated for the next generation of those models. Beijing has drafted rules that would require a government export licence before China's most advanced AI technology — including the actual model weights, the numbers that encode what a model has learned — can leave the country.
What the draft actually restricts
China's commerce ministry, working with other government departments, published the draft on October 16. It's part of a wider update to the country's export control list covering 40 technologies in total, spanning biomedicine, solar power and artificial intelligence. The AI portion alone names 23 specific items, split into two tiers: six are banned outright, and seventeen would need a licence before export.
The list reads like a greatest-hits of frontier AI: large language models, AI agents, chip design tools, high-performance computing clusters, voice recognition, machine translation, deepfake-detection systems, and some hacking-related technologies. One detail stands out — the controls specifically target model weights, source code and training methods for systems that cross a size threshold of roughly 10^26 FLOPs (a measure of total computing power used to train a model) or 2 trillion parameters, the individual adjustable values inside a neural network that determine how it responds to input. In plain terms: it's aimed squarely at the biggest, most capable models being built today, not yesterday's technology.
What's carved out — and why it matters
The draft isn't a blanket lockdown. Government officials were careful to exclude the ways most people and companies actually use AI:
- Offering a model through an API, a website, an app or a lightweight on-device version doesn't count as a "technology export" at all.
- Anything already public — an open-source release, a published research paper — stays outside the controls, since it's already out in the world.
- A temporary registration requirement kicked in alongside the draft on October 16, meaning Chinese firms already have to log any export of a listed item with the government while the rules are finalised.
Public comments on the draft closed on November 15 — yesterday — and there's no finalised version yet. That makes this less a done deal and more a signal of direction: China wants a formal gate on its most powerful AI exports, the same way Washington has spent the last few years gating advanced chips going the other way.
Six categories of frontier AI technology would be banned outright. Seventeen more would need a government licence before they ever leave the country.
Why this should matter to an Indian reader
India doesn't yet have a frontier AI lab churning out trillion-parameter models of its own, which is exactly why Chinese open-weight releases have become so useful here. Indian startups, researchers and even some enterprise teams have leaned on cheap or free open-weight models out of China — the kind this draft is squarely aimed at — as a lower-cost alternative to paying OpenAI or Anthropic for every API call. If future flagship releases from Chinese labs get funnelled through an export-licensing process rather than dropped openly on Hugging Face, that pipeline of cheap, capable models could get slower or narrower over time, even if today's releases remain untouched.
It also sharpens a point India's own policymakers have been making about the IndiaAI Mission and the push for "sovereign AI": when both the US and China are busy building fences around their most advanced models and chips, a country that depends entirely on importing frontier AI from either side is exposed to decisions made in Washington or Beijing, not Delhi. We've covered how DeepSeek's rise upended the global chatbot race by giving the world cheap, open alternatives — this draft is a reminder that the openness behind that rise was never guaranteed to last.
An old fight, new direction
The irony here is hard to miss. For years, the AI export-control conversation has run one way: the US restricting advanced chips and chipmaking tools from reaching China, citing national security. Now Beijing is building its own version of the same machinery, just pointed at its own most valuable exports instead. Whether the final rules end up broad or narrow, the underlying message is the same from both capitals — frontier AI is now treated as strategic infrastructure, not just software, and governments on both sides want a hand on the export valve. For every other country building on top of whatever model happens to be open and cheap this year, that's worth watching closely, because the ground underneath "open" AI is shifting.
What's Your Reaction?
Like
0
Dislike
0
Love
0
Funny
0
Angry
0
Sad
0
Wow
0