China Drafts Export Controls on Open-Weight Models
A draft rule from Beijing's commerce and science ministries puts 23 AI technologies under license, with a fast lane for open-source releases.
In short
China is not banning open-weight models but making them licensable: the draft puts any model above 10^26 FLOPs, 2 trillion parameters, or a global top-10 ranking under export review, while open-source publication gets automatic licensing.
At a glance
- Draft issued October 16, 2026 by China's commerce ministry (Mofcom) and its science ministry.
- Thresholds: 10^26 FLOPs of training compute, 2 trillion parameters, or a global top-10 ranking.
- Open source is the fast lane: publishing weights openly and registering them earns automatic licensing.
- API calls, apps, websites and SaaS stay unrestricted; installing on your own servers abroad needs a license.
- Training data: up to 0.1 percent of cleaned important data is allowed if it does not surface in outputs.
China is not pulling open-weight models off the internet. It is putting their export behind a license. A draft published on October 16, 2026 by the commerce ministry (Mofcom) and the science ministry adds 23 AI technologies to the country's list of banned and restricted export items — but publishing weights as open source and registering them counts as automatically licensed.
Where the thresholds sit
Three triggers decide whether a model is covered: more than 10^26 FLOPs of training compute, 2 trillion parameters, or a place in the global top 10 according to major evaluators. Officials would read that top-10 status off public rankings from domestic and foreign evaluators and industry platforms, weighing overall performance, task-specific scores, safety, or worldwide usage numbers.
The list reaches well past frontier models. It also covers AI agents, vulnerability-finding tools, autonomous malware, compute clusters and data-processing systems. Training data gets its own clause: models trained on China's important data are controlled, with up to 0.1 percent of cleaned data tolerated as long as it does not surface in outputs or materially lift research and development capacity.
Open source as the cleared path
Rather than treating open release as a loophole, the draft turns it into the fast lane — register, and the license counts as granted. Anything already public, through open releases or research papers, stays out of scope. So do API calls, apps, websites and SaaS products, plus trimmed-down versions for local devices. Installing restricted technology on your own servers abroad is the case that still needs an ordinary export license.
Dean W. Ball, who runs strategic futures at OpenAI, reads the draft as groundwork rather than a shutdown.
The move does not mean China is closing down open weights, he argued, but it hands the government "the optionality to shut down future open-weight models at will."
Why the lever matters
Chinese open weights are currently the default for a lot of developers outside the United States. Rest of World reports they trail leading U.S. models by about 3 percent on benchmark scores while costing 50 to 90 percent less than top American frontier systems. DeepSeek, Z.ai, Xiaomi and Tencent sit at the front of the OpenRouter leaderboard.
Graham Webster of Stanford's DigiChina Project puts the dependency bluntly: anyone who needs something close to frontier capability right now is looking at Chinese open models and nothing else. Xi Jinping has framed open models as part of the push for global AI leadership by 2030, and Moonshot AI is negotiating revenue-sharing deals with Microsoft, Amazon and Google for its Kimi K3 model.
What is still unconfirmed
This is a draft, not law. The available reporting gives no effective date, no consultation deadline and no official name for the amended control list. Until final rules land, affected exports are to be registered through the commerce ministry's portal. Which Chinese labs actually cross the thresholds is not spelled out either, so treat any specific naming as unverified for now.
FAQ
Is China banning open-weight AI models?
No. The draft makes exports licensable, but publishing weights as open source and registering them qualifies for automatic licensing.
Which models fall under China's new export controls?
Models above 10^26 FLOPs of training compute, at or above 2 trillion parameters, or ranked in the global top 10 by major evaluators.
When do the rules take effect?
Unclear. It is a draft with no stated effective date; until then, affected exports are to be registered through the Mofcom portal.
