China's Draft AI Export Controls: What They Cover and What Builders Should Know
On 16 October 2026, China's Ministry of Commerce (MOFCOM) and Ministry of Science and Technology published Announcement No. 48 of 2026, a draft that would add 40 technologies to China's catalogue of technologies prohibited or restricted from export. 23 of them are in AI, covering large models, AI agents, computing clusters and related information processing; the rest are 5 biomedicine and 12 solar (photovoltaic) technologies. It is a draft: public comments are open until 15 November 2026. But one part applies right away: exporters of the listed technologies must register before exporting them, starting from the date of the notice.
This explainer sticks to what the official notice says and what careful readers of the annex have reported, with links to each source.
What the draft covers
According to the official MOFCOM notice, the AI section covers large models, agents, computing clusters and information processing. The notice also says that technology which has already entered the public domain through open source or published academic papers is outside the scope of the controls.
The AP reported that a commerce ministry spokesperson called the move "necessary" and cited national security as one of the reasons. The Wall Street Journal framed it as an effort to safeguard a strategically critical industry amid intensifying competition.
The three thresholds for large models
An unofficial English translation of the annex by SouthwardTech says the "core technology" of a large model is restricted if the model meets any one of these:
| Threshold | What the draft says |
|---|---|
| Compute | 10^26 floating-point operations or more (the draft does not say if this means training compute) |
| Size | 2 trillion parameters or more (it does not say total or active) |
| Capability | "Global leading level": a top-10 ranking in mainstream public assessments at the time of export |
"Core technology" includes weights, key checkpoints, source code and architecture design, plus training optimisation and inference acceleration. Separately, models trained on China's "core data" and then used to provide services would be prohibited from export, while those trained on "important data" would be restricted. Only the Chinese text is authoritative.
What is not covered
Based on the notice and the spokesperson's answers as summarised by SouthwardTech:
- Downloading published open weights and running them yourself is not affected, because public-domain technology is excluded.
- Calling a Chinese model through an API, app or website, or using a SaaS product built on one, does not count as an export for restricted-category models.
- Running a Chinese company's model on servers outside China would count as an export for an unpublished restricted model and would need a licence, with "facilitation measures such as automatic licensing" for some cases.
- A Chinese company open-sourcing a new model that meets a threshold would get an automatic licence, which the spokesperson said "basically" allows export after filing.
Several points remain unclear in the draft, such as whether a top-10 place in a single benchmark is enough. Expect clarifications before the rules are final.
What it means if you build with Chinese models
- Nothing changes today for existing open weights. Models already published, such as DeepSeek's (the V4.1-Flash-1017 weights were released on 17 October under the MIT licence), are public-domain technology under the draft.
- API access is not the target. If you call DeepSeek or other Chinese models through an API or through OpenRouter, the spokesperson's answers say that is not an export for restricted-category models.
- Future frontier releases are the open question. If a future top-10 Chinese model is not open-sourced, sharing its weights abroad would need a licence. Commentators such as The Decoder noted this makes relying deeply on Chinese open-weight releases a little less predictable.
- Keep a fallback. If your product depends on one model family, keep a tested alternative. Our free LLM APIs list and open-source local LLM tools are good places to look, and our AI API pricing guide helps compare costs.
Timeline
- 16 October 2026: Announcement No. 48 published; interim pre-export registration starts.
- 15 November 2026: Public comment period closes.
- Later: The revised catalogue takes effect on a date not yet announced.
Sources
- MOFCOM: Announcement No. 48 of 2026 (Chinese)
- AP News: China moves to curb tech exports including for AI and solar
- WSJ: China Proposes Export Controls on AI Technologies
- SouthwardTech: what counts as an export (with annex translation)
FAQ
Does China's draft ban open-source AI models?
No. The notice says technology already in the public domain through open source or published papers is outside the scope. A new model that meets a threshold and is open-sourced would get an automatic licence under the draft.
Can I still use DeepSeek or Qwen through an API?
Yes. According to the commerce ministry spokesperson's answers, calling a restricted-category model through an API, app or website is not an export under the draft.
When do the rules take effect?
The catalogue changes are a draft with comments open until 15 November 2026. The interim pre-export registration requirement took effect on 16 October 2026.
Which models are affected?
The draft names no models. It sets thresholds instead: 10^26 FLOPs of compute, 2 trillion parameters, or a global top-10 ranking, plus rules based on the training data used.