China's AI export draft doesn't ban open-sourcing big models
On Oct 16, 2026, China's commerce and science ministries proposed adding 23 AI technologies to the country's export-control list. Publishing a big model's weights would need an automatic licence: the company files and gets approval without review. Comments close Nov 15.
What each release path needs
| Path abroad | Big model | Other models |
|---|---|---|
| Publish open weights | Automatic licence | Free |
| API, app, website or SaaS | Not an export | Not an export |
| Deploy on own overseas cloud | Automatic licence | Free |
| License or sell weights to a foreign firm | Licence, reviewedRegister first, starting now | Free |
| Released before the rules | Not covered | Not covered |
Most Chinese flagship models meet the top-10 test even though they fall below the compute and parameter limits (Kimi K2 has about 1T parameters). Models trained on China's "core data" can't be exported at all.
Who got it right
X post, "Chinese labs can only open-source small models": wrong. Open weights get an automatic licence. The thresholds it quotes are accurate.
Community note, "open source is excluded": half right. That only covers models already public. Each new big open release still needs a licence, and the government could later withhold it.