China Clears DeepSeek and MiniMax to Ship Models Overseas
China's commerce ministry added AI technologies to its export catalog this month, and DeepSeek and MiniMax have become the first labs to complete pre-export registration under the new rules, according to state media. The registration gives both companies clearance to open-source model weights and license their models overseas, turning what looked like a possible open-weights ban into a formal approval pathway.
On this page
What the rules actually require
On 16 October, Beijing launched new export rules for AI technology, run by the Ministry of Commerce around a draft catalog of 23 AI-related categories with a comment period running through 15 November. Exporting covered technology now requires pre-export registration with the ministry. When the draft appeared, the open-weights community read it as a possible choke point: if releasing weights abroad counts as exporting regulated technology, every Chinese lab's next open release would need government sign-off.
The ministry has since clarified the scope, and the clarification is broader than expected. API access, web and app interfaces, and models already released as open source do not count as regulated exports, which effectively allows existing global deployments to continue while the consultation runs.
Why DeepSeek and MiniMax went first
On 23 October, the state media outlet Yuyuan Tantian, affiliated with China Media Group, reported that DeepSeek and MiniMax have completed registration under the new rules, making them the first AI vendors cleared to export. Reuters and Bloomberg both carried the report, and Reuters framed the clearance as approval for global use of their AI models. The timing for DeepSeek is tight in more ways than one: the company had just reportedly completed a second private financing round of about 100 billion yuan, roughly 15 billion dollars, with Tencent and CATL among the investors, one of China's largest private AI rounds.
The practical effect is a green light with a paper trail. Both labs can continue publishing open weights and licensing models abroad, but each future release now moves through a registered channel rather than a regulatory grey zone.
What it means for open weights
For anyone who downloads Chinese open-weight models, and models from DeepSeek, Qwen, GLM and Kimi anchor a large share of the local-AI ecosystem, the week's signal is reassurance with an asterisk. The worst-case reading of the October 16 draft, that the pipeline of open weights from Chinese labs could be cut off pending approvals, has not happened. Instead there is a formal pathway, two named approvals, and a draft catalog whose final shape is still open for comment until mid-November.
The asterisk is that the pathway is discretionary. A registration regime means each release is a decision, and nothing published so far explains the criteria or the timeline, or whether a model that crosses some capability line could be refused. The companies involved have reasons to cooperate: DeepSeek's week included the reported 15 billion dollar round, and MiniMax trades in Hong Kong, so both have commercial scale to protect. The obvious question is who registers next. GLM, Qwen and Kimi anchor just as much of the local-AI ecosystem, and none has announced registration yet, so the next few open-weights releases from Chinese labs double as a test of whether the pathway is a routine formality or a filter.
Pair this with the other side of the strait: last week Taiwan moved to criminalize AI chip exports to China, and Washington restricts what Nvidia hardware Chinese labs can buy. Now Beijing has its own gate on the models leaving. The weights most people run locally increasingly sit inside a two-sided export-control regime, and both gates are new this month.