Register First, Open-Source Later? Reading China's New AI Model Export Controls

On October 16, 2026, China's Ministry of Commerce (MOFCOM) and Ministry of Science and Technology issued Announcement No. 48 of 2026, releasing a draft revision of the Catalog of Technologies Prohibited or Restricted from Export for public comment, along with an interim pre-export registration measure. The most closely watched change: core technologies of frontier AI models would be placed under export restriction — and once a model crosses any of three red lines covering training compute, parameter scale, or global ranking, exporting its weights and source code must first pass through the gate of export management.

The question on everyone's mind in the industry: does this mean advanced large models can no longer be open-sourced?

The short answer: it is not "no more open source" — it is "register first, then open-source." Open-sourcing is not banned; it simply changes from a purely technical act into a compliance act with export procedures attached.

What the Announcement Says: Three Red Lines and Interim Registration

Per the draft, the technologies proposed for export restriction cover frontier LLM core technologies that meet any one of the following:

One critical note is attached: technologies that have already entered the public domain through open-sourcing or academic publication are excluded from management. This exemption clause is the key to reading the entire proposal.

Does Open-Sourcing Count as "Export"? The Legal Characterization

Under China's technology export regime, "export" means more than shipping a hard drive across the border — transferring technology abroad, or making it available to foreign parties, can constitute a technology export. Uploading a set of model weights to GitHub or HuggingFace for global download makes it available to anyone, anywhere — objectively, an act of providing technology to the world. For models above the thresholds, the act of open-sourcing itself therefore falls within the reach of export management.

This is where the three red lines show their design intent: the thresholds are set extremely high — 2 trillion parameters, 10²⁶ FLOPs — levels only the most advanced flagship models can reach; the "global top 10" clause then catches any model that breaks into the world's first tier on mainstream benchmarks. In other words, the rule targets "national-asset-grade" models, not the open-source ecosystem itself. The vast majority of open models on the market — single-digit billions to a few hundred billion parameters, trained far below the compute threshold — are outside its scope.

The Grandfather Clause: Already Open-Sourced Models Are Unaffected

"Technologies already in the public domain are excluded" means the existing open weights of GLM, DeepSeek, Qwen and other previously released model families remain outside the new rules — community downloads, fine-tuning, and derivative development are unaffected. The clause respects the established facts of the open-source world and avoids retroactive chaos.

What changes is the path for future frontier models: a qualifying new model would, in principle, need to complete export registration or licensing before open release — closing off the "release first, file later" route and adding a compliance prerequisite to the timing decision.

Three Industry Implications

• Tiered open-source strategies: leading labs will likely combine closed flagship operations with freely open-sourced small and mid-size models, plus delayed open-sourcing of flagships. The thresholds are high enough that the mainstay of the open ecosystem — edge models, domain models, mid-scale base models — remains unconstrained.
• From shipping weights to selling services: API access, cloud services, and cooperative on-premise deployment are not technology exports. Chinese model providers going global will shift from releasing weights to delivering services — a net positive for platform companies with engineering and compliance capabilities.
• Compliance awareness in research collaboration: universities, labs, and cross-border joint research involving disclosure of training details will need clearer boundaries and documentation habits to avoid unintentionally triggering export rules.

An International Mirror — and Open Questions

The draft does not come out of nowhere. The United States has imposed restrictions on cross-border deployment of US model weights under its AI diffusion framework, and the EU is weighing similar arrangements — treating frontier models as strategic assets subject to export controls is becoming a shared move among China, the US, and the EU. Opening the comment period now is both a symmetrical institutional response and a bargaining chip for future international rule-making.

Open questions remain: which leaderboards count for "global top 10"? How is a release that precedes registration treated? How is staged open-sourcing (small models first, flagship later) delineated? These gray areas are precisely what the public consultation seeks to address — the final catalog may well be adjusted.

The NineZenith View

In the short run, the practical impact is limited — thresholds are extreme, existing releases are grandfathered, and the rule is still a draft. In the long run, it marks the arrival of the "open-source compliance" era for China's LLM industry: open-sourcing is no longer a purely engineering decision, but a joint decision across technology, business, and compliance. For enterprises, the real question is not "can we open-source" but the compliance profile of their own supply chain — model provenance, deployment mode, and data cross-border flows. NineZenith Research will continue to track the catalog's revision and publish follow-up analysis.