China Proposes Export Controls on 23 AI Technologies — Including Model Weights

China's Ministry of Commerce opened a public consultation today on an export control list covering 23 AI technology categories. For the first time, the list includes trained model weights themselves — not just hardware, not just source code, not just applications. Public comment runs through November 15. Interim pre-export registration is already in force; exporters of in-scope items need to register now even before the final rule publishes.

The framing is "protecting national security in the AI era," which mirrors the language the US Department of Commerce used when it imposed chip export controls in October 2022. The symmetry matters. The two largest AI economies are now both operating under the assumption that model artifacts are dual-use technology subject to sovereign control.

What the List Covers

The 23 categories break into four buckets.

Model artifacts. Trained model weights above a parameter threshold (the consultation draft gives a range; the final threshold is open for comment). Training recipes and curated datasets above a token count. Alignment data and post-training artifacts that materially improve model capability.

Software and code. Source code for training frameworks, inference engines, and agent orchestration systems. Compiler and kernel code for AI accelerators. Evaluation harnesses and benchmark implementations. The list does not name specific projects, which means the categories are broad enough to capture most open-source ML infrastructure.

Hardware designs. Chip designs for AI accelerators above a performance threshold. Reference boards and IP blocks. Manufacturing process recipes tuned for AI compute.

Compute configurations. Cluster configurations above a defined aggregate compute threshold — the specific numbers are in the consultation document, but they are calibrated to capture hyperscale training runs while leaving research-scale training unencumbered.

The list does not cover end-use applications (chatbots, search engines, recommendation systems). It targets the supply side — the artifacts needed to train, run, and reproduce frontier-capability models. The posture is hardware-export-control symmetry, applied to the model layer.

Why This Matters

For the open-weights ecosystem. Models above the threshold now need export registration before international release. DeepSeek-V3-class and larger models would be in scope. The DeepSeek pattern of "release the weights globally, iterate on community fine-tunes" gets harder. Either the model is below the threshold (smaller, less capable) or it is in scope (registration, review, controlled release). The "open weights" tier splits into "research-grade open" and "controlled-release capable."

For US-China AI competition. The October 2022 US chip controls were about denying China access to advanced silicon. The new Chinese controls are about denying the rest of the world access to Chinese frontier models. The two policies are structurally symmetric — both treat the model supply chain as dual-use technology subject to sovereign control. The era of unrestricted international model release is over for both sides.

For builders in third countries. The compliance posture is more complex. A SaaS team in Singapore or India that wants to use a Chinese open-weight model now has to evaluate the export registration status. A team in Germany that wants to distribute a fine-tune of a Chinese base model has to evaluate redistribution rights under both Chinese export law and the model's own license.

For the model release cadence. Expect Chinese model releases to slow for anything above the threshold. The interim pre-export registration is a soft gate now and a hard gate after November 15. That is 30 days for major releases to clear, which is a meaningful calendar change.

The Trade-Off

National security rationales for export controls are credible, and the structural symmetry with US policy is real. The cost is also real: the open-weights ecosystem has been one of the strongest forces democratizing AI access since 2023. Controls at the weights layer push that toward a more bifurcated world, with frontier models under national regimes and smaller open releases in the gaps.

The interim pre-export registration is the most aggressive part of the announcement. Even before the final rule, exporters are expected to register. That is a faster pace than the US October 2022 controls, which had a 60-day comment-to-implementation cycle. The Chinese policy is operating on a tighter clock.

The Bottom Line

China is putting trained model weights on its export control list for the first time. The move is structurally symmetric with the US chip controls from 2022, applied to the model layer instead of the silicon layer. For builders using open-weight Chinese models: check the export registration status of the model you are running before the November 15 deadline. For model release strategy: the open-weights tier is bifurcating into "research-grade open" and "controlled-release capable" — pick your position deliberately.