China drafts export controls covering AI model weights and agents
Half the 40-item draft is AI. Thresholds would gate the Chinese open models builders lean on; pre-export registration is already live, consultation to Nov 15.
Copy markdown
Weights and source code, above a performance line
The draft adds 23 AI items to China's export-control list — 23 of 40 technologies spanning AI, biomedicine and photovoltaics, with 14 marked prohibited and 26 restricted. Model weights and source code get pulled in once the underlying product crosses a (still-unspecified) performance threshold, so the controls bite hardest on the most capable models.
The open weights you pull daily are the open question
Coverage points straight at Qwen, GLM-5.2, DeepSeek and Kimi — the Chinese open-weight models builders download constantly. A threshold-based limit would gate the frontier tiers, but regulators haven't said how it touches checkpoints already posted publicly, which are near-impossible to recall once mirrored. If your stack leans on Chinese open weights, assume the newest and largest releases are the ones at risk.
Registration is already required — the list is still a draft
A pre-export registration requirement for the affected technologies is already in effect as an interim measure, while the full list stays open for public consultation through November 15, 2026. The paperwork friction has started before the rules are even final.
Not just models: agents, compute clusters, chip design
Beyond LLMs, the AI items name intelligent agents, computing-power clusters and chip design. It follows July 2026 consultations with Alibaba and ByteDance that focused on data-transfer limits — a sign Beijing is widening its grip from data to the models and infrastructure themselves.