POLICY · REGULATION · #2133
China drafts tighter AI export controls targeting 23 technologies including open-weight models
China's commerce and science ministries published a draft update to the country's export-control catalogue that restricts 23 AI technologies, covering large-model core assets (weights, checkpoints, code, architecture, training/inference methods), agents, compute clusters and tools that can autonomously write or alter malware. Controls apply when models are trained on China's "important data," exceed thresholds (2 trillion parameters, 10^26 FLOPs) or rank in the global top 10; already public research or open-source releases are exempt and the draft fast-tracks open-weight releases with automatic licensing upon registration.
KEY POINTS
- China's commerce and science ministries published a draft update to the country's export-control catalogue that restricts 23 AI technologies, covering large-model core assets (weights, checkpoints, code, architecture, training/inference methods), agents, compute clusters and tools that can autonomously write or alter malware.
- Controls apply when models are trained on China's "important data," exceed thresholds (2 trillion parameters, 10^26 FLOPs) or rank in the global top 10; already public research or open-source releases are exempt and the draft fast-tracks open-weight releases with automatic licensing upon registration.
- This draft could reshape global model access and supply chains by blocking exports of top-tier and China-trained models while fast-tracking open-source releases, affecting where companies host and source models and EU strategies that rely on open-weight models.
DECISION BRIEF
CONFIDENCE
WHAT CHANGED
China's commerce and science ministries published a draft update to the national export-control catalogue that would restrict 23 AI technologies. The draft explicitly targets large-model core assets (weights, checkpoints, source code, architecture, training/inference methods), AI agents, compute clusters, and tools that can autonomously write or alter malware. It exempts already-public/open-source AI tech and fast-tracks open-weight releases with automatic licensing upon registration.
WHY NOW
The draft sets technical thresholds and procedural rules that could change who can export or access advanced models and tools (e.g., thresholds for size/compute, data-origin restrictions, and fast-track open-weight registration). Because it is a draft from two ministries, it signals a potential near-term regulatory change affecting cross-border model supply chains and model-sharing practices.
WHO IS AFFECTED
Directly affected parties include AI model developers and vendors (especially those producing large models or open-weight releases), entities exporting AI technologies, and organizations using models trained on Chinese "important data." Cloud and compute providers, and firms offering vulnerability/discovery tools or autonomous code-generation tools for malware, are also implicated by the draft's scope.
CONFIRMED
- A draft update to China's export-control catalogue was released by the commerce and science ministries. - The draft targets 23 AI technology areas, including large-model core assets (weights, checkpoints, code, architecture, training/inference methods), agents, compute clusters, and autonomous malware-capable tools. (2525, 2528) - Controls apply when a system was trained on China’s "important data" or when core model technologies meet thresholds: at least 2 trillion parameters, 10^26 FLOPs of training compute, or if the model ranks in the global top 10 by public evaluators/industry platforms. (2525, 2528) - Already-public/open-source AI tech is exempt. The draft fast-tracks open-weight releases via automatic approval upon registration. (2525, 2528) - AI tools that can autonomously write or alter malware aimed at critical infrastructure, military, or government agencies face a total export ban; some vulnerability-finding tools would face export limits rather than an outright ban. (2525, 2528) - The draft permits up to 0.1% of certain "important data" in training under strict conditions (cleaned beforehand, not appearing in outputs, and not materially improving R&D capacity).
UNCERTAIN
Missing or unclear in the supplied excerpts: the draft's final timeline and whether/when the draft will be enacted; the full legal text and precise legal definitions (e.g., operational definition of "important data"); detailed procedures for determining and publishing the ‘global top-10’ lists and which public evaluators/platforms will be used in practice; exact mechanics and eligibility criteria for the promised automatic licensing/registration of open-weight releases; enforcement mechanisms and cross-border compliance processes; whether duplicate media coverage reflects identical drafting or later revisions. The supplied sources are media excerpts from one publisher and do not include the official ministry text or independent confirmation.
WHAT TO WATCH
Observable signals to monitor: publication of the full draft or final rule on official ministry or government gazettes; release of the detailed legal text and definitions (especially for "important data"); launch of a registration/licensing portal or guidance for open-weight releases; lists or announcements of which public evaluators/platforms will determine global top-10 rankings; any initial licensing or export-denial decisions citing the new rules; official Q&A or implementing guidance on the 0.1% data allowance; reactions or clarifications from other governments, cloud providers, and major model developers.
WHY IT MATTERS
This draft could reshape global model access and supply chains by blocking exports of top-tier and China-trained models while fast-tracking open-source releases, affecting where companies host and source models and EU strategies that rely on open-weight models.
EVIDENCE MAP
4Editorial claims linked to specific sources, with support, contradiction and context shown separately.
China's commerce and science ministries published a draft update to the export-control catalogue that targets 23 AI technologies, including frontier/large models, AI agents, and autonomous malware tools.
SUPPORTEDChecked 2026-10-16 · 2 supporting
Article states ministries released a draft targeting 23 key areas including frontier models, agents, and autonomous malware tools.
SUPPORTS The DecoderMEDIA · 2026-10-16Duplicate article text also lists 23 targeted AI areas including models, agents, and autonomous malware tools.
The draft would impose export controls on models trained on China's "important data" when they provide services based on that data.
SUPPORTEDChecked 2026-10-16 · 1 supporting
The draft would also impose controls on core large-model technologies — weights, checkpoints, source code, architecture, and training or inference methods — once a model meets thresholds (≥2 trillion parameters, ≥10^26 FLOPs) or ranks in the global top 10.
SUPPORTEDChecked 2026-10-16 · 1 supporting
Publicly available research and already open-source releases are exempt from these export controls; the draft fast-tracks open-weight releases with automatic approval upon registration.
SUPPORTEDChecked 2026-10-16 · 2 supporting
Excerpt states that already public AI tech is exempt and the plan fast-tracks open-weight releases with automatic approval upon registration.
SUPPORTS The DecoderMEDIA · 2026-10-16Duplicate text likewise notes exemptions for public/open-source releases and automatic approval for registered open-weight releases.
SOURCES & TIMELINE
1China's commerce and science ministries are moving to tighten AI export controls, releasing a draft rule that targets 23 key areas including frontier models, AI agents, and autonomous malware tools. Model restrictions kick in if a system was trained on sensitive Chinese data, or if it hits 2 trillion parameters, 10^26 FLOPs of training compute, or a global top-10 ranking. To keep from slowing down domestic AI labs,…
China's commerce and science ministries are moving to tighten AI export controls, releasing a draft rule that targets 23 key areas including frontier models, AI agents, and autonomous malware tools. Model restrictions kick in if a system was trained on sensitive Chinese data, or if it hits 2 trillion parameters, 10^26 FLOPs of training compute, or a global top-10 ranking. To keep from slowing down domestic AI labs,…