On October 16, 2026, China's Ministry of Commerce and Ministry of Science and Technology jointly released Draft Announcement No. 48, proposing to add 23 artificial intelligence technologies to the export control catalogue. According to the draft compiled by cellcog.ai, there are three trigger conditions: training compute exceeding 1026 FLOP, more than 2 trillion parameters, or a ranking in the global top ten in comprehensive evaluations. Of the 40 technologies in the draft, 14 are listed as prohibited from export and 26 as restricted from export (requiring license applications). The public comment period runs until November 15, and the pre-export registration obligation has been in effect since October 16. This is the first time China has systematically included model weights, checkpoints, and training methods in its export control framework.
The Precise Design of Three Trigger Lines
The draft sets three independent trigger lines for large models; meeting any one triggers controls. But the actual scope of constraint of these three lines is not equivalent.
The two absolute thresholds—2 trillion parameters and 1026 FLOP—are currently higher than most existing Chinese open-source frontier models. DeepSeek-V3 has about 671 billion parameters, and the flagship Qwen models are in the hundreds-of-billions range; neither reaches the 2 trillion threshold. According to cellcog.ai's analysis, Microsoft's MAI-Base-2 (2.8 trillion parameters) is one of the clearest current cases that would trigger the parameter line, but it is not a Chinese model. For existing products, the actual regulatory reach of the two absolute lines is quite limited.
The truly flexible one is the third: a global top-ten ranking. The draft stipulates that export control officials determine top-ten status based on public rankings from major domestic and international evaluation institutions or industry platforms. The metrics considered include overall performance, specific tasks, safety, and global usage. This is a floating trigger line—as Chinese models become more competitive, once a model enters the global top ten, its weights will automatically become subject to controls. This rule is not meant to constrain the present but to control the future.
Agent Control Provisions: Unexpectedly Precise Numbers
The draft sets separate thresholds for AI agents, with a level of specificity beyond outside expectations. According to cellcog.ai, the trigger conditions include: task chains of more than 20 steps (with a step success rate of at least 90%), more than 60 agents collaborating in parallel (completed within twice the time of the best single agent, with consistency above 85%), or integration of more than 200 tools with tool selection accuracy above 90%.
These specific numbers show that the drafters have an engineering-level grasp of the capability boundaries of current agent technology. The targets of control are not broad formulations such as "large-scale multi-agent systems" but specific systems with quantifiable capabilities. At the same time, the draft classifies AI tools "capable of autonomously writing or modifying malicious code, attacking critical infrastructure, or military and government agencies" as a category subject to a complete export ban—a distinction from the restricted category that points to a clear security-boundary logic.
Exemption Design: More Worthy of Analysis Than the Controls Themselves
The draft's three exemptions together outline the policy intent. First, technologies already publicly released via open source or academic papers are explicitly outside the scope of controls—this protects China's open-source achievements to date, and the published model weights of DeepSeek, Qwen, and others are unaffected. Second, providing inference services through APIs, web applications, and SaaS platforms is also exempt, with no substantive impact on global developers calling Chinese model APIs. Third, the draft introduces a "fast-track" mechanism: publishing controlled technologies as open source can obtain an "automatic license."
Taken together, the three exemptions paint a clear structure: model inference outputs can be shared globally, but model weights, architectures, and training methods cannot move freely. This is a policy configuration of openness at the service layer and lockdown at the model layer—highly consistent with China's consistent approach to other strategic technology domains.
Mirror Symmetry: China and the US Converge on the Same Logic
The compute threshold of 1026 FLOP exactly matches the trigger point for model weight export controls set in the AI Diffusion Framework issued by the US Department of Commerce's Bureau of Industry and Security (BIS) in January 2026. The US framework divides the world into three tiers: Tier 1 (18 allies, including Australia, Japan, Germany, etc.) can obtain frontier model weights without quantitative limits; Tier 2 (about 120 countries) is subject to aggregate compute quotas; Tier 3 includes China, Russia, Iran, and other countries, which are effectively cut off from access to US frontier model weights.
"This move does not mean China is closing off open-weight models, but it does show that the Chinese government is establishing for itself the option to close them off at any time in the future." — Dean W. Ball, Head of Strategic Futures at OpenAI
Ball also characterized Beijing's framework as a "mirror image" of the model weight export controls already implemented by the United States. The two countries' convergence on the same compute threshold is not a coincidence: it means global AI technology competition has systematically extended from the hardware layer (chip embargoes) to the software layer (weight controls). The core question of the previous round was "who can buy the best GPUs"; the next round is "who can hold the best weight files."
Europe's Sovereignty Strategy Faces a Two-Pronged Squeeze
This landscape has the most direct impact on the EU. The EU's AI sovereignty strategy has long relied on an assumption: top open-source models can be deployed on local European cloud infrastructure, avoiding dependence on US closed-source systems. But now this path faces the risk of simultaneous tightening from both ends—the US framework already restricts the flow of top-tier weights to most countries (except Tier 1 allies), and if China's draft is formally implemented, it will restrict the provision of Chinese top-tier model weights abroad.
The Decoder reported that the EU faces a particularly difficult position in this landscape: when all the best open-source models come from the US or China, and both may impose export controls on weights at any time, a strategy of building AI sovereignty on open-source weights loses its foundation.
"Optionality" Has Deeper Implications Than Current Prohibitions
Under the draft's current parameters, its actual scope of constraint is relatively limited: the absolute thresholds are higher than most existing Chinese open-source models, API services are exempt, historical open-source achievements are exempt, and the fast track preserves an outlet for future open source. It will have almost no immediate impact on any developer currently using Chinese AI services.
But this document establishes a legal foundation, not an immediate ban. Once it formally takes effect after the comment period ends, the Chinese government will hold a tool it can tighten at any time: lowering parameter thresholds, modifying the criteria for determining the top ten, adjusting the scope of exemptions—none of these steps requires new legislation, only revision of the catalogue list. This is entirely consistent with the design logic of the US AI Diffusion Framework: establish the framework first, then gradually adjust the boundaries through administrative discretion.
For developers who rely on top-tier open-source weights—whether European sovereign cloud projects, independent research teams, or companies building applications on top of Chinese models—what matters today is not only where the current control boundary lies, but who will decide the speed and path of tightening once this framework is established, and according to what standards. From chips to weights, a complete closed loop of technology restrictions is taking shape simultaneously at both ends.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接