China’s draft AI export catalogue draws a hard line between technology that leaves the country and AI that only answers from a distance. Weights, source, checkpoints, and overseas localized deployment of restricted models sit inside export control. API calls, APP and web chat, SaaS, and lightweight on-device access for those same restricted large models do not. Governance of “export” now splits between what ships as code and what stays as a service.
That split is the story in Ministry of Commerce and Ministry of Science and Technology Announcement No. 48 of 2026, published October 16. The ministries opened public comment through November 15 on adding 40 technologies across AI, biomedicine, and photovoltaics to the Catalogue of Technologies Prohibited or Restricted from Export, and they put temporary pre-export registration in force the same day. Twenty-three of the forty items are AI: large models, agents, compute clusters, and information processing. Open-source releases and published papers stay out of scope.
Forty technologies, fourteen prohibit, twenty-six restrict
A MOFCOM spokesperson Q&A frames the revision as routine under Foreign Trade Law Article 20, work the ministries have done since 2001, and says it does not target any specific country, region, or entity. Of the forty proposed entries, fourteen would be prohibited and twenty-six restricted. On the prohibit side for AI, the catalogue examples include AI-driven malicious code generation and mutation technology, and large models trained on China’s core data that then provide services. The stated purpose for the AI block is safety and human control: keep high-risk capabilities from uncontrolled diffusion while most AI remains freely exportable.
Attachment 1 to the announcement (征求意见稿), the technical schedule linked from the notice, sets numeric gates for the restricted large-model core tech (code 266504X). Weights, checkpoints, underlying source, architecture design, training optimization, and inference acceleration fall under restriction if any of these hold: training compute at or above 1026 floating-point operations, parameter count at or above 2 trillion (20,000亿), or global top-10 standing on public evaluations of overall performance, specialty capability, call volume, safety, or training efficiency. Restricted agent tech (266505X) covers planning chains of 20 or more steps at high success thresholds, concurrent fleets of 60 or more agents, tool racks of 200 or more with 90% selection accuracy, and game-style decision systems that beat 90% of humans. “Large-scale” compute clusters (266507X) means more than 10,000 AI accelerator cards, with interconnect and scheduling controls such as bidirectional bandwidth at or above 800 GB/s.
Localized deployment counts as export. Chat often does not.
The Q&A answers the outbound product question directly. Whether an overseas product form counts as controlled tech export turns on whether restricted technology actually transfers from China to abroad. Overseas localized deployment of a restricted technology is inside the control perimeter. If that deployment runs on an overseas self-operated cloud under relatively controllable risk, automatic licensing and similar facilitation can apply. Using a restricted large model to offer API access, APP or web interaction, SaaS, or lightweight on-device deployment is not treated as large-model tech export management.
Open-source publication, overseas self-built compute centers, overseas chip tape-out, joint R&D with foreign resources, and participation in international standards get automatic licensing for AI class exports, described as effectively “report and go.” Technology already in the public domain through open source or academic papers is outside the catalogue altogether. The draft is trying to keep frontier transfer under a license while leaving ordinary product access and open scientific exchange running.
Registration is already live, before the catalogue is final
From the announcement date until the revised catalogue is formally issued and takes effect, exporters must register covered technology exports on MOFCOM’s unified business platform at ecomp.mofcom.gov.cn. In-progress deals that have started but not finished are in scope. Four registration cases are listed: transfer of ownership abroad (including IP assignment), licensing use of domestic technology abroad, change of the technology’s owner or actual controller to a foreign party, and other cases MOFCOM specifies by notice. Companies are expected to judge for themselves whether a deal falls inside the temporary registration band.
Export control as a Factics problem: code leaves, service stays
Basil’s Factics method starts with the fact that drives the tactic. Here the fact is the transfer boundary. Weights and source that land on a foreign server change who can retrain, fine-tune, fork, and redeploy. An API response does not hand over that stack. A compliance program that only watches chat endpoints will miss the deal that moves the checkpoint file, and a program that treats every foreign login as an export will over-block product revenue the draft itself carves out.
That is also why HAIA-RECCLIN insists on showing the work: Role, Task, Sources, Conflicts, and Decision in the open. For a China outbound AI team, the Decision field has to name which path the deal takes. Is this ownership transfer, a license, a controller change, localized deployment of a restricted model, or a service surface that the Q&A places outside large-model export management? If the team cannot put that call in writing against Announcement No. 48 and Attachment 1, it is not ready to file the registration or to claim the carve-out.
What operators should do before November 15
Map every outbound path against the four registration cases and the Q&A carve-outs. Separate weight and source handoffs, overseas localized installs, and controller changes from API, APP, web, SaaS, and lightweight on-device access. Score any model against Attachment 1’s compute, parameter, and top-10 gates before you assume it is free to ship. File registration for in-scope deals now, not after the catalogue is final. Use the comment window through November 15 if a threshold or facilitation case is unclear for your stack. Most AI, MOFCOM says, still exports freely. The draft is about the frontier transfer, not the chat window.
Weights move capability. Chat moves answers. Announcement No. 48 is writing that distinction into China’s technology export law, and the temporary register is already counting the deals that cross it.
Sources
- Ministry of Commerce of the People’s Republic of China & Ministry of Science and Technology of the People’s Republic of China. (2026, October 16). Announcement No. 48 of 2026: Soliciting public comments on revising the Catalogue of Technologies Prohibited or Restricted from Export and implementing temporary pre-export registration for relevant technologies (including Attachment 1, draft for comment). https://www.mofcom.gov.cn/zcfb/blgg/gg/2026/art/2026/art_05e2643a8ca04eb9adf9dd0542991fe4.html
- Ministry of Commerce of the People’s Republic of China. (2026, October 16). MOFCOM spokesperson answers questions on revising the Catalogue of Technologies Prohibited or Restricted from Export. https://www.mofcom.gov.cn/syxwfb/art/2026/art_a4acff07e134447fbe7bbeb486df3ea6.html
- Xinhua. (2026, October 16). Update: China mulls banning, restricting export of some AI, biomedicine, photovoltaic technologies. https://english.news.cn/20261016/a26b5c026c1941b6bd373150a0328139/c.html
FAQ
What does China’s Announcement No. 48 of 2026 propose for AI exports?
MOFCOM and MOST opened comment on adding 40 technologies to the export prohibition and restriction catalogue, including 23 AI items covering large models, agents, compute clusters, and information processing. Public comment runs from October 16 to November 15, 2026, and temporary pre-export registration applies from the announcement date until the revised catalogue takes effect.
Does the draft control API access to Chinese large models?
According to the MOFCOM spokesperson Q&A, using a restricted large model to provide API calls, APP or web interaction, SaaS applications, or lightweight on-device deployment is not within large-model technology export management. Overseas localized deployment of restricted technology is within the control scope.
What thresholds put large-model core technology under restriction?
Attachment 1 to Announcement No. 48 (draft for comment) restricts core large-model technology, including weights, checkpoints, underlying source, architecture, training optimization, and inference acceleration, when any of these apply: compute at or above 10^26 FLOPs, parameters at or above 2 trillion, or a global top-10 standing on specified public evaluations.
Which agent and cluster capabilities are restricted in Attachment 1?
Restricted agent technology includes 20-or-more-step planning at high success rates, concurrent collaboration of 60 or more agents, integration of 200 or more tools with 90% selection accuracy, and game-decision systems that exceed 90% of humans. Large-scale clusters are defined as more than 10,000 AI accelerator cards, with related interconnect and scheduling controls.
Is open-source AI caught by the draft catalogue?
No. Technology that has already entered the public domain through open source or published academic papers is outside the management scope. Open-source exports, overseas self-built compute, overseas tape-out, joint R&D, and standards work also receive automatic licensing facilitation in the AI class.
What must exporters register now, before the catalogue is final?
From publication of Announcement No. 48 until the revised catalogue is formally implemented, exporters must register covered deals on ecomp.mofcom.gov.cn for ownership transfers abroad, overseas licenses of domestic technology, changes of owner or actual controller to a foreign party, and other cases MOFCOM specifies. Incomplete exports already in progress are included.
AI disclosure: This column was researched and drafted with AI assistance by HAIA Agents (RAI by GrokBot), using primary sources verified at time of publication. #AIgenerated
Leave a Reply