MOFCOM Announcement No. 48: China drafts export licences for AI model weights and agents, registration already required

China's Ministry of Commerce (MOFCOM) and Ministry of Science and Technology (MOST) have proposed, in Announcement No. 48 of 16 October 2026, to put AI model weights, AI agents, AI chip and memory design and AI cluster know-how on China's list of technologies that cannot be exported freely. The part that bites today is a second measure in the same notice: a pre-export registration requirement that applies immediately, before the draft list is final. The comment period runs until 15 November 2026. For any company that moves Chinese-developed AI technology across a border, by sale, licence or change of ownership, this is the first Chinese rule written for frontier AI itself.

Key points

• Announcement No. 48 (16 October 2026) proposes 40 additions to China's Catalogue of Technologies Prohibited or Restricted from Export: 23 in AI, 12 in photovoltaics and 5 in biomedicine, of which 14 are outright export bans and 26 need a licence (the sector split is from the announcement text published by Geopolitechs; the ban and licence split is from ChinaTalk and the Cloud Security Alliance).
• Registration is already required: from publication until the revised catalogue takes effect, exporters must register any export of a listed technology, including exports that began but are not finished.
• For model weights and core model know-how, the draft thresholds are any one of: at least 10^26 FLOPs of training compute, at least 2 trillion parameters, or "global leading" capability (a top-10 position on public rankings, call volume or similar measures).
• Technology already public through open-source release or academic publication is excluded, and MOFCOM says API, app, web and SaaS access, and lightweight on-device use, do not count as exporting a large model (per the Cloud Security Alliance's reading of MOFCOM's Q&A).
• Public comments are open from 16 October to 15 November 2026; MOFCOM will then assess them before issuing the final catalogue.

What happened

On 16 October 2026 MOFCOM and MOST jointly published Announcement No. 48 under the Foreign Trade Law and the Regulations on the Administration of Technology Import and Export. It does two things at once. First, it releases for public consultation a revision of the export catalogue covering three sectors: 23 AI technologies, 12 photovoltaic technologies (polysilicon, wafers and solar cells, extending to key equipment that embeds them) and 5 biomedical technologies (cell and gene therapy, small nucleic acid drugs, pathogenic microorganisms). Second, it creates an interim registration regime for every technology listed in the annexes.

The AI items fall into four groups, according to the announcement text: large models, agents, computing clusters and information processing. The annex then sets out control points, mostly as numerical thresholds.

Artificial intelligence23

Photovoltaics12

Biomedicine5

Announcement No. 48, proposed catalogue additions by sector (technologies)

	technologies
Artificial intelligence	23
Photovoltaics	12
Biomedicine	5

Source: MOFCOM/MOST Announcement No. 48 (text published by Geopolitechs)

Registration covers four kinds of transaction: transferring ownership of a listed technology to an overseas party (including intellectual property), licensing it for use overseas, changing the ownership or actual control of a domestic technology so that a foreign entity becomes owner or controller, and any other case MOFCOM names later. The form asks for details of exporter and recipient, destination, the technology and its intended use, and the estimated value, and it is filed through the Technology Trade Management Information System on MOFCOM's Unified Business Platform. Several technologies going to the same recipient in one deal can be registered together.

Background

China keeps two overlapping export regimes. Dual-use goods and related technology fall under the Export Control Law. Technology trade is governed by the Foreign Trade Law and the technology import and export regulations, which classify technologies as prohibited, restricted or freely exportable. In the catalogue, codes ending in J are prohibited and codes ending in X are restricted, according to a note from the law firm HLC. The catalogue has been revised several times since 2001, and in 2020 it added recommendation-algorithm technology, which the Cloud Security Alliance (CSA) notes coincided with the proposed sale of TikTok to a US buyer.

An exporter of a restricted technology must obtain a licence letter of intent before entering into substantive negotiations and an export licence once the contract is signed; the contract takes effect only when the licence is issued (HLC). ChinaTalk adds that provincial authorities administer licensing, though central authorities have intervened in prominent cases.

ChinaTalk argues that this is the first catalogue revision to come with an interim registration regime, and that registration gives Beijing a dataset on what actually leaves the country before it decides what to license. It also reads the third registration trigger, a change of control to a foreign entity, as codifying the case of Manus. In December 2025 Meta agreed to buy the agent startup, which had relocated to Singapore, and the deal was later reversed after Chinese reviews. Under the new draft, ChinaTalk writes, a similar deal would need an export licence, so moving a company's registered home would no longer escape Chinese jurisdiction. That is ChinaTalk's interpretation, not something the announcement itself says.

MOFCOM's spokesperson described the revision as routine and, in the CSA's quotation, said it "does not target any specific country, region, or entity."

The numbers

For model technology, the draft restricts weights, key checkpoints, source code, architecture and training and inference optimisation know-how when a model meets any one of three tests: at least 10^26 FLOPs of training compute, at least 2 trillion parameters, or "global leading" capability. ChinaTalk notes that the compute figure matches the one in the 2023 US executive order, the since-rescinded 2025 US AI Diffusion rule and California's SB 53.

ChinaTalk, citing Epoch AI estimates, says no released Chinese model has crossed the compute line as of 20 October 2026: the largest Chinese training runs sit at roughly 2x10^25 FLOPs.

Draft threshold10

Largest Chinese models (approx.)2

Largest Chinese training compute (estimate) against the draft threshold (10^25 FLOPs)

	10^25 FLOPs
Draft threshold	10
Largest Chinese models (approx.)	2

Source: ChinaTalk citing Epoch AI; MOFCOM Announcement No. 48 via CSA

The parameter test is closer. ChinaTalk, again citing Epoch AI, counts four Chinese models since mid-2024 above 2 trillion parameters (Kimi K3, ERNIE 5.0, Qwen 3.8 Max and Qwen3.8-2.4T-A95B), while models from DeepSeek, Meituan, Xiaomi and Z.ai are below that line but some are near or above one trillion. We have not verified these model counts independently.

The "global leading" test is the vaguest. As the CSA and ChinaTalk both note, the notice names no ranking body, so a model could drop in or out of scope depending on which leaderboard is used on the day of export.

Other control points in the draft, as summarised by the CSA from the annex:

• Models trained on China's "core data" would be prohibited from export; models trained on "important data" restricted, with an exception for trace use of desensitised important data.
• Agents are controlled separately from models: for example, planning and task decomposition above 20-step subtask chains with 80% decomposition accuracy and 90% subtask success; multi-agent systems of at least 60 concurrent agents; tool calling across at least 200 integrated third-party tools. High-performance agents built for critical infrastructure, defence or core government work would be prohibited.
• AI-assisted vulnerability discovery is restricted if it meets any of several tests, including a score of at least 75% on authoritative security benchmarks, a tenfold efficiency gain over senior security experts, or autonomously chaining two or more low or medium severity flaws into a working exploit. Tools that generate and mutate malicious code against highly sensitive systems would be prohibited.
• AI chip design is restricted at 14 nm or below, among other alternative criteria, and cluster technology above 10,000 accelerators with interconnect of at least 800 GB/s.

What it means

For developers and companies using Chinese models. Based on MOFCOM's own Q&A as reported by the CSA, calling a Chinese model through an API, app, web interface or SaaS product is not an export of the model. Running a restricted model locally outside China is, though automatic licensing may apply when a Chinese company deploys on its own overseas cloud. ChinaTalk adds that open-sourcing a restricted model would count as a controlled export and would receive an automatic licence, while models already released with open weights fall outside the rules.

For Chinese AI labs. ChinaTalk points out an incentive problem: because already-public technology is carved out, labs could rush out releases that would otherwise cross the 2 trillion parameter line before the rule is final. That is an analyst's inference, not a reported event. The same piece notes that the compute and parameter tests may be easy to game, since parameter counts are a design choice and mixture-of-experts designs undercount capability when measured by training compute.

For multinationals. HLC says the controls reach beyond products to know-how, source code, model weights and design files, and that firms with R&D, engineering or data operations in China, or that license or acquire technology from Chinese parties, should check exposure. The CSA adds that analysts read the registration form as possibly capturing intra-group code sharing too.

For chips and memory. ChinaTalk reads the chip, memory and cluster items as protecting IP rather than blocking sales of finished hardware: Chinese firms stay free to sell chips and memory, but licensing or selling the underlying designs and know-how would need approval. It also notes that the draft's DRAM and flash thresholds mirror figures from the US October 2022 export controls.

What we don't know yet

• How MOFCOM will treat the "global leading" test, and which rankings count.
• Whether the final list will narrow broad wording that critics have flagged. ChinaTalk says the chip-design clause, as written, appears to cover nearly all digital chips, and calls that either a drafting error or very wide breadth.
• How the registration regime will treat open-sourcing and intra-group code sharing in practice. The form asks about exports not made under a contract, which ChinaTalk takes as a hint that both could be in scope.
• Whether the final catalogue will keep the 40 additions. We have not read the original text on MOFCOM's own website, so the figures above come from the announcement as reproduced by Geopolitechs and from the analyses linked below.

What to watch next

• 15 November 2026: the close of the public comment period.
• The revised catalogue itself, which would end the interim regime and turn the proposals into licensing rules.
• Whether Chinese labs time their next releases around the draft's open-source carve-out.
• Any US or EU response, given that the notice borrows its thresholds from earlier US controls.

Sources

• Geopolitechs: China imposes temporary export registration requirements (full announcement text)
• Cloud Security Alliance: China's First Frontier AI Export Controls, MOFCOM Announcement No. 48
• ChinaTalk: MOFCOM's New Rules and the Future of Chinese AI
• HLC: China proposes new export controls on AI, life sciences and photovoltaic technologies