China opens comment on 23 AI technologies for its export catalogue

On 16 October 2026 the Ministry of Commerce and the Ministry of Science and Technology published Announcement 2026 No. 48, timestamped 16:00, from 安全与管制局. The ministries propose to revise the Catalogue of Technologies Prohibited or Restricted from Export (《中国禁止出口限制出口技术目录》), adding some artificial intelligence, biomedicine, and photovoltaic technologies. Technology already public through open source or academic papers is proposed to stay outside it. Comment is open, and pre-export registration applies.

The draft list

The announcement says the consultation covers 40 technologies. Artificial intelligence is 23, related to large models, agents, compute clusters, and information processing (大模型、智能体、算力集群、信息处理). Technology already in the public domain is outside the management scope. The sentence is: 对已通过开源、发表学术论文等方式进入公共领域的技术，不纳入管理范围.

Biomedicine is five technologies, in cell and gene therapy, small nucleic-acid drugs, and pathogenic microorganisms. Licensing abroad that does not transfer ownership gets automatic licensing and similar facilitation. Photovoltaics is 12 technologies, and the scope includes key and core equipment that carries them. Annex 1 is the WPS scope file. Annex 2 is the registration form. Comment runs from 16 October through 15 November 2026, by email to [email protected], fax 010-65197565, or post to 商务部安全与管制局, 2 East Chang’an Avenue, Beijing 100731.

From publication until the catalogue is formally revised and brought into effect, an exporter registers on ecomp.mofcom.gov.cn before exporting a listed technology. An export already started, and not finished, is included. The four cases are a transfer of ownership abroad, a license for use abroad, a change of owner or actual controller to a foreign party, and other cases MOFCOM requires by notice.

Open source and deployment

The MOFCOM news office published a spokesperson Q&A the same day, timestamped 16:00. National Business Daily carries the AI answer at 16:15:38.

The spokesperson calls the revision routine administration under Article 20 of the Foreign Trade Law, and says it 不针对任何特定国家、地区或实体: it does not target any specific country, region, or entity. Of the 40, the spokesperson says 14 would be prohibited from export and 26 restricted. Prohibited examples named there include AI-driven malicious-code generation and mutation (人工智能驱动的恶意代码生成与变异技术). The AI answer uses the longer name in annex 1, 人工智能自主驱动恶意代码生成与变异技术. Products made with the listed technologies are outside the management scope. A restricted example is polysilicon engineering of at least 30,000 tonnes a year on one line.

For open-sourcing of AI technologies, Chinese firms’ self-built, self-operated compute centers abroad, overseas chip tape-out, joint research using overseas resources, and work on international standards, the spokesperson says automatic licensing applies, basically achieving “即报即出”. Technology already public through open source or papers is, in that answer, proposed to stay outside export administration (拟不纳入出口管理范围). Annex 1 states the same public-domain exclusion and the same automatic-license cases. “即报即出” is the spokesperson’s phrase.

After formal implementation, the spokesperson’s test is whether restricted technology moves from inside China to outside China. The great majority of AI technologies can still be freely exported. Localized overseas deployment of restricted-category AI technology is inside the administration. Deployment on a cloud the firm itself owns abroad can use the automatic license. An API, app or web interaction, SaaS, or local on-device lightweight deployment that uses a restricted large model is outside large-model technology-export administration.

Weights, parameters, and labs already covered here

In annex 1, restricted large-model core technology is covered when any one listed characteristic is met. It includes weight parameters, key checkpoints, underlying source code, architecture, and training optimisation and inference acceleration (权重参数、关键检查点、底层源代码、架构设计，以及训练优化、推理加速). One characteristic is at least 20,000亿 parameters (2 trillion). Another is a public global top 10 at the time of export. A prohibited entry covers models offered as a service after training on China’s “core data.” A restricted entry covers training on “important data.” Both terms are left to the Data Security Law. Training on at most 0.1 percent important data stays outside that entry when it cannot materially raise core capability and does not reappear in outputs, or when it is lawfully public.

The prohibited malicious-code entry sets attack-success thresholds of at least 60 percent, 85 percent, and 90 percent. A large cluster has more than 10,000 AI accelerator cards. AI processor chip design includes a process of 14 nanometres and below.

The announcement, the Q&A, and annex 1 do not name DeepSeek, Qwen, Kimi, or GLM. On this site: Qwen-Image-2.1-Turbo, a checkpoint under a non-commercial license; GLM-5.3, an open-weight model Anthropic and NIST have assessed; a DeepSeek funding report; and a Moonshot report, the lab behind Kimi, per Bloomberg.

What the wires add

Reuters, Yahoo News, Beijing, 16 October, byline Shi Bu, Liz Lee, and Eduardo Baptista, reports the 40 technologies, comment to 15 November, and a ministry line that the move targets no specific country, region, or entity. Reuters itself calls models, agents, and clusters “the three fundamental building blocks of China’s AI industry.” The announcement’s four groupings include information processing. A longer Japanese Reuters item names DeepSeek and Alibaba beside the API and SaaS line. Those names are Reuters’.

South China Morning Post, Xiaofei Xu, 4:31pm on 16 October, headlines “export bans” for a draft it says would prohibit 14 technologies and restrict 26. Bloomberg Law, Nectar Gan and Luz Ding, 10:48 UTC, names Moonshot’s Kimi K3 and DeepSeek. Nikkei Asia, Cissy Zhou, 18:56 JST, reports the proposal.