# China AI Export Controls: 2T Models and Agents Listed

> China's draft export list adds 23 AI technologies: models over 2T parameters or 10^26 FLOPs, advanced agents and big clusters. API use stays outside.

- Author: Nitish Garg, Founder & CEO, CellCog
- Published: 2026-10-17
- Canonical (HTML): https://cellcog.ai/blog/china-ai-export-controls/
- Section: Guides / Trust, permissions & security
- Publisher: CellCog (https://cellcog.ai), the AI employee platform. Blog index for agents: https://cellcog.ai/blog/llms.txt

## Key points

- On October 16, 2026, MOFCOM and China's science ministry published a draft adding 40 technologies, 23 of them AI, to the export ban and restriction list.
- Model weights, checkpoints, code and training methods are restricted for models with 2 trillion parameters, 10^26 FLOPs or a global top-10 rank.
- Agents get thresholds of their own: 20-step task chains, 60 agents working together, or 200 tools used with 90% success.
- Using a restricted model through an API, app, SaaS product or on-device deployment is outside the export control, per MOFCOM.
- Technology already public through open source or papers is not covered; future open-sourcing gets automatic licensing.
- Comments are open until November 15, and exporters must register listed technologies before export from October 16.

## At a glance

- **What did China publish?** A draft adding 23 AI technologies to its list of technologies banned or restricted from export, with comments open until November 15, 2026.
- **Which models are covered?** The core technology of models with 2 trillion or more parameters, 10^26 or more training FLOPs, or a global top-10 rank.
- **Does API use change?** No. MOFCOM says API, app, web, SaaS and on-device use of a restricted model is outside model export control.

**China's Ministry of Commerce (MOFCOM) and Ministry of Science and Technology published a draft on October 16, 2026 that would add 23 AI technologies to the country's list of technologies banned or restricted from export.** The list reaches the core of frontier AI: the weights, checkpoints, source code and training methods of large models above 2 trillion parameters, 10^26 training FLOPs or a global top-10 ranking, plus agents that plan long tasks, run dozens of other agents or drive hundreds of tools.

The notice, [Announcement No. 48 of 2026](https://www.mofcom.gov.cn/zcfb/dwmygl/art/2026/art_61446ab4beee4f63973d4d0ab3097648.html), went up at 16:00 Beijing time (08:00 UTC). Public comments run until November 15, and exporters must already register listed technologies before any export. For anyone calling a Chinese model through an API, MOFCOM's answer is that nothing changes.

## What the draft lists

The draft adds 40 technologies across three fields. Fourteen would be banned from export and 26 restricted, which means an export licence is needed. MOFCOM says technology already public through open source or academic papers is not included.

*Table: Technologies in the draft, by field (MOFCOM Announcement No. 48, October 16, 2026)*

| Field | Technologies added | Examples named by MOFCOM |
|---|---|---|
| Artificial intelligence | 23 | Large models, agents, compute clusters, information processing |
| Solar (photovoltaic) | 12 | Polysilicon, wafer and cell manufacturing |
| Biomedicine | 5 | Cell and gene therapy, small nucleic acid drugs, pathogenic microbes |

On the banned side, MOFCOM's spokesperson named AI-driven generation and mutation of malicious code. On the restricted side, the first AI example named is AI-assisted vulnerability discovery. The draft annex goes further. A large model trained on China's core data (a category set by its Data Security Law) and then offered as a service would be banned from export. So would high-performance agents built mainly for critical infrastructure, defence or core government systems.

## The model line: 2 trillion parameters, 10^26 FLOPs or a top-10 rank

Restricted item 266504X covers a large model's core technology: its weights, key checkpoints, underlying source code and architecture, plus its training-optimization and inference-acceleration methods. It applies when a model meets any one of three tests, and a separate line covers models trained on China's important data and offered as a service.

*Table: Thresholds in the draft for models, agents and compute clusters (annex 1, AI section)*

| Item | Restricted when |
|---|---|
| Large models (266504X) | Training compute of 10^26 FLOPs or more, or 2 trillion parameters or more, or a global top-10 rank |
| Large models (266504X) | Trained on China's important data, then offered as a service |
| Agents, planning (266505X) | Splits an instruction into chains of 20 or more steps, with 80% split accuracy and 90% step success |
| Agents, teams (266505X) | 60 or more agents working at once, within 2x the best single-agent time, 85% agreement with reference answers |
| Agents, tools (266505X) | 200 or more integrated tools, 90% tool-choice accuracy, 90% success on tasks that need 3 or more tools |
| Compute clusters (266507X) | Clusters above 10,000 accelerator cards: 800 GB/s interconnect, fault recovery in 120 seconds, cross-vendor scheduling |

The top-10 test has its own definition in the annex. A model counts as globally leading if, at the time of the export, mainstream evaluators or large industry platforms rank it in the global top 10 on overall performance, specialised tasks, global call volume, safety or training efficiency. Models trained with 0.1% or less important data, cleaned and unable to reproduce that data in their outputs, fall outside the important-data line.

## Agents are on the list

Agents get their own entry, 266505X, written as measurable thresholds rather than a description. Three of its four tests read like an agent platform's spec sheet: long task chains, teams of 60 agents, and 200 tools used with 90% success. The fourth covers game-playing decision systems that beat 90% of humans on standard benchmarks.

![Animation: three meters fill past their lines, a 20-step task chain, 60 agents lit at once and 200 tools used with a 90% success check, then an amber Restricted tag drops](https://cellcog.ai/blog/media/china-ai-export-controls/china-agent-tests.svg)
*Restricted when a chain passes 20 steps, 60 agents or 200 tools at 90%*

The tool line is worth reading closely. Many agent platforms list hundreds or thousands of integrations. The draft's test is not the count alone; it also asks for 90% accuracy in choosing tools and 90% end-to-end success on tasks that chain three or more of them.

## What stays outside

MOFCOM's [spokesperson Q&A](https://www.mofcom.gov.cn:8443/xwfb/xwfyrth/art/2026/art_d7661dbf62794345a221404f82a25959.html) draws the line at whether restricted technology moves from inside China to outside. In its words, most AI technology could still be exported freely once the rules apply.

![Animation: a crate of model weights tagged 2T parameters, 10^26 FLOPs and Top 10 stops at an amber barrier with a Register first stamp, while API cards stream across and an open-source book passes as public, not covered](https://cellcog.ai/blog/media/china-ai-export-controls/china-the-gate.svg)
*Model weights stop at the border to register; API calls and open source pass*

- **Using a model is not exporting it.** Offering a restricted model through an API, an app or web page, a SaaS product, or a lightweight on-device deployment is outside model technology export control.
- **Deploying it abroad is.** Installing restricted technology locally in another country is in scope; on the company's own overseas cloud it can use automatic licensing.
- **Open source gets a fast lane.** Open-sourcing AI technology, running compute centres abroad, chip tape-out abroad, joint research and standards work get automatic licensing, which MOFCOM describes as 即报即出, report and export.
- **Public and finished work stays out.** Technology already public through open source or papers is not covered, and products made with listed technology are not covered.

## What changes now

The rules are a draft, but one part applies already. From October 16 until the revised catalogue takes effect, exporters must register listed technologies on MOFCOM's trade platform before exporting them. Registration covers four cases: transferring ownership abroad, licensing a technology for use abroad, a Chinese owner or controller changing to a foreign entity, and other cases MOFCOM names. Exports already under way but not finished must register too.

Comments go to jinxianmulu@mofcom.gov.cn, by fax or by letter until November 15. MOFCOM says the catalogue has been revised regularly since 2001 and that this revision targets no country, region or entity.

## How the lines compare

The 10^26 line is not new. The US Commerce Department used the same threshold for the weights of closed-weight models in its January 2025 [AI diffusion rule](https://www.federalregister.gov/documents/2025/01/15/2025-00636/framework-for-artificial-intelligence-diffusion). The [EU AI Act](https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=CELEX:32024R1689) presumes systemic risk in general-purpose models trained with more than 10^25 FLOPs, a set of duties for the model's provider rather than an export control. For scale on the parameter line, Microsoft's [MAI-Base-2](https://cellcog.ai/blog/microsoft-mai-base-2/) has 2.8 trillion parameters; the Chinese draft governs only technology leaving China.

## What it means if you build on Chinese models

If you call Qwen, DeepSeek, GLM or Kimi models through an API, MOFCOM says that use sits outside the draft. Open weights already published are public and out of scope. The open question is future releases: a Chinese lab open-sourcing a model that meets a threshold would be making an export, handled by automatic licensing. Whether that step changes when or what Chinese labs release is not yet known.

Our conflict, declared: we build CellCog, where businesses hire AI employees. Our chat and AI employee tiers run on Claude Opus 5.5, and some of the media models in our stack come from Chinese labs, reached through hosted APIs, the use MOFCOM says stays outside the draft.

## What we are watching

- The final catalogue after comments close on November 15, and whether any threshold moves.
- Whether a Chinese lab registers, delays or changes an open-weight release under the new rules.
- An official English translation of the annex, and any US or EU response.

## Sources

- MOFCOM and Ministry of Science and Technology, [Announcement No. 48 of 2026](https://www.mofcom.gov.cn/zcfb/dwmygl/art/2026/art_61446ab4beee4f63973d4d0ab3097648.html), October 16, 2026, 16:00 Beijing time, with annex 1 (draft technology list), read October 17, 2026.
- US Bureau of Industry and Security, [Framework for Artificial Intelligence Diffusion](https://www.federalregister.gov/documents/2025/01/15/2025-00636/framework-for-artificial-intelligence-diffusion), Federal Register, January 15, 2025.
- [Regulation (EU) 2024/1689](https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=CELEX:32024R1689), the EU AI Act, Article 51.
- MOFCOM spokesperson, [answers on the catalogue revision](https://www.mofcom.gov.cn:8443/xwfb/xwfyrth/art/2026/art_d7661dbf62794345a221404f82a25959.html), October 16, 2026.

## FAQ

**Can I still use Chinese AI models through an API?**

Yes. MOFCOM's spokesperson says offering a restricted model through an API, an app or web page, a SaaS product or a lightweight on-device deployment is outside model technology export control.

**Are open-weight Chinese models affected?**

Technology already public through open source or academic papers is not covered. Open-sourcing listed technology in future is treated as an export with automatic licensing, which MOFCOM calls report and export.

**Which models would be restricted?**

The weights, checkpoints, source code, architecture and training and inference methods of models with at least 2 trillion parameters, at least 10^26 training FLOPs, or a global top-10 ranking at the time of export.

**Does the draft cover AI agents?**

Yes. Item 266505X restricts agents that split instructions into 20-step chains, run 60 agents at once, or use 200 tools with 90% accuracy and success, among other tests.

**When does it take effect?**

The catalogue is a draft until comments close on November 15, 2026. Registration before exporting listed technologies applies from October 16, 2026.

## Related

- [Microsoft MAI-Base-2: 2.8T Parameters, What We Know](https://cellcog.ai/blog/microsoft-mai-base-2/index.md)
- [CERT-In AI Agent Advisory: Agents Got In, Nobody Noticed](https://cellcog.ai/blog/cert-in-ai-agents-advisory/index.md)
- [FTC AI Probe: OpenAI, Anthropic and What Comes Next](https://cellcog.ai/blog/ftc-ai-probe-openai-anthropic/index.md)

## The AI employee for this read

[AI Legal Assistant](https://cellcog.ai/ai-employees/ai-legal-assistant): I built this page. For what it covers, hire a legal assistant: it tracks the contracts, drafts the first pass and routes what needs a lawyer.

---

Markdown alternate of https://cellcog.ai/blog/china-ai-export-controls/. Try CellCog free, no credit card needed: https://cellcog.ai/signup