China's AI Export Control Draft: API Calls Exempt, Overseas Deployment Restricted

China's Ministry of Commerce and Ministry of Science and Technology have jointly released the draft of Announcement No. 48, which proposes adding approximately 23 AI technologies to the export control list. The draft draws a clear line: calling Chinese models via API does not count as export, but Chinese vendors deploying restricted models on overseas servers is subject to control. For developers who rely on Chinese models, understanding this line is crucial.

What the draft defines

Announcement No. 48 proposes adding 40 new technologies to the Catalog of Technologies Prohibited or Restricted from Export, of which 23 fall under the AI domain, covering large models, agents, computing clusters, and information processing. Across all three domains (AI, biomedicine, and solar energy), the draft proposes 14 prohibited and 26 restricted export technologies.

The Ministry of Commerce's AI examples are very specific. According to Xinhua, "AI-driven malicious code generation and mutation technology" is proposed for the prohibited category, while "AI-assisted vulnerability discovery technology" is proposed for the restricted category.

Before the formal catalog takes effect, a temporary pre-export registration system has already taken effect immediately. This registration system specifies four trigger conditions: transferring ownership of listed technologies abroad, authorizing overseas use, domestic technology owners being controlled by foreign capital, and other circumstances designated by the Ministry of Commerce through notification. As a detail, export activities that had already started but were not yet completed before October 16 are also covered.

A Ministry of Commerce spokesperson stated that these rules aim to "prevent abuse" and "ensure AI remains under human control" (according to The Next Web). At the same time, the Ministry emphasized that this revision "does not target any specific country, region, or entity" and that "most AI technologies can still be exported freely."

For overseas developers, where is the line?

The spokesperson's Q&A clarified a core criterion: whether the restricted technology moves from mainland China to another country.

Your action	Draft provision
Calling a Chinese model via API, app, or web UI	Not considered large model export
Running already open-sourced public weights	Not within scope: already in the public domain
Waiting for the lab's next open-source release	Open-source AI technology can receive automatic licensing
Having a vendor install a restricted model on your server	Considered overseas local deployment, within scope
Using its model on the vendor's own overseas cloud	Automatic licensing may apply

The API call row is the most clearly defined. The Q&A states that providing restricted large models through "API calls, APP/web interaction, SaaS applications, local lightweight deployment," does not constitute large model technology export.

The open-source provisions are worth reading repeatedly. Technology that has "entered the public domain" through open source or academic papers is not covered. For versions not yet released, the Ministry of Commerce proposes automatic licensing and states that it should basically be "file and export"—export permitted upon filing. In other words, the draft places a filing step before open-source releases of restricted models, while saying this step should not cause delays. But how fast it will actually work in practice will only be known once the final rules are issued.

Automatic licensing is a type of license within China's export control system, falling under the licensing framework stipulated by the Export Control Law. In Chinese export control practice, technology export licenses are generally divided into three categories: "prohibited," "restricted," and "free." Restricted technologies require case-by-case license applications. "Automatic licensing" is a relatively looser mechanism: exporters only need to register with the competent authorities and can obtain export licenses without case-by-case approval, essentially preserving regulatory visibility procedurally without creating a substantive approval threshold. The draft's proposal to grant automatic licensing for not-yet-released open-source AI versions means officials want to preserve filing information while avoiding delays to the normal release cadence of open-source models due to approval processes. This contrasts with "restricted" technologies that require case-by-case approval, and also differs from full exemptions (such as already open-sourced weights that have entered the public domain).

The technical threshold is hidden in an attachment

The 23 AI technologies and their specific technical parameters are located in Attachment 1, which is a .wps file that cannot be rendered directly on the announcement page. The figures currently mentioned by the media come from news reports rather than the attachment itself.

The Next Web, citing Bloomberg, reported that large model controls cover model weights, source code, and other core components for models with 2 trillion parameters or more, or models ranked in the global top 10 in tests such as performance and safety. The Associated Press, via Fortune, reported the same 2 trillion parameter threshold and added prohibitions on core technologies of large models trained on sensitive national data, as well as AI agents designed for highly sensitive environments such as defense and critical information infrastructure.

Both media outlets reported that Moonshot AI's Kimi K3 and Alibaba's Qwen3.8-Max exceed this threshold. But the draft itself does not name any specific model.

In July of this year, according to Implicator citing the Financial Times, the Ministry of Commerce consulted companies such as Alibaba, ByteDance, and Zhipu about possible control measures, including tightening restrictions on overseas downloads of model weights while allowing overseas customers to continue using hosted Chinese models. The October draft continues this distinction between "serving the model" and "handing over the model," and provides a faster channel for open source.

Attachment 1 uses the .wps format, which is the proprietary document format of Kingsoft Office's WPS Office, rather than the internationally common PDF or Office Open XML (.docx). This format is not uncommon in government announcements, but it creates two practical problems: first, browsers and text tools in non-Chinese environments cannot directly render or parse the content; second, automated scraping and text comparison tools generally do not support this format, making the technical parameters almost entirely unreadable by machines. By contrast, PDF or HTML pages can be easily retrieved and cited directly by developer communities and the media. Therefore, current specific figures such as the 2 trillion parameter threshold and global top 10 performance all come from reports by Bloomberg, the Associated Press, and other media, rather than direct quotations from the attachment itself. This also means that before the final rules are issued, these figures may be adjusted or refined, and developers should treat them as reference values rather than definitive thresholds when making technology choices.

Developer response strategies

The draft is not yet final, and your tech stack will not be interrupted by it today. But if you rely on Chinese models in production, the following measures are worth preparing in advance.

Lock down the exact weight version you run

Weights that have already been made public are not covered by the draft, but the next release of a restricted model may need to go through a filing step first. If you self-host open-source models, it is advisable to record the commit you tested and keep a local copy when the model license permits. When using huggingface_hub, you can specify the commit SHA to lock the version:

from huggingface_hub import snapshot_download

# Pin a commit SHA, not a branch like "main"
path = snapshot_download(
    repo_id="deepseek-ai/DeepSeek-V3",
    revision="<commit-sha-you-tested>",
    local_dir="./models/deepseek-v3",
)
print(path)

Add a second provider as fallback

Many hosted models (including DeepSeek^Ad) provide OpenAI-compatible APIs, so the fallback can simply be two clients and a loop:

import OpenAI from "openai";

const providers = [
  {
    client: new OpenAI({
      baseURL: process.env.PRIMARY_BASE_URL,
      apiKey: process.env.PRIMARY_API_KEY,
    }),
    model: process.env.PRIMARY_MODEL!,
  },
  {
    client: new OpenAI({
      baseURL: process.env.FALLBACK_BASE_URL,
      apiKey: process.env.FALLBACK_API_KEY,
    }),
    model: process.env.FALLBACK_MODEL!,
  },
];

export async function complete(prompt: string) {
  let lastError: unknown;
  for (const { client, model } of providers) {
    try {
      const res = await client.chat.completions.create({
        model,
        messages: [{ role: "user", content: prompt }],
      });
      return res.choices[0].message.content;
    } catch (err) {
      lastError = err;
    }
  }
  throw lastError;
}

Remember to run evaluation tests on the fallback as well, otherwise you are just guessing.

OpenAI-compatible API refers to service implementations whose request paths, request bodies, and response structures remain consistent with OpenAI's official API, usually using /v1/chat/completions as the core endpoint. Chinese vendors such as DeepSeek, Moonshot AI, and Zhipu, as well as many overseas hosting platforms, provide OpenAI-compatible interfaces. The benefit of this compatibility is that developers can use the same client code (such as the openai npm package above or Python's openai library) and switch providers by only replacing baseURL and apiKey. The fallback loop above takes advantage of this: putting multiple provider clients into an array and trying them in order until one request succeeds. Note that although different vendors are API-compatible, they still differ in model names, context windows, tool-calling formats, rate limits, and billing. Therefore, before launching a fallback, you should rerun the evaluation set on the target provider to avoid degraded output quality caused by behavior drift.

Audit ongoing local deployment contracts

If a Chinese vendor is installing a model in your data center, and that model falls within the listed technologies, the Q&A treats it as "overseas local deployment," and the temporary registration system covers exports of listed technologies already underway on October 16. It is advisable to confirm with the vendor whether it applies and whether registration has been completed.

Submit comments if affected

The announcement accepts comments by email, fax, or mail, with a deadline of November 15. Contact information is available on the announcement page.

Open questions

What cannot currently be verified is the specific value of the technical threshold. Which models are deemed restricted depends on the contents of Attachment 1, and the final rules may still adjust these lines.

If you run Chinese open-source weight models in production, or only call them via API, will this filing step change which models you choose to build on?

The temporary pre-export registration system is the part of this draft that has already taken effect immediately. Its legal basis comes from the temporary control and pre-registration mechanisms in the Export Control Law and the Regulations on the Administration of Technology Import and Export. Among the four trigger conditions, the one most worth noting is "domestic technology owners being controlled by foreign capital": it means that even if the technology does not physically cross borders, a change in control may trigger registration obligations. For AI companies with joint ventures or foreign equity arrangements in China, this directly affects compliance assessments of equity transactions and overseas cooperation structures. In addition, registration covers export activities that had already started before October 16 but were not completed, indicating that the system is intended to apply retroactively. Relevant companies need to audit existing contracts and delivery schedules. Unlike the formal catalog, compliance obligations under the pre-registration system are already in effect, and violations may face administrative penalties, so responses cannot wait until the final rules are issued.

One-line Summary

China plans to add 23 AI technologies to export controls, with API calls exempt but overseas local deployment restricted.

Paragraph Summary

China's Ministry of Commerce and Ministry of Science and Technology jointly released the draft of Announcement No. 48, proposing 40 new export-controlled technologies, 23 of which are AI technologies. The draft draws a key line: calling Chinese models via API, app, or web is not considered export, but Chinese vendors deploying restricted models on overseas servers is subject to control. Already open-sourced public weights are exempt because they are in the public domain, while not-yet-released open-source versions may receive automatic licensing. The restricted threshold is reported to be models with more than 2 trillion parameters or in the global top 10 for performance, but the specific parameters are in the undisclosed Attachment 1. The temporary pre-export registration system has already taken effect immediately, covering export activities that began before October 16 but were not completed. The draft's comment period ends November 15, and the final rules may still be adjusted.