Regulation

China Advances AI Export Curbs Covering Qwen and Kimi

China's AI export rules for models above 2 trillion parameters move to final phase after November 15, targeting Qwen and Kimi model weights.

Share:XLinkedIn

Key Takeaways

  • 23 AI technologies targeted, comment period closed November 15: MOFCOM and MOST's draft now moves to finalization, with a final catalogue expected in the first quarter of 2027
  • 2 trillion parameters is the size trigger: Bloomberg and the AP reported Alibaba's Qwen3.8-Max and Moonshot AI's Kimi K3 are both above this threshold, putting future weight releases in the licensing regime
  • Interim registration already in force since October 16: companies that transferred covered AI technology without registering are technically in breach of the interim rules right now, before the final catalogue is published
  • API access and open-source are explicitly exempt: Chinese labs can continue cloud API services globally, but distributing weights, source code, or architecture requires a license under the draft rules
  • Both China and the US share the same 10^26 FLOP threshold: the first time both countries have maintained simultaneous outbound AI technology controls at the same computational frontier

China's Ministry of Commerce closed its public comment window on November 15 for a draft export catalogue that would, for the first time, subject AI model weights above 2 trillion parameters to outbound technology controls. Bloomberg and the Associated Press, as cited by The Next Web, reported that Alibaba's Qwen3.8-Max and Moonshot AI's Kimi K3 are both above that threshold. If Beijing keeps the draft numbers in the final catalogue, two of the world's most widely distributed open-weight AI models will require a government license before their weights cross a Chinese border. An interim registration requirement for covered technologies has been in force since October 16, the day the draft was published.

What MOFCOM Announcement No. 48 Actually Controls

On October 16, 2026, China's Ministry of Commerce (MOFCOM) and the Ministry of Science and Technology (MOST) published Announcement No. 48 of 2026, proposing the most sweeping revision of China's dual-use technology export catalogue since it was first introduced. The revision covers 40 technologies across three sectors: 23 in artificial intelligence, 12 in photovoltaics, and 5 in biomedicine. For AI, the catalogue draws two hard lines. Across all 40 items, 14 are proposed for outright prohibition and 26 for restriction. The prohibited AI list covers large models trained on China's core data, AI systems powering high-performance agents deployed in critical infrastructure or national defense, and tools designed to generate or mutate malicious code.

The restricted list is where most of the commercial tension sits. The Cloud Security Alliance's research note on Announcement No. 48 describes the core restricted items as the technical essence of a large language model: trained weights, checkpoints, source code, model architecture, and optimization know-how. Also restricted are multi-agent collaboration architectures, tool-calling systems, AI-assisted vulnerability discovery tools scoring at or above 75 percent on authoritative benchmarks or capable of chaining two or more low-severity vulnerabilities into a working exploit, and deepfake detection technology reaching 98 percent accuracy. The hardware side is equally specific: AI chip designs produced at a process node of 14 nanometers or below fall under restriction, as do computing clusters above 10,000 accelerators with an interconnect bandwidth of 800 gigabytes per second or more, covering most commercially relevant frontier compute clusters built in China today.

The Next Web's reporting on the draft notes that open-source code and published research are outside the controls, and that overseas users reaching a Chinese model through an app, website, API, or lightweight on-device deployment are not treated as receiving an export of the underlying model. The carve-out means Chinese AI labs can maintain and grow cloud API businesses globally while facing new friction for transferring the technical assets, weights, source code, architecture, that allow foreign organizations to run, fine-tune, or build proprietary derivatives from their models independently.

Stay Ahead

Get daily AI signals before the market moves.

Join founders, investors, and operators reading TechFastForward.

Why Qwen and Kimi Fall at the 2 Trillion Threshold

The draft sets three independent triggers for a large model to fall under restriction. Any one of them is sufficient: at least 10^26 training FLOPs, at least 2 trillion parameters, or a global top-10 position on recognized performance or safety rankings at the time of export. That first figure is not an accident. The 10^26 FLOP threshold appeared in the Biden administration's October 2023 executive order on AI and in the since-rescinded 2025 US AI Diffusion Rule. Beijing and Washington have independently arrived at the same computational frontier line, suggesting a convergence in how the two governments define strategically sensitive AI, even as they compete aggressively to control it.

The top-10 global ranking criterion is the most legally uncertain part of the proposal. A model's export classification could change simply because a competitor launches a better model, with no change in the first model itself. The CSA research note describes this criterion as dynamic and potentially retroactive, since the draft names no specific ranking body and no specific benchmark. MOFCOM retains broad discretion to reclassify models at any time, and companies cannot predict their compliance status with confidence years into a product's lifetime.

Bloomberg and the Associated Press reported, as cited by The Next Web, that Alibaba's Qwen3.8-Max and Moonshot AI's Kimi K3 are both above the 2 trillion parameter threshold. Neither company has officially confirmed its model's parameter count, and the draft explicitly names no models. Compliance analysis of the thresholds suggests the 2 trillion parameter figure was chosen precisely to capture models at the current frontier of Chinese AI development, not as a floor that most models fall below. Alibaba has distributed Qwen weights widely through Hugging Face and its own developer platform, and those weights serve as fine-tuning bases for hundreds of derivative models at Western labs. A licensing requirement covering future Qwen weight releases would insert a government approval step into a release pipeline that has until now operated purely as open-source software publishing.

The Interim Registration Requirement Already In Force

Unlike the catalogue additions, which remain a draft pending finalization, an interim registration requirement took effect on October 16, the same day as the announcement. Any covered technology export that starts or is in progress must be registered through MOFCOM's unified business platform before the transfer completes. The form asks for recipient details, destination country, technology description, end use, and estimated value. Four types of transactions trigger the obligation: transferring technology ownership to a foreign party, licensing domestic technology for overseas use, a change in ownership or control that gives a foreign entity de facto control, and other cases MOFCOM designates. That last category is deliberately open-ended.

The CSA research note flags that the registration form language appears to reach code sharing with overseas affiliates and potentially new open-source releases of covered technology, though MOFCOM has not officially confirmed that interpretation. Analysis published at Shattered.io notes that MOFCOM issued a November 4 clarification stating that API calls, app deployments, and SaaS offerings generally do not constitute exports when the underlying technology remains hosted abroad. That clarification is not a binding final rule but signals that the government's intent is to target the transfer of technical control, not the delivery of inference results. Companies operating cloud products built on Chinese models are not, under that reading, in the registration net. Companies that want to run those models on their own infrastructure outside China likely are.

MOFCOM cited its January 2026 review of Meta's then-proposed acquisition of Manus, a Singapore-registered AI agent company whose core technology was developed in China, as a precedent for why redomiciling does not remove technology from Chinese export jurisdiction. That precedent matters for the growing number of Chinese AI startups that have incorporated overseas holding structures while keeping research and training operations in China. Those structures do not insulate the technology from Announcement No. 48.

How Beijing and Washington Now Both Restrict Outbound AI

The emergence of Chinese AI export controls mirrors a dynamic that has been building on the US side for three years. Where US rules try to prevent frontier AI chips from reaching China, Beijing's draft rules try to prevent China's most powerful AI models and chip designs from leaving. The same 10^26 FLOP threshold appearing in both countries' frameworks is the clearest indicator that the two governments, however different their motivations, have converged on a shared definition of what counts as strategically sensitive AI.

The bilateral pressure is building simultaneously. The Korea Times reported on November 13 that Congress is pushing to add four chip-export bills to the annual defense authorization bill after the Nvidia-Megaspeed scandal, in which a Singapore-based neocloud operator imported more than $4.6 billion in Nvidia technology from 2023 to 2025 and rented that capacity to Chinese firms without technically violating then-current rules. Speaker Mike Johnson told reporters there is "a developing consensus in Congress on limiting chip exports to China," and Jensen Huang is expected to testify before the House Foreign Affairs Committee before Thanksgiving on November 26. The Remote Access Security Act, which passed the House 369 to 22 in January 2026, would give the Commerce Department authority to restrict Chinese firms' remote access to US chips in overseas data centers, and Republicans want it in the must-pass defense bill.

China's move reverses the control flow in a complementary direction. Atlantic Council research on the Megaspeed case describes this as the first time both governments have simultaneously maintained outbound AI technology controls at the same computational frontier: the United States trying to prevent China from acquiring AI computing power, China trying to prevent the West from acquiring China's most capable AI model technology. For companies operating in both markets, the practical result is a compliance obligation on both ends of the AI supply chain at the same time, a situation with no historical precedent in the semiconductor or software industries.

The Bear Case: Carveouts May Swallow the Controls

The bear case, however, is that China's new controls are more constrained than they first appear. Critics argue the most consequential transfer has already happened: open-source model weights already distributed globally are explicitly outside the controls. DeepSeek's R2 family, Alibaba's Qwen2.5 series, and other influential Chinese open-weight models are already running at thousands of organizations worldwide. The draft specifically exempts already-public weights, which means Beijing's rules govern what Chinese AI labs may release in the future, not the technology that has already left the country. The practical reach of the controls is prospective, not retrospective.

The exemption for API access compounds the limitation. If a Western organization can receive inference results from a frontier Chinese model through a cloud endpoint without triggering export controls, the practical value of the weight-transfer restrictions depends on whether access to the weights confers a direct technical advantage over API access alone. For most commercial applications, it may not. A developer building a product on top of Qwen through an API does not need the underlying weights to ship a working product. The controls are most consequential for the narrower population of users, primarily research labs, defense contractors, and AI companies that want to understand, fine-tune, or build proprietary derivatives from Chinese model internals rather than simply consuming inference.

US export controls face the same structural limitation. Atlantic Council research on the Megaspeed case documents how third-country neocloud operators provided Chinese firms access to restricted Nvidia chips without physically shipping them to China. Commerce has not added Megaspeed to the Entity List despite documented lobbying from the State, Energy, and Treasury Departments to do so. If Washington cannot enforce its own outbound controls through neoclouds handling $4.6 billion in transactions, Beijing will likely encounter analogous evasion through its API-centric carveout, which provides a legal route to global AI access that the controls explicitly do not close.

What to Watch After November 15

With the public comment period closed, the draft now sits with MOFCOM for revision. No date has been announced for the final catalogue. Based on typical Chinese regulatory timelines, compliance analysts expect a final text in the first quarter of 2027. The critical variable is whether MOFCOM narrows or retains the dynamic top-10 ranking criterion. It is the provision most likely to have drawn the heaviest criticism during the comment period, given that it makes a model's export classification contingent on market conditions outside the model developer's control. A narrowed final text would relieve some compliance uncertainty; a retained or broadened text would create ongoing monitoring obligations for any Chinese AI company with global distribution ambitions.

The interim registration data MOFCOM is now collecting will shape how it enforces the final catalogue. Any Chinese AI company that transferred model weights, fine-tuned checkpoints, or chip design files to a foreign party since October 16 without registering the transaction is technically in breach of the interim rules, regardless of how the final catalogue reads. MOFCOM has announced no enforcement actions yet, but the registration records accumulating since October 16 give the ministry a baseline for identifying which companies and transactions need scrutiny. The first enforcement action under the interim rules, whenever it comes, will tell the market far more about the true scope of the controls than the catalogue text itself.

The next concrete milestone is a MOFCOM final catalogue announcement, expected in the first quarter of 2027. If that announcement retains the 2 trillion parameter threshold and the Senate passes the Remote Access Security Act after Jensen Huang's November 26 testimony, the global AI supply chain will face simultaneously active outbound controls from both the world's two largest AI powers, covering chips flowing into China and model weights flowing out of it, for the first time.

China just drew the same frontier line as Washington, and two of the world's most widely distributed open-weight models are standing on the controlled side of it.


Key Takeaways

  • 23 AI technologies targeted, comment period closed November 15: MOFCOM and MOST's draft now moves to finalization, with a final catalogue expected in the first quarter of 2027
  • 2 trillion parameters is the size trigger: Bloomberg and the AP reported Alibaba's Qwen3.8-Max and Moonshot AI's Kimi K3 are both above this threshold, putting future weight releases in the licensing regime
  • Interim registration already in force since October 16: companies that transferred covered AI technology without registering are technically in breach of the interim rules right now, before the final catalogue is published
  • API access and open-source are explicitly exempt: Chinese labs can continue cloud API services globally, but distributing weights, source code, or architecture requires a license under the draft rules
  • Both China and the US share the same 10^26 FLOP threshold: the first time both countries have maintained simultaneous outbound AI technology controls at the same computational frontier

Questions Worth Asking

  1. If Alibaba already distributed Qwen weights globally before October 16, those copies are outside the controls entirely. Does that make the rules mostly prospective, limiting only future releases while leaving the current state of Chinese AI diffusion unchanged?
  2. The top-10 global ranking trigger means a model's export status can change because a competitor launches a better model, not because the first model changed at all. Who adjudicates disputes over which rankings count, and can a Chinese AI lab legally challenge a reclassification?
  3. Western AI labs that fine-tuned Qwen-based models before October 16 may find that redistributing those derivatives back into China triggers Chinese export jurisdiction on their own work. Have they assessed that exposure?

Sources

  1. China's Ministry of Commerce (MOFCOM) mofcom.gov.cn
  2. Ministry of Science and Technology (MOST) most.gov.cn
  3. Cloud Security Alliance's research note on Announcement No. 48 labs.cloudsecurityalliance.org
  4. The Next Web's reporting on the draft thenextweb.com
  5. Compliance analysis of the thresholds winzheng.com
  6. Analysis published at Shattered.io shattered.io
  7. The Korea Times reported on November 13 koreatimes.co.kr
  8. Atlantic Council research on the Megaspeed case atlanticcouncil.org

Read Next

Google AdMob Adds Agent Skills to Claude Code, Codex, Cursor

Nvidia Rubin HPC Reveals 50,000-Core GPU for Science

Pony.ai Launches Mass-Produced L4 Robotaxi Service in China

Enphase IQ Battery 5P Claims 6,000 Cycles Against All Rivals

Newsletter

Enjoyed this analysis? Join the list for the email edition.

Daily AI signals. No noise. Built for founders, investors, and operators.

Share:XLinkedIn
</> Embed this article

Copy the iframe code below to embed on your site:

<iframe src="https://techfastforward.com/embed/china-advances-ai-export-curbs-covering-qwen-and-kimi" width="480" height="260" frameborder="0" style="border-radius:16px;max-width:100%;" loading="lazy"></iframe>