China’s AI Export-Control Draft Draws a Gate Around the Next Frontier Model

# China's AI Export-Control Draft Draws a Gate Around the Next Frontier Model

*Sophia Chen — October 20, 2026*

China is not closing the door on global access to its artificial-intelligence models. It is installing a gate around what comes next.

A draft revision issued by the Ministry of Commerce (MOFCOM) and the Ministry of Science and Technology on October 16↗ would bring frontier-model intellectual property — including weights, source code and training methods — into China's technology export-control framework. Materials already lawfully released through open-source licenses or academic publication are carved out, while access through an API, website, app or software-as-a-service platform is not treated as a technology export.

Announcement No. 48↗ is therefore not a finalized "AI export ban." Its rules remain subject to consultation through November 15, but it could require authorization when Chinese-developed frontier technology moves abroad while preserving international consumption as a service.

The compliance burden is already real. An interim registration requirement took effect on October 16 for new and unfinished covered transactions, reaching overseas licensing, transfers of technology ownership or intellectual-property rights, and changes placing effective control of a Chinese technology entity in foreign hands. For developers, that means continuity for public releases and hosted use, but potentially more friction around future downloadable frontier weights, offshore deployment and corporate deals.

Forty Technologies, With AI at the Center

The proposed revision covers 40 technologies: 23 in AI, 12 in photovoltaics and five in biomedicine. Fourteen are proposed for prohibition and 26 for restriction through licensing.

The AI entries extend beyond large language models to agents, computing clusters, information-processing systems, advanced chip-design methods and associated infrastructure. For frontier models, controls can reach trained weights, source code and training methods — not merely finished consumer applications.

The Three Capability Triggers

A model enters the restricted category if it meets any one of three thresholds:

• Scale threshold: 2 trillion parameters or more — a bar that currently sits above every publicly released Chinese model, including the 2.8 trillion-parameter Kimi K3, though the exact interpretation of "exceeding" versus "at least" requires clarification in the final text.
• Compute threshold: Training compute of 10^26 floating-point operations or more, roughly ten times the estimated compute used to train GPT-4.
• Ranking threshold: A global top-10 position in specified public evaluations of performance, safety or efficiency — a deliberately flexible criterion that could capture models well below the parameter ceiling.

Data adds another layer. Large-model technologies trained on "core data" under Chinese law may be prohibited from export; "important data" can put them in the restricted category. Similar models could therefore receive different treatment because of their training corpus rather than architecture or benchmark results.

Infrastructure and Agent Provisions

The draft's reach extends well beyond model weights:

• Computing clusters: Installations with 10,000 or more AI accelerator cards and interconnects of at least 800 GB per second fall under the catalogue.
• Chip design: Technologies at or below the 14-nanometer node and advanced 2.5D or 3D packaging are included.
• AI agents: Systems capable of decomposing work into chains of 20 or more steps, coordinating at least 60 agents, or calling 200 or more tools at high accuracy face potential controls — with stricter treatment for systems targeting national defense or critical infrastructure.

The framing is systemic: Beijing is asking not only how large a model is, but what data produced it, what it can autonomously execute, what infrastructure sustains it and whether its capability can be replicated abroad.

The Line Is Between Technology and Service

The draft's central design choice is the line between transferring technology and providing product access.

A foreign developer calling a Chinese model through an API does not receive its weights, source code or training process. Neither does a user of a website, mobile app or SaaS product. Those channels are not classified as technology exports under the reported framework, preserving international inference sales while keeping sensitive assets under the provider's control.

Model weights, code and academic materials already lawfully public are also excluded. The draft neither claws back assets circulating through open-source releases or scholarly publication nor turns existing downloads into prohibited technology.

Future releases differ. A Chinese lab publishing a newly covered frontier model as open source — or, often more precisely, open weight — would enter an automatic or simplified "report-and-export" licensing pathway. A similar mechanism is contemplated for a Chinese company deploying covered technology on its own overseas cloud; international research and standards work may also receive facilitated treatment.

"Automatic" does not mean "unregulated." Reporting gives the state visibility, and licensing creates a point for conditions. Yet a fast track suggests a gate rather than necessarily the end of Chinese model openness.

An open-weight release makes trained files available for download and self-hosting without necessarily disclosing the full training data, code or process needed for reproduction. Fully open-source technology is broader. Because the catalogue explicitly reaches weights, source code and training methods, it could regulate different layers of one release separately.

Registration Starts Before the Consultation Ends

Although the catalogue remains a proposal, MOFCOM's interim registration requirement↗ began on October 16; the November 15 consultation deadline does not postpone all compliance.

Registration applies to new or unfinished covered outbound technology transactions, including:

• Transfers of ownership or intellectual-property rights to an overseas entity.
• Licensing domestic technology for use abroad.
• Ownership or control changes that place a Chinese technology holder under foreign control.

Developers, investors, cloud operators and multinational partners must now determine whether a transaction touches listed technology, whether the asset is lawfully public, and whether the arrangement provides a service or transfers technology.

Definitions, benchmark selection, threshold treatment and licensing procedures may change after consultation. Companies should treat registration as an operative interim obligation without describing the draft catalogue as settled law.

A Symmetric Answer, But Not a Hardware Embargo

Geopolitically, the proposal is China's reciprocal answer to years of US export controls centered on advanced chips and computing capacity. Both treat frontier AI as strategic, but intervene at different points in the stack.

US-style restrictions focus heavily on hardware and the compute needed to build or run advanced systems. China's draft regulates outbound dissemination of Chinese-developed frontier-model intellectual property and know-how; it is not a US-style hardware embargo against foreign recipients.

The distinction reflects China's changing position. Once domestic laboratories hold models global developers want to download, self-host or adapt, Beijing has a new asset to control: trained capability. The draft creates a mechanism for deciding when it may leave Chinese jurisdiction in reproducible form.

Meta released Muse Spark Open↗ on October 15, one day before the draft. The open-weight mixture-of-experts model has 816 billion total parameters and 42 billion active per token, and its license carries commercial conditions — another reason not to equate "open weight" with fully open source. There is no evidence Meta's launch caused the proposal. Their proximity instead shows how model weights have become instruments of ecosystem competition as well as technical artifacts.

What DeepSeek, Qwen, GLM and Kimi Users Should Plan For

For teams using public releases from DeepSeek, Qwen, GLM or Kimi, the draft does not imply an immediate shutdown. Existing lawfully public materials retain the carve-out, while hosted API, web, app and SaaS access remains outside the definition of a technology export.

Risk rises with future frontier releases. Models crossing scale, compute, ranking or data-sensitivity triggers may need reporting and licensing before their weights go abroad. Timing could become less predictable, and conditions could differ between domestic and international availability.

Practical Steps for Developers and Enterprises

• Inventory your artifacts: Distinguish between models accessed via API (unaffected) and downloaded weights (potentially covered for future releases). Pin specific model versions using SHA-256 commit hashes to ensure consistency.
• Assess offshore deployment: A foreign company receiving weights for local operation is closer to a technology transfer than a customer sending prompts to a provider-controlled API. Even a Chinese company deploying into its own overseas cloud may need the simplified pathway.
• Review cross-border agreements: Agreements conveying overseas usage rights, intellectual property or effective foreign control can trigger registration without a public model launch. Cross-border acquisitions and investments may require technical classification alongside ownership review.
• Plan for reproducibility gaps: If weights remain downloadable while training methods, source code or implementation details receive separate treatment, developers may run a model without being able to reconstruct it — widening the gap between open-weight distribution and fully open development.
• Design for substitution: Maintain fallback API providers and evaluate Western open-weight alternatives such as Meta's Muse Spark Open as part of standard operational hygiene.

Sidebar: Tencent's Hy Memory 2.0 Shows Product Velocity Has Not Stopped

On October 16 — the same day the draft landed — Tencent Hunyuan released Hy Memory 2.0↗ under the Apache 2.0 license. The stack packages three components: the Hy-Memory-30B-A3B model↗, the `hy-memory` Python SDK, and MemHub.

Hy-Memory-30B-A3B is a specialized mixture-of-experts memory writer with 30 billion total parameters and roughly 3 billion active per token. Rather than serving primarily as a general chatbot, it decides what an agent should retain, how stored information should be represented and how new evidence should update existing memory. The model is available as a 60 GB BF16 checkpoint on Hugging Face, with an FP8-quantized variant for efficient vLLM deployment.

The Hy-Memory GitHub repository↗ includes the full SDK, which supports three memory scenes:

• Chat memory: Hierarchical storage for long-term user preferences and narrative facts across extended conversations.
• Working memory: Preserves goals, intermediate results and reusable experience for task-oriented productivity workflows.
• Coding memory: Tracks implementation history and version evolution as a diff-chain, enabling agents to reason about software development over time.

Tencent's vendor-reported benchmarks place Hy-Memory-30B-A3B at a macro-average score of 55.49 across four memory frameworks (Mem0, MemOS, EverMemOS and Hy-Memory), compared to 54.62 for Claude Opus 5 and 53.64 for DeepSeek V4.1 Flash. These figures use fixed 20% benchmark subsamples and have not been independently verified.

The release is a useful counterweight to simplistic interpretations of the draft. Chinese labs are still shipping permissively licensed model components and developer infrastructure while the regulatory boundary is being debated.

Beijing's Optionality Over the Next Release

The global open-weight impact depends less on models already available than on the next frontier release that is not.

By protecting public assets and excluding service access, the draft avoids an abrupt break. Chinese providers can keep serving overseas customers, users can keep working with lawfully released weights, and Tencent's same-day Hy Memory 2.0 launch shows open technical distribution has not stopped.

Yet the proposal gives Beijing advance visibility into future frontier releases, offshore deployments and ownership transfers, plus a licensing mechanism that can tighten or loosen as strategic conditions change.

That optionality may shape developer choices even if approvals are routine: uncertain timing and compliance make ecosystems harder to plan around, while predictable automatic licensing could preserve practical access under formal oversight.

The right interpretation is neither "nothing changes" nor "China has banned open models." Existing releases and hosted services retain meaningful continuity; future frontier weights and direct transfers face a sovereign checkpoint.

For global developers, architecture, benchmark performance and license text are no longer enough. The ability to move, reproduce, control and deploy a model across borders is becoming part of the product specification.