China Proposes Export Controls on 23 AI Technologies: Model Weights Brought Under Regulation

China's Ministry of Commerce and Ministry of Science and Technology released a draft revision of the Catalog of Technologies Prohibited or Restricted from Export for public comment on October 16, proposing to add 23 artificial intelligence technologies to the control list. The draft's core focus is not end-user applications but the underlying technologies that support frontier-scale systems — model weights, source code, and training methods.

Control Boundaries: Dual Thresholds of Compute and Parameters

According to The Decoder, the draft targets AI systems trained with more than 10^26 FLOPs of compute or exceeding two trillion parameters. Covered items include model weights, source code, and training methodologies for such systems. Companies seeking to export related technologies must complete registration before export.

The threshold design is noteworthy. 10^26 FLOPs and two trillion parameters roughly correspond to the current training scale of the most advanced closed-source models, meaning the controls target "next-generation" capabilities rather than mid-sized models already in wide circulation. Some analysts note this converges with the compute-threshold logic used in US and EU regulatory frameworks — drawing regulatory boundaries with quantifiable technical metrics rather than vague capability descriptions.

For developers, the immediate impact may be an added layer of compliance review in cross-border collaboration: providing model weights to overseas partners, sharing training code, or even transferring training recipes could trigger registration obligations.

FLOPs (floating-point operations) is a common unit for measuring the computational cost of neural network training. Each multiply-add operation in a forward or backward pass typically counts as about 2 FLOPs, so a model's total training compute is roughly equal to "parameter count × training tokens × a constant." The 10^26 FLOPs scale roughly corresponds to the training cost of models with hundreds of billions to a trillion parameters on trillions of tokens — the same order of magnitude as current leading closed-source models. This is precisely why compute thresholds are easier to verify than subjective descriptions like "model capability": regulators can infer whether a threshold is crossed from training logs, cluster size, or public disclosures, and companies have relatively objective grounds for their filings.

Model weights are the parameter matrices saved after a neural network is trained — essentially a set of floating-point files, commonly in formats such as safetensors or PyTorch's .pt/.bin. Weights determine a model's inference behavior but do not include training data or training code; by contrast, training methodology refers to "recipe" information such as data mixing ratios, learning rate schedules, parallelism strategies, and alignment pipelines. That all three are controlled side by side means regulation covers the full chain from "finished product" to "production process," not just the final deliverable model files.

Open-Source Fast Track: Balancing Controls and Ecosystem

The most telling detail in the draft is the fast-track registration channel established for open-source releases.

China's open-weight ecosystem has considerable global influence — open models from teams like DeepSeek^Ad, Qwen, and Kimi have been downloaded and built upon by a large number of international developers. If control rules imposed the same review burden on open-source releases as on commercial exports, the competitiveness of this ecosystem would take a direct hit. The fast-track design indicates that regulators want visibility and control over frontier technology exports without cutting off the institutional advantages of the open-source route.

Some analysts interpret this dual-track design as a compromise between national security regulation and open-source competitive strategy. But other observers caution that the specific meaning of "fast" — whether it means streamlined procedures or a review exemption — remains unclear in the draft text, and the actual implementation standard will be key.

Open-weight models are models whose parameters are publicly released and that others can download and deploy locally. They are not fully equivalent to "open-source software" — training data and complete training code are often not disclosed, so they are commonly called "open-weight" rather than strictly open source. Models released by Chinese teams such as DeepSeek, Qwen, and Kimi mostly fall into this category, and their licenses typically permit commercial use and redistribution, which is why they have spread rapidly in the international developer community.

It is important to distinguish that "export" in the export-control context is not limited to physical cross-border movement: providing downloadable weight files to overseas entities, granting overseas API access, or hosting models abroad may all be deemed technology exports. This is why the fast-track registration channel has substantive significance — if every weight release required full review, the release cadence of the open-weight route would be significantly slowed.

A New Compliance Layer in the Global Supply Chain

If the draft is ultimately finalized, the impact will propagate along several paths.

Cross-border model distribution will be hit first. International developers who rely on Chinese open-weight models may need to add compliance confirmation at the download, deployment, or redistribution stage. Joint research and infrastructure cooperation also face new variables — for multinational research projects involving frontier-scale models, technology transfer registration obligations could change the pace of collaboration.

The comment period runs through November 15, leaving industry a window for feedback. Observers are watching two focal points: whether the thresholds will be adjusted in the final version, and how the registration system will be enforced in practice. The combination of these two variables will determine the actual flow rate of frontier AI technology out of China.

Continuity in Regulatory Thinking

Listing model weights themselves as controlled items is unusual in export-control practice. Traditional control lists mostly target hardware, equipment, or clearly definable technical solutions, whereas weights are the product of a training process, and their "export" forms can be file downloads, API calls, or model hosting. This ambiguity in the regulatory object may pose enforcement challenges, and it also reflects regulators' efforts to keep up with the speed of AI technology diffusion.

From a broader perspective, the draft continues a trend: major economies are all using compute thresholds as a common yardstick for AI regulation. When technological capability can be defined by quantifiable thresholds, the granularity of regulation depends on the precision of threshold setting. For Chinese AI companies, rising compliance costs are a certainty; whether the open-source channel can truly remain open is the key variable determining the policy's real-world impact.

Background

Compute-threshold regulation is a common tool that major economies have developed in AI governance in recent years. A 2023 US executive order used 10^26 FLOPs as the training threshold for large models requiring government reporting, while the EU AI Act sets a presumption standard of 10^25 FLOPs for general-purpose models and distinguishes additional obligations accordingly. The shared logic of such designs is that capability is hard to measure in advance while compute can be audited, so computational quantity serves as a proxy indicator. Its limitations are equally clear — advances such as distillation, sparsification, and inference-side scaling may enable low-compute models to approach the capabilities of high-compute models, and thresholds need periodic calibration to remain effective.

One-line Summary

China proposes export controls on 23 frontier AI technologies, setting dual compute and parameter thresholds while leaving a fast track for open source.

Paragraph Summary

On October 16, China's Ministry of Commerce and Ministry of Science and Technology opened public comment on a draft revision of the Catalog of Technologies Prohibited or Restricted from Export, proposing to add 23 AI technologies to the control list. The core targets are model weights, source code, and training methods rather than end-user applications. The control threshold is set at training compute exceeding 10^26 FLOPs or more than two trillion parameters — roughly matching the training scale of the most advanced closed-source models and converging with the US and EU approach of drawing regulatory boundaries via compute thresholds. The draft also establishes a fast-track registration channel for open-source releases, seen as a compromise between national security regulation and maintaining the competitiveness of the open-weight ecosystem. If finalized, international developers relying on Chinese open-weight models and multinational joint research will face a new compliance layer. The comment period runs through November 15, with threshold adjustments and enforcement standards as key variables going forward.