China’s Ministry of Commerce (Mofcom) issued Announcement No. 48 on September 28, 2024, imposing the country’s first formal export controls on frontier artificial intelligence models — a move flagged in a Hacker News discussion that garnered 65 points and 40 comments for its technical and geopolitical significance.
Regulation: Mofcom Announcement No. 48 | Effective Date: September 28, 2024 | Scope: AI Models exceeding defined performance thresholds
License: Required for export | Key Metric: >48% score on key benchmarks (MMLU, MATH, GPQA, IFEval)
The Technical Thresholds: What Qualifies as "Controlled"
The controls apply to AI models whose performance exceeds specific benchmarks. A model is subject to licensing requirements if it meets any one of the following four criteria:
| Benchmark | Threshold for Control |
|---|---|
| MMLU (5-shot) | Score > 48% |
| MATH (4-shot) | Score > 12.5% |
| GPQA Diamond (0-shot) | Score > 37.5% |
| IFEval (Instruction Following) | Score > 75% |
These benchmarks target advanced reasoning, mathematical problem-solving, specialized knowledge, and instruction-following capabilities. The MMLU threshold of 48% is a critical line — it captures many current frontier models while exempting older or less capable ones.
Which Models and Companies Are Affected
The regulation directly impacts Chinese AI labs developing state-of-the-art models. Based on published benchmarks, models like DeepSeek-V2, Qwen2.5-72B, and GLM-4-9B likely exceed these thresholds. The controls cover the export of the model weights, associated algorithms, and training techniques.
The rule creates an immediate compliance burden for companies like Alibaba’s Qwen, Baidu’s ERNIE, 01.AI, and Zhipu AI when sharing technology or providing API access to foreign entities. The Cloud Security Alliance research note confirms the controls are retroactive, applying to models already developed that meet the criteria.
Bottom line: Any Chinese-origin model scoring above ~48% on MMLU now requires a government license for international release.
How This Compares to U.S. AI Export Controls
China’s move establishes a parallel regulatory framework to U.S. controls, but with distinct mechanisms.
| Feature | China (Mofcom No. 48) | United States (BIS Rules) |
|---|---|---|
| Basis for Control | Model performance on specific benchmarks | Compute Flops Threshold (≥10^26 FLOPs for training) |
| Primary Target | Export of model weights and algorithms | Export of advanced chips (GPUs) and chip-making equipment |
| Enforcement Point | Licensing at the software/model level | Licensing at the hardware/manufacturing level |
| Key Metric | MMLU >48%, MATH >12.5% | Total Processing Performance (TPP) and Performance Density |
The U.S. strategy focuses on strangling compute supply, while China’s new rules aim to control the output models directly. This creates a bifurcated regulatory landscape where both the hardware to train top models and the models themselves are subject to separate national controls.
Practical Implications for Developers and Companies
For international teams, accessing the latest Chinese model weights via Hugging Face or GitHub may require verification of licensing. API access from Chinese providers to users outside China could be interrupted or gated.
Open-source collaborations face new friction. A Chinese contributor pushing a model that crosses the threshold to a global repository may violate export law. This could lead to the development of region-specific model forks or delayed international releases.
Early commentary on the Hacker News thread highlighted concerns about the fragmentation of the open-source AI ecosystem and the potential for a "splinternet" of AI, where model development forks along geopolitical lines.
"Full Text of the Key Benchmarks"
The official announcement specifies the exact benchmark datasets and versions:
Who Should Pay Attention and Next Steps
AI researchers and developers relying on cutting-edge Chinese models need to audit their supply chain. Monitor official channels from providers like Qwen or DeepSeek for compliance statements.
Enterprise users with contracts for Chinese model APIs should contact providers to confirm export licenses are in place to ensure service continuity.
Global AI labs should view this as a precedent. Other nations may adopt similar performance-based export rules, making benchmark scores a legal compliance metric, not just a research yardstick.
The immediate next step is to check your model's benchmarks. If you are using or developing a model of Chinese origin, verify its scores against the MMLU, MATH, GPQA, and IFEval thresholds. For ongoing projects, consider diversifying model sources or preparing contingency plans for access restrictions.
This is not merely a trade policy—it’s a redefinition of AI as a sovereign technology asset. The era of globally frictionless model distribution is over; the new era requires navigating a patchwork of national performance thresholds.
Top comments (0)