Abstract:
Chinese companies are increasingly using content generated by advanced U.S. artificial intelligence models to train local AI models to cope with the pressure on computing power brought about by the U.S.’s continued tightening of export restrictions on advanced chips. Reuters reviewed more than 300 Chinese academic papers, patents, corporate disclosure documents, recruitment information and government documents and found that this technology, known as "model distillation" (Distillation), has gradually expanded from academic research to practical applications in China's AI industry.

The so-called model distillation refers to using a more powerful AI model as a "teacher model" to answer questions, write code or demonstrate the solution process of complex tasks, and then organize these outputs into training data for training smaller "student models" with lower operating costs. In this way, developers can transfer some of the capabilities already mastered by large models to local models, reducing their continued dependence on expensive computing power and advanced overseas AI services.
Reuters found that Chinese researchers are using U.S. AI models to generate answers to questions, program codes, and solutions to complex problems. Some companies have developed specialized software to automatically collect and process this content, and have professionals further organize training materials. The Chinese government is also promoting distillation technology, hoping to apply AI to manufacturing and other industrial fields at a lower cost.
For Chinese developers constrained by the supply of advanced chips, the appeal of distillation technology is the ability to generate long-term value from an expensive model call. Advanced American models can complete complex tasks first, and developers then save their answers as training samples for local models, without having to re-call overseas models every time they encounter similar problems.
In a paper published in November 2024, researchers from Shanghai Jiao Tong University introduced the use of OpenAI's o1 model to generate training data. The research team screened out about 5,000 pieces of content from the answers of the o1 model, and then translated them through GPT-4o mini. Finally, they sorted out 10,750 bilingual training samples, which were used to further train Alibaba's Qwen2.5-72B model.
The researchers also cautioned that over-reliance on distillation may inhibit original innovation, making developers increasingly dependent on technological breakthroughs made by more advanced model providers. This project also demonstrates how U.S. AI model output can be transformed into training data that can be reused, thereby helping Chinese developers improve local models.
In a paper published in April 2025, Alibaba researchers further introduced a more systematic operation method. They used OpenAI's GPT-4 and GPT-4o to generate training materials in a variety of non-Chinese languages, then screened, optimized and scored the model answers, and then used them to train the smaller Qwen model.
In February 2026, researchers from Shanghai Jiao Tong University and the Chinese food delivery platform Meituan also published a related study. They used GPT-4o to generate task planning solutions for multiple AI agents to collaboratively solve programming problems, and then used these materials to train smaller models so that they can more effectively organize and coordinate multiple agents to complete their work.
Shanghai Jiao Tong University, Alibaba and Meituan did not respond to Reuters' requests for comment.
Model distillation has now become a new friction point in Sino-US scientific and technological relations. In September this year, the U.S. government publicly accused Chinese companies of distilling U.S. AI models in a large-scale and targeted manner. U.S. Department of Commerce officials also raised relevant concerns with Chinese regulators at a meeting held in Chengdu in July this year, warning that certain large-scale collections of advanced U.S. model exports could constitute intellectual property theft.
According to three people familiar with the talks, Chinese officials defended the distillation technology at the time, saying it was a normal AI development method, and said China was willing to take action if the relevant behavior did constitute intellectual property theft. However, a U.S. official familiar with the exchange said China made no specific commitments.
U.S. Treasury Secretary Scott Bessent stated on social media on July 23 that if Chinese companies implement covert, industrial-scale model distillation operations and cross the line of intellectual property theft, the United States will consider imposing sanctions and including relevant companies on the entity list.
Major American AI companies also have corresponding restrictions on the use of model output. OpenAI's terms of service restrict users from using its model output to develop competing products, subject to certain exceptions; Anthropic also restricts the use of its models to train competing systems without permission.
An official familiar with the U.S. government’s July meeting in Chengdu said Washington views industrial-scale model distillation as one of the main reasons why Chinese AI models have narrowed the gap with U.S. rivals in recent years. However, two people familiar with the actual operations of Chinese AI development companies disputed this, believing that distillation is only one factor, and that advances in engineering capabilities, algorithms, data and model design also play a role.
Chinese government documents do not directly require companies to distill American models, but encourage the use of distillation technology as part of promoting the widespread application of AI in the country. In January this year, the artificial intelligence-enabled manufacturing policy document issued by eight government departments in China included guidance to support the application of model training, fine-tuning and distillation technology in the industrial field. Shanghai and Shenzhen are also encouraging companies to adopt technical solutions that reduce computing power requirements.
These policies reflect China’s strategic direction of promoting smaller and cheaper AI systems to various industrial fields. Compared with cutting-edge models that rely on huge computing power to run, distilled lightweight models are easier to deploy in factories, enterprises, and various types of industrial equipment, thereby reducing the cost of AI application.
However, as model distillation gradually involves international intellectual property and technology competition issues, Chinese AI developers have also begun to pay attention to the compliance boundaries of related behaviors. According to a person familiar with the situation, some Chinese AI companies have consulted the Ministry of Industry and Information Technology in recent months, hoping to clarify the boundaries between the normal use of foreign AI models and the organized and large-scale collection of model output for training. China's Ministry of Industry and Information Technology did not respond to Reuters' request for comment.
This controversy comes as the United States further tightens export restrictions on advanced AI chips to China. The U.S. government hopes to slow down China's acquisition of advanced computing capabilities by restricting the supply of high-performance GPUs; while Chinese companies are trying to maximize the efficiency of existing computing power through model distillation, algorithm optimization, and lightweight deployment.
Model distillation cannot completely replicate all the capabilities of leading models, nor can it completely eliminate the need for advanced computing resources, but it can help developers transfer some high-value capabilities to lower-cost models. As this technology moves from laboratories to industrial applications, the actual effects of U.S. chip export controls, the boundaries of intellectual property rights in AI models, and the competition between China and the United States around artificial intelligence technology will all face new challenges.
Related articles:
China's AI model was found to have NVIDIA Blackwell chip "fingerprints". Research estimates that the size of some training clusters reaches 3,000 to 4,500 chips
China accelerates the cultivation of "AI native generation" and aims to popularize artificial intelligence literacy among primary and secondary school students by 2030
Comments