Chinese companies and research institutions are reportedly making increasing use of leading US artificial intelligence (AI) models to train home-grown models through distillation, reducing development costs and reliance on advanced computing resources as the US continues to tighten export restrictions on advanced chips.

A Reuters review of more than 300 Chinese academic papers, patents, corporate disclosures, recruitment notices and government documents found that model distillation has gradually expanded from academic research to widespread use across Chinese industry. Reuters published its in-depth report on Tuesday (17 November).

The documents show that Chinese researchers use US AI models to generate answers to questions, write code and solve complex problems, then turn that content into training data to improve the capabilities of smaller home-grown AI models. Some Chinese companies have also developed software to automate the process.

Model distillation involves a more capable “teacher model” generating answers or demonstrating how to complete tasks. Those results are then used to train a smaller “student model”, allowing it to acquire some of those capabilities at a lower operating cost.

The technique itself is not illegal and is widely used elsewhere in the world. Although distillation cannot fully replicate the capabilities of advanced models or eliminate the need for advanced computing resources, it helps developers build and run models with specific capabilities at a lower cost. This reduces reliance on US technology and could weaken the effectiveness of US chip export controls on China.

For example, in a paper published in November 2024, researchers at Shanghai Jiao Tong University described selecting about 5,000 responses generated by US company OpenAI’s o1 model, then translating them using GPT-4o mini. They ultimately compiled 10,750 bilingual training examples to further train Alibaba’s Qwen2.5-72B model.

Alibaba researchers also published a paper in April last year describing how they used OpenAI’s GPT-4 and GPT-4o to generate multilingual training data, which were then filtered and scored to train smaller Qwen models. In February this year, researchers from Shanghai Jiao Tong University and Meituan also described using GPT-4o to generate plans for multiple AI agents to collaborate on coding, for use in training home-grown models.

As distillation becomes increasingly widespread in China, the US government has begun examining the associated intellectual property issues. Three people familiar with the matter said that, at a meeting in Chengdu in July this year, US Commerce Department officials raised concerns with Chinese regulators about Chinese developers collecting large volumes of output from leading US AI models. They warned that some practices could constitute intellectual property theft.

A US official familiar with the talks said the Chinese side stressed that distillation was a routine technique in AI development and expressed willingness to act where such practices constituted intellectual property theft, but made no specific commitments.

The US official also said Washington believed large-scale model distillation was the main reason Chinese AI models had narrowed the gap with their US rivals in recent years.

However, two people familiar with the operations of Chinese AI developers disagreed, saying distillation was only one factor driving advances in Chinese AI models. Progress in engineering, algorithms, data and model design had played equally important roles, they said.

According to a person familiar with the matter, as distillation has become more geopolitically sensitive, Chinese AI developers have sought guidance from the authorities overseeing industry in recent months. They hope to clarify the boundary between normal use of foreign models and the organised, large-scale collection of model output for training.