Chinese AI Firms Have Cracked The Code For Training LLMs While Dealing With U.S. Chip Tightening Measures; Using OpenAI’s Models For “Industrial-Scale Distillation”

Wccftech

Source published: Site published:

Original article · فارسی

As the U.S. keeps making it harder to export advanced computer chips to China, Chinese AI developers and researchers are finding new ways to work around these trade limits. A recent review of more than 300 Chinese research papers, patents, company reports, job postings, and government documents shows that Chinese companies are using a method called model distillation more often. This method uses answers from powerful American AI models to train smaller Chinese AI models.

Model distillation is no longer just a research experiment. Chinese companies are already using it on a large scale.

Model distillation works by using a powerful AI model called a "teacher" to create answers, computer code, or steps for solving problems. A smaller AI model called a "student" then learns from those results. This is a cheaper option for Chinese AI companies that have trouble getting powerful computer chips, such as NVIDIA GPUs. Instead of using expensive hardware to train their own models from the beginning, they can ask an advanced foreign AI model a question and reuse its answer as training material. This is especially useful now that NVIDIA servers containing eight H200 GPUs reportedly cost $900,000 on China's unofficial gray market because of stricter trade restrictions from Beijing.

Reuters looked at how major Chinese universities and technology companies are using this method in practice. For example, in November 2024, researchers at Shanghai Jiao Tong University collected about 5,000 answers from OpenAI's o1 model. They used GPT-4o mini to translate the answers and created 10,750 training examples in two languages. They then used these examples to further train Alibaba's Qwen2.5-72B AI model. However, the researchers warned that relying too much on distillation could make it harder to develop truly new AI technology.

There are other examples. In 2025, Alibaba researchers created training data in languages other than Chinese. They used OpenAI's GPT-4 and GPT-4o models to rate the answers and choose which ones were good enough to train smaller Qwen models. Later, in February 2026, researchers from Shanghai Jiao Tong University and the Chinese delivery company Meituan used GPT-4o to create plans showing how several AI agents could work together on computer programming tasks. They used these plans to train a smaller Chinese AI model. Distillation itself is not illegal, and researchers and companies around the world use it.

However, AI companies such as OpenAI and Anthropic have rules in their terms of service that do not allow people to use their models to train competing AI systems. This could create more disagreements between the U.S. and China. During meetings in Chengdu in July that had not been reported before, officials from the U.S. Commerce Department warned Chinese regulators that collecting answers from American AI models could count as stealing intellectual property, meaning ideas or technology that belong to someone else.

A U.S. official who knew about the Chengdu meetings said Washington believes that "industrial-scale distillation," or using this training method on a very large scale, is the main reason Chinese AI models are catching up with American models in performance. However, even with tighter restrictions on exports of NVIDIA chips, the U.S. appears to have few ways to slow China's AI progress. Because distillation is not currently considered illegal, Washington might create new laws to stop China from using AI models made by OpenAI and Anthropic for this purpose.

From reportedly smuggling NVIDIA's powerful B300 GPUs into China to using model distillation, Beijing appears to keep finding ways around U.S. efforts to slow the growth of its AI industry.

News Source: Reuters