ByteDance founder Zhang Yiming (장이밍) has made clear that he will not use “distillation” techniques to train the company’s artificial intelligence models.

Cryptopolitan, a blockchain media outlet, reported on Aug. 5 that Zhang has recently been directly involved in ByteDance’s AI strategy, meeting senior researchers to review development progress and also keeping track of hiring.

The stance comes as the United States zeroes in on Chinese AI labs’ use of distillation. Distillation is a method of training a relatively smaller model based on responses from a larger model with higher performance. The technique is commonly used openly in the industry and is often legally permitted, but it has generally been used on a limited scale.

Zhang drew a line at the practice. The outlet reported that Zhang strongly opposes distillation in training ByteDance’s AI models. With parts of China’s AI industry having faced suspicions of using the technique to reduce cost and time burdens, ByteDance has effectively signaled it will choose a different path.

ByteDance’s choice is also linked to the competitiveness of its video generation model, “Seedance.” ByteDance is positioning Seedance as one of the rare Chinese AI products that combines profitability and performance. Seedance 2.0 faced internal disagreement over cost burdens and excessive scalability, but development moved ahead with a target of more than 200 billion parameters. Some staff at the time viewed that scale as overly aggressive.

ByteDance instead concentrated funding on large-scale investments in graphics processing units, data 확보 and recruiting advanced talent. As a result, it was presented that the model’s gross margin stayed at 70 to 90 percent. This is also seen as the backdrop to Zhang choosing a costly in-house development route over a shortcut.

The United States has recently responded more sensitively to Chinese AI companies’ use of distillation. Michael Kratsios (마이클 크라치오스), who leads the White House Office of Science and Technology Policy, claimed that Moonshot AI may have distilled Anthropic’s “Fable” model while building its AI system, “Kimi K3.” He wrote that Moonshot AI operated an internal platform that changed access routes to evade detection and used Nvidia GB300 servers in Thailand that are banned from export to China under U.S. export rules.

Anthropic is said to have tracked 3.4 million Claude conversations linked to Moonshot AI through fake accounts. Moonshot AI has not yet responded to the allegations.

The U.S. government response is also extending into regulatory discussions. Treasury Secretary Scott Bessent said the United States could consider sanctions if Chinese companies adopt distillation at an industrial scale. In the U.S. House of Representatives, a bipartisan bill is also being pursued to make distillation attempts easier to identify and sanction.

In the United States, a sense of crisis is also growing that the AI gap with China is narrowing quickly. Data from Stanford University’s Institute for Human-Centered AI showed the top models’ performance in the two countries was presented as effectively similar. This is the backdrop to observations that the United States is failing to maintain an edge despite spending 23 times more than China on AI investment.

Against this backdrop, ByteDance’s choice goes beyond a dispute over development methods and connects to the broader U.S.-China AI competition. The next focus is what training methods and infrastructure strategies Chinese companies choose under U.S. regulatory pressure, and whether ByteDance will continue its in-house development stance led by Seedance.