ByteDance founder Zhang Yiming says company won’t rely on AI distillation

ByteDance founder Zhang Yiming has reportedly instructed the company not to rely on AI distillation to improve its models, even if that slows its progress compared with competitors. According to The Information, the decision reflects ByteDance’s effort to build AI systems through original research rather than training on the outputs of rival frontier models. The move is also believed to be influenced by the company’s sensitive relationship with U.S. regulators over TikTok ownership, as ByteDance seeks to avoid additional scrutiny surrounding its AI development practices.

AI distillation has become one of the most controversial topics in the global AI race. The technique involves using responses generated by powerful frontier models to train smaller or competing systems. Anthropic has previously accused several Chinese AI companies—including DeepSeek, Moonshot, MiniMax, Z.ai, and Alibaba—of distilling its Claude models. ByteDance, however, has not been publicly accused of engaging in the practice, making Zhang’s reported directive a notable distinction within China’s AI industry.

Industry analysts caution that such a commitment is difficult to verify externally. ByteDance has not explained how the policy will be enforced internally, whether it applies to synthetic data generated by its own AI models, or how much performance the company is prepared to sacrifice by avoiding distillation. As competition intensifies, companies face increasing pressure to release more capable models at lower costs, making original model development significantly more expensive and time-consuming.

Zhang’s comments are particularly noteworthy because he has largely remained out of the public spotlight in recent years. However, recent reports indicate he has become increasingly involved in ByteDance’s AI efforts, participating in technical reviews of the company's Seed AI team and closely following research from leading U.S. AI labs such as OpenAI.

The announcement comes amid rapid advances across China’s AI sector. Alibaba recently introduced Qwen3.8-Max, its largest AI model to date, while DeepSeek continues to attract attention with highly competitive pricing that undercuts many Western rivals. Moonshot’s Kimi K3, a 2.8-trillion-parameter open-weight model, has also emerged as a major competitor in software engineering and autonomous AI tasks, highlighting the growing pace of innovation among Chinese AI developers.

At the same time, Moonshot AI has faced allegations from U.S. officials that it distilled Anthropic’s AI models while developing Kimi K3. White House Office of Science and Technology Policy Director Michael Kratsios claimed the company used specialized infrastructure and overseas computing resources to avoid export restrictions, although those allegations have not been independently verified.

Despite distancing itself from AI distillation, ByteDance continues to invest heavily in original AI research and product development. Earlier this year, the company launched Dreamina Seedance 2.0, a multimodal AI model integrated into CapCut that enables creators to generate and edit synchronized audio and video content using text prompts, images, and reference videos. Zhang’s reported strategy suggests ByteDance is prioritizing long-term credibility and technological independence over short-term competitive gains in the rapidly evolving AI landscape.