Zhang Yiming, the founder of ByteDance, has told the company's AI researchers they may not use distillation to improve its models, even if that leaves the TikTok owner behind its Chinese competitors.
Distillation is a training method in which one model learns from the outputs of a larger, more capable model, cutting the time and computing power required to reach a given level of performance.
Zhang delivered the message at a meeting last month, according to The Information, an event given additional weight by the fact that he rarely addresses the Seed team's gatherings.His stated position is that the team can accept being temporarily behind, but should not distil.Building genuine capability, he argued, requires long-termism and delayed gratification rather than using other people's output to achieve short-term leaderboard rankings.
The commercial cost is real. ByteDance acknowledges internally that without the technique its Seed language model will struggle to close the gap with DeepSeek, Moonshot's Kimi and Alibaba's Qwen in the near term.
ByteDance leads globally in AI video models but trails domestic rivals in large language models for coding and agents.
The political context is unavoidable.
Following the meeting, Seed introduced an explicit internal policy prohibiting distillation of open-weight models, and researchers may only distil from ByteDance's own systems.
Distillation has become the flashpoint of the technology dispute between Washington and Beijing, with US officials warning of possible sanctions against companies found to have derived capabilities from American frontier models.
The White House has accused Moonshot AI of large-scale distillation of Anthropic's models.
For a company that spent years fighting a proposed US ban on TikTok, the incentive to stay clear of that fight needs little explanation.
A Chinese account of the episode disputes the political reading, tracing the policy to 2023, when engineers used data generated through OpenAI's application programming interface on a smaller research model and Zhang rejected the compromise outright.
Two problems remain.
Analysts note that it is close to impossible for outsiders to verify whether any company has honoured a no-distillation commitment, and neither Zhang nor ByteDance has explained how the policy will be enforced at engineering level.
And distillation is not the whole story behind Chinese model performance, which also rests on genuine architectural and training innovations.
Whether a frontier model can be built without the shortcut is now, in effect, being tested in public.
