字節跳動創始人張一鳴在7月召開的Seed團隊全員大會上表示，公司不會採用模型蒸餾作為提升AI能力的捷徑[1]。張一鳴強調應該「願意為長期目標犧牲一部分短期收益」，並指出「做模型要堅持長期主義、延遲滿足感，而不是用別人的輸出，換一時的榜單排名」[1]。
在中國AI大模型競爭日趨激烈、開源模型快速發展的背景下，張一鳴認為蒸餾會干擾真正意義上的長期技術突破[1]。模型蒸餾是指利用複雜、大型模型的輸出，來訓練或改進較小模型的技術手段[1]。基於這一認識，字節跳動內部對開源模型實施了嚴禁蒸餾的政策[1]。
ByteDance founder Zhang Yiming stated at an internal company meeting that the organization will not adopt model distillation as a shortcut to advancing its AI capabilities [1]. Speaking at the ByteDance Seed team all-hands meeting held in July, Zhang emphasized the importance of long-term commitment over short-term rankings [1].
Zhang argued that model distillation—the practice of using outputs from complex, large-scale models to train or improve smaller models—would interfere with genuine technological breakthroughs [1]. He expressed the position that "one should be willing to sacrifice some short-term gains for long-term objectives" and stressed that "model development should adhere to long-termism and delayed gratification, rather than borrowing others' outputs for temporary ranking improvements" [1]. To enforce this principle, ByteDance has instituted an internal ban on distilling open-source models [1].