Zhang Yiming: ByteDance does not take the shortcut of model distillation, preferring to let large models lag behind for now

ByteDance founder Zhang Yiming clearly stated at the Seed team all-hands meeting last month that the company will not use model distillation as a shortcut to catch up with large models, even if it temporarily falls behind domestic competitors. He urged the team to be willing to sacrifice short-term gains for long-term goals.

Distillation refers to allowing a new model to learn from the outputs of a stronger model, which can save training time and computational power. Internally, ByteDance believes that this choice will make it more difficult for Seed's language model to catch up with domestic peers like DeepSeek, Kimi, and Qianwen in the short term. However, ByteDance is more concerned about political risks: the company has long faced scrutiny from the U.S. government due to TikTok, and if it is accused of massively extracting the capabilities of cutting-edge U.S. models, Washington may apply pressure on TikTok again. Previously, Anthropic accused DeepSeek, Dark Side of the Moon, MiniMax, Zhiyu, and Alibaba of massively extracting Claude's capabilities, but did not name ByteDance.