ByteDance founder Zhang Yiming has made it clear: the company won't use model distillation as a shortcut to catch up in the LLM race — even if that means falling behind domestic rivals for now. Speaking at a Seed team all-hands meeting last month, Zhang asked his team to accept short-term losses in exchange for long-term goals.
Distillation is the practice of training a new model to mimic the outputs of a stronger model, saving training time and compute. ByteDance insiders admit that skipping it will make it harder for Seed's language models to catch up with Chinese peers like DeepSeek, Kimi, and Alibaba's Qwen in the near term.
But ByteDance's bigger concern is political risk. The company has been under US government scrutiny for years over TikTok. Even after TikTok's US data security business brought in American investors, ByteDance still holds a stake and continues to run TikTok's global commercial operations. If the company were accused of mass-extracting capabilities from US frontier models, Washington might squeeze TikTok again.
Anthropic has previously accused DeepSeek, Moonshot AI, MiniMax, Zhipu, and Alibaba of broadly extracting Claude's capabilities — though ByteDance wasn't named.