Demystifying Byte Seedance: China’s first large-scale model money printing machine started

Interview | Anita Deng, Xiao Sijia, Li Xiaoxia Text | Zhang Yuxin, Anita Deng Editor | Zhang Yuxin, Yang Xuan At the end of 2025, at a model team dinner, Zeng Yan, the person in charge of the Seedance model, mentioned to the leader that she still wanted to try training a larger size model, at least reaching 200B (200 billion parameters). Zeng Yan is a very young researcher at Byte Seed. She will enroll in the school in 2021. Her research direction has always focused on video understanding and generation. "She has always had her own technical judgment." A person who has dealt with Zeng Yan told 36 Krypton, "She is also very proactive, persistent in what she believes in, and will find ways to obtain resources to achieve it." This evaluation almost replicates Zeng Yan's situation when she was doing Seedance2.0 training. "She wanted to train a model with a larger number of parameters, but there were some differences within the team at the time. Some people felt that directly scaling up the model to the 200B-300B level was a bit aggressive, and the overall training resources were also tight. It might be safer to train a 100B level model first." A person familiar with the matter told 36 Krypton. The above-mentioned person said that senior executives happened to attend the above-mentioned dinner party. After learning about Zeng Yan’s idea, they said that they could coordinate some resources for the training of Seedance 2.0. “Later, at the final formal review meeting within Seed, Seed head Wu Yonghui and Zhou Chang, the person in charge of visual multi-modal generation, finally chose to support Zeng Yan’s idea.” This seemingly radical technical choice at the time was eventually proven to be correct. “Because Zeng Yan insisted that the model must be large enough and the training data must be rich enough, Seedance 2.0 became a reality,” the person familiar with the matter said. We interviewed AI practitioners inside and outside Doubao. Most people believe that Byte’s real reputation reversal in large models began with Seedance 2.0: in terms of large language models, Doubao’s large model will not be iterated to 1.6 until mid-2025.