ByteDance restructures seed department to focus on model development and agent capabilities

According to media reports, ByteDance's Seed department, which oversees large model development, has completed a significant organizational restructuring. The department has established four new first-level units, all reporting to Wu Yonghui. This move aims to consolidate resources across pre-training data, reinforcement learning, and post-training processes, while also separating model capabilities for business-to-end (B-end) and consumer-to-end (C-end) products. The Pretrain Data team, led by Li Chenggang, integrates previously scattered teams responsible for text, programming, visual understanding, and speech data. This team will focus on multimodal data and ultra-large model pre-training for the new Omni model. Horizon RL, under Tang Shengyu, combines post-training, reasoning, and visual understanding teams to advance reinforcement learning and enhance model intelligence. Product Posttrain-Work, led by Qin Yujia, targets B-end applications, handling the integration and release of Agentic models, with an emphasis on optimizing these models for office scenarios. This includes support for task modes in products like Douba and Dola. The original Application team has been renamed Product Posttrain-Chat, led by Zhu Wenjia, focusing on C-end applications and dialogue models. The restructuring highlights ByteDance's effort to strengthen its model development pipeline from data collection to reinforcement learning and product post-training. It also aims to create a differentiated model system tailored for enterprise office agents and consumer dialogue scenarios. ByteDance has not yet issued a formal response to these changes.