ByteSeed Reorganization: Pre-training, RL Consolidation, Office Agent Grouping

Perceiving Beating AI Newsflash: ByteDance's Large Model Team Seed has completed a new round of organizational adjustments. The Seed Foundation Model has established four new primary departments, consolidating the previously scattered data and post-training teams in the text, code, vision, and speech directions. The key change is to unify pre-training data and reinforcement learning, while splitting the application post-training into two lines: Work and Chat.

The Pretrain Data department is responsible for the multimodal data of the Omni model and the data required for training extra-large models. The Horizon RL department is focused on reinforcement learning to enhance the model's core capabilities. The Product Posttrain-Work department is oriented towards office and B-end scenarios, emphasizing the optimization of model invocation tools, computer operation, and the ability to perform long tasks. The original Application team has been renamed Product Posttrain-Chat, continuing to be responsible for C-end conversational models. All four departments report to Wu Yonghui.

In the past, Seed's teams were more divided into text, code, vision, speech, and other directions, each with its own data and post-training personnel. Now ByteDance hopes that the next-generation model will directly evolve towards Omni, allowing different modalities to converge on the same foundation. Seed had previously discussed training a model with over 5 trillion parameters. The larger the model and the more modalities involved, the more likely the previously fragmented development approach is to result in redundant efforts.

Similar adjustments have also occurred at Tencent. In July, Tencent merged its hybrid large language model and multimodal teams, with Yao Shunyu taking unified responsibility. ByteDance has also specifically established a Work post-training department this time, indicating that the office Agent has become a product line independently optimized by the foundational model team.