[ByteDance Seed Restructuring: Pretraining and RL Consolidation, Office Agent Becomes Independent Group] Beating AI Newsflash: ByteDance's large model team, Seed, has completed a new round of organizational restructuring. The Seed Foundation Model has established four primary departments, merging previously dispersed teams focused on text, code, vision, and speech data and post-training. The core change is the unification of pretraining data and reinforcement learning, while splitting post-training applications into two lines: Work and Chat. The Pretrain Data team will be solely responsible for multimodal data for the Omni model, as well as the data required for pretraining ultra-large models. The Horizon RL team will focus on reinforcement learning to enhance the foundational capabilities of the model. The Product Posttrain-Work team will target office and B-end applications, prioritizing the optimization of model tool invocation, computer operations, and execution of long tasks. The original Application team has been renamed Product Posttrain-Chat and will continue to focus on C-end conversational models. All four departments will report to Wu Yonghui. Previously, Seed was more segmented into teams based on text, code, vision, and speech, each with its own data and post-training personnel. Now, ByteDance aims for the next-generation model to directly move toward Omni, integrating different modalities into a single foundation. Seed had also been discussing training models with over 5 trillion parameters. As models grow larger and incorporate more modalities, the previously fragmented R&D approach increasingly leads to redundant investments. Similar adjustments have also occurred at Tencent. In July, Tencent merged its Hunyuan large language model and multimodal teams under unified leadership by Yao Shunyu. ByteDance's establishment of a dedicated Work post-training department further indicates that office Agents have become a distinct product line for optimization within the foundational model team. [Original Link]