独家解读丨字节 Seed 大调整，分支背后谁在操盘？

TL;DR - ByteDance reorganized its Seed foundation-model team around centralized pretraining data, reinforcement learning, and separate post-training groups for workplace agents and consumer chat. The shift reflects growing competition around unified multimodal models, reasoning, and agents that execute real tasks.

• Pretrain Data consolidates text, code, vision, and speech data pipelines to support Omni multimodal and large-scale models.
• Horizon RL centralizes reinforcement learning for scaling, reasoning, and vision, building on Seed’s DAPO and VAPO work.
• Product Posttrain-Work targets GUI and office agents, including tool use, computer interaction, and complex task execution for Doubao and Dola.
• Product Posttrain-Chat focuses on search, dialogue, personalization, safety, and inference costs for consumer-facing products.