Skip to main content

ByteDance's Seed Team Overhauls: New Focus on AI Agents and Chat

ByteDance's large model division, Seed, has just gone through a significant internal shakeup. According to reports, the team completed a new round of organizational restructuring last week, creating four new first-level departments under the Seed Foundation Model umbrella. All of these new department heads will report directly to Wu Yonghui.

This isn't just a simple reshuffle—it's a strategic move to tighten the entire AI development pipeline. The goal is to better integrate pre-training data, reinforcement learning, and post-training resources, while also drawing a clear line between the model capabilities for business (B-end) and consumer (C-end) products.

Let's break down what's new. The Pretrain Data department, now led by Li Chenggang, is pulling together teams that were previously scattered across text models, programming, visual understanding, and speech. This consolidation is aimed at handling the massive multimodal data requirements for the upcoming Omni model and other ultra-large pre-training projects.

Meanwhile, the Horizon RL department, under Tang Shengyu, is merging post-training, reasoning, and visual understanding teams. The focus here is on advancing reinforcement learning to push the boundaries of the model's basic intelligence—essentially, making the AI smarter at its core.

On the product side, things are getting more specialized. The Product Posttrain-Work team, led by Qin Yujia, is targeting B-end applications. They're responsible for integrating and releasing Agentic models, which are designed for office scenarios. This means optimizing AI agents to handle tasks in products like Douba and Dola, making them more useful for enterprise users.

In contrast, the original Application team has been renamed Product Posttrain-Chat and will continue under Zhu Wenjia. This team is all about C-end applications, focusing on dialogue models that power consumer-facing chatbots. The clear division of labor between the Work and Chat teams means that enterprise agents and consumer chat experiences will now have dedicated, tailored development paths.

This restructuring signals that Seed is doubling down on a full-stack approach—from data collection and pre-training, through reinforcement learning, to product-specific post-training. By separating the model development for enterprise office agents and consumer dialogue scenarios, ByteDance is positioning itself to better serve different market segments with more specialized AI solutions.

It's a bold move, but not entirely surprising. As AI competition heats up, companies are realizing that one-size-fits-all models may not cut it. By tailoring models to specific use cases, ByteDance hopes to gain an edge in both the enterprise and consumer markets.

As of now, ByteDance hasn't officially commented on the restructuring. But the industry is watching closely—this could be a sign of where the company is heading in the AI race.

Key Points

  • Seed Foundation Model now has four new first-level departments, all reporting to Wu Yonghui.
  • Pretrain Data, led by Li Chenggang, consolidates scattered data teams for multimodal and ultra-large model pre-training.
  • Horizon RL, led by Tang Shengyu, focuses on reinforcement learning to enhance model intelligence.
  • Product Posttrain-Work (B-end) and Product Posttrain-Chat (C-end) are now separate, targeting enterprise agents and consumer dialogue respectively.
  • ByteDance has not yet responded to the restructuring news.