Skip to main content

ByteDance's Seed Team Restructures: Four Core Departments Aim for 5 Trillion Parameters

ByteDance's Seed division is making big moves. The team behind some of the company's most ambitious AI projects has just completed a sweeping reorganization, merging previously scattered groups into four core departments. The goal? To gear up for the next generation of ultra-large models—some whispers suggest parameter counts could exceed 500 billion.

This isn't just a reshuffle for the sake of change. The old structure divided teams by modality and research direction, which worked well in the early days but started showing cracks as technical boundaries blurred. Cross-team communication became a headache, and there was plenty of duplicated effort. The new setup is all about consolidation: merge similar work, cut overlap, and focus the firepower.

So, what does the new landscape look like? Four first-level departments have emerged, each with a clear mission.

Pretrain Data is led by Li Chenggang. This team brings together pre-training efforts that were previously split across text, programming, visual understanding, and speech. Their job? Handle multi-modal data and pre-train the upcoming Omni model. It's a massive undertaking, but with a unified team, the hope is that things move faster and smoother.

Horizon RL, under Tang Shengyu, is all about reinforcement learning. This group integrates post-training capabilities that used to be scattered across various areas. The core objective is to push the boundaries of model intelligence—essentially, making these models smarter through RL.

For those focused on real-world applications, ByteDance has split the work into two distinct tracks. Product Posttrain-Work, led by Qin Yujia, zeroes in on B-end applications and Agentic models. Think office scenarios, task automation, and supporting tools like Doubao and Dola. Meanwhile, the former Application team, now renamed Product Posttrain-Chat, is dedicated to C-end dialogue models. This team, previously led by Zhu Wenjia, focuses on making chatbots more engaging and useful for everyday users.

All four departments report directly to Wu Yonghui, ensuring a unified vision. Seed has also appointed dedicated leads for frontier areas like AI safety, signaling that they're not just chasing scale but also responsibility.

Why the urgency? Reports suggest ByteDance is in the early stages of training a model with over 500 billion parameters. If that happens, it would dwarf Alibaba's Qwen 3.8-Max and Moonshot's K3, making it the largest known model in China. That's a bold ambition, and this restructuring is the organizational backbone to make it happen.

In the cutthroat world of AI, where every tech giant is vying for supremacy, ByteDance is clearly betting big. By streamlining its teams and sharpening its focus, the company is positioning itself to not just keep up but lead the pack. Whether they hit that 5 trillion parameter mark remains to be seen, but one thing's for sure: they're not holding back.

Key Points

  • ByteDance's Seed division has restructured into four core departments: Pretrain Data, Horizon RL, Product Posttrain-Work, and Product Posttrain-Chat.
  • The reorganization merges previously scattered teams to reduce redundancy and improve R&D efficiency.
  • The move is preparation for a rumored ultra-large model with over 500 billion parameters, which would be the largest in China.
  • All departments report to Wu Yonghui, with dedicated leads for AI safety and other frontier areas.