ByteDance Seed Foundation Model Team Restructures to Target 5 Trillion Parameters: Four Major Departments Take Shape.

ByteDance has carried out another major internal reshuffle within Seed, its large-model division, as the company prepares for the next phase of foundation-model development. The move is widely seen as a step toward supporting a new ultra-large model reportedly being discussed at a scale of more than 5 trillion parameters.

The latest reorganization changes how the Seed base-model team is structured. Instead of continuing with a setup built largely around separate modalities and research tracks, ByteDance has chosen to consolidate previously scattered teams. The goal is to improve coordination, reduce duplicated work, and raise overall R&D efficiency as model development becomes more complex.

Four top-level units now anchor Seed Foundation Model

As part of the restructuring, Seed Foundation Model has formally established four primary departments. Together, they cover pretraining data, reinforcement learning, and post-training work for different product scenarios.

Pretrain Data

This team is led by Li Chenggang. It brings together pretraining groups that were previously split across text, coding, visual understanding, speech, and other domains. Under the new structure, the unit is responsible for multimodal data and pretraining for ByteDance’s new Omni model as well as its next-generation ultra-large models.

Horizon RL

Headed by Tang Shengyu, this department combines post-training capabilities that had been distributed across areas such as post-training, reasoning, and visual understanding. Its central mission is to use reinforcement learning to raise the model’s underlying intelligence ceiling.

Product Posttrain-Work

Led by Qin Yujia, this group focuses on post-training for business-facing applications and the integrated release of agentic models. Its work is centered on improving agent-style capabilities for office and productivity scenarios, while also supporting task-based modes for Doubao and Dola.

Product Posttrain-Chat

This unit is overseen by Zhu Wenjia and is a renamed version of the former Application team. It is mainly responsible for integrating and shipping consumer-facing conversational models, working alongside the Work team with a clearer division of roles.

All four department heads now report to Wu Yonghui. In parallel, Seed has also assigned dedicated leadership for frontier exploration areas including AI safety.

A shift away from modality-based silos

Previously, ByteDance’s Seed organization was split more explicitly by modality and by distinct research directions. That structure helped the company build capabilities quickly during an earlier stage of model development, when teams were still racing to fill gaps across text, vision, speech, and related areas.

But as foundation models have evolved, those boundaries have become less useful. Technical overlap between teams has increased, communication across organizational lines has become more expensive, and duplicated investment has become harder to avoid.

The logic behind the new setup is therefore straightforward: combine closely related work, reduce parallel construction, and strengthen collaboration within larger units. By doing so, ByteDance aims to create an organization better suited to training and refining models at much larger scale.

Organizational groundwork for a larger model push

Earlier media reports said ByteDance was in the early stages of discussing a model with more than 5 trillion parameters. If that effort moves forward, its scale would exceed Alibaba’s Qwen 3.8-Max and Moonshot AI’s K3, making it the largest known model by parameter count among Chinese companies.

Against that backdrop, the restructuring of Seed Foundation Model appears to be more than a routine internal adjustment. It is increasingly viewed as a foundational organizational step for ByteDance’s next wave of large-model development, especially as the company pushes toward a new generation of ultra-scale AI systems.