ByteDance’s Seed foundation-model division completed an internal organizational restructuring in August 2026, establishing four first-level departments designed to streamline responsibilities and accelerate coordination across core training and productization work. The new structure aims to eliminate cross-team communication barriers and cut duplicated efforts that arose under earlier modality-based groupings, according to a TechNode report detailing the changes and their timing.
The reorganization coincides with reports that the company is discussing the training of a foundation model with more than 5 trillion parameters. As AIBase News reported, those plans remain subject to early-stage discussion and have not been officially confirmed or announced as a commercial product by the company. The company’s structural shift places greater emphasis on unified data preparation, reinforcement learning, and product-aligned post-training tracks, while the reported model scale underscores the ambition under consideration.
The reported restructuring arrives after a period of visible competition among Chinese model developers, including ByteDance’s reported five-trillion-parameter model plan. The restructuring arrives as those conversations continue, without a formal product announcement.
Inside the Four New Departments
The Seed team’s Pretrain Data department is led by Li Chenggang and consolidates pretraining teams across text, programming, visual understanding, and speech. Bringing those previously separate efforts into a single first-level group is intended to standardize data pipelines and align pretraining priorities across modalities. By centralizing teams responsible for how data is prepared and fed into training runs, the unit is set up to reduce overlap that accumulated when text, code, vision, and audio matured on parallel tracks.
Horizon RL, led by Tang Shengyu, integrates post-training and reasoning capabilities to enhance basic intelligence through reinforcement learning. Placing reinforcement learning alongside reasoning focuses the organization’s attention on the systems that shape emergent behavior after base pretraining. The department’s mandate links tuning and evaluation with the iterative techniques used to raise model reliability and step-by-step problem solving, creating a direct bridge between pretraining outputs and the behaviors users experience.
Product Tracks for Work and Chat
On the product side, ByteDance created two first-level departments that distinguish enterprise-style and consumer-facing usage. The Product Posttrain-Work department, led by Qin Yujia, focuses on business applications and agentic capabilities for products like Doubao and Dola. That remit emphasizes post-training regimens that reflect productivity demands, workflow integration, and the orchestration of agent behaviors aligned with workplace tasks. By concentrating these expectations in one department, the Seed team can tailor tuning strategies to the needs of professional users and enterprise partners.
Product Posttrain-Chat handles consumer-facing dialogue models. The separation between Work and Chat creates clear lanes for optimizing conversational tone, safety, latency expectations, and feature priorities that differ meaningfully between personal assistants and business copilots. Placing these concerns at the first level of the organization is consistent with the broader restructuring goal of removing friction that previously arose when modality ownership obscured downstream product needs.
Reporting Lines and Strategic Focus
The four new departments report to Wu Yonghui. In parallel, the Seed organization appointed separate leaders for AI safety and for frontier exploration. Establishing those leadership posts distinct from the four first-level teams sets out responsibility for risk management and longer-horizon research without blurring accountability for day-to-day pretraining, tuning, and product delivery.
Beyond new reporting lines, the restructuring is designed to eliminate cross-team communication barriers and reduce duplicated efforts caused by the earlier division of work along modality boundaries. The restructuring consolidates work that had been organized across text, programming, visual-understanding, and speech teams. Consolidating pretraining, reinforcement learning and reasoning, and product-aligned post-training into clear first-level departments is intended to create a more linear path from data to deployment.
The timing of the restructure coincides with external reports that the company is discussing training a foundation model with more than 5 trillion parameters. While such reports highlight the scale under consideration, they remain in early discussion stages and have not been officially confirmed or announced as a commercial product by the company. For now, the confirmed development is the Seed team’s new organizational blueprint.
Each department’s defined scope reflects an effort to match organizational structure to the stages of model development and use. Pretrain Data centralizes the intake of text, programming, visual understanding, and speech data in order to reduce redundancy. Horizon RL aligns reinforcement learning and reasoning capabilities with the goal of enhancing base intelligence after pretraining. Product Posttrain-Work and Product Posttrain-Chat delineate product-facing responsibilities so that business and consumer experiences are tuned for their distinct expectations.
The appointment of separate leaders for AI safety and frontier exploration signals attention to both guardrails and long-term inquiry alongside the immediate delivery focus of the four first-level teams. That arrangement is intended to ensure that risk reduction and exploratory research do not dilute execution on pretraining, reinforcement learning, and product post-training, while still remaining represented at the leadership table.
By shifting from modality-based divisions to first-level departments organized around data, reinforcement learning and reasoning, and product outcomes, the Seed foundation-model team is aiming for clearer communication channels and less duplicated work. The restructured lines of authority place Li Chenggang over Pretrain Data, Tang Shengyu over Horizon RL, and Qin Yujia over Product Posttrain-Work, while Product Posttrain-Chat concentrates on consumer dialogue models, all reporting to Wu Yonghui. As reports circulate about a potential effort exceeding 5 trillion parameters, the company’s confirmed move remains this reorganization, with any larger-scale training plans still unconfirmed and not announced as a commercial product.