The tech giant is consolidating its AI research teams as it prepares for a 5-trillion-parameter Omni model
By Chen Jiahui
ByteDance’s foundation-model division, Seed, completed a new round of organizational changes last week, as the company moves to consolidate its AI research teams and prepare for increasingly large and complex models.
The most significant changes are within Seed Foundation Model, which has created four new top-level departments.
Pretrain Data, led by Li Chenggang, merges previously scattered pre-training data teams from text, coding, vision and speech. The unit will handle multimodal data for new Omni models, which can process different types of information such as text, images, audio and video, as well as data for training very large models.
Horizon RL, led by Tang Shengyu, consolidates post-training teams from reasoning and vision understanding, focusing on reinforcement learning to improve the model’s foundational intelligence ceiling.
Product Posttrain-Work, led by Qin Yujia, serves B2B applications, integrating and releasing Agentic models while optimising their capabilities for office scenarios, including task modes for Doubao and Dola.
The Application team, previously led by Zhu Wenjia, has been renamed Product Posttrain-Chat. It will serve consumer applications, focusing on integrating and releasing conversational models. The Chat and Work teams will operate as counterparts serving different user needs.
All four new departments report to Wu Yonghui. Seed also appointed separate heads for AI safety and other frontier exploration areas.
Omni ambition
The restructuring comes after ByteDance discussed plans in early August to train a model with more than 5 trillion parameters — a scale that would put enormous demands on data, computing power and research infrastructure. The company also wants its next generation of models to be fully multimodal (Omni), capable of handling longer and more complex tasks.
That ambition is forcing Seed to reconsider the boundaries between its existing teams. The restructuring announcement emphasized the need to “consolidate similar work, reduce overlap and bring teams together.”
Previously, Seed organised teams by modality and research area — text, code, vision, reasoning — each with its own data and post-training groups. While this suited early-stage development, growing fragmentation increased duplicate investment and cross-team communication costs as research directions multiplied.
Larger models are a systems-engineering challenge. As parameter counts rise, models require more and richer training data, while the bar for data quality also increases. If high-quality data does not grow alongside model size, additional parameters may simply relearn existing information, sharply reducing the returns on further investment.
Omni models add further complexity. Language, code, vision and speech must feed into a single foundation model, requiring decisions on data mix, timing and modal balance directly affect final model capabilities. This demands deeper collaboration across Seed’s data teams.
“Omni is the industry trend,” said one practitioner. Similar adjustments have occurred elsewhere. In July, Tencent’s Hunyuan merged its multimodal and large-language model departments under LLM head Yao Shunyu.
Competitive pressure
ByteDance has been gradually moving toward greater integration within Seed. After Wu Yonghui took over the division in 2025, he established three virtual organizations — Edge, Focus and Base — to encourage teams to share data and code repositories and reduce duplicated development.
The integration has since extended into formal departments. Seed Robotics, for example, which had been spread across hardware, data and model teams, was brought together under a multimodal interaction and world-modeling group led by Zhou Chang, allowing the company to reuse data, models and algorithms. Zhou’s focus has recently shifted more toward physical AI, according to people familiar with the matter.
The reorganisation also comes as Seed faces mounting competitive pressure. Seedance has shown that leading model capabilities can quickly drive user growth and API revenue. But in the core foundation model battleground, Seed still lags, with Coding and Agent capabilities trailing startups including Moonshot AI and Zhipu AI (Z.AI 2513.HK). ByteDance has also not been fastest among major tech firms pivoting AI product focus toward office applications.
The restructuring notably splits the application post-training team by user task into Work and Chat. Chat continues serving mature consumer needs — search Q&A, dialogue experience, personalisation and safety — while controlling inference costs. Work targets the enterprise market, transferring model capabilities in tool-calling, GUI operations and long-task execution from coding scenarios to broader workplace tasks.
Resource realities
ByteDance has the highest revenue among China’s major technology companies and substantial reserves of talent and computing resources. The group also appears willing to give its research teams more time. At an internal Seed meeting, ByteDance founder Zhang Yiming reportedly said the company could accept its models falling behind for a period, with pushing the upper limits of intelligence a more important objective.
Since its founding, ByteDance has embraced a strategy sometimes summed up internally as “throw enough resources at it and something extraordinary will happen.” When entering a new field, it has often established multiple teams and funded parallel efforts — an approach that helped create Douyin.
Over the past two years, ByteDance recruited China’s largest concentration of top AI talent, dispersed across Seed’s model R&D teams. Each is a “genius” in their domain, promised the opportunity to lead through their own innovation. “Everyone wants to achieve something great,” said a former Seed researcher.
But after two years, too many scattered “fronts” have become an obstacle. Large-parameter models require ever more resources — computing power the most critical, scarce, and valuable. Though ByteDance was among the most committed AI investors, its current compute reserves can no longer sustain multi-track parallel development. Concentrating resources on a few core projects has become the more practical choice.
Source:
LatePost
