ByteDance outlines 4 AI priorities: world models, coding, Seedance, Doubao

ByteDance is entering 2026 with four clear AI priorities: increasing investment in world models, keeping its video model Seedance near the global top tier, strengthening coding capabilities for agent performance, and accelerating commercialization of its Doubao assistant—especially in office productivity. Doubao has already surpassed 200 million daily active users after the 2026 Lunar New Year, giving ByteDance a strong consumer-facing foothold. However, the company acknowledges a significant gap in world models, which are AI systems designed to understand and simulate physical or digital environments for robotics, games, and interactive content. Internal tests as of early 2026 place ByteDance’s world model performance about 10% behind current state-of-the-art alternatives globally.

To close that gap, ByteDance has allocated its largest data budget among all model directions—an eight-figure RMB sum for world model training data across VLA (vision-language-action), long video, and 3D modalities. The company restructured its research teams after the Lunar New Year, merging VLA groups under Zhou Chang and adding a new 3D simulation team led by former Meta FAIR researcher Fan Haoqi, targeting entertainment and gaming. Coding is the second-largest investment area this year, as models like Claude Code and Codex are used to study sample training data. ByteDance sees coding as foundational for agent performance and has purchased data directly to improve its models. The company is also exploring humanoid robotics as a long-term direction, despite early robotics deployments being limited to transport and industrial handling.