ByteDance is entering 2026 with four priorities in artificial intelligence. This includes increasing investment in world models, keeping its video model Seedance near the top tier globally, strengthening its coding foundation, and accelerating commercialization of Doubao, especially in office productivity.
World models
ByteDance’s AI portfolio already includes Seed 2.0, Seedance 2.0, and Doubao, which has built a large user base on the application side. Multiple sources told 36Kr that Doubao’s daily active users (DAUs) surpassed 200 million after the 2026 Lunar New Year.
But ByteDance’s AI lineup still has a gap in world models, a research direction widely seen as central to the next stage of large model development.
World models are AI systems designed to understand and simulate physical or digital environments, making them relevant to robotics, games, and interactive content.
Several people close to the Seed team said ByteDance entered the world model race late. In 2024, Zhou Chang, who had just joined ByteDance from Alibaba, began leading world model research. The internal view at the time was that world model routes and commercial scenarios were still unclear, while securing a lead in video models was more urgent.
ByteDance formed a small research group in 2025 to explore the vision-language-action (VLA) route for world models. VLA models combine visual perception, language understanding, and action planning, and are often used in embodied intelligence. The work was led by two people. One was Li Hang, head of ByteDance AI Lab, which was merged into Seed in April 2025 along with its robotics team to improve coordination between models and embodied intelligence applications. Li’s group focused mainly on simulation data. The other was Seed multimodal researcher Wang Wenqian, whose work focused on natural data.
In 2026, Wu Yonghui, who leads Seed, gave the team a clear target: release at least one world model by year’s end and benchmark its performance against Google’s Genie 3.
Progress has not been fast enough, according to people familiar with the matter. One person close to Seed said Wu has repeatedly told internal meetings that ByteDance’s world model and embodied intelligence results have fallen short of expectations. Another Seed member said internal tests as of early 2026 placed ByteDance’s world model performance about 10% behind current state-of-the-art alternatives globally.
Still, ByteDance sees the field as a long-term priority. World models are closely tied to embodied intelligence, a high-potential market, as well as games and entertainment, where potential use cases are broad. One former Seed researcher said ByteDance’s robotics deployments had mainly focused on transport and industrial handling, but the company sees humanoid robotics as a direction it must enter.
The company has adjusted accordingly. After the Lunar New Year, Seed established a new research group led by Fan Haoqi, a former Meta Fundamental AI Research Lab researcher, reporting to Zhou. At the same time, the VLA groups led by Li and Wang were merged and placed under Zhou. Multiple people familiar with the matter said the earlier groups were focused on VLA for embodied intelligence, while Fan’s team is taking a 3D simulation route aimed at entertainment and gaming.
World models are also receiving ByteDance’s largest data budget among its model directions, including text, coding, and video, according to multiple sources. They said the 2026 budget for world model training data across VLA, long video, and 3D modalities had reached an eight-figure RMB sum. One data supplier said ByteDance’s world model data spending is three to four times that of other companies.
Coding
Coding ability is a foundation for agent performance, a view that has become a consensus across the sector. Several sources said ByteDance has long emphasized coding, with one person close to Seed saying the company’s investment in AI coding is second only to world models this year.
The company has purchased data directly and studied sample training data from leading overseas coding models such as Claude Code and Codex. At the 2025 Volcano Engine Force conference, ByteDance vice president of technology Hong Dingkun said coding is highly structured and logic-intensive, making it useful for pushing models’ limits in semantic understanding, reasoning, algorithm design, and precision.
Externally, ByteDance’s coding business has been less visible. Its Doubao-Seed-Code model, released in November 2025, and Trae, its AI coding tool launched in early 2025, have drawn less attention than products such as Zhipu’s GLM 5 and Moonshot AI’s Kimi K2.
One source said ByteDance’s coding models have struggled to break through because they lack data feedback. Business teams have been reluctant to use Seed-Code because of model limitations. Even Trae was initially connected to DeepSeek, Claude Code, and an internally trained coding model, rather than relying solely on Seed-Code, limiting the feedback available from real use cases.
That is changing in 2026. A Seed employee said multiple application departments are now being required to use Seed models, whereas ByteDance previously did not restrict business teams from using third-party coding models. At the same time, hiring has become more selective. One AI recruiter said ByteDance is signaling that the period of broad, high-salary external hiring is over, with the next phase focused on internal development, promotion of younger talent, and higher compensation for algorithm roles.
Keeping Seedance ahead
ByteDance’s second major AI priority for 2026 is maintaining Seedance’s position in global video generation.
One video generation startup founder described Seedance 2.0 as a victory of data. According to 36Kr, a large training data pool and a review team of more than 2,000 people helped produce its results.
But reliance on scale also creates problems. Some research has identified an “anti-scaling law” in video generation, where more data can encourage models to learn key frames while neglecting full narrative continuity. Two people on the data side said Seedance has reached a ceiling in pretraining and will need cleaner training data and more refined post-training to keep improving.
Dynamic generation is another focus for the Seedance team in 2026. The term refers to interactive video generation, where users can input instructions and adjust generated content and plotlines in real time. One company already active in this area is Vivix AI, founded by former SenseTime senior research director Liu Yu and valued at USD 1.3 billion, sources told 36Kr.
Doubao turns toward monetization and overseas growth
Doubao’s commercialization is the fourth priority. 36Kr previously reported that Doubao is expected to launch paid subscription plans in late June and connect with Douyin E-commerce to expand monetization.
In early May, Doubao updated subscription options in the app store, with monthly plans ranging from free to RMB 500 (USD 73.6). On June 3, Doubao announced an upcoming professional version for workplace productivity needs, including software development, data analysis, design, workflow automation, financial analysis, and scientific research.
Several sources told 36Kr that Doubao had surpassed 200 million DAUs after the Lunar New Year. One source said Doubao’s user acquisition budget is now low, while high DAU has brought heavy inference and operating costs. Monetization at this stage is meant both to slow the pace of growth and help the product generate its own cash flow.
Slide deck generation is viewed as the main entry point for paid user behavior. One person close to Doubao said the company wants to strengthen slide deck generation to build revenue from white-collar professionals in sectors such as finance and law. Doubao also plans to launch an enterprise version and connect with internal company systems, though the exact integration model is still under discussion.
The idea reflects similar trends overseas. Anthropic disclosed that Claude Code reached USD 2.5 billion in annualized recurring revenue (ARR) in February, more than double its level at the start of the year. Cash flow from enterprise development use cases reportedly helped Anthropic surpass OpenAI in ARR earlier this year, despite Anthropic being founded later.
Doubao’s challenge is to reposition itself from a free general-purpose chatbot into a paid office assistant that improves productivity. But the market it wants to enter is already crowded. A Doubao source said ByteDance found through its research that many solution providers already occupy the enterprise AI tool market, meaning a late-arriving Doubao will face higher acquisition costs.
Overseas expansion is also a key task this year. Jiemian News previously reported that Dola, Doubao’s overseas app, surpassed ten million DAUs by the end of 2025. According to 36Kr, Dola’s 2026 target is 30 million DAUs by year’s end. With overseas AI chatbots largely dominated by ChatGPT, Claude, and Gemini, Dola’s strategy is said to focus on more niche markets rather than regions like Europe and the US, where competition is tighter.
KrASIA features translated and adapted content that was originally published by 36Kr. This article was written by Zhou Xinyu for 36Kr.
Note: RMB figures are converted to USD at rates of RMB 6.79 = USD 1 based on estimates as of June 9, 2026, unless otherwise stated. USD conversions are presented for ease of reference and may not fully match prevailing exchange rates.