ByteDance's Zhang Yiming Sets 2029 or 2030 as Earliest Target to Match OpenAI and Anthropic

A new AI assistant soon, a longer wait for the frontier

ByteDance founder Zhang Yiming has told the company's artificial intelligence team that matching the most advanced technology from OpenAI and Anthropic will take until 2029 or 2030 at the earliest, according to accounts of a recent internal meeting published on November 12, 2026. His immediate concern was the smaller scale of the training datasets behind ByteDance's Seed language models.

The assessment comes as the TikTok parent reportedly prepares to introduce a new personal AI assistant within weeks. ByteDance is pursuing a near term consumer product launch while setting a much longer timetable for reaching the leading edge of AI research.

The effort involves a Seed research and engineering team of about 1,000 people, proposed spending of up to $70 billion and a reported $29.6 billion loan for AI initiatives outside China. ByteDance already operates the Doubao chatbot and sells access to its Seedance video generation models.

Zhang's message puts data at the center of that investment. More computing capacity and a popular app can support an AI business, but his assessment identifies the material used to train its core language models as a priority requiring further work.

What Zhang told the AI team

Zhang, ByteDance's founder, reportedly told an internal meeting attended by its AI team that the company had the ability to compete with OpenAI and Anthropic. He placed the earliest point for catching up with frontier technology at:

2029 or 2030 at the earliest

The timetable places the earliest target roughly three to four years after the November 2026 accounts of the meeting. Zhang also said ByteDance's large models should be measured against the leading international group, making the ambition broader than success in China's consumer chatbot market.

Frontier models are the systems at the leading edge of capabilities such as reasoning, coding and understanding different kinds of information. Matching that frontier means improving while competitors continue to develop new models of their own.

Zhang has also urged patience in internal discussions, favoring sustained model development over temporary gains in public rankings. His comments describe a research effort that will need several years even as ByteDance continues releasing products.

Why the training data gap matters

Seed's pretraining datasets are reportedly smaller than those used by leading overseas competitors. Some accounts of Zhang's remarks describe a gap of several orders of magnitude, although no dataset sizes were disclosed.

Pretraining is the initial learning stage in which a model processes large quantities of material and learns patterns in language and other information. Later training can make the system better at following instructions, answering questions or handling particular tasks.

An order of magnitude means a factor of ten. A gap described in several orders of magnitude therefore suggests a much larger difference than a few percentage points. Without published counts or a shared method of measuring the datasets, the size of ByteDance's reported gap cannot be calculated.

ByteDance's vast consumer platforms give it access to a substantial content ecosystem. Building a leading language model, however, requires suitable training material across many subjects and tasks. Zhang's comments indicate that expanding Seed's training foundation remains a central job for the research team.

A preference for independent model development

Zhang reportedly urged the team to reduce its reliance on distillation, with some accounts describing a stronger objection to depending on it. Distillation means training a model using answers or other outputs produced by another model.

A simple comparison is a student learning from a tutor's worked answers. In AI development, the approach can help transfer useful behavior from a stronger system to another model. Zhang's reported concern is that ByteDance should develop its own underlying capabilities rather than focus on quick improvements in benchmark scores.

Benchmarks are standardized tests used to compare models on selected tasks. They provide a useful measure of performance, while research teams also need to assess how their systems handle the wider range of requests people make in everyday use.

Seed's language models reportedly have yet to rank among the world's leading systems on those tests, and their weights have not been made available to outside developers. Model weights are the learned numerical settings that determine how a system produces its responses. Seed 2.1, released in June 2026, was described as the latest major language model update.

A personal agent beyond the chat window

ByteDance's planned assistant is reportedly being developed under the codename Spell by the Doubao mobile assistant and dialogue teams. Its proposed features include voice interaction, animated avatars and the ability to plan and execute tasks across apps.

An AI agent goes beyond producing an answer in a conversation. It can attempt a sequence of actions toward a goal, such as coordinating steps across different services. For users, that makes permissions, reliable execution and control over actions practical concerns alongside the quality of its replies.

The assistant is expected to integrate Doubao models and Seedance video generation technology. That would connect a conversational interface with ByteDance's existing tools for creating visual content, expanding its consumer AI range beyond the Doubao chatbot.

The proposed product has been compared with Meta's Muse, a voice assistant represented by an animated character that launched in September 2026. Muse reportedly attracted more than 10 million users, with smart glasses support and a forthcoming Muse Charm device also described. ByteDance's exact launch date, final feature list and plans for any accompanying hardware have not been disclosed.

Video gives ByteDance another route to users

Seedance is an important part of ByteDance's strategy because the company already sells access to its video generation models. Such systems create moving images from instructions or other inputs, giving ByteDance a commercial AI product outside text conversation.

Seedance reportedly gained an opening after OpenAI discontinued Sora services. ByteDance is seeking to build it into China's leading video AI platform, alongside the broader effort to improve Seed language models.

Doubao provides another established route to consumers and is described as one of China's most popular chatbot products. The proposed assistant could build on that existing presence instead of requiring ByteDance to introduce an entirely unfamiliar AI service.

The company's product plans and research timetable therefore address different needs. Doubao, Seedance and a new agent can reach users now, while the language model team works toward Zhang's longer target. Progress in video generation gives ByteDance a specific area to build around during that period.

How the spending and financing fit together

ByteDance has reportedly discussed spending up to $70 billion on AI. The period attached to that figure varies: it has been described as infrastructure spending in 2026 and as investment over three years. Those descriptions leave the precise budget schedule unsettled.

The proposed spending covers an effort that depends on substantial computing infrastructure. Training large models requires many processors working together, while running them for users requires continuing capacity in data centers.

ByteDance also reportedly secured an unsecured loan of $29.6 billion in September from close to three dozen banks. The facility runs for three years, with proceeds intended for AI initiatives outside China, including data centers under development in Southeast Asia.

An unsecured loan has no specified assets pledged as collateral. The reported borrowing provides a concrete financing detail alongside the broader investment ambition, although the company has not published a breakdown connecting the loan to particular model projects or the full spending proposal.

The researchers, model scale and safety work

Wu Yonghui leads Seed's roughly 1,000 researchers and engineers after joining ByteDance from Google DeepMind in 2025. His team is responsible for advancing the core models behind the company's push toward the international frontier.

Seed is reportedly pretraining a model with as many as 10 trillion parameters. Parameters are the adjustable numerical values a model learns during training. They help describe its scale, while the data, architecture and training process also shape its capabilities.

The reported parameter count points to a very large development effort. ByteDance has not released detailed specifications for that model, including its design or how much of it would be used to process an individual request.

Seed also reportedly established a separate safety and alignment unit in October 2026, hiring about 10 specialists. Safety work examines risks and failures, while alignment aims to make a system follow intended goals and constraints. Those responsibilities become especially relevant when a product can take actions across apps rather than simply return text.

The dates behind the strategy

ByteDance's recent AI developments show a sequence of recruitment, model releases, financing and organizational changes leading up to Zhang's latest assessment:

• 2025: Wu Yonghui joined from Google DeepMind and took charge of Seed.
• June 2026: Seed 2.1 was released, reportedly the latest major language model update.
• September 2026: ByteDance reportedly secured a loan of $29.6 billion for AI initiatives outside China.
• October 2026: Seed reportedly created a separate safety and alignment unit.
• November 12, 2026: Accounts emerged of Zhang's internal remarks on data, competition and the timetable for catching up.
• The following weeks: ByteDance reportedly plans to introduce its new personal AI assistant.
• 2029 or 2030: Zhang's earliest stated target for matching frontier AI technology.

The next development for consumers is the assistant launch, rather than the distant research target. Its eventual release will show which of the reported voice, avatar, video and task execution features ByteDance is ready to put into users' hands.

For the research team, Zhang's priorities are more fundamental: expand the training foundation, build stronger original capabilities and sustain the work over several years. The spending plans, recruitment and safety unit show the different parts of that effort taking shape around an existing consumer AI business.

Key Points

• Zhang Yiming reportedly said ByteDance could match frontier AI technology in 2029 or 2030 at the earliest.
• He identified Seed's smaller pretraining datasets as a priority and urged less dependence on distillation.
• A personal AI assistant, reportedly codenamed Spell, is planned for the coming weeks.
• ByteDance is building on Doubao and Seedance while pursuing stronger core language models.
• Proposed AI spending reaches $70 billion, but the period attached to that figure varies.
• Seed has about 1,000 research and engineering staff and reportedly added a separate safety and alignment unit in October.