Zhang Yiming Sets ByteDance AI Direction: Capable of Competing with Anthropic and OpenAI, Catching Up to the Frontier as Early as 2029
According to Bloomberg, citing people familiar with the matter, ByteDance co-founder Zhang Yiming said at a recent all-hands meeting of the AI team that ByteDance is capable of competing with Anthropic and OpenAI, but truly catching up to frontier technology will not happen until 2029 or 2030 at the earliest. He noted that the pre-training dataset used by the Seed language model is several orders of magnitude smaller than those of leading competitors, and reiterated that it should not rely on distillation. The report also said ByteDance plans to release an AI assistant similar to Meta's Muse in the coming weeks, and is discussing spending up to $70 billion on AI infrastructure in 2026.
Why read this
AI score 60
At an internal meeting, Zhang Yiming laid out a timeline for ByteDance to catch up to the frontier and explained the constraints on data scale and distillation discipline.
Topics
Body · AI translation
Zhang Yiming sets the latest tone: ByteDance is capable of competing with Anthropic and OpenAI, but catching up to the frontier will not happen until 2029 at the earliest Tech Business November 12, 2026 10:16:00 Updated November 12, 2026 11:06:56
Original title: Zhang Yiming sets the latest tone: ByteDance is capable of competing with Anthropic and OpenAI, but catching up to the frontier will not happen until 2029 at the earliest
Amid the feverish rhetoric in which Silicon Valley giants casually claim that artificial general intelligence (AGI) will arrive within two or three years, a pragmatic timetable from the top level of a Chinese internet company stands out as remarkably clear-headed, even somewhat counterintuitive.
On November 11, Bloomberg, citing people familiar with the matter, reported that ByteDance plans to release an AI assistant similar to Meta's Muse within the next few weeks. And at a recent all-hands meeting of its AI team, ByteDance co-founder Zhang Yiming set the tone for this enormous team: ByteDance is capable of competing with Anthropic and OpenAI, but truly catching up to the frontier technology will take "until 2029 or 2030 at the earliest."
"2029 or 2030"—this time point is not only an objective assessment of ByteDance's own AI R&D progress, but also a precise measurement of where China's large-model industry actually stands. Today, with traffic dividends peaking, Zhang Yiming's remarks have torn away a layer of window dressing from the industry: in the frontier exploration of underlying large models, there is no leapfrogging; there is only an extremely long, expensive, and painful process of catching up.
The data gap
Zhang Yiming's concern first stems from a shortage of the most fundamental "provisions" for large models.
According to reports, he acknowledged at the internal meeting that the pre-training dataset used by ByteDance's Seed language model is "several orders of magnitude smaller" than those of leading competitors. Today, with Scaling Laws still governing the evolution of large models, the quality and scale of data directly determine the physical limits of a model's intelligence. Although ByteDance has an extremely rich content ecosystem in China, it faces insurmountable physical barriers and copyright gaps when obtaining high-quality, multilingual, knowledge-dense corpora on a global scale.
Accompanying the data shortage is a difficult choice of technical route. At the meeting, Zhang Yiming reiterated a key discipline: ByteDance should not rely on "distillation," that is, using the outputs of larger models for training.
The bottom line against cheating
This is not an easy decision. In the development path of Chinese large models, distillation is a well-known shortcut—fine-tuning one's own model with high-quality answers from GPT-4 or Claude can quickly narrow the gap in various benchmark scores. However, this practice is being criticized both inside and outside the industry, as well as by peers such as Anthropic CEO Dario Amodei. For a company like ByteDance, which is highly dependent on TikTok for global expansion, once controversy arises over copyright and data sources, the compliance costs and brand risks it faces in overseas markets cannot be underestimated.
The bottom line drawn by Zhang Yiming means that ByteDance must abandon short-term benchmark prosperity and honestly wade into the deep waters of underlying self-research. This also explains why ByteDance's main Seed language model appears somewhat "quiet" in the public eye: among leading Chinese large-model vendors, the Seed flagship model has not made a splash by topping global benchmark tests, nor has it opened frontier weights to developers; its overall posture is relatively low-key. The next major version iteration of the main Seed language model has not yet landed, and the 2.1 version from June 2026 remains its most recent major release at the main model level.
Doubao and Seed
This dormancy at the underlying model layer forms a sharp contrast with the roaring fire at the application layer.
In terms of user data, ByteDance has built China's most popular chatbot—Doubao. With its enormous computing power reserves, free strategy, and Douyin's ubiquitous traffic funneling, Doubao leads far ahead in monthly and daily active users in the consumer market. Yet the awkward fact is that Doubao's success is largely based on ByteDance's extremely strong product operations capabilities, precise recommendation algorithms, and ample user-acquisition budgets, rather than on a generational advantage of the Seed model.
To make up for its shortcomings in underlying technology, ByteDance has demonstrated its signature willingness to bet heavily. In 2025, Zhang Yiming poached Wu Yonghui from Google DeepMind, put him in charge of the Seed department (the core R&D scope) with about 1,000 researchers and engineers, and gave him preferential equity treatment. Wu Yonghui's task is very clear: build a "frontier brain" that truly belongs to ByteDance.
The 2029 timetable given by Zhang Yiming is based precisely on the objective reality that this 1,000-person team is currently starting from a relatively low base and is constrained by the data gap. He knows that there are no myths in the competition over frontier large models, only continuous resource infusions to fill the pits and trial and error. Reshaping the company's DNA from "product-driven" to "underlying-model-driven" is by no means something that can be achieved overnight.
Video generation
According to reports, ByteDance's Seedance model turns the company's vast video resource library into an important content generation platform. As OpenAI pulls back in the AI video field and shuts down its related Sora services, Seedance has successfully filled this market gap. This is by no means a coincidence. Video generation models require massive amounts of high-quality video data for training, and ByteDance, backed by Douyin and TikTok, possesses the world's largest and most vibrant video corpus—a natural barrier that no pure-text company can match.
To consolidate this advantage, ByteDance has once again deployed its favorite "heavy-asset" playbook. According to reports, the company has discussed spending up to $70 billion on AI infrastructure in 2026. This astronomical budget is not just for buying GPUs, but also for building data centers and stockpiling computing power, attempting to use the brute-force aesthetics of physical scale to hedge against the time gap in the technology evolution roadmap.
Once Seedance can stably produce high-quality, long-duration, highly controllable video content, it will directly disrupt the production processes of film and television, advertising, short dramas, and even e-commerce product images. On this track, ByteDance does not need to chase 2029; it is the frontier itself right now. This kind of differentiated competition based on its own resource endowments is ByteDance's most realistic and most reliable cash-flow engine on the long road to AGI.
Agents and compliance
Meta's Muse integrates AI research results into voice-based assistants and animated avatars, and has already been successfully applied to its smart glasses and the upcoming dedicated device Muse Charm. This is an agent form that combines software and hardware and emphasizes companionship and multimodal interaction. If ByteDance launches a similar product, it means its AI strategy is moving from "the chat box on the screen (Doubao)" to "AI worn on the body."
The combination of hardware and multimodal agents places extremely high demands on the model's on-device response speed and multimodal understanding capabilities, and also means that the red lines of data security and privacy compliance are drawing ever closer. It is worth noting that in October 2026, according to reports, ByteDance separately established a safety and alignment team, becoming one of the earlier institutions among China's frontier AI labs to adopt such an arrangement. This move is of great significance. In the past, Silicon Valley giants signed safety commitments one after another to meet compliance requirements; ByteDance's move is not only to meet the compliance requirements of different markets, but also to establish a safety buffer in advance for the large-scale overseas expansion of its AI hardware and agents in the future. In the era of artificial intelligence, safety and alignment are no longer appendages of R&D, but an important prerequisite for whether a product can enter the global market.
Conclusion: A long-term battle
Zhang Yiming's "2029" assertion provides a rare pragmatic coordinate.
It acknowledges that domestic tech companies, when facing frontier technology leaders such as OpenAI and Anthropic, still face multiple constraints including algorithm talent, high-quality data, computing infrastructure, and global market access. But this is not a passive admission of defeat; it is a long-term battle plan based on clear-eyed understanding.
ByteDance's advantage lies in its extremely strong survival adaptability and its vast, suffocating cash flow. Abandoning distillation, spending heavily to poach talent, betting on video generation, investing $70 billion in infrastructure, launching hardware agents, and establishing a safety team—Zhang Yiming is using this combination of moves to try to use the massive cash flow at the application layer and physical computing power to endure a protracted war in underlying technology.
If the timetable remains unchanged, in the agent competition a few years from now, we will see OpenAI, Meta, and a major Chinese competitor go head-to-head. And what decides the outcome may not just be a few decimal points on a model parameter table, but who can truly integrate AI seamlessly into the 24-hour lives of ordinary people. On this track, ByteDance may not need to be the genius that crosses the finish line first; as long as it can carry a sufficiently large user ecosystem and survive to 2029.