Zhang Yiming's latest statement sets the tone: ByteDance has the capacity to compete with Anthropic and OpenAI, but the earliest time it can catch up with the cutting-edge frontier will be 2029.

Against the fanatical context where Silicon Valley giants repeatedly claim that Artificial General Intelligence (AGI) will arrive in just two or three years, a pragmatic timeline from the top leadership of a domestic Chinese internet company appears exceptionally sober, even somewhat counterintuitive.

On November 11, Bloomberg cited sources familiar with the matter saying that ByteDance plans to launch an AI assistant similar to Meta's Muse in the coming weeks. At a recent all-hands meeting for its AI team, ByteDance co-founder Zhang Yiming set a tone for this large team: ByteDance is capable of competing with Anthropic and OpenAI, but to truly catch up with cutting-edge technologies, "the earliest possible time will be 2029 or 2030."

"2029 or 2030" — this time node is not only an objective assessment of ByteDance's own AI R&D progress, but also an accurate measurement of the real position of China's entire large model industry. At a time when traffic dividends have peaked, Zhang Yiming's remarks have torn a layer of window paper for the industry: in the cutting-edge exploration of underlying large models, there is no such thing as "overtaking on a curve", only an extremely long, expensive and painful process of making up for missed lessons.

Data Divide

Zhang Yiming's concerns first stem from the shortage of the most fundamental "supplies" for large models.

According to reports, he admitted frankly at the internal meeting that the pre-training dataset used by ByteDance's Seed language model is "several orders of magnitude smaller" than the datasets of leading competitors. Today, when the Scaling Laws still dominate the evolution of large models, the quality and scale of data directly determine the physical limit of the model's intelligence level. Although ByteDance has an extremely rich content ecosystem in China, it faces insurmountable physical barriers and copyright gaps when accessing high-quality, multilingual, high-knowledge-density corpora on a global scale.

Accompanied by data shortage is the difficult choice of technical routes. Zhang Yiming reaffirmed a key discipline at the meeting: ByteDance should not rely on "distillation", that is, using the outputs of larger models for training.

No Cheating Bottom Line

This is by no means an easy decision. In the development path of domestic Chinese large models, distillation is a well-known shortcut — using high-quality answers from GPT-4 or Claude to fine-tune one's own model can quickly narrow the gap in various benchmark scores. However, this practice is being criticized by people inside and outside the industry, as well as peers such as Dario Amodei, CEO of Anthropic. For an enterprise like ByteDance that relies heavily on TikTok for global expansion, once disputes arise 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 the short-lived boom of high benchmark scores, and wade honestly into the deep waters of underlying independent R&D. This also explains why ByteDance's main Seed language model seems somewhat "quiet" in the public eye: among leading Chinese large model vendors, the flagship Seed model has not generated publicity by topping global benchmark test rankings, nor has it opened cutting-edge weights to developers, with an overall relatively low-profile attitude. The next major version iteration of the main Seed language model has not been launched yet, and version 2.1 released in June 2026 is still the latest major version release at the main model level.

Doubao and Seed

This dormancy of the underlying model forms a strong sense of contrast with the red-hot development at the application layer.

Judging from user data, ByteDance has built the most popular chatbot in China — Doubao. Relying on huge computing power reserves, free strategy, and pervasive traffic diversion from Douyin, Doubao ranks far ahead in monthly active users and daily active user data in the C-end market. However, the awkward truth is that Doubao's success is largely based on ByteDance's extremely strong product operation capabilities, accurate recommendation algorithms and sufficient user acquisition budget, rather than on the generational gap advantage of the Seed model.

To make up for the shortcomings of underlying technologies, ByteDance has demonstrated its signature heavy investment strategy. In 2025, Zhang Yiming poached Wu Yonghui from Google DeepMind, who leads the Seed department (core R&D caliber) with about 1,000 researchers and engineers, and was granted preferential equity treatment. Wu Yonghui's task is very clear: to build a truly ByteDance-owned "cutting-edge brain".

The 2029 timeline given by Zhang Yiming is precisely based on the objective reality that this 1,000-person team currently has a relatively low starting point and is constrained by the data divide. He is well aware that there are no myths in the competition for cutting-edge large models, only continuous resource investment and trial and error. The genetic reshaping from "product-driven" to "underlying model-driven" cannot be achieved overnight.

Video Generation

Although lagging behind in the absolute capability of language models, ByteDance is not without its unique leading edge in the AI field — video generation.

According to reports, ByteDance's Seedance model has transformed the company's huge video resource library into an important content generation platform. As OpenAI contracted in the AI video field and shut down services related to Sora, Seedance successfully filled this market gap. This is no accident. Video generation models require massive amounts of high-quality video data for training, and ByteDance, backed by Douyin and TikTok, has the world's largest and most dynamic video corpus, which is a natural barrier that no pure text company can match.

To consolidate this advantage, ByteDance has once again deployed its best "heavy asset" strategy. According to reports, the company discussed spending up to 70 billion U.S. dollars on artificial intelligence infrastructure in 2026. This sky-high budget is not only for purchasing GPUs, but also for building data centers and hoarding computing power, trying to use the brute force of physical scale to hedge the time difference in the technological evolution route.

Once Seedance can stably output high-quality, long-duration, highly controllable video content, it will directly impact the production processes of film and television, advertising, short dramas, and even e-commerce product images. On this track, ByteDance does not need to wait until 2029 to catch up — it is already at the cutting edge right now. This differentiated competition based on its own resource endowment is the most realistic and stable cash flow engine for ByteDance on the long road to AGI.

Agent and Compliance

Back to the soon-to-be-launched Muse-like AI assistant. The news revealed by Zhang Yiming indicates that ByteDance is preparing to take a key step in the form of AI products.

Meta's Muse integrates AI research results into voice-based assistants and animated avatars, and has 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 multi-modal 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 multi-modal Agent puts extremely high requirements on the model's end-side response speed and multi-modal understanding capability, and also means that the red line of data security and privacy compliance is approaching increasingly. It is worth noting that in October 2026, according to reports, ByteDance set up a separate safety and alignment team, making it one of the earlier institutions among domestic cutting-edge AI labs to take this arrangement. This move is of great significance. In the past, Silicon Valley giants have 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 Agent in the future. In the era of artificial intelligence, safety and alignment are no longer an accessory of R&D, but an important prerequisite for products to enter the global market.

Conclusion: Long-term Combat

Zhang Yiming's "2029" assertion provides a rare pragmatic coordinate.

It acknowledges that local technology companies still face multiple constraints such as algorithm talents, high-quality data, computing power infrastructure, and global market access when facing cutting-edge technology leaders like OpenAI and Anthropic. But this is not passive admission of defeat, but a long-term combat plan based on clear cognition.

ByteDance's advantages lie in its extremely strong survival adaptability and its overwhelming huge cash flow. Abandoning distillation, poaching talents with high salaries, betting on video generation, investing 70 billion U.S. dollars in infrastructure, launching hardware Agents, and setting up a safety team — Zhang Yiming is using this combination of strategies to try to sustain a protracted war for underlying technologies with the huge cash flow at the application layer and physical computing power.

If the timeline remains unchanged, in the Agent competition in a few years, we will see OpenAI, Meta and important Chinese competitors head-to-head. What determines the outcome may not just be a few decimal points on the model parameter list, but who can truly seamlessly integrate AI into the 24-hour life of ordinary people. On this track, ByteDance may not need to be the first genius to cross the finish line, as long as it can survive with a sufficiently huge user ecosystem and reach 2029.

This article is from the WeChat Official Account "Tech Business", written by Hai Yi and Luo Wei, and published with authorization from 36Kr.