Unveiling ByteDance's Seedance: The Second Lucrative Global AI Business | 36Kr

Interview | Deng Yongyi, Xiao Sijia, Li Xiaoxia

Text | Zhang Yuxin, Deng Yongyi

Editor | Zhang Yuxin, Yang Xuan

At the end of 2025, during a team dinner for the model group, Zeng Yan, the lead of the Seedance model, mentioned to her leader that she still wanted to try training a larger model, with at least 200B (200 billion parameters).

Zeng Yan is a very young researcher at ByteDance's Seed division, a 2021 campus hire whose research has always focused on video understanding and generation. "She has always had her own technical judgments," a person who has worked with Zeng Yan told 36Kr, "and she is very proactive, persistent in what she believes in, and will find ways to secure resources to make it happen."

This evaluation almost perfectly mirrored the situation Zeng Yan faced when training Seedance 2.0. "She wanted to train a model with more parameters, but there were disagreements within the team at the time. Some people thought directly scaling the model up to 200B-300B was a bit too aggressive, and overall training resources were tight, so training a 100B-level model first would be safer," an insider told 36Kr.

The source said that coincidentally, a senior executive was at that dinner, learned about Zeng Yan's idea, and said he could coordinate some resources for the training of Seedance 2.0. "Later, at the final official review meeting within Seed, after Wu Yonghui, the head of Seed, and Zhou Chang, the head of visual multimodal generation, discussed it, they finally chose to support Zeng Yan's idea."

This seemingly radical technical choice at the time eventually proved to be correct. "Because Zeng Yan insisted that the model must be large enough and the training data rich enough, Seedance 2.0 succeeded," the insider said.

We interviewed multiple AI practitioners inside and outside ByteDance, and most of them agreed that ByteDance's real reputation reversal in large models started with Seedance 2.0: In terms of large language models, the Doubao large model was not recognized by the industry as entering the domestic first-tier until it iterated to version 1.6 in mid-2025. Its coding capabilities had not made major breakthroughs before the launch of Doubao 2.1, lagging behind the flagship models of domestic startups such as GLM, Kimi, and DeepSeek at that time. Seedance 2.0 was almost ByteDance's first, and currently the only model that can form an absolute lead in performance.

This absolutely leading model is also the first model that truly made money for ByteDance. 36Kr previously exclusively reported that since 2026, Volcano Engine has been continuously raising its MaaS revenue target. Tan Dai, head of Volcano Engine, told 36Kr that more than half of Volcano MaaS's revenue this year comes from Seedance. 

ByteDance has hit a great business opportunity.

There are different claims about the profit margin of Seedance 2.0. LatePost previously reported that the gross profit margin of this model reached 70%, while multiple AI practitioners told 36Kr that they estimated Seedance 2.0's gross profit margin had reached 90% (although Volcano Engine's head Tan Dai said the outside world paid too much attention to Seedance's profits, and the actual margin was not that high). In any case, compared with large language models at this stage, video models are indeed a better business in China: they can support higher pricing and face less industry competition.

Moreover, multiple ByteDance business lines have formed a unique commercial flywheel based on Seedance: content platforms like RedBook Short Drama and Douyin have increased resource support for AI dramas compared to before — more content production companies purchase models from Volcano to produce high-quality AI short dramas — the surge in the number of AI dramas brings more ad or traffic investment revenue to RedBook and Douyin.

A positive cycle has thus formed.

Moments of stepwise improvement in model capabilities often unlock new productivity scenarios, and the flow of money will change as a result. This was verified once after the release of Claude Opus 4.6, and verified again with Seedance 2.0.

Seedance is not only ByteDance's first comeback victory in over three years since it entered the large model field, but it also made the entire Chinese large model industry see a very important fact: Yes, behind large models are sky-high capital expenditures, but they can also generate money with sufficiently high profit margins.

A Victory of Talent and Data

ByteDance has always "invested without counting costs" in AI — full-saturation investment in GPUs, data, and talent.

First is the full-saturation investment in talent.

For a long time, ByteDance has been good at "letting senior employees lead new businesses", and this idea was initially applied to Seed. But after more than a year, there was not much progress. In mid-2024, ByteDance began to place heavy bets on AI talent recruitment. This was also the first important change brought by the core management team deeply engaging in AI business: hunting for large model talents globally, even at the cost of extremely high salaries.

In mid-2024, a dedicated recruitment team serving AI business was established. The special feature of this team is that many HRs have strong business backgrounds, which allows them to better communicate with senior AI talents — both previous heads had led important businesses at ByteDance and had strategic work experience. ByteDance also paid very high salaries to this team — HRs responsible for high-end recruitment often had annual salaries exceeding one million yuan.

An "AI talent list" also began to take shape: top AI talents at home and abroad were all on it, "a few hundred people, updated dynamically at any time", said a high-end recruitment HR, "the Seed department hardly needs to follow ByteDance's standard salary system to issue offers. For anyone we want to recruit, the TP (total compensation package) can be customized."

In addition, Zhang Yiming returned to front-line work, frequently arranging to meet large model talents for "direct chats", and many senior researchers chose to join after seeing the generous compensation and sincerity.

The reason why "people" matter is that: "A responsible leader with experience and correct direction judgment, leading a team of smart young people to steadily conduct training experiments, with sufficient resources provided, it is difficult not to succeed", many industry insiders expressed this view.

With Zhou Chang and Wu Yonghui joining Seed one after another, internal "horse racing" in the same technical direction was reduced, and the judgments on technical directions and training routes became more focused and accurate, which was particularly evident in Seedance.

Previously, there were two independent teams at ByteDance training video generation models: Zeng Yan from AI Lab was working on PixelDance, and Jiang Lu from Seed was working on Seaweed. "At that time, ByteDance had separate teams with their own priorities, and several people working on video generation actually chose the wrong architecture in the early stage," said a person familiar with the early situation.

Take the first version of PixelDance as an example: it chose to extend the 2D UNet architecture to 3D, which can be understood as a video adaptation of an image diffusion model, rather than a native video architecture. At a time when the technical route for video generation was far from being finalized, this seemed like a faster and more stable direction, but later proved to have a low upper limit.

During the same period, Kuaishou's Keling team chose the same DiT (Diffusion Transformer) architecture as Sora, following the native video route, which has a higher upper limit but is more difficult. "If people and data do not hold back, the trained model will not perform poorly," an insider said. The time gap created during this process allowed Keling to take the lead for nearly a year afterward.

At the end of 2024, Jiang Lu left ByteDance to join Apple, and Zeng Yan joined Seed following the organizational restructuring of AI Lab, becoming the main person in charge of the visual generation model Seedance. This is a streamlined team, with only a dozen core algorithm engineers. From this point on, ByteDance's video generation model teams and technical routes began to converge.

The biggest change from the late PixelDance period to Seedance, compared to the earliest stage, was the adjustment of the training architecture: shifting from the UNet architecture to a direction based on DiT. This architecture can better implement the Scaling Law for large models — as the number of parameters, data volume, and computing power increase, the model performance will improve significantly.

However, Seedance 1.0 and 1.5 were not very successful: both versions had obvious gaps with Google Veo, ranking roughly fourth in global market share, and lagging behind Keling in China. A Volcano source said that ByteDance's main competitors during that period were Keling and Runway, "We targeted Keling and competed for quite a long time", but never caught up. A marketing head of a video Agent company described the feeling to 36Kr: Keling once occupied nearly 80% of the video generation market share.

The long-standing rivalry between ByteDance and Kuaishou has expanded from short videos, live streaming, and e-commerce to video models now. This is a battle ByteDance must win.

In December 2025, shortly after Seedance 1.5 was launched, the team began to invest in training version 2.0. This became Seedance's turning point for a comeback.

"Seedance attaches great importance to training data and model structure," a Seed insider said. Simply comparing the number of GPUs, ByteDance lags far behind OpenAI and Google in both quantity and quality. Overseas companies have generally used large clusters of B-series cards (large-scale AI computing clusters built with NVIDIA B200/B300 GPUs) to train video models, while within ByteDance, although Seedance has a high priority, it still ranks behind language models, and uses relatively inferior GPUs for training. Therefore, the team could only focus on optimizing the structure and data.

"After getting these two most important things right, the next step is continuous iteration." This returns to the industry consensus: with the right people, right direction, and sufficient resources, the rest is only a matter of time. And Seedance happened to have all of these.

Regarding "people and organization", a large model entrepreneur who has contacted the Seedance team told 36Kr, "Major technological innovation often requires talented researchers and a relaxed research environment to stimulate creativity. But once the technical route is determined, strict schedule management is needed to ensure that every step of training is done properly. The Seedance team really did every detail well, and its combat effectiveness in large-scale team operations is very strong."

The most critical part of Seedance 2.0's success is considered to be the data used for training. "This is a victory of data." Many practitioners came to this conclusion when talking to 36Kr about the reasons for Seedance 2.0's success.

The pre-training phase of large models requires a large number of small-scale experiments, which requires a lot of people to do work such as data standard formulation, synthesis, sourcing, annotation, cleaning, and evaluation, as well as a sufficiently large, rich, and diverse dataset.

"Seedance's data annotation work is very solid, which allows user input prompts to be accurately matched to specific data," an insider told us. Within ByteDance, there are thousands of people on the team responsible for model data evaluation for Seedance. In contrast, many leading startups in the video field consider an internal evaluation team of dozens of people to be a relatively large investment. Behind every Seedance algorithm engineer, there are often more than ten data colleagues providing support. This large data team must be able to quickly deliver the data required by algorithm engineers for experiments and training, as well as collect user feedback, to ensure that algorithm personnel can quickly capture any signals for model iteration.

There is also a dedicated person in the Seedance algorithm team responsible for coordinating with the data team. Multiple people close to Seed told us that this role is like a data product manager in the training team, which is very critical — this person clearly understands what kind of data training needs, can put forward clear and effective requirements to the data team, and even the algorithm engineers will construct data themselves and do fine-grained data cleaning work.

In terms of specific data selection, Seedance hardly used Douyin data, but purchased a large amount of film and television resources, especially a large number of high-quality film-level materials, and then used language models to break them down into scripts and storyboards. "The final effect is like putting a group of master filmmakers inside the model."

The team conducts targeted data training for different scenarios. For example, "Seedance 2.0 systematically learned from film and television works with rich running and jumping actions, because they can restore many real motion scenes; they also simulate or screen-record games for learning, and use professional scenario data to help the model learn various indoor spaces," a ByteDance data insider told 36Kr. Data from these various scenarios formed Seedance 2.0's huge training dataset.

In fact, not just for video models, ByteDance has made huge investments in the training data of all its models. According to 36Kr's understanding, for the training of world models and coding models, the data budget in early 2026 exceeded tens of millions of dollars, and "if it is not enough, the budget can be increased at any time."

One important reason for the huge investment in AI data is that Seed set the principle of "no distillation" from its very beginning. If all data has to be synthesized, purchased, and cleaned by itself, a huge team is inevitably needed to support this. "ByteDance's goal for almost all its models is to reach the global first tier, or even SOTA, which is difficult to achieve through distillation," a ByteDance insider told 36Kr.

Volcano's Explosive Growth

"The internal morale is extremely high," a Volcano salesperson described to 36Kr. Because of the sudden popularity of Seedance 2.0, the entire Volcano sales team went all out, "everyone is selling Seedance to customers." Faced with customer hesitation, some salespeople confidently say: "Within a year, our model will always maintain SOTA status."

In April 2026, Volcano Engine fully opened the sales of the Seedance 2.0 API to customers, with the condition that customers must sign an annual Seedance usage contract of at least 10 million yuan at one time to obtain the "full-power" usage qualification of 2.0. The so-called full-power version refers to the ability to handle high concurrency and support signed real-person portrait authorization, which are two critical points for Seedance 2.0's main user group — AI content production companies that own real-person IPs and need mass-produce content.

This practice has never appeared in Volcano Engine's previous MaaS sales plans. After DeepSeek R1 was released in February 2025, the price of large language models dropped sharply, and competition among model vendors quickly entered a buyer's market: for buyers, they would use whichever model works well and is cheap. Competing for market share at rock-bottom prices left model vendors with almost no bargaining power. For Volcano Engine, which mainly sells its own models, it was even more impossible to set such a high sales threshold before.

"In 2025, Volcano had a limited number of models to sell, and followed a cost-effective route," a Volcano Engine insider told 36Kr: "Language model competition was too fierce, and there were no high-value scenarios like coding that had taken off yet. The Seedream image model was relatively competitive in China, but its usage volume was not large." Because Seedance 1.0 and 1.5 did not perform well enough, they could not compete with Keling in the market at that time, and not many customers paid for them.

"By early 2026, we put all our hopes on Seedance 2.0. We had to come up with a new story that we could tell customers." According to the aforementioned Volcano Engine insider, they heard internally that Seedance 2.0 was a competitive model, and thought it might have a leading edge of one or two months, "but we didn't expect it to be so competitive. After several months, our leading advantage is still there."

The first wave of feedback that the model performed well came from C-end users. During the 2026 Spring Festival, ByteDance's C-end AI products such as Doubao, Jianying, and Jimeng fully integrated Seedance 2.0. Traffic flooded in like a tide, leading to a situation where generating a 12-15 second video even required waiting in queue for ten hours.

Jimeng, which responded quickly, immediately launched a paid option: subscribing to membership and purchasing points can grant priority queue access — users who could not stand waiting contributed the first batch of revenue for Seedance 2.0. According to estimates from a large company's strategic team to 36Kr, Jimeng's revenue in March was about 140 million yuan, and reached 210-220 million yuan in April, the vast majority of which came from Seedance 2.