The most mysterious company in the AI circle has finally made its "high-stakes gamble" public.

Text | DENG Yongyi

Edited by | ZHANG Yuxin

What on earth is Vivix doing? This has been the biggest curiosity in the AI circle about this mysterious company for more than the past year.

The impression Vivix gives to the outside world is roughly as follows: Liu Yu, a star entrepreneur, has been aggressively recruiting talents across all business lines from algorithms, products to brands. The company's valuation has soared to 1.32 billion US dollars within a year, with a group of investors with impressive backgrounds standing behind it.

But few people can clearly state what exactly this company is doing. Some say it is a pure AI product company that does not develop models, aiming to build the "Douyin of the next era", which is why it has frequently launched interactive entertainment products such as 7verse and TipTap. If you search for Vivix during this period, the only visible introduction to its technical direction is the "next-generation visual generation engine" written on its official website.

During the 19 months since Vivix was founded, the level of competition in the multimodal field has been no less intense than that in the large language model (LLM) sector, and wealth-creating success stories like Seedance have emerged.

"Are we that mysterious?" When we first met Liu Yu, the founder of Vivix, in July, he seemed a little insensitive to various comments from the outside world. "We look mysterious because there is no need for PR when the model is not fully ready yet."

Liu Yu said that all previous products of Vivix are explorations made to verify the model, and what he wants to do has never changed: to build the next-generation real-time interactive model.

The industry's attention to this company largely comes from a number of remarkable labels: Liu Yu, the founder, was born in 1995, is a talented talent from the solid SenseTime system, and has experienced the complete training closed loop of large-scale AIGC and multimodal interactive models. As early as 2023, he led the team to develop a To C AI application called "Miaohua", which gained millions of users in the first month after its launch.

An insider once told us that at that time, there were no more than 5 people in SenseTime who had both algorithm and project experience in the visual field, and Liu Yu, who had led a team of hundreds of people for basic models and managed more than 4,000 GPUs, was one of them.

The huge influence that Vivix has made in the capital circle largely stems from the entrepreneurial direction Liu Yu chose.

In 2024, when the narrative of large language models cooled down, OpenAI Sora on the other side was setting off a wave of video generation models. The route Liu Yu bet on carries the market's expectations for the form of the next-generation multimodal model.

At that time, most video model manufacturers were fully committed to improving video image quality, extending duration and enhancing physical consistency. Liu Yu, who led the entire SenseTime research team at that time, earlier took generation speed and model interactivity as independent technical goals, and explored using consistency models to reduce the number of sampling steps for video generation.

Liu Yu believes that interactive content is a brand-new content category that only emerged in the AI era. If a model can only generate audio and video, it can only reduce costs and increase efficiency for the production end of mature content formats of the previous era. The original motivation for his entrepreneurship is that in the foreseeable future, models will have the opportunity to reach the inflection point where interactive content can be "consumed".

All model training of Vivix is carried out around this goal. "If we want to build a underlying model for interactive content, first, the generation must be fast enough, and second, the cost must be low enough." Liu Yu said.

This also explains why at his first public communication meeting, the benchmark Liu Yu showed to the outside world was not a certain indicator of top-tier models, but the CDN bandwidth cost of YouTube, the giant video platform — which is the basic cost red line for content platforms of this generation.

In July, Vivix will launch two real-time interactive models W1 and A1. W1 is a base model integrating understanding, interaction and generation. A1 is a real-time interactive model centered on character interaction, targeting specific application scenarios such as digital human/character interaction.

Among them, the real-time generation cost of the W1 model is already close to YouTube's CDN bandwidth cost of about 0.2 US dollars per hour. Compared with the current top video generation models, the generation cost of real-time interactive content can be 1-2 orders of magnitude lower.

In terms of speed, when a user operates in an interactive scenario, the model can generate the next frame of picture within 300-600 milliseconds (that is, TTFF, Time to First Frame). This enables the model to truly achieve the next step of personalized distribution — personalized generation.

△ Vivix-W1 demonstrates real-time director capability, supporting real-time changes to the story direction through text, voice and touch input

Vivix initially chose the scenarios of games and live streaming. This comes from the characteristics of these two scenarios: games are many-to-many interactions, while live streaming is one-to-many distribution, with the highest consumption density per token, and higher upper limits of generated GMV and customer unit price.

Real-time interaction further unlocks the possibilities of these scenarios. In the demos co-created by Vivix and some game studios, players can already control the plot progress in real time through Prompt, making a story truly an open-world setting — Liu Yu regards the sense of narrative as one of the core indicators for measuring models. Content that can be interacted with and brings a sense of participation is the core of the next-generation content.

△ Vivix-A1 demonstrates motion performance and interaction with objects and the environment

To match the new content format, all existing "infrastructure" needs to be rebuilt. An essential reason is that "not everything in the world can be expressed through language, images and videos." Liu Yu said, "We need to throw the original raw data of human interaction to the model and let the model understand it on its own."

In the 19 months since its establishment, the Vivix team has almost reworked everything from front-end data collection, data pipeline construction to infrastructure and models. Currently, the size of Vivix's team has reached 150-160 people.

Compared with the ambitious journey Liu Yu has embarked on, he is not an entrepreneur who loves to talk about visions and values, and he is also afraid that people will roughly regard real-time interaction as just another world model. "I just think the concept of world model is too sacred, we are not qualified to talk about it yet." he said.

According to Liu Yu's plan, W1 to be launched in July is a medium-scale version, with the core purpose of verifying that the integration of understanding, interaction and generation can work and be generalized. In Q4 this year, Vivix will also complete the training of W2 — the data volume is two orders of magnitude higher than that of W1, covering all imaginable natural video interaction scenarios.

It can be said that the ultimate goal Vivix wants to achieve is still relatively long-term and extremely difficult. On the one hand, real-time interaction puts higher requirements on computing power and infrastructure — humans consume 5-10 words of language tokens per second, while the token volume of visual content may be thousands to tens of thousands. On the other hand, although real-time interaction has become a consensus for exploration among next-generation model manufacturers, the large technical curve has not yet converged, which requires long-term investment.

However, Liu Yu feels that this has come to his comfort zone instead. "Consensus means mediocrity," he said. Since he was a child, whether it was giving up the admission to pursue a doctoral degree to start a business at SenseTime, or starting internal entrepreneurship at SenseTime all the way to the present, most of the achievements he has made come from non-consensus.

The following is the self-narrative of Liu Yu, founder of Vivix, sorted out by Smart Emergence:

No plans to be a product company 

Since our establishment, we have developed several product prototypes, 7Verse, TipTap, and now it is already the fifth product prototype. All these products only aim to test one clear question: what kind of content format and interaction method are truly consumable.

This is the first-principle question in this field. All companies doing AI entertainment, AI content and interactive content must answer: What kind of AI content is more willing to be consumed than Douyin's content?

We underestimated this question at the beginning, thinking that defining interactive content would work, and everyone would like it as long as we build it. After we actually developed several product prototypes, we found that this is extremely difficult.

It is difficult to define interaction, and it is also difficult to make content have consumption power, and both of these difficulties ultimately return to model capabilities. So products are not our top priority. In the foreseeable future, we will still focus on models, and our core business is still selling model APIs.

Developing these products is also an attempt to point out some directions for the market. At that time, we developed Sofia, the first digital human that can interact in the real world, and found that many otome game companies and AI companion companies also want to do this.

They use our API, and the front-end harness can be completed in two or three days — using Claude Code to build the entire front-end and data system. What customers precipitate is their own harness — how to sell this product, how to set the personality, how to adjust the game mechanism and numerical system. The underlying engine is provided by us.

The Scaling Law for multimodality has emerged 

Before starting my business, I was responsible for training all the underlying basic models of SenseTime. The most famous business line of SenseTime is face recognition. We have about 30 to 40 face recognition business lines — face scanning in subway stations, person-ID verification, mobile phone face unlocking, passive security. We found that even for face recognition tasks, each sub-scenario needs to be trained with a separate model. If you gather all the data to train one model, its accuracy in each scenario will drop.

But this generation of large models is completely different. The self-attention mechanism in the Transformer architecture endows the model with inherent cross-domain generalization capabilities — the more data and the more diverse the scenarios, the stronger it becomes. A single GPT model can do translation, companionship, Q&A, coding and so on, and the marginal cost between these scenarios is almost zero.

In the first half of 2025, we found more than a dozen data sources similar to what next token prediction means to LLM — we call it next reaction prediction. It is essentially a ternary combination, including data of several dimensions: what happened in the past (understanding); how humans interact (interaction); what will happen in the next second (generation).

Games are naturally interactive content. In all game screen recordings and simulation environments, replays can trace the entire game from beginning to end only by recording the initial state and the operation methods of all people. These data naturally show how humans interact with content, why they interact, and what changes will happen to the content after the interaction.

We have tens of millions of hours of carefully screened multimodal data, some of which have interactive attributes. The corpus size of top general video generation is roughly at this level.

This point will also affect my thinking about the Scaling Law of multimodal interactive data.

In the era when we did machine learning training, people needed to label ground truth manually. For example, face recognition needs to label who this person is, and autonomous driving needs to draw bounding boxes to label where the car is.

In the LLM era, the reason why Scaling Law exists is that the Next Token Prediction of natural language is a natural self-supervised "labeling", which does not require any human labeling. In the field of interactive intelligence, we have also found the naturally self-supervised ternary structure of "understanding-interaction-generation", so there is a scalable data paradigm.

Scale data first, then scale parameters and frameworks 

If you look at the demand from C-end users, first, they are hardly willing to pay for tokens; second, their patience for content generation speed is limited. Therefore, if we want to build a underlying model for interactive content, it must have two features — fast enough, and cheap enough.

Cost, speed and effect form an impossible triangle. What we are competing for is not who has lower cost, who is faster, or who has better image quality, but in a four-dimensional space — combining cost, speed, quality and content consumption power, to see whose model and system can go the farthest to the top right corner.

For truly C-end content products to emerge in the AI era, the cost side must break the current cost barrier of utility models. Only when the cost reaches the level of the content cost we spend on browsing Douyin and watching YouTube, can real-time and interactive functions be realized.

The inference cost of our current model is already very close to this state.

Our benchmark is YouTube. When users watch videos on YouTube, the cost they pay for global network CDN and Internet bandwidth is about 0.2 US dollars per hour.

From March 2025 to April this year, the inference cost of Vivix's models has been reduced by two orders of magnitude, and this not only includes generation, but also the understanding of videos and multimodal interactions.

Let's talk about speed. The core indicator I set for the infrastructure team at the beginning of the founding is TTFR (Time to First Reaction) — how long it takes to generate the corresponding reaction content after a person gives a reaction (such as writing a prompt, or tapping on the screen).

Now for the TTFF (Time to First Frame) of model inference, it takes us about 300 to 600 milliseconds, and the full-link TTFR including network communication takes about 1 second. In contrast, the inference time required for top video models to generate video content based on a prompt is hundreds to thousands of seconds.

Considering this is a model with about 30B effective activated parameters and a VAE compression rate of only 8-8-4, it is extremely difficult to achieve this delay under the integration of understanding, interaction and generation.

In July, we will release W1. W1 is a medium-scale model, the core of which is to verify that the integration of understanding, interaction and generation can work and be generalized — a bit like the relationship between GPT-1 and GPT-2. Last month, we started the scaling of W2. Compared with W1, the data volume of W2 is two orders of magnitude larger. We plan to complete the training of W2 and release it to the public in Q4.

From W1 to W2, what we need to scale is data. From W2 to W3, we need to scale the model structure and MoE (Mixture of Experts). For us, we are currently in the stage of scaling data — feed the model sufficiently first, then scale parameters and frameworks after that. Every type of large model — language model, video generation model, audio model, VLM, VLA, including our model integrating understanding, generation and interaction — all need to go through this stage.

The missing part now is: how to integrate all the modalities that humans can naturally interact with models into the native learning of the model.

The upper limit of the model's intelligence should not be bound by human natural language. The understanding of how humans interact with content should not be converted into text first. My current approach is to throw the original raw data of human interaction with the model to the model, and let it understand by itself.

We are not competing with Seedance 

I wanted to do this back in SenseTime in 2022, and it has not changed until now.

Many investors ask us every time they meet us for the first time, how do you compete with Seedance? I said, why should we compete with Seedance?

Many domestic companies that make video tools and