Vivix Founder Liu Yu Discusses Real-Time Interactive Models: W1 and A1 to Be Released in July

By Deng Yongyi

Edited by Zhang Yuxin

What exactly is Vivix doing? This has been the biggest curiosity in AI circles about this mysterious company over the past year or so.

The impression Vivix gives the outside world is roughly this: star entrepreneur Liu Yu has been aggressively hiring across all business lines, from algorithms and products to branding; the company's valuation soared to $1.32 billion in a year; and it is backed by a roster of investors with glittering backgrounds.

But few can clearly explain what the company actually does. Some say it is a pure AI product company that doesn't build models, aiming to create "the TikTok of the next era," which is why it keeps launching interactive entertainment products like 7verse and TipTap. If you searched for Vivix during this period, the only visible description of its technical direction was "next-generation visual generation engine" on its official website.

In the 19 months since Vivix was founded, competition in the multimodal field has become no less intense than in large language models (LLMs), and wealth-creating stories like Seedance have already emerged.

"Are we really that mysterious?" In July, when meeting Vivix founder Liu Yu for the first time, he seemed somewhat numb to the various outside evaluations. "It looks mysterious because when the model isn't very ready yet, there isn't much need for PR."

Liu Yu said that Vivix's previous products were all explorations made to validate the model, and what he wants to do has never changed: the next generation of real-time interactive models.

The industry's attention to this company largely comes from some shiny labels: founder Liu Yu was born in 1995, is a bona fide SenseTime alumnus, and has been through the complete training loop of large-scale AIGC and multimodal interaction models; as early as 2023, he led a team to build a To C AI application called "Miaohua," which gained millions of users in its first month online.

A person familiar with the matter once told us that at the time, in the visual field at SenseTime, there were no more than five people with both algorithm and project experience, and Liu Yu—who had led a hundred-person foundation model team and held more than 4,000 GPUs—was one of them.

And the stir Vivix has caused in capital circles largely stems from the entrepreneurial direction Liu Yu chose.

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

At the time, most video model vendors were going all in on improving video quality, longer duration, and stronger physical consistency; Liu Yu, who was then leading the entire SenseTime research team, had earlier made generation speed and model interactivity independent technical goals, exploring the use of consistency models to reduce the number of sampling steps in video generation.

Liu Yu believes that interactive content is an entirely new content category that emerged only in the AI era. If a model can only do audio and video generation, then it can only reduce costs and increase efficiency for the production side of mature content formats from the previous era. The source of his entrepreneurship is precisely this—in the foreseeable future, models have a chance to reach the inflection point where interactive content "can be consumed."

All of Vivix's model training is carried out around this goal. "If we want to build an underlying model for interactive content, first, generation must be fast enough; second, it must be cheap enough," Liu Yu said.

This also explains why, at his first external communication meeting, the benchmark Liu Yu showed the outside world was not some metric of a top-tier model, but the CDN bandwidth cost of video platform giant YouTube—the baseline red line on cost for this generation of content platforms.

In July, Vivix will launch two real-time interaction models, W1 and A1—W1 is a foundation model integrating understanding, interaction, and generation; A1 is a real-time interaction model centered on character interaction, aimed at 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 per hour, and 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 within 300-600 milliseconds (i.e., TTFF, Time to First Frame, the time to generate the first frame). This enables the model to truly achieve the next step after personalized distribution—personalized generation.

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

The scenarios Vivix initially chose are gaming and livestreaming. This stems from the characteristics of these two scenarios: gaming is many-to-many interaction, while livestreaming is one-to-many distribution; they have the highest token consumption density per unit and also a higher ceiling for generated GMV and average order value.

Real-time interaction further opens up the possibilities of these scenarios. In demos co-created by Vivix and some game studios, players can already control plot progression in real time through prompts, truly turning a story into an open-world setting—Liu Yu regards having a sense of story as one of the most core metrics for evaluating models. Content that can be interacted with and offers a sense of participation is the biggest core of next-generation content.

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

And to match the new content format, all existing “infrastructure” needs to be rebuilt. One fundamental reason is that “not everything in the world can be expressed through language, images, and video.” Liu Yu said, “We need to throw the native raw data of human interaction to the model and let the model understand it on its own.”

In the 19 months since its founding, the Vivix team has redone almost everything from front-end data collection and data pipeline construction to infra and models. Currently, Vivix’s team size has reached 150-160 people.

Compared with the grand journey Liu Yu has embarked on, he is not an entrepreneur who likes to talk about vision and values, and he is also afraid that everyone will crudely understand real-time interaction as yet another world model. “I just think the world model is too sacred, and we are not yet worthy of talking about it,” he said.

In Liu Yu’s plan, W1, launched in July, is a medium-scale version, with the core goal of verifying that the integration of understanding, interaction, and generation can run successfully and can generalize. In Q4 this year, Vivix will also complete the training of W2—with a data volume two orders of magnitude higher than W1, covering every imaginable natural video interaction scenario.

It can be said that the ultimate goal Vivix wants to achieve is still relatively far off, and extremely difficult. On the one hand, real-time interaction places higher demands on compute and infra—humans consume language tokens at a rate of 5-10 characters 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 an exploration consensus among next-generation model vendors, the major technology curve has still not converged and requires long-term investment.

However, Liu Yu feels that this has instead brought him into his comfort zone. “Consensus means mediocrity,” he said. From childhood to now, whether giving up doctoral enrollment to start a business at SenseTime, or repeatedly starting new businesses internally at SenseTime all the way to now, most of the achievements he has made have come from non-consensus.

The following is a self-account by Vivix founder Liu Yu, compiled by Zhineng Yongxian:

Not planning to be a product company

Since our founding, we have made several product prototypes, 7Verse, TipTap, and now the fifth product prototype. All these products are only meant to test one question clearly: what content format and interaction method are truly consumable.

At first we underestimated this problem and thought that defining interactive content would work—as long as we made it, everyone would like it. But after actually making several product prototypes, we found that this is too difficult.

It is hard to define interaction, and it is also hard to determine how content can have consumption power, and both of these difficulties ultimately come back to model capability. So products are not our first priority. In the foreseeable future, we will still focus mainly on models, and the core is still selling model APIs.

Making these products is also an attempt to give the market some direction. At the time, we made Sofia, the first digital human that can interact in the real world, and found that many otome game companies and companies making AI companionship also wanted to do this.

They use our API, and the front-end harness can be finished in two or three days—using Claude Code to build the entire front end and data system. What customers accumulate is their own harness—how this product should be sold, how the persona should be set, and how the game mechanics and numerical system should be adjusted. The underlying engine uses ours.

The Scaling Law for multimodality has already emerged

Before starting the company, I was responsible for training all of SenseTime’s underlying foundation models. SenseTime’s most famous business line is face recognition. We had about thirty or forty face recognition business lines—subway face scanning, identity verification, mobile phone face unlocking, passive security. We found that even if they are all face recognition tasks, each segmented scenario requires training a separate model. If all the data is brought together to train one model, its accuracy in each scenario will decline.

But this generation of large models is completely different. The self attention mechanism in the Transformer architecture gives the model a natural cross-domain generalization ability—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 across these scenarios is almost zero.

In the first half of 2025, we found more than a dozen data sources similar to next token prediction for LLMs—we call it next reaction prediction. It is actually a triplet combination, including data from several dimensions: what happened in the past (understanding); how humans interacted (interaction); and what will happen in the next second (generation).

Games are naturally interactive content. In all game screen recordings and simulation environments, replay only needs to record the initial state and everyone’s operations to trace the entire game from beginning to end. This data naturally shows how people interact with content, why they interact, and what changes happen to the content after the interaction.

We now already have tens of millions of hours of finely filtered multimodal data, some of which has interactive attributes. The corpus for leading general video generation is roughly of the same order of magnitude.

This will also affect my thinking about the Scaling Law approach for multimodal interaction data.

Back when we were doing machine learning training, people had to create ground truth labels—for example, face recognition had to label who this person was, and autonomous driving had to draw boxes to mark where the car was.

In the LLM era, the reason there is a Scaling Law is that Next Token Prediction on natural language is a naturally self-supervised "label" and requires no human annotation. In the field of interactive intelligence, we have also found a naturally self-supervisable triplet structure of "understanding-interaction-generation," so there exists a data paradigm that can be scaled.

Scale data first, then scale parameters and frameworks

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

Cost, speed, and quality form an impossible triangle. What we need to compare is not who has lower cost, who is faster, or who has better image quality, but in a four-dimensional space—cost, speed, quality, and content consumption power combined—who can push their model and system farthest toward the upper right corner.

For true to-C content products to emerge in the AI era, the cost side must break through the current cost barriers of tool-type models. Only when the cost can reach the level of the content cost of scrolling Douyin or watching YouTube can real-time and interactive experiences become possible.

Our current model inference cost is already very close to this state.

Our benchmark is YouTube. When users watch videos on YouTube, the cost they pay for the global CDN and internet bandwidth is roughly $0.2 per hour.

From March 2025 to April this year, Vivix's model inference cost dropped by two orders of magnitude, and this includes not only generation but also understanding of video and multimodal interaction.

Now for speed. At the beginning of the company, the most core metric I set for the infra team was TTFR (Time to First Reaction)—how long after a person gives a reaction (for example, writes a prompt or clicks on the screen) the corresponding reaction content is generated.

Today, the TTFF (time to first frame) of model inference takes us roughly 300 to 600 milliseconds, and including network communication and the full chain, TTFR takes about 1 second. In contrast, leading video models need hundreds to thousands of seconds of inference time to generate video content based on a prompt.

Considering that this is a model with about 30B effective activated parameters and a VAE compression rate of only 8-8-4, achieving this latency while integrating understanding, interaction, and generation is extremely difficult.

In July, we will release W1. W1 is a medium-scale model, and its core purpose is to verify that the integration of understanding, interaction, and generation can work and generalize—somewhat like the meaning of GPT-1 and GPT-2. Last month we already started scaling W2, and compared with W1, W2 has two orders of magnitude more data. We plan to complete W2 training in Q4 and release it externally.

What W1 to W2 needs to scale is data, and what W2 to W3 needs to scale is model structure and MoE (Mixture of Experts). For us, we are now in the stage of scaling data—first feed the model fully, and after it is fully fed, then scale parameters and frameworks. Every type of large model—language models, video generation models, audio models, VLM, VLA, including our integrated understanding-generation-interaction model—needs to go through this stage.

What is relatively missing now is: how to incorporate all the modalities through which humans naturally interact with models into the model's native learning.

The upper limit of a model's intelligence should not be constrained by human natural language. The understanding of how humans interact with content should also not first be converted into text. My current approach is to throw the native raw data of human-model interaction to the model and let it understand on its own.

We are not competing with Seedance

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

Many investors, every time they meet us for the first time, ask how you compete with Seedance. I say, why compete with Seedance?

Many companies in China that make video tools and video models will inevitably have to compete with ByteDance if they choose this track. But first, it is hard to beat them in model capability; second, it is hard to beat the big companies in customer acquisition models and ecosystem, so it is very difficult to win.

All companies making video generation models, such as products for short dramas and for short video creation, are reducing costs and increasing efficiency on the production side for the content formats of the mobile internet era. Any industry that takes cost reduction and efficiency improvement as its starting point and in which China has advantages will ultimately have very low profit margins.

Capabilities that cross tracks are a form of dimensional strike, not direct competition. Just as using expert systems for RAG and using large language models for knowledge base retrieval are both doing knowledge bases, but the technical paradigm has changed, there is no competition—everyone will simply migrate over.

Essentially, we are creating new increments. Right now, everyone making short dramas is doing it offline, and very few people are making real-time interactive short dramas, because there are no corresponding models and the cost is very high. In the AI era, only the emergence of new models can breed new application scenarios. So all our approaches are to continuously add capabilities at the model layer, and then push down how the corresponding infra should be built and how application scenarios can be built.

Startups pursue odds, big companies pursue win rates

Big companies will certainly also build interactive content, but the way they do it, their motivation, and their patience are completely different from those of startups. The decision-making mechanisms of big companies lead to an endless pursuit of probability—if something has a sufficiently high probability, they will commit sufficiently large resources.

The core of a startup is the pursuit of odds. My own understanding of entrepreneurship is this: those who pursue odds find a direction with high odds where a cognitive gap creates a probabilistic advantage.

And before this probabilistic advantage grows to the point where big companies are willing to commit enough resources, you need to build your own data flywheel—that is, sufficiently strong momentum and sufficiently high barriers. Otherwise, without unique insight, a startup that competes with big companies on win rate at the same probability has no chance at all.

In the previous era, Adobe was able to dominate alone because its product features were complex enough—I'm a photography enthusiast, but I've never clicked 70% of the buttons in Photoshop, yet precisely because of this, all workflows were completed within it, with no need to go to other products.

Why was commercialization so poor in the AI 1.0 era? The reason is poor technological generalization: you had to do high-price, high-margin work to subsidize R&D. If I build a model for this company, it can only be used by them and cannot be used by another, because performance would degrade.

The underlying variable of this generation of real-time interaction technology is that model generalization is strong enough, which means a change in business models—the marginal R&D cost of replicating capability is zero. There is no longer a need to rely on the high gross margin of one customer to subsidize the R&D cost of customizing models for another customer. The money we should earn is this added value brought by model-based scaling.

Looking further into the future, there will definitely be more than one company with model training capabilities, and model capabilities will ultimately converge. The business model of homogeneous models will eventually move toward a utility model, the market will enter stock competition with transparent channels, and gross margins will decline. So we need to create added value in scenarios and channels during this process.

But in this era, if you want to reach the first tier of models, you cannot rely on just one genius idea; you must have very strong systematic combat capability—very strong data infrastructure, training infra, and data algorithm capabilities. The devil of ultra-large-scale training tasks is in the details: give the same 10,000 GPUs and 1B raw data to different teams, and the results may differ vastly.

For example, the reason Seedance is so impressive, in my own experience, comes down to two points. First, its legion combat capability is strong enough—there may not be much fancy innovation in data, models, or algorithms, but every detail is done well, and the model team works hard enough; second, automatic storyboarding and prompt following are done well, greatly improving product usability.

In the past two years I have met quite a few people from ByteDance, and I have always been curious about why ByteDance can be like this—no matter who is replaced, the whole team continues to pursue the business with the highest combat effectiveness.

Why does this form of large-legion combat work today? Because in the video generation track at the time, DIT and MoE were already consensus, and there was no need for overly fancy innovation. At that point, what was needed instead was legion combat, pushing details into place step by step—this is where organizational advantages can come into play.

Today, using AI to produce content is still in an early stage. The ARR of the entire video generation track has grown dozens of times compared with last year, and in the coming years it is expected that this track will still see order-of-magnitude growth. Today we are betting on using models to directly change content formats; this track is earlier, more unknown, and also more interesting.

I am someone who is especially afraid of consensus

Many people ask me how I view world models—as if once everyone starts doing it, it becomes consensus? But I don't think so. Consensus means your understanding is consistent with everyone else's, but the starting points from which people are doing this now are obviously still inconsistent.

On a track that has not yet been established, it is wrong to compete over a mistaken consensus metric. Take digital humans: everyone competes over whose lip-sync is better and whose hands are more natural, but I think that is not very essential. I like watching Luo Yonghao's livestreams because he speaks humorously, and I like watching Zhang Xuefeng talk about the college entrance exam because he can explain the gray adult world to me. These are the real focus of consumption.

In my own growth, the positive feedback I received was never brought by consensus. Almost all of my work has required entrepreneurial spirit—starting a business while studying at CUHK, doing research at Microsoft, doing research at Google Brain, going from intern to research director at SenseTime, and then leading a business unit. All of it was from 0 to 1.

In 2015 I applied to a pretty good school for a PhD, and the visa was even processed, but Professor Tang Xiaoou hoped I would stay and start a business together. That year I was 20, and I chose to give up the PhD and stay at SenseTime. Looking at it today, I don't regret it at all.

Consensus means safety, but it also means mediocrity. Since childhood I have been someone who does not particularly pursue safety; safety gives me a great sense of insecurity.

World models are too sacred; we are not yet worthy of discussing them

Some people ask me whether I worry that everyone will define us as a world model company. I do mind that.

I don't really want to talk about world models. On the one hand, the term has been talked about too much and its definition is overused. But the deeper reason is that I think it is too sacred, and we are not yet worthy of discussing it.

One of my views is that multimodal models with language as the hub will not become something that can truly represent this world.

First, the corpora used to train language models all come from the internet—from what was written by those people who once had the ability to leave their voices and records on the internet. These corpora are inherently biased.

I myself am relatively interested in the frontiers of brain neuroscience and high-energy physics. If ChatGPT could really help me understand the most cutting-edge knowledge in these disciplines, it should be a world model. But I find that it is very biased.

Furthermore, suppose the corpus has no bias. Every line a person writes is essentially an electrical signal generated by neural circuits in the brain, which becomes text at the language center. And these electrical signals originate from what he has seen, heard, and thought throughout his life—essentially his sensory system. Vision is essentially electromagnetic waves; hearing, smell, and touch are, in the final analysis, also electromagnetic forces. And this is only one of the four known fundamental forces that constitute this world; it cannot represent the entire world.

All human cognition is some kind of abstraction of perception. Humans' definition of a "world model" also cannot surpass humans' own perception of the "world." For example, on the electromagnetic spectrum, the human eye can perceive only visible light, and so there is a corpus format that records the visible-light band called "video"; and so being able to predict, based on action, the content of the next moment's "video" is defined as a "video world model"—this is narrow. The text that humans have coalesced originates from such a narrow observation of the world.

If humans had not discovered X-rays and, relying only on videos recorded in visible light, said "this can represent the world," wouldn't you find that very absurd? Perhaps in the future there will be Y-rays, Z-rays, or a new fifth fundamental force that can bring us a more essential physical understanding of the world. Then wouldn't the current "physical world model" defined based on visible light and the like seem somewhat narrow?

The "world" is itself defined by humans based on limited observation. And humans are narrow; the things humans define must be even narrower than humans themselves.

Once I thought this through, I actually felt relieved. If we understand world models in a narrow sense, it is actually a kind of functionalism: it can help humans predict what will happen in the next frame, and can make what humans see look somewhat like what people filmed. This "world" is a world meant to satisfy human practicality, not the real, objective world.

After AI agents save humans a great deal of time and generate a great deal of productivity, the time humans have left—their spiritual consumption—should have a more efficient way to be filled. If I can achieve this step, I will be satisfied.

This article comes from the WeChat public account "Intelligent Emergence"; author: Deng Yongyi; published by 36Kr with authorization.