Liu Yu unveiled Vivix for the first time
Liu Yu, born in 1995, is fascinating enough — he graduated with a doctorate from the Multimedia Laboratory of the Chinese University of Hong Kong, and at the age of 26, he became one of the youngest Executive Research Directors and General Managers of Business Units in SenseTime's history. Around 2022, the under-27-year-old Liu Yu managed a team of nearly 200 people, with access to thousands of GPUs of SenseTime.
In October 2024, news that Liu Yu intended to leave his job to start a business spread quickly, sparking a wave of curiosity and uproar. But it wasn't until nearly a year later that the outside world faintly heard that he had founded a company called Vivix. During that period, many investors around were inquiring about this name.
In July this year, Liu Yu, who rarely appeared in public, had an exchange with the investment community.
During the year of silence, Vivix has always focused on real-time interactive content, and its strategic focus has gone through a painful yet rewarding detour from model capabilities, to product definition, and then back to model capabilities. Liu Yu has mentioned more than once: The opportunity for interactive content does not lie in tools or the product side. The richer the modalities, the more immersed consumers will be.
It was during this almost silent period that Vivix secured five rounds of financing, with investors including IDG Capital, HSG, BlueRun Ventures, 5Y Capital, Monolith, JD.com, Alibaba and others, with an overall valuation of 1.32 billion US dollars (approximately 9 billion RMB).
Faced with investors rushing to place bets, Liu Yu would remind them over and over again: "It's quite difficult to do this." But he would also say that Vivix is moving very fast on this path.
Who is Liu Yu?
A post-95s PhD with nine years at SenseTime
Almost all entrepreneurs have been asked a question: Why do you want to do this? The answer is often intertwined with personality and experiences.
It's hard to imagine that Liu Yu, who has published nearly 100 top-tier AI conference papers, once resold smartphones. He thinks that selling products he believes are right to others is a very fulfilling thing, especially when the other person suddenly realizes: "Oh, I didn't know such a thing exists in the world."
In 2014, when Liu Yu was a junior, he participated in a technology competition and developed a drone flight control system that automatically tracks people, using 3D convolutional neural networks, video understanding and real-time tracking. At that time, DJI Phantom 4 had not been released. This year, Liu Yu received the only Microsoft Scholarship in the whole school and went to intern at the Visual Computing Group that was led by Professor Tang Xiao'ou in the early days.
He joined SenseTime at the end of 2015. It was the second year after SenseTime was founded, having just completed its first round of financing, and the team only had dozens of people. Few people know that Liu Yu contributed to SenseTime winning the ImageNet World Championship back then. At that time, while still pursuing his PhD, he flew back and forth between Beijing and Hong Kong almost every week — returning to the Chinese University of Hong Kong on Fridays to submit assignments, and coming to Beijing to work on Mondays. A few years later, Liu Yu was quickly promoted to Executive Research Director and General Manager of the Innovation Business Unit.
A small detail is that during his first five years at SenseTime, although Liu Yu led a team of a certain scale, he received a salary close to that of an intern because he had not yet graduated.
But Liu Yu didn't care at that time. "Whether working at a company or starting a business, it's about finding something that makes you feel consistent with yourself and then doing it right." To get things right, Liu Yu would think about what customers on the downstream business line really want, and where the foundational model team needs to improve. Being able to solve problems ultimately is the source of his sense of achievement.
Intersections in the AI circle are always wonderful.
Around 2022, Yan Junjie resigned to start a business, and the business unit he managed was merged into Liu Yu's jurisdiction. His team suddenly expanded to about 200 people, spanning three sectors including foundational models and the AI Game Business Unit. At that time, 27-year-old Liu Yu managed thousands of SenseTime GPUs and promoted SenseTime's AI content creation community platform, SenseMirage.
It was also during this period that Liu Yu began to realize that not everyone needs tool-based and efficiency-enhancing products, and real-time streaming models can best satisfy content interaction and entertainment.
The news of his departure spread in October 2024. But for the following year, Liu Yu's updates were quiet, and occasional doubts popped up on the Internet, for example, what is Liu Yu doing?
It wasn't until the end of 2025 that the outside world accidentally discovered through Liu Yu's academic personal homepage that he had founded a company called Vivix with a valuation of 1.32 billion US dollars, and the outside world's doubts further turned into: What is Liu Yu doing after receiving several rounds of financing?
In July 2026, Vivix released some answers to the outside world. Over the past year or so, the internal team has iterated more than a dozen versions of models around "interactive content", launched four or five product prototypes overseas, and built a solid training and inference infrastructure.
A more iconic achievement is the two models he quietly polished: one is Vivix-A1, the first real-time full-duplex model for interactive AI characters, which enables characters to continuously perceive the environment, interact with the environment, take actions and respond in real time in an open world; the other is Vivix-W1, a real-time multimodal large model centered on interactive storytelling, which integrates joint audio and video generation, multimodal reference and real-time interaction into a unified streaming generation framework, so that the scene and story development can continuously evolve with user input. We have learned that Vivix will launch these two models and their APIs in the near future.
Obviously, Vivix's focus is on real-time interactive multimodal models. This is a different track from current tool-based model manufacturers such as Seedance and Keling AI — Liu Yu, who has worked on foundational models, is very clear that tool-based model manufacturers will get better and better and make the boundaries clearer, and there is no point in reinventing the wheel.
IDG and HSG approached simultaneously
When he decided to start a business, Liu Yu assumed the most conservative scenario: using the returns from his nine years of work at SenseTime as startup capital to get things off the ground first.
Unexpectedly, IDG Capital and HSG came to him in the same week.
In October 2024, news of his departure to start a business was exposed, and Liu Yu received hundreds of friend requests every day on WeChat from investors, FAs, and headhunters. It was at this point that his few friends in the venture capital circle — two from HSG and IDG — quickly found Liu Yu to communicate.
IDG came first. At that time, Liu Yu hardly had a complete Demo or Prototype, but the IDG team quickly decided: We want to invest, and we can issue the TS right now. Almost in the same week, the friend from HSG called, the week after learning about Liu Yu's entrepreneurial intention, they arranged an online meeting between Liu Yu and Neil Shen, the head of HSG and his partner team, to finalize the investment.
After that, Vivix successively completed 4 rounds of financing in 2025. The investment community obtained a more complete timeline of Vivix's financing:
In February 2025, it received seed round financing from IDG and HSG; in June 2025, BlueRun Ventures and 5Y Capital joined to complete the pre-A round; in August 2025, Monolith participated in the pre-A+ round; then in October and December, JD.com and Alibaba entered the market respectively, completing the pre-A++ round and the A round.
At this point, Vivix's valuation reached 1.32 billion US dollars. At this point in time, a valuation of 1.32 billion US dollars is not the top in the entire video generation field, but it is still not low — after all, Vivix has only been established for just over a year.
It is conceivable that with the upcoming launch of the two real-time interactive multimodal models, it will not be difficult for Vivix to start the next round of financing. What is more certain is that the commercialization capabilities of the entire multimodal track have begun to show, and it is expected that the market size will increase by an order of magnitude in the next few years. As a result, we have also seen that financing news for model and product startups on the track has been endless since the beginning of this year.
It is understood that the size of the Vivix team has expanded to more than 100 people in a year, most of whom are post-95s and post-00s, and Liu Yu is still the first person in charge of the data team and the model team.
Stepping into the Uncharted Territory
There is no doubt that every technological leap will bring a dimensional upgrade of content forms.
At the beginning of last year, users could barely generate videos with a sense of shot framing, lighting and movements based on a single sentence, and extend an image into a dynamic scene, but the effect was in the stage of "impressive but still unstable, able to make sample films but difficult to make finished films", and flaws such as characters having extra hands or feet during generation occurred from time to time. At that time, making old photos "come alive" and move was enough to ignite the public's excitement.
However, with the significant acceleration of model iteration, core breakthroughs have been made in unified multimodality, simultaneous generation of audio and images, and multi-shot narrative. Not only has model generation become more controllable, but aesthetic styles and physical simulation have also improved significantly. In the process of evolution, user expectations are completely different from a year ago.
Standing in today's technological process, a new generation of entrepreneurs is beginning to think: What will the next generation of content and game forms that can truly capture users look like?
As a result, "interactive content" has become a consensus — along this vision, some people are moving towards making products and tools, such as some interactive content platforms in the form of Feed streams and AI digital humans; others prefer to focus on the bottom layer, making the models themselves, such as Vivix's focus on real-time interactive multimodal models.
A painful yet rewarding experience is that Liu Yu, who chose the latter, did not fail to consider starting from the product side.
At first, the focus of the Vivix team was to prove that real-time interaction and personalized generation could be learned by a single model, and the team verified this with a self-developed foundational model. After that, on the issue of whether to define the product first or polish the model first, Vivix experienced a watershed calibration: in the middle of 2025, the team launched a product that mainly allowed users to tap their fingertips to make photos instantly move and turn into a video, but two months after its launch, the product was terminated.
Because during this process, Liu Yu quickly realized that to truly achieve real-time interaction, current model capabilities and infrastructure support are not sufficient, and those interactive gameplay that focus on special effects and rely on viral operations can hardly support scenarios with long-term retention and high payment willingness. At the end of October 2025, Vivix quickly shifted its focus back to model capabilities themselves.
Inevitably, there are questions about the future business model. As we understand it, Vivix has begun commercial exploration, with a preliminary plan to sell APIs to the B-side and charge according to tokens.
Regarding competition, Liu Yu remains optimistic: This track is far from converging. On the way to stepping into the uncharted territory, it is not so lonely even with competition.
This article is from the WeChat official account "pedaily.cn" (ID: pedaily2012), author: Feng Yuchen, published with authorization from 36Kr.