The valuation has soared to 4 billion U.S. dollars, but the real-time interactive video sector has not yet rolled out a "killer application".

"Core Tip" Capital is betting wildly, and valuations are surging. Is this high-stakes gamble on real-time interactive videos betting on technical routes, or the luck of young entrepreneurs?

The world model is one of the most capital-intensive and most distant directions in the field of artificial intelligence in the past two years.

Its goal is to allow AI to build a self-operating virtual world internally for training games, robots and autonomous driving. Judgments that humans can complete with intuition require huge amounts of computing power and data when handed over to AI. For ordinary people, it seems far from daily life.

However, even the seemingly distant world model now has a clear commercialization direction.

In May this year, Decart, a world model company, reached a cooperation with AWS, the cloud computing division of Amazon, packaging its real-time video model Lucy into an API, which is directly called by enterprise customers through AWS's distribution network and billed by tokens.

Lucy's capabilities are also very down-to-earth. You only need to point the camera at yourself, at the room, or at any object at hand, and type a sentence, the picture will change immediately. According to the official video of lucy2.5, you can even try on clothes virtually at any time. When you turn around or raise your hand, the clothes will move accordingly. Perhaps in the near future, the most troublesome problem for us when shopping online will really be solved by AI.

This is real-time interactive video. In the future, videos may no longer be pre-shot finished works, but pictures generated on the spot for you, that can be interrupted at any time and modified at will. On the expensive and long track of world models, it is one of the few directions that can actually generate revenue at present. And because it is close enough to commercialization, a large number of people and funds have poured in, and technical routes parallel to world models are also advancing at the same time.

However, behind the popularity, in terms of products, it is still unknown who will be the first to launch a "killer application".

So in that distant AI-related future, why is real-time interactive video a good commercial landing direction? How many technical routes are running in parallel from the brand-new interaction mode to the landed product, and how far do we have to go?

Three Routes, Three Strategies

Traditional generative AI for videos, such as OpenAI's Sora and ByteDance's Seedance, all take a paragraph of text as input and output a finished video of several seconds to tens of seconds. Users can only watch it, and cannot modify or edit it. Real-time interactive video is different. It can generate continuously based on the already generated picture and the user's interaction, producing an interface that can be explored in real time and changed with the user's operation.

Take a train driving into a tunnel and then driving out as an example:

The world model route chosen by Decart has a simulated world "in mind". When the train drives into the hole, it knows the train is still there and will drive out. This route has strong capabilities, but it is also very expensive, consuming both computing power and data.

Therefore, for the presentation of real-time interactive video, the more mainstream practice at present is the pixel route. Treat each frame in the video as a painting, and draw them frame by frame. When the train drives into the tunnel, the train on the picture disappears, and the model can only guess whether to draw the train back. Therefore, this route is easier to create beautiful and immersive effects. But it does not care whether the object always exists or whether the numbers are accurate. Problems such as an extra finger and a wrong price mark mostly occur here.

The third route is the code route chosen by SigmaZ AI, a domestic real-time video interaction company. SigmaZ AI is a real-time interactive video generation model company driven by diffusion language models.

The route it chooses does not draw pictures, but maintains a scene list in the background, and the train is a record on the list. When the train drives into the tunnel, the background will mark it as "temporarily blocked", so it can steadily drive out from the other end. In fact, what it generates is not a video, but a continuously growing web page with a timeline - everything in it is a piece of code, which you can drag and rotate at will, and you can also call a stop at any time. In the real-time interaction track, it has higher accuracy than the pixel route and is cheaper than the world model approach. But so far, it is temporarily not as good as the pixel route in the dimension of pure visual perception.

The three routes all have their own shortcomings at present, and all are making up for the lessons. The pixel route is improving accuracy, the code route is improving picture quality, and some world model companies such as Decart are trying to reduce computing power costs. Only when they meet at the product end can we know who runs faster.

However, different routes have the same pursuit of fast generation speed. On how to achieve speed, each company gives a different answer. The answer given by SigmaZ AI is a generation method different from mainstream large models: Diffusion Language Model (DLM).

This is a technical path different from the familiar large language models. LLM generates one token at a time, and the next word can only be generated after the previous one is determined, so the longer the sentence, the slower it is. DLM is different. It will first lay out a rough overall draft, and then go through several rounds of iterations to repeatedly correct all positions at the same time until it is clear. Because there is no need to queue up word by word and advance in parallel, it can theoretically achieve faster speed.

In the past, DLMs were mostly small models with billions of parameters. Ant Group's LlaDA 2.0 has for the first time upgraded it to the level of hundreds of billions of parameters, proving that this route can also carry the scale of mainstream large models. In May 2026, ByteDance open-sourced the continuous latent space diffusion language model Cola DLM, and in June of the same year, Google Deep Mind released the diffusion text model Diffusion Gemma with open source weights for the first time. Just like LLM before the advent of ChatGPT, DLM is moving from the laboratory to large-scale application.

Capital is Betting on a Group of Young People

If the route determines the strategy, then a more direct commonality of capital is that it places bets on a group of young people in their twenties.

Dean Lettsdorf, founder and CEO of Decart, is 27 years old this year. The company has completed multiple rounds of financing. In May this year, when it completed about $300 million in financing, its valuation was about $4 billion, and its investors included NVIDIA and others. In August this year, according to media reports, Anthropic intends to acquire Decant at a price of about $6 billion.

Turning to China, money is also flowing in this direction.

For example, Vivix, a domestic company that also does real-time interactive videos but follows the "streamlined DiT transformation" route, its founder Liu Yu is a post-95s, with a doctorate from the Multimedia Laboratory of the Chinese University of Hong Kong, and became one of the youngest research executive directors of SenseTime at the age of 26. According to reports, Vivix currently has a team of more than 100 people, most of whom are post-95s and post-00s, betting on the direction of interactive narrative and entertainment immersion. The company has been established for one year and has won 5 rounds of financing, with a valuation of more than $1.32 billion. Its investors include well-known investment institutions such as IDG and HSG, as well as 5Y Capital and BlueRun Ventures, which have invested in Xiaomi and Li Auto.

SigmaZ AI, which has different technical routes and application landing directions from Vivix, is also favored by BlueRun Ventures. Compared with the direction of entertainment immersion, they are more inclined to build a general platform.

It is reported that the two founders of SigmaZ AI, William (Yang Bolin) and Derek, were born in 2003 and 1995 respectively. The former is a serial entrepreneur who has founded many companies; the latter graduated from the Department of Computer Science of Imperial College London and served as Tech Lead for visual code training at Amazon AGI Lab. In addition to BlueRun Ventures, SigmaZ AI has also received investments from Qiming Venture Partners, DCM, XIAOXIAO FUND and other institutions.

Capital's bet on these young people is actually betting on the same thing: in the cutting-edge directions that large manufacturers are unwilling to go all in, young entrepreneurs may be the first to run out.

An AI practitioner told Baobian that for cutting-edge directions, the exploration space in small startups may actually be larger than that in large manufacturers. Large manufacturers often bet on the most profitable directions, first replicate, amplify and exhaust the places where the business model has been verified. For "sidebet" (high odds and low winning rate) directions, they often only assign a very small team to try. "If it doesn't work, they have many profit points, and they will abandon this direction."

But the directions that large manufacturers have not gone all in are not necessarily wrong.

In May 2026, Decart's Series B lead investor publicly stated that although the entire AI industry generally believes that real-time, production-level generative video world models will take several years to achieve, Decart has turned it into a commercial reality, and Amazon is its core early API customer across multiple product lines.

This company led by people in their twenties delivered something that seemed very distant before large manufacturers, and large manufacturers such as NVIDIA have instead become its investors.

Where is the Next Battlefield?

The excitement of capital is one thing, and whether users buy it is another.

Up to now, Decart has gone the farthest on the product side, and its Lucy model has been delivered externally through cooperation with Amazon AWS. At present, its B-end applications are mainly in scenarios such as Amazon's e-commerce, virtual fitting, live streaming and advertising, provided through API and billed by usage. The founder of Decart once vaguely revealed his yearning for to C applications in an interview: "We could have been a good company sold to NVIDIA or other giants, but what we want to do is an application for a billion consumers."

Vivix may be more inclined to bet on the direction of interactive narrative and entertainment immersion. According to media reports such as Investment, Vivix is currently polishing two models. One is Vivix-A1, known as the first real-time full-duplex model for interactive AI characters; the other is Vivix-W1, a real-time multimodal large model with interactive narrative as the core, and the API will be launched at the same time as the model goes online.

SigmaZ AI puts "accuracy" first. William and Derek said that accuracy will always be the first in their benchmark. This also makes their products eventually oriented to practicality. SigmaZ AI told Baobian that from this perspective, the imagination space of the real-time interactive video track may be far larger than "creating a digital human that can interact".

SigmaZ AI initially made a C-end consumer product, with good data, but the team gradually found that users not only used it for learning, but some people were trying to use it to generate commercial videos.

"Small merchants who upload product detail pages will ask us in the background, can my goods also be rotated 360 degrees to introduce? Those video generation tools on the market can only output 15 seconds at a time, which is expensive, and they always make mistakes on the products. Yours is accurate, if you launch a dedicated video generation version, I am willing to pay three or four times more money." William said.

Therefore, based on the existing real demand, they decided to build a more widely targeted platform Tap8, positioned as a "product laboratory". In the future, Tap8 will be open to C-end consumers and B-end customers at the same time, serving multiple information-intensive scenarios such as marketing, sales, tour guide, and how-to. Perhaps one day, as long as the user asks the camera "how to repair the TV", a rotatable 3D model will pop up on the screen, guiding the user to operate step by step.

At present, we can already see some signs of competition on the product side. Take the trivial matter of "changing the background" as an example. Lucy's ability is that if you say a sentence to a real-time picture, it can change the background on the spot. In Tap8's promotional video, users can ask questions about the real-time generated picture, make the picture change instantly according to instructions, and directly click on the elements in the picture; for example, click on a part of a 3D headphone model, Tap8 will generate the corresponding principle explanation picture in real time.

But this may only be a small part of the competition. The bigger challenge is that these young startups have to make both capital and users believe that what they make with AI is different from the existing products on the market.

For example, in Lucy's promotional video, users can use AI to make their faces very swollen, which looks cool, but Douyin's filters have been able to produce similar effects many years ago. No matter how real-time and smooth the technology is, if users only see it as "just another filter", this track has not really found its own meaning.

Perhaps the real watershed of this track is not who runs faster, but who is the first to create a product that never existed before.

This article is from the WeChat official account "Baobian" (ID: baobiannews), author: Zhan Fangge, published with authorization from 36Kr.