Four Chinese Players Are Building Real-Time Interactive Video Worlds

In 2026, at least four Chinese teams released interactive video world models: Alibaba introduced HappyOyster, Ant Group introduced LingBot, Shengshu introduced Vidu S2, and Aishi Technology introduced PixVerse R2. Their approaches differ, but the goal is shared: video that reacts to the user's actions in real time and turns into a virtual space the user can move through freely. The report comes from Jiqizhixin. The article's headline says one team took first place in two rankings at once, but the available excerpt names neither the team nor the rankings. So what follows covers only what the text confirms.

From clip to space

Conventional video generation is simple: the user writes a prompt, waits, and receives a finished clip. An interactive world model works differently. Frames appear as a continuous stream, and the user steers the camera, a character or objects along the way. The model has to take every action into account and show the result immediately.

This sharply raises the demands on the system. It needs low latency between command and frame. It needs scene consistency over long distances: if a user turns away from a building and comes back, the building must still be there. And the generation itself has to be stable, because a stream can't be "restarted" in the middle of a walk. These are typical engineering problems for this class of model, and each of the four products solves them in its own way.

Who is entering the field

The list of participants shows that the topic has moved beyond niche labs:

The list includes two tech giants and two specialized startups rooted in video generation. For the startups, interactivity is a natural extension of their video model lines. For Alibaba and Ant, it is one more way to show that their foundation models can do more than text and images.

What is left off-screen

The phrase in the original headline hints that the leading team positions its product differently from "yet another world model." What exactly it offers instead, and which rankings it topped, cannot be determined from the excerpt. Without benchmark names, methodology and figures, it is not valid to compare the winner with HappyOyster, LingBot, Vidu S2 and PixVerse R2. Ranking positions should be read together with the conditions under which they were measured.

What to check in practice

Teams considering interactive world models for game prototypes, simulations or advertising formats can skip the loud claims and prepare their own test. It helps to measure response latency on real hardware, not on a demo rig. Run a long session and check whether the scene starts to "drift" after a few minutes. Find out whether API access is open and what a minute of generation costs. Rankings can help build a shortlist, but your own scenarios will be decisive: camera control, scene repeatability and character stability. When the full article or a technical report comes out, these criteria will also be useful for comparing the winner's claims with the rest.