Robbyant, an Ant Group company that builds artificial intelligence for robots, released the last of its LingBot-World Infinity models, including a much smaller version, on Sept. 10, 2026, following a first release in July. Also called LingBot-World 2.0, it draws a video world on the fly as you move through it. As Robbyant's researchers note, most AI worlds like this stay stable for only seconds to a few minutes before textures smear and scenes warp. Robbyant says its version stays steady for hours, long enough to play in, tell stories in or train robots in.
The result plays like a video game with no game engine behind it. Players move with the familiar W, A, S and D keys, and the model paints each next moment of video in response. A second AI acts as a director, suggesting events that suit the scene, such as a whale made of clouds drifting past, which a player can trigger with a single key. The team says the model handles actions like combat, archery and spell-casting, and can make it snow or rain on demand.
Two people can shape the same world at once, Robbyant says, one playing and one directing, which opens the door to shared games and storytelling. The company also sees the model as a practice ground for robots, a place to predict what happens next and to generate training data. Robbyant says the real-time version can already be tried on two partner platforms, and the models are free to download for non-commercial use.
Robbyant also lists the gaps. The world is not truly remembered: walk away from a spot and come back, and the model tends to redraw it rather than recall it. Objects sometimes pass through each other instead of colliding, and characters can slowly change their look over long sessions. The company says it is working on better memory, a firmer grasp of physics and leaner ways to run the model, so that such worlds can one day work on everyday hardware.
For now, the model still needs serious computing power, and Robbyant says it does not plan to release the code behind its own real-time setup. Robbyant's own stability test ran for an hour, and the company says the model runs at full capability in its official setup, calling the partner platforms a convenient way to try it.
