X-Gen Labs wants to simulate an entire persistent world, not just video frames

Most "world models" today are really just glorified video generators. They predict what the next frame should look like, but look away for two seconds and turn around, and the car on the street has magically turned into a trash can. There is no actual persistence behind the pixels.

X-Gen Labs introduced a generative world simulation system designed to fix that problem, split into two distinct layers: DAO and JING.

One runs the engine, the other holds the camera

Basically, you can think of DAO as the background game engine and JING as the first-person camera looking into it.

DAO is the computable world engine. It keeps track of the shared world state, enforces physical and social rules, and updates everything when an action is taken. Crucially, autonomous agents and NPCs keep doing their own thing even when you aren't looking at them. Leave an area and come back later, and the simulation remains consistent.

JING, on the other hand, is the interactive experience model. It takes whatever localized slice of the world DAO gives it and renders what a specific character sees, hears, and touches in real time based on their action history.

This separation enables true multi-observer consistency. In tests with three different characters in the same environment, the shared state stays aligned across every viewpoint. Unlike competitors like VO3 or Genie 3, this system has an explicit state management backbone. As long as your compute holds out, the world can run indefinitely instead of cutting off after a few seconds.

A 67GB download with plenty of rough edges

Right now, Generative World Simulation is purely a research preview, though the team has released the code and instructions for the Jing model on GitHub.

Running it locally is going to hurt your storage. The visual rendering backbone is fine-tuned on the full MiniMax H3 video model, coming in at a massive 67GB. Unless you have access to enterprise-grade silicon, you will likely have to wait for community quantizations.

There are also significant technical hurdles. X-Gen Labs acknowledged that fast camera swings and rapid motions degrade precision. Run the session too long, and you will encounter drift in spatial relationships, scene details, and context consistency. Real-time latency and visual fidelity still have a long way to go before this feels like an actual game.

Why it matters

We think X-Gen Labs is doing something very smart here by refusing to treat world simulation as a pure next-frame computer vision problem.

Video generators are inherently amnesiacs. Trying to force a single model to both imagine photorealistic light transport and remember that an NPC dropped a key behind a rock ten minutes ago is an architectural dead end. Decoupling persistent state logic in DAO from perceptual rendering in JING is how we actually get closer to an OASIS-style persistent universe.

For game developers and interactive creators, this offers a glimpse into a future where environments are neither pre-baked by artists nor limited to scripted loops. You get an evolving society that advances whether the player is logged in or not.

That said, nobody should expect a drop-in Unity replacement anytime soon. Between the 67GB footprint and drift issues during extended runs, it is strictly an experimental playground today. But it officially shifts the conversation from "can AI draw games" to "can AI keep a world running."