← Research

Open Release · 03

A simulator that learned physics — and runs in this tab.

A world model is a neural network that learned to predict what happens next. This one runs its whole loop in your browser: give it your action, and it paints the next frame, feeds that back, and paints the one after — a playable world with no physics engine behind it, just a network that learned the rules by watching. Drive it below.

Loading the world model…
What a world model is

Watch, dream, drive.

The frontier of embodied AI is learning to imagine consequences. This is that idea, shrunk until it fits on your device and you can hold it.

It watched

Learned the physics from pixels

Trained only on rendered rollouts of a rover with momentum in a walled arena — never told the rules. Momentum, friction, and the bounce off a wall are all inferred from watching frames.

It dreams

The network is the engine

At run time there is no simulator. The last two frames and your action go in; the network paints the next frame, then that becomes the input for the one after. It generates the world one step at a time.

You drive

Playable, on your device

Arrow keys are the action. Every frame is a forward pass in plain JavaScript — no server, no GPU. You are playing a physics you can't see, dreamed by a network small enough to fit in a tweet.

The numbers

Small, and honestly measured.

Evaluated on held-out rollouts it never trained on.

22,899parameters
0 physicsengines at run time
next-frameprediction, in plain JS
on-deviceno server, no GPU

It is deliberately tiny, and it is an approximation, not a perfect simulator: over a long unbroken run the network can let the rover soften or drift, because it is guessing every pixel from what it learned rather than solving equations. That honesty is the point. The frontier versions of this idea — playable neural game worlds generated frame by frame — are hundreds of millions of parameters and need a GPU; this one shows the same mechanism, released and runnable, on the device in front of you.

Get it

Weights and the recipe.

Model

nano-world-model

Weights, config, and a runnable example. Trained with a multi-step rollout loss to stay stable when it drives itself.

↗ Hugging Face

The family

Vision-language-action

The released VLA policies — a tiny open-vocab arm, multi-step manipulation, the real 450M SmolVLA — all runnable in the browser.

→ Run them

Learn it

Robot Learning

The in-browser course on how policies and models like these are built, demonstration to release.

→ Take the course