NanoCoinRun

A neural network dreaming a playable game, next to the game itself. Same level, same keys. The left panel is generated one frame at a time by a world model trained in under an hour; the right panel is the real engine.

connecting…

Neural engine

the model, dreaming every frame

Real engine

Procgen CoinRun, ground truth

connecting…

← → move · ↑ or space jump · R new level

Training curve of the selected record entry, from its submitted run logs.

What is this

NanoCoinRun is a speedrun benchmark for tiny neural game engines. Entrants train a model that predicts CoinRun frame by frame from your controls — no game code at play time, just a network hallucinating physics, coins, and deaths it learned from 64×64 gameplay footage.

The race: reach a fixed quality score in the fewest wall-clock minutes on one 8×H100 node, data generation included. The score is the model's own training loss, held out on secret levels — chosen because it was the only metric family that survived validation against blind human play-testing.

Every record checkpoint is playable on this page, next to the real engine it imitates. If you can't tell which panel is which, that's the point.

Rules, harness, and record book on GitHub · Real engine, full screen