A 131,000-Parameter World Model Generates Games on a $5 Pico, With No Frame Rate Measured Yet

A GitHub project called tinyworldformer runs a neural network that generates game frames on a Raspberry Pi Pico 2 W, a microcontroller board with 520 KB of SRAM. The repo, by a developer who goes by SheepSeb, had 223 stars and 8 forks when The Terminal read it on October 30, and the README says the author has not yet measured the frame rate on the Pico itself.

The project reached 214 points on Hacker News. Stars and points measure interest, not adoption.

What runs on the board

Two small networks share the chip. The world model is a looped transformer with 26,000 to 131,000 parameters, stored as int8 weights, and needs 2 to 8 million multiply-adds per frame. It predicts the next frame and reward from the current state and an action.

The second network plays the game the first one generates. It is a multilayer perceptron with two hidden layers, 9,000 to 16,000 parameters in float32, and 9,000 to 16,000 multiply-adds per frame. The README says the actor can be swapped for a "stream-x" agent that learns on the board.

The Raspberry Pi product page lists the board's RP2350 chip as two cores at 150 MHz with 520 KB of SRAM and 4 MB of flash, sold from $5. The README puts both networks on core 1 and leaves USB and the display to core 0. It says it chose the Pico 2 W because it is cheap and has a floating-point unit, a claim the Raspberry Pi page does not address.

How accurate the dreamed games are

The README reports exact-frame accuracy on held-out transitions for five games.

Game	Exact-frame accuracy
Flappy Bird	100%
Breakout	99.99%
Pong	99.45%
Pac-Man lite	86.4%
Doom-lite	83.6%

The README explains the Doom-lite figure: its 3D view has 768 cells, so one wrong shaded pixel marks a whole frame wrong, and the game state is "almost always right". The figures are the author's own, and we found no independent replication.

What is not measured

On the Pico, the frame rate is unmeasured. The README sets core 1 to 66 to 150 milliseconds per frame while a human plays. As arithmetic only, 8 million multiply-adds at one per clock cycle on a 150 MHz core is about 53 milliseconds, so the largest world model would fit that window only if the chip sustains close to one multiply-add per cycle. The README does not claim that.

On a desktop CPU the author's Rust code produces a frame in 0.7 milliseconds for Flappy Bird and 1.5 milliseconds for Breakout. A desktop is not a Pico, and the repo gives no conversion.

The larger caveat is what happened when agents trained only inside the generated game. They scored 210 there, where the real game's maximum is 56, and 28 in the real game. In a 5,000-step check, Pac-Man produced 742 wrong game states, and Breakout's dream sometimes removed bricks incorrectly, which an agent exploited.

For a company funded to build world models at a larger scale, see Dream's $17 million seed. The page shows no licence, so the code's terms for reuse are unclear. For another small model built to run with no connection, see the offline survival LLM.

The README's open item is the on-device measurement. Until the author posts a Pico frame rate, the 150 millisecond ceiling is a setting, not a result.