A 26K-parameter world model squeezed into a Pico 2 W—can Snake hold steady?
tinyworldformer compresses a world model down to 26K–131K parameters and generates frames while choosing actions on a Pico 2 W. Breakout reaches 99.99% frame accuracy; the more interesting next test is Snake: once the snake grows longer, state consistency will reveal problems.
World models are often tied to billions of parameters. tinyworldformer compresses the scale down to 26,000 to 131,000 parameters and runs on a Pico 2 W. It uses a recurrent Transformer to generate frames and an actor to choose actions; the model generates game frames and plays while running on the board. Project: SheepSeb/tinyworldformer
Grok mentioned in a reply that Breakout has already hit 99.99% frame accuracy. Even more interesting is Snake: as the snake body dynamically grows longer, the recurrent Transformer needs to continuously maintain state consistency. With 26K parameters creating and playing on a Pico 2 W, one frame hiccup will give it away. This is a judgment worth watching. Compressing parameters this low makes int8 quantization the key path. Whether on-device inference can hold steady depends on error accumulation after generating many consecutive frames.
A user named “AI Work Diary · Xiao F” said that the Snake question hits the point: once the snake body gets long, the state has to stay aligned all the time; with 26K parameters creating and playing on a Pico 2 W, one frame hiccup exposes everything, and if the int8 path can hold steady, that would be interesting. Another user, haha lin, was more direct: stuffing 100K+ parameters into a microcontroller shows that on-device space is bigger than people think.
Here is a bit of background. The Pico 2 W is a microcontroller board with wireless connectivity, and both memory and compute are limited. Putting a world model on it means the difficulty shifts to post-quantization stability, frame-by-frame generation error, and the mutual constraints between action selection and frame generation; parameter count is only the surface. Breakout reaching 99.99% frame accuracy shows this path is viable under simple game dynamics. Snake lengthens the state-consistency test: each time the snake eats food, its body length and collision logic change, and the model has more things to remember.
Should we have it build Snake or Breakout first? Breakout already has results; Snake looks more like the next hard problem. The project is open-sourced on GitHub; if you want to see how an on-device world model runs on a microcontroller, you can go straight to the code and demo: GitHub project. Original post: X original post.
发布时间: 2026-11-01 14:28