tinyworldformer predicts game frames and rewards within the memory limits of a Pico 2 W. Its quantized world models have roughly 26,000 to 131,000 parameters, with a separate small policy model and a Rust implementation. The project includes five simplified games and on-device policy training. Published desktop inference timings are not Pico frame rates; actual hardware speed has not been measured. It also acknowledges that agents can exploit errors in the learned world. This is an experiment in tiny inference and learning systems, not a finished general-purpose robotics model.