world-models-mario

Weights for a from-scratch PyTorch implementation of the Dreamer 4 world model (Hafner, Yan, Lillicrap 2025), trained on Super Mario Bros levels 1-1 to 1-4. Code: wilsonchenghy/world-models.

This is a world model only: given 8 real frames and a stream of button presses, it imagines the frames that follow. There is no agent.

File	Contents
tokenizer.safetensors	Tokenizer, 12.4M parameters, 40k steps
dynamics.safetensors	Dynamics (EMA), 37.5M parameters, 148k steps
config.json	Constructor arguments for both
sample_1-1.npz	A run the model never trained on: frames (458, 120, 128, 3) uint8 and actions (458,)

git clone https://github.com/wilsonchenghy/world-models && cd world-models
pip install -r requirements.txt
python -m envs.mario.demo demo.mp4

Provided for research. Super Mario Bros is the property of Nintendo; this project is not affiliated with or endorsed by Nintendo.