I trained a tiny 14M-parameter world model of 2D Minecraft in <10 minutes on 1xA100 (~10 cents on google colab). It runs locally in your browser, so you can play on your phone. It requires WebGPU: Chrome, Edge, or Safari on iOS 26 (built a custom runtime to save memory).
There's also a benchmark to score your own models: beat 0.562 and the site switches to your model. Memory is the open problem: walk 10 tiles away from your table and back, and it's gone.
pretty sweet looks stable enough even tho obv its not perfectly persistent
thanks! honestly most of that stability is from the data and tokenizer, not the dynamics model itself. the hardest part was to build a diverse enough but small dataset, that would contain round trips to teach memory and mix non trivial and random behavior on all stages of crafting tree, like, the world model isn't afraid of lava yet :)