Where people are actually running it

The official package targets a normal Python install and can sit on a GPU. The interesting public work is the set of ports that move the same questions onto a laptop or a browser. We only list pages we could open.

• mizorewww/laya-mlx runs the checkpoints with MLX and ships a terminal Snake demo. The package page reports medians of 13.4 ms (English) and 7.4 ms (multilingual) for a short question.
• mizorewww/laya-coreml compiles the same API for Core ML. The ANE note reports 4.98 ms p50 for one short multilingual question on an M3 Max.
• mizchi’s ONNX export is what the browser Space uses for Snake and chess.
• afshinm/laya-mps is a PyTorch Metal runtime with a local Pong demo and more than one memory mode.

Those milliseconds are copied from the authors’ own pages. We have not re-run them. A game that calls the model several times per frame will not match a one-question API number, which is why the MLX repo separates its Snake rate from its short-question benchmark.

Social clips of Breakout, Doom, and Tetris are easy to find and hard to cite: the posts move, and the short links rot. They stay off the gallery until there is a stable URL.