Games played by Laya

Run one locally

Snake on a Mac from laya-mlx:

pip install 'laya-mlx[demo]'
hf download aac6fef/laya-multilingual-mlx
laya-snake

The README reports 75.40 moves/s over 2,400 moves on an M3 Max with --optimize --max-speed, with zero deaths and two visible safety interventions. laya-coreml has a Neural Engine version at 49.1 to 50.0 decisions/s.

Flappy Bird on an Intel CPU from flappy-laya-openvino-cpu, with one noul question per frame:

pip install -r requirements.txt
python flappy_bird.py

Its README reports a mean 28 ms per decision with the INT8 OpenVINO model.

More games

• laya-pong: browser Pong with a native Laya paddle at 18.7 ms p50 on Metal.
• laya-vs-jev: Laya (MLX) and Jev (API) play Chrome's T-Rex side by side.
• laya-vs-jev-arena: a snake race and a fighting game, Laya vs Jev or vs a human.
• open-jev-laya: Gomoku, Big Two and a 3D maze entirely in the browser.
• laya-playground: Flappy, a lane runner and Tetris, plus an editor.
• kevala: Tetris in the browser on WebAssembly and WebGPU.
• Laya4j: Snake and a Spring Boot Tetris in Java.

What the demos teach

• Describe the state in words. In laya-pong, the model matched the reference controller when the state was a sentence and never moved when it was three numbers. The Flappy Bird author saw the same effect: with a JSON dict the model answered about 0.9 on every frame.
• Make options phrases. "move the paddle up" beat "up" in laya-pong: 5/5 correct against 3/5.
• Keep a planner and a safety layer. The Snake and T-Rex demos use planner features and a safety layer that can override unsafe moves. Their scores measure the whole system, not the model alone, and the READMEs say so.