ScreenQuest: screenshot-to-action game agent on an M3 Max, no cloud

The port announcement: Laya on MLX, under 1 GB RAM, choosing the snake’s next move 60 times a second on a local M3 Max.

Atomic.chat ran Laya on a 16 GB MacBook Air against Jev over the API on a Grok-built Tetris. Local won on speed: 11x faster decisions.

Viraj Bhartiya wired both models into the Chrome T-Rex game: Laya local on MLX, Jev over the API, same track and physics, side-by-side latency. Code in virajbhartiya/laya-vs-jev.

Validated Core ML ports of all Laya checkpoints. One short multilingual decision in 4.98 ms P50 on M3 Max with ANE FP16, at 2.78x better energy per decision than MLX FP16.