GitHub · Runtime ports
mizorewww/laya-coreml
The energy-per-decision comparison, if you are choosing between Apple's two runtimes.
What it does
Converts Laya's multilingual checkpoint to Core ML so the Neural Engine runs it, and measures both latency and energy per decision. The packaged Snake demo replays a recorded run at one times speed with the model's own probabilities on screen.
The repository's own description: “Local Laya typed decisions on Apple Core ML and Neural Engine. Validated ports, ~5 ms short decisions on M3 Max, reproducible speed and energy benchmarks.”
How it works
A converted model runs on the Neural Engine in FP16 with a quantised variant alongside it, and the demo drives a three-question game loop through real decisions while a visible safety layer can break a cycle the model proposes.
The repository, by the numbers
Read from the GitHub API on 2026-09-24. Stars and the last push move daily — quote them with the date, the way we do.
What we checked
- the README and both benchmark documents, read 2026-09-24
- the sentence saying which timings include rendering and which exclude terminal painting
- the licence file
What we did not check. The energy comparison. It was measured by the author on one M3 Max against one alternative; we have neither the instrument nor the machine to repeat it.