mizorewww/laya-coreml on GitHub
- Stars
- 1,401
- Forks
- 118
- Language
- Python
- License
- Apache-2.0
- Last push
- 22 Sep 2026
Read from GitHub on . Counts change daily.
Reported by the author
- Short decision
- 4.98 ms P50
4.98 ms P50 / 5.31 ms P95 on M3 Max with ANE FP16
github.com - Snake demo
- 49-50 dec/s
49.1–50.0 decisions/s across three
github.com
Laya-CoreML ports Laya, an open typed-decision model, to Apple’s Core ML framework, including variants that run on the Neural Engine, with zero generated tokens since it reads option probabilities directly. On an M3 Max, the author measured 4.98 milliseconds median and 5.31 milliseconds P95 for one short multilingual decision, and reports the same setup uses 2.78 times less whole-system energy per decision than a compiled MLX FP16 version of the same model; a separately validated 8-bit weight variant reached 4.88 milliseconds with a 3.19 times energy improvement. The author notes plainly that a targeted 10x improvement was not reached.
The Snake demo drives every move through a real local Laya call, sustaining 49.1 to 50.0 decisions per second across three uncapped 600-step episodes with zero deaths and two safety interventions from the visible safety layer. The 4.98 millisecond figures are for single questions, not full Snake frames, and the game-loop timings include rendering.
Following upstream Laya v0.3.5, it clamps fitted calibration temperatures to between 0.5 and 5.0 before use. The shipped value for questions with 11 or more options, 0.1006, would otherwise sharpen the probabilities about tenfold and report a coin flip as near-certain.
