Laya: a multilingual System 1 decision engine

2026-09-21

Laya is an open-weight, non-autoregressive decision model for structured choices, ordinal scores, and calibrated yes-or-no probabilities across more than 100 languages. The published benchmarks report 32.8 millisecond latency on a single GPU, but the page also notes important limits, including weaker performance with more than 20 choices and the need for task-specific fine-tuning and calibration.