Laya: Non-Autoregressive System 1 Decision Engine — 33ms Typed Decisions Over 100 Languages 2026
Introduction There’s a fundamental tension in AI systems: the models that reason best are slow, and the models that are fast don’t reason well. Most developers accept this tradeoff. Laya rejects it. The project achieves what seemed impossible: typed decisions—multiple choice, scoring, yes/no—at 33 milliseconds per question, trained with reinforcement learning against strictly proper scoring rules, across 100+ languages. No text generation. No parsing. No hallucination surface. This isn’t a distilled LLM pretending to decide. It’s a different architecture entirely: a non-autoregressive encoder that evaluates questions in a single forward pass. The implications extend beyond speed. When your decision model outputs a label, not text, you eliminate an entire category of failure modes. ...