Laya the open source version of Jev

A sub-35ms open-weight System 1 decision engine called Laya has emerged in the AI community as a new kind of model that does not generate text and provides lightning-fast probability predictions over structured schemas. This innovation was developed after working on a similar concept for a year, publishing papers, releasing model weights, building datasets, and creating a PyPI package.

In contrast, a well-funded frontier lab, TypeSafe AI, launched a similar model called Jev around the same time without technical papers, open weights, or open training datasets. Laya's architecture is based on reinforcement learning, and it runs in 32.8 milliseconds on a single GPU, making it 6 to 8 times faster than Jev. Laya supports over 100 languages and provides 100% open-source Apache 2.0 weights without any API subscription costs.

It evaluates typed questions over any state in a single forward pass and relies on three primitives. Laya's Router identifies the incoming text's Unicode script and analyzes Latin stopword distributions, ensuring efficient routing with negligible overhead. The model's performance surpasses that of Jev across various enterprise workflows, demonstrating its superiority in production-ready decision quality.