Laya — 33ms Multilingual System 1 Decision Engine

In March 2025, a researcher worked on a sub-35ms open-weight System 1 decision engine named Laya, as described in their March 2025 arXiv paper on sequence conversion trajectories (arXiv:2503.23303). The system utilizes reinforcement learning and provides calibrated probability predictions for structured schemas within 30 to 35 milliseconds on standard hardware. Unlike other models, Laya does not generate text or hallucinate, and it can handle multiple languages and schemas.

The system's architecture consists of three main components: an embedding model, a reinforcement learning component, and a routing mechanism. The embedding model converts input text into a numerical representation, while the reinforcement learning component generates confidence distributions and schema choices. The routing mechanism identifies the appropriate model based on the input language and script.

When comparing Laya to TypeSafe AI's Jev, the researcher notes that Jev launched without technical papers, open weights, or open training datasets. In contrast, Laya is entirely open-source and horizontal, with 32.8 milliseconds latency on a single GPU, making it six to eight times faster than Jev. Both systems support over 100 languages, but Laya includes a built-in Router that eliminates cold-swap penalties for language switching.

Benchmarking against Jev across various enterprise workflows, Laya demonstrates production-ready decision quality. The system's core realization is that generative LLMs are overkill for simple structured decision-making tasks, as they require significant computation time, real money, and parsing to extract clean labels from free-form text. Instead, Laya provides instant, honest, and calibrated probability predictions, making it an ideal solution for modern AI pipelines.