Convai Innovations has released Laya, an open-source non-autoregressive decision engine that achieved significant attention as a leading machine-learning repository in September 2026. The 421-million-parameter encoder model is designed to process text and typed questions in a single forward pass without generating output tokens, returning calibrated probabilities for various options.
A technical guide published in September 2026 demonstrates how developers can utilize Laya on real labeled data using the banking domain of the CLINC150 intent dataset. The tutorial evaluates zero-shot accuracy, the impact of option wording and order, probability calibration, and the integration of abstention gates and pydantic schemas.
This guide matters because it provides developers with practical methodologies for implementing zero-shot decision models and evaluating calibrated probabilities in production environments.