laya: Non-autoregressive model performing text classification, scoring, and binary discrimination simultaneously in 38ms

Given text state and questions, it immediately returns choices, scores, and probability values without text generation. By adopting a non-autoregressive structure, it fundamentally eliminates parsing errors and hallucinations, providing mathematically calibrated confidence scores.

Combining the ModernBERT-large backbone with the RLCD training method, it answers multiple questions simultaneously in a single inference pass. It achieves an average latency of 38ms on GPU, demonstrating processing performance approximately 10 times faster than existing competing models.

It has demonstrated high accuracy in practical workflows such as email triage, customer routing, and content safety verification. Released under the Apache 2.0 license, it is suitable for on-device deployment and building internal systems without data leakage.