Laya open-sourced: 33ms non-autoregressive decision model with routing

duummbass · x · 2026-09-20

convaiinnovations open-sourced Laya on Hugging Face: a multilingual, non-autoregressive System 1 decision model. Given a state (text, email, ticket, or JSON) and typed questions, it returns typed answers with mathematically calibrated probabilities in a single forward pass (33 ms) across 100+ languages. It is trained with RL against strictly proper scoring rules (RLCD), so honest probability reporting is the only reward-maximizing strategy; it never generates text, so there is nothing to parse or hallucinate. The repo ships three checkpoints: English (ModernBERT-large, 421M, 512 ctx, best at guardrails/email triage), multilingual (mmBERT-base, 322M, 1024 ctx, 2.2x faster), and typed-decisions (421M, 0.766 acc on four workflows). Recommended usage is the built-in Router: pip install laya, preload checkpoints, and route in under 35 ms. Users suggested adding it to JevBench-style routing evaluations.

Related event: Open-Source Decision Model Laya Delivers Judgments in 33ms(2 posts)→

Original post →

More from coding & agent

coding & agent channel →