Laya: Open-Weights 421M System 1 Decision Model Outputs Calibrated Probabilities, No Text

victormustar · x · 2026-09-18

Laya, an open-weights (Apache-2.0) non-autoregressive System 1 decision model by convaiinnovations, is trending on Hugging Face. Built on a fully fine-tuned ModernBERT-large backbone (421M total params) with a from-scratch decision head, it takes a state (text, email, ticket, JSON) plus typed questions and returns typed answers — choices with calibrated probabilities, ordinal score distributions, or calibrated P(true) — never generating text, eliminating parsing errors and hallucinations. The checkpoint adds email triage (spam/phishing/routing), trajectory modeling, and per-cardinality temperature calibration.

Related event: Open-Source Decision Model Laya: Trained on a Single GPU, 35ms Inference(2 posts)→

Original post →

More from coding & agent

coding & agent channel →