Laya the open source version of Jev

Laya is an open-source, non-autoregressive decision model family built to make instant, schema-based predictions (not free-form text) with honest, calibrated probabilities. Originating from reinforcement-learning research published in 2025-2026, it implements a System 1 reflex layer that answers typed questions in one forward pass using three primitives - CHOICE (categorical selection with distribution and confidence), SCORE (ordinal expectation and distribution), and NOUL (boolean probability). Models are bidirectional-encoder based, provided as three specialized checkpoints on Hugging Face (English, multilingual, typed-decisions), and include a sub-millisecond pure-Python Router that inspects Unicode script and stopwords to route inputs to the appropriate checkpoint. Laya runs in roughly 32.8 ms single-question latency (7.2 ms/question when batched), supports 100+ languages, ships under Apache 2.0, and is self-hostable at no API cost.

Benchmarks show meaningful gains over a proprietary competitor: higher accuracy on benchmarks (e.g., typed-decisions 0.766 vs 0.727, AG News 0.950 vs 0.910), much better calibration (ECE ~0.081 vs 0.246), and large latency/cost advantages. Real-world workflows (spam, phishing, guardrails) report high accuracy and low ECE. Honest limits include degraded performance for flat CHOICE schemas >20 options (Banking77 stress test), modest zero-shot baseline until fine-tuned, and the need to fit temperature scalars per question type for best calibration. A pip-installable SDK and quickstart router make deployment straightforward.