Laya is a fast, open-source System 1 decision model family designed for instant, calibrated probabilistic decisions over structured schemas, targeting high-volume routing, triage, and classification tasks without generating text.

Honest limitations are acknowledged: performance degrades with many options (e.g., Banking77 with 77 labels), zero-shot performance is limited without fine-tuning, and per-domain calibration can further improve results.

Resources and community links are provided, including Hugging Face hub, live demo, GitHub repository, PyPI package, and Kaggle fine-tuning notebook, emphasizing an open, reproducible ecosystem.

Real-world workflows across nine enterprise scenarios demonstrate production-ready decision quality, including email spam filtering, phishing detection, guardrails/jailbreaking detection, RAG relevance filtering, and multi-way ticket routing.

Conclusion emphasizes year-long development, open-source availability, sub-35ms decision speeds, zero hallucinations, multilingual routing, honest confidence scores, and broad community access.

Three bundled checkpoints (convaiinnovations/laya, laya-multilingual, laya-typed-decisions) with selective subfolder downloads optimize footprint, plus a routing system that detects Unicode scripts across 22 alphabets for language routing and reduces cold-start latency.

A head-to-head comparison shows Laya routed (System 1) outperforms TypeSafe Jev on multiple metrics, including accuracy, calibration (ECE), latency (P50), language coverage (51 languages), and cost (open-source with zero API fees).

A quickstart guide shows how to install and run Laya in under 30 seconds, including an example of multi-schema decisions and language routing using the Router component.

Laya relies on three primitives—choice, score, and noul (boolean probability)—to produce outputs as probabilities and numbers rather than text, eliminating hallucinations and malformed outputs.

System 1 emphasizes rapid, calibrated decisions and contrasts with System 2 autoregressive text generation, arguing most AI pipelines bottleneck on unnecessary generative models and proposing Laya as a faster alternative running in roughly 30–35 milliseconds on a single GPU.