Laya Releases 33ms Multilingual System 1 Decision Engine

Developed over the past year following 2025 research publications, developers at convaiinnovations have released Laya, an open-source decision engine designed to replace autoregressive language models for structured classification tasks. Following industry developments from TypeSafe AI’s proprietary Jev, Laya operates as a System 1 architecture, delivering calibrated probability outputs in under 35 milliseconds without generating text or relying on streaming inference. The framework addresses a persistent AI bottleneck: the computational inefficiency and hallucination risks of deploying generative models for simple reflex decisions. Instead of predicting tokens, Laya employs a bidirectional encoder to evaluate inputs directly against structured schemas. The system supports three core primitives: choice for categorical selection, score for ordinal ranking, and noul for calibrated binary probabilities. Strict numerical outputs eliminate schema validation errors and fabrications. Laya ships with three specialized checkpoints for English classification, multilingual processing across over 100 languages, and enterprise decision-making like customer service routing. All models are consolidated in a single repository, enabling selective weight downloads. A built-in script-aware router automatically detects input character sets, dispatching queries to the appropriate model in under two milliseconds. Preloading eliminates cold-start latency, maintaining consistent performance across mixed-language workloads. Independent benchmarks position Laya as a direct competitor to TypeSafe AI’s Jev engine. Laya demonstrates a 7.8x reduction in single-query latency, averaging 32.8 milliseconds compared to Jev’s 236 to 276 milliseconds. Batched processing achieves 20x speed improvements, while calibration error metrics improve threefold. The framework eliminates per-token API fees, offering full commercial rights under an Apache 2.0 license. Laya also outperforms Jev in standard classification benchmarks and extends usable multilingual support to 45 of 51 tested languages. The team acknowledges specific architectural constraints. Performance on categorical choice questions degrades when schema options exceed twenty, prompting recommendations for hierarchical routing in complex tasks. Base models require domain-specific fine-tuning to achieve peak accuracy, with temperature calibration essential for production-ready confidence scores. For enterprises managing high-volume triage and real-time guardrails, Laya provides a lightweight alternative to generative AI. By decoupling structured decision-making from autoregressive generation, the framework enables sub-second inference, zero-hallucination outputs, and transparent probability scoring directly integrated into application logic.