From Convaiinnovations · · 1m 50s
Laya: 32ms Multilingual Decision Engine
Open-source system that answers structured questions in 32ms across 100+ languages without hallucinations
In brief
Open-source system that answers structured questions in 32ms across 100+ languages without hallucinations Laya uses bidirectional encoders for 32ms decisions with calibrated probabilities in 100+ languages Originally reported by Convaiinnovations.
System 1 vs System 2
Modern AI pipelines use slow generative models for simple decisions. Laya provides instant, calibrated responses without hallucinations.
Three Decision Primitives
Laya uses choice, score, and noul to answer structured questions with probabilities from 0.0 to 1.0.
Three Checkpoints
Laya offers English, multilingual, and typed-decision models in a single Hugging Face repository with selective downloads.
Language Routing
Laya detects scripts in 0.09ms for English and 0.54ms for Indic languages, routing to the right model without cold swaps.
Laya vs Jev
Laya runs 7.8x faster than Jev with 45 of 51 languages supported and zero API costs.
Real-World Use
Laya filters spam with 99.3% accuracy and detects phishing with 98% accuracy in production workflows.
Honest Limitations
Laya struggles with >20 options and requires fine-tuning for best results, but provides honest confidence scores.
Quickstart
Run Laya in 30 seconds with pip install and a single forward pass for multi-schema decisions.
Open Source
Laya is 100% open-source Apache 2.0 with free Hugging Face models and self-hosted deployment.
The Core Idea
Laya solves high-volume decisions with sub-35ms responses using bidirectional encoders and reinforcement learning.
Why It Matters
Laya avoids hallucinations and provides calibrated probabilities for real-world decisions without expensive LLMs.
The Competition
TypeSafe AI launched Jev with similar concepts but without open weights or training data.
The Timeline
Laya started in March 2025 with an arXiv paper and evolved through September 2025 and 2026.
The Numbers
Laya runs in 32.8ms on a single GPU with 100% open-source weights and 45 of 51 languages supported.
The big picture
Laya uses bidirectional encoders for 32ms decisions with calibrated probabilities in 100+ languages
