Laya: The Open-Source Alternative to Jev's Proprietary AI Decision Engine
A sub-35ms open-weight System 1 decision model with multilingual routing and RLCD calibration.
1Open-source decision engine breakthrough
Laya represents a new class of non-autoregressive AI models focused solely on fast, calibrated decision-making rather than text generation. Built on bidirectional encoders, it achieves 32.8ms latency on single GPU, making it 6-8x faster than proprietary competitor Jev. The architecture outputs probabilities and structured choices without text generation, eliminating hallucination risks while maintaining state-of-the-art accuracy across 9 enterprise workflows and 51 languages.
2Reinforcement learning foundation
The system's core innovation is RLCD (Reinforcement Learning for Calibrated Decisions), initially developed in March 2025 research. Unlike traditional LLMs that predict tokens, Laya uses PPO over sequence representations to output turn-by-turn conversion trajectories (0.0-1.0 probabilities). This approach was formalized in October 2025 research, establishing schema-based decisions guided by reinforcement learning rather than autoregressive patterns.
3Multilingual routing system
Laya's built-in Router analyzes Unicode scripts across 22 alphabets with only 2% overhead, solving critical non-Latin script failures observed in benchmarks. The system automatically routes inputs to specialized checkpoints (English ModernBERT-large or multilingual mmBERT-base) while maintaining sub-35ms latency. Preloading eliminates 7-10 second cold-swap penalties when handling alternating languages in production environments.
4Performance benchmarks
Independent benchmarks show Laya outperforms Jev across key metrics: 7.8x faster execution (32.8ms vs 150ms), 100% open-source availability versus proprietary API ($0.042/million tokens), and superior calibration. On the MASSIVE benchmark's 51-language sweep, Laya maintained consistent accuracy while exposing English models' critical failure to recognize non-Latin script inputs despite high confidence scores.
5Architectural advantages
The model's three primitives (choice, score, and noul) enable complex decision workflows without text generation. Outputs are constrained to valid schema structures, making malformed JSON physically impossible. The system handles typed questions over raw text, emails, tickets or JSON documents in a single forward pass, with batched processing achieving 7.2ms/question throughput.
6Resource efficiency
Laya's modular design allows downloading only required subfolders (808MB for English vs 647MB multilingual) rather than the full 2.5GB bundle. The PyPI package includes complete SDK for multi-schema decisions with automatic routing. Community resources include Hugging Face checkpoints, demo space, GitHub repository, and Kaggle notebook for custom model training in ~4 hours on free GPUs.
7Enterprise workflow applications
Practical implementations demonstrate Laya's production readiness: ticket routing (96% confidence in queue assignment), urgency scoring (2.87/3.0 precision), and churn risk detection (91.4% accuracy). The system handles complex state evaluation across infrastructure, billing, security and support domains while maintaining sub-35ms latency for high-volume classification and triage tasks.
8Open AI ecosystem contribution
Laya's full open-source release (Apache 2.0) contrasts with proprietary alternatives, providing the community with production-grade decision architecture. The project demonstrates that not all AI problems require autoregressive chatbots, offering a specialized solution for high-throughput, low-latency decision workflows with mathematically honest confidence scores suitable for branching in production code.