laya
Non-autoregressive System 1 decision engine for typed choices across 100+ languages.
Typed choices, scores, and yes/no evaluations over text in a single forward pass.
Laya is a non-autoregressive decision engine that performs typed choices, scoring, and yes/no evaluations on any text in a single forward pass. It is designed for developers and data professionals needing fast, deterministic text routing and classification without autoregressive text generation.
What it is
The tool processes text inputs in over 100 languages via an automatic router that selects between ModernBERT and mmBERT checkpoints. It evaluates custom question schemas with criteria for classification, scoring, and boolean decisions in roughly 33 ms per request on a T4 GPU.
Who it's for
It suits teams building multilingual triage, routing, or classification pipelines who want to avoid the latency and hallucination risks of large autoregressive language models. It does not suit those looking for conversational agents, text generation, or complex multi-turn reasoning.
Key features
Single Forward Pass
Evaluates typed choices and scores instantly without text generation or parsing overhead.
Automatic Checkpoint Routing
Routes requests dynamically between English, multilingual, and fine-tuned domain checkpoints.
Long Document Support
Reads inputs up to 8,192 tokens using the multilingual checkpoint for extensive text analysis.
Framework Integration
Includes optional extras for LangChain, LlamaIndex, CrewAI, MCP servers, and HTTP deployment.
Up and running in 3 steps
Our honest verdict
What's great
- Performs classification and scoring in a single forward pass without generation hallucinations.
- Native multilingual support across 100+ languages with automatic script detection.
- Supports fine-tuning on domain-specific decisions using reinforcement learning against proper scoring rules.
- Provides optional integrations with LangChain, LlamaIndex, and CrewAI workflows.
Where it falls short
- Requires managing and hosting PyTorch model checkpoints and handling hardware acceleration.
- Long documents above 1,024 tokens require explicitly passing a higher max length and tuning accuracy.
- Fine-tuning requires domain datasets and compute resources like a multi-GPU setup or Apple Silicon.
Pricing & license
Free under its open-source license. The only cost is running it yourself on your own hardware or cloud infrastructure.
Open remote AI-training roles
880 open remote AI-training roles, updated November 15, 2026; most pay $20–$175 an hour. Zekai lists them free; the application goes to the employer’s platform.
Alternatives
Use it, or have us set it up
FAQ
How does Laya avoid text generation hallucinations?
It uses non-autoregressive encoders to output structured choice, score, and yes/no probabilities directly in a single forward pass rather than generating text tokens sequentially.
Can I run Laya on non-English text?
Yes, the built-in router detects non-English scripts and automatically routes requests to the multilingual mmBERT checkpoint supporting 100+ languages.
What hardware do I need to run Laya?
It runs on CPU or GPU via PyTorch, with ONNX Runtime and TileLang GPU fast paths available as optional installation extras.
