Laya is a non-autoregressive System 1 decision engine. It makes typed choice, score, and yes/no decisions in one forward pass. It supports 100+ languages. A router picks the best checkpoint per request. It processes up to 8,192 tokens with a max_len setting. Fine-tuning on domain data boosts accuracy from 0.362 to 0.766. It runs in 33 ms per decision. It works zero-shot out of the box. The README now leads with installation, quickstart, and docs. A new documentation site includes an API reference.
What changed
Previous versions handled shorter documents and fewer languages with lower accuracy on typed decisions.
What it unlocks
Making typed choice, score, and yes/no decisions in a single forward pass across 100+ languages with a router selecting the best model per request.
- 33 ms per decision
- 100+ languages supported
- 8,192 tokens max document length
- 0.766 fine-tuned accuracy vs 0.362 base
What you need to act on it
- Python 3.10 or newer
- Optional extras for HTTP, MCP, LangChain, ONNX, or GPU fast path
Sources