Laya

Laya is a 0.4B-parameter text-classification model built for System 1 style decisions rather than generation. You hand it a state — text, email, ticket, or JSON — plus typed questions, and it returns typed answers with mathematically calibrated probabilities in about 33 milliseconds, across 100+ languages. It is trained with reinforcement learning against strictly proper scoring rules (RLCD), so reporting honest probabilities is the only way for it to maximize reward. Because it never generates text, there is nothing to parse and nothing to hallucinate, which makes it a natural fit for routing, scoring, moderation, and guardrail pipelines. The laya-multilingual checkpoint reads up to 8,192 tokens. It was released on 2026-09-18 under Apache 2.0 and has pulled roughly 50k downloads and 5.6k likes on Hugging Face.

• Long-document accuracy: 16–18 of 20 requests correct with up to ~4,000 tokens of preceding text, degrading beyond that — vendor-reported (own research scripts)
• Single forward pass latency ~33 ms — vendor-reported