Laya runs typed classification, scoring and confidence gates locally in one model pass, with separate English, multilingual and task-tuned checkpoints.
Yes, Laya is open source under the Apache-2.0 license.
Laya is free to use.
Yes, Laya can be self-hosted (the source is available under the Apache-2.0 license).
Apache-2.0 local decision engine for a narrow job generative models are often overused for: take text or JSON, answer a fixed set of typed questions, and return labels, scores and probabilities without generating prose. The Python package exposes choice, ordinal-score and boolean-probability primitives, plus presets for model routing, prompt guardrails, moderation and support-ticket triage. Three open checkpoints cover English, multilingual input and the project's task-tuned workflows.
Support routing, lead classification, churn flags and prompt screening are decision problems before they are writing problems. Laya gives those steps a local, inspectable probability instead of paying a large model to write a sentence that another rule has to parse. That makes it useful ahead of a CRM, support queue or agent tool call where low confidence should hand the case to a person. The honest read is in the project's own limitations: use the task-tuned checkpoint for its advertised workflows, fit calibration on your data, and do not treat a confident score as proof when the input or label space falls outside what the model handles well.