Laya is an open-source family of small encoder models from Convai Innovations that answers typed questions about a text, such as yes or no, pick one, or rate on a scale, without generating any words.
In plain words
Chat models answer by writing text, which you then have to parse and which can be made up. Laya instead reads the input once and returns answers with probabilities, never generating text. You define the questions and allowed answers when you send the request, so new question sets need no retraining. The English model is built on ModernBERT-large; a multilingual one uses mmBERT-base. It works out of the box, but its makers say accuracy rises most after fine-tuning, and its probabilities start out over-confident.
A quick example
The model card's example is a support email about being billed twice, with a threat to cancel. Three questions go in together: which department should handle it, how urgent it is on a three-step scale, and whether the customer threatens to leave. All three come back from one pass, each with a probability.
Where you'll meet it
- Triage of support tickets and emails.
- Guardrail checks that need a quick yes or no.
- A base model to fine-tune for specific decisions.
Related
Sources
- Docsconvaiinnovations/laya, Convai Innovations (Hugging Face) · read 29 Sept 2026
- DocsNandhaKishorM/laya, Laya (GitHub) · read 29 Sept 2026
Checked by WhatIsAI against the sources above · 29 Sept 2026. Spotted an error? Report it.