Laya
Open decision model from India that classifies text, emails or JSON and answers fixed questions with probabilities instead of free text.
What it does
Laya answers structured questions about a piece of content without writing text. You pass in a ticket, an email or a JSON object, for example, and set the type of each question: a choice between options, yes/no, or a level on a scale. You get the answers back with probabilities, all questions in a single pass. Typical tasks are routing requests to the right department, rating urgency, or checking a step before an AI agent carries it out. There are three variants: an English one (421M parameters), a multilingual one for more than 100 languages (322M parameters, reads up to 8,192 tokens) and one fine-tuned for four business workflows. A built-in router detects the language and picks the variant. You use Laya as a Python package, as your own HTTP server, as an MCP server or with LangChain; the server accepts the same requests as TypeSafe’s closed decision model Jev.
Who it suits
Development teams that want to automate many similar decisions, for instance in customer service, invoice handling or security alerts, and are looking for a small model they can run in-house. It pays off most when you have your own example data for fine-tuning.
Data protection for businesses
Laya runs entirely on your own hardware; no data goes to a vendor, so no data processing agreement is needed. The models are downloaded once from Hugging Face. The bundled HTTP server accepts requests without authentication unless you set an API key – always do so in production.
Limits
According to the vendor, the base models are barely better than guessing on business decisions without fine-tuning: in its own test they get 36% right, the fine-tuned model 77%. Out of the box the probabilities are too high and only become reliable after calibration on your own data. Accuracy drops sharply for questions with very many options, and scale questions are the weakest question type. The quoted 33 milliseconds apply to a GPU; on a CPU the vendor reports 193 to 464 milliseconds. The vendor did not measure the comparison figures for Jev itself.