Laya is an open-source model that reads a message and answers the questions you set, such as which team should handle it, without writing a reply.
A ticket, three questions, no essay
The project's own quickstart uses one billing email: "Hi, we were billed twice for March. Please refund the duplicate today or we will cancel our plan." You hand Laya that text plus the questions. It does not draft the refund reply.
- Which department: billing, technical, or other. On this email the documented answer is billing.
- How urgent, on a scale you define, such as not urgent, soon, or blocking.
- A yes-or-no probability, here "does the user threaten to cancel." The field is a probability, not a sentence.
Three checkpoints, and a router picks one
- English, 421M parameters, 512-token context. The router uses this for the billing email above.
- Multilingual, 322M parameters, for 100+ languages. The same department question in Hindi or Spanish is documented as routing here.
- A third checkpoint fine-tuned on four workflows: invoices, security incidents, customer service, and agent traces.
- Install with pip install laya. Apache 2.0. Weights are on Hugging Face under convaiinnovations/laya, and only the checkpoint you ask for is downloaded.
The base checkpoint is a starting point
On the project's typed-decisions benchmark, the base English checkpoint scores 0.362 and the fine-tuned one scores 0.766. Those are Laya's own numbers. The README also says one question takes about 33 ms on a T4, and about 7.2 ms per question when batched. That timing is their measurement, not an independent test.