You give Laya two things: a piece of text and a list of questions. Each question has a fixed shape. A yes or no question ("Does the customer threaten to cancel?"), a choice between options you name ("Which team should handle this: billing, technical, or sales?"), or a score on a scale you describe ("How urgent is this, from not urgent to blocking?"). It answers all of them in one pass, roughly 30 to 40 milliseconds on a modest GPU, and every answer comes with a probability. If it says billing at 0.94, that number is meant to be trusted: the model was trained so that reporting an honest probability is the only way to score well.
That is the part that makes it different from asking a chat model. Laya never generates text. There is no reply to parse, no made-up detail to check, and no waiting for tokens. The answer space is whatever you define in the question, so you can add a new label or a new question without retraining anything.
It comes in three checkpoints under one repo. The English one is the default. A multilingual one covers more than 100 languages and reads up to about 8,000 tokens of text, so a long document works too. A third is fine-tuned for four business workflows (invoices, security incidents, customer service, agent logs) and shows how much accuracy jumps once you train it on decisions from your own domain. A built-in router picks the right checkpoint by detecting the language of the input.