Laya
An open-source decision model that generates no text: you give it a state (a message, a ticket, a JSON document) and typed questions, it returns a choice, a score or a boolean with its probability, in a single pass. Three checkpoints from 322 to 421 million parameters, `pip install laya`, Apache 2.0 licence, and it fits in a gigabyte of local memory.
Strengths
- Typed output (choice, score, boolean) with a probability: nothing to parse, no answer outside the list possible
- A decision in 32.8 ms median on a T4, 7.2 ms per question batched, against several hundred milliseconds for a generative call
- Apache 2.0, `pip install laya`, no key and no waitlist: the measurement can be reproduced on your own cases
- A multilingual 322-million-parameter version covering over a hundred languages, and one gigabyte of memory is enough
Limitations
- Generates nothing: no text, no code, no summary. It decides, full stop
- Past twenty possible options, the repository itself admits Jev takes the lead again
- Repository opened on 18 September 2026: the numbers are its author's own and the project has no release history
- Questions must be written as closed lists, which is specification work rather than prompting
Best for
- Classifying, routing or scoring at volume without paying for a text generation per decision
- Putting in place the guardrail that decides to hand over to a human, with a calibrated probability
- Testing the decision-model category with no waitlist and no vendor commitment