Independent, unofficial model guide · verified 22 Sep 2026

A decision model.
Not another chatbot.

Laya reads a state, evaluates typed questions and returns choices, scores and probabilities in one forward pass. It is open-weight, self-hostable and built for the decisions inside software.

Apache 2.0322M–421M parametersthree checkpointslocal inference
01 / state “I was charged twice. Please refund the duplicate.”
02 / typed question Which team should handle this?

choice: billing · technical · sales · other

03 / distribution
billing0.94
technical0.03
sales0.02
other0.01

Illustrative output shape, not a live model response.

421Mlargest checkpointModernBERT-large encoder
~33msreported single-question latencymultilingual checkpoint on a Tesla T4
3decision primitiveschoice · score · noul
45/51languages above the stated usability barMASSIVE intent benchmark

The interface

Three outputs software can use directly.

Laya does not draft replies or explanations. Your application defines the allowed output shape before inference, then decides what to do with the returned probability.

01 / CHOICE

Pick one label

Route a ticket, classify intent or select a tool from options supplied with the request.

How choice works
02 / SCORE

Rate an ordered scale

Estimate urgency, severity or frustration across levels, returning a distribution and expected score.

How score works
03 / NOUL

Return P(true)

Answer a bounded yes/no question such as whether a message explicitly requests a refund.

How noul works

Positioning

Same problem shape as Jev. A different trade.

Jev is a managed TypeSafe API. Laya publishes downloadable weights and training code. That changes privacy, operations, customization and where latency is spent.

Laya makes sense when…

  • data should remain local or offline;
  • you have domain labels and can fine-tune;
  • high-volume local inference offsets serving costs;
  • you want control over the weights and runtime.

Jev makes sense when…

  • you want a managed API rather than model operations;
  • the label set is large or changes often;
  • stronger out-of-box behavior matters more than customization;
  • network access is acceptable.

Read this before adopting

Open does not mean universally capable.

Laya is promising, young and highly task-dependent. The project’s own documentation is unusually direct about where the current checkpoints fail.

Strong claims need context

The 0.766 typed-decisions result comes from a checkpoint fine-tuned on that benchmark’s training split. The base checkpoints score near chance on the same suite.

Confidence needs validation

The shipped checkpoints can be over-confident. Fit calibration temperatures on held-out domain data before using a probability threshold to automate actions.

Large label sets are a weak point

At default token budgets, 50–77 option questions can become hard to distinguish. Jev currently leads clearly on Banking77.

“No hallucination” is narrow

A typed model cannot produce an answer outside your schema. It can still confidently select the wrong answer inside that schema.

Independent, unofficial guide. This site is not affiliated with or endorsed by Convai Innovations, the Laya maintainers, TypeSafe AI or Jev. Metrics are labelled by source and should not replace evaluation on your own traffic.