Skip to content

What Is the Laya Decision Model?

Laya takes text and a question whose answers you already listed. It returns the answer and a probability. It does not write the paragraph a chat model would write.

Three outputs

Name in the APIWhat you seeWhat it is for
choiceOne category, plus a probability for each optionRoute a ticket or pick a feedback topic
scoreA level on a scale you orderedUrgency or severity, from the levels you wrote
noulA probability that a statement is trueYes/no checks, such as “does this ask for a refund?”

A score is not a forecast of any number you can imagine. A yes/no probability is not proof. Structured output still needs review. “It does not generate text” does not mean “it cannot pick the wrong label”, and it can be confident when it is wrong.

Try one choice before the rest

The support ticket classifier starts with a single department question. After that run, you can add urgency and a refund yes/no. The playground exposes the JSON for all three types.

Three checkpoints

Figures below are from the upstream model card and README, checked on 29 Sep 2026. This site did not remeasure them. Package laya 0.3.21 on PyPI that day is a software release. The 0.3.20 notes say the checkpoints themselves were unchanged.

CheckpointEncoderParamsContextRole
layaModernBERT-large421M512 tokensEnglish
laya-multilingualmmBERT-base322M1,024 tokens, configurable toward 8,192Multilingual. “100+ languages” refers to this checkpoint
laya-typed-decisionsModernBERT-large421M1,024 tokensFine-tune for four typed-decision workflows

The upstream router chooses between the English and multilingual checkpoints. It does not automatically select typed-decisions. On 29 Sep 2026, English examples sent through the public Space playground returned the model field laya.

0.766 is not the base model

On the typed-decisions benchmark (400 cases, 2,000 decisions, four workflows), the upstream card reports 0.766 accuracy for the fine-tuned laya-typed-decisions checkpoint. The English base on the same decisions is 0.362. The multilingual base is 0.342 in the checkpoint table and 0.352 in the honest-limits note and the PyPI 0.3.21 text. Both sit under the 0.461 per-question majority baseline. Installing the package does not apply that fine-tune, and 0.766 is not an accuracy claim for support-ticket routing on this website.

The upstream card also says the base model can miss negations on yes/no questions, and that action.act_probability is not a usable gate. Our recorded “not a feature request” call missed. See the feedback classifier.

Speed numbers are not this website

Upstream reports about 33 ms to 39.5 ms for one question on a GPU under their test conditions, and 32.8 ms for the multilingual checkpoint on a T4 in the router table. A visit to this site also waits on the network, the Space queue, and the browser. The tool prints the Space latency_ms field and the time for that click. Those are measurements of one call, not a guarantee that every request finishes in 33 ms.

Not LayaAir, and not an official site

LayaAir is a game engine from Layabox. This page is about the Apache-2.0 decision model published by Convai Innovations and maintained in NandhaKishorM/laya. Model card. Laya Model Online is unofficial.

A fair reading of the published Jev comparison, including the rows Jev leads, is on Laya vs Jev.