Independent community site · not LayaAir

Test Laya on your support decisions.

Run Laya AI on the official Convai Hugging Face Space, or install it and run locally. It is an open-source System-1 decision model: choice, score, and noul in one forward pass. Inference via official Convai HF Space uses their queue and their hardware.

On typed-decisions (2,000 decisions) the English base is 0.362 and the multilingual base is 0.352, both under a 0.461 majority baseline. 0.766 is laya-typed-decisions after fine-tuning. Those figures are on the Hugging Face model card and the README. They are not a zero-shot result of pip install laya.

What this System-1 decision model does

Laya AI is an open-source System-1 decision model from Convai Innovations / NandhaKishorM. One forward pass returns typed answers over text. It does not generate a paragraph you have to parse. You define the questions. The checkpoint fills choice, score, and noul. Run it on the official Hugging Face Space or locally with python -m pip install laya.

choice
One label, a probability per option, and a confidence. Routing, intent, topic.
score
Expected level on an ordered rubric. Urgency, frustration, severity.
noul
P(true) from 0 to 1. Phishing, churn, “does this statement hold?”.

Laya vs Jev, from the README

These cells are the upstream README’s routed comparison. This site did not remeasure them. The 0.766 cell is laya-typed-decisions after fine-tuning. On the same 2,000 decisions the English base is 0.362 and the multilingual base is 0.352, under a 0.461 majority baseline. pip install laya does not apply that fine-tune. The full split, including third-party cards, is on Laya vs Jev.

Source: NandhaKishorM/laya README, “Laya (with routing) vs Jev”. Jev latency citations in that README: AbdelStark/jev-benchmarks and nibzard/decision-model-benchmark.
Jev 1.13.0Laya (routed)
typed-decisions, 2,000 decisions0.7270.766
AG News, 4 labels0.9100.950
DAIR Emotion, 6 labels0.4800.595
Banking77 (72 vs 77 labels)0.8700.425
ECE, lower better, post-temperature0.2460.081
p50 latency, 1 question236–276 ms32.8 ms

Run locally

Python 3.10 or newer, then python -m pip install laya. The README does not publish a minimum VRAM. After preload, one multilingual question is 32.8 ms on a Tesla T4 and 193–464 ms on CPU. max_loaded=1 reloads on a language switch (7.4 s median on CPU, 10.3 s on a T4). Details: run locally.

Laya AI is not LayaAir

LayaAir is the Layabox game engine. This page is about the Apache-2.0 decision checkpoint in NandhaKishorM/laya. The site is an unofficial community project and is not affiliated with Convai Innovations. The 中文页面 repeats that distinction.

Read this before the first real run

Three checkpoints, one router

Router().predict picks a checkpoint before the forward pass. The typed-decisions checkpoint is not selected automatically.

From the upstream README checkpoint table. Context for multilingual can be raised to 8,192 tokens with max_len=8192.
CheckpointEncoderParamsContextUse it for
layaModernBERT-large421M512English
laya-multilingualmmBERT-base322M1024, up to 8192100+ languages
laya-typed-decisionsModernBERT-large421M1024Four typed-decision workflows, after fine-tuning

Questions

Is this LayaAir?

No. LayaAir is a game engine from Layabox. This site documents Laya, the open-source System-1 decision model published by Convai Innovations and maintained by NandhaKishorM. The projects are unrelated.

Does this website run the official Laya weights?

The Live (HF Space) frame and the Run live button call the official Convai Hugging Face Space. Inference via official Convai HF Space: the checkpoint stays on their hardware, and answers depend on that Space being up and on its queue. This site does not download the weights. Offline demo is a keyword heuristic labeled “Offline demo” with the Router().predict shape, for when the Space is queued or down. A local install is python -m pip install laya.

Does pip install laya score 0.766 on typed decisions?

No. The upstream README reports 0.766 for the fine-tuned laya-typed-decisions checkpoint on that benchmark’s 2,000 decisions. The base English and multilingual checkpoints score 0.362 and 0.352 on the same decisions, under the 0.461 majority-class baseline. Installing the package does not fine-tune the model.

What are choice, score, and noul?

They are the three typed questions Laya answers in one forward pass. choice returns a label and a probability per option. score returns an expected level on an ordered rubric. noul returns P(true), a probability from 0 to 1. The model does not generate text.

When does the README say Jev leads?

On high-cardinality label sets. The README’s Banking77 row is 0.870 for Jev 1.13.0 on 72 labels and 0.425 for routed Laya on 77 labels at the default head budget. Jev also has higher soft accuracy on typed-decisions (0.580 vs 0.471) and better raw ECE before Laya’s temperature fit (0.144 vs 0.213).

Is this site affiliated with Convai Innovations?

No. This is an unofficial community documentation and demo site. It is not affiliated with Convai Innovations, NandhaKishorM, or TypeSafe. The model and the laya package are Apache-2.0.

Laya AI — Open Source System-1 Model You Can Run Locally