# Laya

Laya is an open-weight, non-autoregressive AI decision model released by the Indian startup Convai Innovations in September 2026 under the Apache 2.0 license. Like TypeSafe AI's closed Jev, it is a "System One" model: given a piece of state and typed questions it returns typed answers with calibrated probabilities instead of generating text, in a single forward pass of roughly 33 milliseconds, across more than 100 languages.

**Laya** is an open-weight artificial-intelligence decision model released in September 2026 by **Convai Innovations**, a startup based in Kasaragod, Kerala, India, founded by Nandakishor M. Distributed under the **Apache 2.0** license with weights on Hugging Face, Laya is positioned as the open, self-hostable counterpart to [jev](https://www.wikiprompt.org/wiki/jev), TypeSafe AI's closed "System One" decision model released the same month.

## How it works

Laya is **non-autoregressive**: instead of generating text token by token like a large language model, it classifies. Given a piece of state (an email, a support ticket, an agent trace, a JSON object) and a list of typed questions with fixed answer sets, it returns the chosen answers with **calibrated probabilities** in a single forward pass. Because it produces no free-form text, there is nothing to hallucinate and no prose or JSON to parse; downstream code branches directly on the typed result and its confidence.

The main English checkpoint is built on a **ModernBERT-large** backbone with roughly **421 million parameters**, small enough to run on modest local hardware. A multilingual variant covers more than 100 languages.

## Performance claims

Convai reports a median (p50) latency of about **32.8 milliseconds** on a single Tesla T4 GPU, which it compares to 236-276 ms measured for Jev's hosted API, a roughly 7.8x speed advantage on the vendor's own benchmark, at zero marginal cost when self-hosted. As with Jev's launch numbers, these figures come from the vendor's own evaluations and had limited independent verification in the weeks after release.

An important caveat acknowledged by the project itself: **out of the box the base model scores near random on decision tasks** - the headline accuracy numbers are achieved after fine-tuning on the target task. Laya is best understood as an open, efficient decision-model architecture you adapt to your workload, rather than a zero-shot drop-in.

## Laya and Jev

Laya and [jev](https://www.wikiprompt.org/wiki/jev) define an emerging category sometimes called System One models, after the fast, intuitive "System 1" of Daniel Kahneman's Thinking, Fast and Slow: models built for high-volume, low-latency classification, scoring, routing and next-step decisions that sit alongside generative [large-language-model](https://www.wikiprompt.org/wiki/large-language-model)s rather than replacing them. The two differ mainly in distribution: Jev is a closed hosted API, while Laya's Apache-2.0 weights can be run locally (community runtimes exist for Node.js/TypeScript via ONNX Runtime), fine-tuned freely, and deployed without per-call costs.

## See also

- [jev](https://www.wikiprompt.org/wiki/jev)
- [large-language-model](https://www.wikiprompt.org/wiki/large-language-model)
- [generative-ai](https://www.wikiprompt.org/wiki/generative-ai)
- [ai-agent](https://www.wikiprompt.org/wiki/ai-agent)

---
Source: https://www.wikiprompt.org/wiki/laya
License: CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/)
Last updated: 2026-09-22T06:34:46.567222+00:00