What Is Laya AI? Open-Source Decision Model Explained

Laya is an open-source family of non-autoregressive decision models created by ConvAI Innovations. Instead of writing a free-form answer token by token, Laya evaluates structured questions over a state and returns typed outputs that application code can use directly.

The upstream project describes three decision primitives:

• choice — select among named options and return probabilities.
• score — place an input on an ordered rubric.
• noul — return a probability that a proposition is true.

Primary source: Laya on GitHub.

Laya is not a general-purpose text generator

A generative LLM is useful when the output itself is language: an explanation, email, summary, code block, or conversation.

Laya targets a different part of an application. It is designed for places where software needs a compact decision such as:

• Which support queue should receive this ticket?
• How urgent is this request?
• Is this message likely to be phishing?
• Which model should handle the next task?
• Does this user appear likely to churn?

That means Laya is better understood as a decision layer than a chatbot.

How Laya works

The application provides two things:

1. A state, such as text, JSON, a ticket, an email, or another document-like payload.
2. One or more typed questions about that state.

The model evaluates those questions in a forward pass and returns structured answers.

choice

Use choice when the output must be one label from a defined set.

Examples:

• billing / technical / sales
• safe / suspicious / blocked
• small-model / frontier-model

score

Use score when the output belongs on an ordered scale.

Examples:

• low / medium / high urgency
• mild / moderate / severe frustration
• 1–5 risk level

noul

Use noul for a yes/no proposition where the application also benefits from a probability.

Examples:

• refund requested?
• likely phishing?
• user threatens to cancel?

The current Laya model family

The upstream project currently publishes three checkpoints:

See our Laya Models page for the tradeoffs.

Who makes Laya?

Laya is built by ConvAI Innovations, a small AI company in Kerala, India, led by its founder and CEO Nandakishor Mukkunnoth (Nandakishor M.). He started working on Laya in early 2025 while looking for ways to catch LLM hallucinations before a response is finished, and released the code and weights under the Apache 2.0 license. In a September 2026 interview he said he has no plans to raise funding specifically for Laya (Analytics India Magazine).

Official resources: the Laya page by ConvAI Innovations, the GitHub repository and the Hugging Face models. laya-ai.com is an independent guide and is not run by ConvAI Innovations; see About.

Why Laya is attracting attention

Several properties make the project unusual for a newly released decision model:

• The model weights are published.
• The code is Apache 2.0 licensed.
• It can be self-hosted.
• The upstream project includes a Python package, Router, benchmarks, and a fine-tuning workflow.
• Community ports and runtimes started appearing within days of release.

The important point is not that every benchmark already favors Laya. It does not. The opportunity is that a new open decision-model ecosystem is forming quickly.

Current limitations matter

The upstream project is unusually explicit about several limitations:

• The base checkpoints are weak zero-shot on its typed-decisions benchmark; the stronger 0.766 result comes from the fine-tuned laya-typed-decisions checkpoint.
• High-cardinality choice tasks can degrade at default token-budget settings.
• Calibration needs to be evaluated before using probabilities as production thresholds.
• The English and multilingual checkpoints have different strengths.

Read the upstream BENCHMARKS.md before treating a single headline number as representative.

Where Laya fits

Laya can make sense when an application needs frequent, typed decisions and one or more of these matter:

• local inference,
• open weights,
• fine-tuning,
• predictable structured outputs,
• privacy-sensitive routing,
• high-throughput classification,
• confidence-based escalation.

It is less obviously suitable when the task requires long-form reasoning or when a decision contains a very large label set without additional design work.

Security testing is one early research use: a September 2026 arXiv paper studies Jev and Laya as the decision layer of LLM-driven penetration-testing agents, to cut false positives, inflated severity ratings and wasted LLM calls. Its case study is a single run, so treat it as a direction rather than a result (arXiv 2609.28940).

Laya vs Jev

Laya and TypeSafe Jev target a similar problem category, but their product models differ. Laya is open and self-hostable; Jev is provided as a hosted model API. Capability comparisons depend heavily on checkpoint, tuning, dataset, language, and option count.

See Laya vs Jev for a source-backed comparison.

Because the base checkpoints are weak zero-shot on custom workflows, most real deployments fine-tune first: see How to fine-tune Laya.

Ready to try it? How to use Laya walks through a real task step by step.

FAQ

Who created Laya?

Nandakishor Mukkunnoth, founder and CEO of ConvAI Innovations in Kerala, India. He began the project in early 2025 and open-sourced it under Apache 2.0.

Is Laya free?

Yes. The code and weights are released under Apache 2.0, so you can use, modify and self-host Laya, including commercially, without per-request fees. You only pay for the hardware you run it on.

Is Laya available on OpenRouter?

No. Laya is not offered as a hosted API on OpenRouter; you run it yourself. OpenRouter serves TypeSafe's hosted model, Jev (see What is Jev?). To run Laya, follow Install Laya or Laya and Ollama, or try it in the browser on the Playground.

Last verified: September 22, 2026.