Open-source System 1 decision model

Laya AI — Open-Source Decision Model

Fast typed decisions for routing, intent classification, urgency scoring, risk detection, and probability outputs—run locally in a single forward pass without generating text.

Install Laya View Laya on GitHub

Apache 2.0Open weights3 checkpoints100+ languagesPython package

State

“Customer was charged twice and asked for a refund.”

Laya

choice

billing

0.94

score

urgent

2.6 / 3

noul

refund?

0.91

One forward pass, typed outputs your application can consume directly.

What Can You Build with Laya AI?

Common Laya AI use cases include agent and model routing, support triage, guardrails, intent classification, risk scoring, and workflow escalation.

Agent and model routing

Choose which tool, agent, or model should handle a request based on the current state and typed criteria.

Support ticket triage

Route tickets by department, score urgency, and flag refund or churn signals in the same decision call.

Guardrails and safety gates

Evaluate jailbreak, policy, abuse, or escalation conditions before a workflow continues.

Intent classification

Map text or JSON inputs to a small, defined set of intents without generating and parsing free-form output.

Risk and fraud scoring

Return structured scores or probabilities for fraud, phishing, churn, or other application-specific risks.

Workflow escalation

Use calibrated decision outputs to automate clear cases and send uncertain cases to a larger model or human review.

How Laya works

How Laya AI Works: Choice, Score, and Noul

Give Laya a state plus typed questions. It evaluates choice, score, and noul decisions in one forward pass and returns structured outputs your application can use directly.

choicelabel + probabilities

Pick one option from named criteria, such as a support department or intent.

Output

→billing · 0.94

scoreordered score

Place an input on an ordered rubric such as urgency, severity, or frustration.

Output

→urgent · 2.6 / 3

noulP(true)

Estimate the probability that a proposition is true, such as churn risk or fraud.

Output

→refund requested · 0.91

Laya AI models

Laya AI Models: Choose the Right Checkpoint

Laya currently publishes three checkpoints. The best choice depends on language and whether the workflow resembles the typed-decisions training setup.

Checkpoint	Encoder	Params	Context	Best for
laya	ModernBERT-large	421M	512	English workloads
laya-multilingual	mmBERT-base	322M	1024	100+ languages
laya-typed-decisions	ModernBERT-large	421M	1024	Typed-decision workflows

Laya model routing

Route each workload to the right checkpoint

Choose the checkpoint that matches your language mix and output type.

English requests

classification, routing, moderation

Multilingual requests

non-English intent and cross-lingual classification

Typed decisions

choice, score, noul, and probability outputs

Laya Router

Base checkpoint

general English decision tasks

laya

Multilingual checkpoint

100+ language decision tasks

laya-multilingual

Typed-decisions checkpoint

structured choice, score, and noul workflows

laya-typed-decisions

Inputs, language, and output shape determine the best checkpoint.

Compare all Laya models

Quickstart

Run Laya AI Locally with Python

The upstream package installs from PyPI and provides a Router that selects a checkpoint based on the request.

API examples are checked against the upstream repository and may change while Laya is evolving quickly.

Read the installation guide

pip install laya

from laya import Router

router = Router(preload=True)

state = {"body": "I was charged twice. Please refund the duplicate."}
questions = {
    "department": {
        "type": "choice",
        "instructions": "Which department should handle this?",
        "criteria": {
            "billing": "payments and refunds",
            "technical": "bugs and outages"
        }
    },
    "refund_requested": {
        "type": "noul",
        "instructions": "Does the user explicitly request a refund?"
    }
}

result = router.predict(state, questions)

Resultdepartment=billing · refund_requested=0.91

Laya vs Jev

Laya vs Jev: Open Weights, Local Inference, and Tradeoffs

Both target typed decisions, but the deployment model, openness, and current tradeoffs differ.

LayaJev

Weights

LayaOpen weights · Apache 2.0

JevHosted model / API

Deployment

LayaLocal or self-hosted

JevManaged API

Fine-tuning

LayaWeights and tuning workflow available

JevNo public weights

Large choice sets

LayaUpstream docs flag default-setting degradation

JevDesigned for larger option sets

Benchmark claims are only useful when checkpoint, tuning, hardware, dataset, and prompt conditions are preserved.

Read the full Laya vs Jev comparison

Laya GitHub ecosystem

Laya Projects and Community Runtimes

Track the upstream model plus fast-moving community runtimes, ports, integrations, demos, and benchmarks. GitHub metrics below are a Sep 22, 2026 snapshot.

Upstream

NandhaKishorM/laya

Reference implementation, models, benchmarks, Router, and fine-tuning workflow.

Stars15.1k

Forks1.2k

PythonPyTorch

Last push Sep 21, 2026

Community

mizorewww/laya-mlx

A community MLX port aimed at running Laya efficiently on Apple Silicon.

Stars5.1k

Forks357

PythonMLX

Last push Sep 22, 2026

Community

receptron/laya

A JavaScript / TypeScript oriented community runtime around Laya.

Stars260

Forks18

TypeScriptONNX

Last push Sep 21, 2026

Explore Laya projects

Laya benchmarks

Laya Benchmarks: Accuracy, Latency, and Limitations

Laya is new and benchmark numbers are moving quickly. We separate upstream claims from independent measurements and keep tuning conditions visible.

Typed decisions0.766

Accuracy reported for the fine-tuned laya-typed-decisions checkpoint on 2,000 decisions.

Upstream benchmark

Single question latency32.8 ms

Reported for laya-multilingual on a Tesla T4 in the upstream benchmark.

Upstream benchmark

Important limitation>20 choices

The upstream project documents degradation on high-cardinality choice tasks at default settings.

Upstream limitation

The 0.766 result is from a checkpoint fine-tuned for the benchmark family; the base checkpoints are much weaker zero-shot on that benchmark.

How to read the numbers

Benchmark results only make sense with their conditions

Keep checkpoint, tuning status, dataset, and hardware attached to every result.

Accuracy example

Checkpoint

laya-typed-decisions

Tuning

fine-tuned

Dataset

2,000 decisions

Result

0.766 accuracy

Latency example

Checkpoint

laya-multilingual

Model state

warm

Hardware

Tesla T4

Result

32.8 ms / query

Large choice sets can degrade at default settings, so test high-cardinality tasks separately.

Explore Laya benchmarks

Laya AI Guides and Examples

Short, source-backed guides for the questions developers are searching for first.

Install Laya locally

Python requirements, package installation, model loading, and the first Router call.

Laya Python guide

Build choice, score, and noul questions and consume typed outputs safely.

What is Laya?

Understand System 1 decision models and where Laya fits next to generative LLMs.

Laya AI FAQ

Answers to the first questions developers have about the model and this site.

Independent community resource. Not affiliated with ConvAI Innovations. Laya names, logos, and related marks belong to their respective owners.

What is Laya AI?

Laya is an open-source family of non-autoregressive decision models that answers typed choice, score, and noul questions over a state instead of generating free-form text.

Is Laya open source?

The upstream project publishes model weights and code under Apache 2.0.

Can Laya run locally?

Yes. The upstream Python package can load Laya checkpoints locally, and community runtimes are already appearing for additional environments.

What is the difference between Laya and Jev?

They target similar typed-decision workflows, but Laya publishes weights for self-hosting while Jev is provided as a hosted model API. Current benchmark tradeoffs vary by task and test conditions.

How do I install Laya AI?

Install the Python package with pip install laya, then use the Router or load a specific checkpoint locally. The installation guide covers Python requirements, preloading, and the first prediction call.

Can I use Laya from TypeScript or on Apple Silicon?

Yes, community runtimes are already available for JavaScript or TypeScript through ONNX and for Apple Silicon through MLX. These are community projects rather than official ConvAI Innovations releases.

Which Laya checkpoint should I use?

Start with laya for English, laya-multilingual for multilingual workloads, and evaluate laya-typed-decisions when your task matches typed-decision workflows. Measure on your own data before production use.

Is this the official Laya website?

No. Laya AI is an independent community resource and is not affiliated with ConvAI Innovations.