FIELD NOTES / SEPTEMBER 2026

Laya Model: Decisions for AI Workflows

Laya is an open-weight model for choosing among defined options, scoring an input, or estimating a yes/no answer. Use it when a workflow needs a decision rather than a written response.

Sources: developer model card and developer repository. Source-based guidance; this site has not completed an independent model benchmark.

Where it fits

Incoming message → Laya classifies the request → your workflow selects a queue → a person reviews uncertain cases.

A decision model is one component of an agent system. Your application still needs to handle permissions, tool execution and errors.

Three useful experiments

Choose the checkpoint deliberately

The project publishes English, multilingual and task-tuned checkpoints. Its documentation describes different accuracy and context limits across them. Record the exact checkpoint and version; a result from one is not evidence about all of them.

What to measure before using it

Build a small labelled set from your actual task. Include short messages, mixed intents, unfamiliar language and missing context. Measure per-category mistakes as well as overall accuracy. A high confidence value alone does not establish that a decision is correct.

Compare against a simple rules baseline. If a few reliable rules solve the task, a model may add unnecessary latency and maintenance.

Our first local check

We ran eight synthetic English email-routing examples on a Windows CPU. All matched their expected labels in this one smoke test. Inspect the cases, raw outputs and timing limitations; this is not a general accuracy benchmark.

Start a reproducible test

Local setup checklist

Download an email-routing experiment with labelled examples and a script that records raw outputs.

Common questions

Does Laya write email replies?

Its documented purpose is structured decisions. Use a separate writing model or template after routing if you need a reply.

Is this a hosted Laya demo?

No. These pages provide a guide and a downloadable local experiment. They do not send your text to a model.