Laya

An open-source alternative to Jev. If you do not want a closed, vendor-hosted model, self-host this one and call the API: typed yes/no, choice and score decisions with probabilities, on CPU, in your own project.

What it does

Laya is a non-autoregressive decision engine. You POST a piece of state — an email, a ticket, a support message, a JSON object — together with typed questions, and it answers each one in a single forward pass.

• choice — picks one of your options, with a probability for each
• score — an expected value over an ordered scale
• noul — the probability that something is true

No text is generated, so there is nothing to parse and no format to repair. Typical uses: routing tickets, classifying email, scoring urgency, gating an LLM call.

Jev is TypeSafe AI's proprietary decision model. Laya gives you the same three primitives with an open Apache-2.0 model, on hardware you control. Different wire format, and not affiliated with or endorsed by TypeSafe.

Why run it yourself

• Your data stays in your project. Tickets, emails and documents are scored on your own machine, not posted to a model vendor.
• No per-token bill. You pay for a running container, which is the difference that matters when the job is "label every inbound message".
• Nothing is downloaded at runtime. The 842 MB checkpoint is baked into the image and the container never reaches the Hugging Face Hub, so what you deploy is what was built and tested.
• Interactive docs at /docs, where you can compose a request and send it without writing a client.

Deploy

1. Paste an API key of your choosing. That is the only input; nothing to download.
2. Click Deploy. The service is live in about 2 minutes.
3. Most of that is the model loading. /decide answers 503 until it finishes, and /healthz says when.

Use it

curl -X POST https://YOUR-SERVICE-URL/decide \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" -d '{
  "state": "We were billed twice for March. Refund it today or we cancel.",
  "questions": [
    {"type": "choice", "instructions": "Which department?", "options": ["billing", "technical", "sales"]},
    {"type": "score", "instructions": "How urgent?", "options": ["not urgent", "soon", "blocking"]},
    {"type": "noul", "instructions": "Does the user threaten to cancel?"}
  ]}'

Each answer carries its probabilities and a confidence; the response carries latency_ms. Basic auth sends the same secret in RFC 7617 form: curl -u api:YOUR_API_KEY ... (the username is always api).

{"answers": [
  {"type": "choice", "choice": "billing", "probabilities": {"billing": 0.86, "technical": 0.06, "sales": 0.08}},
  {"type": "score", "score": 1.73, "legend": {"0": "not urgent", "1": "soon", "2": "blocking"}},
  {"type": "noul", "noul": 0.76}
], "latency_ms": 321.9}

Act on the numbers in your own code. The thresholds are yours, not the model's.

GET /healthz needs no credential and reports model_loaded. Poll it if you are scripting against the service.

Configuration

Everything else is set for you.

Limits

• Input past ~1K tokens is silently truncated. Keep states short.
• English only. The original project's multilingual checkpoint is not in this template.
• confidence is miscalibrated, by the original authors' own ECE numbers. Rank by it; do not read it as a probability of being right.
• amd64 only.

The first is the deployment fork's measured finding (deploy/DEPLOY-CLOUD.md); the third is the original authors' (BENCHMARKS.md).

FAQ

Is this compatible with Jev's API? No. Same three primitives, different wire format. Not affiliated with or endorsed by TypeSafe.

Why is the first call slow? The 842 MB checkpoint loads into memory on boot, which takes about 45 seconds. It is baked into the image, so nothing is downloaded; the wait is the load itself. It happens once per boot, so only a restart or a redeploy pays it again.

Does it stop when nobody is using it? No. It stays up, so a request never waits for a boot. That also means it is charged the whole time; delete the service when you are done with it.

Why is the API key not generated for me? A generated value would be stored write-only and you could never read it back. The deploy form starts the field empty on purpose: paste a key of your choosing.

Why is there no volume? Nothing is written between requests. A request carries the state it asks about.

Which checkpoint is this? typed-decisions, ModernBERT-large, 421M parameters, 1024-token context - built by Convai Innovations, the original Laya authors. The one the deployment fork's deploy/ directory ships.

Links

• Original project: Laya by Convai Innovations, who built and trained the model; model card on Hugging Face
• Deployment source: https://github.com/tonychang04/laya-template, pinned to c9dcaab (packages the original project's model behind a deploy directory of its own)
• License: Apache-2.0