About this site

What we run

Laya answers typed questions in one forward pass. A call returns a choice label, a score on an ordered rubric, or a noul probability, which is P(true) from 0 to 1. It does not write a paragraph. This site exposes that contract as POST /v1/decide. The same body is accepted at POST /v1/systemone and at POST /api/decide. The browser never talks to the upstream inference host. The server holds the inference key and proxies the request.

You can try a ticket in the tools catalog or on the home playground before you create a key. The first anonymous call is free. After that, sign in with Google and create a laya_ key, or keep calling with a signed-in session. The first successful Google registration starts a monthly grant of 10,000 input tokens. Prepaid packs add balance on top. Output tokens are not billed. Prices, pack sizes, and the request-per-minute ceiling for each pack are on pricing.

The tools are the same decide call with the questions filled in. Support triage, email triage, a prompt guard, content moderation, and scam spotting share one balance. The field list for each tool is on its page, and the HTTP paths are in the API reference. POST /v1/gate is separate: it applies a local allow, review, or block policy to answers you already have. It does not call the model and does not spend tokens.

If the server has no inference key, decide returns 503 and the playground can label a run as an Offline demo. That demo does not load the weights. The status page shows whether this deploy is configured, which model id it will send, the anonymous free-run allowance, and whether Creem checkout is on.

What we do not claim

The weights and the laya package are Apache-2.0, published upstream by their maintainers. Installing the package and calling Router().predict is documented on get started. That local path is not billed by this website. A local router can also pick the English or multilingual checkpoint from the script of the text. This hosted API does not add that routing for you. You pass a model alias, or you accept the server default, convaiinnovations/laya, unless the deploy sets a different model id.

The number 0.766 is not what a fresh install scores. The upstream README and the model card report 0.766 for the fine-tuned laya-typed-decisions checkpoint on that benchmark. The English and multilingual bases on the same set are 0.362 and 0.352. The fine-tune guide and the benchmarks page keep those sources separate. On this gateway the alias typed-decisions resolves to convaiinnovations/laya, the same id as english. It does not silently switch the hosted API onto the fine-tuned checkpoint.

This site is not affiliated with Convai Innovations, NandhaKishorM, or TypeSafe, and it is not an official Convai product. LayaAir is a game engine from Layabox. The projects are unrelated. Upstream source, the model card, and the package page stay linked for attribution: the GitHub repository, the Hugging Face model card, and PyPI.

Who operates it, and how to write

The site is operated as laya-model.com. Sign-in is Google. Checkout is Creem when that provider is configured, and Stripe only when the deploy is switched to Stripe. Analytics is Google Analytics 4. Advertising, when shown, may be served by third-party vendors including Google. The privacy policy states what is stored and how those cookies work. The terms cover the prepaid API.

There is no phone line and no chat widget. Email support@laya-model.com, or use the contact page. Do not send API keys, card numbers, or ticket text that contains secrets. Longer explanations of the request shape, the meter, and a support-triage workflow are in the guides.