Laya AI

FIVE MODELS · ONE ACCOUNT

Which model should read your input?

Laya AI started with two compact Laya models for short text. It now also runs Jev, Mercury Decide and Jev-Omni, so longer documents and media get typed answers from the same account, wallet and API key.

Open the playgroundModel IDs in the docs

START FROM YOUR INPUT

If your input is…

The question types are the same everywhere. What changes is how much each model can read and what kind of input it accepts.

  1. One message, review or comment in English

    Laya EnglishSmallest budget, built for a sentence or two.
    Try it

  2. A short message in Chinese, Spanish, Arabic or another language

    Laya MultilingualCovers 100+ languages within 1,024 tokens per question.
    Try it

  3. A full ticket thread, contract clause or JSON record

    Jev64k tokens of context, so nothing needs trimming.
    Try it

  4. Mid-length text where you want a second, independent model

    Mercury DecideSame request shape as Jev, 32k tokens, built for throughput.
    Try it

  5. A photo, voice note or short clip

    Jev-OmniReads the media and picks one of your answers.
    Try it

SIDE BY SIDE

What each model accepts

ModelMade byInputSize limitQuestionsAPINow
Laya English
laya-english		Convai InnovationsShort English text512 tokens per questionYes/no, choice, scoreYes…
Laya Multilingual
laya-multilingual		Convai InnovationsShort text, 100+ languages1,024 tokens per questionYes/no, choice, scoreYes…
Jev
jev-latest		TypeSafeLong text and JSON64k tokens per requestYes/no, choice, scoreYes, also jev-1.13.0 and jev-preview…
Mercury Decide
mercury-decide		InceptionText and JSON32,768 tokens per requestYes/no, choice, scoreYes…
Jev-Omni
jev-omni		Independent, open weights on Gemma 4 12BImage, audio, video or text8,192 tokens; 6 MiB fileOne choice per requestPlayground only…

Oversized input is rejected before any charge on every model; nothing is silently cut. If a model is offline, requests to it fail before billing and are never answered by another model.

SWITCHING

Change one field, keep the rest

Every text model uses the same request: a state, a model and typed questions. Start on Laya English, and when a case turns out to need more context, change model to jev-latest or mercury-decide without touching your questions or parsing code. The Decisions-compatible endpoint also accepts ~typesafe/jev-latest and inception/mercury-decide.

{
  "model": "jev-latest",
  "state": "…the full support thread…",
  "questions": {
    "escalate": { "type": "noul", "instructions": "Escalate to engineering now?" }
  }
}

BILLING

One wallet for every model

Playground

Each run uses one credit first; if credits run out, it uses the actual input tokens instead. This is the same for all five models.

API

Requests use paid input tokens first, then one credit for the whole request when tokens are insufficient. Output tokens are free. Failed requests are refunded.

See plans →