Multimodal Jev

Jev went multimodal with jev-1.14.0: a decide call can now carry images alongside your text state, and a new Box question type returns where a described target sits in a picture. It's still a decision model — you get back typed, calibrated values, not a paragraph — only now the context it judges can include what's in an image. These are opt-in preview features: you pick the model and send one header, and nothing changes for your existing text calls.

Multimodal Jev doesn't turn Jev into an image generator or a vision-language chatbot. Jev is TypeSafe AI's System One model: you hand it state plus one or more typed questions, and it returns structured values — choice, score, noul, and now box — in a single non-autoregressive parallel pass. "Multimodal" here means the state it reasons over can include images, so the same calibrated decide call can judge a photo, a screenshot, or a scanned document, not just text.

What's new in Jev 1.14

jev-1.14.0 (also reachable as jev-preview) is the version that adds the vision path. It introduces the box question type and is required for it; images themselves are accepted by both jev-1.13 and jev-1.14. The long-context ceiling also lifts in this line. Everything here is marked preview by TypeSafe — the field and header names are theirs, and the exact shapes and limits may still move — so pin the version explicitly rather than relying on jev-latest if you depend on vision.

• Images — send pictures in a top-level assets map and reference them from your state or questions
• box — a new question type that locates a described target in an image as fractional coordinates
• Long context — raise both input budgets to ~262k tokens with one header, at the same per-token price
• Opt-in — every vision feature is gated behind an official TypeSafe-Danger-Zone-* header; text calls are unaffected

Send an image

Put each image in a top-level assets map keyed by a reference string, then mention that reference from state, a question's instructions, or its criteria. Over HTTP an asset is a base64 payload with a format; you opt in with the TypeSafe-Danger-Zone-Images header — without it the request is rejected. Through this site's gateway you send exactly the same request you'd send to TypeSafe directly; the danger-zone header is forwarded verbatim.

curl https://jevtypesafeai.com/api/v1/decide \
  -H "Authorization: Bearer $JEV_KEY" \
  -H "Content-Type: application/json" \
  -H "TypeSafe-Danger-Zone-Images: true" \
  -d '{
    "model": "jev-1.14.0",
    "state": { "photo": "shot" },
    "assets": {
      "shot": { "type": "image", "data_base64": "<base64 jpg>", "format": "jpg" }
    },
    "questions": {
      "tidy": { "type": "noul", "instructions": "Is the kitchen in `photo` tidy and free of hazards?" }
    }
  }'

The Box question — locate something in an image

A box question answers where, not whether. It returns a box with top / bottom / left / right as 0–1 fractions measured from the top-left corner, plus a confidence. It always returns a box — there is no "not found" answer — so pair it with a noul question that checks the target is actually present before you trust the coordinates. box requires jev-1.14.0 (or jev-preview) and the same images header.

// Locate a target in an image, and check it's really there first.
const res = await fetch("https://jevtypesafeai.com/api/v1/decide", {
  method: "POST",
  headers: {
    "Content-Type": "application/json",
    Authorization: `Bearer ${process.env.JEV_KEY}`,   // jv_live_...
    "TypeSafe-Danger-Zone-Images": "true",            // forwarded verbatim
  },
  body: JSON.stringify({
    model: "jev-1.14.0",                               // box needs 1.14.0 / jev-preview
    state: { photo: "shelf" },
    assets: { shelf: { type: "image", data_base64: imgB64, format: "png" } },
    questions: {
      present: { type: "noul", instructions: "Is there a red mug in `photo`?" },
      where:   { type: "box",  instructions: "the red mug in `photo`" },
    },
  }),
});
const { answers } = await res.json();
if (answers.present.noul > 0.8) crop(answers.where.box); // {top,bottom,left,right}, 0–1

Opt-in headers, model, and limits

• Model: pin model: "jev-1.14.0" (or "jev-preview") for box; images also work on jev-1.13. We never override a model you set.
• Header: send TypeSafe-Danger-Zone-Images: true for any request with images or a box question, or it's rejected.
• Long context: add TypeSafe-Danger-Zone-Long-Context: true to raise the input budgets to ~262k tokens at the same per-token price.
• Size: the whole request body (images included) is capped near 32 MiB — an over-size request comes back 413, not a silent truncation.

What it costs

Images are billed as ordinary input tokens — the same per-token rate as text, with output free. A single image decision is still a fraction of a cent; a client-side character count won't estimate image tokens well, so the reservation is a floor that reconciles to real usage after the call. For current per-million rates (they fall as your prepaid pack gets larger) see the pricing page.

FAQ

See also: Jev for computer use · Jev context window & long context · Docs — Vision & advanced · API reference · Pricing