What is Jev 1.14?
Jev 1.14 is the second published version of Jev, released by TypeSafe AI on 14 October 2026 with the model id jev-1.14.0. It is out as a preview under the jev-preview alias. TypeSafe describes the update as “small, subtle improvements to everything” and expects the biggest gain in multilingual support. Price and limits match 1.13: $0.042 per million input tokens, free output, 64k tokens per request and 32k per question.
The same release post added a set of opt-in “danger zone” features: image input, a new box question type, much longer context and opt-in data sharing. TypeSafe says they still have too much jaggedness for production. Image input and long context also work on 1.13. Box needs 1.14.
What changed in the model
TypeSafe has not published a changelog, a 1.14 jaggedness page or benchmark numbers for 1.14. The release post gives two things: improvements spread across the model, and a likely gain in multilingual support. That matters because TypeSafe’s docs name English as the primary training language and say other languages are “handled but not equally well.”
The post also mentions an internal eval called “Measuring Intelligence” that compares Jev with open-weight models. It is vendor-run, the post gives no numbers, and TypeSafe itself warns that it may favor its own models. Your own tests are the way to judge the change. Jev benchmarks covers the independent boards.
No 1.14 page existed at TypeSafe’s jaggedness URL on 15 October 2026. Until one appears, the 1.13 jaggedness list on the Jev model page is the only published account of what Jev does badly.
When does jev-latest move to 1.14?
About a week after release. TypeSafe’s post says jev-latest will point to jev-1.14.0 “in one week,” which puts the switch around 21 October 2026. On 15 October the docs still showed:
| Alias | Points to | Docs definition |
|---|---|---|
jev-latest |
jev-1.13.0 |
The most recent stable, official release, and the SDK default |
jev-preview |
jev-1.14.0 |
The most recent release, official or not |
If your code sends jev-latest, or relies on the SDK default, it moves to 1.14 on that date without any change on your side.
How to pin a Jev version
Send the versioned id in the model field instead of an alias: "model": "jev-1.13.0" to stay where you are, or "model": "jev-1.14.0" to move now. The response’s model field names the version that answered, so you can log it and check.
GET /v1/models lists only the aliases at the moment, but the docs say versioned ids work whether or not they appear in that list. Getting access and making your first call shows where model goes in a full request.
The danger zone features
All four are off until you opt in, and TypeSafe says none of them is ready for production use. Images and long context are switched on per request with a header.
Images: TypeSafe-Danger-Zone-Images
Send TypeSafe-Danger-Zone-Images: true and you can put images into a request on 1.13 or 1.14. Images go in a top-level assets map, keyed by a reference string you choose, such as a UUID. Wherever that string appears in state, instructions or criteria, the API swaps in the image.
The limits:
- PNG or JPEG stills only, up to 16 MiB per image.
- 256 to 1568 pixels per side, with a minimum total pixel count that rules out a 256 by 256 image.
- 32 MiB for the whole request body.
- Images count as input tokens at the normal input price. Each use of a reference adds another copy of the image, and you pay for each copy.
A request with an image but no header fails with a 400. TypeSafe’s post says image performance still trails text-only performance.
Box: a new question type
box is a fourth question type next to Choice, Score and Noul. It finds one thing in an image and returns its bounding box: top, bottom, left and right as fractions from 0 to 1, measured from the top-left corner, plus a confidence value. It takes instructions and no criteria.
Box needs both Jev 1.14 (jev-1.14.0 or jev-preview) and the images header. A request using it looks like this:
{
"model": "jev-1.14.0",
"state": { "photo": "3f6c2a9e-8b1d-4e57-a0c4-9d2f61b7e835" },
"assets": {
"3f6c2a9e-8b1d-4e57-a0c4-9d2f61b7e835": {
"type": "image",
"format": "jpg",
"data_base64": "..."
}
},
"questions": {
"mug": { "type": "box", "instructions": "the coffee mug in `photo`" },
"has_mug": { "type": "noul", "instructions": "Is there a coffee mug in `photo`?" }
}
}The second question is there because Box has no “not found” answer. Ask about something that is not in the picture and you still get four edges back. The docs suggest pairing each Box with a Noul about the same target and ignoring the box when the Noul comes back low. Use a reference string that cannot turn up by accident. If the reference were photo, the instruction text above would get the image swapped in too.
Long context: TypeSafe-Danger-Zone-Long-Context
Send TypeSafe-Danger-Zone-Long-Context: true and both context limits rise to 262k tokens, from 64k per request and 32k per question. It works on 1.13 and 1.14 at the same per-token price. The docs say long context is slower and less accurate, and the post still recommends keeping context short. To price a 262k-token call, run the arithmetic in Jev speed and pricing with the bigger input.
Opt-in data sharing
You can choose to share your queries so TypeSafe can find weaknesses in the model. The post says shared queries won’t be used for training, and the docs already state that Jev is not trained on customer requests.
How long will Jev 1.13 stay available?
TypeSafe promises 60 days’ notice before jev-1.13.0 is formally deprecated, and 30 days’ notice for jev-1.14.0. The docs’ models page carried no deprecation notice on 15 October 2026. A formal deprecation policy is promised but not yet published.
What to do if you are on Jev 1.13
- Check which id your code sends. If it is
jev-latestor the SDK default, you move to 1.14 around 21 October 2026. - If you want to stay on 1.13 for now, pin
jev-1.13.0. You get at least 60 days’ notice before it goes. - Run your own evaluation set against
jev-1.14.0before you switch. Any threshold you tuned on 1.13’s probabilities should be re-checked, since TypeSafe has not said how calibration changed. - If you process non-English text, test that first. It is where TypeSafe expects 1.14 to improve most.
- Keep danger zone headers out of production paths until TypeSafe says otherwise.
The Jev model page tracks both versions, prices and limits. What is Jev AI? covers the model itself if you are new to it.
FAQ
Is Jev 1.14 more expensive than 1.13?
No. Both cost $0.042 per million input tokens with free output, and both share the same context limits. Images and long context use the same per-token rate too, though an image adds input tokens and a long-context call sends more of them, so those requests cost more in total.
Does Jev 1.14 accept images?
Yes, with the TypeSafe-Danger-Zone-Images: true header. So does 1.13. Only the box question type is limited to 1.14. Images must be PNG or JPEG stills between 256 and 1568 pixels per side, and a request without the header fails with a 400.
Is Jev 1.14 production-ready?
TypeSafe released it as a preview and plans to make it jev-latest, the alias it defines as the latest stable, official release, about a week later. The danger zone features are a separate matter: TypeSafe says they are not ready for production on either version.