Jev 1.14 in preview: images and the new Box question, not yet for production

TypeSafe AI released a new version of its decision model, labelled jev-1.14.0, on 14 October 2026. S01 It is not yet the default stable version: the jev-latest alias still points to jev-1.13.0, and the new release is available through the jev-preview alias. S01 Alongside it, the documentation gained a Box question type, which locates an object in an image and works only with Jev 1.14. S52 This matters for online shops, because a large part of their data consists of product photos and photos attached to return claims. The vendor, however, explicitly labels both images and Box as experimental for now.

What version 1.14 brings

The models page lists the same price for both versions, 0.042 USD per million input tokens, with output tokens free. The context limits are also identical: 64,000 tokens per request and 32,000 tokens per question. S01 You will not find a detailed list of changes compared with version 1.13 there. S01 In the model table both versions are listed with text input; working with images is described on a separate documentation page. S01S51

The documentation points out that aliases can move without any change to your code, and that the model field in the response shows which specific version answered. Anyone who has tuned thresholds against a particular version should send its exact identifier in the request. S01 In practice, this means that an application calling jev-preview has been getting answers from a different model since 14 October than before. confidence thresholds set for version 1.13 need to be re-checked on your own data before switching. The procedure is described in the guide confidence and calibration.

Images: an experimental mode with strict limits

According to the documentation, both versions, 1.13 and 1.14, accept images. S51 The mode is off by default, and every request containing an image must carry the header TypeSafe-Danger-Zone-Images: true. S51 The vendor labels it “Dreamer-ready”, meaning suitable for prototypes, not for production use. It warns that the request format, the answers and the limits may change between releases, and that accuracy will also change as the models are updated. S51S52

An image can be used in every question type described in the guide Choice, Score and Noul. With Choice, example photos can sit directly in the option descriptions, with Score in the level descriptions, and with Noul in the definitions of both the yes and the no answer. S51 Images are not charged separately: they count as input tokens at the same price as text and take up room in the context limit. The documentation so far only announces guidance on how to estimate the number of these tokens. S51

The technical requirements are strict. The API accepts only still PNG and JPEG images; formats such as GIF, WebP or HEIC have to be converted first. A single file may be at most 16 MiB, width and height must each be between 256 and 1568 pixels, and the body of the whole request must not exceed 32 MiB. S51 The API rejects an image outside these limits and does not shrink it. A single rejected image is enough for the whole request to fail, with no partial answers. S51 The API also strips EXIF metadata and does not rotate the image according to it, so the photo has to be straightened before it is sent. S51

Box: exactly where the object is

The answer to a Box question is the four edges of a rectangle (top, bottom, left, right). Each one is expressed as a fraction of the image size from 0 to 1, measured from the top-left corner, so to crop you simply multiply the value by the width or height in pixels. S52 The answer also includes confidence. The documentation does not yet give a recommended threshold; it uses 0.6 only as an example and advises setting it based on your own images. S52

An important caveat: Box always returns a rectangle, even when the object you are looking for is not in the image at all. The documentation therefore recommends adding a Noul question that first establishes whether the object is there. S52 One question looks for one target, and the model can miss a small object in a very large image; in that case it helps to identify the region first and search only within it. S52 The documentation also reports a vendor test on 36 crops of paintings where part of the picture was known to be missing. Four questions about the individual edges caught 30 cases, a single overall question 26, and drawing the rectangle onto the full image 20. S52 This is a vendor figure, not an independent measurement.

The Python SDK, in version 0.7.5 of 14 October 2026, added support for Box questions. Images can be created in it from bytes, a file path, a base64 string or a PIL object, and the request body is now compressed with the zstd algorithm by default. S53

What this means for a Slovak online shop

The API will accept neither supplier photos in WebP format nor phone photos in HEIC format. An online shop therefore has to convert them to JPEG or PNG before sending, resize them so that the longer side is at most 1568 pixels, and rotate them into the correct orientation. S51 Otherwise the request ends in an error. Candidates for a prototype include sorting a photo from a return claim with a Choice question that has example images for the options S51, or checking whether a product photo shows the whole item with a Noul question combined with Box. S52 These are suggestions from JEV.sk, not proven deployments.

For production, the stable jev-latest remains the right choice for now. The vendor treats images and Box as experimental, and the models documentation also states that Jev is most accurate in English and that other languages should be tried on your own content. S01S51 That applies to Slovak descriptions in questions as well. How many tokens images actually consume can currently only be read from the usage.input_tokens field in the response. S51 The cost calculator estimates the cost of text questions, and the guide pricing and limits summarises the details of the price.