I Tested Jev’s Visual AI On 30 Images Of My Own

🔍 Read the full analysis: I Tested Jev’s Visual AI On 30 Images Of My Own on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get the little things that make your day delivered free with Prime

• Fast, free delivery on millions of items
• Prime Video, Amazon Music and more included
• Member-only deals all year

Start your free Prime trial →Free trial for eligible customers · Cancel anytime

As an affiliate, we earn on qualifying purchases.Hide

Thorsten Meyer tested TypeSafe’s new Jev image features on 30 images from his website after the company released Jev 1.14 on October 14, 2026. In the reported results, Jev matched product photos to their associated headlines with strong separation, while TypeSafe cautioned that the image capabilities remain experimental and are not recommended for production use.

Jev is a decision model that responds to structured questions with outputs such as probabilities, selections from a list or scores. Before the release, those questions were text-only. TypeSafe’s update allows users to provide images in a request, and Jev can return a bounding box for a target object. Both features require explicit opt-in headers labeled as danger-zone capabilities.

Meyer tested the features on the 30 most recent featured images on ThorstenMeyerAI.com: 10 product roundups, 15 news posts and five essays. He asked whether each image suited its own headline, a randomly selected headline from another post, and an unrelated plant-care headline. The 10 product images also received a multiple-choice question asking which product each image showed. Meyer says he ran the test on Jev 1.14 and Jev 1.13, for about 60 requests in total.

For the product roundup images, Meyer reports that Jev 1.14 gave each image a higher fit score for its own headline than for the randomly selected alternatives. The correct-headline scores ranged from 0.65 to 0.89, with a median of 0.76; the other headlines scored no higher than 0.18, with a median of 0.06. The plant-care headline scored at most 0.10. Jev 1.14 identified the product in nine of the 10 images. Jev 1.13 identified all 10, though two answers had confidence below 0.5.

Where Visual Classification Could Help

The test points to a possible use for Jev: repeatable image checks that can be expressed as structured decisions, such as whether a featured image fits a headline or which product appears in a photo. Such checks could help publishers and other teams review large sets of images, though Meyer’s small, self-selected test does not establish how well the system would perform across other sites, image types or workflows.

The release also shows TypeSafe making experimental capabilities available without turning them on by default. Users must opt in, and the company says the features still have too much “jaggedness” for it to endorse production use. That warning matters for anyone considering automated decisions based on visual output: the reported scores are an individual test, not a guarantee of accuracy.

Jev’s Opt-In Visual Release

TypeSafe published Jev 1.14 on October 14. The company says image input also works with Jev 1.13, describing that earlier model as multimodal despite the capability not being publicly presented that way before. The new box response type, which returns a target’s bounding box and a confidence score, is limited to Jev 1.14.

Image support is enabled with the TypeSafe-Danger-Zone-Images header. A separate header enables longer context, increasing the stated request limit from 64,000 to 262,000 tokens and the per-question limit from 32,000 to 262,000. TypeSafe retains the lower defaults and advises against longer inputs because reliability can decline. The company also allows users to opt in to sharing queries to help identify weaknesses, while saying it will not use those queries to train the model.

The image feature has file constraints: PNG and JPEG still images between 256 and 1,568 pixels per side, with a 16 MB limit per file and 32 MB per request. TypeSafe says images are billed as input tokens at Jev’s usual rate, rather than with an added image surcharge. The current model is available as jev-preview; TypeSafe said the jev-latest alias was due to move to it about a week after release.

“There is still way too much jaggedness for us to vouch for these for production use cases.”

— TypeSafe, in its release notes as quoted by Thorsten Meyer

Limits of the Image Test

The published test covers only 30 images from one website, and the supplied source material ends just as Meyer begins discussing results for the 15 news images. It does not provide those results or findings for the five essays, nor does it report outcomes from the bounding-box test. Those parts of the evaluation cannot be assessed from the available account.

The test also does not establish accuracy across other datasets or production conditions. TypeSafe has not endorsed the new visual features for production use, and the reported scores come from Meyer’s own evaluation rather than a public benchmark. The company has also cautioned that its internal comparisons with open-weight models may favor Jev and advised users to conduct their own evaluations.

Model Rollout and Further Testing

TypeSafe made Jev 1.14 available under jev-preview after the October 14 release and said jev-latest would move to the model around October 21. The company has described deprecation notice periods of 60 days for Jev 1.13 and 30 days for Jev 1.14, with a formal policy still to come.

For now, the image and bounding-box tools remain opt-in, and users weighing them for a specific task will need to test them against their own images and error tolerances. More complete results from Meyer’s evaluation, particularly for news images and object localization, are not included in the supplied report.

Key Questions

What changed in Jev 1.14?

TypeSafe added image input and a new bounding-box output type. Image input also works with Jev 1.13, according to the company; the box type is for Jev 1.14.

What did Meyer test?

He tested 30 featured images from his website against headline-fit questions and, for 10 product images, a question asking which product appeared in each image. He says he ran the tests on Jev 1.13 and 1.14.

How did Jev perform on product images?

In Meyer’s account, Jev 1.14 scored each product image higher against its own headline than against the comparison headlines. It selected the correct product in nine of 10 cases; Jev 1.13 got all 10 but assigned two answers confidence below 0.5.

Is Jev’s image support ready for production?

TypeSafe says it is not ready to vouch for the features in production use cases. Image support and longer context require explicit opt-in headers, and the company warns that reliability can vary.

What remains unknown about the test?

The available source does not include the results for the 15 news images or five essays, or findings from the bounding-box test. It also does not show how well the model would perform on images outside this one-site sample.

Source: ThorstenMeyerAI.com