A Personal Image Test For Jev’s AI Vision

🔍 Read the full analysis: A Personal Image Test For Jev’s AI Vision on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get the latest gadgets 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

TypeSafe added opt-in image input and a new bounding-box output to Jev, its structured decision model, in a release published October 14, 2026. The next day, Thorsten Meyer tested Jev 1.14 and its predecessor on 30 featured images from ThorstenMeyerAI.com, reporting strong results on product photos but cautioning that his small, site-specific test does not establish production reliability.

Jev is designed to answer typed questions about supplied information, such as yes-or-no questions with probabilities, choices from a list, and scores on a scale. Before this release, those questions were text-based. TypeSafe says image input can now be used in the request’s state, a question’s instructions, or criteria for possible answers. The feature requires the explicit header TypeSafe-Danger-Zone-Images: true, and TypeSafe says it works with both jev-1.13.0 and jev-1.14.0.

Jev 1.14 also introduces a box response type, which returns one bounding box and a confidence score for a described target. TypeSafe restricts the new capabilities behind opt-in headers, saying image support and other experimental functions have too much “jaggedness” for it to recommend them in production. Image requests accept PNG and JPEG stills within specified dimensions and file-size limits; the source says WebP and HEIC images must first be converted.

Meyer tested the 30 most recent featured images on his site: 10 product roundups, 15 news posts and five essays. For each, he asked whether it fit its own headline, a randomly chosen other headline and an unrelated headline about repotting a monstera plant. Product images also received a multiple-choice identification question. Meyer says he ran the test on both model versions, for about 60 requests overall.

What the Product-Photo Scores Show

For the 10 product images, Meyer reports that Jev 1.14 separated matching headlines from mismatches in his test. The model gave each image’s own headline a score from 0.65 to 0.89, with a median of 0.76. A randomly selected different headline scored no higher than 0.18, with a median of 0.06; the unrelated plant headline scored no higher than 0.10.

On product identification, Jev 1.14 selected the right item in nine of 10 cases. Meyer described the miss as understandable: a mesh Wi-Fi photograph showed both an antenna-equipped router and a cylindrical mesh unit, and the model labeled it a Wi-Fi 7 router with 0.60 confidence. He said all correct answers had confidence scores of at least 0.77. Jev 1.13 identified all 10 correctly, but two of those answers had confidence below 0.5. These observations suggest confidence may help flag uncertain outputs in this particular test; they do not show how well that approach would work across other images or tasks.

The practical relevance is that image questions can be put into the same structured request format as text questions, potentially making visual checks easier to automate. But the evidence here comes from one publisher’s small collection, and the results are not an independent benchmark. TypeSafe itself advises against relying on the new functions in production.

How TypeSafe Is Gating Image Input

In its October 14 release, TypeSafe framed the opt-in controls as a way to let developers experiment while leaving existing production behavior unchanged by default. Image input is enabled with a dedicated danger-zone header. The bounding-box feature is available only on Jev 1.14, while image input also works on version 1.13. TypeSafe’s post described the current model as having been multimodal already, though image support had not previously been exposed to users in this way.

The release also adds an optional longer-context setting. The stated defaults are 64,000 tokens per request and 32,000 per question; a separate header raises both limits to 262,000 tokens. TypeSafe cautions that reliability declines with longer inputs and advises against using the larger limits without a reason. The model itself is available as jev-preview, with jev-latest scheduled to move to it about a week after release. The source describes TypeSafe’s internal estimates as comparisons only with open-weight models and warns that they may favor Jev, recommending that users run their own evaluations.

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

— TypeSafe, in its Jev 1.14 release post, as quoted by Thorsten Meyer

Limits of the Image Test

The supplied account ends just as Meyer begins discussing the 15 news images. It says, “News images: Jev was right, my images weren’t,” but provides no supporting scores, examples or explanation. Those results cannot be reported in detail from the available material, and the test’s overall performance across news images remains unclear.

The test covers only 30 images from one website and does not establish performance on other publishers’ images, image types or real-world workloads. The source does not provide the full request data, an independent review, or a larger validation set. TypeSafe has also said it cannot yet vouch for production use. Its claims about model behavior and feature availability are company statements, while the image-test results are Meyer’s own reported observations.

Jev Rollout and Further Testing

TypeSafe said jev-preview was available with the release and planned to move jev-latest to version 1.14 about a week later, around October 21, 2026. The company’s stated deprecation notice is 60 days for version 1.13.0 and 30 days for version 1.14.0; it said a formal policy was still to come.

For developers, the immediate next step is to test image inputs against their own representative data and account for the documented format, size and request limits. Meyer’s report does not state that TypeSafe has changed its production warning or that the publisher plans a follow-up test. Broader results, particularly the missing account of news-image performance, remain necessary to judge how consistently Jev handles visual matching beyond this small trial.

Key Questions

What changed in Jev 1.14?

TypeSafe added opt-in image input and a new box response type that returns one bounding box with a confidence score. The image feature also works with Jev 1.13.0, according to TypeSafe.

How did Jev perform on product images?

In Thorsten Meyer’s test of 10 product photos, Jev 1.14 matched each photo to its own headline with reported scores from 0.65 to 0.89 and identified the product correctly in nine cases. This is a small, publisher-run test, not a general benchmark.

Does TypeSafe recommend using image support in production?

No. TypeSafe said the new features remain too inconsistent for it to vouch for production use and placed them behind an explicit opt-in header.

What happened in the news-image portion of the test?

The available source begins a section saying Jev was right while the publisher’s news images were not, but it cuts off before giving results or examples. The details and scores are therefore unknown from the material provided.

Which image formats does Jev accept?

The source lists PNG and JPEG still images, subject to stated dimension and file-size limits. It says WebP and HEIC files need to be converted first.

Source: ThorstenMeyerAI.com