Thirty Images Later: My Experience Testing Jev’s Vision

🔍 Read the full analysis: Thirty Images Later: My Experience Testing Jev’s 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 bounding-box answer type to Jev in version 1.14, released October 14, 2026. In a test reported the following day, Thorsten Meyer ran the model against 30 featured images from his own site and found strong product-image matching, while warning that the results came from a small, single-site evaluation rather than a production assessment.

Meyer tested the 30 most recent featured images on ThorstenMeyerAI.com: 10 product roundups, 15 news posts and five essays. For each image, he asked whether it fit its own headline, a randomly selected different post’s headline and an unrelated headline about repotting a monstera plant. He also asked Jev to identify which of 10 products appeared in each product image. He ran the questions on both jev-1.14.0 and jev-1.13.0, for about 60 requests in total.

For the 10 product images, Jev 1.14 gave each image’s own headline a score of 0.65 to 0.89, with a median of 0.76. Random other headlines had a median score of 0.06 and never exceeded 0.18; the unrelated plant headline never scored above 0.10. Jev 1.14 identified the product correctly in nine of 10 images. In the one error, it described a mesh Wi-Fi photo showing a router and cylindrical mesh unit as a Wi-Fi 7 router, with 0.60 confidence. Meyer says the correct product identifications had confidence of at least 0.77. Jev 1.13 identified all 10 products, though two answers had confidence below 0.5.

TypeSafe says the image feature works with both jev-1.13.0 and jev-1.14.0, while the new bounding-box type is limited to 1.14. The release also includes an opt-in longer-context setting. TypeSafe describes the version as a set of broad, small improvements, with multilingual support expected to benefit most. The source account begins to report the news-image results but is cut off before giving their details, so no conclusion about that group can be drawn from the supplied material.

What the Image Scores Can Show

The test examines whether a structured decision model can assess whether an image matches a headline, rather than generate a caption or free-form description. Meyer describes image-headline checking as a repetitive editorial task that may be skipped. In this test, the model assigned higher scores to product images paired with their own headlines than to images paired with unrelated headlines.

The results do not establish how well Jev works across other publishers, image styles or more ambiguous editorial choices. The sample includes only 30 images from one site, and the report provides no independent comparison or wider validation. The product-image scores are an initial observation and do not establish that the system can make editorial decisions reliably at scale.

Jev’s Opt-In Image Release

Jev is a decision model that answers typed questions using formats such as yes-or-no probabilities, choices from a list and scores on a scale, each with a confidence value. Before the October 14 update, TypeSafe’s questions had to concern text. The release lets users put images in a request’s state, question instructions or criteria used to define options and scoring levels.

Images are referenced through an assets map in the request. TypeSafe’s published limits allow PNG and JPEG still images from 256 to 1,568 pixels per side, with a maximum of 16 MiB per image and 32 MiB per request. WebP and HEIC files must be converted. TypeSafe says image input is billed as ordinary input tokens at Jev’s usual rate, rather than carrying a separate image surcharge.

TypeSafe has put the new features behind explicit headers labeled as a “danger zone”. The company said the capabilities still have too much “jaggedness” for it to vouch for production use. The default context limit remains 64,000 tokens per request and 32,000 per question; an opt-in header raises both limits to 262,000 tokens, though TypeSafe advises against the larger setting because reliability can decline on long inputs.

“The results were more honest about my images than I expected.”

— Thorsten Meyer

Limits of the Thirty-Image Test

The supplied report does not include the complete results for the 15 news images and five essays; the source text ends as Meyer begins discussing news-image performance. It also does not provide image-by-image scores for those groups, independent reviewers, or an evaluation against a larger, varied dataset. The test’s figures describe Meyer’s own setup and should not be treated as a benchmark of general model performance.

TypeSafe has not endorsed the new features for production use. The test also does not establish how confidence scores should be calibrated for other tasks, or whether the model’s performance changes with different prompts, image quality or subject matter. Those questions remain open on the information provided.

Release Rollout and Further Testing

TypeSafe made Jev 1.14 available as jev-preview on October 14. The company said the jev-latest alias would move to the new version one week after release, around October 21, 2026. It also described deprecation notice periods of 60 days for version 1.13.0 and 30 days for 1.14.0, with a formal policy still to come.

Further testing across more sites and image types would be needed to assess whether the results apply beyond Meyer’s sample. TypeSafe has invited users to opt in to sharing queries to help identify model weaknesses, while saying it will not train on those queries. The reliability of the image and bounding-box features for production use remains undetermined.

Key Questions

What changed in Jev 1.14?

TypeSafe added opt-in image input, a new bounding-box answer type in Jev 1.14, and an opt-in longer-context setting. Image input also works on Jev 1.13.0, according to TypeSafe.

How did Jev 1.14 perform on the product images?

In Meyer’s test, Jev 1.14 matched 10 product images to their own headlines with scores from 0.65 to 0.89 and identified nine of 10 products correctly. These are results from one site’s small test, not a general performance benchmark.

Was Jev 1.14 more accurate than Jev 1.13?

The versions had different results in the product identification task: Jev 1.14 got nine of 10 correct, while Jev 1.13 got all 10 correct. Meyer reported that two Jev 1.13 answers had confidence below 0.5, while Jev 1.14’s correct answers were at least 0.77 confident. The sample is too small to establish which version is generally more accurate.

Is TypeSafe recommending Jev’s image features for production?

No. TypeSafe said the capabilities remain too inconsistent for it to vouch for production use and placed them behind explicit opt-in headers.

What happened in the test of news images?

The supplied source text starts to discuss the 15 news images but cuts off before reporting results. It does not provide enough information to summarize how Jev performed on that group.

Source: ThorstenMeyerAI.com