I Turned 30 Of My Images Into A Test For Jev

🔍 Read the full analysis: I Turned 30 Of My Images Into A Test For Jev on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get smart everyday buys 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 released Jev 1.14 on October 14, adding opt-in image input and a new bounding-box response type to its decision model; the next day, Thorsten Meyer tested image matching on 30 featured images from his site. Meyer reported strong results for product images, but TypeSafe warns that the new capabilities remain too rough for production use, and the available account does not include the test’s full results for news and essay images.

Jev is designed to answer typed questions about supplied information, such as choosing among options, assigning a score or estimating a yes-or-no answer with confidence. Before the release, those questions were text-based. TypeSafe says images can now be placed in a request’s state, a question’s instructions or the criteria used to define choices and scores. Image input is available on jev-1.13.0 and jev-1.14.0, but callers must explicitly enable it using the TypeSafe-Danger-Zone-Images header.

Jev 1.14 also adds a box type, which returns one bounding box for a specified target, with a confidence score. That feature is limited to 1.14 and also requires an opt-in header. TypeSafe separately offers a long-context header that raises request limits from 64,000 to 262,000 tokens and question limits from 32,000 to 262,000, while advising against the larger context because reliability can decline. The source says the new model is available as jev-preview; jev-latest was due to move to it about a week after release.

For his test, Meyer used the 30 most recent featured images on ThorstenMeyerAI.com: 10 product roundups, 15 news posts and five essays. Each image was tested against its own headline, a randomly selected headline from another post and an unrelated houseplant headline. The product images also faced a question asking Jev to identify which of 10 products each image showed. Meyer says he ran the tests on both Jev 1.13 and 1.14, for about 60 requests overall.

What Image Decisions Can Show

The test offers an early look at whether Jev’s image support can handle a narrow, repeatable editorial task: checking whether an image matches a headline. For publishers and developers, that could help identify mismatches or route uncertain cases for human review. Meyer’s results suggest product photos may be a promising use case, but a small test on one site cannot establish how the system performs across different image styles, topics or workflows.

The release also matters for teams considering production use. Image analysis and bounding boxes are now available, but TypeSafe has deliberately put them behind an explicit opt-in and says it cannot yet vouch for them in production. That warning puts the reported test in the category of an initial evaluation, not evidence that the capability is ready for automated publishing decisions.

How TypeSafe Gated Jev Images

TypeSafe describes the release as balancing builders who need reliability with users who want to experiment. Existing behavior stays in place unless a request includes the relevant danger-zone header. The company said its previous model, jev-1.13.0, had image capability already, even though image input was not previously exposed in this way.

Requests use an assets map: an image is assigned a reference string, and that reference can be inserted where the image should be used. The source advises using a fresh UUID because the substitution is plain text. An image placed in the shared state can be reused across questions; placing it in individual instructions may send it repeatedly and increase input use. Supported still images are PNG and JPEG, with sides from 256 to 1,568 pixels, a 16 MiB per-image limit and a 32 MiB request limit. The source says WebP and HEIC files need conversion and that one invalid image can fail the whole request.

TypeSafe says images do not carry a separate surcharge and are billed as ordinary input tokens at the model’s listed rate. It recommends multipart uploads because base64 encoding increases file size by roughly one-third and that added size counts toward the request limit. These implementation details shape both the cost and reliability of any practical trial.

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

— TypeSafe, as quoted in Meyer’s account

The Missing Test Results

The supplied source ends as Meyer begins discussing the 15 news images, saying, “The news posts were a dif…” It does not provide the rest of that section or the results for his five essays. It is therefore not possible to report how Jev performed across those categories, or to summarize the test’s overall accuracy.

For the product images, Meyer reports that Jev 1.14’s score for an image’s own headline ranged from 0.65 to 0.89, with a median of 0.76. Random other headlines had a median score of 0.06 and did not exceed 0.18; the unrelated plant headline stayed at or below 0.10. Jev 1.14 identified nine of the 10 products correctly. The remaining image showed a router beside a cylindrical mesh unit; Jev selected “Wi-Fi 7 router” with 0.60 confidence. Meyer says Jev 1.13 named all 10 correctly, though two answers had confidence below 0.5. These are results from a small, self-run evaluation, not independently verified findings.

TypeSafe’s post also did not, according to the source, provide public benchmark scores for its internal estimates. The company compared Jev with open-weight models and reportedly cautioned that the comparison could favor Jev. The test setup, scoring behavior on other kinds of images and performance in real production traffic remain open questions.

Release and Evaluation Milestones

TypeSafe said jev-preview was available with the release and planned to move jev-latest to Jev 1.14 about one week later, around October 21, 2026. The company also described deprecation notice periods of 60 days for version 1.13.0 and 30 days for 1.14.0, while saying a formal policy was still to come.

For users, the immediate next step is to test the features on their own images and tasks, keeping the opt-in headers and TypeSafe’s production warning in view. Meyer advises running independent evaluations rather than relying on the company’s internal comparisons. The missing portions of his image-test report would be needed to assess the news and essay results and to judge whether the product-image pattern held across the full set.

Key Questions

What changed in Jev 1.14?

TypeSafe added a bounding-box response type to Jev 1.14 and made image input available through an opt-in header. Image input also works with Jev 1.13.0, according to the company.

What did Meyer’s image test find?

For 10 product images, Meyer reports that Jev 1.14 matched images to their own headlines more strongly than to unrelated headlines and identified nine of the 10 products correctly. The source provided here cuts off before the news-image and essay results.

Is Jev’s image capability ready for production?

TypeSafe said there was still too much “jaggedness” for it to vouch for the new features in production. Image input and bounding boxes require explicit opt-in, and the company’s warning should be treated separately from Meyer’s limited test results.

What image files does Jev accept?

The source lists PNG and JPEG still images, with image sides between 256 and 1,568 pixels. It gives a 16 MiB limit per image and 32 MiB per request; WebP and HEIC files must be converted first.

Source: ThorstenMeyerAI.com