From Pixels To AI: Jev Looks At 30 Of My Images

🔍 Read the full analysis: From Pixels To AI: Jev Looks At 30 Of My Images on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on office and shipping supplies

• Business-only prices and quantity discounts
• Tax-exempt purchasing
• Multiple users, one account, clear invoices

Create a free Business account →Free to register

As an affiliate, we earn on qualifying purchases.Hide

TypeSafe released Jev 1.14 on October 14, adding optional image input and a new bounding-box output to its decision model. In a test published the next day, Thorsten Meyer of ThorstenMeyerAI.com ran the model against 30 featured images from his site, reporting strong results on product images while noting that the model’s image capabilities are not yet endorsed for production use.

Jev is designed to return structured decisions rather than free-form prose: users provide a state and typed questions, and the model responds with outputs such as yes-or-no probabilities, choices from a list, or scores with confidence. Before the release, those questions concerned text. TypeSafe says image support also works with jev-1.13.0, describing that version as multimodal despite not having previously presented it that way.

Image handling is enabled with the request header TypeSafe-Danger-Zone-Images: true. Images can be referenced in the request state, question instructions, or criteria for choices and scores. Jev 1.14 adds a box answer type, which returns one bounding box and a confidence score for a specified target. That feature is exclusive to 1.14, according to the source report.

Meyer tested the two model versions on images accompanying 10 product roundups, 15 news posts and five essays. He asked whether each image fit its own headline, a randomly selected different headline, and an unrelated houseplant headline. The 10 product images also received a multiple-choice product-identification question. In that subset, Jev 1.14 identified the product correctly in nine of 10 cases; its correct answers had confidence of at least 0.77, while the one error had confidence 0.60.

Image Checks Move Into Structured AI

The release extends Jev from text-based classification to tasks where the input itself is visual. That could let developers build structured checks for image relevance or locate a named object without asking a model to produce a general description. The new bounding-box output gives a specific, machine-readable result rather than a prose explanation.

Meyer’s test suggests the feature may be useful for narrow, repeatable checks, particularly product-image matching. But it is a small test on one site’s images, not an independent benchmark or evidence that the model performs reliably across different image types. The report’s product findings also show why confidence matters: the one mistaken product label had lower confidence than the correct labels, but that pattern alone does not establish a dependable error-detection rule.

For developers, TypeSafe’s decision to place image and other experimental features behind explicit headers is relevant to deployment. It keeps existing requests on their prior behavior unless users opt in, while signaling that these capabilities are still being tested rather than presented as production-ready.

Opt-In Release and Model Limits

TypeSafe described the release as balancing builders seeking reliability with users who want to experiment. It said production behavior remains unchanged unless a request includes the relevant opt-in header. The company warned that image support has too much “jaggedness” for it to vouch for production use cases, and the report says TypeSafe asks users to evaluate the model for themselves.

The release also adds an opt-in long-context setting. The default limits are 64,000 tokens per request and 32,000 per question; a separate header raises both limits to 262,000. TypeSafe advises against the larger context because reliability can fall on long inputs. Users may also opt in to sharing queries to help TypeSafe identify weaknesses; according to the report, the company says it will not train on those queries.

Images must be PNG or JPEG stills, with each side between 256 and 1,568 pixels. A file may be up to 16 MiB, and a request up to 32 MiB; one invalid image can fail the whole request. The report says image input has no surcharge and is billed as ordinary input tokens at Jev’s stated rate of $0.042 per million input tokens. Jev 1.14 is available as jev-preview; TypeSafe planned to move jev-latest to it about a week after release.

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

— Thorsten Meyer

News-Image Results Are Missing

The supplied source ends just as Meyer begins the section on news images, after the words “News images: Jev was right, my images weren’t a dif”. It does not include the rest of that finding, so the results for the 15 news images cannot be reported here. The source also does not provide the full test outputs, confidence scores for every image, or enough methodological detail to independently assess the evaluation.

TypeSafe’s warning leaves open how reliably image input and bounding boxes work across production tasks. The company reportedly shared internal intelligence estimates rather than public benchmark scores, and Meyer cautioned that those estimates may favor Jev. The source does not establish whether the model’s performance generalizes beyond the images and questions in his test.

Preview Rollout and Testing

TypeSafe made Jev 1.14 available under the jev-preview name on release. According to Meyer’s report, the company planned to move jev-latest to 1.14 around October 21, one week after release. Users testing image input or bounding boxes must opt in with the designated header and account for the documented file and request limits.

TypeSafe has said it will give 60 days’ deprecation notice for Jev 1.13.0 and 30 days for 1.14.0, with a formal policy still to come. For now, the practical next step for developers is to run their own evaluations and decide whether the experimental behavior fits their needs; the source provides no later benchmark or production-readiness announcement.

Key Questions

What changed in Jev 1.14?

TypeSafe added opt-in image input and a bounding-box answer type. The box type returns one box and a confidence score for a specified target.

Can Jev 1.13 process images?

TypeSafe says image support also works on jev-1.13.0, describing that version as multimodal. The new bounding-box type is limited to Jev 1.14, according to the report.

How did Jev perform on the 30-image test?

For 10 product images, Jev 1.14 matched images to their own headlines with scores from 0.65 to 0.89 and identified the pictured product correctly in nine of 10 cases. The supplied source cuts off before reporting the findings for the 15 news images.

Is TypeSafe recommending the image features for production?

No. TypeSafe said the capabilities still had too much “jaggedness” for the company to vouch for production use. The features are behind an explicit opt-in header.

Source: ThorstenMeyerAI.com