Jev Can See Now: 30 Images From My Personal Archive

🔍 Read the full analysis: Jev Can See Now: 30 Images From My Personal Archive on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

• 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 image input and a bounding-box response type to its decision model, which previously handled text-based questions. The new capabilities are behind explicit opt-in headers, and TypeSafe says the features remain too unreliable for it to recommend in production. Thorsten Meyer tested the image feature on 30 images from his site’s archive, reporting strong results for product photos while warning that the available account does not include the full findings for news images.

Jev is designed to return structured decisions rather than generate prose. Users provide a state and typed questions; responses can be yes-or-no probabilities, choices from a list, or scores, each with confidence information. According to Meyer, before this release those questions had to concern text. TypeSafe’s release post says image support also works with jev-1.13.0, describing that model as multimodal even before the capability was documented publicly.

Images are supplied through an assets map and referenced in the request wherever they are needed. TypeSafe says an image can appear in the state, a question’s instructions, or criteria used to define options and scores. The new box type is limited to Jev 1.14 and returns one bounding box for a specified target, along with a confidence value. Image support accepts PNG and JPEG files within specified size and dimension limits; unsupported formats such as WebP and HEIC must be converted first.

Meyer tested the 30 most recent featured images on ThorstenMeyerAI.com: 10 product roundups, 15 news posts and five essays. He asked whether each image matched its own headline, a randomly selected different headline, and an unrelated plant-care headline. For product images, he reports that Jev 1.14 gave each image’s own headline a higher score than the alternatives. The scores for matching headlines ranged from 0.65 to 0.89, with a median of 0.76; random headlines reached no more than 0.18, and the unrelated headline no more than 0.10. On a separate product-identification question, Jev 1.14 selected the right product in nine of 10 cases.

Testing Jev Against Real Site Images

The release expands Jev from text-only decision tasks to requests that can include visual information. That could let developers build structured checks around image contents—for example, whether a picture fits a headline or which listed item it depicts—without asking a general-purpose model to write an explanation. The bounding-box output also offers a defined way to identify a target’s location in an image, though TypeSafe has not endorsed it for production use.

Meyer’s test offers an early, small-scale example rather than independent validation. The product-photo results suggest Jev could distinguish relevant images from unrelated headline prompts in this particular archive. The product identification results also show why confidence matters: Jev 1.14 missed one item, while Meyer reports that every correct answer in that test had confidence of at least 0.77. That pattern may inform further evaluation, but it does not establish a generally reliable threshold.

The practical impact is therefore experimental. Developers can try visual decision workflows now, but should not treat this one archive test as proof of performance across other image types, languages, or operating conditions. TypeSafe’s opt-in design makes the boundary explicit: the feature is available to test, while the company says its reliability still needs work.

Opt-In Features and Release Limits

TypeSafe describes the new options as a way to serve both builders seeking reliability and users who want to experiment. Existing production behavior remains unchanged unless a request includes the relevant header. Image requests require TypeSafe-Danger-Zone-Images: true; longer context has a separate opt-in header. TypeSafe sets the default context limits at 64,000 tokens per request and 32,000 per question, with the header raising both to 262,000. The company advises against the larger limit because it says reliability declines on long inputs.

TypeSafe also says users may opt in to letting the company use their queries to identify model weaknesses, while stating that it will not train on those queries. Image input has no separate surcharge, according to the source: it is billed as ordinary input tokens at Jev’s stated rate of $0.042 per million input tokens, while output tokens are free. The source recommends multipart uploads because base64 encoding increases file size by about a third, counting toward the request limit.

Jev 1.14 is available as jev-preview. Meyer says TypeSafe plans to move the jev-latest alias to the new model about one week after release, around October 21. TypeSafe also provided internal intelligence estimates rather than public benchmark scores, and, as Meyer reports it, cautioned that comparisons with open-weight models may favor Jev. The source recommends running independent evaluations.

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

— Thorsten Meyer

What the Image Test Cannot Establish

The source supplied for this report ends as Meyer begins discussing the 15 news images, after the subheading “News images: Jev was right, my images weren’t.” It does not include the results for that group or the five essays, nor the full comparison between Jev 1.14 and 1.13. Those findings cannot be inferred from the product-photo results. The sample is also limited to one site’s archive and does not establish how the feature performs on broader image collections.

TypeSafe has not published public benchmark scores for these capabilities in the material provided, and it says it cannot yet vouch for production use. Details of the formal deprecation policy are also pending: Meyer reports that TypeSafe gave 60 days’ notice for Jev 1.13.0 and 30 days for Jev 1.14.0, with a formal policy still to come. It is not clear from the source how the model will perform across different tasks, image formats after conversion, or long requests.

Jev Preview and Alias Schedule

Developers can test Jev 1.14 through the jev-preview model name. TypeSafe’s reported schedule moves jev-latest to version 1.14 around October 21, one week after the release. Users considering the image or extended-context features will need to opt in explicitly, and TypeSafe’s warning means results should be checked against task-specific evaluations before any production use.

Further clarity will depend on TypeSafe’s forthcoming deprecation policy and any additional reliability information it publishes. The complete results of Meyer’s archive test are not present in the supplied source, so the reported product-image findings should not be treated as a full account of the 30-image evaluation.

Key Questions

What did TypeSafe add in Jev 1.14?

Jev 1.14 adds opt-in image input and a new bounding-box response type. Image support is also reported to work with Jev 1.13.0, while the box type is specific to Jev 1.14.

Is TypeSafe recommending the image features for production?

No. TypeSafe says there is still too much unreliability for it to vouch for the capabilities in production. The features require explicit opt-in headers.

How did Jev perform on the 30-image test?

For the 10 product images, Meyer reports that Jev 1.14 scored each image’s own headline between 0.65 and 0.89, with a median of 0.76, and identified the featured product correctly in nine cases. The supplied source does not include the full results for the news images or essays.

When does Jev 1.14 become the latest model?

It is available now as jev-preview. TypeSafe’s reported plan is to move the jev-latest alias to Jev 1.14 about one week after its October 14 release, around October 21, 2026.

Source: ThorstenMeyerAI.com