I Put 30 Of My Images In Front Of Jev

🔍 Read the full analysis: I Put 30 Of My Images In Front Of Jev on ThorstenMeyerAI.com

Age 18–24?Offer from Amazon

Prime made for students and young adults

• Fast, free delivery for dorm and study essentials
• Prime Video and Amazon Music included
• Member-only deals

Try Prime for Young Adults →Free trial for eligible 18–24 year olds

As an affiliate, we earn on qualifying purchases.Hide

TypeSafe added opt-in image input to its Jev decision model on October 14, and Thorsten Meyer tested the feature on 30 images from his website. In the product-image portion of the test, Jev 1.14 matched images to their own headlines more strongly than to unrelated headlines, but the published source excerpt ends before giving results for news images and essays.

Jev is designed to answer structured questions rather than generate prose: users provide a state and typed questions, and the model returns choices, scores or yes-or-no answers with confidence values. Until this release, the questions were limited to text, according to Meyer. TypeSafe says image input also works with jev-1.13.0, describing that model as multimodal even before the new release.

Images are enabled with the request header TypeSafe-Danger-Zone-Images: true. Developers can place an image reference in the request state, question instructions or option criteria. Jev 1.14 also adds a box response type that returns a bounding box for one specified target, with a confidence score. That capability is limited to Jev 1.14.

Meyer tested the 30 most recent featured images on ThorstenMeyerAI.com: 10 product roundups, 15 news stories and five essays. He asked whether each image fit its own headline, a randomly selected different headline, and an unrelated headline about repotting a monstera. The 10 product images also received a multiple-choice question asking which product each showed. Meyer says the test involved about 60 requests across Jev 1.13 and 1.14.

Testing Jev Beyond Text

The release gives developers a way to apply Jev’s structured outputs to visual checks, such as deciding whether an image matches a headline or locating a named object. That could make the model useful for repeatable image review tasks that are difficult to handle through text-only questions. But the feature is explicitly marked experimental: TypeSafe says it is not ready to vouch for these capabilities in production.

Meyer’s results offer an early, limited example rather than an independent benchmark. In the product subset, Jev 1.14’s confidence scores separated the matching headline from the alternatives in his test. The finding may help builders decide what to evaluate, but it does not establish performance across other image types, datasets or real-world workflows. The incomplete results in the supplied report also prevent a full assessment of the 30-image test.

An Opt-In Release With Limits

TypeSafe framed the release as serving both builders seeking reliability and people experimenting with new capabilities. The company kept existing production behavior unchanged unless a request sends an explicit danger-zone header. It says the image and bounding-box features still have too much “jaggedness” for production use cases.

The release also adds a separate opt-in for longer context. The default request limit remains 64,000 tokens, with a 32,000-token limit per question; the long-context header raises each limit to 262,000 tokens. TypeSafe advises against the larger setting because reliability can fall on long inputs. Jev 1.14 is available as jev-preview, with jev-latest scheduled to move to it about a week after release, around October 21.

For images, TypeSafe lists support for PNG and JPEG stills from 256 to 1,568 pixels per side. Each file can be up to 16 MiB, and a request can contain up to 32 MiB. The company says images are billed as ordinary input tokens at Jev’s stated rate of $0.042 per million input tokens, with output tokens free. Meyer also reports that TypeSafe shared internal intelligence estimates rather than public benchmark scores and cautioned that the estimates may favor Jev.

““jev-1.13.0 was multimodal all along.””

— TypeSafe’s release post, as reported by Meyer

Results Missing for Other Images

The supplied report excerpt ends as Meyer begins discussing the 15 news images, so it does not provide their scores or the results for the five essays. It is also unclear whether the test used repeated runs, how the images and headlines were selected beyond the stated method, or whether the outcomes were independently checked. The product-image results are from one site and a small sample, so they cannot establish general accuracy.

TypeSafe has not vouched for production reliability, and the release materials described by Meyer do not provide public benchmark scores for image performance. The company’s internal estimates were limited to comparisons with open-weight models and, by its own account, may be biased toward Jev. Broader performance, error rates and suitability for high-stakes applications remain unestablished.

Watch for Wider Evaluation

The immediate next product milestone is Jev 1.14’s planned move from jev-preview to jev-latest, which Meyer said was expected around October 21, 2026. TypeSafe also indicated it would provide 60 days’ deprecation notice for Jev 1.13 and 30 days for Jev 1.14, while a formal deprecation policy was still to come.

For readers judging the image feature, the next useful evidence would be complete results from Meyer’s news-image and essay tests, followed by evaluations on more varied images and tasks. TypeSafe’s own guidance is to run evaluations suited to each use case and treat the new capabilities as experimental unless and until reliability is demonstrated.

Key Questions

What changed in Jev 1.14?

TypeSafe added opt-in image input and a new bounding-box response type. Image input is also available in Jev 1.13, according to the company; the box type is limited to Jev 1.14.

How did Jev perform on the product images?

Meyer reports that Jev 1.14 gave each product image a higher score for its own headline than for a randomly chosen other headline or the unrelated plant headline. It identified the pictured product correctly in 9 of 10 cases; Jev 1.13 identified all 10 correctly, though two answers had confidence below 0.5.

Does the test show that Jev is ready for production?

No. It was a small test using one website’s images, and TypeSafe says the new capabilities are not yet reliable enough for it to vouch for production use. The reported results do not establish performance across other settings.

What happened in the news-image and essay tests?

The supplied source excerpt does not include those results. Meyer had 15 news images and five essays in the test, but the report cuts off as he begins discussing the news-image findings.

Source: ThorstenMeyerAI.com