Seeing What Jev Sees: 30 Images From My Own Collection

🔍 Read the full analysis: Seeing What Jev Sees: 30 Images From My Own Collection on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get privacy and security gear 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 with support for image-based questions and a new answer type that returns a bounding box, opening the decision model to visual tasks. The company has placed the capabilities behind opt-in headers and says the technology is still too unreliable for production use; a test published the following day examined 30 featured images from ThorstenMeyerAI.com.

Jev is designed to answer structured questions rather than generate prose: users provide a state and typed questions, and the model returns outputs such as a yes-or-no probability, a choice from a list, or a score, along with confidence. Before this release, the questions described in the account were text-based. TypeSafe says image input also works with jev-1.13.0, which it described as multimodal, as well as 1.14.0.

Image support requires the request header TypeSafe-Danger-Zone-Images: true. Images can be placed in the state, question instructions, or the criteria for options and scores. The new box type, available only in Jev 1.14, returns one bounding box for a specified target and a confidence score. TypeSafe also introduced a separate opt-in for longer context, lifting request and question limits from 64,000 and 32,000 tokens to 262,000 tokens each, while warning that reliability declines on long inputs.

Meyer tested three yes-or-no questions on each of 30 recent site images: whether the image fit its own headline, a randomly selected different headline, and an unrelated plant-care headline. The 10 product images also received a multiple-choice identification question. Across roughly 60 requests on the two model versions, Jev 1.14 gave product images their own headlines scores from 0.65 to 0.89, with a median of 0.76; mismatched headlines scored no higher than 0.18, and the plant headline no higher than 0.10. It named the correct product in 9 of 10 cases.

Image Checks Move Beyond Text

The release extends Jev’s structured decision format to visual inputs without turning it into a general-purpose image description tool. Tasks such as checking whether a product photo matches a headline, selecting among image-based options, or locating a named object can be framed as defined questions with confidence scores. That could make visual checks easier to automate in workflows built around repeatable decisions.

The test offers an early example, not evidence that the model is dependable across image tasks. Meyer’s results were strongest for product images, and one of the 10 product identifications was wrong. He reports that all correct identifications had confidence of at least 0.77, while the mistaken answer had confidence of 0.60; that pattern may help guide evaluation, but it does not establish a reliable threshold for other datasets. The key practical point is TypeSafe’s own warning: image support is experimental and should not be treated as production-ready based on a small test.

A Guarded Jev 1.14 Rollout

TypeSafe presented the release as a balance between users seeking reliability and those wanting to experiment. Existing production behavior remains unchanged unless a caller sends the relevant opt-in header. The image and long-context features carry a “danger zone” label, and TypeSafe said there was “still way too much jaggedness” for it to endorse them in production use cases.

The release also includes broad, small improvements to Jev 1.14, with TypeSafe expecting multilingual support to benefit most. The model is available as jev-preview; the company said jev-latest would move to it about a week after release, around October 21. TypeSafe shared internal intelligence estimates rather than public benchmark scores and cautioned that comparisons against open-weight models may favor Jev, making independent evaluation relevant.

Image inputs have practical constraints: only PNG and JPEG still images are supported, each side must measure 256 to 1,568 pixels, and files are limited to 16 MiB each and 32 MiB per request. The request fails if any image is invalid. TypeSafe says there is no image surcharge; input is billed at the standard rate of $0.042 per million input tokens. Base64 encoding adds about one-third to file size, so the company recommends multipart uploads and says its Python and JavaScript SDKs handle them.

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

— TypeSafe, describing image and long-context reliability

How Reliable Is Image Judging?

The source account is incomplete: it ends as Meyer begins discussing the 15 news-post images, so their scores and findings are not available here. It also does not provide the results for the five essays, a full comparison of Jev 1.13 and 1.14 across every question, or enough methodology to independently reproduce the test. The reported product results come from a small, site-specific sample and cannot establish accuracy on other image collections or tasks.

TypeSafe has not supplied public benchmark scores for these capabilities and says the features remain rough. It is not clear how performance changes with different image quality, more complex scenes, or requests containing multiple images. The release account also does not give a general reliability threshold for confidence scores or evidence that a threshold would hold beyond Meyer’s sample.

Independent Testing and Model Rollout

Users can test image requests now through jev-preview, provided they explicitly enable the image header and stay within the published file limits. TypeSafe said jev-latest was due to move to Jev 1.14 around October 21, 2026. The company also set out deprecation notice periods of 60 days for Jev 1.13.0 and 30 days for 1.14.0, while saying a formal policy was still to come.

The next useful evidence will be broader evaluations that report results by task and image type, alongside TypeSafe’s later guidance on production reliability. Until then, the available account supports a limited conclusion: Jev 1.14 handled this site’s product-image checks well, but the release and test do not show that visual judgments are dependable across uses.

Key Questions

What changed in Jev 1.14?

Jev 1.14 added an opt-in image input feature and a new box answer type that returns one bounding box and a confidence score. It also offers an opt-in longer context limit.

Can Jev 1.13 process images?

According to TypeSafe, image support also works on jev-1.13.0, which the company said was multimodal. Image requests require the opt-in header described for the feature.

How did Jev perform on the 30 images?

In the reported test, Jev 1.14 gave 10 product images higher fit scores for their own headlines than for unrelated headlines and identified 9 of 10 products correctly. The supplied account does not include the results for the 15 news images or five essays.

Is TypeSafe recommending image support for production?

No. TypeSafe said reliability remains too uneven to vouch for image and long-context features in production. Both are opt-in, and the company advises caution.

Source: ThorstenMeyerAI.com