Jev’s New Vision Meets 30 Images From My Camera Roll

🔍 Read the full analysis: Jev’s New Vision Meets 30 Images From My Camera Roll on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get the latest gadgets 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 image input and a bounding-box response type to its decision model, while keeping both behind explicit opt-in headers because the company says the features are not ready for production use. In a test published the next day, Thorsten Meyer of ThorstenMeyerAI.com ran Jev against 30 images from his site and reported strong results for product-image matching, though the supplied account does not include the full findings for its news-image tests.

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, with confidence values. Before this release, those questions were limited to text. TypeSafe says image input can now be used in the state, question instructions, and criteria for choices or scores. The image feature works with jev-1.13.0 and jev-1.14.0; the new box type is limited to Jev 1.14 and returns one bounding box for a specified target, along with confidence.

The changes are disabled by default. Image requests require the TypeSafe-Danger-Zone-Images: true header, while long-context requests require a separate opt-in header. The image interface uses references in an assets map, with the referenced image inserted wherever its identifier appears in the request. TypeSafe’s published constraints include PNG and JPEG still images from 256 to 1,568 pixels per side, a 16 MiB limit per file, and 32 MiB per request. One invalid image causes the full request to fail; WebP and HEIC files need conversion first.

Meyer tested the two model versions on 30 recent featured images: 10 product roundups, 15 news posts, and five essays. Each image was assessed against 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. Meyer reports that Jev 1.14 separated the product images from mismatched headlines: scores for each image’s own headline ranged from 0.65 to 0.89, with a median of 0.76, while randomly paired headlines had a median of 0.06. The off-topic headline scored no higher than 0.10. Jev 1.14 identified nine of the 10 products correctly; Jev 1.13 identified all 10, although two answers had confidence below 0.5.

Image Checks for Structured AI

The release extends Jev from text-only classification and scoring to requests that include visual evidence. That could support tasks such as checking whether a product photo matches a listing, finding a named object in an image, or comparing an image with a set of criteria. The bounding-box type adds a localized visual result rather than only a label or score, though TypeSafe has limited it to a single target and a single box per response.

Meyer’s test offers an early, narrow example rather than a general performance evaluation. The product-photo results suggest that Jev can distinguish strongly matching images from unrelated ones in this small sample. But the test was conducted by the site’s author on its own featured images, and the source does not provide independent replication or a complete account of the news-image results. Readers considering the feature for business workflows should treat the figures as an informal trial, not proof that the model is reliable across image types.

TypeSafe’s caution also matters for users evaluating costs and risk. The company says image input is billed as ordinary input tokens at its standard rate of $0.042 per million input tokens, with no separate image surcharge, but images placed inside individual questions may be sent repeatedly. The company has not endorsed the new functions for production use, so price alone does not establish that they are suitable for operational decisions.

Opt-In Release and Model Timeline

TypeSafe describes the release as a balance between users seeking reliability and those wanting to experiment. Its production behavior remains unchanged unless users send the relevant header. The company said the prior model, jev-1.13.0, was already multimodal, even though image support had not previously been available through the described interface. That is TypeSafe’s characterization; the supplied material does not describe separate independent verification of the claim.

Jev 1.14 is available under the jev-preview name. TypeSafe said the jev-latest alias would move to the new version one week after release, around October 21, 2026. The company described 1.14 as a set of broad, small improvements and said it expected multilingual performance to gain the most. It also shared internal intelligence estimates rather than public benchmark scores, and cautioned that those comparisons were limited to open-weight models and may favor Jev.

The image interface has practical constraints that shape how it can be tested. Users can place an image in shared request state so it is available to several questions, or include it in individual instructions, which can cause it to be sent again for each question. TypeSafe recommends multipart uploads because base64 encoding increases file size by about one-third, counting toward the request limit. Its Python and JavaScript SDKs handle that upload method.

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

— TypeSafe, in its Jev 1.14 release post, as quoted by Thorsten Meyer

Test Results and Reliability Gaps

The available source material ends as Meyer begins describing the news-image results, so it does not establish how Jev performed on the 15 news images or five essays. It also does not provide full results for Jev 1.13 on the headline-matching questions, or enough information to assess the test’s scoring setup, image selection, or repeatability. The product-image figures are the author’s reported results from one site’s sample, not an independent benchmark.

TypeSafe has not specified in the supplied release details what level of accuracy would make image analysis or bounding-box output production-ready. Its warning that results remain jagged leaves open how often the model misidentifies objects, assigns misleading confidence, or fails on images outside the tested formats and dimensions. The longer-context option carries a separate caveat: TypeSafe says reliability drops on longer inputs, but the source gives no detailed performance measurements.

Other implementation questions remain, including how the company will formalize its stated deprecation notice policy. TypeSafe says it gives 60 days’ notice for Jev 1.13.0 and 30 days for 1.14.0, with a formal policy still to come. The release material also says the company may use opted-in queries to find model weaknesses but will not train on them; the supplied account does not detail retention or other data-handling terms.

Alias Update and Further Evaluation

TypeSafe said jev-latest would move to Jev 1.14 around October 21, one week after the October 14 release. Until then, the new version is available through jev-preview, and users can test image input or bounding boxes only after sending the required opt-in header. TypeSafe has not provided a confirmed date for its formal model deprecation policy in the source material.

For readers interested in the reported test, the next useful development would be the complete results for news images and essays, along with the test method and any repeated trials. Those details would help show whether the product-photo separation generalizes beyond one small collection. TypeSafe’s own release advice, as reported by Meyer, is to run independent evaluations rather than rely on the company’s internal estimates.

Key Questions

What changed in Jev 1.14?

TypeSafe added opt-in image input and a new box response type. Image input also works with Jev 1.13.0; the box type is available only in Jev 1.14.0.

Can Jev image features be used in production?

TypeSafe says it is not ready to vouch for the image and long-context features in production. They are disabled unless a user sends the relevant danger-zone opt-in header.

What did the 30-image test find?

Meyer reported strong separation for 10 product images: Jev 1.14 matched each product image to its own headline with scores from 0.65 to 0.89 and identified nine products correctly. The source supplied here cuts off before the full results for news images and essays.

What image files does Jev accept?

The release details list PNG and JPEG still images, 256 to 1,568 pixels per side, up to 16 MiB per file and 32 MiB per request. WebP and HEIC images need conversion, and one invalid image fails the request.

When does Jev 1.14 become the default latest model?

TypeSafe said the jev-latest alias would move to Jev 1.14 around October 21, 2026, one week after the release. The new version is available as jev-preview in the meantime.

Source: ThorstenMeyerAI.com