🔍 Read the full analysis: Jev Looks At My Pictures: A Test With 30 Images on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
TypeSafe added opt-in image input and bounding-box output to its Jev decision model in release 1.14 on October 14, 2026. In a test of 30 images published by Thorsten Meyer, Jev 1.14 matched product photos to their headlines and identified 9 of 10 products, while the supplied report cuts off before giving complete results for news and essay images.
TypeSafe added image input and a new bounding-box output to its Jev decision model on October 14, with both features disabled unless users explicitly opt in. In a test published the following day, Thorsten Meyer reported strong results for product photos, including correct product identification in 9 of 10 images with Jev 1.14, but the source account available here ends before reporting full findings for news and essay images.
Jev is a model designed to answer structured questions rather than generate prose. Users provide a state and typed questions; responses can be yes-or-no probabilities, selections from a list, or scores, accompanied by confidence values. Before this release, the questions were limited to text. TypeSafe says image input works with jev-1.13.0 and jev-1.14.0, while the new box question type is limited to Jev 1.14.
The image capability is opt-in through the request header TypeSafe-Danger-Zone-Images: true. Images can be referenced in the request state, question instructions or criteria. The box type lets a user describe one target and receive a bounding box with a confidence score. TypeSafe also added an opt-in long-context setting that raises request and question limits from 64,000 and 32,000 tokens to 262,000 tokens each; the company cautions that reliability can fall with longer inputs.
Meyer tested the two model versions on 30 featured images from his site: 10 product roundups, 15 news posts and five essays. Each image was assessed against its own headline, a randomly selected unrelated post headline and an off-topic houseplant headline. The 10 product images also received a multiple-choice product-identification question. Meyer says Jev 1.14 matched each product image to its own headline more strongly than to the comparison headlines, and identified the pictured product correctly in nine cases.
Jev 1.14 · images + box
Jev can see now. Here is what 30 of my images showed.
I asked about every featured image three times: does it fit its own headline, a random other post’s headline, an off-topic headline? Both Jev versions, about 60 requests.
Does this image fit the headline?
Median “yes” probability on Jev 1.14, by kind of post
- Product roundups10 images0.760.061.13: 0.79 / 0.05
- News posts15 images0.330.161.13: 0.51 / 0.25
- My essays5 images0.510.161.13: 0.62 / 0.33
Off-topic headline (“How to Repot a Monstera Plant”): never above 0.10 on either version, 60 of 60 rejected.
Why news is weak: almost all news images show the same scene, people at desks looking at monitors. Jev could not tell them apart, and it was right not to.
1.14 is stricter than 1.13
Same 30 images, same questions
| Result | 1.14 | 1.13 |
|---|---|---|
| Fits own headline (≥ 0.5) | 15/30 | 23/30 |
| Fits a wrong headline (≥ 0.5) | 1/30 | 4/30 |
| Own beats random other (AUC) | 0.91 | 0.89 |
| Names the product (10 options) | 9/10 | 10/10 |
1.14’s one product miss: mesh system read as “Wi-Fi 7 router”, confidence 0.60. Re-tune thresholds before jev-latest moves.
What shipped on Oct 14
- headerImagesin state, instructions or criteria · 1.13 + 1.14 ·
TypeSafe-Danger-Zone-Images - 1.14 onlyBox typeone bounding box + confidence. 6 of 6 clear targets boxed (0.66–0.90); an absent bicycle still got a box (0.03)
- headerLong context64k / 32k → 262k tokens, slower and less accurate
- opt-inHelp find weaknessesqueries used for testing, not training
now: jev-preview = 1.14~Oct 21: jev-latest → 1.14notice: 60 days (1.13), 30 days (1.14)
Source: own test on October 15, 2026. 30 featured images from ThorstenMeyerAI.com (10 product roundups, 15 news posts, 5 essays), JPEG at up to 1,568 px, jev-1.14.0 and jev-1.13.0 with the image header. Bars show medians. AUC = probability that an image’s own headline scores higher than a random other headline. Features, prices and limits: TypeSafe blog post “Jev 1.14” and docs.typesafe.ai, as of October 15, 2026.
Powered by Thorsten Meyer AI
What the Product Photo Test Shows
The test offers an early, concrete example of how a structured decision model might help with routine image checks, such as flagging whether a featured image fits a headline. Meyer describes that task as one that is often not checked manually. If the reported pattern holds in broader evaluations, teams could use confidence scores to prioritize images for human review rather than treating every model answer as equally reliable.
The results also show why a successful example should not be read as general proof of image understanding. The product test involved a small, specific set of images and questions, and the available account does not provide the complete news- and essay-image results. Meyer reports one product-identification error: Jev 1.14 labeled a photo showing a router beside a cylindrical mesh unit as a Wi-Fi 7 router, with confidence of 0.60. He says the nine correct identifications each had confidence of at least 0.77. That pattern suggests confidence thresholds may help screen outputs in this test, but it does not establish a dependable threshold for other uses.
product photo identification AI tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
A Cautious Opt-In Release
TypeSafe presented Jev 1.14 as a release balancing experimentation with production reliability. The company said the image and long-context functions have enough rough edges that it would not vouch for them in production use cases. Keeping them behind explicit headers leaves the default behavior unchanged and requires developers to choose whether to test the features.
Image references are inserted wherever a matching reference string appears in a request, according to Meyer’s account of the implementation. He advises using a fresh UUID because matching is plain text: a short reference could unintentionally match part of another word. Images placed in shared state are available to all questions; images embedded in individual question instructions may be sent again for each question that uses them, affecting input use.
The supported formats and limits are also narrow: PNG and JPEG still images, with each side between 256 and 1,568 pixels, up to 16 MB per image and 32 MB per request. WebP and HEIC files must be converted first, and the source says one invalid image can fail the entire request. TypeSafe bills image input as ordinary input tokens at its stated rate of $0.042 per million input tokens; base64 encoding adds roughly one-third to file size, while multipart upload avoids that encoding overhead.
“There is still way too much jaggedness for us to vouch for these for production use cases.”
— TypeSafe, describing the status of the new capabilities in its release post, as quoted by Thorsten Meyer
image recognition software for product images
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The Missing News and Essay Results
The source material supplied for this report stops during Meyer’s section on news images, after the heading “News images: Jev was right, my images weren’t” and the incomplete phrase “The news posts were a dif”. It does not provide the completed news-image findings, any results for the five essays, or a final comparison between Jev 1.13 and 1.14 across all 30 images. Those results cannot be inferred from the product-image figures.
The test is an individual evaluation on one site’s featured images, not an independent benchmark. The account also does not establish whether the images, headlines or scoring criteria were reviewed by other evaluators, or how well performance carries over to different image types and production workflows. TypeSafe itself cautions that it cannot yet vouch for the new features in production, and its stated confidence values should not be treated as a guarantee of correctness.
AI image bounding box annotation tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Jev 1.14 Rollout and Further Testing
TypeSafe made Jev 1.14 available as jev-preview on October 14, 2026. The company said the jev-latest alias would move to that version one week after release, around October 21. It also outlined deprecation notice periods of 60 days for Jev 1.13.0 and 30 days for Jev 1.14.0, while saying a formal policy is still to come.
For now, TypeSafe’s recommendation is to treat the capabilities as experimental and evaluate them against a developer’s own use case. Meyer likewise cautioned readers to run their own evaluations, noting that TypeSafe’s internal intelligence estimates compare Jev only with open-weight models and may be biased in Jev’s favor. The complete news- and essay-image results from his test, if published, would help clarify how far the product-photo findings extend.
As an affiliate, we earn on qualifying purchases.
Key Questions
What changed in Jev 1.14?
TypeSafe added opt-in image input and a new bounding-box question type. Image input is supported on Jev 1.13.0 and 1.14.0; box output is available only on 1.14. The release also offers an opt-in setting for longer context.
How did Jev 1.14 perform on the product images?
In Meyer’s test, Jev 1.14 matched each of the 10 product images more strongly to its own headline than to the comparison headlines. It named the correct product in 9 of 10 cases. These figures come from a small test on images from one site, not a general benchmark.
Did the test show how Jev performed on news and essay images?
Not in the source material provided. It cuts off during the news-image discussion and contains no completed results for the 15 news posts or five essays, so their performance cannot be reported here.
Are Jev’s image features ready for production?
TypeSafe said it would not yet vouch for the new capabilities in production use cases because of reliability issues. Image support requires the explicit TypeSafe-Danger-Zone-Images: true header, giving developers a way to test it without changing default requests.
Source: ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
