🔍 Read the full analysis: Jev Can See Now—Here’s My 30-Image AI Experiment on ThorstenMeyerAI.com
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
TL;DR
TypeSafe released Jev 1.14 on October 14, adding opt-in image inputs and a bounding-box answer type; image input also works with Jev 1.13. In Thorsten Meyer’s test of 30 site images, Jev 1.14 separated product images from unrelated headlines and identified 9 of 10 products, but the source account is cut off before reporting the news-image results.
Jev is a decision model that answers structured questions with outputs such as yes-or-no probabilities, choices from a list, or scores with confidence values. Before this release, its questions were text-based. TypeSafe says image input works with both jev-1.13.0 and jev-1.14.0; the new box type, which returns one bounding box and a confidence score for a target, is limited to Jev 1.14.
Image use requires the request header TypeSafe-Danger-Zone-Images: true. Meyer tested the feature on the 30 most recent featured images on ThorstenMeyerAI.com: 10 product roundups, 15 news posts and five essays. He asked whether each image fit its own headline, a randomly selected different headline, and an unrelated headline about repotting a plant. The 10 product images also received a question asking the model to select which product each showed.
For the product images, Meyer reported that Jev 1.14 assigned each image’s own headline a score of 0.65 to 0.89, with a median of 0.76. Random other headlines scored no higher than 0.18, with a median of 0.06; the plant headline scored at most 0.10. Jev 1.14 selected the right product for 9 of 10 images. Jev 1.13 selected all 10 correctly, though two answers had confidence below 0.5. These are results from one publisher’s small test, not an independent benchmark.
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
Testing Image Matching in Practice
The experiment offers an early example of where image-enabled structured models may be useful: repeatable checks such as verifying whether a featured image matches a headline, or classifying which product appears in a photo. Such checks could help publishers review large image libraries, although Meyer’s test does not establish how well the approach works across other sites, image styles or subject areas.
The confidence scores also point to a possible review aid, rather than a guarantee. In Meyer’s product-identification results, the one incorrect answer had a confidence of 0.60, while correct answers were at least 0.77. That pattern suggests a threshold might have flagged the error in this sample; it does not show that the same cutoff would reliably separate correct and incorrect answers in other datasets.
For users, the release’s opt-in design matters as much as the feature. TypeSafe says image and long-context capabilities remain experimental, so they do not change existing behavior unless a request explicitly enables them. That gives developers a way to test the features while leaving production requests on the prior default.
As an affiliate, we earn on qualifying purchases.
Jev’s Opt-In Release Design
TypeSafe published Jev 1.14 on October 14, 2026, describing the release as a balance between users seeking reliability and those wanting to experiment. The company said Jev 1.13 had multimodal capability, although image support had not previously been exposed for these requests. Jev 1.14 is available as jev-preview; TypeSafe said jev-latest would move to it about a week after release, around October 21.
The release also offers an opt-in long-context header that raises limits from 64,000 tokens per request and 32,000 per question to 262,000 tokens. TypeSafe advises against the expanded limit because reliability can decline on long inputs. The company also says users may allow queries to be used to identify model weaknesses, while stating that it will not train on those queries.
For image requests, the documented limits are PNG or JPEG stills, 256 to 1,568 pixels per side, up to 16 MiB per file and 32 MiB per request. TypeSafe says images are billed as regular input tokens at $0.042 per million input tokens, with no separate image surcharge. Meyer also cautioned that the company’s internal comparisons against open-weight models are not public benchmark results and may favor Jev.
““There is still way too much jaggedness for us to vouch for these for production use cases.””
— TypeSafe, in its release notes as quoted by Thorsten Meyer
As an affiliate, we earn on qualifying purchases.
Limits of the 30-Image Test
The supplied account ends as Meyer begins describing results for the 15 news images, so those findings are not available here. It is also unclear how the model performed on the five essays, whether the image-matching results hold on a larger or more varied dataset, and how often confidence scores would flag mistakes outside this sample.
The report does not provide an independent evaluation, repeated trials, or a comparison with human judgments. TypeSafe’s own warning says the new capabilities are not yet suitable for production use in its view. Its internal estimates are not public benchmark scores, and the company says they may be biased in Jev’s favor.
Other practical details matter to deployment: images must meet the file and format limits, and one invalid image can cause the whole request to fail. The reported results do not establish whether the new bounding-box feature is accurate; Meyer’s described experiment focused on headline matching and product selection.
As an affiliate, we earn on qualifying purchases.
Preview Rollout and Further Testing
TypeSafe made Jev 1.14 available as jev-preview and said jev-latest was scheduled to switch to it around October 21, 2026. The company’s stated deprecation windows are 60 days for Jev 1.13.0 and 30 days for Jev 1.14.0, with a formal policy still to come.
For now, TypeSafe’s position is that developers should enable the new features deliberately and test them against their own tasks. Further evidence would need broader evaluations, including results for news and editorial images, before the reported product-image performance could support wider conclusions about production reliability.
AI-powered image analysis software
As an affiliate, we earn on qualifying purchases.
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 box response type that returns one bounding box with a confidence score. Image input also works with Jev 1.13.0; the box type is Jev 1.14-only.
How did Jev perform on the 30 images?
In Meyer’s product-image tests, Jev 1.14 gave each image’s own headline a higher score than the tested alternatives and identified 9 of 10 products. The provided account does not include the results for the 15 news images or five essays.
Can developers use the image features in production?
TypeSafe says the capabilities remain too jagged for it to vouch for production use. Image support is behind an explicit danger-zone opt-in header, leaving existing request behavior unchanged by default.
Does the test prove Jev reliably understands images?
No. It was a 30-image test from one website, not an independent or large-scale benchmark. The results offer an early example, but do not establish performance across other datasets or production settings.
Source: ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
