Jev And 30 Images: A Personal Test Of AI Vision

🔍 Read the full analysis: Jev And 30 Images: A Personal Test Of AI Vision on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on monitors, keyboards and dev gear

• Business-only prices and quantity discounts
• Tax-exempt purchasing
• Multiple users, one account, clear invoices

Create a free Business account →Free to register

As an affiliate, we earn on qualifying purchases.Hide

TypeSafe released Jev 1.14 on October 14, adding opt-in image input and a new bounding-box response type to its decision model. In a personal test published the next day, Thorsten Meyer of ThorstenMeyerAI.com tried the features on 30 images from his site; the supplied account reports strong results for product photos but ends before giving the outcomes 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 outputs such as a probability, a selection from a list, or a score. Before the release, those questions had to be about text. TypeSafe says images can now be placed in the state, a question’s instructions, or criteria used to define options and scoring levels. Image input works with jev-1.13.0 and jev-1.14.0; bounding-box output is limited to Jev 1.14.

The image capability and longer-context option are gated by explicit request headers TypeSafe labels “danger zone.” The company says its new features remain too inconsistent for it to recommend them for production use. The long-context header raises request limits from 64,000 to 262,000 tokens, and per-question limits from 32,000 to 262,000, but TypeSafe advises against the larger setting because reliability declines on long inputs.

Meyer tested the model on the site’s 30 most recent featured images: 10 product roundups, 15 news posts and five personal essays. Each image was tested against its own headline, a randomly selected other headline and an unrelated headline about repotting a monstera. The product images also went through a selection question asking which of 10 products each image showed. Meyer says he ran about 60 requests across both model versions.

What the Image Test Revealed

The reported product-photo results suggest Jev’s image input may be useful for narrow, repeatable classification checks, such as checking whether a product image matches a headline. Jev 1.14 assigned the correct headline scores from 0.65 to 0.89, with a median of 0.76. Random alternative headlines scored no higher than 0.18, with a median of 0.06, while the unrelated plant headline stayed at or below 0.10.

In the product identification task, Jev 1.14 selected the right product in nine of 10 images. Its one reported error was a mesh Wi-Fi photograph that showed both a router with antennas and a cylindrical mesh unit; the model chose “Wi-Fi 7 router” with 0.60 confidence. Meyer says the correct answers each had confidence of at least 0.77, suggesting a threshold might have flagged the error in this small sample. That is a test observation, not evidence that a confidence cutoff will reliably catch mistakes in other settings.

The results matter to teams considering automated image checks, but the release’s opt-in design and TypeSafe’s warning are a reminder that a promising demonstration is not the same as production validation. The test involved one site’s images and a limited set of prompts. Its available account does not include the results for the 15 news posts or five essays, so it cannot support a broader conclusion about image understanding.

How TypeSafe Gated Image Input

TypeSafe describes Jev as a decision model for structured tasks, rather than a general-purpose text generator. In September, Meyer wrote about the model as an example of what he called “System One” AI and later outlined 24 potential uses. Until the October 14 update, Jev’s typed questions applied to text only, according to his account.

TypeSafe says the image feature was also available in Jev 1.13, describing that version as having been “multimodal all along.” Its new bounding-box type, which returns one box and a confidence for a named target, is available only in 1.14. The company also offers an optional way for it to use customer queries to identify model weaknesses, while saying it will not train on those queries.

The release has practical image restrictions: PNG and JPEG stills only, with each side between 256 and 1,568 pixels. A file can be up to 16 MiB and a request up to 32 MiB; one invalid image fails the whole request. TypeSafe says images are billed as ordinary input tokens at Jev’s stated rate of $0.042 per million input tokens, with no image surcharge. Meyer recommends multipart uploads because base64 encoding increases file size by about a third.

“was multimodal all along”

— TypeSafe, describing Jev 1.13.0

Results Still Missing From the Test

The source material ends as Meyer begins discussing the 15 news images, at the words “The news posts were a dif”. It does not provide the results for those posts or the five essays, nor does it report a complete comparison between Jev 1.14 and Jev 1.13 across all 30 images. The stated total of about 60 requests is not accompanied by full task-level results.

The sample is also a personal evaluation, not an independent benchmark or a peer-reviewed study. It covers images from one website, uses a small set of prompts and does not establish how the model performs on different image types, image quality, or real-world workflows. TypeSafe’s own warning indicates that it does not yet endorse these features for production use.

Jev 1.14 Rollout and Further Testing

TypeSafe made Jev 1.14 available under the jev-preview model name. According to the release details reported by Meyer, jev-latest is scheduled to move to 1.14 around October 21, one week after launch. TypeSafe also outlined deprecation notice periods of 60 days for version 1.13 and 30 days for 1.14, with a formal policy still to come.

For users weighing the image features, the next useful evidence would include the missing news-image and essay results, repeat tests on larger and more varied samples, and evaluations tied to each intended use. TypeSafe’s internal intelligence estimates compare Jev only with open-weight models, and the company says those figures may favor Jev. Meyer recommends running independent evaluations rather than relying on those estimates.

Key Questions

What changed in Jev 1.14?

TypeSafe added opt-in image input and a new bounding-box response type. The box feature is available in Jev 1.14; image input also works with Jev 1.13, according to the release account.

How did Jev perform on the product photos?

In Meyer’s test, Jev 1.14 separated the 10 product images from headline alternatives, and identified the pictured product correctly in nine of 10 cases. These are results from a small personal test, not a general performance guarantee.

Did the test cover all 30 images?

Meyer says he selected 30 images and tested them, but the supplied source cuts off during the section on news images. It does not include the reported outcomes for the news posts or essays.

Is TypeSafe recommending the image feature for production?

No. TypeSafe says the features remain too inconsistent for it to vouch for them in production. Image support and longer context require explicit opt-in headers.

When will Jev 1.14 become the latest model?

The release account says Jev 1.14 is available as jev-preview, with jev-latest expected to move to it about one week after the October 14 launch, around October 21, 2026.

Source: ThorstenMeyerAI.com