Jev’s New Ability To See, Tested On 30 Of My Images

🔍 Read the full analysis: Jev’s New Ability To See, Tested On 30 Of My Images on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

• 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 image input and a bounding-box answer type to its decision model, and Thorsten Meyer tested the features on 30 images from his website. His published account reports strong results for product images, but TypeSafe says the new capabilities remain experimental and should be enabled only through explicit request headers.

Jev is a decision model that returns structured answers such as yes-or-no probabilities, selections from a list, or scores on a scale. Before the release, its questions had to concern text. TypeSafe’s October 14 post says image support also works with jev-1.13.0, while the new box answer type is limited to jev-1.14.0. That type returns one bounding box for a described target, along with a confidence score.

Image input is disabled by default and requires the TypeSafe-Danger-Zone-Images: true header. A separate header opts requests into longer context, raising the stated limits from 64,000 tokens per request and 32,000 per question to 262,000 for both. TypeSafe says reliability can decline on long inputs and advises against using the expanded limit without testing. The company also says customers may opt in to sharing queries to help identify model weaknesses, but that those queries will not be used to train the model.

Meyer tested 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 monstera. The product images also received a multiple-choice identification question. He says the roughly 60 requests were run across Jev 1.13 and 1.14. The supplied account gives detailed product results but ends before describing the news and essay results.

Image Checks Move Into Jev

The release gives developers a way to ask a structured model to assess images without turning every task into an open-ended description. That distinction matters for routine checks such as whether a featured image appears relevant to a headline, or whether a specified object is present and where it sits. In Meyer’s product-image test, Jev 1.14’s own-headline scores ranged from 0.65 to 0.89, with a median of 0.76; randomly paired headlines scored no higher than 0.18, with a median of 0.06. The unrelated plant headline scored below 0.10 for every product image.

Those figures are a small, self-run test, not an independent benchmark. Meyer says Jev 1.14 identified the right product in nine of 10 images. The one error was a mesh Wi-Fi image that showed both a router with antennas and a cylindrical mesh unit; Jev selected “Wi-Fi 7 router” with 0.60 confidence. Meyer reports that all correct identifications had confidence of at least 0.77, suggesting a confidence cutoff might flag uncertain results in this set. That observation does not establish how well such a cutoff would work on other images or tasks.

For readers and developers, the release is a practical capability change, but not a claim of production readiness. TypeSafe’s opt-in approach leaves existing behavior unchanged unless a request explicitly enables the feature. Anyone considering image-based automation still needs to test its own content, define how to handle low-confidence answers and account for the limits of this early release.

How TypeSafe Is Limiting the Release

In September, Meyer described Jev as a model for “System One” AI and outlined ways to use it for structured decisions. Its prior text-only interface asked developers to supply a state and typed questions, with answers returned in constrained formats rather than prose. The October update extends that interface to images. TypeSafe says jev-1.13.0 was multimodal, although image support had not previously been exposed in the documented workflow described by Meyer.

TypeSafe has put the image and longer-context features behind headers bearing the label “Danger Zone.” The company says the features still have “too much jaggedness” for it to vouch for production use. Jev 1.14 is available as jev-preview; TypeSafe said the jev-latest alias would move to it one week after release, around October 21. The company also shared internal intelligence estimates rather than public benchmark scores and cautioned that comparisons with open-weight models may be biased in Jev’s favor.

Image requests have technical limits: the documented formats are PNG and JPEG still images, with each side between 256 and 1,568 pixels, a maximum of 16 MiB per file and 32 MiB per request. The source says images are billed as ordinary input tokens at Jev’s stated rate of $0.042 per million input tokens, with no image surcharge. Base64 encoding adds about one-third to file size, while multipart uploads avoid that added encoding overhead; the Python and JavaScript SDKs support the workflow.

“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

What the 30-Image Test Leaves Open

The source account supplied for this article is incomplete: it ends as Meyer begins discussing the 15 news images and does not provide their scores, the results for the five essays, or his full comparison of Jev 1.13 and 1.14. Those findings cannot be inferred from the product-image results. The described test also uses a small set drawn from one website, and it is not an independent evaluation.

TypeSafe has not vouched for image features in production, and the source gives no public benchmark results for them. The practical reliability of bounding boxes, image interpretation across different subject matter, and the effect of image placement on cost and consistency remain dependent on further testing. The source also does not report whether the test used a confidence threshold in actual deployment or whether any errors were checked by a person.

Preview Rollout and Further Testing

At the time of Meyer’s October 15 report, Jev 1.14 was available through the jev-preview name, with TypeSafe’s planned jev-latest change expected around October 21. The release post described 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.

For developers, the immediate next step is evaluation on their own image sets before relying on the new output. Meyer’s account advises running independent checks rather than treating TypeSafe’s internal estimates or a 30-image sample as a substitute. Further reporting would need the omitted news and essay results, a clearer comparison across both model versions, and evidence on how the features perform across more varied images and real workflows.

Key Questions

What changed in Jev 1.14?

TypeSafe added opt-in image input and a new bounding-box answer type. The box type is available on Jev 1.14; TypeSafe says image input also works with Jev 1.13.0.

How did Jev perform on the product images?

In Meyer’s test, Jev 1.14’s product images scored 0.65 to 0.89 for matching their own headlines, with a median of 0.76. It identified the right product in nine of 10 images. These results come from a small, self-run test, not a general benchmark.

Are Jev’s image features ready for production?

TypeSafe says it cannot yet vouch for the image features in production use cases. They are off by default and require an explicit request header, so teams should test them on their own tasks before relying on the results.

What happened in the test of news images?

The supplied source ends just as Meyer begins reporting on the 15 news images. It does not include their results, so their performance cannot be stated from the available account.

Source: ThorstenMeyerAI.com