Can AI Jev Make Sense Of Pictures I Created?

🔍 Read the full analysis: Can AI Jev Make Sense Of Pictures I Created? on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on office and shipping supplies

• 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 with opt-in support for image inputs and a new bounding-box response, allowing its decision model to answer structured questions about pictures. In a report published the next day, Thorsten Meyer said his test found strong headline matching for product images. TypeSafe said the new capabilities remain too unreliable for production use without further evaluation.

Jev is designed to return structured decisions rather than free-form prose: depending on the question, it can provide a probability, choose from listed options or assign a score, along with confidence. Before this release, users had to ask about text. TypeSafe says image input also works with jev-1.13.0, describing that model as having been multimodal already. The new bounding-box type, which returns one box for a specified target, is limited to Jev 1.14.

Both features require an explicit request header. Image input uses TypeSafe-Danger-Zone-Images: true; the box output is also presented as experimental. The release adds a separate opt-in header for longer context, raising the limits from 64,000 tokens per request and 32,000 per question to 262,000 for each. TypeSafe says reliability can decline on long inputs and advises caution with the higher limits.

Meyer tested 30 featured images from his site: 10 product roundups, 15 news posts and five essays. He asked whether each image suited its own headline, a randomly selected different headline and an unrelated plant-care headline. For the 10 product images, he also asked Jev to identify which product was shown. He says the test involved about 60 requests across Jev 1.13 and 1.14. The available results cover the product-image portion, not the full set.

What Image Decisions Could Enable

Structured image judgments could let developers build repeatable checks into workflows that previously relied on a person inspecting each picture. A publisher, for example, might ask whether an image appears relevant to a headline or identify a product in a photo. Jev’s format can return a score or selection rather than a paragraph, which may be easier to use in an automated process.

Meyer’s results come from a small test and do not establish broad accuracy. In the product-image sample, Jev 1.14 gave stronger matches to the images’ own headlines than to mismatched headlines and identified nine of the 10 products correctly. The report does not establish whether confidence scores or thresholds would reliably detect errors in other image sets. TypeSafe says the capabilities need further evaluation before production use.

Jev’s Opt-In Release Model

TypeSafe describes the release as balancing the needs of developers seeking predictable behavior with those experimenting on new capabilities. Existing behavior remains unchanged unless a request includes the relevant header. The company labels the image feature a “danger zone” and says there is still too much inconsistency for it to endorse production use.

Image references are declared once in an assets map and inserted wherever the reference appears in the request, including the state, question instructions or criteria. The documentation described by Meyer limits inputs to PNG and JPEG still images measuring 256 to 1,568 pixels per side, with a 16 MiB limit per file and 32 MiB per request. A failed image can fail the entire request; WebP and HEIC files need conversion. TypeSafe says images incur ordinary input-token charges, at $0.042 per million input tokens, with no separate image surcharge. Base64 encoding increases file size by about one-third and counts toward the request limit.

Jev 1.14 is available as jev-preview. TypeSafe said jev-latest would move to that version about a week after release, around October 21. The company also shared internal intelligence estimates rather than public benchmark scores and cautioned that its comparisons with open-weight models may favor Jev. Meyer recommends independent evaluation rather than treating those estimates as settled evidence.

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

— TypeSafe, as quoted in Meyer’s report

How Far the Test Results Reach

The supplied report ends as Meyer begins discussing the 15 news images, so it provides no results for those posts or the five essays. It also does not include a complete breakdown of the image-matching questions across both model versions. The product results should not be treated as a general accuracy rate: the sample was small, drawn from one website and designed by the author.

Other practical questions remain open, including how performance changes with different image types, image quality, wording or larger collections. TypeSafe’s warning about production reliability is explicit, but the report does not quantify the error rate that developers should expect. The company’s internal estimates are not public benchmark results, and it says their comparisons may be biased toward Jev.

Preview Rollout and Independent Testing

The next scheduled change is the planned move of jev-latest to Jev 1.14 around October 21, one week after the release. TypeSafe said it would give 60 days’ deprecation notice for version 1.13.0 and 30 days for 1.14.0, with a formal policy still to come.

Developers can evaluate the preview on their own images and workflows using the opt-in headers and review uncertain responses. TypeSafe has not supplied a firm date for removing the experimental status or a public benchmark that measures image-feature performance across broader use cases.

Key Questions

What changed in Jev 1.14?

Jev 1.14 adds opt-in image input and a new bounding-box answer type. The box feature is limited to Jev 1.14; TypeSafe says image input also works with Jev 1.13.0.

What did Meyer’s image test find?

For 10 product images, Meyer reports that Jev 1.14 matched each image to its own headline more strongly than to mismatched headlines and identified the product correctly in nine of 10 cases. The supplied report does not give the results for the news images or essays.

Is Jev’s image feature ready for production?

TypeSafe says it is not ready for production use without further testing, citing inconsistent results. The image capability also requires an explicit opt-in header.

How much does image input cost?

TypeSafe says image input is billed as ordinary input tokens at $0.042 per million input tokens, with no separate image surcharge. The amount depends on the input-token usage.

Which image formats does Jev accept?

The limits reported by Meyer cover PNG and JPEG still images. WebP and HEIC images need conversion first; each image must meet the stated size and dimension limits.

Source: ThorstenMeyerAI.com