Jev’s Image Vision: What It Did With 30 Of My Pictures

🔍 Read the full analysis: Jev’s Image Vision: What It Did With 30 Of My Pictures on ThorstenMeyerAI.com

TypeSafe’s Jev 1.14 decision model began accepting image inputs on October 14, 2026, and added a bounding-box answer type, according to the company’s release notes. In a test published the next day, Thorsten Meyer of ThorstenMeyerAI.com ran two versions of Jev against 30 featured images; the reported product-photo results were strong, while the supplied account does not include the results for news and essay images.

Jev is designed to answer structured questions rather than generate prose. Users provide a state and typed questions, and the model responds with a probability for a yes-or-no decision, a selection from a list, or a score, with confidence information. Before this release, the questions were limited to text, Meyer wrote. TypeSafe said image support also works with jev-1.13.0, describing that version as “multimodal all along.”

Images require the TypeSafe-Danger-Zone-Images: true header; the new box response type is available only in Jev 1.14. The image feature lets developers place a reference in the model’s state, question instructions, or criteria. TypeSafe’s documentation, as summarized by Meyer, says it accepts still PNG and JPEG files from 256 to 1,568 pixels per side, with a 16 MiB per-image limit and a 32 MiB per-request limit. A request containing one invalid image fails as a whole.

Meyer tested the 30 most recent images on his site: 10 product roundups, 15 news posts and five essays. He asked whether each image suited its own headline, a randomly chosen different headline, and an unrelated headline about repotting a plant. The product images also faced a question asking which of 10 products each showed. He ran the tests on Jev 1.14 and 1.13, for about 60 requests in total, according to his account.

Image Decisions Need Testing

The release gives developers a way to use images in structured classification and selection tasks, rather than asking a general-purpose chatbot to produce an open-ended description. That could be useful where a workflow needs a consistent answer, such as checking whether a picture matches a headline or identifying an item in a product image.

Meyer’s reported product results suggest that confidence scores may help flag uncertain answers, but the small test is not evidence of general performance across image types or business settings. In one example, Jev 1.14 selected the wrong product from a mesh Wi-Fi image, identifying a “Wi-Fi 7 router” with 0.60 confidence. Meyer said all correct product identifications scored at least 0.77, a pattern that could have exposed the mistake in this sample. That does not establish a reliable cutoff for other uses.

TypeSafe itself cautioned against relying on the new capabilities in production. Keeping them behind an explicit opt-in header lets users experiment while production behavior remains unchanged by default, the company said. For teams considering image-based automation, the release is a reason to run task-specific evaluations—not a substitute for them.

How Jev’s Image Inputs Work

TypeSafe describes Jev as a decision model: the caller defines the questions and acceptable answer formats, and the model returns structured results. For image requests, developers declare each image once in an assets map and use its reference wherever they want the image inserted. Meyer warned that references are matched as plain text, so short strings could unintentionally match parts of other words; he recommends using a fresh UUID.

Where an image is placed affects how often it is processed. An image in the shared state can be used across questions, while an image inserted into separate question instructions is sent again for each question that uses it. Images can also form part of a set of answer options. The release notes, as described by Meyer, say image inputs are charged as ordinary input tokens at $0.042 per million input tokens, with no separate image surcharge. Base64 encoding increases file size by about one-third; TypeSafe recommends multipart uploads, which its Python and JavaScript SDKs support.

The release also included a longer-context option, enabled with a separate danger-zone header. TypeSafe’s default limits remain 64,000 tokens per request and 32,000 per question; the opt-in setting raises both to 262,000. The company advised against the larger context because reliability can fall on long inputs. Jev 1.14 is available as jev-preview, with jev-latest scheduled to move to it about a week after release, around October 21, 2026.

„jev-1.13.0 was multimodal all along.“

— TypeSafe, as described by Meyer

Results Beyond Product Photos

The supplied source ends partway through Meyer’s discussion of the 15 news images, after the words “The news posts were a dif”. It provides no complete findings for those images or for the five essays, so their performance cannot be reported here. It also does not provide the full request data, image set, question wording, or evaluation method needed for an independent review.

The product test covered only 10 images from one site and a narrow set of questions. Its results do not show how Jev performs across different image sources, languages, lighting conditions, or production workloads. TypeSafe’s reported concern about reliability remains relevant, and the test does not establish whether the box response type performs accurately: Meyer’s account in the supplied material gives no bounding-box evaluation results.

TypeSafe also shared internal intelligence estimates rather than public benchmark scores, according to Meyer. The company compared Jev with open-weight models and acknowledged that the estimates may favor Jev. The source does not include the figures, so no broader performance comparison can be drawn.

Preview Rollout and Evaluations

TypeSafe made Jev 1.14 available under the jev-preview name on October 14, with the jev-latest alias expected to switch about one week later, around October 21. The company said users should evaluate the model against their own tasks and treat the new image features as experimental rather than production-ready.

TypeSafe also set out deprecation notice periods: 60 days for Jev 1.13.0 and 30 days for 1.14.0, with a formal policy still to come, Meyer reported. Developers assessing the release will need to test their own images, define confidence thresholds carefully, and account for file-format and request-size limits. The full news-image and essay-image findings from Meyer’s test are not present in the supplied source, leaving its overall 30-image conclusion incomplete.

Key Questions

What changed in Jev 1.14?

TypeSafe added opt-in image inputs and a new box answer type that returns one bounding box and a confidence score. The box type is limited to Jev 1.14.

Did the test show that Jev can match product images to headlines?

In Meyer’s test, Jev 1.14 gave the 10 product images higher scores for their own headlines than for randomly selected alternatives or an unrelated plant-care headline. This was a small test on one site, not a general performance guarantee.

How accurate was Jev at identifying the 10 products?

Meyer reported that Jev 1.14 identified 9 of 10 correctly. Jev 1.13 identified all 10, although two answers had confidence below 0.5.

Are the image features ready for production?

TypeSafe said there was “still way too much jaggedness” to vouch for the features in production use cases. The company placed them behind an explicit opt-in header, and Meyer’s limited test does not establish production reliability.

What were the results for news and essay images?

The supplied account cuts off during the discussion of news images and does not include results for the news or essay groups. Those findings are not confirmed by the material available here.

Source: ThorstenMeyerAI.com