Evidence at a glance
From Seeing Images to Giving Software an Answer
TypeSafe AI released Jev 1.14.0 on October 14, 2026, initially making it available through the `jev-preview` alias. Jev is what the company calls a System One decision model: callers submit a `state` and typed questions, and the model returns structured answers such as Choice, Score, or Noul rather than treating generated prose as its main interface. This release adds experimental image input, a Box question type for bounding boxes, and an optional 262k-token context.
The interesting change is not that the model can finally “see images.” TypeSafe says Jev 1.13 already accepted images. The new step is bringing visual localization into the existing typed decision interface. For teams integrating a model into software workflows, coordinates and decision values are easier for downstream code to consume than an open-ended description. But a structured interface does not make its answers reliable by itself. TypeSafe marks images, Box, and long context as experimental and says they are not ready for production.
Box Returns Coordinates, but Cannot Say “Not Found”
Images are defined once in the request's top-level `assets` map, then referenced in `state`, `instructions`, or `criteria`. The API replaces each reference with the image before the model reads the request. Image requests require the `TypeSafe-Danger-Zone-Images: true` header. With Box, the caller describes the target, and the model returns `top`, `bottom`, `left`, and `right` edges, normalized to values from 0 to 1 relative to the image dimensions, along with a confidence score.
There is an important gap in this interface: Box always returns a box. It has no “not found” answer, even when the target is absent. TypeSafe recommends separating localization from existence checking, using a Noul question to check whether the target exists before, or alongside, using the coordinates. This is not merely a calling convention. The application has to own that logic. If it consumes Box output alone, an invalid coordinate can be passed to later steps as if it were a valid localization.
What 36 Cropped Images Show—and What They Do Not
TypeSafe's Box test used 36 paintings in which the target was partially cropped. Four edge-by-edge Noul questions detected 30 crops. A single question asking whether the target was entirely inside the image detected 26, while drawing a box and then asking about it detected 20. The result supports a limited conclusion: on this sample and for this task, breaking completeness into separate edge checks worked better than either one broad question or asking about a drawn box.
It does not show that Box will perform similarly in industrial inspection, retail shelves, or other real-world settings. This was a small internal test published by TypeSafe, not an independent replication, and the release provides no public benchmark score table for comparison. The company describes Jev 1.14's overall gains as small and expects the largest improvement in multilingual support, but provides no specific scores. Its materials also warn that image performance is worse than performance on pure text and recommend keeping questions simple. The 30-of-36 result is therefore a useful lead for designing evaluations, not an accuracy guarantee for deployment.
262k Is an Input Limit, Not a Reliability Promise
By default, Jev allows up to 64k tokens per request and 32k per question. Adding `TypeSafe-Danger-Zone-Long-Context: true` raises both limits to 262k. The same switch works with Jev 1.13. In other words, the version number can obscure what actually changed: image input is not exclusive to 1.14, and neither is long context. Box is the capability in this release that specifically requires 1.14.
A larger window solves the question of how much input can fit, not whether the model can make reliable judgments across a long input. TypeSafe says reliability declines on long inputs for Jev and other models, and still recommends shorter contexts. As with Box's need for a separate existence check, this leaves reliability work to the caller. Images are limited to static PNG and JPEG files, with a maximum of 16 MiB per image, 32 MiB per request body, and image dimensions between 256 and 1568 pixels per side. Input costs $0.042 per million tokens and output is free, but neither price nor capacity substitutes for task-level evaluation.
Ready for Evaluation, Not Yet for Production
For technical leads, the most sensible way to try Jev 1.14 is to add Box as a new output type in an existing decision workflow, not to treat it as a complete vision system. Separate “does the target exist?” from “where is it?”, set acceptable false-positive and false-negative rates for each, and test coordinate error, cropped boundaries, and performance across the image conditions that matter to your application. TypeSafe's 36-painting test can inform test design, but cannot replace those evaluations.
The deployment boundaries are clear. All three new capabilities require explicit danger-zone headers, and TypeSafe says they are not production-ready. The model is available only through TypeSafe's hosted API, with no disclosed parameter count, open weights, or hardware variants. Those constraints matter if your team cannot accept a hosted dependency or needs to audit visual failure modes closely. The practical judgment for now is to use the preview to test whether the interface fits your workflow, include Noul-based existence checks in the design, and treat neither the coordinates nor the 262k context as a production guarantee until your own evaluations support it.