Jev Day shipped two things at once: a routine model bump and a deliberate widening of what the API accepts. Jev 1.14.0 is the routine part — small improvements everywhere, the largest expected in multilingual support. The rest is new surface area for builders who want to prototype: images, the box question type, and 262k context, all behind headers TypeSafe calls the danger zone.
What is actually stable
jev-1.14.0 ships as jev-preview at the same price and default context as 1.13.0: $0.042 per million input tokens, output free, 64k per request and 32k per question. jev-latest stays on 1.13.0 for about a week, then moves. TypeSafe intends 60 days' warning before 1.13.0 is formally deprecated and 30 days for 1.14.0, with a deprecation policy landing first. If you tuned thresholds against a version, pin the versioned ID and move on your schedule — the model alias glossary covers why.
Images were there all along
The surprise in the post is that jev-1.13.0 was multimodal the whole time; TypeSafe just did not expose it. Images now ride in an assets map keyed by a reference string, and the reference can sit anywhere the JSON interface allows — in state, in instructions, or in criteria. That composability is the interesting part: an option description can carry a reference image, and so can a Noul's true/false criteria.
The limits are real: PNG or JPEG stills, 256 to 1568 pixels per side, 16 MiB per file, 32 MiB per request body shared after processing. And TypeSafe's own warning is blunt — all models, Jev included, perform worse on images than on text, so decompose into the simplest possible questions.
Box is the first type that is not a decision
Every Jev question type until now mapped to a branch: Choice to a switch, Noul to an if, Score to a sort. Box returns four edges as fractions of the image plus a confidence — it answers where, not which. It needs 1.14.0 (or jev-preview) and the images header, and it takes no criteria.
One design detail decides how you use it: Box always returns a box. An absent target still gets four edges, so the documented pattern pairs it with a Noul presence check. TypeSafe's own small test on 36 painting crops, each cutting off part of the target, shows the shape of the problem: four per-edge Noul questions caught 30 cut-off crops, a single "is all of it inside" question caught 26, and drawing the box as an outline and asking about it caught 20.
The danger zone is a contract marker
The headers — TypeSafe-Danger-Zone-Images: true and TypeSafe-Danger-Zone-Long-Context: true — are TypeSafe's way of splitting reliability from experimentation. Features behind them will persist and improve, but the docs say plainly that format, limits, and accuracy can change between releases. For a production system the header is a flag that a request sits outside the reliability bar; for a prototype it is a cheap way to test whether a workflow is worth building properly later.
262k is permission, not advice
The context change is the most tempting one to misuse. Both caps rise to 262k with the long-context header, on 1.13.0 and 1.14.0, at the same per-token price — but the docs say the mode is slower and less accurate than staying inside the defaults, and the release post still strongly recommends limiting context. The shorter cap was a product choice; treat the header as a pressure valve for the rare long request, not a new default.
What to do this week
- Pin
jev-1.13.0orjev-1.14.0explicitly if thresholds depend on a version; log the response'smodelfield. - Try images and Box on a throwaway workflow — the box study and imajev-4b are two community starting points.
- Keep the rejection and presence checks around any Box work: a box for a missing target is still a box.
- Re-run your evals against 1.14.0 before
jev-latestmoves, since the multilingual claims are TypeSafe's own and its internal chart is self-described as biased.
The headline is not the model bump. It is that Jev's interface now reaches past text and past decisions — carefully, behind headers, with the company saying out loud that the features are not ready. That is a more honest launch posture than most.
