Jev 1.14.0 Arrives as a Preview With Opt-In Flags for Images and Long Context

Typesafe has released Jev 1.14.0 under the jev-preview tag. The company announced the version on Oct. 14, 2026, and said it brings small improvements across the board, with the largest gain in multi-lingual support Typesafe.

The tag is the key detail. This is a preview for testing, not the default version for production.

The release adds experimental "danger zone" flags, sent as API headers, or small settings attached to a software request, for features Typesafe does not approve for production use. Developers turn on a feature by sending its header. Send no flag and you stay on the unchanged stable path.

Image support is one opt-in, turned on with TypeSafe-Danger-Zone-Images: true. Typesafe said it works by find-and-replace in assets and can be used in state, instructions, or criteria. The company also said the prior model jev-1.13.0 was multimodal, able to handle images and text, so this flag continues that work. A new box type, which returns a single bounding box, or one rectangle marking an object in an image, is enabled only for jev-1.14.0.

Longer context, or how much text the model can consider at once, uses the same opt-in pattern. It is turned on with TypeSafe-Danger-Zone-Long-Context: true, without changing the default for existing callers. Typesafe also added an opt-in to share user queries to help find model jaggedness, or uneven behavior. The company said it will still not train on those queries. The scope is diagnostic, not data reuse.

The broader context here is worth spelling out for teams planning tests. In my view, Typesafe is keeping the interface stable while letting behavior behind flags change. That avoids extra integration work for users on defaults and limits risk to those who opt in.

Looking at what this means for preview users, keep tests separate. Images, long context, and the single-box output each need their own coverage, because combined flags make failures hard to trace. Multilingual handling needs its own pass, since that is the main claimed improvement. Sharing queries needs a governance decision, even with the no-training promise. Production use is out of scope. The value now is in small, reversible trials, like trying image replacement only in criteria, trying long context on a little traffic, and checking if one box simplifies parsing, before any path becomes supported.