Jev 1.14 introduces experimental object bounding boxes

How the image reaches the model

The image documentation describes supplying an image once in the assets collection, then referencing it in the shared input, question instructions or criteria. The API replaces those references with the image; each request containing an image or Box question requires the TypeSafe-Danger-Zone-Images: true header. Repeated references count the image again toward the expanded request limit.

A rectangle is not proof of presence

For developers locating an object, Box takes instructions naming the target and image, then returns one rectangle per question. Its four edges are fractions of image width or height, accompanied by a confidence value indicating how sure the model is. An absent target still produces edges: TypeSafe advises asking a separate Noul presence question in the same request and ignoring the rectangle when that answer is low.

Longer input is a separate opt-in

The model documentation makes longer context, meaning a larger allowance for processed input, available to both Jev 1.13 and 1.14. Default allowances are 64,000 tokens per request and 32,000 per question; the TypeSafe-Danger-Zone-Long-Context: true header raises both to 262,000 for that request. TypeSafe says this mode is slower and less accurate than staying within the default limits, and recommends enabling it only when needed.

MarkTechPost’s overview summarizes the release and links the provider’s announcement and technical documentation.