TypeSafe’s Jev 1.14 release changes the API’s scope, not just its model ranking. Images can appear in the shared state, question instructions, or answer criteria; a new box type locates a target in an image. Both features are explicitly experimental.

I read the complete announcement and the linked image, Box, and model documentation. These are vendor-documented capabilities and limits, not independently reproduced API or accuracy results.

Images are composable, but placement matters

  • Send TypeSafe-Danger-Zone-Images: true; TypeSafe documents image support for both Jev 1.13 and 1.14.
  • Put shared images in state; repeating an image in several questions adds it repeatedly to the expanded input.
  • With raw HTTP, assets maps reference strings to uploaded images; use collision-resistant references, not ordinary words.
  • Prepare upright PNG/JPEG files yourself: the documented API strips metadata and does not apply EXIF orientation or resize oversized images.

A valid box is not proof that the object exists

  • box requires jev-1.14.0 or the currently matching jev-preview alias.
  • The documented answer is one rectangle with normalized edges—not a list of detected objects.
  • There is no ‘not found’ answer: asking about an absent target still returns a box.
  • Ask a separate Noul yes/no question about presence and ignore the box when that answer is low; validate that gate on your own images.
  • TypeSafe gives no recommended Box confidence threshold. Its sample threshold is an illustration, not a calibration guarantee.

Pin before retesting

At capture time, TypeSafe’s model page maps jev-latest to 1.13 and jev-preview to 1.14. The announcement says the stable alias will move in a week. Pin a version and log the response’s model ID if your workflow depends on tuned thresholds.

TypeSafe documents default limits of 64k tokens per request and 32k per question, rising to 262k for both with TypeSafe-Danger-Zone-Long-Context: true. Those are vendor limits, not measured usable-context guarantees; TypeSafe explicitly warns that longer context is slower and less accurate.

The release claims broad improvements, especially multilingual performance, but its internal comparisons use a guarded validation set and acknowledge possible bias. Typed output is not a correct decision—and the new Box type makes that distinction particularly concrete.