Jev 1.14: experimental images, Box and longer context
What changed on October 14, how model aliases resolve, and which new capabilities remain experimental.
The October 14 release
TypeSafe's model reference dates Jev 1.14 to October 14, 2026. The release announcement describes broad incremental improvements, particularly multilingual support, alongside experimental image inputs, Box questions and longer context.
As checked on October 16, jev-preview points to jev-1.14.0, while jev-latest still points to jev-1.13.0. Pin a version when comparing results; aliases can move.
What multimodal means here
Jev 1.13 and 1.14 accept images when explicitly enabled. A picture can be part of state, question instructions or criteria. Choice, Score and Noul still return typed judgments; the new Box type locates one target and requires 1.14. These are image-understanding capabilities, not image generation or native audio/video input.
Images and Box require TypeSafe-Danger-Zone-Images: true on each request. TypeSafe labels them experimental and below its production-readiness standard. See our Images and Box guide for formats, limits and presence checks.
Longer context is a separate opt-in
The default limits remain 64k tokens per request and 32k per question. The separate TypeSafe-Danger-Zone-Long-Context: true header raises both to 262k, with slower responses and lower accuracy according to the model reference.
This is independent coverage, checked October 16. Our local explorer demonstrates output shapes; it does not call the model.