TypeSafe has released Jev 1.14.0 alongside experimental features designed for creative users while maintaining a distinction between production-ready and experimental capabilities.
The company introduced what it calls “danger zone” flags—API headers that enable features TypeSafe considers less suitable for production but valuable for exploration and innovation. According to the blog post, these experimental features “will be here to stay for future models” and represent “early glimpses into what we’re focused on making reliable enough to meet TypeSafe standards.”
Three main experimental features are now available. First, composable image support has been enabled across Jev’s JSON interface, allowing users to incorporate images in state, instructions, or criteria. The company notes that “all models (Jev included) perform worse on multimodality over pure text,” recommending users break down complex questions into simpler decompositions. Image support is activated via the TypeSafe-Danger-Zone-Images: true header.
Second, TypeSafe introduced a new “box” type that returns a single bounding box without type errors. According to the announcement, this represents the company’s “first type that’s not purely meant for smart-if statements” and signals exploration into capabilities beyond decision-making.
Third, users can now access longer context lengths than previously enabled in production through the TypeSafe-Danger-Zone-Long-Context: true header. TypeSafe had intentionally limited context length because “Jev, like all models, is less reliable on longer contexts,” but has removed this constraint for experimental users.
The core Jev 1.14.0 release focuses on stability improvements. TypeSafe describes the update as delivering “small, subtle improvements to everything,” with the largest gains expected in multi-lingual support. The company has committed to a deprecation policy providing 60 days’ notice before removing Jev 1.13.0 and 30 days for Jev 1.14.0.
TypeS also announced an opt-in program allowing users to help identify model issues, though the company clarified it will not train on user queries. Additionally, TypeSafe is sharing internal intelligence measurements with developers, though it cautions these are “best guesses” that may evolve and should be validated against individual use cases.
Key facts
- Jev 1.14.0 includes experimental image support, accessible via the TypeSafe-Danger-Zone-Images header
- A new ‘box’ type has been introduced for returning bounding boxes, marking the first type beyond decision-making
- Longer context lengths are now available through the TypeSafe-Danger-Zone-Long-Context header, though TypeSafe notes models are less reliable with extended contexts
- The release emphasizes small improvements across stability and multi-lingual support rather than major feature additions
- TypeSafe is sharing internal intelligence measurements with developers while cautioning against over-reliance on the metrics