TypeSafe AI has released Jev 1.14.0, a new iteration of its model designed to balance production reliability with creative experimentation. The update introduces a unique 'Danger Zone' mechanism, which allows users to access experimental features via specific API headers. These features are intended for 'dreamers'—users seeking creative exploration—rather than production-ready automation, as the company notes these capabilities still possess 'jaggedness' that prevents them from meeting the strict reliability standards required for enterprise deployment.
The image represents the Jev model's dual focus on production reliability and creative experimentation.
One of the primary additions in this release is composable image support. While the current Jev model was already multimodal, the new update enables users to integrate images directly into the JSON interface within the `state`, `instructions`, or `criteria` fields using a find-and-replace method in assets. Additionally, the update introduces a new data type called `box`. Unlike previous types designed for simple 'smart-if' logic, the `box` type returns a single bounding box, serving as a preview of the company's exploration into capabilities beyond basic decision-making.
For users requiring more data, the update also unlocks extended context lengths. While TypeSafe AI intentionally limits production context to encourage more effective usage patterns, the 'Danger Zone' flag allows users to bypass these constraints. The company also introduced a new `jev-preview` version for subtle, multi-lingual improvements and an opt-in system for users to help identify model 'jaggedness' without their queries being used for training. Despite these advancements, the company maintains that while these features are a glimpse into future reliable models, they should be tested thoroughly against specific use cases before deployment.