We’re releasing Jev 1.14.0 along with experimental features for dreamers

You can now experiment with image inputs and longer context windows in the Jev AI model.

TypeSafe AI has released Jev 1.14.0, introducing experimental features designed for creative exploration, accessible through opt-in flags in API headers. These "danger zone" features include support for composable images, a new "box" data type for bounding box output, and the ability to utilize significantly longer context lengths. While these functionalities are considered more advanced than current production-ready offerings, TypeSafe AI states they represent future capabilities and are actively being improved. The company also noted that models, including Jev, generally perform worse with multimodality and longer contexts. The release of Jev 1.14.0 under the `jev-preview` tag includes subtle improvements, particularly in multilingual support, with `jev-latest` expected to update to 1.14.0 in one week, and `jev-1.13.0` will be deprecated with advance notice. Additionally, users can now opt-in to share query data to aid in finding and debugging model issues.

The release of experimental features like image composability and extended context lengths signifies TypeSafe AI's strategy to balance production-grade reliability with fostering developer creativity. By labeling these advanced capabilities as "danger zone" features, the company allows early adopters to push boundaries and provide feedback, thereby accelerating development towards future stable releases. This approach acknowledges that innovation often emerges from exploring less-tested functionalities, especially in the rapidly evolving multimodal AI landscape.

The introduction of the "box" data type suggests a broader exploration beyond traditional decision-making outputs, hinting at System One Models' potential to support more diverse AI applications. TypeSafe AI's concurrent focus on subtle, incremental improvements in Jev 1.14.0, alongside its unconventional stance on public benchmarks, indicates a pragmatic approach to AI development centered on developer utility and internal validation for real-world deployment.