TypeSafe Releases Jev 1.14.0 AI Model and Experimental Features for Developers
TypeSafe has released version 1.14.0 of its Jev AI model. Alongside it come experimental features: image support, a new "box" type and longer context length.
AI developer TypeSafe has announced the release of version 1.14.0 of its Jev language model. At the same time, the company has opened up experimental capabilities to developers through special API header flags it calls "danger zone": working with images, a new output type and longer context length. TypeSafe does not recommend using these features in production, since they still perform unevenly.
Why "danger zone"
TypeSafe explains that its team has debated the balance between reliability and creativity. Production builders are offered thoroughly tested models, while experimenters get features the company considers promising and expects to keep in future models, but which are not yet stable enough.
New capabilities
- Images. Support works with the current jev-1.13.0 model too, which the company reveals was multimodal all along. Images can be placed anywhere the JSON interface supports: in the state, instructions or criteria. TypeSafe warns that all models perform worse with images than with plain text.
- The "box" type. It returns a single bounding box and is the first type not intended purely for conditional decisions. It is available only for jev-1.14.0.
- Longer context. Jev can handle a longer context than is currently allowed in production; the company limited it deliberately because models become less reliable with longer contexts. The limit can now be lifted with a separate flag.
Versions and updates
Jev 1.14.0 is available as jev-preview. TypeSafe describes it as a set of small improvements, with the largest gain expected in multilingual support. The jev-latest alias will be updated to 1.14.0 in a week. Users will get 60 days' notice before jev-1.13.0 is formally deprecated and 30 days for jev-1.14.0; the company has promised to publish a deprecation policy beforehand.
The company has also added an opt-in that lets users share their queries to help find model issues. TypeSafe stresses that it still will not train on those queries.
Internal measurements
TypeSafe, which dislikes public benchmarks, shared its own internal intelligence measurements for the first time, including only open-weight models. It admits the results may be biased toward its own models and should not be taken at face value. For future releases, TypeSafe expects mostly small improvements, better stability and features for real deployment.
