TypeSafe AI Releases Jev 1.14.0 with Experimental Multimodal and Long-Context Features

TypeSafe AI has released Jev 1.14.0, introducing experimental features designed for creative exploration rather than production stability. The company distinguishes between its battle-tested models for enterprise reliability and these new capabilities, which are gated behind specific API headers to warn users of potential "jaggedness."

Experimental Features

The release includes three primary experimental capabilities, each requiring explicit opt-in via API headers:

• Composable Images: Multimodal support is now available for the current jev-1.13.0 and new jev-1.14.0 models. Images can be inserted into state, instructions, or criteria fields. The company notes that multimodal performance is generally lower than pure text and recommends decomposing complex questions.
  • Header: TypeSafe-Danger-Zone-Images: true
• New box Type: A new return type that outputs a single bounding box. This is the first type not intended for simple conditional logic, signaling a move toward broader "System One Model" capabilities. It is exclusive to jev-1.14.0.
• Longer Context: Users can now enable longer context windows, which were previously limited to ensure reliability. The company advises caution, noting that reliability decreases with context length.
  • Header: TypeSafe-Danger-Zone-Long-Context: true

Model Updates and Deprecation

The new model, jev-1.14.0, is available under the jev-preview tag. It focuses on subtle, broad improvements, with the most significant gains expected in multilingual support.

• jev-latest will automatically update to jev-1.14.0 in one week.
• jev-1.13.0 will remain available for a period, with a formal deprecation warning expected 60 days prior to removal.
• jev-1.14.0 will have a 30-day deprecation warning window.

Internal Metrics and Philosophy

TypeSafe AI shared internal intelligence measurements, explicitly stating these are "best guesses" and not public benchmarks. The data includes only open-weight models to avoid data policy conflicts with other providers. The company emphasizes that these metrics are biased toward their own models and encourages developers to rely on their own evaluations. Future releases are expected to focus on stability and "boring" reliability improvements rather than flashy new capabilities.