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

We're excited to announce the release of Jev 1.14.0, along with some experimental features specifically designed for our "dreamers" – those who like to push the boundaries of technology and explore new possibilities. These features include image support, a new type, and longer context lengths. However, please be aware that these experimental features are still in the "danger zone" and should be used with caution in production environments.

Our team has been debating the balance between reliability and creativity. While we aim to ship models that meet our quality bar for intelligence and have been battle-tested, we also recognize the need to spark delight and build something exciting for the world. To address this, we've shared dreamer-ready features under "danger zone" flags in our API headers.

These features are here to stay for future models and are expected to improve significantly over time. But remember, we're not entirely confident in their performance for production use cases yet.

One of the key changes in Jev 1.14.0 is the introduction of image support. This feature is fully composable via a find-and-replace in assets, allowing you to put images anywhere that our JSON interface supports, such as in the state, instructions, or criteria. We're also introducing a new type called "box," which returns a single bounding box with no type errors. This type is our first step into exploring beyond simple decisions and could pave the way for future advancements in System One Models.

We've also listened to your feedback regarding context lengths. While we intentionally limited the context length to ensure better reliability in production, Jev now supports a much longer context length than we've allowed before. This choice encourages more effective usage patterns in production environments. However, if you're willing to experiment, you can now check out the full docs to see the updated context length guidelines.

We value your contributions to improving our models and have added an opt-in feature to help you find model jaggedness with your queries. While we won't train on your queries, this opt-in can help us identify issues and debug your use cases more efficiently. Additionally, we're releasing a new version of Jev under the jev-preview tag, focusing on small, subtle improvements across the board. Most of these enhancements aim to provide better multilingual support in the upcoming release.