Jev 1.14.0 is here, along with a set of experimental features aimed at developers who want to push boundaries. The team describes a long-running debate between reliability and creativity. Rather than impose its extreme reliability standards on every user, it is sharing dreamer-ready features behind "danger zone" flags in its API headers. These capabilities are expected to remain in future models and are, by the company's account, better than existing alternatives. However, the team warns that there is still too much jaggedness to vouch for them in production use cases. The features are framed as a continuation of two years of stealth development, and are "less ready for automation, but very fun to play with."
The first feature is composable images. Jev 1.13.0 was already multimodal, and image support can now be added anywhere the JSON interface accepts input, including state, instructions, or criteria, via a find-and-replace in assets. The company cautions that all models, Jev included, perform worse on multimodal tasks than on pure text, and recommends breaking questions into the simplest possible decompositions. A new box type returns a single bounding box with no type errors allowed. It is the first type not purely intended for smart-if statements and is enabled only for jev-1.14.0.
The release also lifts the production context-length limit through the Type Safe-Danger-Zone-Long-Context flag. The company intentionally capped context because Jev, like other models, is less reliable on long inputs, and it still strongly recommends limiting length. An opt-in lets users help identify model weaknesses through their queries, though the company says it will not train on those queries.
A new preview version, jev-preview, offers subtle improvements, with the largest gains expected in multilingual support. jev-latest will move to jev-1.14.0 in a week. jev-1.13.0 remains available for now, with 60 days' notice before deprecation and 30 days for jev-1.14.0.
Source: Hacker News · Summarized by HeadlinesBriefing