Jev: For the builders AND dreamers
typesafe.ai ● Covered by 2 sources
Jev 1.14.0 adds images, a new box type, and longer context behind “danger zone” flags. It’s powerful, but the company says it’s still too rough for production.
Based on reporting by typesafe.ai — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
TL;DR is shipping an updated Jev 1.14.0, plus a set of experimental features it clearly does not want people to confuse with production-ready software. The new stuff lives behind “danger zone” flags in API headers: images, a new type called box, and longer context lengths. The pitch is simple enough. Builders get the safer path; dreamers get access to the sharp edges.
The image support is the most visible change. Jev’s current model, jev-1.13.0, was already multimodal, even if that wasn’t obvious at first glance. The company says image support is composable through a find-and-replace in assets, and can be placed anywhere the JSON interface allows, including state, instructions, or criteria. But it also warns that every model, Jev included, performs worse on multimodality than on pure text.
Then there’s box, which returns a single bounding box and is described as the first type that is not just for smart if statements. TL;DR frames it as a preview of work beyond simple decisions, tied to what it calls System One Models. This one is only enabled for jev-1.14.0, which seems to be the company’s way of saying it has seen enough internal progress to let people poke at it without pretending it is finished.
Context length gets the same treatment. TL;DR says Jev can handle more context than it currently exposes in production, but it limited that on purpose because longer contexts are less reliable and can encourage better usage patterns. Now users can override that with TypeSafe-Danger-Zone-Long-Context: true, though the company still recommends keeping it short.
Alongside the features, TL;DR is also opening an opt-in to help find jaggedness in queries. It says it will not train on those queries, only use them to surface issues and debug use cases. And on the release side, jev-preview now points to jev-1.14.0, jev-latest will move there in a week, and 1.13.0 will stick around for a while. The company says it will give 60 days warning before 1.13.0 is formally deprecated, and 30 days for 1.14.0, with a deprecation policy still coming.
There’s one more twist: TL;DR says it will share its own internal way of measuring intelligence, while admitting the metric will evolve and may not stay consistent. It also says the ranking only covers open weight models. The closing message is pretty clear. Expect fewer fireworks and more boring, steady improvements — which, in AI, is almost rebellious.
My take — AI-written commentary, not fact-checked reporting
This is the right kind of annoying: useful features, clear warning labels, and no fake macho talk about “shipping to prod” when the edges are still jagged. The best AI teams are starting to act less like hype machines and more like airline manufacturers, which is a relief. If only every model vendor would admit that “danger zone” is sometimes the most honest product name in the room.
Read more about this at: typesafe.ai