TypeSafe opened up image support for Jev and told developers in the same post that it works worse than text. “All models (Jev included) perform worse on multimodality over pure text,” the company wrote on its blog. Jev is the model TypeSafe built to make decisions inside software. The company says the version developers already use was capable of reading images all along, and it has now switched that ability on behind a flag.
The sentence deserves a careful reading. It covers all models, not just the two Jev versions that now accept images, and it comes with no figures. TypeSafe does not say how large the gap is, so developers have a warning but no measurement to size it against.
The accompanying advice is more telling: “Breaking down questions into the simplest possible decompositions helps a lot.” That is a usage instruction, and it implies where the trouble sits. A question that bundles several judgments about a picture is where the model is likely to stumble, so TypeSafe wants one small question at a time. The post does not state this outright; it is the inference the guidance supports.
Images work in both the current jev-1.13.0, which the company says “was multimodal all along,” and the new jev-1.14.0. Pictures can go into the criteria, the instructions or the state that Jev is given, and the swap happens with a find-and-replace handled by the assets. Switching it on takes an API header, TypeSafe-Danger-Zone-Images.
The second addition, a new output type called box, is for 1.14.0 only. Instead of describing what is in an image, the model returns a single bounding box, a rectangle marking where something sits in the frame. The post names no use case; an obvious one is flagging where a defect or a label sits in a photo. One box per answer means it locates one thing, and the post does not explain how to find several. TypeSafe also says the output will never come back malformed (“no type errors allowed”).
The box type is the strategic detail. The company calls it “our first type that’s not purely meant for smart-if statements,” and a preview of what else its “System One Models” could support. TypeSafe sells a model that picks from a list, and this is its first step toward doing more than that.
Longer context arrives behind a second flag. TypeSafe says Jev handles far more text than the production limit allows, and that it capped the limit on purpose because models grow less reliable as inputs lengthen. The post gives neither the current cap nor the new ceiling, so the headline question of how long goes unanswered.
The company is candid about why all of this sits behind “danger zone” flags: it sees “way too much jaggedness” to vouch for production use. The company draws its line at opt-in flags, so the experiments stay out of the default path for customers who depend on it.
The 1.14.0 model itself is a modest upgrade, with the largest gain expected in languages other than English. The jev-latest alias moves to it in a week. TypeSafe promises 60 days of warning before retiring 1.13.0 and 30 days for 1.14.0, with a deprecation policy to follow. Developers can also opt in to sharing queries to help find failures; the company says it will not train on them.
The post gives no pricing and no licence terms.
For anyone prototyping with Jev this month, the practical plan is to test images in small, single-question steps, and to expect jev-latest to change underneath them within a week.
Based on TypeSafe’s blog post “Jev: For the builders AND dreamers” (publication date not shown in the source text)
