Jev 1.14 can read images. Should you care?

On 14 October 2026, TypeSafe AI published a blog post titled "Jev: For The Builders AND Dreamers". It announces a new version, Jev 1.14.0, and three experimental features. TypeSafe's models page says 1.14 has been available since 14 October.

The company says it wants to keep the main product reliable for people who run it in production, and to let others play with newer capabilities behind what it calls "danger zone" settings. Those are extra headers you add to an API request. TypeSafe says these features will stay in future models, and also that there is "way too much jaggedness" for it to vouch for them in production.

Jev 1.14.0 is available now under the name jev-preview. TypeSafe describes it as small, subtle improvements to everything, with the largest gain in languages other than English. It does not publish numbers for that here.

The name jev-latest still points to Jev 1.13.0. TypeSafe says it will move to 1.14.0 "in a week", and that 1.13.0 will stay available "for a while". It also says it will give 60 days' warning before 1.13.0 is formally deprecated and 30 days for 1.14.0, and will publish a deprecation policy first. If you want results that never change under you, pin a version such as jev-1.13.0 instead of using jev-latest.

Each one is switched on with its own header, so nothing changes unless you ask for it:

• Images. TypeSafe says image support works on the current model too: "the Jev you all know and love (jev-1.13.0) was multimodal all along". You put an image anywhere the JSON interface takes text: in the state, the instructions or the criteria. The header is TypeSafe-Danger-Zone-Images: true. TypeSafe warns that every model, Jev included, does worse on images than on text, and that splitting questions into the simplest possible pieces helps a lot.
• A new answer type, box. It returns a single bounding box, which is a rectangle around something in an image. TypeSafe calls it its first type that is not purely for if-statements, and enables it only on jev-1.14.0.
• Longer context. Jev can accept more text than TypeSafe lets it by default. The header TypeSafe-Danger-Zone-Long-Context: true raises the limits to 262,000 tokens, up from 64,000 per request and 32,000 per question. TypeSafe still recommends keeping context short, because models are less reliable on long inputs.

TypeSafe added an option that lets you share your queries to help it find weak spots. It says it still will not train on them.

The post also shows an internal chart of how it measures intelligence. TypeSafe says the chart includes only open-weight models, is probably biased towards its own models, and should not be believed entirely. We have not reproduced it and do not quote its numbers here.

The price is the same: $0.042 per million input tokens, with free output, on both versions, according to TypeSafe's models page. Audio and video are still not accepted.

The models page we read on 15 October still lists the input as text. The blog post is newer in spirit, so treat images as an opt-in experiment that the docs have not caught up with yet, and check both before you build on it.

This is a tidy release for people already using Jev: a drop-in improvement, a clear promise about deprecation, and nothing forced on you. The image option is the headline, but TypeSafe itself calls it unfit for production, so for now it is something to try, not something to ship.

It does narrow one gap that we have pointed out in our comparisons: OpenAI's Decisions API accepts images as a normal feature, while Jev's image input is experimental. If you need to look at images in production, that difference still matters today.

• Jev vs OpenAI Decisions API
• What is Jev?
• How to use Jev