Jev 1.14 is a big improvement, but it still has a limit of 255 classes.
What if you need to classify across more than 255 classes, such as time zones?
TypeLLM supports up to 512 classes natively.
Try it now: typellm.ai/cookbook/class…
What if you need to classify across more than 255 classes, such as time zones?
TypeLLM supports up to 512 classes natively.
Try it now: typellm.ai/cookbook/class…
Oct 15, 2026 · 3:47 AM UTC
2
1
13
1,533
@TypeLLM We actually went deeper on the new Jev features here: x.com/AGTPinsights/s…
TypeSafe AI just shipped Jev 1.14.0 for JevDay, adding image inputs and longer context. Here's what you need to know.
The news was teased on X with a post asking "what's this?" and a town hall with founder Diogo Almeida and Allie at 1:14 PM PST. The post got 44.2K views. Almeida JevDay. He says the first Jev release favored extremely stable infrastructure. This one is for early adopters, tinkerers and dreamers.
Images are the big change. Almeida says Jev was multimodal all along. It was an image model the whole time. TypeSafe disabled images because models tend to be less intelligent on other modalities. You can turn them on with the header TypeSafe-Danger-Zone-Images: true.
Context is longer too. Jev already supported far longer contexts than TypeSafe allowed. The cap was a choice, not a limit. TypeSafe thinks the best products shouldn't need it and should manage state more carefully. Long context is also behind an experimental header, TypeSafe-Danger-Zone-Long-Context: true.
There's a new output type called box. It returns a single bounding box. Almeida calls it a gray area between decisions, a reminder of why they're called System One Models. It's only on jev-preview.
The new model is jev-1.14.0, shipped as jev-preview. It's better overall and should mostly be a drop-in replacement. It's especially better at multilingual support. jev-1.13.0 stays available so production apps don't break, and it isn't deprecated yet.
There's also a new opt-in flag. It lets users share data to help TypeSafe find model jaggedness. TypeSafe says it still won't train on your data.
Developers are already reacting. One says Jev can answer hundreds of questions in parallel in a single call, since output is free. On a 32k token document, 100 questions cost about $0.00001 and take about 1ms each. Another developer likes Jev for websites, pointing to a new Expand Fetch product that uses System One vision models including Iris.
Key numbers:
- Model: jev-1.14.0 (jev-preview)
- Old model kept: jev-1.13.0
- Town hall: 1:14 PM PST, on Discord
- Base price at launch: $0.042 per million input tokens, output free
- Launch context cap: 32K tokens
- Reported cost: ~$0.00001 and ~1ms per question on a 32k document
Image input, longer context and the box type are all experimental for now.
0
0
0
54