Jev 1.14 adds images, boxes and 262k context at flat price

TypeSafe shipped Jev 1.14, announced by @typesafeai on 10.14 at 1.14pm PT, and the three headline features are all preview only. Image input works anywhere text did, there is a new bounding box primitive, and context length goes up to 262k, which the company describes as four times as much. Each one is gated behind an opt-in header, TypeSafe-Danger-Zone-Images and TypeSafe-Danger-Zone-Long-Context, which is TypeSafe's way of saying it does not consider them production grade.

OpenRouter listed the model the same day and summarized it as everything a tiny bit better, multilingual as the biggest upgrade, multimodal experimental, price unchanged at $0.042 per million input tokens with output free, and drama none.

OpenRouter

@OpenRouter

X

Price: unchanged ($0.042/M input, output free)

The model could already see

The more interesting detail came from Diogo Almeida, who said the capability was never added, it was switched off. He wrote that TypeSafe disabled it deliberately because models tend to be less intelligent on other modalities. The same applies to context length. Almeida said Jev already supported much longer contexts and the limit was a product opinion, that the best use cases should not need it and should do more thoughtful state management instead.

Diogo Almeida

@CompleteSkeptic

X

Yep, the Jev you know and love was an image model this whole time.

Diogo Almeida

@CompleteSkeptic

X

We just chose to disable it because models tend to be less intelligent on other modalities.

The bounding box type is the odd one out. Almeida put it in what he called the gray area of decisions, a single box rather than a classification, available only on jev-preview.

The eval is TypeSafe's own

There is no public benchmark attached to this release. Almeida, who has said repeatedly that he dislikes public benchmarks, called the intelligence numbers in the blog post an experiment in communicating TypeSafe's internal measure. He told readers not to treat it like gospel, said it is likely biased toward TypeSafe's models, and suggested paying attention only to the relative ranking of open models. Hosted models were excluded, he said, over what their providers do with submitted data and a desire to keep the validation set secure.

For people building on it, jev-preview now points to jev-1.14.0 and TypeSafe calls it mostly a drop-in replacement, but 1.13.0 stays available and is not yet deprecated, so production apps do not have to move. There is also a new opt-in flag for people who want to help TypeSafe find model jaggedness, with the company saying it still will not train on your data. The honest read today is that the multilingual gains are the only part of this release anyone outside TypeSafe has a reason to trust, and even those are unquantified.