Jev 1.14.0 can read images — if you open its 'Danger Zone'
TypeSafe's decision model learned to look at pictures this morning, but the company shipped the capability behind a warning label instead of a launch announcement.
Jev 1.14.0 adds image input, a new question type called Box, and a 262,000-token context window — all three shipped as experimental features behind a program TypeSafe calls the "Danger Zone." The features are off by default and have to be enabled per request through the API; TypeSafe's own documentation says they do not meet production standards and developers use them at their own risk. The model itself is a small upgrade — TypeSafe describes it as an unshowy release, broad but subtle, with the clearest expected gain in non-English languages. Chinese is explicitly still behind English, and the docs tell developers to test with their own content before shipping. Pricing is unchanged at $0.042 per million input tokens, with output free. The 1.14.0 release currently carries the preview alias and is expected to become the default within a week, while 1.13.0 stays available for a while longer.
Image support is the interesting part because of how it composes. TypeSafe says images can appear in the state information, in the question instruction, and even in the criteria Jev grades against — and all three question types accept them. The format is narrow: PNG or JPEG only, sides between 256 and 1,568 pixels, no more than 16 megabytes per image, billed as input tokens. TypeSafe also confirms that Jev 1.13.0 was already multimodal — the image path just was never opened to developers. Box is the genuinely new question type — it returns normalized bounding-box coordinates between 0 and 1 for a target in an image, so an agent can point at a cup on a counter, a signature on a scanned page, or a person in a doorway. Box is also the first Jev question type that does not exist to serve a yes/no judgment. The context bump runs from 64,000 tokens per request and 32,000 per question up to 262,000 for both, with the same caveat the field always gives: longer context tends to be slower and less accurate, so open it only when you need it.
Then there is the benchmark chart, which is where this release tells you more than the feature list does. TypeSafe's internal comparison puts Jev 1.14.0 first among eight models on a composite intelligence score at 60.0, ahead of Jev 1.13.0's 58.9, with the best open-weight model, kev-27b, at 46.7. Jev also posts the best calibration error at 1.6. The same document warns that the ranking covers open-weight models only, may favor TypeSafe's own model, and could be revised — and every chart carries the line "Warning: AI benchmarks, do not trust them lightly." A vendor printing that on its own marketing is unusual, and it is the right instinct for a product whose pitch is calibrated confidence. The whole category rests on the claim that a model can say how sure it is; the moment a lab starts treating its own scores as settled, the pitch collapses.
What matters beyond the feature list is the pace. TypeSafe exited stealth with Jev on September 16 after two years of quiet work, raised $870 million at a $7.5 billion valuation on October 10, and has now shipped its first update five days later. The category it created is crowded already: Microsoft put out Microsoft-Decision-1 built on an open-weight base, the vLLM community shipped six open decision models with vision, and Jev-shaped projects like Kev, Laya and Decider are multiplying — the wave we flagged in Open Source Radar — September 21: decision models go local. The Danger Zone framing is the company's answer to that pressure while keeping its production promise intact: give developers the capability before it is reliable, and make them sign for it.
What to watch: whether Box and image input graduate out of the experimental switch, and whether anyone outside TypeSafe reruns that 60.0.
Would you wire a "use at your own risk" model into an agent's decision path — and what would you test before you did? Tell us in the comments.
