Jev 1.14 from TypeSafe AI is now available through Vercel's AI Gateway. Jev is a decision model that takes state plus a set of questions and returns typed Choice, Score, and Boolean answers with probabilities rather than free-form text, answering all questions in one parallel request for classification, routing, scoring, and verification tasks. Version 1.14 adds image input support (passing files alongside text in the state) and an experimental long-context mode for oversized documents, logs, or conversation histories. It supports Zero Data Retention and No Training options per request, and usage shows up in Gateway logs, custom reporting, and budgets. Use the model name typesafe-ai/jev-1.14.0.
Questions this post answers
How do I pass an image to the Jev decision model on Vercel AI Gateway?
Include the image as a file part alongside text in the state array when calling the decide function from the ai SDK, specifying type 'file', the mediaType such as 'image/png', and the raw file data. The model name used is typesafe-ai/jev-1.14.0, and the image is evaluated together with typed questions like Boolean checks to produce an answer with a probability. Teams wiring image understanding into decision pipelines follow releases like this via daily.dev.
How do I enable long-context mode for the Jev 1.14 model on Vercel AI Gateway?
Set providerOptions.typesafe.longContext to true in the decide call. This experimental mode lets jev-1.14.0 process state that is too large for a standard request, such as full support logs, long documents, or entire conversation histories, while still returning typed Choice, Score, or Boolean answers with probabilities. Developers handling oversized logs or transcripts track gateway model updates like this on daily.dev.
What is a decision model and how does it differ from a typical text-generating LLM call?
A decision model like Jev takes a state input plus a set of defined questions and returns typed answers, Choice, Score, or Boolean, each with an associated probability, instead of free-form generated text. It answers every question in a single parallel request, making it suited for classification, routing, scoring, and automated verification tasks rather than open-ended generation. Engineers choosing between generative and structured model calls compare approaches like this via daily.dev.
Share this post