Vercel's Python SDK adds evaluate() for Jev

Jev comes out of an older idea: the classifier, the classic primitive of machine learning. Traditionally you would train a classifier from scratch for each new task — collect labeled emails for spam filtering, for example, and tune the model's weights to that domain. The team behind Jev instead turned a large language model into a classifier, so it already carries broad human knowledge and does not need task-specific training. That is why the model is cheap and fast, and why it returns structured JSON that conforms to the types and choices you define rather than free-flowing text.

Vercel's Python team has wrapped Jev's API almost verbatim into the AI SDK, keeping the surface small — one evaluate() function plus a few supporting types. The blog walks through two hands-on tests. In the first, a Vercel engineer revisited a Python REPL feature that needed to tell English from Python; a hand-rolled classifier trained over two days felt jittery, and while Jev fared better, it still labeled the half-typed expression "what's" + " up as English. In the second, the engineer tried to make Jev write Python code by picking characters, then abstract syntax tree nodes; it produced syntactically valid but mostly incorrect code, and only after an LLM expanded the prompt into a detailed plan.

The takeaway is narrow but useful: Jev is built for tight decisions, not for general writing. Whether it is accurate enough for a given use case is something teams will have to test themselves — and the two examples suggest the answer depends heavily on how well the question can be pinned down to concrete choices.