Developers are trying it in place of LLM calls for classification, from command-safety checks to business-email sorting, because it runs cheaper and faster. An open-source imitation followed two days after launch.
TypeSafe AI opened Jev in early access on 15 September. It is built on the same transformer architecture as large language models, but it is not one and writes no text. A developer sends a question and the answers it may choose from. Jev puts a probability on each. TypeSafe, founded two years ago by a former OpenAI researcher who helped build ChatGPT, calls the result a calibrated decision.
TechCrunch, which reported the first deployments, found developers treating it as a sturdier way to build judgment into software. Input costs $0.042 per million tokens, output is free, and TypeSafe says a decision comes back in 70 to 500 milliseconds. Demand briefly overwhelmed the API.
At Vercel, Jev replaced OpenAI's Luna 5.6 in the classifier that screens commands for safety, and an engineer there said answers came back 5 to 18 times faster and were more accurate. Another company's CTO ran it against Gemini on sorting business email: Gemini was slightly more accurate and cost 10 to 20 times as much.
Armin Ronacher, who builds open-source AI tooling, told TechCrunch that acting on a probability moves part of the hallucination problem to the user. A 50% answer may be a coin toss to discard, 95% something to rely on. He expects the same speed and price to open up jobs that are expensive to hand an LLM, such as deciding in real time which model a task should go to.
Two days after the launch, Jared Palmer published Kev, an open-source model built on Qwen that does the same job, and larger versions followed on 20 September. It keeps Jev's API, so TypeSafe's own SDK can be pointed at a locally run copy. Palmer measured his two largest models six to seven percentage points below Jev in accuracy on data from sources they had not seen, and says their probabilities are not well calibrated there.