Jev is the new System One model from Typesafe.AI. Unlike the chat models we’re used to, Jev only responds in structured data. You can’t have a conversation with it. But you can ask it a question and get back a very fast, typed answer. Think JSON.
What it actually does
You send Jev a chunk of context (text or JSON) plus a list of questions. Each question is yes/no, pick-one-of-N, or a score on a scale. Each answer comes back with a probability and a confidence value. Ask “Is this a refund request?” and you might get 0.93. Route a support ticket and you might get billing at 84%, technical at 15%. Your code decides what happens next: act automatically above a threshold, escalate to a human below it. One early tester called it “a smart if statement,” which is about right.
Why it matters
Much of what companies do with LLMs today is classification, routing, and scoring. That means paying for prose you throw away and then parsing it. Jev skips the prose. It’s non-autoregressive, so it answers all questions in parallel instead of token by token. Typesafe claims 70 to 500 ms end-to-end, $0.042 per million input tokens with free output, and, in its own workflow benchmarks, results 193.6x faster and 444.6x cheaper than frontier LLMs. Because outputs are typed, schema errors can’t happen.
The takeaway
Jev isn’t a replacement for a chat model. It’s a fast, cheap decision layer that sits next to one. For voice AI, where every 100 ms of latency is felt, that’s worth watching: intent routing, call disposition, and escalation calls are exactly this kind of work.
Click below to see the actual Jev UI asking if a hotdog is a sandwich:
https://console.typesafe.ai/playground?share=shr_1574edaebbb9a4849b389fd631923c2af43