Meet Xev β Jev-compatible decisions, served from Europe π―
I did hint about this in the last announcement π
Here's what's going on.
Jev, from TypeSafe, made a new kind of model popular. You don't ask it to write text, you ask it to decide things for you.
- Yes or no.
- Which of these options.
- Where on this scale.
And you get a probability back instead of a paragraph you have to parse, read and work with (take that claude! with your annoying walls of text..).
Today we're launching our own: Xev.
- Same format as Jev. Xev speaks the System One API, so if you already use Jev you change the URL and the key, and that's it.
- Served from Europe. It runs on our own infrastrukture in Stockholm, with zero prompt retention.
- β¬0.039 per million input tokens. Output is free.
- Up to 32 questions per call about the same data and the data is only read once.
Xev is powered by Qwen3.8 27B on our endpoint, so there's a 27-billion-parameter model behind every decision.
In our first measurements on public benchmarks, 150 questions each, Xev scores 87% on yes/no questions about a text (BoolQ), 87% on topic classification (AG News) and 84% on general-knowledge questions (MMLU). With the confidence threshold set at 80%, it still decides two-thirds to four-fifths of the questions and gets 91β96% of those right. The rest come back marked undecided instead of guessed. It costs β¬0.039 per million input tokens, output is free, and it runs in Sweden with nothing stored. A full benchmark is coming.
So it's not perfect, but it's a new tool for everyone to make use of, and i'm happy we can offer a first beta version of it.
What would you use it for? Routing agents, sorting support tickets, checking another model's answer before it goes out, scoring leads, branching workflows. We're already using it ourselves β the assistant in the console now asks Xev what you want before it answers.
It's live now. Grab a key in the console, read the docs, or look at Xev and Jev side by side.
Less prompting. More deciding. Let's see what you build with it.