Mario Zechner calls TypeSafe's Jev a general purpose classifier
Same-day impressions, one price tag, and no evals yet.
TypeSafe's Jev landed, and the engineers who got to it first are not describing it as another chat model. Mario Zechner called it "basically a general purpose classifier" and said he is curious what the training set looks like. Thorsten Ball praised the onboarding, saying the launch knew it had to overcome the question "wait, so it's not an LLM?" and threw everything at it. Malte Ubl said it is now on AI Gateway.
Armin Ronacher played with it before bed and came away impressed, with the specific claim that it fits applications where traditional LLMs were not viable on either speed or cost. He expects fast followers.
The price tag
Wes Bos left a preview browser open overnight and watched Jev burn through his credits at $0.000232 a call. That is the only concrete number anyone in these posts has put on it, and it arrived as a complaint about a runaway loop rather than as a benchmark.
What nobody has yet
No evals. What is circulating is same-day impressions from people who just got access, plus one per-call price. The interesting claim is about the small classify-and-route work inside pipelines, not about reasoning, and that is a category where cost per item and latency decide whether you make the call at all or batch it and hope.
Zechner is still asking what the data regime is. In a follow-up he said he did not expect 100% synthetic data and suspects nobody will ever see a high-level description of it, while adding that he is really excited about Jev.
Ronacher, meanwhile, has moved on to the part that actually bites. He said he is "getting Clanker psychosis" trying to work out how to build a sensible API around Jev like models. That is the honest state of things. A model shape that is not an LLM does not slot into the request and response conventions everyone has spent two years building around chat completions, and the first people through the door are still figuring out what the wrapper should look like.
For working engineers, treat the enthusiasm as what it is. Four people liked it on day one, one of them measured a bill, and none of them has published a comparison against the classifier you already run. If you have a pipeline where you skip a classification step because a chat model call costs too much or takes too long, that is the case worth testing yourself.
jev is basically a general purpose classifier. and i think that is amazing.
Left a preview browser open last night and jev burned though all my credits. $0.000232 at a time
I'm getting Clanker psychosis trying to figure out how to create a sensible API around Jev like models.