Jev is the first model in a long while that I wanted in production the same week I tested it. It is also the first one I have to explain by what it refuses to do, because it is different.
It is not your normal llm. It does not chat. It does not write. It does not reason. What it does is something different - it judges, scores and categorizes and it does so at a fraction of the cost and incredible speed. It hands back typed values: a choice from a list, a score on a scale, or a yes/no probability. Every answer carries a confidence number. That's all.
Diogo Almeida founded TypeSafe AI with Erik Gafni and Sasha Sheng, and they came out of stealth on 15 September with $40 million led by DCVC. Almeida is ex-OpenAI, one of the people behind RLHF and InstructGPT. So it is kinda cool that they are taking this entire industry into a different direction.
I have long been a proponent of the idea that specialized tiny models at a fraction of the cost should become the industry norm. I don't always want a reasoning model arguing for minutes with itself, sometimes I just want a confident answer and Jev promises this.
Where it is genuinely fantastic
So I ran a test and replaced some of the core routines in various of my projects. Jev delivered whenever I asked it to do micro-decisions, routing, and performed as a real judge. It did better than Mistral-small or similar models, it was faster, more accurate.
The case I benchmarked hardest on was an intent gate. I often add these to my projects - they check what the user "wants to do" and then reroute accordingly to a proper model with a selection of tools. It helps keep things slim (prompts included) and reduces tokens. And I often use it to check if the user is sneakily trying to do something I don't want them to - ie, misuse my software for something.
I had close to 70 cases, and reran the script three times. Jev got every single one right in all three rounds. The median was 250 ms and on average it ran between 453 and 688 ms. Out of every model I put in that bench (I tried a few) it was the only one that never once blew the 1.5 second budget. That was what I had hoped for - accurate performance for something other models would take longer for.
The second case was a judge with no incumbent at all. One yes/no question on a slide deck, 40 slides, one call, 1.41 seconds. It caught claims in that deck that I had read past myself - Jev spotted them accurately.
Where it does not fit
I cannot talk about Jev without talking about the downsides. Like I said in the intro, Jev is not your normal llm and you should not treat it as such. It is also incapable of handling many use cases, such as:
No images. The input is always text, objects or arrays.
No prose output, which sounds obvious until you go through your own candidate list and find how many of them need one human-readable sentence at the end.
It is not OpenAI compatible, either. The api endpoint is different at the moment. Though I bet that'll get solved quickly, for now it means that you cannot swap it in as a model name in a config file, so every single use is new code. That is a real cost and it is worth knowing before you get excited.
The business model is unqiue
Lastly, I should talk about the business model. Jev costs only $0.042 per million input tokens and the Output is free, which they describe as too cheap to meter.
They named the model after William Stanley Jevons, which is not subtle and not meant to be. Jevons paradox says that making a resource cheaper raises total consumption rather than lowering it. So the pricing is not a launch discount, it is the thesis: at four cents a million tokens, you may start wondering which "micro decision" you can now actually route to an llm, when it was simply too costly before.
The transactions are tiny and so are the costs. They are in it for the long run rather than for this quarter and it gets me excited to play around with Jev some more and see if I can come up with new business models, too.
Conclusion if there is one
I think that Jev is a fantastic addition to the model space. It ventures into a region that opens up new possibilities for all of us and it gets me excited over having a new lever to pull during my implementations. Kudos to the team!
Top comments (1)
It's impressive that Jev felt production‑ready right out of the private beta, especially given how rare that immediate confidence is with new models. Could you share which evaluation metrics or latency benchmarks convinced you it was safe to deploy that quickly?