Jev
TypeSafe's Jev, a decision model that returns probabilities for yes/no, multiple-choice or score questions.
HN commenters like Jev for fast, cheap yes/no and multiple-choice decisions123 such as classification, spam filtering and picking tools or page elements for agents.456 The most common complaint is that its accuracy is only average next to purpose-built classifiers789, and that it struggles with novel domains and unfamiliar tasks.1011
Why people recommend it
Low cost · 3
Many users report large cost savings over the LLMs they used before.
“It really does just work. And it works so well I already integrated it into my product.”
“I used Jev to make my AI game 6x faster and 8.7x cheaper.”
Show 1 more
“This is a great use case for such a model.”
Speed · 3
Decisions come back in well under a second, much faster than general LLMs in users' tests.
“It's fast and cheap: about 0.35 s per decision, about 2 cents per hour of audio.”
“Jev averaged 227 ms (p95 400 ms). Gemini 3.2 s, Haiku 2.5 s, GPT-5.6 Sol 3.5 s.”
Show 1 more
“Hosted p50: Clef ~850ms, Jev ~110ms”
Classification quality · 3
Users find it roughly matches LLMs for filtering and classification in their own tests.
“It's loads cheaper and faster with roughly equivalent performance on my own internal tests.”
“I used the model and it was good enough for email classification.”
Show 1 more
“Cost, latency and accuracy were all near the top for Jev.”
Agent decisions · 3
It works well for choosing tools or page elements in agent and browser workflows.
“We use Jev to decide which subset of tools to send to the LLM.”
“21–23 decisions for six benchmark cards, all correct (!), ~$0.001 total. Really good.”
Show 1 more
“act: 4.3x faster, 97% fewer LLM calls. Pass rate: 97.5% -> 98.3%.”
Calibrated confidence · 2
Some users find its confidence scores reasonably well calibrated.
“It's loads cheaper and faster with roughly equivalent performance on my own internal tests.”
“it performs on par with small LLMs while being faster and cheaper”
Production-ready · 2
Several users say it works reliably enough to ship in production.
“I use Jev in production. Hype or no hype it’s working.”
“It really does just work. And it works so well I already integrated it into my product.”
Caveats and dissent
Average accuracy · 3
Some find purpose-built or simple trained classifiers beat it, sometimes by a wide margin.
“Jev is the general purpose ML pipeline for when you want average results.”
“but with 40.7% accuracy it's basically useless for this use case.”
Show 1 more
“I just know that I can't think of when I'd reach for Jev 1.*[4] for a prod usecase”
Novel domains · 2
Users report it fails to understand unfamiliar domains and to make connections in multi-step tasks.
“jev has failed to understand any novel domain I've given it”
“jev as it is now is really bad at making connections and understanding it's input”
Confidence pitfalls · 3
It can pick a wrong option with high confidence, and its scores may be uncalibrated for niche tasks.
“the absence of essential options can cause jev to choose the wrong option with high confidence”
“Jev, on the other hand, has a generic model, not trained on in-domain examples, so its confidence scores are uncalibrated for any non-generic task.”
Show 1 more
“It undermines any confidence I have in the decision models (at least Official Jev)”
No text output · 3
It cannot write free-form text, so another model is needed for that.
“21–23 decisions for six benchmark cards, all correct (!), ~$0.001 total. Really good.”
“result: in 6 cases (button, form, dropdown, read table) = 3 secs and 0,02 Cent.”
Show 1 more
“This sparked an idea for me to use Jev in my recently launched Panel Review tool”
Not novel · 1
Some commenters see it as BERT-style classification with more data rather than a breakthrough.
“it’s just BERT with more data”
Rate limits · 1
One user needs higher rate limits for heavy concurrent use.
“Main thing I need from Jev is high rate limits.”
Often compared with
Who recommends it in their own words
From the makers not counted in the ranking
“Jev is weak at some kinds of tasks, so you should check the calibration for your own use case”
“Jev always chose face 1 and the probability it returned was about 83%”
Threads where it came up
- We’re releasing Jev 1.14.0 along with experimental features for dreamers 5 mentions
- Typesafe AI raises $870M at $7.5B 3 mentions
- I built non-autoregressive decision models with RL a year ago 3 mentions
- The rise and fall of Jev hype 2 mentions
- Decisions API is in public beta 2 mentions
- Jev Does Not Play Dice: 83% probability, 19% accuracy on a hidden fair die roll 2 mentions
Also written as JEV. Data: jev.json. Quotes are verbatim; the verdict and theme labels are written by a model from these comments, and every clause links to its sources.
All comments 41
“I use jev, but it does not solve all my problems.”
“I use Jev in production. Hype or no hype it’s working.”
“Cost, latency and accuracy were all near the top for Jev.”
“how would Jev and other decision models handle the Trolley Problem?”
“Main thing I need from Jev is high rate limits.”
“And i already sub to jev and i think this launch is going to make me a much more loyal user now lol”
“I just know that I can't think of when I'd reach for Jev 1.*[4] for a prod usecase”
“We use Jev to decide which subset of tools to send to the LLM.”
“It's loads cheaper and faster with roughly equivalent performance on my own internal tests.”
“I used the model and it was good enough for email classification.”
“I use jev to classify 'has statement' and 'can be a question' paragraphs/chunks taken from cli script to reduce token usage.”
“They moved from gemini 2.5 flash lite type models to jev.”
“my guess is that the primitive Jev provides is likely to be considered essential in the future development of software.”
“Jev is the general purpose ML pipeline for when you want average results.”
“It undermines any confidence I have in the decision models (at least Official Jev)”
“I'd argue the values provided by Jev were far more representative of my subjective assessment over Lunas”
“jev has failed to understand any novel domain I've given it”
“This is a great use case for such a model.”
“I do find it disappointing that Jev conflates “the probabilities are all equal” with “I have no clue””
“Hosted p50: Clef ~850ms, Jev ~110ms”
“This sparked an idea for me to use Jev in my recently launched Panel Review tool”
“I use Jev for my app (Greer - open source tool to build a community from 0), to sort things fast and cheap”
“result: in 6 cases (button, form, dropdown, read table) = 3 secs and 0,02 Cent.”
“On multi-step reasoning it's behind: BBH 64-68% against Jev's 94%”
“it performs on par with small LLMs while being faster and cheaper”
“Is Jev just that with nice caching?”
“It really does just work. And it works so well I already integrated it into my product.”
“Jev is a great model, unblocks things you could not do before when you're in builder mode”
“Jev, on the other hand, has a generic model, not trained on in-domain examples, so its confidence scores are uncalibrated for any non-generic task.”
“It's fast and cheap: about 0.35 s per decision, about 2 cents per hour of audio.”
“Zero-shot Laya tracks Jev on 2-4 labels and falls off a cliff on 77”
“Now that Jev is out I can finally have a true AI sysadmins managing my "cloud in the basement" homelab”
“the absence of essential options can cause jev to choose the wrong option with high confidence”
“I used Jev to make my AI game 6x faster and 8.7x cheaper.”
“I use my Jev BYOK”
“jev as it is now is really bad at making connections and understanding it's input”
“I use it mainly to check whether this code fits the rules I defined”
“but with 40.7% accuracy it's basically useless for this use case.”
“Jev seems pretty cool! I just got access and have only gotten to do minimal experiments”
“act: 4.3x faster, 97% fewer LLM calls. Pass rate: 97.5% -> 98.3%.”
“21–23 decisions for six benchmark cards, all correct (!), ~$0.001 total. Really good.”