# The Free Fake-Detector Setup
Open-Jev-2B is a free, open-source decision AI you run on your own Mac. You give it a
scam text, a review, a job offer, or a link, plus a short fixed list of possible answers, and
it picks one and hands back a confidence score, in well under half a second. It's an
independent open-source rebuild of TypeSafe AI's paid Jev ("System One") API, built by
someone else, not affiliated with TypeSafe AI, and not claiming to match it.
Code is MIT-licensed (github.com/Zefan-Cai/Open-Jev), weights are CC0
(ZefanCai/Open-Jev-2B on Hugging Face): no account, no key, no bill, at any point. It sat
at 246 stars as of 2026-09-23, two days after TypeSafe AI's own Jev went public.
What it actually is
It is not a chatbot. It's a small, narrow decision model: the base is Qwen/Qwen3.5-2B, a real
~2 billion-parameter model, fine-tuned for exactly one shape of task. Here's a situation,
here's a short list of fixed answers, pick one and say how sure you are. No open-ended
conversation, no long reply to read. Just a verdict and a number.
We downloaded it, ran it on an Apple M5 Pro over the mps backend, and put it through five
small test sets of our own: scam texts, fake reviews, job offers, risky links, and general
inbox triage, to see what it actually gets right, and what it doesn't.
What we measured
| Task | Right | What it missed |
|---|---|---|
| Scam texts | 6 / 6 | none |
| Fake reviews | 7 / 8 | a keyword-stuffed blender review, called it genuine |
| Job offers | 7 / 8 | "You're hired, send your bank login," called it real |
| Risky links | 6 / 8 | a fake FedEx delivery link, called it safe at 76% confidence |
| Inbox triage (4-way) | 4 / 12 | defaulted to "reply-today" on every one of the 8 misses |
These are small hand-written test sets (six to twelve items each, written by us), not a
public benchmark. Read them as "on our items, this once," not a guaranteed hit rate.
Speed held up cleanly. Across every task except inbox triage (49 decisions in total),
every response came back between 109 ms and 512 ms on the M5 Pro. Under half a second, every
single time, right or wrong.
Where it's strong, and where it isn't
It caught every scam text we threw at it, and three of the four fake reviews: the ones with
obvious five-star-spam phrasing ("AMAZING PRODUCT!!! BEST THING I EVER BOUGHT!!!"). What it
missed was a fake review with no exclamation points and no punctuation discipline at all,
just keyword after keyword, a subtler fake than tone alone catches.
The job offer it missed is the one that would have actually cost someone money: "You're
hired! To set up payroll direct deposit, we just need your bank login username and
password." It called that real, at a near-coin-flip 56% confidence, its least confident
call in the whole set, which is at least an honest tell if you're watching for it.
The link it missed is worse, because it was confident about being wrong: a text reading
"Your FedEx package is delayed. Update delivery preferences here," pointing at a fake
tracking domain, called safe at 76% confidence, its most self-assured wrong answer of the
whole run. It also over-corrected once the other way: a genuine bank one-time-passcode text
with no link at all got flagged "risky." A false alarm costs you nothing; a confident false
"safe" on a real phishing link is the one that would.
Inbox triage was a real failure, not a near-miss. Handed four choices (urgent,
reply-today, later, junk-scam) instead of two, it got 4 of 12 right, and on every one of
the 8 wrong answers, it landed on "reply-today," including a service-disconnection notice it
should have flagged urgent and a phishing email it should have flagged junk-scam. That's a
model defaulting to one label under uncertainty, not reading each message on its own terms.
It's a fixed binary-choice specialist, not a general inbox sorter. Don't ask it to be one.
Install
git clone https://github.com/Zefan-Cai/Open-Jev.git
cd Open-Jev
/opt/homebrew/bin/python3.11 -m venv .venv && source .venv/bin/activate
python -m pip install -e '.[train]'
hf download ZefanCai/Open-Jev-2B --local-dir models/Open-Jev-2B
python -m jev.server --checkpoint models/Open-Jev-2B/package/checkpoint --device mps \
--max-length 4096 --batch-size 1 --no-prefix-cache --port 8792
About 8.5 GB to download, no account needed. Apple Silicon gets you the sub-second speed
above; it also runs on CPU (any OS), just 8-17 seconds per decision instead. The kit
includes check.py, a ready command-line script built from the exact call we used to
generate the numbers in this article: paste any text, get SCAM or SAFE back with a
confidence score.
The honest bottom line
Fast and free, and it caught the obvious fakes cleanly. It missed the subtler ones: a fake
review without the tell-tale tone, a phishing job offer at coin-flip confidence, a fake
delivery link it was confident about. Use it as a first check on your own Mac, never as the
last word where money, credentials, or a bank login are actually on the line.
The keyword
Comment FAKE on the video for the setup kit, or grab it straight from the button above.
The kit has the exact install commands, check.py, and the full honest breakdown of what it
missed and how confidently.
The repo
github.com/Zefan-Cai/Open-Jev (code) · huggingface.co/ZefanCai/Open-Jev-2B (weights)