Jev and openjev are not the same thing

01What openjev actually is

openjev is not a client for Jev, not an open release of Jev, and not a TypeSafe project. It is an independent effort that started within days of the Jev launch, asking whether the same idea — typed probabilistic answers instead of generated text — can run on hardware you already own.

The main repository is TheoLeeCJ/openjev, written in Python, and its stated question is whether something like Jev can run on a single consumer GPU at home. A second project, ekzhang/openjev-sglang, exposes a Jev-compatible endpoint backed by open models, which is the closest thing to a drop-in local substitute.

The name confuses people, and the confusion runs both ways. Searching for one reliably returns the other, and at least one widely shared thread claiming to have “built Jev a year ago” is about neither.

02How they differ

03When a local reproduction makes sense

If the data cannot leave your network, if you need to classify volumes where per-token pricing stops being trivial, or if you want to study how the approach works, a local reproduction is worth the setup. The tradeoff is that you own the quality problem: nothing guarantees a local cross-encoder is calibrated the way a hosted model claims to be, and calibration is the whole value of a probability.

04When it does not

For most applications the hosted model is cheaper than the GPU you would buy to avoid it. At the published rate, a million judgments costs about fifteen dollars, which is less than a day of the instance you would rent to serve them yourself.

Star counts on these repositories moved by hundreds within a single day around launch. Any figure quoted here, including in this paragraph, is stale the week after it is written — check the repositories directly before making a decision on them.