OpenJev vs Jev: the same question, two different models
We host one of these and not the other, so read this knowing that. The short version: they take the same request and return the same shape, they are unrelated models by unrelated people, we expect ours to be the weaker one, and nobody has measured the gap — including us.
The names are the problem
OpenJev is not an open version of Jev. It is not a smaller checkpoint, a free tier, or a community port. It is a separate MIT-licensed model that took the name because it answers the same kind of question.
If you arrived expecting the open weights behind TypeSafe's model, they are not here and they are not anywhere. We do not resell Jev; TypeSafe sells it directly.
Side by side, only where a comparison is real
| OpenJev | Jev | |
|---|---|---|
| Who makes it | AlexWortega, open source | TypeSafe |
| Licence | MIT | Commercial, theirs |
| Where we serve it | Free on our own GPU, with a registered key | We do not serve it at all |
| Request shape | state + typed questions | state + typed questions — the same shape |
| Question types | yes/no, choice, score | yes/no, choice, score |
| What the number is | A share of the entailment found across your options | Probabilities fitted against real outcomes |
| Input | Text and JSON here; the checkpoint also reads images, we do not expose that | Text |
| Quality against each other | Not measured by anyone, including us | Not measured by anyone, including us |
The last row is the one most comparison pages would quietly fill in. No published head-to-head exists, we have not run one, and a figure invented to fill that cell would be the least trustworthy thing on the page.
The one difference that is concrete
Forget which is "better" for a moment. There is a difference in what the numbers mean, and it changes how you are allowed to use them.
Jev is trained so its probabilities line up with outcomes: across many predictions, the ones it calls seven-in-ten happen about seven times in ten. That is what calibration buys you, and it is why you can set a threshold from the number itself.
OpenJev is an entailment model, and for a labelled choice its scores are divided by their total so they sum to one. Three options it supports equally weakly and two it supports equally strongly produce the same even split. The number is a share of the agreement found, not a probability the answer is right, so a threshold has to come from your own labelled examples instead of from the number's face value.
That is not a quality claim. It is a difference in kind, and it is the one thing on this page you can check for yourself in an afternoon.
Which to use
When the probability has to mean something — a threshold you will defend, a risk budget, anything where "act above 0.85" is a policy rather than a hunch. Also when you want the strongest available model for this task and are willing to pay for it. Go to TypeSafe.
When you want to find out whether typed judgments help at all before committing to anything, when you are ranking or filtering rather than gating, or when free and no-signup-beyond-a-key matters more than the last few points of accuracy. It costs nothing to find out.
A third option people skip: try both against a chat model with a JSON schema on your own data. On BlockRun that is one key and a changed string, which is the whole reason this comparison is worth our time to write.
Read next
- OpenJev — what the model is, its authors' own benchmark figures, and the licence.
- Decide — the endpoint, the three question types, and how to call it free.
- API reference — the request shape, the response fields, and the arithmetic worked through.
OpenJev vs Jev questions: differences, calibration, which to use
- What is the difference between OpenJev and Jev?
- They are unrelated models by unrelated people that take the same request shape. Jev is TypeSafe's commercial model; OpenJev is MIT-licensed and open. The concrete difference is what the numbers mean, not just who made them.
- Is OpenJev as good as Jev?
- We expect it to be materially weaker, and we have not measured the gap. No published head-to-head exists and we are not going to invent a figure to fill that cell — we host one of the two, which is exactly why we will not.
- What does calibration mean when comparing Jev and OpenJev?
- A calibrated model's probabilities line up with outcomes: across many predictions the ones it calls seven-in-ten happen about seven times in ten, so a threshold can come from the number itself. OpenJev's choice scores are entailment shares normalised to sum to one, so a threshold has to come from your own labelled examples instead.
- When should I use Jev instead of OpenJev?
- When the probability has to mean something — a threshold you will defend, a risk budget, a policy rather than a hunch — or when you want the strongest available model for the task and will pay for it. TypeSafe sells it directly.
- When is OpenJev the right choice?
- When you are finding out whether typed judgments help at all, when you are ranking or filtering rather than gating on an absolute number, or when free and no signup beyond a key matters more than the last few points of accuracy.
- Can I get Jev through BlockRun?
- No. We do not resell it and there is no gated path waiting to be switched on. One backend serves this endpoint and the response header names it on every call.
- Are OpenJev's weights the open version of Jev?
- No, and this is the most common wrong assumption. TypeSafe's weights are not public anywhere. OpenJev is a separate model that adopted a similar name.
- OpenJev vs Jev vs a reasoning model — how do I choose?
- Judgment models are for the fast, narrow, typed question; a reasoning model is for the case that needs deliberation. The useful shape is to judge cheaply, act where the margin is wide, and escalate the narrow cases. Testing all three against your own data is one key and a changed string here.
- Why does a page comparing them come from a company that hosts one?
- It should make you read it more carefully. The page concedes the things that cost us something — that we expect ours to be weaker, that nobody has measured it, and that Jev is the right answer when the probability has to be defensible — because a comparison from an interested party is worth nothing unless it can be caught being fair.