Jev · open source
What "open-source Jev" actually means
TypeSafe has not released the model. Several different kinds of project get written up under that headline anyway, and they answer different questions — "can I run this", "can I train it", "how does it work", "can I test the pattern without a key".
In short
- No open release of the model. Jev is a hosted API; everything below is somebody else's build.
- The closest substitute publishes its own losing score: 72.0% vs 96.6% macro, and it wins the real-time loop at 9.00 vs 5.62 kills.
- You can own the whole stack — Kev is a Jev-shaped family on Qwen3.5, 0.8B to 9B, with the training path included.
- You can test the pattern today without a TypeSafe key by giving an open model the same request shape.
- None of them promises what the hosted model promises: an answer that cannot leave your option set.
Four questions, four kinds of project
Sorting them by the question they answer is the only way this stops being confusing, because these projects are not competing with each other — they are competing with different parts of your problem.
The five we read, and what each gives up
Every row is a repository we opened on 2026-09-22; the numbers are the publishers' own, and where a project publishes a head-to-head against the model it replaces, the losing side is printed too.
| Project | What it is | What you get | What it gives up |
|---|---|---|---|
| Von 399 ★ | an open non-autoregressive decision model — 395M parameters, 1.5GB, runs locally | publishes its own head-to-head against the model it replaces: 72.0% macro on a 49-task suite vs 96.6%, and 9.00 kills vs 5.62 in ViZDoom | 24 points of accuracy for a real-time loop it wins at 25 ms a decision |
| Kev 2,857 ★ | a family of Jev-like decision models on Qwen3.5 — 0.8B, 4B and 9B — with the training path included | you own the model and the weights; measured on Kev-4B on an L4 in bf16, 118 ms for a 101-token request | you also own the training, the serving and the calibration |
| SemIf 3,496 ★ | semantic ifs from open models, on a single 3090 at home | the boundary drawn in your own code, with an open model behind it | consumer hardware and your own judgement about the boundary |
| NanoJev 1,910 ★ | a small replica of the architecture: parallel decisions, dynamic option sets | the clearest read of how the thing works — a replica is an explanation | it is a teaching implementation, not a service |
| simple-jev 469 ★ | turns any open model into a classifier endpoint with the same calling shape | the pattern without the vendor: testable today, at zero cost | the option-set guarantee is a prompt and a parser here, not a property of the model |
Von is the honest one, and that is why it is the useful one
The nearest thing to a substitute publishes the comparison that makes it look worse. On a 49-task suite it reports 72.0% macro against Jev's 96.6%, and in the ViZDoom Defend-the-Center protocol 9.00 kills against 5.62 — a 24-point accuracy gap and a real-time win, in one table, from the people who built it.
That is the shape of the trade, and it is worth more than a benchmark somebody else ran: the open model is fast and cheap enough to sit in a loop, and it is wrong more often in ways that matter. Whether that is acceptable is a property of your loop. In a game the 395M model plays at 25 ms a decision and nobody is billed for a mistake; in a payment routing decision the same gap is not a trade, it is a defect.
Both sides of Von's comparison are the publisher's, including the number that flatters the hosted model. We have not reproduced either.
The thing none of them copies
Every project here reproduces the shape: a state goes in, an option comes out, in one pass. What none of them can reproduce from outside is the guarantee that made the hosted model worth writing about — that the answer cannot be something you did not offer. In simple-jev that guarantee is a prompt and a parser; in a replica it is whatever the evaluation showed.
We tried to break the hosted model's version of it and could not, in thirteen attempts across four runs — and we published every attempt, including the two that tried to widen the answer set itself. So the honest summary of this page: if you need the guarantee you are buying a service and these projects are how you learn what it does; if you need the shape, several of them will give it to you this afternoon, and one will tell you where it fails.
What this page cannot tell you
It cannot rank these projects, because they were not measured on the same tasks — one publishes a 49-task suite, another publishes latency on an L4, a third publishes nothing but a README. It cannot tell you which is cheaper over a month, since the honest comparison is your tokens against your electricity and your weekend. And it leaves out the forks that exist only as a description: a repository we cannot read is not a project, and this page is built from repositories we opened.
Is there an open-source Jev?
Not a release of the model. TypeSafe's Jev is a hosted API. What exists are three different things with the same name in the headline: local substitutes trained from scratch, small replicas built to study the architecture, and projects that turn another open model into a decision endpoint with the same calling shape.
What is the closest local substitute?
Von, by wfzyx: a 395M-parameter non-autoregressive decision model, published with its own head-to-head against Jev. It reports 72.0% macro on a 49-task suite against Jev's 96.6%, and 9.00 kills against 5.62 in the ViZDoom protocol — a clear loss on accuracy and a win on the real-time loop.
Can I just call a small open model instead?
That is what featherless-ai/simple-jev does: it turns an open model into a classifier endpoint with the same request shape. It is the cheapest way to test the pattern without a TypeSafe key, and it does not carry the guarantee the hosted model makes — a fixed option set cannot be left.
repository READMEs Verified 2026-09-22 Each project links to its own page, where the dossier says what we checked and what we did not.Scores are the publishers' own; we reproduced none of them.