Is Jev open source?
Updated 24 Sep 2026
No. TypeSafe AI serves Jev only through its hosted API and has not published the weights, so you cannot self-host, fine-tune or inspect it. Open models such as Laya answer the same Choice, Score and Noul questions on your own hardware, with weaker accuracy until you train them.
What can’t I do with Jev?
- Run it offline or on your own hardware.
- Keep your data from leaving your servers.
- Fine-tune it on your own labels.
- Pin a model version you control.
Which open models can I run instead?
| Model | What it is | Our verdict |
|---|---|---|
| Laya Convai Innovations | Open-weight model with the same three question types, Apache 2.0. Checkpoints: English (421M), multilingual (322M), and one tuned on typed decisions (421M). | Real and runnable. Its headline accuracy claim is misleading. |
| Kev | Jev-like decision models built on Qwen, with a TypeSafe-compatible API. | Plausible |
| openjev (SemIf) | A Jev-style decision model that runs on one consumer RTX 3090. | Verified |
| jevlike | An open reimplementation you train yourself, with Doom, chess and Wikispeedia demos. | Verified |
Verified means the project exists and does what its author says. It does not mean it matches Jev’s accuracy.
Which should I pick?
- No labeled data and you need it working today: Jev, if a hosted API is acceptable.
- Private data or high volume, and you have labeled examples: Laya, fine-tuned on your labels.
- You want a local server your Jev code can call: Kev, after you compare its answers with Jev’s on your own questions.
- You want to understand how the architecture works: jevlike.
Is an open replica as good as Jev?
Not on the published evidence. Laya’s base checkpoint is near chance on typed decisions without training, and its 0.766 headline score came from a checkpoint fine-tuned on that benchmark’s own training split. The replicas report their own numbers, and we have not compared them with Jev on the same questions.
Sources and test status
Reviewed 24 September 2026 against TypeSafe’s documentation, the Laya model card and each project’s repository. We have not yet run any of these models ourselves.