What OpenJEV is
OpenJEV is an open-weights decision model published on Hugging Face as openjev/openjev. You describe a decision in plain words at request time, with your own labels, and it returns a choice, a yes/no probability (noul), or a score — the same three shapes as Jev.
It exposes POST /v1/systemone with model, state, and questions, so code written for TypeSafe’s endpoint can be pointed at a self-hosted OpenJEV server. The project also runs a hosted API at api.openjev.sh with the default model alias openjev.
- Independent project, not a TypeSafe release.
- Weights: CC BY-NC 4.0. Helper and serving code: Apache 2.0.
- Up to 52 options per question in one pass; prompts up to 16,384 tokens; one image per request.
OpenJEV vs hosted Jev: the published numbers
The OpenJEV authors ran the hosted Jev API and OpenJEV on the same 10,000 text questions. Jev answered 85.4% correctly, OpenJEV 84.0%, and the untuned base model 80.4%. On 100 MiniWoB browser tasks both completed 39.
On one H100 with FP8, the authors report about 80 ms for a short text decision and about 210 ms for a web decision with roughly 1,460 prompt tokens. Hosted Jev on OpenRouter shows a p50 latency around 0.23 s, including the network.
- Intent / routing: 92.8% both.
- Ethics / policy judgement: OpenJEV 78.1%, Jev 75.8%.
- Science / facts / claims: OpenJEV 82.5%, Jev 89.0%.
When to pick which
Pick OpenJEV for research, data that cannot leave your hardware, or when you already have an 80 GB GPU idle. Check the non-commercial licence before you put it in a paid product.
Pick hosted Jev when you want no GPU to run, commercial terms, and TypeSafe’s version pins such as jev-1.13.0. Input is $0.042 per million tokens direct; on jevmodel.org you can start with 100,000 free tokens and packs from $9.90.
curl https://jevmodel.org/v1/systemone \
-H "Authorization: Bearer $JEVMODEL_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"jev-latest","state":"I was charged twice.",
"questions":{"team":{"type":"choice","instructions":"Which team?",
"criteria":{"billing":null,"technical":null}}}}'FAQ
Is OpenJEV the same as Jev?
No. OpenJEV is an independent project and says so in its model card: it is not affiliated with TypeSafe, and Jev is TypeSafe’s product. It copies the request and response shape of the hosted Jev API so a Jev client can point at an OpenJEV server.
Is OpenJEV free for commercial use?
Not by default. The OpenJEV weights are CC BY-NC 4.0 — research and other non-commercial use with attribution. The helper and serving code are Apache 2.0. For commercial use the authors ask you to open a discussion on the Hugging Face repo.
How accurate is OpenJEV compared with Jev?
On the authors’ 10,000-question text benchmark (34 public sources), OpenJEV scored 84.0% and the hosted Jev API 85.4%. OpenJEV was ahead on sentiment and ethics/policy, behind on commonsense and science/facts. These are the authors’ numbers; test on your own data.
What hardware does OpenJEV need?
The bfloat16 weights are about 54 GB and the documented recipe serves them with vLLM and --quantization fp8 on one 80 GB GPU. There is also a 29 GB FP8 checkpoint and a 27 GB MLX 8-bit build for Apple silicon (text only).
Can I try Jev without running anything?
Yes. The jevmodel.org playground calls the hosted Jev model in your browser — 10,000 input tokens without signup and 100,000 after sign-in — so you can compare against your own OpenJEV answers on the same inputs.