Frequently asked questions.
What can I use eu/jev for?
Turn context into a choice, a probability, or a score. Use it to route support tickets, classify content, or evaluate options in your application. Try an example in the playground before connecting the API.
Where is my data processed?
Inference runs on Hetzner servers in Germany. Prompts and answers are kept out of our database and logs. Our privacy policy explains transient processing, account records, and usage metadata. A self-serve data processing agreement is available for business use.
How much does it cost?
€37 per billion input tokens. Output tokens are free, and new accounts receive €5 in starting credits. Sign up with Google to get started. For extra usage, use our request form. There is no automatic billing when credits run out.
Can I use the same request format as Jev?
eu/jev supports Jev-style requests to /v1/systemone, with shared context and choice, noul, or score questions. Use eu/jev’s base URL, API key, and a supported model ID. Supported fields, limits, token counts, and outputs can differ; see the compatibility documentation.
What model are you using in the back?
eu/jev is built on top of Winnow, the top open-source model in the JevBench v1.5.4 Jev-class capability ranking.
Original Winnow. Published results.
Winnow-12B Q8 evaluation score on JevBench.
- Capability / 100
- 79.3
- Intelligence / 100
- 74.4
- Calibration / 100
- 84.1
- Measured median latency
- 95 ms
| Model | Capability / 100 |
|---|---|
| Jev 1.13.0Closed model | 80.0 |
| Winnow-12B Q8Open source | 79.3 |
| CygnetOpen source | 79.0 |
Scores run from 0 to 100. Capability averages intelligence and calibration; it is not an accuracy percentage.
1,624 decisions · 904 open + 720 sealed. Latency uses 450 samples. JevBench’s adjusted median is 340 ms: 2 × measured latency + 150 ms, an estimate rather than a measured hosted API result.
JevBench v1.5.4Original resultsChecked 1 October 2026