Does Jev Train on Your Data? Privacy and Retention
Last checked · Independent guide, not affiliated with TypeSafe AI
According to TypeSafe, no: Jev is not trained on customer requests or responses, and every customer uses the same model weights. TypeSafe offers zero data retention (ZDR) to enterprise customers. If you call Jev through Vercel, OpenRouter or Cloudflare, that platform's data policy applies to your request as well.
What TypeSafe commits to
Section titled “What TypeSafe commits to”TypeSafe’s model documentation and legal page make three statements:
- No training on your data. Jev is not trained on customer requests or responses. The privacy policy repeats this commitment.
- No per-customer models. Jev is not fine-tuned or adapted with any customer’s data; the same weights serve every account. Your content shapes answers only through what you put in the request.
- Zero data retention on request. Enterprise customers can get ZDR by contacting [email protected].
The legal page also links a Data Processing Agreement, which covers how TypeSafe processes customer data and how long it is retained, and a Master Customer Agreement. Read those documents for the details that matter to your compliance team; this page is a summary, not legal advice.
What “not trained on” does not mean
Section titled “What “not trained on” does not mean”- It does not mean TypeSafe never stores requests. Unless you have ZDR, assume requests may be kept for some period for operations, abuse prevention or debugging, as with most APIs; the DPA sets the terms.
- It does not cover other companies in the path. When you use a reseller, your request passes through that platform first.
Through other platforms
Section titled “Through other platforms”Hacker News commenters raised this about OpenRouter specifically: routing requests through a marketplace adds another party that sees your data. Each platform has its own policy:
| Channel | Who handles your request | Where to check |
|---|---|---|
| TypeSafe API | TypeSafe | TypeSafe’s privacy policy and DPA |
| Vercel AI Gateway | Vercel, then TypeSafe | Vercel lists ZDR and “no training” per model and provider, and links TypeSafe’s terms and privacy policy |
| OpenRouter | OpenRouter, then TypeSafe | OpenRouter’s privacy and logging settings, plus TypeSafe’s terms |
| Cloudflare Workers AI | Cloudflare, then TypeSafe | Cloudflare’s model terms, linked from the Jev catalog page |
If data handling is a hard requirement, the simplest setup is TypeSafe’s API directly, with an enterprise ZDR agreement.
Practical steps
Section titled “Practical steps”- Send less. Jev only needs the text relevant to the question, and TypeSafe notes that smaller states are more accurate anyway. Strip names, emails and account numbers when the decision does not depend on them.
- Use IDs instead of personal data. Keep a mapping in your own database and send the model a reference.
- Check your own obligations. If you process data about people in the EU or other regulated contexts, you need a data processing agreement with every processor in the chain.
- Keep keys server-side. Calls should come from your backend, not the browser, so you control exactly what is sent.