Explainers · Last updated 2026-09-21

Jev language support

Jev language support is short on the official models page, and we will not pad it with invented per-language leaderboards. English is the primary training language and where TypeSafe says accuracy is currently best. Other languages, including CJK scripts, are handled but not equally well.

This page is for builders who already heard of Jev / System One and want a production-minded mental model before they send non-English state. jev.pro is an independent unofficial field guide — not affiliated with TypeSafe AI. Confirm the models page on docs.typesafe.ai. We do not sell keys. This site is English-only by design.

Definition

Jev is TypeSafe’s first public System One model: unstructured text state in, typed probabilistic decisions out. It understands language enough to judge it. It does not translate, draft, or localize. Input is text only — no images, audio, or video that might “carry” another language as pixels.

If the models page later lists more languages, update this explainer from that source. Do not invent search volumes or accuracy percentages this site did not measure.

Mental model: test, then threshold — do not assume transfer

Their guidance: test on your own content before relying on Jev for a non-English workload, and pay close attention to confidence when routing. We do not publish localized copies that could drift from the official docs.

If you must evaluate non-English state:

1. Keep the questions in clear English or in the language of the state, and test both.
2. Log versioned model IDs (jev-1.13.0, not only jev-latest).
3. Send uncertain cases to humans. Do not assume a Noul of 0.9 means the same thing across languages until you measure it.
4. Filter state. Unused tokens still bill, and 1.13 loses accuracy as distractors grow.

Schema-safe ≠ correct in every language. A Choice still cannot leave criteria, and can still pick the wrong label.

Cost and context (vendor claims)

You pay input tokens at the public list price TypeSafe publishes ($0.042 per million; output free — confirm on the models page). Scripts that tokenize longer cost more — measure usage. Context remains 64k for state plus all questions and 32k for state plus the longest question, regardless of script.

Latency figures (70ms–500ms) and “40x–200x” are vendor claims about System One shaped queries, not a promise that CJK traffic matches English p95.

What language support is not

• Not a translation API.
• Not a reason to generate replies in the customer’s language. Jev does not generate text. Compose an LLM if you need a localized email after the decision.
• Not a ChatGPT replacement with a language dropdown.
• Not official documentation of jev.pro as a product. We only restate the models page.

Jaggedness still applies: weak numeric precision, extra indirection, literal wording. Keep arithmetic in code even when the ticket is in another language.

See also

• Explainers hub
• State
• Jev 1.13
• Confidence

FAQ

Does Jev translate?

No. It evaluates text you send. Translation is another system.

Should criteria be English if state is not?

Test both. TypeSafe does not publish a per-language leaderboard we will invent.

Does language change price?

You pay input tokens (vendor list price). Scripts that tokenize longer cost more — measure usage.

Can I trust confidence the same way in Japanese as in English?

Not until you measure calibration on that traffic. RLCD is a named objective, not a per-language certificate.

Disclaimer

This is an independent unofficial site and is not affiliated with TypeSafe AI; official documentation is available at https://docs.typesafe.ai. Never treat jev.pro as TypeSafe official documentation. We do not sell, issue, or proxy API keys.