Lesson 08 · Jev 101
The model card: jev-1.13.0
A complete 101 has to say which model you are calling. This page restates TypeSafe’s models.md. Numbers move — re-read the official page the day you ship.
Official source:models.md. Every model on that page uses the same endpoint,POST /v1/systemone. The request’s model field picks which one answers.
Current public model
Pin jev-1.13.0. The response’s model field reports the versioned ID that actually answered, even if you sent an alias. Log it if you tune thresholds.
| Field | Official models.md | Beginner note |
|---|---|---|
| Price | $42 / Btok · $0.042 / Mtok · output free | Company list price Charged on input tokens only. |
| Rate limits | 250,000 tokens / second · 1,200 requests / minute | Over either limit → HTTP 429. Official: limits adjust dynamically; higher via sales@typesafe.ai. |
| Context | 64k tokens for state + all questions · 32k tokens for state + the longest question | State is ingested once; questions run in parallel against it. Vercel / Netlify / AI/ML API cards that say “32k” are platform summaries — official models.md still publishes both 64k and 32k.PRESS |
| Input | Text only. String, JSON object, or array of text. No image, audio, or video. | Pre-process images, audio, video, binaries into text or fields first. |
Aliases move. Versioned IDs do not.
jev-latest → jev-1.13.0
Most recent stable official release. Official: default in client SDKs and the name used in docs examples.
jev-preview → jev-1.13.0
Most recent release, official or not. Moves ahead of jev-latest when a preview exists. Official today: no preview build — same as jev-latest.
Official: an alias moves when a new release ships, so answers can change without a change on your side. If you tuned confidence against a version, pin that ID and migrate on your schedule.
GET /v1/models lists names your account can send. Official: it currently lists the aliases. Versioned IDs such as jev-1.13.0 are accepted whether or not they appear in that list.
You do not get a customer fine-tune
Official: Jev is not fine-tuned or LoRA-adapted with customer data. Same weights for every account. You shape answers through the request:
- Put proprietary content in state — not in a fine-tune.
- Encode domain rules in instructions and criteria.
- Decompose broad judgments; combine in code (or train a classical model on the probabilities).
Language and data
- English is the primary training language and currently the most accurate.
- Other languages, including CJK, are accepted but not equally well. Test, and watch confidence.
- Official: Jev is not trained on customer requests or responses.
- Official legal.md: enterprise zero data retention is offered — contact privacy@typesafe.ai. Vercel Gateway’s per-request ZDR / No Training flags are a platform control, not a substitute for TypeSafe’s enterprise DPA.
jev-1.13 without the patch as HTTP 400 “Unknown model.” Pin the dotted ID from models.md.PRESS