Typed yes/no, choice and score decisions with probabilities, built for routing and triage inside software instead of text generation.
Output Speed *
Intelligence Index *
Context Window *
Input price
Output price
Performance
Parallel typed answers replace a generate-and-parse loop
Technical Specifications
What the model supports
Comparison
Jev 1.14 or Jev 1.13: What Actually Changes?
Jev 1.13 is the direct predecessor and shares the same state-plus-typed-questions interface. TypeSafe describes 1.14 as small improvements across the board, with the largest expected gain in multilingual support. Box questions are exclusive to 1.14. Both models accept images behind the danger-zone header and both support the opt-in long-context mode. The input price is unchanged. The decision is whether the newer model's gains justify re-checking thresholds you tuned on 1.13.
Choose Jev 1.14 when your inputs include non-English text, when you want to test Box questions on images, or when you are starting a new integration. Choose Jev 1.13 when confidence thresholds are already tuned against it and you want results to stay stable. TypeSafe keeps 1.13 available and has promised 60 days' deprecation notice, against 30 days for 1.14. Pin a versioned ID rather than a moving alias either way.
Limitations & Trade-offs
Best-Fit Workloads
Where this model earns its place
Routing and triage
Jev 1.14 is positioned for routing, classification and other decision points where a fast, predictable answer matters more than prose. One request can return department, urgency and sentiment together, each with its own probability. It is the decision model behind AI/ML API's Jev Router, which uses it to pick the bucket a chat request belongs to. Use it for support triage, model routing and content tagging. Handle low-confidence cases with a human or an LLM.
Confidence-gated automation
Every Jev answer carries probabilities and a confidence estimate, so code can decide when to act autonomously and when to ask for review. That suits approvals, moderation and agent guardrails where the cost of a wrong action is uneven. Pydantic AI's docs say a Jev-based guard belongs alongside deterministic checks. Injected text in the state can sway its answers, so test with adversarial inputs before removing human review.
Multilingual classification
TypeSafe expects the largest gain of 1.14 over 1.13 in multilingual support, which makes this release the one to evaluate for non-English tickets, comments and moderation queues. TypeSafe has published no per-language numbers, so treat the gain as a direction to test. Run your own labeled sample in each target language before routing production traffic on 1.14 thresholds.
Image questions and element location
With the images header, 1.14 can answer yes/no or choice questions about PNG or JPEG images, and Box questions return one normalized bounding box with a confidence. Candidate uses include screenshots, scanned forms and locating a field or button. TypeSafe flags this as experimental and recommends the simplest possible questions. Each Box answer is a single box, so several targets need one question each.