Jev is TypeSafe AI’s AI Model Serving & Inference API. It is a System One model: software sends a state and a map of typed questions; Jev returns structured answers, probabilities, and (for Choice and Score) confidence. There is no generated prose to parse. Application code branches, sorts, and routes on those values. Jev is not a chat or code-completion LLM.

Key Capabilities

Three primitives, one request: Choice selects from a closed set and returns choice, per-option probabilities, and confidence. Score rates the state on an ordered rubric and returns score, per-level probabilities, and confidence. Noul is a yes/no question and returns noul in 0–1 (probability that the answer is yes). Questions share one state, run in parallel and in isolation, and mix in a single POST. Adding questions does not grow a shared generation context.
State: A string, JSON object, or array of text (chat logs, records, application state). Nested fields can be referenced from instructions. Text only: no image, audio, or video. Non-text inputs must be converted before the call.
Confidence-gated control: Probability is the answer; confidence is whether to act. Documented patterns include speculative fan-out, confidence-gated routing, composite scoring (atomic scores combined with weights in code), and intent routing to deterministic logic, a specialist LLM, or a human.
Documented jobs: Ticket routing and urgency; RAG passage scoring and citation checks; BM25 shortlist re-rank; line-by-line semantic find; hierarchical classification; function-name and argument mapping for typed tools; skill selection for agent turns; entity alignment; LLM input/output screening (pass, review, block); structured-data extraction cascades; date and span extraction after regex candidates.
What it does not do: Write replies, code, or explanations. Docs state it is not a drop-in model for Claude Code, Cursor, Copilot, or similar agents. A TypeSafe agent skill exists so those agents can generate TypeSafe integrations.

Audience & Use Cases

Audience: Application and agent developers who need a semantic decision inside code, not a conversational model.
Use Case: Classify and route messages; score and rank retrieval candidates; gate LLM traffic; map language onto typed function calls; batch judgments over large text corpora; keep control flow in code.

Technical Specifications

Model: jev-1.13.0; aliases jev-latest and jev-preview currently resolve to that ID.
Context: 64k tokens per request (state plus all questions); 32k for state plus the longest single question.
Limits (docs, subject to change): 100K tokens/s and 80 requests/s; 429 when exceeded. SDKs retry with backoff and honor Retry-After. Higher limits on custom/enterprise plans.
Billing (vendor list): Input billed; output tokens free. Jev 1.13 listed at $42 per billion input tokens ($0.042 per million).
Training (vendor): Reinforcement Learning for Calibrated Decisions (RLCD). Calibration is measured across groups of predictions; an individual answer is not guaranteed correct.

Jev

What is Jev?

Categories & Use Cases