AI modelsTypeSafe AI

Jev

Decision matrix

Jev model decision matrix
Overview

Use Jev when your application needs to choose, score or check a well-scoped question and act on a typed answer. TypeSafe's hosted System One model returns probabilities and, for Choice and Score, confidence. It does not generate text or replace a chat LLM. Start with one decision in your own workflow and keep thresholds and final actions in code.

Other good fits
  • Scoring a piece of text against a defined rubric before your application decides what happens next.
  • Checking a narrow yes/no condition within an existing software workflow.
Pros
  • Typed, constrained decisions avoid parsing free-form generated answers.
  • Several independent questions can share a state in one request; the response names each result.
  • Choice and Score include distributions and confidence so code can route uncertain cases for review.
Cons
Read cons
  • No generated prose, code, chat or non-text input; use a generative model or preprocess other media.
  • TypeSafe documents literal readings, poor counting and date arithmetic, sensitivity to irrelevant state and adversarial content for Jev 1.13.
  • A confidence value does not prove one answer is right; assess errors on examples from your own task before automating consequential actions.

The publisher's Jev 1.13 limitations (opens in a new tab) explicitly put math, exact dates and generation elsewhere. Choice and Score expose confidence; Noul exposes a yes-probability without the separate confidence field. Early access and cloud-only service are additional adoption constraints.

Avoid if

  • Your task is exact arithmetic, counting, date comparison or deterministic validation; keep that logic in code.
  • You require offline or self-hosted inference; TypeSafe documents a hosted API, not downloadable Jev weights.
Context

The Jev 1.13 model page (opens in a new tab) documents 64,000 tokens across state and all questions in one request, plus a 32,000-token limit on the state and longest single question. Input is text only, supplied as a string, JSON object or array of text. These are capacity limits, not a claim that distant or irrelevant state is recalled reliably; TypeSafe warns that long irrelevant state can reduce accuracy. Pin jev-1.13.0 when a changing alias would make your thresholds hard to compare.

Requirements

TypeSafe documents a hosted POST https://api.typesafe.ai/v1/systemone endpoint, an account/API key and Python or JavaScript clients. Use model jev-1.13.0 for the documented version or jev-latest to follow future releases; the response reports the version that answered. No local weights, self-hosted runtime, quantization or hardware memory requirement is documented for Jev. The documented request-rate limits are service limits, not inference-hardware measurements.

Licence & cost

On 24 September 2026 the model page (opens in a new tab) lists Jev 1.13 at $0.042 per million input tokens ($42 per billion); output tokens are free. It lists 250,000 tokens per second and 1,200 requests per minute but warns those limits can change. Billing for repeated calls and engineering to handle authentication, retries, review paths and state selection remain your responsibility. This is a hosted service price, not a downloadable-weights licence; no rights to run or redistribute Jev weights are established here.

First task

Documentation-based suggestion — not run here.

Input: one fictional request and the destinations that make sense in your workflow. Expected output: a Choice answer with one selected destination and per-option probabilities. Ask a human to examine ambiguous examples before setting a threshold; do not treat a single successful call as proof of calibration.

Links & evidence
Documentation-based · not hands-on · Sep 24, 2026

Read TypeSafe's official model, System One, quick-start, API, confidence and Jev 1.13 jaggedness pages on 24 September 2026. Pricing and limits are dated publisher claims. No Jev credentials were used, no request was sent, no model weights or licence text were located, and no comparative performance or calibration test was performed.

  • TypeSafe model versions and limits · Sep 24, 2026 — Jev 1.13 model ID, aliases, 64k/32k context limits, text-only input, $0.042 per million input tokens, output pricing and changing rate limits.
  • System One concept and output scope · Sep 24, 2026 — Choice, Score and Noul typed decisions; distinction from generated replies and code; model uncertainty does not guarantee correctness.
  • Official Jev quick start · Sep 24, 2026 — Hosted endpoint, API key, example state/questions, and response fields; published sample is not our own executed run.
  • Confidence definition · Sep 24, 2026 — Choice and Score confidence is derived from probability distributions; Noul lacks a separate confidence field.
  • Jev 1.13 limitations · Sep 24, 2026 — Publisher-documented literalness, weak arithmetic/counting and date comparison, long irrelevant state and adversarial-text limitations.
Further review notes & alternatives

What it unlocks

Jev 1.13 is TypeSafe AI's System One model: it evaluates text state against narrow, typed questions and returns values your software can use directly. A Choice selects among named options and returns their probabilities; a Score rates against your ordered rubric; a Noul returns a yes-probability. Choice and Score also include confidence derived from the probability distribution. Your code owns the thresholds and actions. This is not a conversational or code-generating LLM. See the System One explanation (opens in a new tab) and API reference (opens in a new tab).

A first useful project

Pick a real routing decision from your own workflow, such as assigning a synthetic request to billing, technical support or a person to review. Define a Choice question with those three options and send a short fictional message as state using the official quick start (opens in a new tab). Read the selected choice, its probabilities and confidence; let your code decide whether to route or request review. A useful result is a routed example plus a visible uncertain path. This is a documentation-based suggestion, not an executed API run or a guaranteed accuracy threshold.

Alternatives and trade-offs

GPT-5 mini is a hosted generative model when you need explanations or free-form text as well as classification; its structured outputs are a different interface and require task-specific validation. MiniCPM5-2B is a smaller open-weight text model when self-hosting is essential, but running it requires your own inference infrastructure and its text/tool output is not Jev's typed decision API. Neither alternative is claimed to match Jev's probability calibration; evaluate the same learner-chosen decision and operational constraints before switching.

Learn more

Catalogue facts

Jev is TypeSafe AI's hosted System One model for making typed decisions inside software. Give it text or structured text state and narrow Choice, Score or Noul questions; it returns probabilities for routing, classification and scoring. It does not generate prose or code and cannot replace a chat or coding-agent LLM. The documented Jev 1.13 version is available through TypeSafe's API in early access. Start with a decision in your own workflow and keep the final action in your application code.

TypeSafe AI · paid · cloud

  • Choice, Score and Noul questions in one API request
  • Typed answers and probability distributions
  • Confidence values for Choice and Score
  • Text or structured text input via the System One API
  • Typed decision routing in applications
  • Classification against a fixed choice set
  • Scoring text against a defined rubric
  • Confidence-aware review queues

Browse all tools