TypeSafe launches Jev, a text-free decision model

Key Takeaways

1TypeSafe AI launches Jev, an AI model that returns probabilistic decisions based on typed questions and the state of an application, without text generation.

2Unveiled on September 15, 2026, and funded with $40 million led by DCVC, Jev offers dedicated primitives, RLCD calibration, and API/SDK integrations with multiple workflows.

3Its limitations include the lack of arithmetic, world knowledge, and an evaluation based on self-reported benchmarks.

Why it matters : Jev provides an alternative to traditional generative models for applications requiring structured and rapid judgments.

TypeSafe AI, founded by Diogo Almeida, presents Jev, an AI model unveiled on September 15, 2026, and funded with $40 million led by DCVC. Unlike autoregressive chat models, Jev answers typed questions by returning probabilistic decisions in a single pass. However, comparisons of speed, cost, and reliability rely on self-reported metrics from the publisher.

Explicit Limits and Caution in Measurements

The comparison of speed, cost, and reliability presented is based on benchmarks provided by TypeSafe itself and evaluates concordance rather than independently verified accuracy. In terms of capabilities, Jev struggles with literal answers and is unable to perform calculations or arithmetic. Dates are treated only as text, and the model has no knowledge of the world or generative ability. By default, its state management is non-adversarial.

Parallel Decisions, Typed Primitives, and RLCD Calibration

Jev deviates from autoregressive text generation: it takes the state of an application and structured, typed questions to produce probabilistic decisions in a single parallel pass. Three primitives serve as an interface for queries: Choice, Score, and Noul. A calibration approach, RLCD, transforms confidence into usable statistical signals. This setup eliminates schema parsing and output issues while allowing for the dissemination of multiple decisions per query at a reduced cost.

Le brief IA que lisent les pros

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

Typed APIs, SDKs, and Ready-to-Use Decision Workflows

Integration occurs via typed API endpoints as well as SDK models. Various decision workflows are offered, including broadcasting, confidence-based routing, cascades, and retrieval-then-judge.

Launch, Team, and Targeted Use Cases

Developed by TypeSafe AI, Jev is the first public release from the "System One" class, named after economist William Stanley Jevons. The startup, founded by Diogo Almeida, a former OpenAI researcher who worked on RLHF/InstructGPT, presented the model on September 15, 2026. Its funding amounts to $40 million, led by DCVC. Highlighted use cases include triage and routing, filtering, guardrailing, model routing, agent control signals, and batch classification.