TypeSafe AI launches Jev, a tokenless decision model

TypeSafe AI unveils Jev, a model that does not produce text but makes probabilistic decisions based on application states and typed questions. Announced on September 15, 2026, the project comes with $40 million in funding led by DCVC. The first public version of the System One family aims for rapid product integrations, promising efficiency while setting clear limitations from the outset.

Assumed Limitations and Safeguards on Evaluation

The system comes with caveats regarding autonomous evaluation and the search for a consensual truth. The model does not perform "thinking" in natural language and exhibits weaknesses in counting and dates. It lacks knowledge about the world and retrieval mechanisms. Exposure to adversarially constructed application states is identified as a vulnerability point.

Targeted Uses and Integration in Three Steps

The cited use cases focus on frequent and structured decisions: ticket triage, routing, filtering for retrieval, and high-frequency loops. The described integration involves installing an SDK, using a single endpoint, and integrating typed outputs within the code. TypeSafe AI indicates that this method could reduce latency, costs, and schema errors compared to the most advanced models.

A Decision-Making Model with Typed Primitives

Instead of generating tokens, the model takes the state of an application and structured, typed questions, then returns probabilistic decisions in a parallel pass. Three primitives are proposed: Choice, Score, and Noul. The RLCD training aims to calibrate confidence as a statistical signal. Developed by TypeSafe AI, Jev represents the first public version of the System One lineage, named in reference to William Stanley Jevons. The startup was founded by Diogo Almeida, a former OpenAI researcher involved in RLHF/InstructGPT. The launch occurred on September 15, 2026, accompanied by $40 million in funding led by DCVC.