Jev (AI model)

No. The article states TypeSafe has published no exact architecture, no weights and no technical paper. It describes the model as transformer-based and trained only on synthetic data, and notes that outside observers have suggested it may be built on an open-weight LLM.

Hosted access is the only documented route. The article lists the license as proprietary and the release as limited early access, with the model returning typed values, probability estimates and confidence scores to software rather than downloadable weights.

No. Jev is described as a discriminative model for classification or regression that does not produce natural-language text, and whose output is meant to be read by software rather than a person. It only answers typed questions you define at prediction time.

No. The 70 to 500 ms end-to-end response times are TypeSafe's own figures, drawn from vendor posts and press coverage. The entry's own note is that such numbers likely sit at the high end of real-world performance, and no third-party benchmark appears on the page.

Limited early access on 15 September 2026, alongside a $40 million seed round led by DCVC. The stable release listed is jev-1.13.0, the license is proprietary, and it was developed in stealth for roughly two years beforehand.

Founded in 2024 by Diogo Almeida, Erik Gafni and Sasha Sheng. Almeida, the chief executive, spent about four years at OpenAI on RLHF, InstructGPT, ChatGPT and GPT-4 before leaving in 2024.

The System One model works on Reinforcement Learning for Calibrated Decisions, or RLCD. The name comes from 19th-century economist William Stanley Jevons, whose Jevons paradox describes how more efficient use of a resource can increase consumption.