Diogo Almeida, a former OpenAI researcher, launched the Jev model through his startup TypeSafe AI. Unlike traditional LLMs, Jev is a transformer-based model that outputs probabilities for predefined decisions rather than generating text. This architecture is designed to eliminate hallucinations and improve speed for automation tasks.
The model has already seen significant adoption, with Vercel reporting that replacing OpenAI models with Jev increased instruction verification speed by 5 to 18 times. In other benchmarks, Jev demonstrated cost savings of 10 to 20 times compared to Gemini, while providing numerical probability scores for its outputs.
Jev is intended for use in AI agent monitoring, request routing, and security compliance. TypeSafe AI trains the model using synthetic data through a proprietary method called 'calibrated decision reinforcement learning.' While the company keeps the architecture private, it plans to expand the model to support additional modalities in the future.
The name Jev references the Jevons paradox, which suggests that increased efficiency leads to higher consumption. TypeSafe AI aims to lower the cost of intelligent decision-making, allowing developers to integrate these capabilities directly into software applications. The company currently offers free output tokens for its users.
