TypeSafe AI, founded by former OpenAI researcher Diogo Almeida, has released a new transformer-based model called Jev. Unlike large language models, Jev is designed for software automation and does not generate text. Instead, it outputs calibrated probabilities to facilitate decision-making processes.
The model is trained exclusively on synthetic data using a technique the company calls reinforcement learning from calibrated decisions. Because users define outputs in advance, the system is designed to avoid hallucinations. Input tokens are metered by the billion, making the model faster and cheaper than traditional LLMs.
Early testing shows significant performance gains in automation tasks. Vercel reported that Jev processed commands 5 to 18 times faster than OpenAI’s Luna 5.6. Meanwhile, Bryo AI found the model to be 10 to 20 times more cost-effective than Gemini for email classification, while providing confidence scores for its outputs.
Developers are exploring Jev for use cases such as monitoring LLM agents to prevent jailbreaks and performing real-time model routing. By providing statistical probabilities, the model allows users to set thresholds for automated workflows. TypeSafe AI plans to develop additional versions of the model in new modalities.
