Jev

Jev is a transformer-based model from Type Safe AI, a startup founded by former OpenAI researcher Diogo Almeida. It is not a large language model and does not generate text. Instead, developers define the possible outputs in advance (for example a choice, a score, true/false or null), and the model returns calibrated probabilities for them. The company calls it a "System One" model, focused on fast intuition rather than reasoning, and says it was trained only on synthetic data using "reinforcement learning from calibrated decisions".

Because outputs are constrained to the options the user defines, the model cannot invent free-form answers. The user must still decide how to act on the confidence scores. Reported use cases include safety classification of commands, email triage, content moderation, file sorting, support ticket routing, model routing, and monitoring LLM agent traces. Developer tests cited by TechCrunch include Vercel, which reported results 5 to 18 times faster with greater accuracy than an LLM classifier. Another developer found Gemini slightly more accurate but 10 to 20 times more expensive. Creators also showed demos such as playing Doom and Wiki Race. It is reported to be weak at chess compared with chat models and is not intended for coding or interactive chat.

Access is through an API. Output tokens are free and input tokens are metered, with creators citing roughly 4 cents per million input tokens. Sign-ups were reported as waitlisted or paused after demand briefly overwhelmed the API. The architecture is undisclosed, and outside observers suspect it is built on an open-weight LLM. The speed (about 70-500 ms) and the "200x faster / 400x cheaper" and "0% hallucination" figures are company or creator claims.