Jev AI Model Claims 194x Speed, 445x Cost Cut

TypeSafe AI, a company founded by a former ChatGPT engineer, released a model called Jev this week that never generates a single word of text. Send it a state and a set of typed questions, and it returns typed decisions with probabilities attached, the kind of output a piece of software can act on directly, no parsing or prompt engineering required. The company says the model runs 194 times faster and 445 times cheaper than comparable text-based large language models, a claim tied to its decision to skip token generation entirely.

Jev is available now through a hosted API in early access, gated behind a waitlist. TypeSafe hasn't released model weights, a parameter count, or any option to self-host. The company calls Jev a "System One Model," a label borrowed from Daniel Kahneman's distinction between fast intuitive judgment and slow deliberate reasoning.

TypeSafe argues that RLHF, the technique behind most chat models, optimized for human preference rather than calibrated accuracy, leaving models overconfident and prone to dropping viable answers. Jev instead relies on a new training approach the company calls Reinforcement Learning for Calibrated Decisions, paired with an architecture TypeSafe has not disclosed.