Introducing System One Models and Jev

TypeSafe AI has unveiled its inaugural System One Model, dubbed Jev, designed for rapid, structured decision-making that software can readily utilize. The groundbreaking stack encompasses a novel model architecture, a parallel sampler optimized for efficiency, and a training method known as Reinforcement Learning for Calibrated Decisions (RLCD).

Jev emerges as an impressive contender, matching the intelligence of leading LLMs on specific System One tasks while delivering two orders of magnitude in speed and efficiency. Distinct from conventional language models, Jev is fine-tuned for structured outputs and eschews string generation, effectively eliminating the risk of hallucinations.

To substantiate these claims, TypeSafe AI provides extensive evidence, including side-by-side comparisons to GPT-5.6 Terra demonstrating Jev's superiority in specific scenarios. The company also conducts rigorous evaluations to quantify the performance of AI systems within code, employing a unique approach that treats every model equally by using the workflow of the most sophisticated external models as a reference.

The results reveal Jev's exceptional capabilities, with evaluations showing it to be 193.6 times faster and 444.6 times more cost-effective than its competitors. TypeSafe AI remains transparent about the pricing of Jev, acknowledging potential subsidies but emphasizing long-term sustainability.