Jev introduces a new shape of LLM: System One decision models

Last week TypeSafe AI unveiled Jev, described by the company as the first example of a new model category it calls “System One models.” Simon Willison, reporting the announcement, noted he prefers an alternative name suggested by Maggie Appleton on social media, “decision models,” calling it a better fit for the category.

According to Willison’s account, Jev departs from the usual large language model format in one key respect. It still accepts text inputs in the ordinary way, but rather than generating output token by token as a stream of text, the model is built to produce decisions directly. The available material stops short of detailing the mechanism or the training approach behind that shift, and Willison’s post does not elaborate on benchmarks, release timing, or availability.

The provided source material does not include commentary from TypeSafe AI beyond the naming of the category, nor does it specify what domains or tasks the company intends Jev to serve. No pricing, licensing, or technical specification details appear in the excerpt.

Willison frames the development as noteworthy chiefly for its departure from the standard LLM output format, rather than for any stated performance claim. Appleton’s proposed naming, “decision models,” is presented as a competing label for the same category, offered publicly in response to TypeSafe AI’s own terminology. No further detail from either party is given in the source material reviewed.