Intel / article

Jev introduces a new shape of LLM - System One, aka Decision Models

Source: simonwillison.net
Date: Published Sep 21, 2026

Context

Jev makes AI output easier for ordinary software to consume by returning decisions as numbers and predefined choices. Simon Willison describes TypeSafe AI's model in his assessment: an application supplies information and questions, then receives probabilities, category choices or scores. A program can use those values to label records or rank search results without extracting a decision from generated prose.

The same interface gives developers less explanatory material to inspect. Willison argues that task-specific experiments become especially important because a numerical answer does not reveal why the model chose it. The useful shift is a simpler decision interface whose behavior still needs to be measured.

Key takeaways · AI-distilled

Jev outputs only floating-point confidence scores, not text. No output means input-only pricing: $0.042 per million input tokens, undercutting GPT-5 Nano's $0.05/million input rate.
Jev supports three query types: yes/no ("Noul," named for the Bernoulli distribution per TypeSafe's CEO on Hacker News), multiple-choice with a probability distribution, and score questions along a numeric range.
Questions run in parallel against one input document, so sending hundreds of questions takes about as long as sending one. One use case: scoring 100 BM25-retrieved candidates for relevance in a search reranker.
Willison flags a bias risk: asking Jev to rate whether Bay Area cities are a 'Good city?' returned Cupertino highest and East Palo Alto lowest, with no explanation for the score - a concern for uses like ranking job applicants.
Within a week of release, clones appeared, including 'Kev' (0.8B/4B/9B models built on Qwen 3.5) and a community called JevBench for comparing decision models.

Terms in this piece · Glossary

token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.
benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.

Read the source simonwillison.net