Decision models

Models that answer with a probability for every option instead of a paragraph β€” ranked by JevBench on smarts, calibration, speed and cost per 1,000 decisions.

A decision model reads a state β€” a log, a diff, a ticket, a web page β€” and a set of typed questions: pick one of these options, yes or no, or where on this scale. One forward pass returns a probability for every option. Nothing is generated, so there's no text to parse and no answer outside the options you gave it.

They sit inside agent loops where a chat model would be slow and expensive: routing, triage, "did the tests pass", "which button next", "is this safe to run". Most are small open-weight fine-tunes you can host yourself; the original is TypeSafe AI's hosted Jev.

β€œCheckout has been returning 503 for twelve minutes. Rollback is staged.”
What should on-call do next?
rollback_now99.1%
page_database0.7%
keep_investigating0.2%
close_ticket0.0%
Real answer from decider-4b on our DGX Spark, 86 ms.

πŸ’£ At β€” the Per month column shows what that volume costs.
Kind Decision models Chat LLMs Classifiers
Weights Open Closed
Size < 1B 1–4B 5–14B 15B+
Model ↕ Score ↕ Intelligence ↕ Calibration ↕ Per 1k decisions ↕ Per month ↕ Speed p50 ↕