Teaching a small decision model to think cost it general knowledge.
PostHog released Jeeves, a nine-billion-parameter model that reasons before it labels a piece of text. On the decision benchmark it scores 0.935 against 0.866 for Jev, the model it extends. On general knowledge it falls to 0.793 where Jev reaches 0.900, and thinking adds three seconds.
Why it mattersSmall models are now tuned for one job instead of being made broadly smarter. That trade works for sorting support tickets and fails for answering real questions.
Read in the app — free, in 9 languages
Related stories
OpenAI's new model nearly matches its best at a fifth of the price.2026-09-30A model trained at home in two hours tied a paid decision service.
2026-09-29Kuaishou's video unit is now worth more than Kuaishou itself.
2026-09-29Your cloned voice now speaks 90 languages in a native accent.
2026-09-29An open 27B model now clicks through desktop apps on its own.
2026-09-29