Benchmarks & research
Jeff is an open 0.8B decision model that runs on your own hardware, using swappable add-ons to make quick choices in front of a larger local model.
Only you can see your notes.
Screenshot unavailable. Open the experiment ↗
It uses Jev's request format but is an independent project, not affiliated with TypeSafe. The maker reports Jeff with task adapters averaging 94.6% across eight tasks against 86.6% for Qwen3.8-27B, about 35 times faster on a Mac. Jeff-Code puts it inside a coding agent and reportedly cuts time per task by about a third at the same pass rate.
Jeff suits teams that run their own AI models and want quick yes-or-no or pick-one decisions, such as spotting spam or sorting support requests. Choose the add-on for that task, let Jeff answer when it is confident, and pass unsure cases to the bigger model. The speed and accuracy figures come from the maker’s own tests.