Jeff – Jev-compatible 0.8B decision models, trained at home, ~30 ms
Jeff is pitched as a small front-end decider that can skip most 27B calls while improving accuracy on the project’s adapter tests.
The repo says Jeff v1.2 is a 0.8B open “System 1” model that returns calibrated option probabilities, not generated text. With eight adapters in the authors’ comparison, Jeff plus fallback to Qwen3.8-27B scored 95.3% versus 86.6% for the 27B alone, with mean decision time falling from 8.1 seconds to 0.25 seconds. The nine LoRA adapters are about 41 MB each and cover jobs like guard, triage, tool choice, grounding, spam, and legal clauses. The authors also warn that adapters are tied to a base version, small models do not reason, and v1.3 will require retrained adapters. HN · GitHub AI's note
The repo says Jeff v1.2 is a 0.8B open “System 1” model that returns calibrated option probabilities, not generated text. With eight adapters in the authors’ comparison, Jeff plus fallback to Qwen3.8-27B scored 95.3% versus 86.6% for the 27B alone, with mean decision time falling from 8.1 seconds to 0.25 seconds. The nine LoRA adapters are about 41 MB each and cover jobs like guard, triage, tool choice, grounding, spam, and legal clauses. The authors also warn that adapters are tied to a base version, small models do not reason, and v1.3 will require retrained adapters. HN · GitHub AI's note
score 6
Discussions
- hn · 575 points · 225 comments