tl;dr: Jeff is a set of small (0.8B–2B) fine-tuned Qwen and Gemma models for zero-shot classification, compatible with the Jev API format, returning calibrated probabilities in ~22-30ms per decision on GPU/Apple Silicon. Trained entirely on local hardware using synthetic data from open models, it matches or beats Jev on classification/grounding benchmarks but lags on reasoning-heavy tasks. Limitations include a 26-option cap, English-only support, and no multi-step reasoning—though a short fine-tune reportedly boosted one task's accuracy from 31.7% to 95.8%.