Jeff – Jev-compatible 0.8B decision models, trained at home, ~30 ms

Jeff is a family of very small, Jev-compatible decision models (notably a 0.8B Qwen3.5 fine-tune) that return calibrated probabilities for a set of user-provided options in a single forward pass. They perform zero-shot classification over arbitrary option lists (support tickets, intents, moderation labels, game moves) using the same request format as Jev, with three question types (choice, noul yes/no, score). Trained entirely on local hardware with synthetic data from open models, the 0.8B model trains in about two hours on one RTX PRO 6000 and runs in roughly 22 ms on that GPU and ~28 ms on an Apple M4 Max (weights ~1.7 GB). Benchmarks across five datasets show Jeff-Qwen3.5-0.8B reaching ~79% overall accuracy, approaching Jev’s published figures and sometimes beating it on classification and grounding tasks, though it lags on reasoning-heavy benchmarks.

Practical guidance emphasizes that Jeff is a fast classifier, not a planner: describe each option’s consequence and keep option keys short. Zero-shot gameplay tests (Doom, Frogger, Pac-Man) demonstrate competent, low-latency decisions (29-49 ms per move on M4), and a short fine-tune can massively improve domain accuracy (a voice-navigation fine-tune raised held-out accuracy from 31.7% to 95.8% in under 30 minutes on one GPU). Models and checkpoints are published for local use; the project is independent, open-source, and optimized for slotting calibrated, low-latency decision models directly into production code.