Jeff – Jev-compatible 0.8B decision models, trained at home, ~30 ms
Summary
Jeff is an open-source project that delivers Jev-compatible 0.8B decision models trained entirely on local hardware, designed for zero-shot classification with very fast per-decision times. It showcases a spectrum of small models (0.8B and 2B) trained with synthetic data, a local pipeline, and benchmarks comparing against Jev, with code and weights published under MIT/Apache licenses. The repo includes setup instructions, benchmarking results, and a clear emphasis on on-device inference without cloud GPUs.