Released by Mathias Strasser (GitHub firelex) on 28 September 2026, Jeff fine-tunes Qwen3.5 0.8B/2B and Gemma 4 E2B using a full-weight AutoJev-derived recipe: cross-entropy over option letters plus a fitted temperature, not TypeSafe's hosted Jev. The project is not affiliated with TypeSafe. Its own five-benchmark panel lists Jeff-Qwen3.5-2B at 83.1% versus Jev's published 83.0% on a different sample, with much larger gaps on BBH and JevBench hard; training code is MIT.