⬢github Python · 490 ★ +473 since we first saw it · pushed 8 h ago · MIT
firelex/jeff
Fine-tunes of Qwen3.5 and Gemma 4 for zero-shot classification
Jeff provides small (0.8B–2B) fine-tuned Qwen3.5 and Gemma 4 models for zero-shot classification: you describe options in plain language and it returns calibrated probabilities in a single forward pass, no text generation or parsing. It runs locally in ~22–30 ms, supports choice, yes/no, and score questions, and is Jev API-compatible.
Why now: Recently released and discussed on Hacker News as home-trained, ~30 ms decision models that approach Jev's published accuracy at a tiny size, built entirely on local hardware with synthetic data.
Who it is for: Developers who need fast, local, cheap classification routing or judgement calls inside applications without cloud LLM calls.
Stars over our 32 snapshots: 17 to 490, since 7 h ago.
Where people talked about it
API: https://socialmediatrends-api.osmike.com/v1/repos/firelex/jeff