Ozan-12B
Open-weights language model, 2026
A creative-writing model fine-tuned from Mistral-Nemo with QLoRA and anti-slop DPO, on $40 of GPU. It has the lowest slop score of any 12B measured on EQ-Bench and beats bigger models like Gemma 27B. The weights, the quants, and the full training recipe are public.
#1 trending on Hugging Face · 2.3K downloads in month one