trained on 600K curated RL samples from 59 datasets across 6 visual reasoning categories.
https://huggingface.co/zlab-princeton/Vero-Qwen3I-8B
Original is quite good at all, including nsfw content.
now + heretic for me is best (small model) for image description (2026).
Vero-Qwen3I-8B-heresy-ara
2nd version vero-ara + nsfw merged with 0.8 weights lora from
https://huggingface.co/thesby/Qwen3-VL-8B-NSFW-Caption-V4.5 (original source deleted)
https://huggingface.co/mradermacher/Qwen3-VL-8B-NSFW-Caption-V4.5-GGUF
(maybe a bit more understanding i didn't tested that much)
Vero-Qwen3-8B_VL_Heresy_ARA_NSFW
This is a decensored version of zlab-princeton/Vero-Qwen3I-8B, made using Heretic v2.1.0 with the Arbitrary-Rank Ablation (ARA) method
fewer refusals (6/100 Uncensored vs 89/100 Original) while preserving very good model quality (0.004 KL divergence)