A vision-language model built on Qwen3's 4B foundation, this variant has been 'abliterated' — a community fine-tuning technique that removes refusal behaviors, making it more permissive in what it will discuss or generate. It handles both text and image inputs, but the abliteration process can introduce unpredictability in outputs compared to the base model.