A community-modified variant of Qwen3's architecture, this model has been through several post-processing steps — abliteration (removal of refusal behaviors), quantization, and other optimizations — that distinguish it from the base release. It accepts both text and image inputs, inheriting multimodal capabilities from its Qwen3 lineage. The layered modifications mean behavior may differ noticeably from standard Qwen3 models, particularly around content boundaries.