Qwen3-VL fits 12GB local rigs

The real story here is that “NSFW-capable” is less about a magic model and more about choosing an open-weight vision stack that fits your hardware and frontend. Qwen3-VL looks like a plausible fit for a 12GB card, but image generation is still a separate model class.

• –The official GGUF card lists Q4_K_M at about 5.03 GB and Q8_0 at about 8.71 GB, so a 12GB GPU is workable with quantization.
• –Qwen3-VL is image-text-to-text, not an image generator; if the user wants generation too, they’ll need a separate diffusion or T2I model.
• –The thread’s symptom looks like integration friction: the model card says to use the latest llama.cpp stack and load the vision `mmproj` file correctly.
• –For “NSFW + usual stuff,” the checkpoint policy and UI filters matter as much as model size; open weights help, but they do not guarantee uncensored behavior.
• –In practical terms, this is a local-inference stack question more than a single-model question.