Gemma 4 12b unified as a q6 gguf fits with about 24k context inside your vram. koboldcpp+sillytavern works well on the software side.