Users with limited VRAM should not insist on high-precision quantization, but should choose a balance between inference speed and quality based on their hardware conditions.