Quantisation can be unpredictable. Some model architectures and even some specific weights just quantise better, others not so much.