Do we have some studies on the impact of extreme quantization on these models? In my understanding, the efforts to make them run in commodity hardware involves the quantization of the weights to 8 bits or even 4 bits. Is the model still reliable after this?