That may well be true. I know that earlier models like Llama 1 65B could tolerate more aggressive quantization, which supports that idea.