arxiv How Good Are Low-bit Quantized LLaMA3 Models? An Empirical Study