The field of large language models (LLMs) is moving towards more efficient and scalable training and inference methods, with a focus on quantization techniques. Recent developments have shown that advanced low-bit quantization methods can achieve comparable or even superior performance to traditional full-precision methods, paving the way for more robust and scalable LLM training. Notable advancements include the development of novel quantization frameworks, hardware-efficient kernels, and dynamic grouping strategies, which have led to significant improvements in training stability, model performance, and computational efficiency. These innovations have the potential to greatly reduce the memory footprint and computational requirements of LLMs, making them more accessible and deployable in real-world applications. Noteworthy papers include: Metis, which introduces a training framework that enables stable and unbiased low-bit training, surpassing full-precision baselines. LiquidGEMM, which presents a hardware-efficient W4A8 GEMM kernel that achieves up to 2.90x speedup over state-of-the-art kernels. Binary Quantization For LLMs Through Dynamic Grouping, which proposes a novel optimization objective and algorithms for binary quantization, achieving an average bit length of just 1.007 bits while maintaining high model quality.