The field of large language models is moving towards more efficient training and inference methods, with a focus on reducing computational power and memory resources. Researchers are exploring various techniques, including quantization, novel hardware architectures, and optimized software solutions. These innovations aim to improve the scalability and accessibility of large language models, enabling wider adoption and further advancements in the field. Notable papers in this area include AxLLM, which proposes a hardware accelerator architecture for quantized models, and InfiR2, which introduces a comprehensive FP8 training recipe for reasoning-enhanced language models. Additionally, papers like Pretraining Large Language Models with NVFP4 and SAIL demonstrate significant improvements in training efficiency and inference speed, respectively. Microscaling Floating Point Formats for Large Language Models also presents a promising approach to reducing memory footprint and computational costs.