In the field of artificial intelligence, the DeepSeek team recently released their latest research成果, introducing an innovative sparse attention mechanism named NSA (Native Sparse Attention). The core objective of this technology is to enhance the speed of long context training and inference, particularly optimized for modern hardware, significantly improving the efficiency of training and inference. The launch of NSA technology brings significant changes to the training of artificial intelligence models. Firstly, it utilizes a series of advancements tailored for modern computational hardware.