TechFlow reported on February 18 that, according to Jinshi News, DeepSeek has introduced NSA. DeepSeek stated that NSA is a hardware-aligned and natively trainable sparse attention mechanism designed for ultra-fast long-context training and inference. Through optimization tailored for modern hardware, NSA accelerates inference speed while reducing pretraining costs, without compromising performance. It performs on par with or even better than full-attention models across general benchmarks, long-context tasks, and instruction-based reasoning.
Add to Favorites
Share to Social Media
