Feb. 19, 2025
DeepSeek and Moonshot AI Unveil New Attention Mechanisms: NSA and MoBA
TMTPOST -- DeepSeek on Tuesday released a new paper detailing an improved attention mechanism called NSA. This groundbreaking research was spearheaded by DeepSeek's founder and CEO, Liang Wenfeng. On the same day, another paper with a similar theme was published by Moonshot AI, a tech company led by founder and CEO Kimi Yang Zhilin. Yang is also a co-author of this paper. The focus of their paper is MoBA, an innovative attention mechanism that integrates the principles of Mixture of Experts (MoE) into the attention process. According to the paper, MoBA adheres to a "less structure" principle, meaning it doesn't impose predefined biases. Instead, the model autonomously determines which positions to focus on, offering greater flexibility and adaptability in processing information.
More News

  • Subscribe To Our News