Kimi Delta Attention is framed as a regression rule mapping queries to keys

burny_tech · x · 2026-07-28

The post points readers to a video explaining Kimi Delta Attention and linear attention more broadly. The core takeaway is that the differences between Mamba2, Gated DeltaNet, and Kimi Delta Attention may be smaller than they look: the key question is how the state update rule is derived.

The update is described as a cumulative sum of the outer product q kᵀ, and in DeltaNet terms it is likened to an SGD step with MSE loss. In other words, the intuition is that these methods can be seen as training a regression model that maps queries to keys.

Original post →

More from Research

Research channel →