Tag
This blog post derives Kimi Delta Attention step by step from standard softmax attention through linear attention and DeltaNet variants, explaining the state update equations used by recent Qwen and Kimi models.