“The DeepSeek and MiMo models use sparse attention mechanisms.”