DeepSeek V4.1 Flash’s reported fourth-place sparse-attention benchmark ranking

A user reports that DeepSeek V4.1 Flash placed fourth for DeepSeek Native Sparse Attention on KernelBench-CUDA, behind Fable 5.1, Opus 5 and Fable 5. At an 8K context length, they say its kernel processes 78% of the causal attention blocks even though only about 14% are needed. They attribute the entire performance gap to that excess work. A separate commenter calls V4.1 Flash a major improvement in kernel engineering and says GLM-5.3 Flash is “nowhere close.”