Definition

Attention mechanism assigning each layer a static mode (Full, Reindex, or Reuse) to share main KV and sparse-attention indices across layers, reducing memory footprint.

Key Points

  • deepseek-v41-flash uses pure CSA2 with Hierarchical Sparse Indexer bounding deeper indexer cost independent of context length
  • Combined with FP4 KV caching (E2M1 format) achieves 890 bytes/token global KV
  • ~4x runtime and ~8x persistent KV storage reduction vs DeepSeek-V4-Flash per official claims

Sources