Long Context & Efficient Sequences · 2025

Native Sparse Attention: Hardware-Aligned and Natively Trainable Sparse Attention (NSA)

Jingyang Yuan, Huazuo Gao, Damai Dai, et al. (DeepSeek-AI)

Makes sparse attention trainable end-to-end and arithmetic-intensity-balanced for GPUs, cutting long-context train and decode cost without the quality loss of post-hoc sparsity.

Editorial record

Plain-language summary

Makes sparse attention trainable end-to-end and arithmetic-intensity-balanced for GPUs, cutting long-context train and decode cost without the quality loss of post-hoc sparsity.

Knowledge graph

Relationships

Descendants

  • Depends onEvidence: Strongly supported

    Attention Is All You Need

    Builds on the long context lineage in the archive

    freshness sweep 2026

Source record

Provenance

Record ID
P-606
Record created
2026-07-13
Last reviewed
2026-07-14
Record version
2

Citation caveat: Citation metadata is approximate and marked unverified in the source dataset.