Long Context & Efficient Sequences · 2025

DeepSeek-V3.2: DeepSeek Sparse Attention (DSA)

DeepSeek-AI

Adds a lightning indexer plus fine-grained token selection on top of MLA to replace O(L^2) attention with scalable sparse attention, roughly halving per-token long-context cost at parity.

Editorial record

Plain-language summary

Adds a lightning indexer plus fine-grained token selection on top of MLA to replace O(L^2) attention with scalable sparse attention, roughly halving per-token long-context cost at parity.

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-613
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.