Revision History
Native Sparse Attention (NSA) · 5 revisions
Sizes are character counts of the article source. The signed number is the change from the previous revision.
Recent edit summaries
Detailed summaries recorded by editors. Generic maintenance summaries are omitted here; every recorded revision remains below. Dates describe the edit, not necessarily the event it covers.
- Correction: decoding speedup (11.6x) is an expected figure from KV-cache memory access, not a measured FlashAttention comparison; baseline is Triton FlashAttention-2 on A100s at 64k (65,536) tokens; DSA no longer called NSA's 'research predecessor' (DeepSeek's reports cite NSA only for kernel-level KV sharing); API price cut is 50%+. Update: full method notation, training setup, all result tables, rejected alternatives, limitations, implementations and dated follow-up work.
Version 5 · Oct 31, 2026, 03:09 AM
- Improve internal links (2026-10-13)
Version 4 · Oct 13, 2026, 10:12 AM
- Added 6 contextual internal links
Version 3 · Jul 23, 2026, 04:02 PM