Read only the blocks that matter
Dense
NSA
DSA
pause
restart
new context
top-n blocks
4
reveal answer
cost per query, at real context lengths
context L
1M
log–log. Illustrative NSA config matching the strip above (64-token blocks, 128-token window). DSA k = 2048 as in DeepSeek-V3.2. The indexer line's height is illustrative; its slope is not.