DeepSeek V4.1 Flash's global attention: compressed attention with three KV-reuse variants

nrehiew_ · x · 2026-09-11

nrehiew explains DeepSeek V4.1 Flash's "global attention": an updated compressed-attention scheme. All three variants run an indexer over smaller keys/values, do selection on the main KV, concatenate with windowed KV, then attend. Variants differ in:

The 20-layer encoder + 20-layer decoder design cuts KV cache in half: the bottom half builds a KV cache shared with upper half layers via per-layer projection — inspired by YOCO, hence "decoder-decoder" (SWA layers excluded).

Related event: DeepSeek V4.1 Flash Deep Dive: KV Cache Compression at the Frontier(9 posts)→

Original post →

More from Models

Models channel →