DeepSeek V4.1 Flash halves KV cache with YOCO-style decoder-decoder design

nrehiew_ · x · 2026-09-11

nrehiew explains the core design of DeepSeek V4.1 Flash: 20-layer encoder plus 20-layer decoder. Since prefill is expensive, the bottom half builds the KV cache, which is shared with upper half layers via per-layer projection — roughly a state read individually by later layers. Inspired by YOCO, hence "decoder-decoder" with both modules causally masked; this doesn't apply to the SWA layers.

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

Original post →

More from Infra

Infra channel →