DeepSeek v4.1 flash's insane cache hit rate shines in multi-billion-token sessions

teortaxesTex · x · 2026-09-13

Commenters highlight DeepSeek v4.1 flash's remarkably high prompt-cache hit rate in the Zcode coding harness, attributing it to DeepSeek's cache engineering — performance holds even in extremely long sessions burning billions of tokens.

Original post →

More from Infra

Infra channel →