DeepSeek V4.1 Flash architecture reset cuts KV cache to a quarter

DeepSeek has released V4.1 Flash, which—despite the model being roughly twice the size of its predecessor—compresses the working KV cache to 1/4 and persistent cache storage to 1/8, with measured speeds approaching 420 tokens/s. Multiple independent commentators call it "a masterpiece of AI engineering and research," and one architecture analyst even argues it deserves to be called "DeepSeek-V5 Flash." The model achieves an architectural reset by actively deleting its own older designs, sharply cutting cache-holding costs for long agent sessions, though the community remains puzzled by the mechanics of the 190B+ engram lookup table.

Confirmed

Zhihu author Exhalation's teardown shows that although V4.1 Flash has more total and activated parameters than the previous generation, the working KV cache dropped to 1/4 and persistent cache storage to 1/8 (bendee983 puts it as KV storage falling from 3514 bytes per token to 890 bytes).
Core approach: remove old modules, share KV state, and recompute local context.
The model uses a detached encoder-decoder architecture with extremely strong prefill optimization and a very large KV cache.
Measured speed approaches 420 tokens/s.
Blogger zartbot published a long-form architecture analysis of the technical report, arguing it's substantial enough to be called "DeepSeek-V5 Flash."

Not yet confirmed

Poster NoAfternoon4260 asked (help-seeking) what the 500B+ weights and the 190B+ "engram lookup table" each do; this core question remains unanswered authoritatively.

Why it matters

bendee983 points out that lower cache-holding costs for long agent sessions are a direct win for long-context agent applications.
teortaxesTex's take: DeepSeek keeps pushing the quadratic parts of attention upward and compressing them into more refined operations—not incremental patching, but a frank embrace of the "bitter lesson"—searching for the irreducible minimal primitives of global computation within its own scientific framework, an exploratory architecture research path.

2026-09-15 ~ 2026-09-17 · 5 related posts

Primary sources

DeepSeek V4.1 Flash Is an Architecture Reset: KV Cache Cut to 1/4 by Deleting Its Own Ideas — teortaxesTex · 2026-09-16 19:11 (UTC+08:00)
DeepSeek-V4.1-Flash cuts KV cache 4x with split encoder-decoder architecture — bendee983 · 2026-09-17 15:29 (UTC+08:00)
DeepSeek keeps moving the quadratic component up, a 'bitter lesson' science play — teortaxesTex · 2026-09-16 20:14 (UTC+08:00)

DeepSeek V4.1 Flash's 190B Engram Lookup Table Confuses Builders: Can It Store Knowledge Without Fine-Tuning? — No_Afternoon_4260 · 2026-09-15
[source] DeepSeek V4.1 Flash Is an Architecture Reset: KV Cache Cut to 1/4 by Deleting Its Own Ideas — teortaxesTex · 2026-09-16
[source] DeepSeek keeps moving the quadratic component up, a 'bitter lesson' science play — teortaxesTex · 2026-09-16
DeepSeek-V4.1 Flash deep dive: pushing KV cache compression to the limit at 420 tok/s — teortaxesTex · 2026-09-17
[source] DeepSeek-V4.1-Flash cuts KV cache 4x with split encoder-decoder architecture — bendee983 · 2026-09-17