DeepSeek V4.1 Flash resets architecture, shrinking KV cache to a quarter
DeepSeek released V4.1 Flash: despite being roughly twice the size of its predecessor, working KV cache shrinks to 1/4 and persistent cache storage to 1/8, with measured speeds near 420 tokens/s. Commenters call it a masterwork of AI engineering; zartbot's architecture deep-dive argues it deserves the name "DeepSeek-V5 Flash". By proactively deleting its own older designs, the model resets the architecture and cuts cache holding costs for long agent sessions, though the community remains puzzled by the 190B+ engram lookup table.
Confirmed
- Zhihu author Exhalation's teardown: total and activated parameters both grew, yet working KV cache dropped to 1/4 and persistent cache to 1/8; bendee983 reports per-token KV storage falling from 3514 bytes to 890 bytes.
- Core approach: removing old modules, sharing KV state, and recomputing local context.
- The model uses a detached encoder-decoder architecture with strong prefill optimization and a very large KV cache.
- Measured speed near 420 tokens/s.
- Blogger zartbot published a long architectural analysis of the technical report, deeming it worthy of the name "DeepSeek-V5 Flash".
Unconfirmed
- Poster NoAfternoon4260 asked what roles the 500B+ weights and the 190B+ engram lookup table each play; no authoritative answer yet.
Why it matters
- bendee983 notes lower cache holding costs for long agent sessions, a direct win for long-context agent applications.
- teortaxesTex argues DeepSeek keeps pushing the quadratic component of attention upward into more refined operations—not incremental patching but embracing the "bitter lesson", searching for irreducible minimal primitives of global computation within its own scientific framework.
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 ·
- DeepSeek-V4.1-Flash cuts KV cache 4x with split encoder-decoder architecture — bendee983 ·
- DeepSeek V4.1 Flash's 190B Engram Lookup Table Confuses Builders: Can It Store Knowledge Without Fine-Tuning? — No_Afternoon_4260 ·
- [source] 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
- 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