conv.

All stories
AIQuiet 10d · day 10

Blogger dissects DeepSeek-V4.1 Flash's 4x KV cache compression, calls it V5 in disguise

A detailed technical teardown argues DeepSeek's 'Flash' update is really a major architecture overhaul built around extreme KV cache compression for long-context agents.

newest coverage 9d agosummary written 9d ago5 pieces127 likes & upvotes10 comments MastodonNewswiresHacker News

What to know

  • Outside analysis frames DeepSeek-V4.1 Flash as a substantially larger architectural change than its version number implies, effectively a V5-class model.
  • The core technical goal is compressing KV cache 4x to relieve HBM/SSD storage and interconnect bandwidth bottlenecks created by long-horizon agent workflows.
  • DeepSeek reportedly took its DeepSeek-V4 Pro models offline around the same time, reinforcing the sense of a major transition.
  • The design is compared to a Recursive Transformer architecture that modifies queries while reusing key-value cache across recursive steps.

DeepSeek AI lab developing the modelCui DeepSeek-affiliated figure cited in the blog postzartbot Independent technical blogger

How it unfolded 1 development · click the chart to see its coverage articlesposts

Peak 4 pieces in 3h at Sep 16, 9 PM; 5 pieces over 10 days (1 article · 4 posts) Sep 16, 9 PM — 4 pieces · 1 article · 3 posts — Mastodon 2, Hacker News 1, Newswires 1Sep 17, 12 AM — 1 piece · 1 post — Mastodon 1Sep 17, 3 AM — quietSep 17, 6 AM — quietSep 17, 9 AM — quietSep 17, 12 PM — quietSep 17, 3 PM — quietSep 17, 6 PM — quietSep 17, 9 PM — quietSep 18, 12 AM — quietSep 18, 3 AM — quietSep 18, 6 AM — quietSep 18, 9 AM — quietSep 18, 12 PM — quietSep 18, 3 PM — quietSep 18, 6 PM — quietSep 18, 9 PM — quietSep 19, 12 AM — quietSep 19, 3 AM — quietSep 19, 6 AM — quietSep 19, 9 AM — quietSep 19, 12 PM — quietSep 19, 3 PM — quietSep 19, 6 PM — quietSep 19, 9 PM — quietSep 20, 12 AM — quietSep 20, 3 AM — quietSep 20, 6 AM — quietSep 20, 9 AM — quietSep 20, 12 PM — quietSep 20, 3 PM — quietSep 20, 6 PM — quietSep 20, 9 PM — quietSep 21, 12 AM — quietSep 21, 3 AM — quietSep 21, 6 AM — quietSep 21, 9 AM — quietSep 21, 12 PM — quietSep 21, 3 PM — quietSep 21, 6 PM — quietSep 21, 9 PM — quietSep 22, 12 AM — quietSep 22, 3 AM — quietSep 22, 6 AM — quietSep 22, 9 AM — quietSep 22, 12 PM — quietSep 22, 3 PM — quietSep 22, 6 PM — quietSep 22, 9 PM — quietSep 23, 12 AM — quietSep 23, 3 AM — quietSep 23, 6 AM — quietSep 23, 9 AM — quietSep 23, 12 PM — quietSep 23, 3 PM — quietSep 23, 6 PM — quietSep 23, 9 PM — quietSep 24, 12 AM — quietSep 24, 3 AM — quietSep 24, 6 AM — quietSep 24, 9 AM — quietSep 24, 12 PM — quietSep 24, 3 PM — quietSep 24, 6 PM — quietSep 24, 9 PM — quietYesterday, 12 AM — quietYesterday, 3 AM — quietYesterday, 6 AM — quietYesterday, 9 AM — quietYesterday, 12 PM — quietYesterday, 3 PM — quietYesterday, 6 PM — quietYesterday, 9 PM — quietToday, 12 AM — quietToday, 3 AM — quietToday, 6 AM — quietToday, 9 AM — quietToday, 12 PM — quietToday, 3 PM — quiet 1
Sep 17Sep 18Sep 19Sep 20Sep 21Sep 22Sep 23Sep 24yesterdaynow · 5:15 PM ET
  1. 1

    Blogger zartbot argues the model is effectively DeepSeek-V5 Flash

    A detailed architecture write-up interprets the report's compression-and-editing (CED) design as a Recursive Transformer variant that modifies the query while reusing KV across recursive steps, and shares it on Hacker News and Mastodon.

    “It can be seen as a kind of Recursive Transformer architecture, a way of modifying Q and reusing KV during the recursive process.”
    — zartbot
    1. 1 outlet DeepSeek-v4.1 Flash: Pushing the Limits of KV Cache Compression

      first by HN Frontpage, 9d ago

  2. background

    DeepSeek publishes technical report on KV cache compression — DeepSeek released the technical report "DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression," describing architectural work to compress KV cache 4x to ease HBM/SSD storage and interconnect bandwidth pressure from long-horizon agent workflows.

  3. background

    DeepSeek-V4.1 Flash ships, V4 Pro models pulled offline — After DeepSeek-V4.1 Flash's release, a blogger noticed inference speeds near 420 tokens/s and reported that DeepSeek figure Cui said all DeepSeek-V4 Pro models would be taken offline, signaling a bigger change than a routine update.