Thread

DeepSeek-v4.1 Flash: Pushing the Limits of KV Cache Compression

131 points · 10 comments · mfiguiere

arikrahman · 1789614756 · focus · HN ↗
I am very impressed with the KV Cache Compression work as well as the prefix cacheing making queries converge on practically free.
smy20011 · 1789615547 · focus · HN ↗
Removed
1. sebmellen · 1789616453 · focus · HN ↗
  The writing feels human to me… and I call out AI slop as much as possible.
2. girvo · 1789617325 · focus · HN ↗
  The website design definitely is, but I don’t know if the content is? This reads pretty human to me and is quite interesting to boot!
3. yunfei · 1789617799 · focus · HN ↗
  you don't know him?
vivzkestrel · 1789617370 · focus · HN ↗
404 on the blog page? <a href="https://zartbot.github.io/blog/" rel="nofollow">https://zartbot.github.io/blog/
1. sanufar · 1789621206 · focus · HN ↗
  Yeah, for some reason their /blog/ is 404ing but for anyone interested in their other articles, <a href="https://github.com/zartbot/blog/" rel="nofollow">https://github.com/zartbot/blog/ has all of them, just sub out anything after /blog/.* with the folder path (e.g <a href="https://zartbot.github.io/blog/arch/jalapeno/" rel="nofollow">https://zartbot.github.io/blog/arch/jalapeno/ from <a href="https://github.com/zartbot/blog/tree/main/arch/jalapeno" rel="nofollow">https://github.com/zartbot/blog/tree/main/arch/jalapeno)
N_Lens · 1789620047 · focus · HN ↗
[dead]
mmastrac · 1789625462 · focus · HN ↗
I've been working with an automatic incremental context compactor enabled and it's been surprisingly helpful. It was particularly effective with DS41f - I think I was running at an effective session length of 5M, with the model running around 300k-400k and it was holding on both speed and intelligence.
TBH I also ran the 400tok/s preview and that was just nuts. I just let the thing compact over and over over the course of a day attacking a couple of tough problems
1. nchmy · 1789642769 · focus · HN ↗
  Can you share a link to thr automatic compactor? I've been noticing that when I get to around 60% context window, the cache will simply break and suddenly I've paid 50x more than expected. The only solution seems to be to compact or start a new session.
  1. mmastrac · 1789651746 · focus · HN ↗
    Send me an email- it's not public just yet

Open on Hacker News to reply ↗

Unofficial Hacker News client; not affiliated with Y Combinator.