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

zartbot.github.io

131 points·mfiguiere··10 comments

I am very impressed with the KV Cache Compression work as well as the prefix cacheing making queries converge on practically free.

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

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.

Send me an email- it's not public just yet

404 on the blog page? https://zartbot.github.io/blog/

Yeah, for some reason their /blog/ is 404ing but for anyone interested in their other articles, https://github.com/zartbot/blog/ has all of them, just sub out anything after /blog/.* with the folder path (e.g https://zartbot.github.io/blog/arch/jalapeno/ from https://github.com/zartbot/blog/tree/main/arch/jalapeno)

Removed

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!

you don't know him?

The writing feels human to me… and I call out AI slop as much as possible.

Reply on news.ycombinator.com