Two Minute Papers · 2026-09-18 · notable
Two Minute Papers — 'DeepSeek Just Made AI Memory 4x Smaller!'
Two Minute Papers posted 'DeepSeek Just Made AI Memory 4x Smaller!' on 18 September 2026. The subject is DeepSeek-V4.1-Flash's KV cache work, which cuts the global cache to 890 bytes per token — about a quarter of DeepSeek-V4-Flash.
Two Minute Papers walks through DeepSeek-V4.1-Flash's KV cache compression in an episode posted on 18 September 2026.
What is it?
'DeepSeek Just Made AI Memory 4x Smaller!' went up on the Two Minute Papers channel on 18 September 2026. The paper behind the title is 'DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression', submitted to arXiv by DeepSeek-AI on 17 September 2026, which sat second on Hugging Face's trending papers the next day.
How does it work?
The method in the paper is called Compressed Sparse Attention 2, or CSA2, and it pairs cross-layer reuse of the KV cache with storing that cache in FP4. Together they bring the global KV cache down to 890 bytes per token, roughly a quarter of DeepSeek-V4-Flash. A separate trick the authors call SWA Bounded Replay takes the persistent cache to about an eighth of the earlier model.
Why does it matter?
The KV cache decides how many long conversations one GPU can hold at once, so shrinking it raises throughput without buying new hardware. DeepSeek reports that the smaller footprint still delivers substantially better performance than the baseline it is measured against, which is the part that makes the result interesting rather than just cheap. A channel that explains papers to a general audience puts the idea in front of people who will never open an arXiv page.
Who is it for?
people serving LLMs at scale