Two Minute Papers · 2026-07-07 · notable
Two Minute Papers: 'DeepSeek's New AI Speed Hack Is Amazing'
Two Minute Papers walks through DeepSeek's newest inference speedup, positioning it as another round of the sparse-attention plus speculative-decoding tricks that let V4 serve 1M-token contexts without blowing up serving cost.
Károly Zsolnai-Fehér reviews DeepSeek's newest inference optimization on Two Minute Papers, arguing it changes the cost math of long-context serving.
What is it?
This Two Minute Papers video, uploaded 7 July 2026, reviews DeepSeek's newest inference-speed work — the same lineage as DeepSeek Sparse Attention in V3.2-Exp / V4 and the DSpark speculative-decoding release from late June 2026. Két perce Kártya-style: pull the paper, cover the pretty pictures, land on the throughput claim.
How does it work?
DeepSeek's recent speed hacks all target the same bottleneck: long-context inference cost. Sparse attention drops the per-token cost from O(L²) to roughly O(L·k) by scoring tokens with a small indexer and computing full attention only over the top-k, and DSpark stacks speculative decoding on top so a small draft model proposes many tokens per V4 verification pass. Two Minute Papers walks through those mechanics with the paper's own benchmark charts.
Why does it matter?
Long-context serving cost is what decides whether a 1-million-token window is a demo feature or a real product. Making sparse attention and speculative decoding land in the same production stack, on open weights, is exactly the pressure that pushed the recent Wafer/Meituan/Zhipu numbers into the open — and Two Minute Papers is the channel that translates the paper for practitioners who won't read it.
Who is it for?
Inference engineers, ML researchers, anyone tracking open-weight model serving cost
Try it
Watch: https://www.youtube.com/watch?v=1yBU41auQhw