72.4× prefill faster speedup on Qwen-3.8-27B.
This is a dev post classified by AGK Intelligence as AI dev tools (a tool drop), kept by the Dev Radar because it carries real work, not commentary.
72.4× prefill faster speedup on Qwen-3.8-27B. Open-sourced Latent Sparse Attention (LSA) hierarchical clustering that finds the right keys without scanning the entire context. you know, At 10M tokens, 98% of sparse attention’s time is wasted just finding which tokens to attend to. LSA fixes this, Up to 22.7× faster than existing sparse attention and 72.4× faster than full attention, with accuracy that stays on par. Matches full attention accuracy. Paper, code, kernels out now. https://x.com/jcrwhittington/status/2117317159777177602/video/1
Posted by Md Ismail Šojal 🕷️ (64.9k followers) 1 h ago · 17 likes · 1.1k views · view the original post on X. Kept by the Dev Radar as AI dev tools.
More dev work like this
- A New 65B dense model just Beats 900B models on agentic coding & You can run locally. — @0x0SojalSec
- This is gonna suck performance wise, and cost-wise to operate because it's built with… — @gakonst
- fal CTO @gorkem reveals a new way to build video games: — @MTSlive
- Exciting news: @Kimi_Moonshot has released the open weights for Kimi K3.1 Pro, now #1… — @arena
- We started this project only 6 weeks ago. It's the first model we trained with a… — @silasalberti
- Grok Build v1.0.75 is out with a few practical updates. Standouts include permission… — @blankspeaker
- Want to learn how AI systems are actually trained, deployed, and optimized? — @sgl_project
- Intelligence company Block has introduced Builderlab, an enterprise intelligence toolkit… — @entrepreneur_cm
Every post is read and classified by AGK Intelligence: what it is, which market it belongs to, and whether the link is a real tool. 31.4k posts from 5k X accounts over the last 21 days, 9.3k tools, 12 markets. New posts are collected several times an hour and the full ranking is recomputed every hour — last update 2026-11-04 21:57 UTC. Full method.