Xiaohongshu Open-Sources DoSA, Cutting First-Token Latency by Up to 48%

Xiaohongshu's Dots team open-sourced DoSA, which accelerates the indexer bottleneck in sparse attention for long-context prefill. It cuts first-token latency by 37-48% at around 1M-token contexts while preserving the original Top-K selections losslessly.

2026-10-20 ~ 2026-10-21 · 2 related posts

• Xiaohongshu's DoSA cuts 1M-context TTFT by 37-48% by speeding up sparse-attention indexer — 机器之心 · 2026-10-20

1 near-duplicate retellings: 小红书技术REDtech