According to @_avichawla on Twitter, DeepSeek’s V3.2 introduces DeepSeek Sparse Attention (DSA) that reduces attention complexity from O(L²) to O(Lk) by selecting only the top‑k key‑value pairs per query, capped at 2048 tokens regardless of a 128K context. As reported by @_avichawla, a lightweight Lightning Indexer ranks salient tokens using a small number of FP8 heads, enabling a compute‑cheap preselection step before running the expensive attention on the subset. According to the tweet, this design concentrates GPU FLOPs on useful tokens, offering lower latency and cost for long‑context inference and enabling scalable retrieval‑augmented generation and document intelligence workloads. As reported by the same source, the fixed k makes memory and compute predictable, which can translate into higher throughput per GPU and improved serving economics for enterprise long‑context applications.

According to @deepseek_ai, DeepSeek introduced token-wise compression combined with DeepSeek Sparse Attention (DSA) to deliver world-leading long‑context efficiency with sharply reduced compute and memory costs, and set 1M tokens as the default context across all official services. As reported by DeepSeek’s official announcement on X, the structural innovations target lower latency and lower total cost of ownership for long-context workloads such as multi-document RAG, long-form codebases, and enterprise archives. According to the same source, the move standardizes million-token windows for production, creating business opportunities for enterprises to consolidate retrieval, summarization, and compliance audit pipelines into a single pass, potentially cutting inference spend and hardware footprint.