{"id":29002,"date":"2025-02-18T19:56:52","date_gmt":"2025-02-18T11:56:52","guid":{"rendered":"https:\/\/www.1ai.net\/?p=29002"},"modified":"2025-02-18T19:56:52","modified_gmt":"2025-02-18T11:56:52","slug":"deepseek-%e5%86%8d%e6%94%be%e9%99%8d%e6%9c%ac%e5%a4%a7%e6%8b%9b%ef%bc%9ansa-%e5%ae%98%e5%ae%a3%e5%8f%91%e5%b8%83%ef%bc%8c%e5%8a%a0%e9%80%9f%e6%8e%a8%e7%90%86%e9%99%8d%e4%bd%8e%e6%88%90%e6%9c%ac","status":"publish","type":"post","link":"https:\/\/www.1ai.net\/en\/29002.html","title":{"rendered":"DeepSeek Releases Another Cost-Reducing Move: NSA Announces Release, Accelerates Inference to Reduce Costs and Doesn't Sacrifice Performance"},"content":{"rendered":"<p>February 18th.<a href=\"https:\/\/www.1ai.net\/en\/tag\/deepseek\" title=\"[View articles tagged with [DeepSeek]]\" target=\"_blank\" >DeepSeek<\/a> Today's official announcement of the launch of\u00a0<strong>NSA (Native Sparse Attention)<\/strong>, which is a hardware-aligned and natively trainable sparse attention mechanism for ultra-fast long context training and inference.<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"alignnone size-full wp-image-29003\" title=\"34c13fd5j00srvnu40052d000v900ggp\" src=\"https:\/\/www.1ai.net\/wp-content\/uploads\/2025\/02\/34c13fd5j00srvnu40052d000v900ggp.jpg\" alt=\"34c13fd5j00srvnu40052d000v900ggp\" width=\"1125\" height=\"592\" \/><\/p>\n<p>The core components of the NSA include:<\/p>\n<ul>\n<li>Dynamic Hierarchical Sparse Strategy<\/li>\n<li>Coarse-grained token compression<\/li>\n<li>Fine-grained token selection<\/li>\n<\/ul>\n<p>DeepSeek officials say the mechanism optimizes modern hardware designs.<strong>Accelerating inference while reducing pre-training costs without sacrificing performance<\/strong>.. Performance is comparable to or better than the full-attention model on generic benchmarks, long context tasks, and instruction-based reasoning.<\/p>\n<p>Attached paper link:<\/p>\n<p>https:\/\/arxiv.org\/abs\/2502.11089<\/p>","protected":false},"excerpt":{"rendered":"<p>February 18, 2012 - DeepSeek today announced the launch of NSA (Native Sparse Attention), a hardware-aligned and natively trainable sparse attention mechanism for ultra-fast long context training and inference. The core components of NSA include: Dynamic Hierarchical Sparse Policies Coarse-Grained Token Compression Fine-Grained Token Selection DeepSeek says the mechanism is optimized for modern hardware designs, accelerating inference while reducing pre-training costs, without sacrificing performance. It performs as well as or better than full-attention models on generalized benchmarks, long context tasks, and instruction-based reasoning. With link to the paper: https:\/\/arxiv.org\/ab<\/p>","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[146],"tags":[3606],"collection":[],"class_list":["post-29002","post","type-post","status-publish","format-standard","hentry","category-news","tag-deepseek"],"acf":[],"_links":{"self":[{"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/posts\/29002","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/comments?post=29002"}],"version-history":[{"count":0,"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/posts\/29002\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/media?parent=29002"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/categories?post=29002"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/tags?post=29002"},{"taxonomy":"collection","embeddable":true,"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/collection?post=29002"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}