{"id":13273,"date":"2026-10-13T05:07:52","date_gmt":"2026-10-12T21:07:52","guid":{"rendered":"https:\/\/infernews.com\/blog\/visionweave\/"},"modified":"2026-10-13T05:09:09","modified_gmt":"2026-10-12T21:09:09","slug":"visionweave","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/visionweave\/","title":{"rendered":"VisionWeave \u8b93\u591a\u6a21\u614b\u6a21\u578b\u6309\u756b\u9762\u5206\u914d\u7d30\u7bc0"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/10\/VisionWeave.jpg\" alt=\"VisionWeave logo\"><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u9577\u5f71\u7247\u3001\u5716\u8868\u548c\u6587\u5b57\u5bc6\u96c6\u756b\u9762\u4e0d\u9700\u8981\u6bcf\u500b\u4f4d\u7f6e\u90fd\u7528\u540c\u6a23\u7cbe\u7d30\u5ea6\u8655\u7406\u3002VisionWeave \u5c6c\u65bc\u591a\u6a21\u614b\u6a21\u578b\uff08MLLMs\uff09\u6548\u7387\u512a\u5316\u9805\u76ee\uff0c\u8b93\u6a21\u578b\u6309\u8996\u89ba\u5167\u5bb9\u6c7a\u5b9a\u4fdd\u7559\u7d30\u7bc0\u6216\u58d3\u7e2e\u8868\u793a\uff0c\u8655\u7406\u8996\u89ba tokens \u904e\u591a\u6240\u9020\u6210\u7684\u6210\u672c\u8207\u5ef6\u9072\u554f\u984c\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">VisionWeave \u4e0d\u662f\u56fa\u5b9a\u6bd4\u4f8b\u522a\u6e1b tokens\uff0c\u800c\u662f\u628a fine-grained \u8207 coarse-grained representations \u7de8\u7e54\u8d77\u4f86\uff0c\u8b93\u91cd\u8981\u5340\u57df\u4fdd\u7559\u8f03\u591a\u8cc7\u8a0a\u3001\u80cc\u666f\u4f7f\u7528\u8f03\u7dca\u6e4a\u7684\u8868\u793a\u3002\u9805\u76ee\u91dd\u5c0d dense Qwen3.5 \u63d0\u4f9b\u5f71\u50cf\u53ca\u5f71\u7247 inference\uff0c\u8a55\u6e2c\u5247\u6db5\u84cb Qwen3.8-27B \u8207 Qwen3.5-4B\uff0c\u4e26\u6bd4\u8f03 FastV\u3001VisionZip \u7b49\u56fa\u5b9a\u9810\u7b97\u65b9\u6cd5\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Qwen3.8-27B \u5728\u516b\u9805 benchmark \u5e73\u5747\u7bc0\u7701 43.0% visual tokens\uff0c\u540c\u6642\u4fdd\u7559 98.9% \u539f\u751f\u6a21\u578b\u8868\u73fe\u3002\u65bc SGLang\u3001\u5169\u5f35 A100-80GB \u548c BF16\u3001TP=2 \u689d\u4ef6\u4e0b\uff0cLongVideoBench throughput \u7531\u6bcf\u5206\u9418 0.90 \u500b\u8acb\u6c42\u5347\u81f3 2.07 \u500b\uff0c\u5e73\u5747 TTFT \u4ea6\u7531 169.90 \u79d2\u964d\u81f3 77.54 \u79d2\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u6309\u756b\u9762\u5167\u5bb9\u81ea\u52d5\u8abf\u7bc0\u8996\u89ba\u8868\u793a\u7c92\u5ea6<\/li>\n<li>\u9069\u5408\u9577\u5f71\u7247\u3001\u81ea\u7136\u5834\u666f\u3001\u5716\u8868\u53ca\u6587\u5b57\u5bc6\u96c6\u5f71\u50cf<\/li>\n<li>\u5728\u591a\u9805\u4efb\u52d9\u8207\u4e0d\u540c\u8f38\u5165\u89e3\u6790\u5ea6\u4e0b\u8861\u91cf\u6548\u7387\u8207\u8cea\u7d20\u53d6\u6368<\/li>\n<li>README \u63d0\u4f9b\u7684 checkpoint \u9700\u53e6\u884c\u53d6\u5f97\uff0c\u5b8c\u6574\u5b89\u88dd\u6d41\u7a0b\u5728\u76ee\u524d\u8cc7\u6599\u4e2d\u672a\u4ea4\u4ee3<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u7814\u7a76\u5718\u968a\u540c\u6642\u63d0\u4f9b image\/video inference\uff0c\u4f46\u5132\u5b58\u5eab\u8cc7\u8a0a\u672a\u5217\u51fa\u4e0b\u8f09\u9023\u7d50\u3001\u786c\u4ef6\u6700\u4f4e\u8981\u6c42\u6216\u5b8c\u6574\u6e2c\u8a66\u6b65\u9a5f\uff0c\u56e0\u6b64\u8f03\u9069\u5408\u4f5c\u70ba\u7814\u7a76\u8a55\u4f30\u548c SGLang serving \u5be6\u9a57\u8d77\u9ede\u3002\u5c0d\u9700\u8981\u8655\u7406\u9577\u5f71\u7247\u3001\u591a\u5f35\u5f71\u50cf\u6216\u9ad8\u89e3\u6790\u5ea6\u8f38\u5165\u7684\u6a21\u578b\u5718\u968a\uff0cVisionWeave \u63d0\u4f9b\u4e86\u6bd4\u56fa\u5b9a token \u9810\u7b97\u66f4\u9748\u6d3b\u7684\u58d3\u7e2e\u65b9\u5411\uff1b\u4ee3\u50f9\u662f\u8981\u914d\u5408\u6307\u5b9a Qwen \u6a21\u578b\u8207\u984d\u5916 checkpoint\uff0c\u672a\u5fc5\u80fd\u76f4\u63a5\u5957\u7528\u5230\u5176\u4ed6\u67b6\u69cb\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/huggingface.co\/papers\/2610.07987\" rel=\"noopener noreferrer\"><strong>\u9805\u76ee\u4e3b\u9801<\/strong><\/a> \u00b7 <a href=\"https:\/\/github.com\/FFY0\/VisionWeave\" rel=\"noopener noreferrer\"><strong>GitHub<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>VisionWeave \u6309\u5167\u5bb9\u4fdd\u7559\u91cd\u8981\u8996\u89ba\u7d30\u7bc0\uff0c\u6e1b\u5c11\u5197\u9918\u8996\u89ba tokens\uff0c\u517c\u9867\u9577\u5f71\u7247\u7406\u89e3\u6548\u7387\u8207\u4efb\u52d9\u8cea\u7d20\u3002<\/p>\n","protected":false},"author":8,"featured_media":13272,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"wpai_generated_summary":"","wpai_meta_description":"","footnotes":""},"categories":[133,119,185,31,30,150,199],"tags":[],"class_list":["post-13273","post","type-post","status-publish","format-standard","hentry","category-133","category-119","category-qwen","category-video","category-image","category-150","category-dataset-"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/13273","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/users\/8"}],"replies":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/comments?post=13273"}],"version-history":[{"count":1,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/13273\/revisions"}],"predecessor-version":[{"id":13275,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/13273\/revisions\/13275"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/13272"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=13273"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=13273"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=13273"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}