{"id":39623,"date":"2025-03-25T03:02:48","date_gmt":"2025-03-24T19:02:48","guid":{"rendered":"https:\/\/17aitech.com\/?p=39623"},"modified":"2025-03-25T03:02:48","modified_gmt":"2025-03-24T19:02:48","slug":"%e8%85%be%e8%ae%af%e6%b7%b7%e5%85%83%e3%80%81%e8%8b%b1%e4%bc%9f%e8%be%be%e9%83%bd%e5%8f%91%e6%b7%b7%e5%90%88%e6%9e%b6%e6%9e%84%e6%a8%a1%e5%9e%8b%ef%bc%8cmamba-transformer%e8%a6%81%e5%b4%9b%e8%b5%b7","status":"publish","type":"post","link":"https:\/\/17aitech.com\/?p=39623","title":{"rendered":"\u817e\u8baf\u6df7\u5143\u3001\u82f1\u4f1f\u8fbe\u90fd\u53d1\u6df7\u5408\u67b6\u6784\u6a21\u578b\uff0cMamba-Transformer\u8981\u5d1b\u8d77\u5417\uff1f"},"content":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2025-03-24-6\" target=\"_blank\">\u817e\u8baf\u6df7\u5143\u3001\u82f1\u4f1f\u8fbe\u90fd\u53d1\u6df7\u5408\u67b6\u6784\u6a21\u578b\uff0cMamba-Transformer\u8981\u5d1b\u8d77\u5417\uff1f<\/a><\/p>\n<p>\u5728\u8fc7\u53bb\u7684\u4e00\u4e24\u5e74\u4e2d\uff0cTransformer \u67b6\u6784\u4e0d\u65ad\u9762\u4e34\u6765\u81ea\u65b0\u5174\u67b6\u6784\u7684\u6311\u6218\u3002<\/p>\n<p>\u5728\u4f17\u591a\u975e Transformer \u67b6\u6784\u4e2d\uff0cMamba \u65e0\u7591\u662f\u58f0\u91cf\u8f83\u5927\u4e14\u540e\u7eed\u53d1\u5c55\u8f83\u597d\u7684\u4e00\u4e2a\u3002\u7136\u800c\uff0c\u4e0e\u6700\u521d\u53d1\u5e03\u65f6\u90a3\u79cd\u4eff\u4f5b\u300c\u6c34\u706b\u4e0d\u5bb9\u300d\u7684\u5c40\u9762\u4e0d\u540c\uff0c\u6700\u8fd1\u4e00\u6bb5\u65f6\u95f4\uff0c\u8fd9\u4e24\u79cd\u67b6\u6784\u4f3c\u4e4e\u6b63\u5728\u8d70\u5411\u878d\u5408\u3002<\/p>\n<p>\u4e0a\u5468\u4e94\uff0c\u817e\u8baf\u5ba3\u5e03\u63a8\u51fa\u81ea\u7814\u6df1\u5ea6\u601d\u8003\u6a21\u578b\u300c\u6df7\u5143 T1\u300d\u6b63\u5f0f\u7248\uff0c\u8fd9\u662f\u4e00\u4e2a\u80fd\u79d2\u56de\u3001\u5410\u5b57\u5feb\u3001\u64c5\u957f\u8d85\u957f\u6587\u5904\u7406\u7684\u5f3a\u63a8\u7406\u6a21\u578b\u3002\u800c\u4e4b\u6240\u4ee5\u5177\u5907\u8fd9\u4e9b\u4f18\u52bf\uff0c\u5f88\u5927\u7a0b\u5ea6\u4e0a\u662f\u56e0\u4e3a\u817e\u8baf\u91c7\u7528\u4e86 Hybrid-Mamba-Transformer \u878d\u5408\u67b6\u6784\u3002\u8fd9\u4e00\u67b6\u6784\u6709\u6548\u964d\u4f4e\u4e86\u4f20\u7edf Transformer \u67b6\u6784\u7684\u8ba1\u7b97\u590d\u6742\u5ea6\uff0c\u51cf\u5c11\u4e86 KV-Cache \u7684\u5185\u5b58\u5360\u7528\uff0c\u4ece\u800c\u663e\u8457\u964d\u4f4e\u4e86\u8bad\u7ec3\u548c\u63a8\u7406\u6210\u672c\uff0c\u8ba9\u6df7\u5143 T1 \u5b9e\u73b0\u9996\u5b57\u79d2\u51fa\uff0c\u5410\u5b57\u901f\u5ea6\u6700\u5feb\u53ef\u8fbe 80 token\/s\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-f48cf64127ec6258cbf90a37c27be146.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-f48cf64127ec6258cbf90a37c27be146.png\"><\/a><\/p>\n<p>\u4e0e\u6b64\u540c\u65f6\uff0c\u82f1\u4f1f\u8fbe\u4e5f\u63a8\u51fa\u4e86\u4e00\u4e2a\u91c7\u7528 Mamba-Transformer \u6df7\u5408\u67b6\u6784\u7684\u6a21\u578b\u5bb6\u65cf \u2014\u2014Nemotron-H\uff0c\u5176\u901f\u5ea6\u662f\u540c\u4f53\u91cf\u7ade\u54c1\u6a21\u578b\u7684\u4e09\u500d\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-5bb03e8589b6eba5743832c9f704f278.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-5bb03e8589b6eba5743832c9f704f278.png\"><\/a><\/p>\n<p>\u901f\u5ea6\u7684\u63d0\u5347\u4e0e\u6210\u672c\u7684\u964d\u4f4e\uff0c\u662f AI \u5927\u6a21\u578b\u8fc8\u5411\u66f4\u5e7f\u6cdb\u5e94\u7528\u4e0e\u666e\u53ca\u7684\u5fc5\u7ecf\u4e4b\u8def\u3002\u5982\u4eca\uff0c\u817e\u8baf\u3001\u82f1\u4f1f\u8fbe\u7b49\u79d1\u6280\u5de8\u5934\u5bf9 Mamba-Transformer \u6df7\u5408\u67b6\u6784\u7684\u9ad8\u5ea6\u5173\u6ce8\u4e0e\u6295\u5165\uff0c\u91ca\u653e\u51fa\u4e00\u4e2a\u6781\u4e3a\u91cd\u8981\u7684\u4fe1\u53f7\uff1a\u6b64\u7c7b\u67b6\u6784\u6240\u8574\u542b\u7684\u5de8\u5927\u4ef7\u503c\uff0c\u503c\u5f97\u6211\u4eec\u6df1\u5165\u6316\u6398\u4e0e\u63a2\u7d22\u3002<\/p>\n<p>\u6070\u597d\uff0c\u6ed1\u94c1\u5362\u5927\u5b66\u8ba1\u7b97\u673a\u79d1\u5b66\u52a9\u7406\u6559\u6388\u9648\u6587\u864e\uff08Wenhu Chen\uff09\u6700\u8fd1\u53d1\u5e16\uff0c\u76d8\u70b9\u4e86\u4e00\u4e0b\u6700\u8fd1\u7684\u51e0\u6b3e Mamba-Transformer \u6a21\u578b\u3002\u6211\u4eec\u5bf9\u4ed6\u7684\u76d8\u70b9\u8fdb\u884c\u4e86\u6574\u7406\uff0c\u5e0c\u671b\u80fd\u5e26\u7ed9\u5927\u5bb6\u4e00\u4e9b\u542f\u53d1\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-d679f6849267d261524a75a976176df2.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-d679f6849267d261524a75a976176df2.png\"><\/a><\/p>\n<p><strong>Mamba-Transformer \u7b80\u4ecb<\/strong><\/p>\n<p>Mamba-Transformer \u6df7\u5408\u67b6\u6784\uff0c\u987e\u540d\u601d\u4e49\uff0c\u5c31\u662f\u5c06 Mamba \u4e0e Transformer \u67b6\u6784\u7ec4\u5408\u5230\u4e00\u8d77\u3002<\/p>\n<p>Transformer \u60f3\u5fc5\u5927\u5bb6\u5df2\u7ecf\u975e\u5e38\u719f\u6089\u4e86\uff0c\u7b80\u5355\u6765\u8bf4\uff1aTransformer \u67b6\u6784\u662f\u4e00\u79cd\u4ee5\u81ea\u6ce8\u610f\u529b\u673a\u5236\u4e3a\u6838\u5fc3\u7684\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\uff0c\u81ea 2017 \u5e74\u7531 Ashish Vaswani \u7b49\u4eba\u63d0\u51fa\u4ee5\u6765\uff0c\u4fbf\u9769\u65b0\u4e86\u4f20\u7edf\u5e8f\u5217\u6a21\u578b\u7684\u8bbe\u8ba1\u7406\u5ff5\u3002\u5176\u5173\u952e\u5728\u4e8e\u591a\u5934\u81ea\u6ce8\u610f\u529b\u673a\u5236\u80fd\u591f\u5728\u5168\u5c40\u8303\u56f4\u5185\u6355\u6349\u8f93\u5165\u5e8f\u5217\u4e2d\u5404\u5143\u7d20\u4e4b\u95f4\u7684\u590d\u6742\u4f9d\u8d56\u5173\u7cfb\uff0c\u540c\u65f6\u901a\u8fc7\u4f4d\u7f6e\u7f16\u7801\u4fdd\u7559\u5e8f\u5217\u4f4d\u7f6e\u4fe1\u606f\uff0c\u52a0\u4e0a\u6b8b\u5dee\u8fde\u63a5\u548c\u5c42\u5f52\u4e00\u5316\u7b49\u6280\u672f\u786e\u4fdd\u4e86\u9ad8\u6548\u7684\u5e76\u884c\u8ba1\u7b97\u548c\u7a33\u5b9a\u7684\u8bad\u7ec3\u8fc7\u7a0b\u3002\u8fd9\u79cd\u67b6\u6784\u4e0d\u4ec5\u663e\u8457\u63d0\u5347\u4e86\u673a\u5668\u7ffb\u8bd1\u3001\u6587\u672c\u751f\u6210\u7b49\u81ea\u7136\u8bed\u8a00\u5904\u7406\u4efb\u52a1\u7684\u6027\u80fd\uff0c\u4e5f\u4e3a\u540e\u7eed BERT\u3001GPT \u7b49\u9884\u8bad\u7ec3\u6a21\u578b\u7684\u5feb\u901f\u53d1\u5c55\u5960\u5b9a\u4e86\u575a\u5b9e\u57fa\u7840\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-9fff8dcc18368b8a8521f249bf9c3661.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-9fff8dcc18368b8a8521f249bf9c3661.png\"><\/a><\/p>\n<p><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0\u7ecf\u5178 Transformer \u67b6\u6784<\/sup><\/em><\/p>\n<p>\u800c Mamba \u5219\u662f\u4e00\u79cd\u72b6\u6001\u7a7a\u95f4\u6a21\u578b\uff08SSM\uff09\u2014\u2014 \u8be5\u67b6\u6784\u7684\u4e00\u5927\u663e\u8457\u4f18\u52bf\u662f\u80fd\u9ad8\u6548\u5730\u6355\u83b7\u5e8f\u5217\u6570\u636e\u4e2d\u7684\u590d\u6742\u4f9d\u8d56\u5173\u7cfb\uff0c\u5e76\u7531\u6b64\u6210\u4e3a Transformer \u7684\u4e00\u5927\u5f3a\u52b2\u5bf9\u624b\u3002<\/p>\n<p>\u7ecf\u5178\u7684\u72b6\u6001\u7a7a\u95f4\u6a21\u578b\u53ef\u88ab\u89c6\u4e3a\u5faa\u73af\u795e\u7ecf\u7f51\u7edc\uff08RNN\uff09\u548c\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\u7684\uff08CNN \u878d\u5408\u6a21\u578b\u3002\u5b83\u4eec\u53ef\u4f7f\u7528\u5faa\u73af\u6216\u5377\u79ef\u8fd0\u7b97\u8fdb\u884c\u9ad8\u6548\u5730\u8ba1\u7b97\uff0c\u4ece\u800c\u8ba9\u8ba1\u7b97\u5f00\u9500\u968f\u5e8f\u5217\u957f\u5ea6\u800c\u7ebf\u6027\u6216\u8fd1\u7ebf\u6027\u5730\u53d8\u5316\uff0c\u7531\u6b64\u5927\u5e45\u964d\u4f4e\u8ba1\u7b97\u6210\u672c\u3002<\/p>\n<p>\u4f5c\u4e3a SSM \u6700\u6210\u529f\u7684\u53d8\u4f53\u67b6\u6784\u4e4b\u4e00\uff0cMamba \u7684\u5efa\u6a21\u80fd\u529b\u5df2\u7ecf\u53ef\u4ee5\u6bd4\u80a9 Transformer\uff0c\u540c\u65f6\u8fd8\u80fd\u7ef4\u6301\u968f\u5e8f\u5217\u957f\u5ea6\u7684\u7ebf\u6027\u53ef\u6269\u5c55\u6027\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-2f4da3d63bab06b6a4b5267e1fa2f0ee.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-2f4da3d63bab06b6a4b5267e1fa2f0ee.png\"><\/a><\/p>\n<p>Mamba \u9996\u5148\u5f15\u5165\u4e86\u4e00\u4e2a\u7b80\u5355\u5374\u6709\u6548\u7684\u9009\u62e9\u673a\u5236\uff0c\u53ef\u6839\u636e\u8f93\u5165\u5bf9 SSM \u8fdb\u884c\u91cd\u65b0\u53c2\u6570\u5316\uff0c\u4ece\u800c\u53ef\u8ba9\u6a21\u578b\u5728\u6ee4\u9664\u4e0d\u76f8\u5173\u4fe1\u606f\u7684\u540c\u65f6\u65e0\u9650\u671f\u5730\u4fdd\u7559\u5fc5\u8981\u548c\u76f8\u5173\u7684\u6570\u636e\u3002\u7136\u540e\uff0cMamba \u8fd8\u5305\u542b\u4e00\u79cd\u786c\u4ef6\u611f\u77e5\u578b\u7b97\u6cd5\uff0c\u53ef\u4f7f\u7528\u626b\u63cf\uff08scan\uff09\u800c\u975e\u5377\u79ef\u6765\u5faa\u73af\u5730\u8ba1\u7b97\u6a21\u578b\uff0c\u8fd9\u5728 A100 GPU \u4e0a\u80fd\u8ba9\u8ba1\u7b97\u901f\u5ea6\u63d0\u5347 3 \u500d\u3002<\/p>\n<p>\u51ed\u501f\u5f3a\u5927\u7684\u5efa\u6a21\u590d\u6742\u957f\u5e8f\u5217\u6570\u636e\u7684\u80fd\u529b\u548c\u8fd1\u4e4e\u7ebf\u6027\u7684\u53ef\u6269\u5c55\u6027\uff0cMamba \u5df2\u7ecf\u5d1b\u8d77\u6210\u4e3a\u4e00\u79cd\u91cd\u8981\u7684\u57fa\u7840\u6a21\u578b\u67b6\u6784\u3002\u5173\u4e8e Mamba \u67b6\u6784\u7684\u66f4\u8be6\u7ec6\u4ecb\u7ecd\u53ef\u53c2\u9605\u673a\u5668\u4e4b\u5fc3\u6587\u7ae0<a data-itemshowtype=\"0\" data-linktype=\"2\" href=\"https:\/\/mp.weixin.qq.com\/s?__biz=MzA3MzI4MjgzMw==&amp;mid=2650930854&amp;idx=3&amp;sn=9eb22131ff3ddb060e00bfdb88358ead&amp;scene=21#wechat_redirect\" target=\"_blank\">\u300a\u4e00\u6587\u770b\u61c2 Mamba\uff0cTransformer \u6700\u5f3a\u7ade\u4e89\u8005\u300b<\/a>\u4ee5\u53ca<a data-itemshowtype=\"0\" data-linktype=\"2\" href=\"https:\/\/mp.weixin.qq.com\/s?__biz=MzA3MzI4MjgzMw==&amp;mid=2650920390&amp;idx=1&amp;sn=46a34c9edf886e58b6d18f28c8ec438a&amp;scene=21#wechat_redirect\" target=\"_blank\">\u300a\u518d\u6218 Transformer\uff01\u539f\u4f5c\u8005\u5e26\u961f\u7684 Mamba 2 \u6765\u4e86\uff0c\u65b0\u67b6\u6784\u8bad\u7ec3\u6548\u7387\u5927\u5e45\u63d0\u5347\u300b<\/a>\u3002<\/p>\n<p>\u800c\u8981\u5c06 Mamba \u4e0e Transformer \u6df7\u5408\u5230\u4e00\u8d77\uff0c\u5e38\u89c1\u7684\u65b9\u5f0f\u6709\u4e24\u79cd\uff1a\u5c42\u7ea7\u6df7\u5408\u4e0e\u5e8f\u5217\u7ea7\u6df7\u5408\u3002\u5982\u4e0b\u56fe\u6240\u793a\uff0c\u5f53\u524d\u7684 Mamba-Transformer \u6df7\u5408\u67b6\u6784\u6a21\u578b\u4f1a\u6839\u636e\u81ea\u8eab\u9700\u8981\u9009\u62e9\u4e0d\u540c\u7684\u6df7\u5408\u7b56\u7565\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-d91a7159e96ab4472a546d89de2881c1.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-d91a7159e96ab4472a546d89de2881c1.png\"><\/a><\/p>\n<p><strong>Nemotron-H<\/strong><\/p>\n<p>3 \u6708 21 \u65e5\uff0c\u82f1\u4f1f\u8fbe\u63a8\u51fa\u4e86 Nemotron-H \u7cfb\u5217\u7684 Mamba-Transformer \u6df7\u5408\u67b6\u6784\u6a21\u578b\uff0c\u5176\u4e2d\u5305\u542b\u591a\u79cd\u89c4\u6a21\u7684\u591a\u79cd\u6a21\u578b\uff0c\u6bd4\u5982 Nemotron-H-8B-Base\u3001Nemotron-H-8B-Instruct\u3001Nemotron-H-8B-VLM\u3001Nemotron-H-47B-Base\u3001 Nemotron-H-56B-Base\u3001Nemotron-H-56B-VLM\u3002\u5176\u4e2d 47B \u7248\u672c\u53ef\u4ee5\u5728\u5355\u53f0\u5546\u54c1\u7ea7 NVIDIA RTX 5090 GPU \u4e0a\u4ee5 FP4 \u7cbe\u5ea6\u652f\u6301 100 \u4e07 token \u957f\u5ea6\u4e0a\u4e0b\u6587\u7684\u63a8\u7406\u3002<\/p>\n<p>\u5b9e\u9645\u4e0a\uff0cNemotron-H \u5c31\u662f\u82f1\u4f1f\u8fbe\u8fd1\u671f\u53d1\u5e03\u7684\u7528\u4e8e\u7269\u7406 AI \u7684\u5f3a\u5927 VLM \u6a21\u578b Cosmos-Reason 1 \u80cc\u540e\u7684\u9aa8\u5e72\u7f51\u7edc\u3002<\/p>\n<p>\u636e\u4ecb\u7ecd\uff0c\u901a\u8fc7\u91c7\u7528 Mamba-Transformer \u6df7\u5408\u67b6\u6784\uff0c\u76f8\u6bd4\u4e8e SOTA \u7684\u5f00\u6e90\u7eaf Transformer \u6a21\u578b\uff0cNemotron-H \u5728\u4fdd\u8bc1\u4e86\u76f8\u5f53\u4e43\u81f3\u66f4\u597d\u7684\u51c6\u786e\u5ea6\u7684\u540c\u65f6\uff0c\u53ef\u4ee5\u63d0\u4f9b\u8fdc\u8fdc\u66f4\u5feb\u7684\u63a8\u7406\u901f\u5ea6\uff08\u9ad8\u8fbe 3 \u500d\uff09\u3002\u4e0b\u9762\u4e24\u56fe\u5c55\u793a\u4e86\u5728 MMLU-Pro \u57fa\u51c6\u4e0a\uff0cNemotron-H \u4e0e\u540c\u7b49\u6570\u91cf\u7ea7\u53c2\u6570\u7684\u5f00\u6e90 Transformer \u6a21\u578b\u7684\u51c6\u786e\u5ea6\u4e0e\u541e\u5410\u91cf\u5bf9\u6bd4\u3002\u53ef\u4ee5\u770b\u5230\uff0cNemotron-H \u5177\u6709\u975e\u5e38\u660e\u663e\u7684\u6548\u7387\u4f18\u52bf\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-aff168e079862acbe52ed33279139a28.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-aff168e079862acbe52ed33279139a28.png\"><\/a><\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-4019b77655bde6e35a86a1690fa1c35b.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-4019b77655bde6e35a86a1690fa1c35b.png\"><\/a><\/p>\n<p>\u6b64\u5916\uff0c\u82f1\u4f1f\u8fbe\u4e5f\u516c\u5e03\u4e86 Nemotron-H \u7684\u5176\u5b83\u4e00\u4e9b\u7ec6\u8282\uff0c\u5305\u62ec Nemotron-H-56B-Base \u7684\u8bad\u7ec3\u4f7f\u7528\u4e86 6144 \u53f0 H100 GPU \u548c 20 \u4e07\u4ebf token\uff0c\u8bad\u7ec3\u7cbe\u5ea6\u4e3a FP8\uff08\u662f FP8 \u9884\u8bad\u7ec3\u7684\u4e00\u6b21\u5927\u89c4\u6a21\u5c55\u793a\uff09\uff1bNemotron-H-47B-Base \u662f Nemotron-H-56B-Base \u7684\u84b8\u998f\u7248 \u2014\u2014 \u84b8\u998f\u7cbe\u5ea6\u4e3a FP8 \u5e76\u4f7f\u7528\u4e86 630 \u4ebf\u8bad\u7ec3 token\u3002<\/p>\n<p><strong>Hunyuan-Turbo-S \u548c Hunyuan-T1<\/strong><\/p>\n<p>3 \u6708 21 \u65e5\u540c\u4e00\u5929\uff0c\u817e\u8baf\u4e5f\u5ba3\u5e03\u63a8\u51fa\u4e86\u6df1\u5ea6\u601d\u8003\u6a21\u578b\u6df7\u5143 T1 \u6b63\u5f0f\u7248\uff0c\u5e76\u540c\u6b65\u5728\u817e\u8baf\u4e91\u5b98\u7f51\u4e0a\u7ebf\u3002\u800c\u5728\u6b64\u4e4b\u524d\uff0c\u4ed6\u4eec\u5df2\u4e8e 3 \u6708\u521d\u53d1\u5e03\u4e86\u6df7\u5143 Turbo S \u57fa\u7840\u6a21\u578b\u4ee5\u53ca\u4e4b\u540e\u7684\u6df7\u5143 T1-preview \u6a21\u578b\u3002<\/p>\n<p>\u8fd9\u4e9b\u6a21\u578b\u90fd\u91c7\u7528\u4e86 Mamba-Transformer \u6df7\u5408\u67b6\u6784\uff0c\u5176\u4e2d T1 \u66f4\u662f\u4e00\u6b3e\u4f7f\u7528\u4e86\u5927\u89c4\u6a21\u5f3a\u5316\u5b66\u4e60\u7684\u5f3a\u63a8\u7406\u6a21\u578b\uff0c\u5728\u6570\u5b66\u3001\u903b\u8f91\u63a8\u7406\u3001\u79d1\u5b66\u548c\u4ee3\u7801\u7b49\u7406\u79d1\u96be\u9898\u5177\u6709\u975e\u5e38\u660e\u663e\u7684\u4f18\u52bf\u3002\u5f53\u7136\uff0cMamba-Transformer \u9ad8\u6548\u7387\u7684\u4f18\u52bf\u4e5f\u5728\u8fd9\u91cc\u6709\u975e\u5e38\u660e\u663e\u7684\u4f53\u73b0\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-c32b0c8eff6d73cf36dc68c70140dc90.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-c32b0c8eff6d73cf36dc68c70140dc90.png\"><\/a><\/p>\n<p>\u6bd4\u5982\u5728\u5927\u8bed\u8a00\u6a21\u578b\u8bc4\u4f30\u589e\u5f3a\u6570\u636e\u96c6 MMLU-PRO \u4e0a\uff0c\u6df7\u5143 T1 \u53d6\u5f97\u4e86 87.2 \u5206\uff0c\u4ec5\u6b21\u4e8e o1\u3002\u5728 CEval\u3001AIME\u3001Zebra Logic \u7b49\u4e2d\u82f1\u6587\u77e5\u8bc6\u53ca\u7ade\u8d5b\u7ea7\u6570\u5b66\u3001\u903b\u8f91\u63a8\u7406\u7684\u516c\u5f00\u57fa\u51c6\u6d4b\u8bd5\u4e2d\uff0c\u6df7\u5143 T1 \u7684\u6210\u7ee9\u4e5f\u8fbe\u5230\u4e1a\u754c\u9886\u5148\u63a8\u7406\u6a21\u578b\u7684\u6c34\u5e73\u3002<\/p>\n<p>\u636e\u4ecb\u7ecd\uff0c\u6df7\u5143 T1 \u6b63\u5f0f\u7248\u6cbf\u7528\u4e86\u6df7\u5143 Turbo S \u7684\u521b\u65b0\u67b6\u6784\uff0c\u91c7\u7528 Hybrid-Mamba-Transformer \u878d\u5408\u6a21\u5f0f\u3002\u8fd9\u662f\u5de5\u4e1a\u754c\u9996\u6b21\u5c06\u6df7\u5408 Mamba \u67b6\u6784\u65e0\u635f\u5e94\u7528\u4e8e\u8d85\u5927\u578b\u63a8\u7406\u6a21\u578b\u3002<\/p>\n<p>\u8fd9\u4e00\u67b6\u6784\u6709\u6548\u964d\u4f4e\u4e86\u4f20\u7edf Transformer \u7ed3\u6784\u7684\u8ba1\u7b97\u590d\u6742\u5ea6\uff0c\u51cf\u5c11\u4e86 KV-Cache \u7684\u5185\u5b58\u5360\u7528\uff0c\u4ece\u800c\u663e\u8457\u964d\u4f4e\u4e86\u8bad\u7ec3\u548c\u63a8\u7406\u6210\u672c\uff0c\u8ba9\u6df7\u5143 T1 \u5b9e\u73b0\u9996\u5b57\u79d2\u51fa\uff0c\u5410\u5b57\u901f\u5ea6\u8fbe\u5230\u6700\u5feb 80 token\/s\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-f3f42c7c1685db0e41ab3aa86896d086.gif\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-f3f42c7c1685db0e41ab3aa86896d086.gif\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0\u6df7\u5143 T1 \u4e0e DeepSeek-R1 \u7684\u5410\u5b57\u901f\u5ea6\u5bf9\u6bd4<\/sup><\/em><\/section>\n<section><\/section>\n<section>\u6df7\u5143 T1 \u5728\u8d85\u957f\u6587\u672c\u63a8\u7406\u9886\u57df\u4e5f\u5c55\u73b0\u51fa\u72ec\u7279\u4f18\u52bf\u3002\u57fa\u4e8e\u51fa\u8272\u7684\u957f\u6587\u6355\u6349\u80fd\u529b\uff0c\u6df7\u5143 T1 \u80fd\u6709\u6548\u89e3\u51b3\u957f\u6587\u63a8\u7406\u4e2d\u5e38\u89c1\u7684\u4e0a\u4e0b\u6587\u4e22\u5931\u548c\u957f\u8ddd\u79bb\u4fe1\u606f\u4f9d\u8d56\u95ee\u9898\u3002\u540c\u65f6\uff0c\u6df7\u5408 Mamba \u67b6\u6784\u9488\u5bf9\u957f\u5e8f\u5217\u5904\u7406\u8fdb\u884c\u4e86\u4e13\u9879\u4f18\u5316\uff0c\u901a\u8fc7\u9ad8\u6548\u7684\u8ba1\u7b97\u65b9\u5f0f\uff0c\u5728\u786e\u4fdd\u957f\u6587\u672c\u4fe1\u606f\u6355\u6349\u80fd\u529b\u7684\u540c\u65f6\u5927\u5e45\u964d\u4f4e\u8d44\u6e90\u6d88\u8017\uff0c\u5728\u76f8\u8fd1\u7684\u6fc0\u6d3b\u53c2\u6570\u91cf\u4e0b\uff0c\u5b9e\u73b0\u4e86\u89e3\u7801\u901f\u5ea6\u63d0\u5347 2 \u500d\u3002<\/section>\n<section>\u00a0<\/section>\n<section>\u5728\u4f53\u73b0\u63a8\u7406\u6a21\u578b\u57fa\u7840\u80fd\u529b\u7684\u5e38\u89c1 benchmark \u4e0a\uff0c\u5982\u5927\u8bed\u8a00\u6a21\u578b\u8bc4\u4f30\u589e\u5f3a\u6570\u636e\u96c6 MMLU-PRO \u4e2d\uff0c\u6df7\u5143 T1 \u53d6\u5f97 87.2 \u5206\uff0c\u4ec5\u6b21\u4e8e o1\u3002\u5728 CEval\u3001AIME\u3001Zebra Logic \u7b49\u4e2d\u82f1\u6587\u77e5\u8bc6\u53ca\u7ade\u8d5b\u7ea7\u6570\u5b66\u3001\u903b\u8f91\u63a8\u7406\u7684\u516c\u5f00\u57fa\u51c6\u6d4b\u8bd5\u4e2d\uff0c\u6df7\u5143 T1 \u7684\u6210\u7ee9\u4e5f\u8fbe\u5230\u4e1a\u754c\u9886\u5148\u63a8\u7406\u6a21\u578b\u7684\u6c34\u5e73\u3002<\/section>\n<section>\u00a0<\/section>\n<section><strong>STORM<\/strong><\/section>\n<section><\/section>\n<section>STORM \u540c\u6837\u6765\u81ea\u82f1\u4f1f\u8fbe\uff0c\u662f\u4e00\u6b3e\u57fa\u4e8e\u89c6\u9891\u7684\u591a\u6a21\u6001\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08Video-LLM\uff09\u3002\u76f8\u5173\u7814\u7a76\u8bba\u6587\u5df2\u4e8e\u672c\u6708\u521d\u53d1\u5e03\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-0220aa893cbbb159451a8ccdd6e0f53b.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-0220aa893cbbb159451a8ccdd6e0f53b.png\"><\/a><\/section>\n<ul>\n<li>\n<section>\u8bba\u6587\u6807\u9898\uff1aToken-Efficient Long Video Understanding for Multimodal LLMs<\/section>\n<\/li>\n<li>\n<section>\u8bba\u6587\u5730\u5740\uff1ahttps:\/\/arxiv.org\/pdf\/2503.04130<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<section>\u5177\u4f53\u6280\u672f\u4e0a\uff0cSTORM \u5728\u56fe\u50cf\u7f16\u7801\u5668\u548c LLM \u4e4b\u95f4\u5f15\u5165\u4e86\u4e00\u79cd\u65b0\u7684\u65f6\u95f4\u7f16\u7801\u5668\uff0c\u4ece\u800c\u5c06\u89c6\u89c9\u548c\u8bed\u8a00\u8868\u5f81\u6574\u5408\u5230\u4e86\u4e00\u8d77\u3002\u8be5\u67b6\u6784\u5728\u5de5\u4f5c\u6d41\u7a0b\u65e9\u671f\u5f15\u5165\u4e86\u65f6\u95f4\u52a8\u6001\uff0c\u4ece\u800c\u53ef\u4ee5\u663e\u8457\u63d0\u5347 Video-LLM \u7684\u65f6\u95f4\u63a8\u7406\u80fd\u529b\uff0c\u540c\u65f6\u8fd8\u80fd\u5927\u5e45\u63d0\u5347\u7684\u4e0b\u6e38\u8ba1\u7b97\u6548\u7387\u3002\u8be5\u56e2\u961f\u8868\u793a\uff1a\u300c\u901a\u8fc7\u5c06\u65f6\u95f4\u4fe1\u606f\u76f4\u63a5\u6ce8\u5165\u89c6\u89c9 token\uff0c\u53ef\u4ee5\u51cf\u8f7b LLM \u7684\u65f6\u95f4\u63a8\u7406\u8d1f\u62c5\uff0c\u4f7f\u5176\u80fd\u591f\u4e13\u6ce8\u4e8e\u66f4\u9ad8\u7ea7\u7684\u8bed\u8a00\u4efb\u52a1\u3002\u300d<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-743b0157f943026b1a78b63e2285cdea.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-743b0157f943026b1a78b63e2285cdea.png\"><\/a><\/section>\n<section>\u67b6\u6784\u4e0a\uff0c\u4ed6\u4eec\u91c7\u7528\u4e86 Mamba \u72b6\u6001\u7a7a\u95f4\u6a21\u578b\u4f5c\u4e3a\u65f6\u95f4\u5c42\u7684\u6838\u5fc3\uff0c\u8fd9\u65e2\u80fd\u9ad8\u6548\u5904\u7406\u957f\u89c6\u9891\uff0c\u540c\u65f6\u4e5f\u53ef\u589e\u5f3a\u5bf9\u7ecf\u8fc7\u6269\u5c55\u7684\u65f6\u95f4\u4e0a\u4e0b\u6587\u7684\u6cdb\u5316\u3002\u8fd9\u4e2a\u65f6\u95f4\u5c42\u5bf9\u56fe\u50cf\u548c\u89c6\u9891\u8f93\u5165\u7684\u5904\u7406\u65b9\u5f0f\u4e0d\u540c \u2014\u2014 \u5bf9\u4e8e\u56fe\u50cf\uff0c\u5b83\u7684\u4f5c\u7528\u662f\u7a7a\u95f4\u626b\u63cf\u5668\uff0c\u901a\u8fc7\u7ed3\u5408\u5168\u5c40\u7a7a\u95f4\u4e0a\u4e0b\u6587\u6765\u589e\u5f3a token\uff1b\u5bf9\u4e8e\u89c6\u9891\uff0c\u5b83\u540c\u65f6\u6267\u884c\u7a7a\u95f4\u548c\u65f6\u95f4\u626b\u63cf\uff0c\u4ee5\u6355\u83b7\u5168\u9762\u7684\u65f6\u7a7a\u4fe1\u606f\u3002<\/section>\n<section><\/section>\n<section>Mamba \u5c42\u7684\u4e3b\u8981\u4f18\u52bf\u5728\u4e8e\u5b83\u80fd\u591f\u5c06\u5386\u53f2\u4fe1\u606f\u538b\u7f29\u4e3a\u72b6\u6001\u8868\u5f81\u3002\u7531\u4e8e\u89c6\u9891\u8f93\u5165\u4e2d\u7684\u8fde\u7eed\u5e27\u901a\u5e38\u5305\u542b\u5197\u4f59\u4fe1\u606f\uff0c\u56e0\u6b64\u65f6\u95f4\u7f16\u7801\u5668\u53ef\u4ee5\u6709\u6548\u5730\u5904\u7406\u548c\u4f20\u64ad\u6574\u4e2a\u89c6\u9891\u5e8f\u5217\u7684\u65f6\u95f4\u4fe1\u606f\u3002\u8fd9\u6837\u4e00\u6765\uff0c\u751f\u6210\u7684\u89c6\u89c9 token \u672c\u8eab\u5c31\u5c01\u88c5\u4e86\u65f6\u95f4\u5386\u53f2\uff0c\u5e76\u6709\u6548\u5730\u603b\u7ed3\u4e86\u89c6\u9891\u7684\u65f6\u95f4\u52a8\u6001\u3002<\/section>\n<section><\/section>\n<section>\u6709\u4e86\u8fd9\u4e00\u7279\u6027\uff0cLLM \u5904\u7406\u6240\u9700\u7684\u89c6\u89c9 token \u5c31\u66f4\u5c11\u4e86\uff0c\u540c\u65f6\u8fd8\u80fd\u4fdd\u7559\u5173\u952e\u4fe1\u606f\u3002<\/section>\n<section><\/section>\n<section>\u4ed6\u4eec\u7684\u5b9e\u9a8c\u8868\u660e\uff0c\u65b0\u63d0\u51fa\u7684\u65b9\u6cd5\u6548\u679c\u5f88\u597d\uff0c\u5e76\u4e14\u53ef\u4ee5\u6210\u529f\u6269\u5c55\u5230\u957f\u4e0a\u4e0b\u6587\u89c6\u9891\u7406\u89e3\u6a21\u578b\uff0c\u800c\u4e0d\u4f1a\u5f71\u54cd\u8bad\u7ec3\u6548\u7387\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-cc70edb025cf2abd75f79ac94f052bc1.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-cc70edb025cf2abd75f79ac94f052bc1.png\"><\/a><\/section>\n<section><strong>Vamba<\/strong><\/section>\n<section><\/section>\n<section>Vamba \u5219\u662f\u4e00\u4e2a\u7528\u4e8e\u957f\u89c6\u9891\u7406\u89e3\u7684 Mamba-Transformer \u6df7\u5408\u6a21\u578b\uff0c\u6765\u81ea\u524d\u6587\u4e2d\u63d0\u5230\u7684\u9648\u6587\u864e\u56e2\u961f\u3002<\/section>\n<section><\/section>\n<section>\u6211\u4eec\u77e5\u9053\uff0c\u7531\u4e8e\u56e0\u679c\u81ea\u6ce8\u610f\u529b\u8fd0\u7b97\u7684\u4e8c\u6b21\u590d\u6742\u5ea6\uff08\u4f1a\u5bfc\u81f4\u8bad\u7ec3\u548c\u63a8\u7406\u7684\u8ba1\u7b97\u6210\u672c\u5f88\u9ad8\uff09\uff0c\u5f53\u524d\u9886\u5148\u7684 Transformer LMM \u5176\u5b9e\u5f88\u96be\u5904\u7406\u957f\u89c6\u9891\uff0c\u5c24\u5176\u662f\u5f53\u89c6\u9891\u957f\u5ea6\u957f\u8fbe 1 \u4e2a\u5c0f\u65f6\u4ee5\u4e0a\u65f6\u3002\u73b0\u5728\u867d\u6709\u4e00\u4e9b token \u538b\u7f29\u65b9\u6cd5\u53ef\u4ee5\u51cf\u5c11\u89c6\u9891 token \u7684\u6570\u91cf\uff0c\u4f46\u5f80\u5f80\u4f1a\u4e22\u5931\u4e00\u4e9b\u4fe1\u606f\uff0c\u5e76\u4e14\u5904\u7406\u6781\u957f\u5e8f\u5217\u7684\u6548\u7387\u5f88\u4f4e\u3002<\/section>\n<section><\/section>\n<section>\u4e8e\u662f Vamba \u56e2\u961f\u63a2\u7d22\u4e86\u53e6\u4e00\u4e2a\u65b9\u5411\uff1a\u4e0d\u5bf9\u89c6\u9891 token \u8fdb\u884c\u538b\u7f29\uff0c\u800c\u662f\u4f7f\u7528\u53e6\u4e00\u79cd\u6a21\u578b\u67b6\u6784\u6765\u63d0\u9ad8\u5728\u8bad\u7ec3\u548c\u63a8\u7406\u9884\u586b\u5145\u9636\u6bb5\u5904\u7406\u89c6\u9891 token \u7684\u6548\u7387\u3002<\/section>\n<section><\/section>\n<section>VAMBA \u7531\u6b64\u8bde\u751f\uff01\u8fd9\u6b3e\u6df7\u5408 Mamba-Transformer \u6a21\u578b\u53ef\u4ee5\u5bf9\u957f\u8fbe 1 \u4e2a\u5c0f\u65f6\u7684\u89c6\u9891\u5b9e\u73b0\u9ad8\u6548\u7684\u7406\u89e3\u3002\u8be5\u56e2\u961f\u8bf4\u660e\u4e86\u8bbe\u8ba1\u8fd9\u4e00\u67b6\u6784\u7684\u5173\u952e\u89c1\u89e3\uff1a\u53ef\u4ee5\u8bbe\u8ba1\u9ad8\u6548\u7684\u6a21\u5757\u6765\u8fd1\u4f3c\u57fa\u4e8e Transformer \u7684 LMM \u4e2d\u6587\u672c\u548c\u89c6\u9891 token \u7684\u56e0\u679c\u81ea\u6ce8\u610f\u529b\u8fd0\u7b97\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-12ced095540ba7e68fd8cd874a823b6a.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-12ced095540ba7e68fd8cd874a823b6a.png\"><\/a><\/section>\n<ul>\n<li>\n<section>\u8bba\u6587\u6807\u9898\uff1aVAMBA: Understanding Hour-Long Videos with Hybrid Mamba-Transformers<\/section>\n<\/li>\n<li>\n<section>\u8bba\u6587\u5730\u5740\uff1ahttps:\/\/arxiv.org\/pdf\/2503.11579<\/section>\n<\/li>\n<li>\n<section>\u9879\u76ee\u5730\u5740\uff1ahttps:\/\/tiger-ai-lab.github.io\/Vamba\/<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<section>\u5177\u4f53\u6765\u8bf4\uff0c\u4ed6\u4eec\u63d0\u51fa\uff1a\u00a0<\/section>\n<section><\/section>\n<section>1\u3001\u5229\u7528\u4ea4\u53c9\u6ce8\u610f\u529b\u673a\u5236\u6839\u636e\u89c6\u9891 token \u66f4\u65b0\u6587\u672c token\uff0c\u7531\u4e8e\u6587\u672c token \u7684\u957f\u5ea6\u8f83\u77ed\uff0c\u8fd9\u79cd\u65b9\u6cd5\u6210\u672c\u8f83\u4f4e\uff1b<\/section>\n<section>2\u3001\u91c7\u7528 Mamba-2 \u4ee5\u7ebf\u6027\u590d\u6742\u5ea6\u5904\u7406\u5927\u91cf\u89c6\u9891 token\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-e7542dfcc106283b677ba6addb3b3c2f.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-e7542dfcc106283b677ba6addb3b3c2f.png\"><\/a><\/section>\n<section>\u5047\u8bbe\u8f93\u5165\u5e8f\u5217\u4e3a M + N \u4e2a token\uff0c\u5176\u4e2d M \u662f\u89c6\u9891 token \u7684\u6570\u91cf\uff0cN \u662f\u6587\u672c token \u7684\u6570\u91cf\uff0c\u8be5\u56e2\u961f\u53d1\u73b0\u5728\u8bb8\u591a\u957f\u89c6\u9891\u4efb\u52a1\uff08M \u226b N\uff09\u4e0a\uff0cM \u81f3\u5c11\u662f N \u7684 100 \u500d\u3002<\/section>\n<section><\/section>\n<section>\u800c\u8be5\u56e2\u961f\u63d0\u51fa\u7684\u65b0\u6a21\u578b\u53ef\u4ee5\u5c06\u8bad\u7ec3 \/ \u9884\u586b\u5145\u8ba1\u7b97\u590d\u6742\u5ea6\u4ece O (d (M + N)\u00b2) \u964d\u4f4e\u5230 O (dMN + d\u00b2M)\uff0c\u5176\u4e2d d \u662f\u9690\u85cf\u7ef4\u5ea6\u3002\u4e0d\u8fc7\u8be5\u56e2\u961f\u4e5f\u6307\u51fa\uff0c\u5728\u5b9e\u8df5\u4e2d\uff0c\u7531\u4e8e\u76ee\u524d Mamba \u7684\u786c\u4ef6\u4f18\u5316\u4e0d\u8db3\uff0c\u8fd9\u79cd\u7406\u8bba\u4e0a\u7684\u6539\u8fdb\u53ef\u80fd\u65e0\u6cd5\u5b8c\u5168\u5b9e\u73b0\u3002<\/section>\n<section><\/section>\n<section>\u5c3d\u7ba1\u5982\u6b64\uff0c\u4ed6\u4eec\u4ecd\u7136\u89c2\u5bdf\u5230\uff0c\u5728\u957f\u89c6\u9891\u8f93\u5165\u7684\u8bad\u7ec3\u548c\u63a8\u7406\u8fc7\u7a0b\u4e2d\uff0cGPU \u5185\u5b58\u4f7f\u7528\u91cf\u548c FLOP \/ \u8fd0\u884c\u65f6\u95f4\u51cf\u5c11\u4e86 50% \u4ee5\u4e0a\u3002<\/section>\n<section><\/section>\n<section>VAMBA \u53ef\u4ee5\u4f7f\u7528 8 \u53f0 A100 GPU \u8fdb\u884c\u6709\u6548\u8bad\u7ec3\uff0c\u800c LongVU \u548c LongLLaVA \u5206\u522b\u9700\u8981 64 \u548c 24 \u53f0 GPU \u8fdb\u884c\u8bad\u7ec3\u3002\u901a\u8fc7\u8fdb\u884c\u4e24\u9636\u6bb5\u8bad\u7ec3\uff0c\u65b0\u63d0\u51fa\u7684 VAMBA \u5728\u9ad8\u96be\u5ea6\u7684\u5c0f\u65f6\u7ea7\u89c6\u9891\u7406\u89e3\u57fa\u51c6 LVBench \u4e0a\u6bd4\u4e4b\u524d\u6700\u9ad8\u6548\u7684\u89c6\u9891 LMM \u63d0\u9ad8\u4e86 4.3%\u3002\u5728\u5176\u4ed6\u957f\u89c6\u9891\u7406\u89e3\u6570\u636e\u96c6\u4e0a\uff0c\u5982 Video-MME\u3001MLVU \u548c LongVideoBench \uff0cVAMBA \u4e5f\u8868\u73b0\u51fa\u8272\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-c74798ee79790a42d07e02c60b275172.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-c74798ee79790a42d07e02c60b275172.png\"><\/a><\/section>\n<section>\u770b\u8d77\u6765\uff0cMamba-Transformer \u6df7\u5408\u67b6\u6784\u5df2\u7ecf\u5f00\u59cb\u5728\u63a8\u7406\u4ee5\u53ca\u591a\u6a21\u6001\u4efb\u52a1\u4e2d\u5c55\u9732\u5934\u89d2\u3002\u901a\u8fc7\u5408\u7406\u7684\u8bbe\u8ba1\uff0cTransformer \u5f3a\u5927\u7684\u6027\u80fd\u53ef\u4ee5\u4e0e Mamba \u5728\u957f\u4e0a\u4e0b\u6587\u548c\u6548\u7387\u4e0a\u7684\u4f18\u52bf\u6574\u5408\u5230\u4e00\u8d77\uff0c\u4e3a\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u548c\u5927\u578b\u591a\u6a21\u6001\u6a21\u578b\uff08LMM\uff09\u5e26\u6765\u65b0\u7684\u53ef\u80fd\u6027\u3002<\/section>\n<section><\/section>\n<section>\u673a\u5668\u4e4b\u5fc3\u4e5f\u66fe\u62a5\u9053\u8fc7\u5176\u5b83\u4e00\u4e9b Mamba-Transformer \u6df7\u5408\u67b6\u6784\u6a21\u578b\uff0c\u8be6\u60c5\u8bbf\u95ee\u4ee5\u4e0b\u94fe\u63a5\uff1a<\/section>\n<section><\/section>\n<section><a data-itemshowtype=\"0\" data-linktype=\"2\" href=\"https:\/\/mp.weixin.qq.com\/s?__biz=MzA3MzI4MjgzMw==&amp;mid=2650935336&amp;idx=4&amp;sn=1d0f7c1abd46f072741e563050151045&amp;scene=21#wechat_redirect\" target=\"_blank\">\u300a\u9996\u4e2a Mamba+Transformer \u6df7\u5408\u67b6\u6784\u591a\u6a21\u6001\u5927\u6a21\u578b\u6765\u4e86\uff0c\u5b9e\u73b0\u5355\u5361\u5343\u56fe\u63a8\u7406\u300b<\/a><\/section>\n<section><a data-itemshowtype=\"0\" data-linktype=\"2\" href=\"https:\/\/mp.weixin.qq.com\/s?__biz=MzA3MzI4MjgzMw==&amp;mid=2650912593&amp;idx=2&amp;sn=ed4b73d0303495105a8e3f511134acfe&amp;scene=21#wechat_redirect\" target=\"_blank\">\u300aAttention isn\u2019t all you need\uff01Mamba \u6df7\u5408\u5927\u6a21\u578b\u5f00\u6e90\uff1a\u4e09\u500d Transformer \u541e\u5410\u91cf\u300b<\/a><\/section>\n<section><\/section>\n<section><em><sup>\u53c2\u8003\u5185\u5bb9<\/sup><\/em><\/section>\n<section><sup><em>https:\/\/x.com\/WenhuChen\/status\/1903656455036715311<\/em><\/sup><\/section>\n<section><em><sup>https:\/\/research.nvidia.com\/labs\/adlr\/nemotronh\/<\/sup><\/em><\/section>\n<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2025-03-24-6\" target=\"_blank\">\u817e\u8baf\u6df7\u5143\u3001\u82f1\u4f1f\u8fbe\u90fd\u53d1\u6df7\u5408\u67b6\u6784\u6a21\u578b\uff0cMamba-Transformer\u8981\u5d1b\u8d77\u5417\uff1f<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:\u817e\u8baf\u6df7\u5143\u3001\u82f1\u4f1f\u8fbe\u90fd\u53d1\u6df7 [&hellip;]<\/p>\n","protected":false},"author":3,"featured_media":0,"comment_status":"open","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"default","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"var(--ast-global-color-4)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"footnotes":""},"categories":[27],"tags":[73,68,82,65,67],"class_list":["post-39623","post","type-post","status-publish","format-standard","hentry","category-news","tag-73","tag-68","tag-82","tag-65","tag-67"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v26.4 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>\u817e\u8baf\u6df7\u5143\u3001\u82f1\u4f1f\u8fbe\u90fd\u53d1\u6df7\u5408\u67b6\u6784\u6a21\u578b\uff0cMamba-Transformer\u8981\u5d1b\u8d77\u5417\uff1f - \u4e00\u8d77AI\u6280\u672f<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/17aitech.com\/?p=39623\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/17aitech.com\/?p=39623\",\"url\":\"https:\/\/17aitech.com\/?p=39623\",\"name\":\"\u817e\u8baf\u6df7\u5143\u3001\u82f1\u4f1f\u8fbe\u90fd\u53d1\u6df7\u5408\u67b6\u6784\u6a21\u578b\uff0cMamba-Transformer\u8981\u5d1b\u8d77\u5417\uff1f - \u4e00\u8d77AI\u6280\u672f\",\"isPartOf\":{\"@id\":\"https:\/\/17aitech.com\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\/\/17aitech.com\/?p=39623#primaryimage\"},\"image\":{\"@id\":\"https:\/\/17aitech.com\/?p=39623#primaryimage\"},\"thumbnailUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-f48cf64127ec6258cbf90a37c27be146.png\",\"datePublished\":\"2025-03-24T19:02:48+00:00\",\"author\":{\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d\"},\"breadcrumb\":{\"@id\":\"https:\/\/17aitech.com\/?p=39623#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/17aitech.com\/?p=39623\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/17aitech.com\/?p=39623#primaryimage\",\"url\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-f48cf64127ec6258cbf90a37c27be146.png\",\"contentUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-f48cf64127ec6258cbf90a37c27be146.png\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/17aitech.com\/?p=39623#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/17aitech.com\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u817e\u8baf\u6df7\u5143\u3001\u82f1\u4f1f\u8fbe\u90fd\u53d1\u6df7\u5408\u67b6\u6784\u6a21\u578b\uff0cMamba-Transformer\u8981\u5d1b\u8d77\u5417\uff1f\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/17aitech.com\/#website\",\"url\":\"https:\/\/17aitech.com\/\",\"name\":\"\u4e00\u8d77AI\u6280\u672f\",\"description\":\"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca\",\"alternateName\":\"\u4e00\u8d77AI\u6280\u672f\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/17aitech.com\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d\",\"name\":\"AI\u5c0f\u52a9\u624b\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/image\/\",\"url\":\"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png\",\"contentUrl\":\"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png\",\"caption\":\"AI\u5c0f\u52a9\u624b\"},\"description\":\"\u8fd9\u4e2a\u4eba\u5f88\u61d2\uff0c\u4ec0\u4e48\u90fd\u6ca1\u6709\u7559\u4e0b\uff5e\",\"url\":\"https:\/\/17aitech.com\/?page_id=33738&user=3\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"\u817e\u8baf\u6df7\u5143\u3001\u82f1\u4f1f\u8fbe\u90fd\u53d1\u6df7\u5408\u67b6\u6784\u6a21\u578b\uff0cMamba-Transformer\u8981\u5d1b\u8d77\u5417\uff1f - \u4e00\u8d77AI\u6280\u672f","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/17aitech.com\/?p=39623","schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/17aitech.com\/?p=39623","url":"https:\/\/17aitech.com\/?p=39623","name":"\u817e\u8baf\u6df7\u5143\u3001\u82f1\u4f1f\u8fbe\u90fd\u53d1\u6df7\u5408\u67b6\u6784\u6a21\u578b\uff0cMamba-Transformer\u8981\u5d1b\u8d77\u5417\uff1f - \u4e00\u8d77AI\u6280\u672f","isPartOf":{"@id":"https:\/\/17aitech.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/17aitech.com\/?p=39623#primaryimage"},"image":{"@id":"https:\/\/17aitech.com\/?p=39623#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-f48cf64127ec6258cbf90a37c27be146.png","datePublished":"2025-03-24T19:02:48+00:00","author":{"@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"breadcrumb":{"@id":"https:\/\/17aitech.com\/?p=39623#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/17aitech.com\/?p=39623"]}]},{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/?p=39623#primaryimage","url":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-f48cf64127ec6258cbf90a37c27be146.png","contentUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-f48cf64127ec6258cbf90a37c27be146.png"},{"@type":"BreadcrumbList","@id":"https:\/\/17aitech.com\/?p=39623#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/17aitech.com\/"},{"@type":"ListItem","position":2,"name":"\u817e\u8baf\u6df7\u5143\u3001\u82f1\u4f1f\u8fbe\u90fd\u53d1\u6df7\u5408\u67b6\u6784\u6a21\u578b\uff0cMamba-Transformer\u8981\u5d1b\u8d77\u5417\uff1f"}]},{"@type":"WebSite","@id":"https:\/\/17aitech.com\/#website","url":"https:\/\/17aitech.com\/","name":"\u4e00\u8d77AI\u6280\u672f","description":"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca","alternateName":"\u4e00\u8d77AI\u6280\u672f","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/17aitech.com\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d","name":"AI\u5c0f\u52a9\u624b","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/#\/schema\/person\/image\/","url":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","contentUrl":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","caption":"AI\u5c0f\u52a9\u624b"},"description":"\u8fd9\u4e2a\u4eba\u5f88\u61d2\uff0c\u4ec0\u4e48\u90fd\u6ca1\u6709\u7559\u4e0b\uff5e","url":"https:\/\/17aitech.com\/?page_id=33738&user=3"}]}},"_links":{"self":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/39623","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/users\/3"}],"replies":[{"embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=39623"}],"version-history":[{"count":0,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/39623\/revisions"}],"wp:attachment":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=39623"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=39623"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=39623"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}