{"data":{"id":14725,"name":"Liquid AI\u63a8\u51fa\u65b0\u201c\u6db2\u6001\u201d\u8fb9\u7f18\u6a21\u578bHyena Edge\uff0c\u5b9e\u6d4b\u6548\u7387\u8d28\u91cf\u5747\u8d85Transformer","name_en":null,"summary":"Liquid AI \u6b63\u5f0f\u53d1\u5e03\u4e86\u9762\u5411\u8fb9\u7f18\u8bbe\u5907\u7684\u5168\u65b0 AI \u67b6\u6784\u201cHyena Edge\u201d\u3002","type":495,"labels":[{"id":211,"name":"AI"}],"authors":[{"id":364,"username":"DeepTech\u6df1\u79d1\u6280","avatar":"https:\/\/image.deeptechchina.com\/article\/2025082809412530504.jpg","summary":"\u53d1\u73b0\u65b0\u5174\u79d1\u6280\u3002"}],"cover":"https:\/\/image.deeptechchina.com\/article\/2025042818033326770.jpg","start_time":1746067440,"look_num":3453,"share_num":0,"status":1,"create_time":1745834652,"update_time":1746067483,"collect_num":0,"comment_num":0,"is_vip":2,"platform":"1,2","information_id":14725,"content_str":"\u8fd8\u8bb0\u5f97\u4e4b\u524d\u90a3\u4e2a\u63d0\u51fa\u4e86\u65b0\u7684\u6db2\u4f53\u57fa\u7840\u6a21\u578b\uff08LFM\uff0cLiquid Foundation Model\uff09\u7684 Liquid AI \u5417\uff1f\u5728\u88ab AMD \u6295\u8d44\u6570\u6708\u540e\uff0c\u8fd9\u5bb6\u6765\u81ea\u7f8e\u56fd\u9ebb\u7701\u7406\u5de5\u5b66\u9662\u7684\u521d\u521b\u516c\u53f8\u53c8\u5e26\u6765\u65b0\u6210\u679c\u4e86\u30024 \u6708 25 \u65e5\uff0cLiquid AI \u6b63\u5f0f\u53d1\u5e03\u4e86\u9762\u5411\u8fb9\u7f18\u8bbe\u5907\u7684\u5168\u65b0 AI \u67b6\u6784\u201cHyena Edge\u201d\u3002\u4e0e\u76ee\u524d\u4e3b\u6d41\u7684\u57fa\u4e8e Transformer \u7684\u6a21\u578b\u67b6\u6784\u4e0d\u540c\uff0cHyena Edge \u662f\u4e00\u79cd\u57fa\u4e8e\u5377\u79ef\u7684\u591a\u6df7\u5408\u6a21\u578b\uff0c\u4e13\u4e3a\u667a\u80fd\u624b\u673a\u7b49\u8fb9\u7f18\u8bbe\u5907\u4f18\u5316\u8bbe\u8ba1\u3002\u201c\u4eba\u5de5\u667a\u80fd\u6b63\u8fc5\u901f\u53d8\u5f97\u65e0\u5904\u4e0d\u5728\uff0c\u4ece\u5927\u89c4\u6a21\u4e91\u7aef\u90e8\u7f72\u5230\u667a\u80fd\u624b\u673a\u548c\u7b14\u8bb0\u672c\u7535\u8111\u7b49\u8d44\u6e90\u53d7\u9650\u7684\u8fb9\u7f18\u8bbe\u5907\u3002\u201dLiquid AI \u79d1\u5b66\u56e2\u961f\u6210\u5458 Armin Thomas\u3001Stefano Massaroli \u548c Michael Poli \u5728\u7814\u7a76\u62a5\u544a\u4e2d\u8868\u793a\uff0c\u201c\u5c3d\u7ba1\u6709\u4ee4\u4eba\u5370\u8c61\u6df1\u523b\u7684\u8fdb\u6b65\uff0c\u5927\u591a\u6570\u4e3a\u8fb9\u7f18\u90e8\u7f72\u4f18\u5316\u7684\u5c0f\u578b\u6a21\u578b\uff0c\u5982 SmolLM2\u3001Phi \u6a21\u578b\u548c Llama 3.2 1B\uff0c\u4e3b\u8981\u4f9d\u8d56\u4e8e\u57fa\u4e8e\u6ce8\u610f\u529b\u8fd0\u7b97\u7b26\u7684 Transformer \u67b6\u6784\u3002\u201d\u8fd9\u4e9b\u4f20\u7edf\u67b6\u6784\u867d\u7136\u5177\u6709\u53ef\u5e76\u884c\u8ba1\u7b97\u548c\u9ad8\u6548\u5185\u6838\u7684\u7279\u70b9\uff0c\u4f46\u5728\u8fb9\u7f18\u8bbe\u5907\u4e0a\u4ecd\u9762\u4e34\u6548\u7387\u74f6\u9888\u3002\u800c Hyena Edge \u4f5c\u4e3a\u4e00\u79cd Liquid \u67b6\u6784\uff0c\u5929\u751f\u5c31\u5728\u8ba1\u7b97\u6548\u7387\u4e0a\u5177\u6709\u4f18\u52bf\uff0c\u975e\u5e38\u9002\u7528\u4e8e\u8fb9\u7f18\u90e8\u7f72\u3002\u636e Liquid AI \u5ba3\u79f0\uff0cHyena Edge \u5728\u771f\u5b9e\u786c\u4ef6\u6d4b\u8bd5\u4e2d\uff0c\u4e8e\u8ba1\u7b97\u6548\u7387\u548c\u6a21\u578b\u8d28\u91cf\u4e24\u65b9\u9762\u5747\u5c55\u73b0\u51fa\u8d85\u8d8a Transformer \u57fa\u7ebf\u7684\u6027\u80fd\u3002\u4ed6\u4eec\u5728\u4e09\u661f Galaxy S24 Ultra \u4e0a\u5bf9 Hyena Edge \u8fdb\u884c\u4e86\u6d4b\u8bd5\uff0c\u7ed3\u679c\u663e\u793a\u8be5\u6a21\u578b\u5728\u591a\u4e2a\u5173\u952e\u6307\u6807\u4e0a\u5747\u8d85\u8d8a\u5f3a\u5927\u7684\u57fa\u4e8e Transformer \u7684\u57fa\u51c6\u6a21\u578b\u3002\u5728\u6548\u7387\u65b9\u9762\uff0cHyena Edge \u5c55\u793a\u4e86\u66f4\u5feb\u7684\u9884\u586b\u5145\uff08prefill\uff09\u548c\u89e3\u7801\u5ef6\u8fdf\u3002\u7279\u522b\u662f\u5bf9\u4e8e\u8d85\u8fc7 256 \u4e2a\u6807\u8bb0\u7684\u5e8f\u5217\uff0c\u89e3\u7801\u548c\u9884\u586b\u5145\u5ef6\u8fdf\u6700\u9ad8\u53ef\u63d0\u9ad8 30%\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u5176\u5728\u77ed\u5e8f\u5217\u957f\u5ea6\u4e0b\u7684\u9884\u586b\u5145\u5ef6\u8fdf\u4e5f\u4f18\u4e8e Transformer \u57fa\u7ebf\uff0c\u8fd9\u4e00\u70b9\u5bf9\u54cd\u5e94\u5f0f\u8bbe\u5907\u5e94\u7528\u81f3\u5173\u91cd\u8981\u3002\u5728\u5185\u5b58\u4f7f\u7528\u65b9\u9762\uff0cHyena Edge \u5728\u6240\u6709\u6d4b\u8bd5\u5e8f\u5217\u957f\u5ea6\u4e0b\u5747\u4f7f\u7528\u66f4\u5c11\u7684\u5185\u5b58\u3002\u56fe\u4e28\u5728\u4e09\u661f S24 Ultra \u667a\u80fd\u624b\u673a\u4e0a\u6536\u96c6\u7684\u5ef6\u8fdf\u548c\u5185\u5b58\u914d\u7f6e\u6587\u4ef6\uff08\u6765\u6e90\uff1aLiquid AI\uff09\u5728\u6a21\u578b\u8d28\u91cf\u65b9\u9762\uff0cHyena Edge \u5728\u7ecf\u8fc7 1000 \u4ebf tokens \u7684\u8bad\u7ec3\u540e\uff0c\u5728\u5305\u62ec Wikitext\u3001Lambada\u3001Hellaswag\u3001Winogrande\u3001Piqa\u3001Arc-easy \u548c Arc-challenge \u5728\u5185\u7684\u5404\u79cd\u5e38\u89c1\u8bed\u8a00\u5efa\u6a21\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u8868\u73b0\u51fa\u8272\u3002\u4f8b\u5982\uff0c\u5728 Wikitext \u4e0a\u7684\u56f0\u60d1\u5ea6\uff08perplexity\uff09\u4ece 17.3 \u964d\u81f3 16.2\uff0c\u5728 Lambada \u4e0a\u4ece 10.8 \u964d\u81f3 9.4\uff0c\u5728 PiQA \u4e0a\u7684\u51c6\u786e\u7387\u4ece 71.1% \u63d0\u5347\u81f3 72.3%\uff0c\u5728 Hellaswag \u4e0a\u4ece 49.3% \u63d0\u5347\u81f3 52.8%\uff0c\u5728 Winogrande \u4e0a\u4ece 51.4% \u63d0\u5347\u81f3 54.8%\u3002\u201c\u8fd9\u4e9b\u7ed3\u679c\u8868\u660e\uff0c\u6a21\u578b\u7684\u6548\u7387\u63d0\u5347\u5e76\u4e0d\u4ee5\u9884\u6d4b\u8d28\u91cf\u4e3a\u4ee3\u4ef7\u2014\u2014\u8fd9\u662f\u8bb8\u591a\u8fb9\u7f18\u4f18\u5316\u67b6\u6784\u5e38\u89c1\u7684\u53d6\u820d\u3002\u201d\u7814\u7a76\u56e2\u961f\u8868\u793a\u3002\u800c Hyena Edge \u7684\u6838\u5fc3\u6280\u672f\uff0c\u6b63\u5728\u4e8e\u56e2\u961f\u6b64\u524d\u63d0\u51fa\u7684 STAR\uff08Synthesis of Tailored Architectures\uff09\u6846\u67b6\u53ca\u5176\u4f18\u5316\u6280\u672f\u3002STAR \u6846\u67b6\u7684\u6838\u5fc3\u601d\u60f3\u662f\u5229\u7528\u8fdb\u5316\u7b97\u6cd5\uff08Evolutionary Algorithms\uff09\u548c\u7ebf\u6027\u8f93\u5165\u53d8\u5316\u7cfb\u7edf\uff08LIVs\uff0cLinear Input-Varying Systems\uff09\u7684\u6570\u5b66\u7406\u8bba\uff0c\u6765\u9ad8\u6548\u5730\u63a2\u7d22\u5e7f\u9614\u7684\u795e\u7ecf\u7f51\u7edc\u67b6\u6784\u7a7a\u95f4\uff0c\u5e76\u81ea\u52a8\u5408\u6210\u51fa\u9488\u5bf9\u7279\u5b9a\u76ee\u6807\uff08\u5982\u4f4e\u5ef6\u8fdf\u3001\u5c0f\u5185\u5b58\u5360\u7528\u3001\u9ad8\u6a21\u578b\u8d28\u91cf\u3001\u5c0f\u53c2\u6570\u91cf\u7b49\uff0c\u53ef\u591a\u76ee\u6807\u540c\u65f6\u4f18\u5316\uff09\u7684\u201c\u91cf\u8eab\u5b9a\u5236\u201d\u7684\u67b6\u6784\u3002\u56fe\u4e28 STAR \u6846\u67b6\u8bf4\u660e\uff08\u6765\u6e90\uff1aLiquid AI\uff09\u4e0e\u4f20\u7edf\u7684\u4f9d\u8d56\u4eba\u5de5\u7ecf\u9a8c\u548c\u76f4\u89c9\u8fdb\u884c\u6a21\u578b\u8bbe\u8ba1\uff0c\u6216\u5728\u6709\u9650\u7a7a\u95f4\u5185\u8fdb\u884c\u81ea\u52a8\u5316\u641c\u7d22\u7684\u65b9\u6cd5\u4e0d\u540c\uff0cSTAR \u63d0\u4f9b\u4e86\u4e00\u4e2a\u66f4\u5168\u9762\u7684\u89e3\u51b3\u65b9\u6848\u3002LIV \u7406\u8bba\u662f\u4e00\u4e2a\u5173\u952e\u7684\u7406\u8bba\u57fa\u7840\uff0c\u5b83\u80fd\u591f\u7edf\u4e00\u5730\u63cf\u8ff0\u548c\u6cdb\u5316\u6df1\u5ea6\u5b66\u4e60\u4e2d\u5e38\u89c1\u7684\u591a\u79cd\u8ba1\u7b97\u5355\u5143\uff0c\u5305\u62ec\u5404\u79cd\u6ce8\u610f\u529b\u53d8\u4f53\u3001\u7ebf\u6027\u5faa\u73af\u7f51\u7edc\u3001\u5377\u79ef\u7f51\u7edc\u4ee5\u53ca\u5176\u4ed6\u7ed3\u6784\u5316\u7b97\u5b50\u3002STAR \u57fa\u4e8e LIV \u7406\u8bba\uff0c\u6784\u5efa\u4e86\u4e00\u4e2a\u65b0\u9896\u7684\u3001\u5206\u5c42\u7684\u67b6\u6784\u641c\u7d22\u7a7a\u95f4\u3002\u5728\u8fd9\u4e2a\u6846\u67b6\u4e0b\uff0c\u6a21\u578b\u67b6\u6784\u88ab\u7f16\u7801\u6210\u4e00\u79cd\u201c\u57fa\u56e0\u7ec4\uff08Genome\uff09\u201d\u3002\u8fd9\u79cd\u57fa\u56e0\u7ec4\u5305\u542b\u4e86\u67b6\u6784\u7684\u591a\u4e2a\u5c42\u7ea7\u4fe1\u606f\uff0c\u4ece\u5e95\u5c42\u7684\u7279\u5f81\u5316\uff08Featurization\uff09\u65b9\u5f0f\u3001\u7b97\u5b50\u7ed3\u6784\uff08Operator Structure\uff0c\u5b9a\u4e49\u4e86 Token \u548c Channel \u7684\u6df7\u5408\u65b9\u5f0f\uff09\uff0c\u5230\u9876\u5c42\u7684\u9aa8\u5e72\u7f51\u7edc\uff08Backbone\uff0c\u5b9a\u4e49\u4e86 LIV \u5355\u5143\u4e4b\u95f4\u7684\u8fde\u63a5\u548c\u7ec4\u5408\u65b9\u5f0f\uff09\u3002STAR \u7684\u57fa\u56e0\u7ec4\u8bbe\u8ba1\u5177\u6709\u826f\u597d\u7684\u5c42\u6b21\u5316\u548c\u6a21\u5757\u5316\u7279\u6027\u3002\u968f\u540e\uff0cSTAR \u8fd0\u7528\u8fdb\u5316\u7b97\u6cd5\u7684\u539f\u7406\uff0c\u5bf9\u8fd9\u4e9b\u67b6\u6784\u57fa\u56e0\u7ec4\u8fdb\u884c\u8fed\u4ee3\u4f18\u5316\u3002\u4e3b\u8981\u5305\u62ec\u8bc4\u4f30\uff08\u6839\u636e\u9884\u8bbe\u76ee\u6807\u8861\u91cf\u67b6\u6784\u6027\u80fd\uff09\u3001\u91cd\u7ec4\uff08\u7ec4\u5408\u4f18\u79c0\u7236\u4ee3\u67b6\u6784\u7684\u7279\u5f81\uff09\u548c\u53d8\u5f02\uff08\u5f15\u5165\u968f\u673a\u53d8\u5316\u4ee5\u63a2\u7d22\u65b0\u67b6\u6784\uff09\u7b49\u6b65\u9aa4\u3002\u8be5\u6846\u67b6\u652f\u6301\u591a\u76ee\u6807\u4f18\u5316\uff0c\u80fd\u591f\u540c\u65f6\u8003\u8651\u6a21\u578b\u8d28\u91cf\u3001\u53c2\u6570\u91cf\u3001\u63a8\u7406\u7f13\u5b58\u5927\u5c0f\u3001\u5ef6\u8fdf\u7b49\u591a\u4e2a\u53ef\u80fd\u76f8\u4e92\u51b2\u7a81\u7684\u6307\u6807\uff0c\u4ee5\u5bfb\u627e\u5728\u8fd9\u4e9b\u76ee\u6807\u95f4\u8fbe\u5230\u8f83\u597d\u5e73\u8861\u7684\u67b6\u6784\u65b9\u6848\u3002\u6839\u636e\u5176\u6280\u672f\u6587\u6863\uff0cSTAR \u5728\u4f18\u5316\u5927\u578b\u8bed\u8a00\u6a21\u578b\u67b6\u6784\u65f6\uff0c\u65e0\u8bba\u662f\u5728\u7eaf\u8d28\u91cf\u4f18\u5316\u3001\u8d28\u91cf\u4e0e\u53c2\u6570\u91cf\u8054\u5408\u4f18\u5316\uff0c\u8fd8\u662f\u8d28\u91cf\u4e0e\u63a8\u7406\u7f13\u5b58\u5927\u5c0f\u8054\u5408\u4f18\u5316\u65b9\u9762\uff0c\u5176\u751f\u6210\u7684\u67b6\u6784\u5747\u80fd\u5728\u5404\u9879\u6307\u6807\u4e0a\u663e\u8457\u4f18\u4e8e\u9ad8\u5ea6\u4f18\u5316\u7684 Transformer++ \u548c StripedMamba \u7b49\u57fa\u51c6\u6a21\u578b\u3002\u4f8b\u5982\uff0c\u5728\u4f18\u5316\u8d28\u91cf\u548c\u53c2\u6570\u91cf\u65f6\uff0c7\/8 \u7684 STAR \u8fdb\u5316\u67b6\u6784\u5728\u4e0b\u6e38\u4efb\u52a1\u57fa\u51c6\u4e0a\u8d85\u8d8a\u4e86 Transformer++ \u548c\u6df7\u5408\u6a21\u578b\uff0c\u540c\u65f6\u53c2\u6570\u91cf\u51cf\u5c11\u9ad8\u8fbe 13%\u3002\u5728\u4f18\u5316\u8d28\u91cf\u548c\u7f13\u5b58\u5927\u5c0f\u65f6\uff0c7\/8 \u7684 STAR \u8fdb\u5316\u67b6\u6784\u5b9e\u73b0\u4e86\u6bd4\u6df7\u5408\u6a21\u578b\u5c0f 37%\u3001\u6bd4 Transformer \u5c0f 90% \u7684\u7f13\u5b58\uff0c\u540c\u65f6\u8d28\u91cf\u76f8\u5f53\u751a\u81f3\u66f4\u597d\u3002\u5728 Hyena Edge \u7684\u8bbe\u8ba1\u4e2d\uff0cLiquid AI \u56e2\u961f\u5c31\u5e94\u7528\u4e86 STAR \u6846\u67b6\u3002\u4ed6\u4eec\u4ece\u4e00\u4e2a\u5305\u542b 16 \u4e2a\u5019\u9009\u67b6\u6784\u7684\u521d\u59cb\u79cd\u7fa4\u5f00\u59cb\uff0c\u8fdb\u884c\u4e86 24 \u4ee3\u7684\u8fdb\u5316\u8fed\u4ee3\u3002\u5176\u641c\u7d22\u7a7a\u95f4\u8bbe\u8ba1\u5f97\u975e\u5e38\u4e30\u5bcc\uff0c\u5305\u542b\u4e86\u591a\u79cd\u5377\u79ef\u7b97\u5b50\u7684\u53d8\u4f53\uff0c\u8fd9\u4e9b\u53d8\u4f53\u4e3b\u8981\u53d7\u5230 Hyena \u67b6\u6784\u7684\u542f\u53d1\uff1aHyena\uff08Full\uff09\uff1a\u5728\u95e8\u63a7\u673a\u5236\uff08Gating Mechanism\uff09\u548c Hyena \u5185\u90e8\u5377\u79ef\u4e2d\u90fd\u5305\u542b\u5377\u79ef\u64cd\u4f5c\u3002Hyena-X\uff1a\u6392\u9664\u4e86\u5185\u90e8\u5377\u79ef\u3002Hyena-Y\uff1a\u6392\u9664\u4e86\u7279\u5f81\u7ec4\uff08\u95e8\uff09\u4e2d\u7684\u5377\u79ef\u3002\u9664\u4e86\u8fd9\u4e09\u79cd\u4e3b\u8981\u7684 Hyena \u7c7b\u578b\uff0c\u641c\u7d22\u7a7a\u95f4\u8fd8\u8003\u8651\u4e86\u5b83\u4eec\u5b66\u4e60\u5230\u7684\u77ed\u8ddd\u79bb\u3001\u663e\u5f0f\u5377\u79ef\u6838\u7684\u957f\u5ea6\u53d8\u5316\uff08\u4ece 3 \u5230 128\uff09\uff0c\u603b\u5171\u6784\u6210\u4e86 18 \u79cd\u4e0d\u540c\u7684\u5377\u79ef\u7b97\u5b50\u3002\u6b64\u5916\uff0c\u641c\u7d22\u7a7a\u95f4\u8fd8\u5305\u62ec\u4e86\u5206\u7ec4\u67e5\u8be2\u6ce8\u610f\u529b\u7684\u53d8\u4f53\uff08\u5177\u6709\u4e0d\u540c\u7684 KV \u5934\u6570\u91cf\uff09\u548c SwiGLU\uff08\u5177\u6709\u4e0d\u540c\u7684\u5185\u90e8\u5bbd\u5ea6\uff09\u7b49\u5e38\u89c1\u7684 Transformer \u7ec4\u4ef6\u3002\u5728\u8fdb\u5316\u8fc7\u7a0b\u4e2d\uff0cSTAR \u6846\u67b6\u901a\u8fc7\u5bf9\u6bcf\u4e2a\u5019\u9009\u67b6\u6784\u5728\u4e09\u661f S24 Ultra \u4e0a\u7684\u521d\u59cb\u5ef6\u8fdf\u548c\u5185\u5b58\u5360\u7528\u8fdb\u884c\u6027\u80fd\u5206\u6790\uff0c\u5e76\u7ed3\u5408\u6a21\u578b\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u7684\u56f0\u60d1\u5ea6\u8868\u73b0\uff0c\u4e0d\u65ad\u5730\u5c06\u67b6\u6784\u79cd\u7fa4\u5411\u7740\u6548\u7387-\u8d28\u91cf\u524d\u6cbf\uff08Efficiency-Quality Frontier\uff09\u8fdb\u884c\u4f18\u5316\u3002\u6709\u610f\u601d\u7684\u662f\uff0c\u968f\u7740\u8fdb\u5316\u8fc7\u7a0b\u7684\u63a8\u8fdb\uff0c\u5f53\u67b6\u6784\u63a5\u8fd1\u6548\u7387-\u8d28\u91cf\u6700\u4f18\u8fb9\u754c\u65f6\uff0cSTAR \u660e\u663e\u5730\u504f\u7231\u9009\u7528 Hyena-Y \u7c7b\u578b\u7684\u5377\u79ef\u3002\u8fd9\u8868\u660e Hyena-Y \u5377\u79ef\u5728\u5ef6\u8fdf\u3001\u5185\u5b58\u548c\u6a21\u578b\u8d28\u91cf\u4e4b\u95f4\u53d6\u5f97\u4e86\u66f4\u4f18\u8d8a\u7684\u5e73\u8861\u3002\u57fa\u4e8e\u8fd9\u4e00\u53d1\u73b0\uff0c\u6700\u7ec8\u7684 Hyena Edge \u67b6\u6784\u5728\u4e00\u4e2a GQA-Transformer++ \u57fa\u51c6\u6a21\u578b\u7684\u57fa\u7840\u4e0a\uff0c\u5c06\u90e8\u5206 GQA \u7b97\u5b50\u66ff\u6362\u4e3a\u7531 STAR \u4f18\u5316\u9009\u51fa\u7684 Hyena-Y \u95e8\u63a7\u5377\u79ef\u3002\u800c\u6700\u7ec8\u7ed3\u679c\u5982\u57fa\u51c6\u6d4b\u8bd5\u6240\u8868\u660e\u7684\uff0cHyena Edge \u5728\u63d0\u5347\u6548\u7387\u7684\u540c\u65f6\uff0c\u4fdd\u6301\u4e86\u8f83\u9ad8\u7684\u6a21\u578b\u8d28\u91cf\uff0c\u8fd9\u5bf9\u4e8e\u6027\u80fd\u548c\u8d44\u6e90\u90fd\u53d7\u9650\u7684\u8fb9\u7f18\u8bbe\u5907\u5e94\u7528\u800c\u8a00\u662f\u4e00\u4e2a\u91cd\u8981\u7684\u7279\u6027\u3002Liquid AI \u5df2\u8868\u793a\u8ba1\u5212\u5728\u672a\u6765\u6570\u6708\u5185\u5f00\u6e90\u5305\u62ec Hyena Edge \u5728\u5185\u7684\u4e00\u7cfb\u5217\u57fa\u7840\u6a21\u578b\uff0c\u5176\u76ee\u6807\u662f\u6784\u5efa\u80fd\u591f\u9002\u5e94\u4ece\u4e91\u7aef\u5230\u8fb9\u7f18\u591a\u79cd\u73af\u5883\u7684 AI \u7cfb\u7edf\u3002\u9664\u4e86\u6a21\u578b\u672c\u8eab\uff0c\u5176\u6240\u5c55\u73b0\u51fa\u7684\u8bbe\u8ba1\u65b9\u6cd5\uff0c\u6216\u8bb8\u66f4\u503c\u5f97\u6211\u4eec\u7684\u671f\u5f85\u3002\u53c2\u8003\u8d44\u6599\uff1a1.https:\/\/www.liquid.ai\/research\/convolutional-multi-hybrids-for-edge-devices2.https:\/\/arxiv.org\/abs\/2411.178003.https:\/\/venturebeat.com\/ai\/liquid-ai-is-revolutionizing-llms-to-work-on-edge-devices-like-smartphones-with-new-hyena-edge-model\/\u8fd0\u8425\/\u6392\u7248\uff1a\u4f55\u6668\u9f99","content":"

\u8fd8\u8bb0\u5f97\u4e4b\u524d\u90a3\u4e2a\u63d0\u51fa\u4e86\u65b0\u7684\u6db2\u4f53\u57fa\u7840\u6a21\u578b\uff08LFM\uff0cLiquid Foundation Model\uff09\u7684 Liquid AI \u5417\uff1f\u5728\u88ab AMD \u6295\u8d44\u6570\u6708\u540e\uff0c\u8fd9\u5bb6\u6765\u81ea\u7f8e\u56fd\u9ebb\u7701\u7406\u5de5\u5b66\u9662\u7684\u521d\u521b\u516c\u53f8\u53c8\u5e26\u6765\u65b0\u6210\u679c\u4e86\u3002<\/span><\/span><\/p>

<\/p>

4 \u6708 25 \u65e5\uff0cLiquid AI \u6b63\u5f0f\u53d1\u5e03\u4e86\u9762\u5411\u8fb9\u7f18\u8bbe\u5907\u7684\u5168\u65b0 AI \u67b6\u6784\u201cHyena Edge\u201d\u3002\u4e0e\u76ee\u524d\u4e3b\u6d41\u7684\u57fa\u4e8e Transformer \u7684\u6a21\u578b\u67b6\u6784\u4e0d\u540c\uff0cHyena Edge \u662f\u4e00\u79cd\u57fa\u4e8e\u5377\u79ef\u7684\u591a\u6df7\u5408\u6a21\u578b\uff0c\u4e13\u4e3a\u667a\u80fd\u624b\u673a\u7b49\u8fb9\u7f18\u8bbe\u5907\u4f18\u5316\u8bbe\u8ba1\u3002<\/span><\/span><\/p>

<\/p>

\u201c\u4eba\u5de5\u667a\u80fd\u6b63\u8fc5\u901f\u53d8\u5f97\u65e0\u5904\u4e0d\u5728\uff0c\u4ece\u5927\u89c4\u6a21\u4e91\u7aef\u90e8\u7f72\u5230\u667a\u80fd\u624b\u673a\u548c\u7b14\u8bb0\u672c\u7535\u8111\u7b49\u8d44\u6e90\u53d7\u9650\u7684\u8fb9\u7f18\u8bbe\u5907\u3002\u201dLiquid AI \u79d1\u5b66\u56e2\u961f\u6210\u5458 Armin Thomas\u3001Stefano Massaroli \u548c Michael Poli \u5728\u7814\u7a76\u62a5\u544a\u4e2d\u8868\u793a\uff0c\u201c\u5c3d\u7ba1\u6709\u4ee4\u4eba\u5370\u8c61\u6df1\u523b\u7684\u8fdb\u6b65\uff0c\u5927\u591a\u6570\u4e3a\u8fb9\u7f18\u90e8\u7f72\u4f18\u5316\u7684\u5c0f\u578b\u6a21\u578b\uff0c\u5982 SmolLM2\u3001Phi \u6a21\u578b\u548c Llama 3.2 1B\uff0c\u4e3b\u8981\u4f9d\u8d56\u4e8e\u57fa\u4e8e\u6ce8\u610f\u529b\u8fd0\u7b97\u7b26\u7684 Transformer \u67b6\u6784\u3002\u201d<\/span><\/span><\/p>

<\/p>

\u8fd9\u4e9b\u4f20\u7edf\u67b6\u6784\u867d\u7136\u5177\u6709\u53ef\u5e76\u884c\u8ba1\u7b97\u548c\u9ad8\u6548\u5185\u6838\u7684\u7279\u70b9\uff0c\u4f46\u5728\u8fb9\u7f18\u8bbe\u5907\u4e0a\u4ecd\u9762\u4e34\u6548\u7387\u74f6\u9888\u3002\u800c Hyena Edge \u4f5c\u4e3a\u4e00\u79cd Liquid \u67b6\u6784\uff0c\u5929\u751f\u5c31\u5728\u8ba1\u7b97\u6548\u7387\u4e0a\u5177\u6709\u4f18\u52bf\uff0c\u975e\u5e38\u9002\u7528\u4e8e\u8fb9\u7f18\u90e8\u7f72\u3002\u636e Liquid AI \u5ba3\u79f0\uff0cHyena Edge \u5728\u771f\u5b9e\u786c\u4ef6\u6d4b\u8bd5\u4e2d\uff0c\u4e8e\u8ba1\u7b97\u6548\u7387\u548c\u6a21\u578b\u8d28\u91cf\u4e24\u65b9\u9762\u5747\u5c55\u73b0\u51fa\u8d85\u8d8a Transformer \u57fa\u7ebf\u7684\u6027\u80fd\u3002<\/span><\/span><\/p>

<\/p>

\u4ed6\u4eec\u5728\u4e09\u661f Galaxy S24 Ultra \u4e0a\u5bf9 Hyena Edge \u8fdb\u884c\u4e86\u6d4b\u8bd5\uff0c\u7ed3\u679c\u663e\u793a\u8be5\u6a21\u578b\u5728\u591a\u4e2a\u5173\u952e\u6307\u6807\u4e0a\u5747\u8d85\u8d8a\u5f3a\u5927\u7684\u57fa\u4e8e Transformer \u7684\u57fa\u51c6\u6a21\u578b\u3002<\/span><\/span><\/p>

<\/p>

\u5728\u6548\u7387\u65b9\u9762\uff0cHyena Edge \u5c55\u793a\u4e86\u66f4\u5feb\u7684\u9884\u586b\u5145\uff08prefill\uff09\u548c\u89e3\u7801\u5ef6\u8fdf\u3002\u7279\u522b\u662f\u5bf9\u4e8e\u8d85\u8fc7 256 \u4e2a\u6807\u8bb0\u7684\u5e8f\u5217\uff0c\u89e3\u7801\u548c\u9884\u586b\u5145\u5ef6\u8fdf\u6700\u9ad8\u53ef\u63d0\u9ad8 30%\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u5176\u5728\u77ed\u5e8f\u5217\u957f\u5ea6\u4e0b\u7684\u9884\u586b\u5145\u5ef6\u8fdf\u4e5f\u4f18\u4e8e Transformer \u57fa\u7ebf\uff0c\u8fd9\u4e00\u70b9\u5bf9\u54cd\u5e94\u5f0f\u8bbe\u5907\u5e94\u7528\u81f3\u5173\u91cd\u8981\u3002\u5728\u5185\u5b58\u4f7f\u7528\u65b9\u9762\uff0cHyena Edge \u5728\u6240\u6709\u6d4b\u8bd5\u5e8f\u5217\u957f\u5ea6\u4e0b\u5747\u4f7f\u7528\u66f4\u5c11\u7684\u5185\u5b58\u3002<\/span><\/span><\/p>

<\/p>

\u56fe\u4e28\u5728\u4e09\u661f S24 Ultra \u667a\u80fd\u624b\u673a\u4e0a\u6536\u96c6\u7684\u5ef6\u8fdf\u548c\u5185\u5b58\u914d\u7f6e\u6587\u4ef6\uff08\u6765\u6e90\uff1aLiquid AI\uff09<\/div>
<\/a>

\u5728\u6a21\u578b\u8d28\u91cf\u65b9\u9762\uff0cHyena Edge \u5728\u7ecf\u8fc7 1000 \u4ebf tokens \u7684\u8bad\u7ec3\u540e\uff0c\u5728\u5305\u62ec Wikitext\u3001Lambada\u3001Hellaswag\u3001Winogrande\u3001Piqa\u3001Arc-easy \u548c Arc-challenge \u5728\u5185\u7684\u5404\u79cd\u5e38\u89c1\u8bed\u8a00\u5efa\u6a21\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u8868\u73b0\u51fa\u8272\u3002\u4f8b\u5982\uff0c\u5728 Wikitext \u4e0a\u7684\u56f0\u60d1\u5ea6\uff08perplexity\uff09\u4ece 17.3 \u964d\u81f3 16.2\uff0c\u5728 Lambada \u4e0a\u4ece 10.8 \u964d\u81f3 9.4\uff0c\u5728 PiQA \u4e0a\u7684\u51c6\u786e\u7387\u4ece 71.1% \u63d0\u5347\u81f3 72.3%\uff0c\u5728 Hellaswag \u4e0a\u4ece 49.3% \u63d0\u5347\u81f3 52.8%\uff0c\u5728 Winogrande \u4e0a\u4ece 51.4% \u63d0\u5347\u81f3 54.8%\u3002<\/span><\/span><\/p>

\u201c\u8fd9\u4e9b\u7ed3\u679c\u8868\u660e\uff0c\u6a21\u578b\u7684\u6548\u7387\u63d0\u5347\u5e76\u4e0d\u4ee5\u9884\u6d4b\u8d28\u91cf\u4e3a\u4ee3\u4ef7\u2014\u2014\u8fd9\u662f\u8bb8\u591a\u8fb9\u7f18\u4f18\u5316\u67b6\u6784\u5e38\u89c1\u7684\u53d6\u820d\u3002\u201d\u7814\u7a76\u56e2\u961f\u8868\u793a\u3002<\/span><\/span><\/p>

\u800c Hyena Edge \u7684\u6838\u5fc3\u6280\u672f\uff0c\u6b63\u5728\u4e8e\u56e2\u961f\u6b64\u524d\u63d0\u51fa\u7684 STAR\uff08Synthesis of Tailored Architectures\uff09\u6846\u67b6\u53ca\u5176\u4f18\u5316\u6280\u672f\u3002STAR \u6846\u67b6\u7684\u6838\u5fc3\u601d\u60f3\u662f\u5229\u7528\u8fdb\u5316\u7b97\u6cd5\uff08Evolutionary Algorithms\uff09\u548c\u7ebf\u6027\u8f93\u5165\u53d8\u5316\u7cfb\u7edf\uff08LIVs\uff0cLinear Input-Varying Systems\uff09\u7684\u6570\u5b66\u7406\u8bba\uff0c\u6765\u9ad8\u6548\u5730\u63a2\u7d22\u5e7f\u9614\u7684\u795e\u7ecf\u7f51\u7edc\u67b6\u6784\u7a7a\u95f4\uff0c\u5e76\u81ea\u52a8\u5408\u6210\u51fa\u9488\u5bf9\u7279\u5b9a\u76ee\u6807\uff08\u5982\u4f4e\u5ef6\u8fdf\u3001\u5c0f\u5185\u5b58\u5360\u7528\u3001\u9ad8\u6a21\u578b\u8d28\u91cf\u3001\u5c0f\u53c2\u6570\u91cf\u7b49\uff0c\u53ef\u591a\u76ee\u6807\u540c\u65f6\u4f18\u5316\uff09\u7684\u201c\u91cf\u8eab\u5b9a\u5236\u201d\u7684\u67b6\u6784\u3002<\/span><\/span><\/p>

\u56fe\u4e28 STAR \u6846\u67b6\u8bf4\u660e\uff08\u6765\u6e90\uff1aLiquid AI\uff09<\/div>
<\/a>

\u4e0e\u4f20\u7edf\u7684\u4f9d\u8d56\u4eba\u5de5\u7ecf\u9a8c\u548c\u76f4\u89c9\u8fdb\u884c\u6a21\u578b\u8bbe\u8ba1\uff0c\u6216\u5728\u6709\u9650\u7a7a\u95f4\u5185\u8fdb\u884c\u81ea\u52a8\u5316\u641c\u7d22\u7684\u65b9\u6cd5\u4e0d\u540c\uff0cSTAR \u63d0\u4f9b\u4e86\u4e00\u4e2a\u66f4\u5168\u9762\u7684\u89e3\u51b3\u65b9\u6848\u3002LIV \u7406\u8bba\u662f\u4e00\u4e2a\u5173\u952e\u7684\u7406\u8bba\u57fa\u7840\uff0c\u5b83\u80fd\u591f\u7edf\u4e00\u5730\u63cf\u8ff0\u548c\u6cdb\u5316\u6df1\u5ea6\u5b66\u4e60\u4e2d\u5e38\u89c1\u7684\u591a\u79cd\u8ba1\u7b97\u5355\u5143\uff0c\u5305\u62ec\u5404\u79cd\u6ce8\u610f\u529b\u53d8\u4f53\u3001\u7ebf\u6027\u5faa\u73af\u7f51\u7edc\u3001\u5377\u79ef\u7f51\u7edc\u4ee5\u53ca\u5176\u4ed6\u7ed3\u6784\u5316\u7b97\u5b50\u3002STAR \u57fa\u4e8e LIV \u7406\u8bba\uff0c\u6784\u5efa\u4e86\u4e00\u4e2a\u65b0\u9896\u7684\u3001\u5206\u5c42\u7684\u67b6\u6784\u641c\u7d22\u7a7a\u95f4\u3002<\/span><\/span><\/p>

\u5728\u8fd9\u4e2a\u6846\u67b6\u4e0b\uff0c\u6a21\u578b\u67b6\u6784\u88ab\u7f16\u7801\u6210\u4e00\u79cd\u201c\u57fa\u56e0\u7ec4\uff08Genome\uff09\u201d\u3002\u8fd9\u79cd\u57fa\u56e0\u7ec4\u5305\u542b\u4e86\u67b6\u6784\u7684\u591a\u4e2a\u5c42\u7ea7\u4fe1\u606f\uff0c\u4ece\u5e95\u5c42\u7684\u7279\u5f81\u5316\uff08Featurization\uff09\u65b9\u5f0f\u3001\u7b97\u5b50\u7ed3\u6784\uff08Operator Structure\uff0c\u5b9a\u4e49\u4e86 Token \u548c Channel \u7684\u6df7\u5408\u65b9\u5f0f\uff09\uff0c\u5230\u9876\u5c42\u7684\u9aa8\u5e72\u7f51\u7edc\uff08Backbone\uff0c\u5b9a\u4e49\u4e86 LIV \u5355\u5143\u4e4b\u95f4\u7684\u8fde\u63a5\u548c\u7ec4\u5408\u65b9\u5f0f\uff09\u3002STAR \u7684\u57fa\u56e0\u7ec4\u8bbe\u8ba1\u5177\u6709\u826f\u597d\u7684\u5c42\u6b21\u5316\u548c\u6a21\u5757\u5316\u7279\u6027\u3002<\/span><\/span><\/p>

\u968f\u540e\uff0cSTAR \u8fd0\u7528\u8fdb\u5316\u7b97\u6cd5\u7684\u539f\u7406\uff0c\u5bf9\u8fd9\u4e9b\u67b6\u6784\u57fa\u56e0\u7ec4\u8fdb\u884c\u8fed\u4ee3\u4f18\u5316\u3002\u4e3b\u8981\u5305\u62ec\u8bc4\u4f30\uff08\u6839\u636e\u9884\u8bbe\u76ee\u6807\u8861\u91cf\u67b6\u6784\u6027\u80fd\uff09\u3001\u91cd\u7ec4\uff08\u7ec4\u5408\u4f18\u79c0\u7236\u4ee3\u67b6\u6784\u7684\u7279\u5f81\uff09\u548c\u53d8\u5f02\uff08\u5f15\u5165\u968f\u673a\u53d8\u5316\u4ee5\u63a2\u7d22\u65b0\u67b6\u6784\uff09\u7b49\u6b65\u9aa4\u3002\u8be5\u6846\u67b6\u652f\u6301\u591a\u76ee\u6807\u4f18\u5316\uff0c\u80fd\u591f\u540c\u65f6\u8003\u8651\u6a21\u578b\u8d28\u91cf\u3001\u53c2\u6570\u91cf\u3001\u63a8\u7406\u7f13\u5b58\u5927\u5c0f\u3001\u5ef6\u8fdf\u7b49\u591a\u4e2a\u53ef\u80fd\u76f8\u4e92\u51b2\u7a81\u7684\u6307\u6807\uff0c\u4ee5\u5bfb\u627e\u5728\u8fd9\u4e9b\u76ee\u6807\u95f4\u8fbe\u5230\u8f83\u597d\u5e73\u8861\u7684\u67b6\u6784\u65b9\u6848\u3002<\/span><\/span><\/p>

\u6839\u636e\u5176\u6280\u672f\u6587\u6863\uff0cSTAR \u5728\u4f18\u5316\u5927\u578b\u8bed\u8a00\u6a21\u578b\u67b6\u6784\u65f6\uff0c\u65e0\u8bba\u662f\u5728\u7eaf\u8d28\u91cf\u4f18\u5316\u3001\u8d28\u91cf\u4e0e\u53c2\u6570\u91cf\u8054\u5408\u4f18\u5316\uff0c<\/span>\u8fd8\u662f\u8d28\u91cf\u4e0e\u63a8\u7406\u7f13\u5b58\u5927\u5c0f\u8054\u5408\u4f18\u5316\u65b9\u9762\uff0c\u5176\u751f\u6210\u7684\u67b6\u6784\u5747\u80fd\u5728\u5404\u9879\u6307\u6807\u4e0a\u663e\u8457\u4f18\u4e8e\u9ad8\u5ea6\u4f18\u5316\u7684 Transformer++ \u548c StripedMamba \u7b49\u57fa\u51c6\u6a21\u578b\u3002<\/span><\/strong>\u4f8b\u5982\uff0c\u5728\u4f18\u5316\u8d28\u91cf\u548c\u53c2\u6570\u91cf\u65f6\uff0c7\/8 \u7684 STAR \u8fdb\u5316\u67b6\u6784\u5728\u4e0b\u6e38\u4efb\u52a1\u57fa\u51c6\u4e0a\u8d85\u8d8a\u4e86 Transformer++ \u548c\u6df7\u5408\u6a21\u578b\uff0c\u540c\u65f6\u53c2\u6570\u91cf\u51cf\u5c11\u9ad8\u8fbe 13%\u3002\u5728\u4f18\u5316\u8d28\u91cf\u548c\u7f13\u5b58\u5927\u5c0f\u65f6\uff0c7\/8 \u7684 STAR \u8fdb\u5316\u67b6\u6784\u5b9e\u73b0\u4e86\u6bd4\u6df7\u5408\u6a21\u578b\u5c0f 37%\u3001\u6bd4 Transformer \u5c0f 90% \u7684\u7f13\u5b58\uff0c\u540c\u65f6\u8d28\u91cf\u76f8\u5f53\u751a\u81f3\u66f4\u597d\u3002<\/span><\/p>

\u5728 Hyena Edge \u7684\u8bbe\u8ba1\u4e2d\uff0cLiquid AI \u56e2\u961f\u5c31\u5e94\u7528\u4e86 STAR \u6846\u67b6\u3002\u4ed6\u4eec\u4ece\u4e00\u4e2a\u5305\u542b 16 \u4e2a\u5019\u9009\u67b6\u6784\u7684\u521d\u59cb\u79cd\u7fa4\u5f00\u59cb\uff0c\u8fdb\u884c\u4e86 24 \u4ee3\u7684\u8fdb\u5316\u8fed\u4ee3\u3002\u5176\u641c\u7d22\u7a7a\u95f4\u8bbe\u8ba1\u5f97\u975e\u5e38\u4e30\u5bcc\uff0c\u5305\u542b\u4e86\u591a\u79cd\u5377\u79ef\u7b97\u5b50\u7684\u53d8\u4f53\uff0c\u8fd9\u4e9b\u53d8\u4f53\u4e3b\u8981\u53d7\u5230 Hyena \u67b6\u6784\u7684\u542f\u53d1\uff1a<\/span><\/span><\/p>

  • Hyena\uff08Full\uff09\uff1a\u5728\u95e8\u63a7\u673a\u5236\uff08Gating Mechanism\uff09\u548c Hyena \u5185\u90e8\u5377\u79ef\u4e2d\u90fd\u5305\u542b\u5377\u79ef\u64cd\u4f5c\u3002<\/span><\/span><\/li><\/ul>
    • Hyena-X\uff1a\u6392\u9664\u4e86\u5185\u90e8\u5377\u79ef\u3002<\/span><\/span><\/li><\/ul>
      • Hyena-Y\uff1a\u6392\u9664\u4e86\u7279\u5f81\u7ec4\uff08\u95e8\uff09\u4e2d\u7684\u5377\u79ef\u3002<\/span><\/span><\/li><\/ul>

        \u9664\u4e86\u8fd9\u4e09\u79cd\u4e3b\u8981\u7684 Hyena \u7c7b\u578b\uff0c\u641c\u7d22\u7a7a\u95f4\u8fd8\u8003\u8651\u4e86\u5b83\u4eec\u5b66\u4e60\u5230\u7684\u77ed\u8ddd\u79bb\u3001\u663e\u5f0f\u5377\u79ef\u6838\u7684\u957f\u5ea6\u53d8\u5316\uff08\u4ece 3 \u5230 128\uff09\uff0c\u603b\u5171\u6784\u6210\u4e86 18 \u79cd\u4e0d\u540c\u7684\u5377\u79ef\u7b97\u5b50\u3002\u6b64\u5916\uff0c\u641c\u7d22\u7a7a\u95f4\u8fd8\u5305\u62ec\u4e86\u5206\u7ec4\u67e5\u8be2\u6ce8\u610f\u529b\u7684\u53d8\u4f53\uff08\u5177\u6709\u4e0d\u540c\u7684 KV \u5934\u6570\u91cf\uff09\u548c SwiGLU\uff08\u5177\u6709\u4e0d\u540c\u7684\u5185\u90e8\u5bbd\u5ea6\uff09\u7b49\u5e38\u89c1\u7684 Transformer \u7ec4\u4ef6\u3002<\/span><\/span><\/p>

        \u5728\u8fdb\u5316\u8fc7\u7a0b\u4e2d\uff0cSTAR \u6846\u67b6\u901a\u8fc7\u5bf9\u6bcf\u4e2a\u5019\u9009\u67b6\u6784\u5728\u4e09\u661f S24 Ultra \u4e0a\u7684\u521d\u59cb\u5ef6\u8fdf\u548c\u5185\u5b58\u5360\u7528\u8fdb\u884c\u6027\u80fd\u5206\u6790\uff0c\u5e76\u7ed3\u5408\u6a21\u578b\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u7684\u56f0\u60d1\u5ea6\u8868\u73b0\uff0c\u4e0d\u65ad\u5730\u5c06\u67b6\u6784\u79cd\u7fa4\u5411\u7740\u6548\u7387-\u8d28\u91cf\u524d\u6cbf\uff08Efficiency-Quality Frontier\uff09\u8fdb\u884c\u4f18\u5316\u3002<\/span><\/span><\/p>

        \u6709\u610f\u601d\u7684\u662f\uff0c<\/span>\u968f\u7740\u8fdb\u5316\u8fc7\u7a0b\u7684\u63a8\u8fdb\uff0c\u5f53\u67b6\u6784\u63a5\u8fd1\u6548\u7387-\u8d28\u91cf\u6700\u4f18\u8fb9\u754c\u65f6\uff0cSTAR \u660e\u663e\u5730\u504f\u7231\u9009\u7528 Hyena-Y \u7c7b\u578b\u7684\u5377\u79ef\u3002\u8fd9\u8868\u660e Hyena-Y \u5377\u79ef\u5728\u5ef6\u8fdf\u3001\u5185\u5b58\u548c\u6a21\u578b\u8d28\u91cf\u4e4b\u95f4\u53d6\u5f97\u4e86\u66f4\u4f18\u8d8a\u7684\u5e73\u8861\u3002<\/span><\/strong><\/span><\/p>

        \u57fa\u4e8e\u8fd9\u4e00\u53d1\u73b0\uff0c\u6700\u7ec8\u7684 Hyena Edge \u67b6\u6784\u5728\u4e00\u4e2a GQA-Transformer++ \u57fa\u51c6\u6a21\u578b\u7684\u57fa\u7840\u4e0a\uff0c\u5c06\u90e8\u5206 GQA \u7b97\u5b50\u66ff\u6362\u4e3a\u7531 STAR \u4f18\u5316\u9009\u51fa\u7684 Hyena-Y \u95e8\u63a7\u5377\u79ef\u3002<\/span><\/span><\/p>

        \u800c\u6700\u7ec8\u7ed3\u679c\u5982\u57fa\u51c6\u6d4b\u8bd5\u6240\u8868\u660e\u7684\uff0cHyena Edge \u5728\u63d0\u5347\u6548\u7387\u7684\u540c\u65f6\uff0c\u4fdd\u6301\u4e86\u8f83\u9ad8\u7684\u6a21\u578b\u8d28\u91cf\uff0c\u8fd9\u5bf9\u4e8e\u6027\u80fd\u548c\u8d44\u6e90\u90fd\u53d7\u9650\u7684\u8fb9\u7f18\u8bbe\u5907\u5e94\u7528\u800c\u8a00\u662f\u4e00\u4e2a\u91cd\u8981\u7684\u7279\u6027\u3002<\/span><\/span><\/p>

        Liquid AI \u5df2\u8868\u793a\u8ba1\u5212\u5728\u672a\u6765\u6570\u6708\u5185\u5f00\u6e90\u5305\u62ec Hyena Edge \u5728\u5185\u7684\u4e00\u7cfb\u5217\u57fa\u7840\u6a21\u578b\uff0c\u5176\u76ee\u6807\u662f\u6784\u5efa\u80fd\u591f\u9002\u5e94\u4ece\u4e91\u7aef\u5230\u8fb9\u7f18\u591a\u79cd\u73af\u5883\u7684 AI \u7cfb\u7edf\u3002<\/span>\u9664\u4e86\u6a21\u578b\u672c\u8eab\uff0c\u5176\u6240\u5c55\u73b0\u51fa\u7684\u8bbe\u8ba1\u65b9\u6cd5\uff0c\u6216\u8bb8\u66f4\u503c\u5f97\u6211\u4eec\u7684\u671f\u5f85\u3002<\/span><\/strong><\/span><\/p>

        \u53c2\u8003\u8d44\u6599\uff1a<\/span><\/span><\/p>

        1.https:\/\/www.liquid.ai\/research\/convolutional-multi-hybrids-for-edge-devices<\/span><\/span><\/p>

        2.https:\/\/arxiv.org\/abs\/2411.17800<\/span><\/span><\/p>

        3.https:\/\/venturebeat.com\/ai\/liquid-ai-is-revolutionizing-llms-to-work-on-edge-devices-like-smartphones-with-new-hyena-edge-model\/<\/span><\/span><\/p>

        <\/p>

        \u8fd0\u8425\/\u6392\u7248\uff1a\u4f55\u6668\u9f99<\/span><\/span><\/p><\/div><\/div>","content_en":null,"content_all":null,"article_url":"http:\/\/www.sciphi.cn\/article\/view\/14725","typeName":"AI"},"code":10000,"message":"\u6210\u529f"}