{"id":749110,"date":"2026-03-18T21:08:22","date_gmt":"2026-03-18T13:08:22","guid":{"rendered":"https:\/\/www.alphaseek.cn\/?p=749110"},"modified":"2026-03-18T21:08:22","modified_gmt":"2026-03-18T13:08:22","slug":"%e9%a9%ac%e6%96%af%e5%85%8b%e4%b8%8b%e5%9c%ba%e7%82%b9%e8%b5%9e%ef%bc%81kimi-%e8%bf%99%e7%af%87%e8%ae%ba%e6%96%87%e6%92%ac%e5%8a%a8%e4%ba%86%e5%a4%a7%e6%a8%a1%e5%9e%8b%e7%9a%84%e7%a5%96%e4%bc%a0","status":"publish","type":"post","link":"https:\/\/www.alphaseek.cn\/?p=749110","title":{"rendered":"\u9a6c\u65af\u514b\u4e0b\u573a\u70b9\u8d5e\uff01Kimi \u8fd9\u7bc7\u8bba\u6587\u64ac\u52a8\u4e86\u5927\u6a21\u578b\u7684\u7956\u4f20\u5730\u57fa"},"content":{"rendered":"

\"\u9a6c\u65af\u514b\u4e0b\u573a\u70b9\u8d5e\uff01Kimi<\/p>\n

\u540c\u6837\u7684\u7b97\u529b\uff0c\u540c\u6837\u7684\u6570\u636e\uff0c\u51ed\u4ec0\u4e48\u6548\u679c\u4e0d\u4e00\u6837\uff1f\u5927\u591a\u6570\u4eba\u7684\u76f4\u89c9\u662f\uff1a\u6a21\u578b\u66f4\u5927\u3001\u6570\u636e\u66f4\u597d\u3001\u5de5\u7a0b\u5e08\u66f4\u5389\u5bb3\u3002\u4f46 Kimi \u7ed9\u51fa\u4e86\u4e00\u4e2a\u66f4\u51fa\u4eba\u610f\u6599\u7684\u7b54\u6848\u3002<\/p>\n

3 \u6708 16 \u65e5\uff0c\u6708\u4e4b\u6697\u9762 Kimi \u53d1\u5e03\u4e86\u4e00\u9879\u91cd\u78c5\u6280\u672f\u62a5\u544a\u300aAttention Residuals\u300b\uff08\u6ce8\u610f\u529b\u6b8b\u5dee\uff09\u3002<\/p>\n

\u8fd9\u9879\u6280\u672f\u9488\u5bf9\u51e0\u4e4e\u6240\u6709\u73b0\u4ee3\u5927\u6a21\u578b\u90fd\u5728\u4f7f\u7528\u7684\u6b8b\u5dee\u8fde\u63a5\u7ed3\u6784\u8fdb\u884c\u4e86\u6539\u9020\uff0c\u5e76\u5728\u5b9e\u9a8c\u4e2d\u8bc1\u660e\uff0c\u7528\u540c\u6837\u591a\u7684\u7b97\u529b\uff0c\u65b0\u65b9\u6cd5\u8bad\u7ec3\u51fa\u7684\u6a21\u578b\u6548\u679c\u76f8\u5f53\u4e8e\u57fa\u7ebf\u6a21\u578b\u82b1\u8d39 1.25 \u500d\u7b97\u529b\u624d\u80fd\u8fbe\u5230\u7684\u6548\u679c\u3002<\/p>\n

\u62a5\u544a\u53d1\u5e03\u540e\uff0c\u4e5f\u6beb\u65e0\u610f\u5916\u5f97\u5230\u4e86\u8bb8\u591a\u7845\u8c37<\/a>\u9876\u5c16 AI \u4eba\u7269\u7684\u70b9\u8d5e\u80cc\u4e66\u3002<\/p>\n

\"\u9a6c\u65af\u514b\u4e0b\u573a\u70b9\u8d5e\uff01Kimi<\/p>\n

\u25b2\u9644 GitHub \u5f00\u6e90\u5730\u5740\uff1agithub.com\/MoonshotAI\/Attention-Residuals<\/p>\n

\u6bd4\u5982\u9a6c\u65af\u514b<\/a>\u901a\u8fc7\u793e\u4ea4\u5a92\u4f53<\/a>\u8868\u793a\u300c\u300cImpressive work from Kimi\u300d\uff08\u4ee4\u4eba\u5370\u8c61\u6df1\u523b\u7684\u5de5\u4f5c\uff09\u300dOpenAI<\/a> o1 \u4e3b\u8981\u53d1\u660e\u8005 Jerry Tworek \u79f0\u5176\u4e3a\u300c\u6df1\u5ea6\u5b66\u4e60 2.0\u300d\u7684\u5f00\u7aef\u3002<\/p>\n

\"\u9a6c\u65af\u514b\u4e0b\u573a\u70b9\u8d5e\uff01Kimi<\/p>\n

\u524d OpenAI \u8054\u521b Andrej Karpathy \u8bf4\u300c\u770b\u6765\u6211\u4eec\u8fd8\u6ca1\u628a\u300eAttention is All You Need\u300f\u8fd9\u53e5\u8bdd\u6309\u5b57\u9762\u610f\u601d\u7406\u89e3\u900f\u3002\u300d\u4f46\u6bd4\u8d77\u8fd9\u4e9b\u5938\u5956\uff0c\u6280\u672f\u8bba\u6587\u80cc\u540e\u7684\u4fe1\u53f7\u6216\u8bb8\u66f4\u503c\u5f97\u5173\u6ce8\uff1a\u6df1\u5ea6\u5b66\u4e60\u6700\u57fa\u7840\u7684\u8303\u5f0f\uff0c\u6b63\u5728\u53d1\u751f\u53d8\u5316\u3002<\/p>\n

\"\u9a6c\u65af\u514b\u4e0b\u573a\u70b9\u8d5e\uff01Kimi<\/p>\n

\u5341\u5e74\u6ca1\u4eba\u52a8\u8fc7\u7684\u5730\u57fa\uff0c\u88ab\u64ac\u52a8\u4e86<\/p>\n

\u8fc7\u53bb\u4e24\u5e74\uff0c\u5927\u6a21\u578b\u7684\u7ade\u4e89\u4e3b\u8981\u5728\u300c\u4e0a\u5c42\u5efa\u7b51\u300d\u5c55\u5f00\uff1a\u66f4\u597d\u7684\u6ce8\u610f\u529b\u53d8\u4f53\u3001\u66f4\u806a\u660e\u7684 MoE \u8def\u7531\u7b56\u7565<\/a>\u3001\u66f4\u7cbe\u5de7\u7684\u5bf9\u9f50\u65b9\u6cd5\uff0c\u5927\u5bb6\u90fd\u5728 Transformer \u8fd9\u680b\u5927\u697c\u7684\u9ad8\u5c42\u7cbe\u88c5\u4fee\u3002<\/p>\n

\u552f\u72ec\u6709\u4e00\u6837\u4e1c\u897f\uff0c\u4ece 2015 \u5e74 ResNet \u8bba\u6587\u53d1\u8868\u4ee5\u6765\uff0c\u51e0\u4e4e\u6ca1\u4eba\u52a8\u8fc7\uff1a\u6b8b\u5dee\u8fde\u63a5\uff08Residual Connections\uff09\u3002<\/p>\n

\u8981\u7406\u89e3\u8fd9\u9879\u6280\u672f\uff0c\u5f97\u5148\u77e5\u9053\u5927\u6a21\u578b\u5185\u90e8\u7684\u57fa\u672c\u7ed3\u6784\u3002<\/p>\n

\u73b0\u4ee3\u5927\u6a21\u578b\uff0c\u5176\u5b9e\u90fd\u662f\u7531\u5f88\u591a\u5c42\u795e\u7ecf\u7f51\u7edc\u53e0\u52a0\u800c\u6210\u7684\uff0c\u5c11\u5219\u51e0\u5341\u5c42\uff0c\u591a\u5219\u4e0a\u767e\u5c42\u3002\u4fe1\u606f\u4ece\u5e95\u90e8\u8f93\u5165\uff0c\u4e00\u5c42\u4e00\u5c42\u5f80\u4e0a\u4f20\u9012\uff0c\u6bcf\u4e00\u5c42\u90fd\u5bf9\u4fe1\u606f\u505a\u4e00\u6b21\u52a0\u5de5\uff0c\u6700\u7ec8\u5728\u9876\u90e8\u8f93\u51fa\u7ed3\u679c\u3002<\/p>\n

\u53ef\u4ee5\u628a\u5b83\u60f3\u8c61\u6210\u4e00\u6761\u6d41\u6c34\u7ebf\u4e0a\u7684\u5de5\u4eba\uff1a\u539f\u6750\u6599\u4ece\u7b2c\u4e00\u9053\u5de5\u5e8f\u8fdb\u6765\uff0c\u6bcf\u4e2a\u5de5\u4eba\u5bf9\u5b83\u52a0\u5de5\u4e00\u904d\uff0c\u518d\u4f20\u7ed9\u4e0b\u4e00\u4e2a\uff0c\u6700\u7ec8\u51fa\u6765\u6210\u54c1\u3002\u95ee\u9898\u662f\uff0c\u6d41\u6c34\u7ebf\u8d8a\u957f\uff0c\u8d8a\u96be\u8bad\u7ec3\u3002<\/p>\n

\u5047\u8bbe\u7b2c 50 \u9053\u5de5\u5e8f\u7684\u5de5\u4eba\u72af\u4e86\u9519\uff0c\u4f60\u60f3\u7ea0\u6b63\u4ed6\uff0c\u5c31\u5f97\u628a\u8fd9\u4e2a\u300c\u7ea0\u9519\u4fe1\u53f7\u300d\u4e00\u8def\u5f80\u56de\u4f20\uff0c\u7ecf\u8fc7 49 \u4e2a\u5de5\u4eba\u624d\u80fd\u4f20\u5230\u7b2c 1 \u4e2a\u3002\u4f20\u7740\u4f20\u7740\uff0c\u4fe1\u53f7\u5c31\u6d88\u5931\u4e86\uff0c\u5e95\u5c42\u7684\u5de5\u4eba\u6839\u672c\u4e0d\u77e5\u9053\u81ea\u5df1\u54ea\u91cc\u51fa\u4e86\u95ee\u9898\u3002<\/p>\n

\u4e3a\u4e86\u8ba9\u8fd9\u4e48\u6df1\u7684\u7f51\u7edc\u80fd\u591f\u8bad\u7ec3\u8d77\u6765\uff0c\u77e5\u540d\u5b66\u8005\u4f55\u607a\u660e\u56e2\u961f\u5728 2015 \u5e74\u53d1\u8868\u4e86\u4e00\u7bc7\u9898\u4e3a\u300aDeep Residual Learning for Image Recognition\u300b\u7684\u8bba\u6587\uff0c\u5f15\u5165\u4e86\u4e00\u4e2a\u5173\u952e\u8bbe\u8ba1\uff0c\u53eb\u505a\u6b8b\u5dee\u8fde\u63a5\uff08Residual Connections\uff09\uff1a<\/p>\n

\u6bcf\u4e00\u5c42\u5728\u52a0\u5de5\u4fe1\u606f\u7684\u540c\u65f6\uff0c\u8fd8\u4f1a\u4fdd\u7559\u4e00\u6761\u300c\u76f4\u901a\u9053\u300d\uff0c\u628a\u539f\u59cb\u8f93\u5165\u539f\u5c01\u4e0d\u52a8\u5730\u52a0\u5230\u52a0\u5de5\u7ed3\u679c\u4e0a\uff0c\u518d\u5f80\u4e0b\u4f20\u3002\u8fd9\u6761\u76f4\u901a\u9053\u8ba9\u68af\u5ea6\u5728\u53cd\u5411\u4f20\u64ad\u65f6\u53ef\u4ee5\u7ed5\u8fc7\u4e2d\u95f4\u7684\u53d8\u6362\uff0c\u4e00\u8def\u6d41\u56de\u5e95\u5c42\uff0c\u4ece\u6839\u672c\u4e0a\u89e3\u51b3\u4e86\u6df1\u5c42\u7f51\u7edc\u96be\u4ee5\u8bad\u7ec3\u7684\u95ee\u9898\u3002<\/p>\n

\"\u9a6c\u65af\u514b\u4e0b\u573a\u70b9\u8d5e\uff01Kimi<\/p>\n

\u6bd4\u8f83\u901a\u4fd7\u7684\u7406\u89e3\u662f\uff0c\u5728\u6bcf\u9053\u5de5\u5e8f\u65c1\u8fb9\u52a0\u4e00\u6761\u300c\u76f4\u901a\u9053\u300d\uff0c\u628a\u539f\u6750\u6599\u539f\u5c01\u4e0d\u52a8\u5730\u7ed5\u8fc7\u8fd9\u9053\u5de5\u5e8f\uff0c\u76f4\u63a5\u548c\u52a0\u5de5\u7ed3\u679c\u5408\u5e76\uff0c\u518d\u5f80\u4e0b\u4f20\u3002\u8fd9\u6837\u7ea0\u9519\u4fe1\u53f7\u5c31\u53ef\u4ee5\u6cbf\u7740\u76f4\u901a\u9053\u4e00\u8def\u7545\u901a\u65e0\u963b\u5730\u4f20\u56de\u5e95\u5c42\uff0c\u4e0d\u4f1a\u6d88\u5931\u3002<\/p>\n

\u8fd9\u7bc7\u8bba\u6587\u540e\u6765\u6210\u4e3a\u8ba1\u7b97\u673a\u89c6\u89c9\u4e43\u81f3\u6574\u4e2a\u6df1\u5ea6\u5b66\u4e60\u9886\u57df\u5f15\u7528\u6b21\u6570\u6700\u591a\u7684\u8bba\u6587\u4e4b\u4e00\uff0c\u6b8b\u5dee\u8fde\u63a5\u4e5f\u6cbf\u7528\u81f3\u4eca\uff0c\u662f\u51e0\u4e4e\u6240\u6709\u5927\u6a21\u578b\u7684\u57fa\u77f3\u3002<\/p>\n

\u6b8b\u5dee\u8fde\u63a5\u867d\u7136\u597d\u7528\uff0c\u4f46\u5b83\u505a\u4fe1\u606f\u805a\u5408\u7684\u65b9\u5f0f\u975e\u5e38\u7c97\u66b4\uff1a\u628a\u6240\u6709\u524d\u9762\u5c42\u7684\u8f93\u51fa\uff0c\u65e0\u5dee\u522b\u5730\u7b49\u6743\u76f8\u52a0\u3002<\/p>\n

\u8fd8\u662f\u7528\u6d41\u6c34\u7ebf\u6765\u6bd4\u55bb\u3002\u5230\u4e86\u7b2c 51 \u9053\u5de5\u5e8f\uff0c\u8fd9\u4e2a\u5de5\u4eba\u624b\u91cc\u62ff\u5230\u7684\uff0c\u662f\u524d\u9762 50 \u9053\u5de5\u5e8f\u6240\u6709\u4ea7\u51fa\u7269\u7684\u7b49\u91cf\u6df7\u5408\uff0c\u6bcf\u9053\u5de5\u5e8f\u7684\u4ea7\u51fa\u5404\u5360\u4e00\u4efd\uff0c\u4e0d\u591a\u4e0d\u5c11\u3002\u4ed6\u6ca1\u6709\u529e\u6cd5\u8bf4\u300c\u6211\u60f3\u591a\u8981\u4e00\u70b9\u7b2c 3 \u9053\u5de5\u5e8f\u7684\u539f\u6599\u300d\uff0c\u4e5f\u6ca1\u6709\u529e\u6cd5\u8bf4\u300c\u7b2c 20 \u9053\u5de5\u5e8f\u7684\u4e1c\u897f\u5bf9\u6211\u6ca1\u7528\uff0c\u5c11\u7ed9\u6211\u4e00\u70b9\u300d\u3002<\/p>\n

\u8fd9\u5e26\u6765\u4e86\u4e00\u4e2a\u540d\u4e3a PreNorm \u7a00\u91ca\u7684\u5b9e\u9645\u95ee\u9898 \uff1a\u968f\u7740\u7f51\u7edc\u8d8a\u6765\u8d8a\u6df1\uff0c\u7d2f\u79ef\u53e0\u52a0\u7684\u4fe1\u606f\u8d8a\u6765\u8d8a\u591a\uff0c\u6bcf\u4e00\u5c42\u81ea\u5df1\u7684\u8d21\u732e\u5728\u5e9e\u5927\u7684\u603b\u91cf\u91cc\u8d8a\u6765\u8d8a\u5fae\u4e0d\u8db3\u9053\u3002\u8d8a\u9760\u540e\u7684\u5c42\uff0c\u60f3\u8981\u8ba9\u81ea\u5df1\u7684\u58f0\u97f3\u88ab\u300c\u542c\u89c1\u300d\uff0c\u5c31\u5f97\u8f93\u51fa\u8d8a\u6765\u8d8a\u5927\u7684\u6570\u503c\uff0c\u5426\u5219\u5c31\u4f1a\u88ab\u6df9\u6ca1\u3002<\/p>\n

\u7ed3\u679c\u5c31\u662f\uff0c\u5f88\u591a\u4e2d\u95f4\u5c42\u5176\u5b9e\u6ca1\u5728\u8ba4\u771f\u5e72\u6d3b\u3002\u5df2\u6709\u7814\u7a76\u53d1\u73b0\uff0c\u5927\u6a21\u578b\u91cc\u76f8\u5f53\u4e00\u90e8\u5206\u5c42\u76f4\u63a5\u5220\u6389\uff0c\u6548\u679c\u51e0\u4e4e\u4e0d\u53d8\uff0c\u8fd9\u8bf4\u660e\u8fd9\u4e9b\u5c42\u7684\u8d21\u732e\u5b9e\u9645\u4e0a\u6781\u4e3a\u6709\u9650\u3002<\/p>\n

\u5927\u591a\u6570\u56e2\u961f\u65e9\u5c31\u77e5\u9053\u8fd9\u4e2a\u95ee\u9898\uff0c\u9009\u62e9\u7ed5\u5f00\u5b83\uff0c\u8f6c\u800c\u5728\u5728\u73b0\u6709\u67b6\u6784\u4e0a\u53e0\u52a0\u66f4\u597d\u7684\u6570\u636e\u914d\u6bd4\u3001\u66f4\u7cbe\u5de7\u7684\u8bad\u7ec3\u7b56\u7565\u3001\u66f4\u957f\u7684\u4e0a\u4e0b\u6587\u7a97\u53e3\u3002\u8fd9\u4e9b\u5de5\u4f5c\u5f53\u7136\u6709\u4ef7\u503c\uff0c\u4f46\u672c\u8d28\u4e0a\u662f\u5728\u4e00\u4e2a\u5df2\u6709\u7684\u6280\u672f\u6846\u67b6\u5185\u505a\u589e\u91cf\u4f18\u5316\u3002<\/p>\n

Kimi \u9009\u62e9\u7684\u662f\u4e00\u6761\u66f4\u5b64\u72ec\u4e5f\u66f4\u96be\u7684\u8def\uff1a\u56de\u5230\u6700\u57fa\u7840\u7684\u7ed3\u6784\uff0c\u7528\u7b2c\u4e00\u6027\u539f\u7406\u91cd\u65b0\u5ba1\u89c6\u90a3\u4e9b\u300c\u7406\u6240\u5f53\u7136\u300d\u7684\u8bbe\u8ba1\u3002<\/p>\n

\u4eca\u5929\u51cc\u6668\uff0cKimi \u521b\u59cb\u4eba\u6768\u690d\u9e9f\u5728 GTC 2026 \u6f14\u8bb2\u4e2d\u63d0\u5230\uff1a\u300c\u884c\u4e1a\u76ee\u524d\u666e\u904d\u4f7f\u7528\u7684\u5f88\u591a\u6280\u672f\u6807\u51c6\uff0c\u672c\u8d28\u4e0a\u662f\u516b\u4e5d\u5e74\u524d\u7684\u4ea7\u7269\uff0c\u6b63\u9010\u6e10\u6210\u4e3a Scaling \u7684\u74f6\u9888\u3002\u300d<\/p>\n

\u6768\u690d\u9e9f\u8ba4\u4e3a\uff0c\u8981\u63a8\u52a8\u5927\u6a21\u578b\u667a\u80fd\u4e0a\u9650\u7684\u6301\u7eed\u7a81\u7834\uff0c\u5fc5\u987b\u5bf9\u4f18\u5316\u5668\u3001\u6ce8\u610f\u529b\u673a\u5236\u53ca\u6b8b\u5dee\u8fde\u63a5\u7b49\u5e95\u5c42\u57fa\u77f3\u8fdb\u884c\u91cd\u6784\u3002<\/p>\n

\"\u9a6c\u65af\u514b\u4e0b\u573a\u70b9\u8d5e\uff01Kimi<\/p>\n

\u4e00\u6b21\u4f18\u96c5\u7684\u300c\u65cb\u8f6c\u300d<\/p>\n

Kimi \u56e2\u961f\u8fd9\u7bc7\u8bba\u6587\u7684\u6838\u5fc3\u7a81\u7834\uff0c\u5176\u5b9e\u4e5f\u6765\u81ea\u4e00\u4e2a\u4f18\u96c5\u7684\u7c7b\u6bd4\u53d1\u73b0\u3002<\/p>\n

\u5904\u7406\u6587\u5b57\u5e8f\u5217\u65f6\uff0c\u65e9\u671f\u7684\u5faa\u73af\u795e\u7ecf\u7f51\u7edc\uff08RNN\uff09\u4e5f\u6709\u7c7b\u4f3c\u7684\u989d\u5916\u95ee\u9898\uff1a\u8bb0\u6027\u5dee\u3002\u5b83\u8bfb\u5b8c\u4e00\u6574\u6bb5\u8bdd\u4e4b\u540e\uff0c\u65e9\u671f\u8bfb\u5230\u7684\u5185\u5bb9\u4f1a\u88ab\u540e\u6765\u7684\u5185\u5bb9\u4e0d\u65ad\u8986\u76d6\uff0c\u7b49\u8bfb\u5230\u6700\u540e\u4e00\u4e2a\u8bcd\uff0c\u524d\u51e0\u53e5\u8bf4\u4e86\u4ec0\u4e48\u5df2\u7ecf\u6a21\u7cca\u4e86\u3002<\/p>\n

\u540e\u6765 Transformer \u7528\u6ce8\u610f\u529b\u673a\u5236\u89e3\u51b3\u4e86\u8fd9\u4e2a\u95ee\u9898\uff0c\u76f8\u5f53\u4e8e\u7ed9\u6a21\u578b\u914d\u4e86\u4e00\u5f20\u300c\u5168\u6587\u7b14\u8bb0\u300d\uff0c\u5904\u7406\u6bcf\u4e2a\u8bcd\u7684\u65f6\u5019\uff0c\u90fd\u53ef\u4ee5\u7ffb\u56de\u53bb\u67e5\u4efb\u610f\u4e00\u4e2a\u4e4b\u524d\u51fa\u73b0\u8fc7\u7684\u8bcd\uff0c\u800c\u4e14\u67e5\u54ea\u91cc\u3001\u67e5\u591a\u5c11\uff0c\u7531\u5f53\u524d\u7684\u5185\u5bb9\u81ea\u5df1\u51b3\u5b9a\u3002<\/p>\n

\u7814\u7a76\u4eba\u5458\u53d1\u73b0\uff0c\u6b8b\u5dee\u8fde\u63a5\u5728\u6df1\u5ea6\u65b9\u5411\u4e0a\u78b0\u5230\u7684\u95ee\u9898\uff0c\u548c RNN \u5728\u65f6\u95f4\u65b9\u5411\u4e0a\u78b0\u5230\u7684\u95ee\u9898\uff0c\u6570\u5b66\u7ed3\u6784\u5b8c\u5168\u4e00\u6837\u3002\u6362\u53e5\u8bdd\u8bf4\uff0c\u628a Transformer \u60f3\u8c61\u6210\u4e00\u5f20\u4e8c\u7ef4\u7684\u7f51\u683c\uff1a<\/p>\n

\u6a2a\u8f74\u662f\u5e8f\u5217\u65b9\u5411\uff0c\u4e00\u53e5\u8bdd\u91cc\u4ece\u5de6\u5230\u53f3\u7684\u6bcf\u4e2a\u8bcd\uff1b\u7eb5\u8f74\u662f\u6df1\u5ea6\u65b9\u5411\uff0c\u4ece\u5e95\u5c42\u5230\u9876\u5c42\u7684\u6bcf\u4e00\u5c42\u7f51\u7edc\u3002\u4f20\u7edf\u7684\u6ce8\u610f\u529b\u673a\u5236\u662f\u6cbf\u7740\u6a2a\u8f74\u5de5\u4f5c\u7684\uff0c\u5904\u7406\u67d0\u4e2a\u8bcd\u65f6\u53bb\u67e5\u540c\u4e00\u5c42\u91cc\u5176\u4ed6\u8bcd\u7684\u4fe1\u606f\u3002<\/p>\n

\u800c Attention Residuals \u505a\u7684\u4e8b\u60c5\uff0c\u5c31\u662f\u628a\u5b8c\u5168\u76f8\u540c\u7684\u673a\u5236\u8f6c\u5230\u7eb5\u8f74\u4e0a\u53bb\uff0c\u5904\u7406\u67d0\u4e00\u5c42\u65f6\u53bb\u67e5\u524d\u9762\u6240\u6709\u5c42\u7684\u8f93\u51fa\uff0c\u51b3\u5b9a\u8981\u53c2\u8003\u54ea\u4e9b\u5c42\u3001\u53c2\u8003\u591a\u5c11\u3002\u64cd\u4f5c\u5bf9\u8c61\u4ece\u300c\u540c\u4e00\u5c42\u91cc\u7684\u4e0d\u540c\u8bcd\u300d\u53d8\u6210\u4e86\u300c\u540c\u4e00\u4e2a\u8bcd\u5728\u4e0d\u540c\u5c42\u91cc\u7684\u72b6\u6001\u300d\uff0c\u673a\u5236\u672c\u8eab\u4e00\u6a21\u4e00\u6837\uff0c\u597d\u6bd4\u65b9\u5411\u8f6c\u4e86 90 \u5ea6\u3002<\/p>\n

\"\u9a6c\u65af\u514b\u4e0b\u573a\u70b9\u8d5e\uff01Kimi<\/p>\n

\u65e2\u7136\u6ce8\u610f\u529b\u673a\u5236\u89e3\u51b3\u4e86\u5e8f\u5217\u65b9\u5411\u7684\u95ee\u9898\uff0c\u65cb\u8f6c\u4e00\u4e0b\u642c\u5230\u6df1\u5ea6\u65b9\u5411\u4e0a\uff0c\u540c\u6837\u6709\u6548\u3002<\/p>\n

\u8fd9\u91cc\u6709\u4e00\u4e2a\u66f4\u6df1\u5c42\u7684\u7406\u8bba\u53d1\u73b0\u503c\u5f97\u4e00\u63d0\u3002\u7814\u7a76\u4eba\u5458\u901a\u8fc7\u6570\u5b66\u5206\u6790\u53d1\u73b0\uff0c\u8fc7\u53bb\u5341\u5e74\u91cc\u6240\u6709\u5bf9\u6b8b\u5dee\u8fde\u63a5\u7684\u6539\u8fdb\uff0c\u5305\u62ec\u6807\u51c6\u6b8b\u5dee\u3001Highway \u7f51\u7edc\u3001mHC \u7b49\u5404\u79cd\u53d8\u4f53\uff0c\u5728\u6570\u5b66\u4e0a\u5176\u5b9e\u90fd\u662f\u540c\u4e00\u4ef6\u4e8b\u7684\u4e0d\u540c\u5f62\u5f0f\uff0c\u90fd\u7b49\u4ef7\u4e8e\u67d0\u79cd\u300c\u6df1\u5ea6\u65b9\u5411\u7684\u7ebf\u6027\u6ce8\u610f\u529b\u300d\u3002\u6362\u53e5\u8bdd\u8bf4\uff0c\u5927\u5bb6\u4e00\u76f4\u5728\u671d\u540c\u4e00\u4e2a\u65b9\u5411\u52aa\u529b\uff0c\u53ea\u662f\u5f53\u65f6\u6ca1\u610f\u8bc6\u5230\u3002<\/p>\n

\u800c AttnRes \u7684\u6838\u5fc3\u601d\u8def\u5728\u4e8e\uff0c\u628a\u6ce8\u610f\u529b\u673a\u5236\u4ece\u300c\u5904\u7406\u6587\u5b57\u5e8f\u5217\u300d\u7684\u7ef4\u5ea6\uff0c\u79fb\u690d\u5230\u300c\u8de8\u8d8a\u7f51\u7edc\u6df1\u5ea6\u300d\u7684\u7ef4\u5ea6\u4e0a\u3002<\/p>\n

\u5177\u4f53\u505a\u6cd5\u662f\uff0c\u7ed9\u6bcf\u4e00\u5c42\u914d\u5907\u4e00\u4e2a\u5c0f\u5c0f\u7684\u300c\u67e5\u8be2\u5411\u91cf\u300d\uff0c\u5c31\u50cf\u7ed9\u6bcf\u9053\u5de5\u5e8f\u7684\u5de5\u4eba\u914d\u4e86\u4e00\u5f20\u9700\u6c42\u5355\u3002\u5de5\u4eba\u5728\u5f00\u5de5\u524d\uff0c\u5148\u62ff\u7740\u9700\u6c42\u5355\u53bb\u7ffb\u6240\u6709\u524d\u9762\u5de5\u5e8f\u7684\u4ea7\u51fa\uff0c\u6839\u636e\u76f8\u5173\u5ea6\u7b97\u51fa\u4e00\u5957\u53d6\u7528\u6bd4\u4f8b\uff0c\u518d\u6309\u8fd9\u4e2a\u6bd4\u4f8b\u628a\u9700\u8981\u7684\u539f\u6599\u6df7\u5408\u8d77\u6765\u3002<\/p>\n

\u8fd9\u6837\u4e00\u6765\uff0c\u6bcf\u4e00\u5c42\u4e0d\u518d\u662f\u88ab\u52a8\u63a5\u53d7\u6240\u6709\u524d\u9762\u5c42\u8f93\u51fa\u7684\u7b49\u6743\u53e0\u52a0\uff0c\u800c\u662f\u4e3b\u52a8\u3001\u6709\u9009\u62e9\u6027\u5730\u51b3\u5b9a\u8981\u4ece\u54ea\u4e9b\u5c42\u63d0\u53d6\u591a\u5c11\u4fe1\u606f\uff0c\u6bd4\u4f8b\u8fd8\u4f1a\u6839\u636e\u5f53\u524d\u4efb\u52a1\u7684\u5185\u5bb9\u52a8\u6001\u53d8\u5316\u3002\u6bcf\u5c42\u53ea\u65b0\u589e\u4e00\u4e2a\u5411\u91cf\u548c\u4e00\u4e2a\u5f52\u4e00\u5316\u64cd\u4f5c\uff0c\u53c2\u6570\u91cf\u7684\u589e\u52a0\u5bf9\u6574\u4e2a\u6a21\u578b\u6765\u8bf4\u51e0\u4e4e\u53ef\u4ee5\u5ffd\u7565\u4e0d\u8ba1\u3002<\/p>\n

\u4e3a\u4e86\u4fdd\u8bc1\u8bad\u7ec3\u521d\u671f\u7a33\u5b9a\uff0c\u8fd9\u4e2a\u67e5\u8be2\u5411\u91cf\u5fc5\u987b\u521d\u59cb\u5316\u4e3a\u5168\u96f6\uff0c\u76f8\u5f53\u4e8e\u8ba9\u5de5\u4eba\u4e00\u5f00\u59cb\u4ec0\u4e48\u504f\u597d\u90fd\u6ca1\u6709\u3001\u5e73\u7b49\u5bf9\u5f85\u6240\u6709\u524d\u5e8f\u4ea7\u51fa\uff0c\u7b49\u8bad\u7ec3\u63a8\u8fdb\u4e86\u518d\u6162\u6162\u5f62\u6210\u81ea\u5df1\u7684\u5224\u65ad\u3002<\/p>\n

\u503c\u5f97\u4e00\u63d0\u7684\u662f\uff0c\u7814\u7a76\u4eba\u5458\u4e5f\u6d4b\u8bd5\u8fc7\u4e00\u4e2a\u66f4\u6fc0\u8fdb\u7684\u7248\u672c\uff1a\u8ba9\u67e5\u8be2\u5411\u91cf\u4e0d\u518d\u662f\u56fa\u5b9a\u53c2\u6570\uff0c\u800c\u662f\u6839\u636e\u6bcf\u4e00\u5c42\u5f53\u524d\u7684\u8f93\u5165\u5185\u5bb9\u52a8\u6001\u751f\u6210\u3002\u8fd9\u4e2a\u7248\u672c\u6548\u679c\u786e\u5b9e\u66f4\u597d\uff0c\u635f\u5931\u503c\u8fdb\u4e00\u6b65\u4e0b\u964d\u3002<\/p>\n

\u4f46\u6700\u7ec8\u6ca1\u6709\u91c7\u7528\uff0c\u539f\u56e0\u662f\u63a8\u7406\u65f6\u8fd9\u79cd\u65b9\u5f0f\u9700\u8981\u987a\u5e8f\u8bfb\u53d6\u5185\u5b58\uff0c\u4f1a\u589e\u52a0\u5ef6\u8fdf\u3002\u8fd9\u4e2a\u53d6\u820d\u4f53\u73b0\u4e86\u8d2f\u7a7f\u6574\u7bc7\u8bba\u6587\u7684\u5de5\u7a0b\u54f2\u5b66\uff0c\u7406\u8bba\u4e0a\u66f4\u4f18\u7684\u65b9\u6848\uff0c\u4e0d\u4e00\u5b9a\u662f\u5b9e\u7528\u4e0a\u5e94\u8be5\u9009\u7684\u65b9\u6848\u3002<\/p>\n

\u5927\u6a21\u578b\u7684\u65b0\u6280\u672f\uff0c\u6700\u540e\u90fd\u5f97\u8fc7\u8fd9\u4e00\u5173<\/p>\n

\u5168\u91cf AttnRes \u5728\u5c0f\u89c4\u6a21\u5b9e\u9a8c\u4e2d\u5f88\u597d\u7528\uff0c\u4f46\u4e00\u5230\u5927\u89c4\u6a21\u8bad\u7ec3\u5c31\u9047\u5230\u4e86\u9ebb\u70e6\u3002<\/p>\n

\u5b83\u9700\u8981\u6bcf\u4e00\u5c42\u90fd\u80fd\u8bbf\u95ee\u6240\u6709\u524d\u9762\u5c42\u7684\u8f93\u51fa\u3002\u6a21\u578b\u6709\u4e00\u767e\u591a\u5c42\uff0c\u6bcf\u5c42\u7684\u8f93\u51fa\u90fd\u5f97\u4fdd\u5b58\u5728\u5185\u5b58\u91cc\uff0c\u8fd8\u8981\u5728\u4e0d\u540c\u8ba1\u7b97\u8282\u70b9\u4e4b\u95f4\u6765\u56de\u4f20\u8f93\uff0c\u5185\u5b58\u548c\u901a\u4fe1\u5f00\u9500\u968f\u5c42\u6570\u7ebf\u6027\u589e\u957f\uff0c\u5728\u5927\u6a21\u578b\u4e0a\u6839\u672c\u627f\u53d7\u4e0d\u8d77\u3002<\/p>\n

Kimi \u56e2\u961f\u7684\u89e3\u6cd5\u5f88\u5b9e\u5728\uff1aBlock AttnRes\u3002\u628a\u7f51\u7edc\u6240\u6709\u5c42\u5212\u5206\u4e3a\u82e5\u5e72\u4e2a Block\uff0848B \u6a21\u578b\u4e2d\u5206\u4e86 8-9 \u4e2a Block\uff0c\u6bcf\u4e2a Block \u7ea6 6 \u5c42\uff09\uff0cBlock \u5185\u90e8\u6cbf\u7528\u4f20\u7edf\u6b8b\u5dee\u8fde\u63a5\uff0cBlock \u4e4b\u95f4\u4f7f\u7528 softmax \u6ce8\u610f\u529b\u3002\u6253\u4e2a\u6bd4\u65b9\u2014\u2014\u4e0d\u5fc5\u7ed9\u6bcf\u5c42\u697c\u90fd\u88c5\u7535\u68af\uff0c\u5728\u5173\u952e\u697c\u5c42\u4e4b\u95f4\u67b6\u8bbe\u5feb\u901f\u901a\u9053\u5c31\u591f\u4e86\u3002<\/p>\n

\"\u9a6c\u65af\u514b\u4e0b\u573a\u70b9\u8d5e\uff01Kimi<\/p>\n

\u8fd9\u6837\uff0c\u9700\u8981\u4fdd\u5b58\u548c\u4f20\u8f93\u7684\u6570\u636e\u91cf\uff0c\u4ece\u300c\u6240\u6709\u5c42\u7684\u6570\u91cf\u300d\u964d\u4f4e\u5230\u300c\u5757\u7684\u6570\u91cf\u300d\uff0c\u5f00\u9500\u5927\u5e45\u7f29\u5c0f\u3002\u5b9e\u9a8c\u53d1\u73b0\uff0c\u5206\u6210\u7ea6 8 \u4e2a\u5757\u5c31\u80fd\u4fdd\u7559\u5168\u91cf\u65b9\u6cd5\u7edd\u5927\u90e8\u5206\u7684\u6027\u80fd\u63d0\u5347\u3002<\/p>\n

\u5728\u5177\u4f53\u7684\u5de5\u7a0b\u5b9e\u73b0\u4e0a\uff0c\u56e2\u961f\u8fd8\u505a\u4e86\u4e24\u9879\u4f18\u5316\u3002<\/p>\n

\u8bad\u7ec3\u7aef\u8bbe\u8ba1\u4e86\u8de8\u9636\u6bb5\u7f13\u5b58\u673a\u5236\uff0c\u5728\u6d41\u6c34\u7ebf\u5e76\u884c\u8bad\u7ec3\u4e2d\u6bcf\u6b21\u5207\u6362\u9636\u6bb5\u65f6\u53ea\u4f20\u8f93\u65b0\u589e\u7684\u90a3\u4e00\u5c0f\u90e8\u5206\u5757\u6570\u636e\uff0c\u800c\u4e0d\u662f\u6bcf\u6b21\u90fd\u628a\u5168\u90e8\u5386\u53f2\u91cd\u65b0\u4f20\u4e00\u904d\uff0c\u5b9e\u6d4b\u6574\u4f53\u8bad\u7ec3\u989d\u5916\u5f00\u9500\u4e0d\u8d85\u8fc7 4%\u3002<\/p>\n

\u63a8\u7406\u7aef\u8bbe\u8ba1\u4e86\u4e24\u9636\u6bb5\u8ba1\u7b97\u7b56\u7565\uff0c\u628a\u4e00\u4e2a\u5757\u5185\u6240\u6709\u5c42\u7684\u67e5\u8be2\u6253\u5305\u6210\u4e00\u6b21\u77e9\u9635\u8fd0\u7b97\u7edf\u4e00\u5904\u7406\uff0c\u628a\u91cd\u590d\u7684\u5185\u5b58\u8bbf\u95ee\u644a\u9500\u6389\uff0c\u6700\u7ec8\u63a8\u7406\u5ef6\u8fdf\u589e\u52a0\u4e0d\u8d85\u8fc7 2%\u3002<\/p>\n

\u90a3\u5b9e\u9a8c\u6548\u679c\u600e\u4e48\u6837\u5462\uff1f\u7814\u7a76\u4eba\u5458\u6d4b\u4e86\u4e94\u4e2a\u4e0d\u540c\u89c4\u6a21\u7684\u6a21\u578b\u3002<\/p>\n

\u7ed3\u679c\u663e\u793a\uff0cBlock AttnRes \u5728\u5168\u90e8\u89c4\u6a21\u4e0a\u5747\u4ee5\u66f4\u4f4e\u7684\u9a8c\u8bc1\u635f\u5931\u9886\u5148\u4e8e\u57fa\u7ebf\uff0c\u4e14\u6539\u5584\u5e45\u5ea6\u968f\u89c4\u6a21\u589e\u5927\u800c\u7a33\u5b9a\u4fdd\u6301\u3002\u6309\u62df\u5408\u66f2\u7ebf\u63a8\u7b97\uff0c\u5728\u76f8\u540c\u7684\u8ba1\u7b97\u91cf\u4e0b\uff0cBlock AttnRes \u76f8\u5f53\u4e8e\u57fa\u7ebf\u6a21\u578b\u7528 1.25 \u500d\u7b97\u529b\u624d\u80fd\u8fbe\u5230\u7684\u6548\u679c\u3002<\/p>\n

\u5728 48B \u53c2\u6570\uff083B \u6fc0\u6d3b\uff09\u89c4\u6a21\u7684 Kimi Linear \u67b6\u6784\u5b9e\u9a8c\u4e2d\uff0cBlock AttnRes \u5c55\u73b0\u4e86\u6781\u5f3a\u7684\u6cdb\u5316\u6027\uff1a\u5728\u5168\u90e8 15 \u9879\u4e3b\u6d41\u8bc4\u6d4b\u57fa\u51c6\u4e2d\uff0c\u5176\u8868\u73b0\u5747\u6301\u5e73\u6216\u4f18\u4e8e PreNorm \u57fa\u7ebf\u6a21\u578b\u3002<\/p>\n

\"\u9a6c\u65af\u514b\u4e0b\u573a\u70b9\u8d5e\uff01Kimi<\/p>\n

\u4f8b\u5982\uff0c\u5728\u535a\u58eb\u7ea7\u79d1\u5b66\u63a8\u7406 GPQA-Diamond \u4e0a\u5b9e\u73b0\u4e86 7.5% \u7684\u98de\u8dc3\uff0c\u5728\u6570\u5b66 Math (+3.6%) \u53ca\u4ee3\u7801\u751f\u6210 HumanEval (+3.1%) \u4efb\u52a1\u4e2d\u4e5f\u5f55\u5f97\u4e86\u663e\u8457\u589e\u76ca \u3002<\/p>\n

\u4ece\u8bad\u7ec3\u8fc7\u7a0b\u6765\u770b\uff0c\u57fa\u7ebf\u6a21\u578b\u7684\u5404\u5c42\u8f93\u51fa\u6570\u503c\u968f\u6df1\u5ea6\u5355\u8c03\u589e\u5927\uff0c\u5370\u8bc1\u4e86 PreNorm \u7a00\u91ca\u95ee\u9898\uff1b\u800c AttnRes \u7684\u5404\u5c42\u8f93\u51fa\u6570\u503c\u5728\u5757\u8fb9\u754c\u5904\u5f97\u5230\u91cd\u7f6e\uff0c\u5448\u73b0\u5468\u671f\u6027\u53d8\u5316\uff0c\u5404\u5c42\u68af\u5ea6\u5206\u5e03\u4e5f\u66f4\u52a0\u5747\u5300\uff0c\u8bf4\u660e\u66f4\u591a\u7684\u5c42\u771f\u6b63\u53c2\u4e0e\u5230\u4e86\u6709\u6548\u7684\u5b66\u4e60\u4e2d\u3002<\/p>\n

\"\u9a6c\u65af\u514b\u4e0b\u573a\u70b9\u8d5e\uff01Kimi<\/p>\n

\u6b64\u5916\uff0c\u7814\u7a76\u4eba\u5458\u8fd8\u53ef\u89c6\u5316\u4e86\u8bad\u7ec3\u540e\u6a21\u578b\u5b66\u5230\u7684\u6ce8\u610f\u529b\u6743\u91cd\uff0c\u53d1\u73b0\u4e86\u51e0\u4e2a\u6709\u8da3\u7684\u89c4\u5f8b\u3002<\/p>\n

\u6bcf\u4e00\u5c42\u4ecd\u7136\u6700\u4f9d\u8d56\u76f4\u63a5\u524d\u4e00\u5c42\u7684\u8f93\u51fa\uff0c\u5c40\u90e8\u6027\u4f9d\u7136\u662f\u4e3b\u8981\u7684\u4fe1\u606f\u6d41\u901a\u65b9\u5f0f\u3002\u4f46\u540c\u65f6\u51fa\u73b0\u4e86\u4e00\u4e9b\u8df3\u8dc3\u6027\u7684\u8fde\u63a5\uff0c\u6bd4\u5982\u67d0\u4e9b\u5c42\u4f1a\u7a33\u5b9a\u5730\u56de\u6eaf\u5230\u5f88\u65e9\u671f\u7684\u5c42\uff0c\u8fd8\u6709\u4e9b\u5c42\u4f1a\u7279\u522b\u5173\u6ce8\u6700\u521d\u7684\u8bcd\u5d4c\u5165\u8f93\u51fa\u3002<\/p>\n

\u53e6\u4e00\u4e2a\u89c4\u5f8b\u662f\uff0c\u6ce8\u610f\u529b\u5c42\u548c MLP \u5c42\u7684\u300c\u56de\u671b\u300d\u6a21\u5f0f\u4e0d\u540c\uff1a\u6ce8\u610f\u529b\u5c42\u503e\u5411\u4e8e\u5173\u6ce8\u66f4\u5e7f\u6cdb\u7684\u5386\u53f2\uff0cMLP \u5c42\u5219\u66f4\u4f9d\u8d56\u8fd1\u90bb\u5c42\u3002\u8fd9\u4e0e\u4e24\u8005\u5728\u6a21\u578b\u4e2d\u7684\u529f\u80fd\u5206\u5de5\u662f\u543b\u5408\u7684\u3002<\/p>\n

AttnRes\u8fd8\u5e26\u6765\u4e86\u4e00\u4e2a\u5bf9\u672a\u6765\u6a21\u578b\u8bbe\u8ba1\u6709\u53c2\u8003\u4ef7\u503c\u7684\u53d1\u73b0\u3002\u7814\u7a76\u4eba\u5458\u5728\u56fa\u5b9a\u603b\u8ba1\u7b97\u91cf\u548c\u53c2\u6570\u91cf\u7684\u524d\u63d0\u4e0b\uff0c\u679a\u4e3e\u4e86 25 \u79cd\u4e0d\u540c\u7684\u6df1\u5ea6\u4e0e\u5bbd\u5ea6\u7ec4\u5408\uff0c\u5bf9\u6bd4\u57fa\u7ebf\u6a21\u578b\u548c AttnRes \u5404\u81ea\u504f\u597d\u7684\u6700\u4f18\u67b6\u6784\u3002<\/p>\n

\"\u9a6c\u65af\u514b\u4e0b\u573a\u70b9\u8d5e\uff01Kimi<\/p>\n

\u7ed3\u679c\u53d1\u73b0\uff0c\u6807\u51c6\u6b8b\u5dee\u8fde\u63a5\u504f\u597d\u300c\u66f4\u5bbd\u3001\u5c42\u6570\u66f4\u5c11\u300d\u7684\u6a21\u578b\uff0c\u800c AttnRes \u7684\u6700\u4f18\u70b9\u504f\u5411\u300c\u66f4\u7a84\u3001\u5c42\u6570\u66f4\u591a\u300d\u7684\u6a21\u578b\u3002\u8fd9\u8bf4\u660e AttnRes \u80fd\u591f\u66f4\u6709\u6548\u5730\u5229\u7528\u6df1\u5ea6\uff0c\u8ba9\u6bcf\u589e\u52a0\u4e00\u5c42\u90fd\u771f\u6b63\u4ea7\u751f\u4ef7\u503c\uff0c\u800c\u4e0d\u662f\u8ba9\u6df1\u5ea6\u53d8\u6210\u4e00\u79cd\u8fb9\u9645\u6548\u76ca\u9012\u51cf\u7684\u5806\u780c\u3002<\/p>\n

\u8fd9\u4e2a\u53d1\u73b0\u7684\u542b\u4e49\u4e0d\u6b62\u4e8e\u6b64\u3002\u5b83\u610f\u5473\u7740 AttnRes \u4e0d\u53ea\u662f\u5728\u539f\u6709\u67b6\u6784\u4e0a\u6253\u4e86\u4e00\u4e2a\u8865\u4e01\uff0c\u800c\u662f\u4ece\u6839\u672c\u4e0a\u6539\u53d8\u4e86\u7f51\u7edc\u6df1\u5ea6\u7684\u5229\u7528\u6548\u7387\uff0c\u4e5f\u4e3a\u672a\u6765\u8bbe\u8ba1\u5927\u6a21\u578b\u65f6\u5982\u4f55\u5206\u914d\u6df1\u5ea6\u4e0e\u5bbd\u5ea6\u7684\u8d44\u6e90<\/a>\u63d0\u4f9b\u4e86\u65b0\u7684\u53c2\u8003\u4f9d\u636e\u3002<\/p>\n<\/p>\n

\u6768\u690d\u9e9f\u66fe\u63d0\u5230\uff0c\u5341\u5e74\u524d\u4e0d\u662f\u6ca1\u6709\u597d\u60f3\u6cd5\uff0c\u800c\u662f\u6ca1\u6709\u7b97\u529b\u53bb\u9a8c\u8bc1\u3002\u73b0\u5728\u6709\u4e86\u8db3\u591f\u7684\u8d44\u6e90\u548c\u300c\u7f29\u653e\u9636\u68af\uff08Scaling Ladder\uff09\u300d\uff0c\u90a3\u4e9b\u88ab\u6401\u7f6e\u7684\u95ee\u9898\u624d\u7ec8\u4e8e\u80fd\u88ab\u8ba4\u771f\u7b54\u4e00\u904d\u3002<\/p>\n

\u5927\u4f6c\u70b9\u8d5e\u7684\u80cc\u540e\uff0c\u662f\u4e00\u4e2a\u65f6\u4ee3\u5728\u8f6c\u5f2f<\/p>\n

\u4e00\u4e2a\u4e2d\u56fd\u56e2\u961f\u5728\u6700\u5e95\u5c42\u7684\u67b6\u6784\u521b\u65b0<\/a>\u4e0a\u83b7\u5f97\u7845\u8c37\u9876\u7ea7\u4eba\u7269\u7684\u5b9e\u8d28\u6027\u8ba4\u53ef\uff0c\u8fd9\u4ef6\u4e8b\u672c\u8eab\u5341\u5206\u7f55\u89c1\uff0c\u4ed6\u4eec\u8ba4\u53ef\u7684\u4e0d\u53ea\u662f\u8bba\u6587\u6210\u679c\u672c\u8eab\uff0c\u66f4\u5728\u4e8eKimi \u8fd9\u7bc7\u8bba\u6587\u6307\u5411\u4e86\u4e00\u4e2a\u5168\u65b0\u7684\u65b9\u5411\uff1a\u4f18\u5316\u5df2\u7ecf\u4ece attention\u3001MoE \u8fd9\u4e9b\u4e0a\u5c42\u6a21\u5757\uff0c\u6df1\u5165\u5230\u4e86\u6700\u5e95\u5c42\u7684\u6b8b\u5dee\u8fde\u63a5\u3002<\/p>\n

\u5728 GTC 2026 \u6f14\u8bb2\u4e2d\uff0c\u6768\u690d\u9e9f\u8fd8\u62ab\u9732\u4e86\u4e00\u8fde\u4e32\u5e95\u5c42\u6280\u672f\u521b\u65b0\uff1aMuonClip \u4f18\u5316\u5668\u5b9e\u73b0\u4e86\u76f8\u6bd4 AdamW 2 \u500d\u7684\u8ba1\u7b97\u6548\u7387\u63d0\u5347\u2014\u2014\u8981\u77e5\u9053 Adam \u4f18\u5316\u5668\u81ea 2014 \u5e74\u4ee5\u6765\u51e0\u4e4e\u672a\u88ab\u64bc\u52a8\uff0c\u5c5e\u4e8e\u6df1\u5ea6\u5b66\u4e60\u7684\u300c\u4e0d\u53ef\u89e6\u78b0\u4e4b\u7269\u300d\uff1bKimi Linear\uff08KDA \u67b6\u6784\uff09\u5728 128K \u5230\u767e\u4e07\u7ea7\u8d85\u957f\u4e0a\u4e0b\u6587\u4e0b\u5b9e\u73b0 5-6 \u500d\u7684\u89e3\u7801\u52a0\u901f\uff1bVision RL \u7684\u8de8\u6a21\u6001\u8bad\u7ec3\u751a\u81f3\u8ba9\u7eaf\u6587\u672c benchmark \u4e5f\u63d0\u5347\u4e86\u7ea6 2.1%\u3002<\/p>\n

\u6768\u690d\u9e9f\u628a\u8fd9\u4e9b\u521b\u65b0\u6982\u62ec\u4e3a\u4e09\u4e2a\u7ef4\u5ea6\u7684 Scaling \u6846\u67b6\uff1aToken \u6548\u7387 × \u957f\u4e0a\u4e0b\u6587 × Agent Swarms\u3002<\/p>\n

\u300c\u5f53\u524d\u7684 Scaling \u5df2\u7ecf\u4e0d\u518d\u662f\u5355\u7eaf\u7684\u8d44\u6e90\u5806\u780c\uff0c\u800c\u662f\u8981\u5728\u8ba1\u7b97\u6548\u7387\u3001\u957f\u7a0b\u8bb0\u5fc6\u548c\u81ea\u52a8\u5316\u534f\u4f5c\u4e0a\u540c\u65f6\u5bfb\u627e\u89c4\u6a21\u6548\u5e94\u3002\u300d<\/p>\n

\u4e00\u5bb6\u516c\u53f8\uff0c\u540c\u65f6\u5728\u4f18\u5316\u5668\u3001\u6b8b\u5dee\u8fde\u63a5\u3001\u6ce8\u610f\u529b\u67b6\u6784\u3001\u8de8\u6a21\u6001\u8bad\u7ec3\u8fd9\u4e9b\u5e95\u5c42\u6218\u573a\u4e0a\u5168\u7ebf\u63a8\u8fdb\uff0c\u8fd9\u79cd\u6253\u6cd5\u5728\u884c\u4e1a\u91cc\u76f8\u5f53\u7279\u7acb\u72ec\u884c\u3002<\/p>\n

\u8fd9\u4e5f\u662f\u4e3a\u4ec0\u4e48 Jerry Tworek \u4f1a\u8bf4\u51fa\u300c\u6df1\u5ea6\u5b66\u4e60 2.0\u300d\u8fd9\u6837\u7684\u5224\u65ad\u3002\u5f53\u7136\u4e0d\u662f\u8bf4 Attention Residuals \u8fd9\u7bc7\u8bba\u6587\u5c31\u80fd\u98a0\u8986\u4e00\u5207\uff0c\u66f4\u591a\u662f\u5b83\u4ee3\u8868\u4e86\u4e00\u79cd\u65b9\u6cd5\u8bba\u7684\u56de\u5f52\uff1a\u4e0d\u518d\u6ee1\u8db3\u4e8e\u5728\u5df2\u6709\u6846\u67b6\u4e0a\u4fee\u4fee\u8865\u8865\uff0c\u53bb\u91cd\u65b0\u5ba1\u89c6\u90a3\u4e9b\u88ab\u6240\u6709\u4eba\u5f53\u4f5c\u300c\u5df2\u89e3\u51b3\u95ee\u9898\u300d\u7684\u57fa\u7840\u8bbe\u65bd\u3002<\/p>\n

\u5982\u679c\u6b8b\u5dee\u8fde\u63a5\u53ef\u4ee5\u88ab\u91cd\u65b0\u8bbe\u8ba1\uff0c\u90a3\u4e48 Adam \u4f18\u5316\u5668\u5462\uff1f\u5c42\u5f52\u4e00\u5316\u5462\uff1f\u4f4d\u7f6e\u7f16\u7801\u5462\uff1f\u6df1\u5ea6\u5b66\u4e60\u7684\u57fa\u7840\u8303\u5f0f\u672c\u8eab\u6b63\u5728\u53d1\u751f\u53d8\u5316\uff0c\u8fd9\u6247\u95e8\u4e00\u65e6\u63a8\u5f00\uff0c\u540e\u9762\u7684\u6545\u4e8b\u5c31\u4e0d\u518d\u662f\u7ebf\u6027\u5916\u63a8\u80fd\u9884\u6d4b\u7684\u4e86\u3002<\/p>\n

Karpathy \u90a3\u53e5\u300cAttention is All You Need \u8fd8\u6ca1\u88ab\u7406\u89e3\u900f\u300d\u7684\u611f\u6168\uff0c\u5927\u6982\u4e5f\u662f\u8fd9\u4e2a\u610f\u601d\u3002<\/p>\n

\u8fc7\u53bb\u51e0\u5e74\uff0c\u4e2d\u56fd AI \u56e2\u961f\u7684\u8d21\u732e\u66f4\u591a\u96c6\u4e2d\u5728\u5de5\u7a0b\u843d\u5730\u548c\u5e94\u7528\u521b\u65b0\u4e0a\uff0c\u5728\u5e95\u5c42\u67b6\u6784\u7406\u8bba\u65b9\u9762\u7684\u539f\u521b\u6027\u7a81\u7834\u76f8\u5bf9\u7a00\u7f3a\u3002Kimi \u8fd9\u7bc7\u8bba\u6587\u8d70\u7684\u662f\u4e00\u6761\u5b8c\u5168\u4e0d\u540c\u7684\u8def\u7ebf\u2014\u2014\u4e00\u4e2a\u7edf\u4e00\u7684\u7406\u8bba\u6846\u67b6\uff0c\u4e00\u4e2a\u4f18\u96c5\u7684\u5de5\u7a0b\u5b9e\u73b0\uff0c\u52a0\u4e0a\u4e25\u8c28\u7684\u5927\u89c4\u6a21\u5b9e\u9a8c\u9a8c\u8bc1\u3002<\/p>\n

\u5f53\u7136\uff0cKimi \u8fd9\u7bc7\u8bba\u6587\u8fd8\u6709\u7559\u4e0b\u4e0d\u5c11\u9700\u8981\u89e3\u51b3\u7684\u95ee\u9898\u3002\u8bba\u6587\u7684\u5927\u89c4\u6a21\u9a8c\u8bc1\u662f\u5728 48B \u603b\u53c2\u6570\uff083B \u6fc0\u6d3b\u53c2\u6570\uff09\u7684\u6a21\u578b\u4e0a\u5b8c\u6210\u7684\uff0c\u8fd9\u4e2a\u89c4\u6a21\u653e\u5728\u4eca\u5929\u7684\u7b2c\u4e00\u68af\u961f\u91cc\u5e76\u4e0d\u7b97\u5927\u3002\u5728\u771f\u6b63\u7684\u5343\u4ebf\u4e43\u81f3\u4e07\u4ebf\u53c2\u6570\u6a21\u578b\u4e0a\uff0c1.25 \u500d\u7684\u7b49\u6548\u4f18\u52bf\u80fd\u5426\u7a33\u4f4f\uff0c\u76ee\u524d\u8fd8\u662f\u4e2a\u95ee\u53f7\u3002<\/p>\n

\u540c\u65f6\u8bba\u6587\u5c55\u793a\u7684\u4e5f\u53ea\u662f\u9884\u8bad\u7ec3\u9636\u6bb5\u7684\u6536\u76ca\uff0c\u7ecf\u8fc7\u6307\u4ee4\u5fae\u8c03\u3001RLHF \u7b49\u540e\u8bad\u7ec3\u6b65\u9aa4\u540e\uff0cAttnRes \u7684\u4f18\u52bf\u662f\u5426\u4f1a\u88ab\u7a00\u91ca\uff0c\u7f3a\u4e4f\u6570\u636e\u3002<\/p>\n

\u4f46\u8bdd\u8bf4\u56de\u6765\uff0c\u8fd9\u4e9b\u5c40\u9650\u6070\u6070\u4e5f\u662f\u60f3\u8c61\u529b\u7684\u6765\u6e90\u3002\u4e00\u4e2a\u4ec5\u9700\u7ea6 100 \u884c\u4ee3\u7801\u6539\u52a8\u3001\u589e\u52a0\u4e0d\u5230 4% \u8bad\u7ec3\u5f00\u9500\u7684\u8f7b\u91cf\u4fee\u6539\uff0c\u5c31\u80fd\u5728 48B \u89c4\u6a21\u4e0a\u5e26\u6765\u8fd9\u6837\u7684\u63d0\u5347\u3002<\/p>\n

\u5f53\u5b83\u88ab\u5e94\u7528\u5230\u66f4\u5927\u89c4\u6a21\u7684\u4e0b\u4e00\u4ee3\u6a21\u578b\u4e0a\u65f6\uff0c\u6536\u76ca\u7684\u5929\u82b1\u677f\u5728\u54ea\u91cc\uff0c\u8c01\u4e5f\u8bf4\u4e0d\u51c6\u3002<\/p>\n

Attention Residuals \u62ac\u9ad8\u4e86 Token \u6548\u7387\u7684\u5929\u82b1\u677f\uff0cKimi Linear \u62d3\u5c55\u4e86\u957f\u4e0a\u4e0b\u6587\u7684\u8fb9\u754c\uff0cAgent Swarms \u6307\u5411\u667a\u80fd\u4f53\u534f\u4f5c\u7684\u672a\u6765\u3002\u5f53\u8fd9\u4e09\u6761\u6280\u672f\u7ebf\u5728\u4e0b\u4e00\u4ee3\u6a21\u578b\u4e2d\u6c47\u5408\uff0c\u5448\u73b0\u51fa\u7684\u53ef\u80fd\u5c31\u662f\u65b0\u7684\u8303\u5f0f\u8f6c\u53d8\u3002<\/p>\n

\u5728 AI \u8fd9\u5ea7\u901a\u5929\u5854\u7684\u5de5\u7a0b\u4e0a\uff0c\u6240\u6709\u4eba\u90fd\u5728\u4e89\u7740\u5f80\u4e0a\u6dfb\u7816\u52a0\u74e6\uff0c\u800c Kimi \u4f4e\u5934\u5f80\u8def\u57fa\u91cd\u91cd\u5730\u51ff\u4e86\u4e00\u9539\uff0c\u6070\u597d\u64ac\u52a8\u4e86\u6df1\u5ea6\u5b66\u4e60\u7684\u5730\u57fa\u3002<\/p>\n

\u4f5c\u8005\uff1a\u83ab\u5d07\u5b87\uff0c\u674e\u8d85\u51e1<\/p>\n

(\u6765\u6e90\uff1a\u65b0\u6d6a\u79d1\u6280)<\/p>\n","protected":false},"excerpt":{"rendered":"

\u540c\u6837\u7684\u7b97\u529b\uff0c\u540c\u6837\u7684\u6570\u636e\uff0c\u51ed\u4ec0\u4e48\u6548\u679c\u4e0d\u4e00\u6837\uff1f\u5927\u591a\u6570\u4eba\u7684\u76f4\u89c9\u662f\uff1a\u6a21\u578b\u66f4\u5927\u3001\u6570\u636e\u66f4\u597d\u3001\u5de5\u7a0b\u5e08\u66f4\u5389\u5bb3\u3002\u4f46 Kimi \u7ed9 […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":[],"categories":[6],"tags":[378,252,280,157,88,238,376],"views":20,"_links":{"self":[{"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=\/wp\/v2\/posts\/749110"}],"collection":[{"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=749110"}],"version-history":[{"count":0,"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=\/wp\/v2\/posts\/749110\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=749110"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=749110"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=749110"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}