{"id":688,"date":"2026-07-28T10:54:34","date_gmt":"2026-07-28T02:54:34","guid":{"rendered":"https:\/\/aitimes.link\/?p=688"},"modified":"2026-07-28T11:18:53","modified_gmt":"2026-07-28T03:18:53","slug":"47-ye-kimik3-ji-shu-bao-gao-jiang-gei-pu-tong-kai-fa-zhe","status":"publish","type":"post","link":"https:\/\/aitimes.link\/index.php\/2026\/07\/28\/47-ye-kimik3-ji-shu-bao-gao-jiang-gei-pu-tong-kai-fa-zhe\/","title":{"rendered":"47 \u9875 Kimi K3 \u6280\u672f\u62a5\u544a\uff0c\u8bb2\u7ed9\u666e\u901a\u5f00\u53d1\u8005\u542c\uff1a2.8T \u5f00\u6e90\u5de8\u517d\u7684\u201c\u5077\u61d2\u201d\u827a\u672f"},"content":{"rendered":"

30 \u79d2\u7248\u672c\uff1a<\/strong>Kimi K3 \u662f Moonshot \u6700\u65b0\u5f00\u6e90\u7684\u65d7\u8230\u6a21\u578b\u2014\u20142.8 \u4e07\u4ebf\u53c2\u6570\u7684 MoE\uff0c\u6bcf\u4e2a token \u53ea\u6fc0\u6d3b 1040 \u4ebf\uff08\u7ea6 1\/27\uff09\uff0c\u539f\u751f\u89c6\u89c9\uff0c100 \u4e07 token \u4e0a\u4e0b\u6587<\/strong>\u3002\u5b9a\u4f4d\u4e00\u53e5\u8bdd\uff1a\u80fd\u529b\u54ac\u4f4f\u95ed\u6e90\u524d\u4e24\u540d\uff08Claude Fable 5\u3001GPT-5.6 Sol\uff09\uff0c\u7a33\u5750\u5f00\u6e90\u7b2c\u4e00\uff0c\u6210\u672c\u5374\u53ea\u6709\u5bf9\u624b\u7684\u51e0\u5206\u4e4b\u4e00\u3002\u672c\u6587\u628a 47 \u9875\u6280\u672f\u62a5\u544a\u62c6\u6210\u4e09\u4e2a\u7075\u9b42\u95ee\u9898\uff1a<\/p>\n

    \n
  1. 2.8T \u7684\u5927\u5bb6\u4f19\uff0c\u600e\u4e48\u8bad\u5f97\u8d77\u3001\u63a8\u5f97\u52a8\uff1f<\/strong> \u2192 \u67b6\u6784”\u4e09\u677f\u65a7”<\/li>\n
  2. 100 \u4e07 token \u4e0a\u4e0b\u6587\uff0c\u600e\u4e48\u4e0d\u8ba9\u663e\u5b58\u7206\u70b8\uff1f<\/strong> \u2192 \u6362\u4e00\u79cd”\u8bb0\u5fc6\u65b9\u5f0f”<\/li>\n
  3. base \u6a21\u578b\u600e\u4e48\u7ec3\u6210\u80fd\u5e72\u6d3b\u7684 agent\uff1f<\/strong> \u2192 RL + \u4e5d\u4e13\u5bb6\u84b8\u998f<\/li>\n<\/ol>\n

    \u5148\u770b\u5168\u666f\uff1a\u67b6\u6784\u5c31\u4e09\u53e5\u8bdd<\/h2>\n

    <\/p>\n

    \n \n \n <\/marker>\n <\/defs>\n \n MoonViT-V2<\/text>\n 0.4B \u89c6\u89c9\u7f16\u7801\u5668<\/text>\n \n \n Embedding<\/text>\n \u5171\u4eab\u5d4c\u5165\u7a7a\u95f4<\/text>\n \n \n Block \u00d78\uff1a93 \u5c42\u5207\u6210 8 \u5757<\/text>\n token \u6df7\u5408\uff1a3 KDA : 1 MLA<\/text>\n \n KDA<\/text>\n \n KDA<\/text>\n \n KDA<\/text>\n \n MLA<\/text>\n \n \n \n \n \n MoE<\/text>\n \n MoE<\/text>\n \n MoE<\/text>\n \n MoE<\/text>\n \u6bcf\u5c42 attention \u540e\u90fd\u8ddf\u4e00\u4e2a LatentMoE<\/text>\n 896 \u9009 16 \u00b7 \u538b\u5230 3584 \u7ef4 latent \u7a7a\u95f4<\/text>\n \n \n AttnRes\uff1a\u6df1\u5ea6\u65b9\u5411\u7684”\u67e5\u6863”\u673a\u5236<\/text>\n \u6bcf\u5c42\u5e26 query \u68c0\u7d22 embedding + \u6240\u6709\u524d\u5e8f Block \u7684\u8f93\u51fa<\/text>\n \n \n \n \u8f93\u51fa<\/text>\n \u5faa\u73af \u00d78<\/text>\n<\/svg><\/p>\n

    \u4e09\u53e5\u8bdd\u6982\u62ec\u6574\u4e2a\u67b6\u6784\uff1atoken \u7ef4\u5ea6<\/strong>\u9760 3 \u5c42 KDA + 1 \u5c42 MLA \u6df7\u642d\uff0c\u6df1\u5ea6\u7ef4\u5ea6<\/strong>\u9760 AttnRes \u8de8\u5c42\u67e5\u6863\uff0c\u5bbd\u5ea6\u7ef4\u5ea6<\/strong>\u9760 LatentMoE \u7a00\u758f\u8c03\u7528\u4e13\u5bb6\u3002\u5148\u611f\u53d7\u4e00\u4e0b K2 \u2192 K3 \u7684\u4f53\u683c\u53d8\u5316\uff1a<\/p>\n\n\n\n\n\n\n\n\n
    <\/th>\nKimi K2<\/th>\nKimi K3<\/th>\n\u53d8\u5316<\/th>\n<\/tr>\n
    \u603b\u53c2\u6570 \/ \u6fc0\u6d3b\u53c2\u6570<\/td>\n1.04T \/ 32.6B<\/td>\n2.78T \/ 104.2B<\/td>\n\u00d72.7 \/ \u00d73.2<\/td>\n<\/tr>\n
    \u5c42\u6570<\/td>\n61<\/td>\n93<\/td>\n+52%<\/td>\n<\/tr>\n
    \u4e13\u5bb6\u6570\uff08\u6bcf token \u6fc0\u6d3b\uff09<\/td>\n384\uff088\uff09<\/td>\n896\uff0816\uff09<\/td>\n\u7a00\u758f\u5ea6 56<\/td>\n<\/tr>\n
    \u8bad\u7ec3\u4e0a\u4e0b\u6587<\/td>\n128K<\/td>\n1M<\/td>\n\u00d78<\/td>\n<\/tr>\n
    \u6ce8\u610f\u529b<\/td>\n\u5168 MLA<\/td>\n69 KDA + 24 MLA<\/td>\n\u6df7\u5408\u67b6\u6784<\/td>\n<\/tr>\n
    \u7efc\u5408 Scaling \u6548\u7387<\/td>\n\u57fa\u51c6<\/td>\n2.5\u00d7<\/strong><\/td>\n\u540c\u7b49 loss \u7701 60% \u7b97\u529b<\/td>\n<\/tr>\n<\/table>\n

    \u7b2c\u4e00\u677f\u65a7\uff1aKDA \u2014\u2014 \u628a”\u65e0\u9650\u589e\u957f\u7684\u4f1a\u8bae\u8bb0\u5f55”\u6362\u6210”\u56fa\u5b9a\u5927\u5c0f\u7684\u7b14\u8bb0\u672c”<\/h2>\n

    \u4f20\u7edf attention \u7684\u75db\u70b9\uff1a\u6bcf\u6765\u4e00\u4e2a token\uff0c\u90fd\u8981\u548c\u4e4b\u524d\u6240\u6709<\/strong> token \u91cd\u65b0\u7b97\u4e00\u904d\u5173\u7cfb\uff0cKV cache \u65e0\u9650\u53d8\u957f\u2014\u2014\u50cf\u5f00\u4e00\u573a\u6c38\u4e0d\u6563\u4f1a\u7684\u4f1a\uff0c\u6bcf\u9875\u53d1\u8a00\u8bb0\u5f55\u90fd\u8981\u6c38\u4e45\u5b58\u6863\u3002KDA\uff08Kimi Delta Attention\uff09\u662f\u7ebf\u6027\u6ce8\u610f\u529b\u7684\u4e00\u79cd\uff1a\u4e0d\u7ba1\u5e8f\u5217\u591a\u957f\uff0c\u53ea\u7ef4\u62a4\u4e00\u4e2a\u56fa\u5b9a\u5927\u5c0f\u7684\u7b14\u8bb0\u672c S<\/strong>\uff0c\u6210\u672c O(n)\u3001\u7f13\u5b58\u6052\u5b9a\u3002<\/p>\n

    <\/p>\n

    \n \n \n <\/marker>\n <\/defs>\n \n \u5168\u91cf\u6ce8\u610f\u529b\uff08\u5982 MLA\uff09<\/text>\n \u6bcf\u6765 1 \u4e2a token\uff0c\u8bb0\u5f55 +1 \u9875<\/text>\n KV cache<\/text>\n \n \n \n \n \n \n \u6210\u672c O(n\u00b2) \u00b7 \u7f13\u5b58\u65e0\u9650\u6da8<\/text>\n \n KDA\uff08\u7ebf\u6027\u6ce8\u610f\u529b\uff09<\/text>\n \u6bcf\u6765 1 \u4e2a token\uff0c\u53ea\u66f4\u65b0\u7b14\u8bb0\u672c S<\/text>\n \n \n \n \n \n \u2460 \u892a\u8272 \u03b1\uff1a\u65e7\u7b14\u8bb0\u6309\u901a\u9053\u53d8\u6de1<\/text>\n \u2461 \u64e6\u9664\uff1a\u5148\u7cbe\u51c6\u64e6\u6389\u65e7\u503c<\/text>\n \u2462 \u5199\u5165 \u03b2\uff1a\u65b0\u5185\u5bb9\u6309\u5f3a\u5ea6\u843d\u7b14<\/text>\n \u6210\u672c O(n) \u00b7 \u7f13\u5b58\u56fa\u5b9a<\/text>\n \n K3 \u7684\u6298\u4e2d\uff1a3 \u5c42 KDA + 1 \u5c42 MLA<\/text>\n 75% \u7684\u5c42\u8bb0\u7b14\u8bb0\u672c\uff08\u4fbf\u5b9c\uff09\uff0c25% \u5f00\u5168\u5458\u5927\u4f1a\uff08\u4fdd\u5168\u5c40\uff09<\/text>\n<\/svg><\/p>\n

    \u7b14\u8bb0\u672c\u6bcf\u6765\u4e00\u4e2a token \u505a\u4e09\u4e2a\u52a8\u4f5c\uff1a\u892a\u8272<\/strong>\uff08forget gate \u03b1 \u8ba9\u6bcf\u4e2a\u901a\u9053\u7684\u65e7\u7b14\u8bb0\u6309\u5404\u81ea\u901f\u5ea6\u53d8\u6de1\uff09\u3001\u64e6\u9664<\/strong>\uff08delta rule\u2014\u2014\u5199\u5165\u524d\u5148\u628a\u65e7\u503c\u7cbe\u51c6\u64e6\u6389\uff0c\u8fd9\u662f\u5b83\u548c\u666e\u901a\u7ebf\u6027\u6ce8\u610f\u529b\u7684\u5173\u952e\u533a\u522b\uff0c\u9632\u6b62\u5185\u5bb9\u8d8a\u53e0\u8d8a\u7cca\uff09\u3001\u5199\u5165<\/strong>\uff08\u65b0\u5185\u5bb9\u6309\u5f3a\u5ea6 \u03b2 \u843d\u7b14\uff09\u3002<\/p>\n

    \u62a5\u544a\u91cc\u6211\u6700\u559c\u6b22\u7684\u4e00\u4e2a\u5de5\u7a0b\u7ec6\u8282\uff1a\u6570\u5b66\u4e0a”1\/\u7d2f\u8ba1\u8870\u51cf”\u4f1a\u6307\u6570\u7ea7\u81a8\u80c0\uff0cBF16 \u88c5\u4e0d\u4e0b\u5c31\u6ea2\u51fa\u3002K3 \u7684\u89e3\u6cd5\u662f\u7ed9\u892a\u8272\u901f\u5ea6\u8bbe\u4e2a\u4e0b\u9650\uff08\u7528 sigmoid \u628a log-decay \u538b\u8fdb (-5, 0)\uff09\uff0c16 \u4e2a token \u7684\u7d2f\u8ba1\u8870\u51cf\u5012\u6570\u6700\u5927 e\u2078\u2070\uff0c\u521a\u597d\u5361\u5728 BF16 \u52a8\u6001\u8303\u56f4\u5185\u2014\u2014\u4e00\u4e2a\u6570\u503c\u7a33\u5b9a\u6027\u7684\u5c0f\u6539\u52a8\uff0c\u8ba9\u6574\u6761\u8ba1\u7b97\u8def\u5f84\u90fd\u80fd\u8dd1\u6ee1 Tensor Core\uff0c\u76f4\u63a5\u6d88\u706d\u4e86\u6700\u6162\u7684 kernel \u5206\u652f<\/strong>\u3002<\/p>\n

    \u4e24\u4e2a\u914d\u5957\u8bbe\u8ba1\u4e5f\u503c\u5f97\u8bb0\u4f4f\uff1a<\/p>\n

      \n
    • 3:1 \u6df7\u642d<\/strong>\uff1a\u5149\u9760\u7b14\u8bb0\u672c\u4f1a\u4e22\u5168\u5c40\u7ec6\u8282\uff0c\u6240\u4ee5\u6bcf 4 \u5c42\u63d2 1 \u5c42 Gated MLA \u5168\u91cf\u6ce8\u610f\u529b”\u5f00\u5168\u5458\u5927\u4f1a”<\/li>\n
    • NoPE\uff08\u65e0\u4f4d\u7f6e\u7f16\u7801\uff09<\/strong>\uff1a\u4e0d\u663e\u5f0f\u5b58”\u7b2c\u51e0\u4e2a\u4f4d\u7f6e”\uff0c\u9760\u7b14\u8bb0\u672c\u7684\u8870\u51cf\u81ea\u7136\u611f\u77e5\u8fdc\u8fd1\u3002\u4e0a\u4e0b\u6587\u4ece 8K \u4e00\u8def\u6269\u5230 1M \u5b8c\u5168\u4e0d\u7528\u6539\u4f4d\u7f6e\u7f16\u7801<\/strong><\/li>\n<\/ul>\n

      \u7b2c\u4e8c\u677f\u65a7\uff1aAttnRes \u2014\u2014 \u628a attention \u7684\u601d\u60f3\u642c\u5230”\u6df1\u5ea6”\u65b9\u5411<\/h2>\n

      \u666e\u901a\u6b8b\u5dee\u8fde\u63a5\u50cf\u4f20\u8bdd\u6e38\u620f\uff1a\u7b2c 93 \u5c42\u53ea\u80fd\u62ff\u5230\u7b2c 92 \u5c42\u4f20\u8fc7\u6765\u7684\u8bdd\uff0c\u524d\u9762 91 \u5c42\u7684\u7ec6\u8282\u5168\u88ab\u538b\u6241\u5728\u4e00\u4e2a\u5411\u91cf\u91cc\u3002AttnRes \u7684\u601d\u8def\u5f88\u76f4\u63a5\uff1a\u6bcf\u5c42\u5e26\u4e00\u4e2a\u53ef\u5b66\u4e60\u7684”\u67e5\u8be2\u8bcd”\uff0c\u4e3b\u52a8\u53bb\u68c0\u7d22 embedding \u548c\u6240\u6709\u524d\u5e8f Block \u7684\u8f93\u51fa<\/strong>\uff0c\u60f3\u8981\u8c01\u7684\u4fe1\u606f\u5c31\u62ff\u8c01\u7684\u300293 \u5c42\u5207\u6210 8 \u4e2a Block\uff0c\u5757\u5185\u5148\u6c42\u548c\u538b\u7f29\u3001\u5757\u95f4\u518d\u505a\u68c0\u7d22\uff0c\u5f00\u9500\u53ef\u5ffd\u7565\uff0c\u5b9e\u9a8c\u4e0a N=8 \u5c31\u80fd\u62ff\u56de\u5927\u90e8\u5206\u6536\u76ca\u3002<\/p>\n

      \u7b2c\u4e09\u677f\u65a7\uff1aLatentMoE \u2014\u2014 896 \u4e2a\u4e13\u5bb6\u7684\u5206\u8bca\u827a\u672f<\/h2>\n

      <\/p>\n

      \n \n \n <\/marker>\n <\/defs>\n \n token<\/text>\n 7168 \u7ef4<\/text>\n \n \n W\u2193 \u538b\u7f29<\/text>\n 7168\u21923584<\/text>\n \n \n Router \u5206\u8bca\u53f0<\/text>\n QB \u8d1f\u8f7d\u5747\u8861<\/text>\n \u504f\u7f6e b \u53ea\u8c03\u5206\u8bca\u3001\u4e0d\u6539\u6743\u91cd<\/text>\n \n \n 896 \u4e2a routed \u4e13\u5bb6<\/text>\n \u6bcf\u4e2a token \u53ea\u6fc0\u6d3b\u5176\u4e2d 16 \u4e2a<\/text>\n \n \n \n \n \n \n \n \n <\/g>\n \u4e13\u5bb6\u5728 3584 \u7ef4 latent \u7a7a\u95f4\u5de5\u4f5c<\/text>\n \u901a\u4fe1\u3001\u663e\u5b58\u90fd\u6309”\u7b80\u62a5”\u5927\u5c0f\u7b97<\/text>\n \n \n \u6c47\u603b \u2192 RMSNorm \u2192 W\u2191<\/text>\n \u6295\u5f71\u56de 7168 \u7ef4\uff0c\u4e0e shared \u76f8\u52a0<\/text>\n \n \n 2 \u4e2a shared \u4e13\u5bb6\uff08\u5168\u79d1\uff09<\/text>\n \u5168\u5bbd\u5ea6\u76f4\u63a5\u5904\u7406\u6bcf\u4e2a token<\/text>\n 896 \u9009 16 = \u7a00\u758f\u5ea6 56\uff1a\u53ea\u82b1 1\/56 \u7684\u7b97\u529b\uff0c\u4eab\u53d7 56 \u500d\u7684\u4e13\u5bb6\u5bb9\u91cf<\/text>\n<\/svg><\/p>\n

      MoE \u7684\u6838\u5fc3\u77db\u76fe\uff1a\u4e13\u5bb6\u8d8a\u591a\u77e5\u8bc6\u5bb9\u91cf\u8d8a\u5927\uff0c\u4f46\u6bcf\u4e2a\u88ab\u9009\u4e2d\u7684\u4e13\u5bb6\u90fd\u8981\u6536\u4e00\u4efd 7168 \u7ef4\u7684\u5b8c\u6574 token \u8868\u793a\uff0c\u901a\u4fe1\u548c\u663e\u5b58\u8ddf\u7740\u7206\u70b8\u3002K3 \u7684\u4e09\u8fde\u89e3\u6cd5\uff1a<\/p>\n

        \n
      1. \u5148\u538b\u7f29\u518d\u9001\u8bca<\/strong>\uff1atoken \u5148\u88ab W\u2193 \u538b\u6210 3584 \u7ef4”\u7b80\u62a5”\uff0c\u4e13\u5bb6\u4eec\u5728\u7b80\u62a5\u7a7a\u95f4\u91cc\u5e72\u6d3b\u3002\u4e8e\u662f\u4e13\u5bb6\u6570\u7ffb\u500d\uff08384\u2192896\uff09\u3001\u6fc0\u6d3b\u6570\u7ffb\u500d\uff088\u219216\uff09\uff0c\u901a\u4fe1\u91cf\u5374\u6ca1\u6da8<\/li>\n
      2. SiTU-GLU \u6fc0\u6d3b\u51fd\u6570<\/strong>\uff1aSwiGLU \u7684\u4e24\u4e2a\u76f8\u4e58\u56e0\u5b50\u90fd\u65e0\u754c\uff0c2.8T \u5c3a\u5ea6\u4e0b\u4f1a\u6570\u503c\u7206\u70b8\u3002SiTU-GLU \u7ed9\u4e24\u4e2a\u5206\u652f\u90fd\u88c5\u4e0a tanh”\u9650\u9ad8\u6746”\uff08\u8f93\u51fa\u786c\u4e0a\u9650 100\uff09\uff0c\u539f\u70b9\u9644\u8fd1\u548c SwiGLU \u4e00\u9636\u7b49\u4ef7\u2014\u2014\u5f62\u72b6\u4e0d\u53d8\uff0c\u6c38\u8fdc\u4e0d\u5931\u63a7<\/li>\n
      3. QB \u5206\u4f4d\u6570\u5747\u8861<\/strong>\uff1a\u4f20\u7edf MoE \u9760\u8f85\u52a9\u635f\u5931\u903c\u8d1f\u8f7d\u5747\u8861\uff08\u4f24\u6027\u80fd\uff09\uff0cK3 \u6539\u6210\u7ed9\u6bcf\u4e2a\u4e13\u5bb6\u52a0\u4e00\u4e2a”\u6392\u73ed\u504f\u7f6e b”\u2014\u2014b \u53ea\u5f71\u54cd\u8c01\u88ab\u9009\u4e2d\u3001\u4e0d\u8fdb\u68af\u5ea6\u3001\u4e0d\u6539\u6743\u91cd\u3002\u6bcf\u6b65\u8bad\u7ec3\u540e\u6309\u5206\u4f4d\u6570\u8c03\uff1a\u592a\u706b\u7684\u4e13\u5bb6\u964d\u5206\u3001\u51b7\u95e8\u7684\u52a0\u5206\uff0c\u76f4\u5230\u6bcf\u4e2a\u4e13\u5bb6\u63a5\u5f85\u91cf\u7cbe\u786e\u8fbe\u6807\u3002\u5de5\u7a0b\u4e0a\u7528\u76f4\u65b9\u56fe\u4f30\u5206\u4f4d\u6570\uff0c\u4e00\u6b21 all-reduce \u53ea\u4f20\u51e0\u767e\u4e2a bin\uff0c\u51e0\u4e4e\u514d\u8d39<\/li>\n<\/ol>\n

        \u540e\u8bad\u7ec3\uff1a\u5148\u5206\u5934\u7ec3\u51fa 9 \u4e2a\u4e13\u5bb6\uff0c\u518d\u84b8\u998f\u56de\u4e00\u4e2a\u6a21\u578b<\/h2>\n

        <\/p>\n

        \n \n \n <\/marker>\n <\/defs>\n \n \u2460 SFT \u51b7\u542f\u52a8<\/text>\n \u6253\u57fa\u7840 agent \u80fd\u529b<\/text>\n XTML \u7edf\u4e00\u6a21\u677f<\/text>\n \u5168\u7a0b\u91cf\u5316\u611f\u77e5<\/text>\n \n \n \u2461 RL \u7ec3\u4e13\u9879\u4e13\u5bb6<\/text>\n 3 \u9886\u57df \u00d7 3 \u601d\u8003\u6863<\/text>\n = 9 \u4e2a\u4e13\u5bb6\u6a21\u578b<\/text>\n \u8d85\u9884\u7b97\u5956\u52b1 -1<\/text>\n \n \n \u2462 MOPD \u84b8\u998f<\/text>\n 9 \u4e2a\u8001\u5e08\u6559 1 \u4e2a\u5b66\u751f<\/text>\n \u9010 token \u6253\u5206<\/text>\n \u5408\u6210\u7edf\u4e00\u6a21\u578b<\/text>\n \n \n \u2463 \u4e00\u4e2a K3<\/text>\n low\/high\/max \u4e09\u6863<\/text>\n MXFP4 \u4f4e\u7cbe\u5ea6\u90e8\u7f72<\/text>\n \u8349\u7a3f\u6a21\u578b\u52a0\u901f\u89e3\u7801<\/text>\n \n reasoning effort = \u7ed9\u6a21\u578b\u53d1”\u601d\u8003\u9884\u7b97”<\/text>\n \u5148\u8bad max \u6863\uff08\u9884\u7b97\u5bbd\uff09\uff0c\u518d\u9000\u706b\u51fa high \/ low \u6863 \u2014\u2014 \u8ba9\u6a21\u578b\u5b66\u4f1a\u8be5\u60f3\u60f3\u3001\u8be5\u7701\u7701<\/text>\n<\/svg><\/p>\n

        \u540e\u8bad\u7ec3\u91cc\u51e0\u4e2a\u6709\u610f\u601d\u7684\u7ec6\u8282\uff1a<\/p>\n

          \n
        • \u601d\u8003\u9884\u7b97\u63a7\u5236<\/strong>\uff1a\u6bcf\u9898\u5148\u53d1\u9884\u7b97 b\u2080(x)\uff0c\u8f93\u51fa\u8d85 \u03c4\u00b7b\u2080 \u76f4\u63a5\u5956\u52b1 -1\u3002\u5148\u8bad max \u6863\uff0c\u518d\u9000\u706b\u51fa high\/low \u6863\u2014\u2014\u6a21\u578b\u81ea\u5df1\u5b66\u4f1a”\u7b80\u5355\u9898\u522b\u8fc7\u5ea6\u601d\u8003”<\/li>\n
        • \u9632\u8bdd\u75e8<\/strong>\uff1aGRM \u6253\u5206\u65f6\u56de\u7b54\u957f\u5ea6\u8d85 \u03c3\u00b7\u2113\u2080 \u76f4\u63a5\u5224\u8d1f\uff0c\u4e13\u6cbb reward hacking \u7684\u5570\u55e6\u503e\u5411<\/li>\n
        • \u8bad\u7ec3\u4efb\u52a1\u591f\u91ce<\/strong>\uff1a\u77e5\u8bc6\u56fe\u8c31\u81ea\u52a8\u6269\u5c55\u51fa\u9898\u3001mock \u7248 Gmail\/Notion\/Slack \u7ec3”\u8fde\u7eed\u5de5\u4f5c\u597d\u51e0\u5929”\u7684\u52a9\u7406\u4efb\u52a1\u3001GPU kernel \u4f18\u5316\uff08\u5956\u52b1=\u6b63\u786e\u6027+\u903c\u8fd1 roofline \u7a0b\u5ea6\uff0c\u8fd8\u914d\u53cd\u4f5c\u5f0a\u7cfb\u7edf\uff09\u3001\u9ed1\u76d2\u7cfb\u7edf\u590d\u523b\u7b49<\/li>\n<\/ul>\n

          \u57fa\u5efa\uff1a2.8T \u4e0d\u662f\u5806\u5361\u5c31\u884c\u7684<\/h2>\n\n\n\n\n\n\n\n\n
          \u5de5\u7a0b\u96be\u9898<\/th>\nK3 \u7684\u89e3\u6cd5<\/th>\n<\/tr>\n
          KDA \u9012\u5f52\u672c\u8d28\u662f\u4e32\u884c\uff0cGPU \u4e0d\u559c\u6b22<\/td>\nFlashKDA kernel\uff1achunk \u5185\u5e76\u884c\u3001chunk \u95f4\u6d41\u6c34\u91cd\u53e0<\/td>\n<\/tr>\n
          1M \u4e0a\u4e0b\u6587\u5e76\u884c\u8bad\u7ec3<\/td>\nKCP\uff1a\u628a\u6bcf\u6bb5\u7684\u72b6\u6001\u8f6c\u79fb\u5206\u89e3\u6210\u53ef\u672c\u5730\u8ba1\u7b97\u7684\u4e24\u90e8\u5206\uff0c\u4e00\u6b21 all-gather \u641e\u5b9a\u2014\u2014\u7ebf\u6027 attention \u7684\u901a\u4fe1\u91cf\u4e0e\u5e8f\u5217\u957f\u5ea6\u65e0\u5173<\/strong><\/td>\n<\/tr>\n
          \u4e13\u5bb6\u5e76\u884c\u8d1f\u8f7d\u4e0d\u5747\u3001\u663e\u5b58\u788e\u7247<\/td>\nMoonEP\uff1a\u6bcf\u4e2a rank \u7cbe\u786e\u6536 S\u00d7K \u4e2a token\uff0c\u5197\u4f59\u4e13\u5bb6 \u2264 E\/R\uff08\u6709\u6570\u5b66\u8bc1\u660e\uff09\uff0cshape \u5168\u9759\u6001\u514d\u540c\u6b65<\/td>\n<\/tr>\n
          1M agentic RL \u7684 KV cache \u5b58\u4e0d\u4e0b<\/td>\n\u5916\u90e8 KV \u6c60\uff1aGPU \u9010\u51fa\u7684\u524d\u7f00\u5199\u56de CPU DRAM\uff0cKDA state \u4e0e MLA KV \u751f\u547d\u5468\u671f\u5bf9\u9f50<\/td>\n<\/tr>\n
          agent \u628a\u5bb9\u5668\u6c99\u7bb1\u641e\u5230\u5185\u6838 panic<\/td>\nAgentENV\uff1aFirecracker microVM\uff0c\u652f\u6301\u6682\u505c\/\u5206\u53c9\/\u5feb\u7167\uff08checkpoint 133ms\uff09\uff1b\u6574\u4e2a\u8bad\u7ec3\u5171\u521b\u5efa 5122 \u4e07\u4e2a\u6c99\u7bb1\u3001150 \u4e07\u4e2a\u955c\u50cf<\/strong><\/td>\n<\/tr>\n
          1M token \u957f\u8bf7\u6c42\u6324\u6b7b 2K \u77ed\u8bf7\u6c42<\/td>\n\u9884\u7b97\u5236\u51c6\u5165\u63a7\u5236\uff1a\u957f\u77ed\u8bf7\u6c42\u5404\u6709\u72ec\u7acb\u8d44\u6e90\u6c60\uff0c\u4e92\u4e0d\u997f\u6b7b<\/td>\n<\/tr>\n<\/table>\n

          \u6210\u7ee9\u5355\uff08\u62a5\u544a\u5b98\u65b9\u6570\u636e\uff0c\u622a\u81f3 2026-07-23\uff09<\/h2>\n\n\n\n\n\n\n\n
          \u7ef4\u5ea6<\/th>\n\u8868\u73b0<\/th>\n<\/tr>\n
          \u7f16\u7a0b<\/td>\nSWE-Marathon \u7b2c\u4e00\uff0842.0%\uff0c\u9886\u5148 Fable 5 \u4e03\u5206\uff09\uff1bTerminal-Bench 88.3% \u51e0\u4e4e\u5e73 GPT-5.6 Sol\uff1bWebDev Arena \u603b\u699c\u7b2c\u4e00<\/strong>\uff081678 Elo\uff0c\u9996\u4e2a\u767b\u9876\u7684\u5f00\u6e90\u6a21\u578b\uff09<\/td>\n<\/tr>\n
          Agent<\/td>\nBrowseComp 91.2% \u5168\u573a\u7b2c\u4e00\uff1bMCPMark 94.5% \u7b2c\u4e00\uff1bGDPval-AA \u7b2c\u4e09\uff08\u524d\u4e8c\u4e3a Fable 5\u3001GPT-5.6 Sol\uff09<\/td>\n<\/tr>\n
          \u77e5\u8bc6\u63a8\u7406<\/td>\nGPQA 93.5%\uff1b\u4f46 HLE\u3001CritPt \u7b49\u7814\u7a76\u7ea7\u4efb\u52a1\u4ecd\u843d\u540e\u524d\u4e24\u540d\u2014\u2014\u62a5\u544a\u81ea\u5df1\u627f\u8ba4\u8fd9\u662f\u77ed\u677f<\/td>\n<\/tr>\n
          \u89c6\u89c9<\/td>\nOmniDocBench 91.1% \u7b2c\u4e00\uff1bZeroBench \u5e73 Fable 5<\/td>\n<\/tr>\n
          \u6210\u672c<\/td>\nBrowseComp \u6bcf\u9898 $2.03\uff0c\u662f GPT-5.6 Sol \u7684\u4e00\u534a\u3001Claude \u7cfb\u7684\u7ea6 1\/10\uff1bKCB 2.0 \u7528 Fable 5 \u7ea6 38% \u7684\u6210\u672c\u62ff\u5230\u53ea\u5dee 4 \u5206\u7684\u6210\u7ee9<\/td>\n<\/tr>\n<\/table>\n

          Case study \u91cc\u6700\u5938\u5f20\u7684\u4e09\u4e2a\uff1a24 \u5c0f\u65f6\u5185\u628a 4 \u4e2a GPU kernel \u4f18\u5316\u5230\u5339\u914d Fable 5 \u7684\u6c34\u5e73\uff1b\u72ec\u7acb\u5199\u51fa\u4e00\u4e2a\u80fd\u7aef\u5230\u7aef\u8bad GPT \u7684 Triton \u7c7b\u7f16\u8bd1\u5668 MiniTriton\uff08matmul \u8fbe cuBLAS \u7ea6 90%\uff09\uff1b48 \u5c0f\u65f6\u7528\u5f00\u6e90 EDA \u94fe\u8bbe\u8ba1\u4e86\u9897 INT4 \u63a8\u7406\u82af\u7247\uff08RTL \u4eff\u771f 8700 tokens\/s\uff09\u3002<\/p>\n

          \u666e\u901a\u5f00\u53d1\u8005\u80fd\u5e26\u8d70\u4ec0\u4e48<\/h2>\n
            \n
          1. “\u6df7\u5408”\u6253\u8d25”\u7eaf\u7cb9”<\/strong>\uff1a75% \u4fbf\u5b9c\u5c42 + 25% \u8d35\u5c42 \u2248 \u5168\u8d35\u5c42\u3002\u7f13\u5b58\/\u6570\u636e\u5e93\/\u7d22\u5f15\u7684\u5206\u5c42\u8bbe\u8ba1\u662f\u540c\u4e00\u4e2a\u9053\u7406<\/li>\n
          2. \u6570\u503c\u7a33\u5b9a\u6027\u5c31\u662f\u6027\u80fd<\/strong>\uff1a\u4e00\u4e2a sigmoid \u4e0b\u9650\u6362\u6765\u6574\u6761 kernel \u8def\u5f84\u4e0a Tensor Core\uff0c\u8fb9\u754c\u6761\u4ef6\u522b\u5c0f\u770b<\/li>\n
          3. \u8d1f\u8f7d\u5747\u8861\u53ef\u4ee5”\u8c03\u504f\u7f6e”\u800c\u4e0d\u662f”\u52a0\u60e9\u7f5a”<\/strong>\uff1a\u4e0d\u6539\u76ee\u6807\u51fd\u6570\u3001\u53ea\u8c03\u51c6\u5165\u95e8\u69db\u2014\u2014\u4efb\u4f55\u8c03\u5ea6\u7cfb\u7edf\u90fd\u597d\u4f7f<\/li>\n
          4. agent \u80fd\u529b \u2248 \u73af\u5883\u590d\u6742\u5ea6<\/strong>\uff1aK3 \u7684 agent \u80fd\u529b\u662f 5122 \u4e07\u4e2a\u6c99\u7bb1\u5582\u51fa\u6765\u7684\u3002\u4f60\u7684 agent \u4ea7\u54c1\u80fd\u7ec3\u591a\u72e0\uff0c\u53d6\u51b3\u4e8e\u4f60\u80fd\u9020\u591a\u771f\u7684\u6f14\u7ec3\u573a<\/li>\n
          5. \u6210\u672c\u62d0\u70b9<\/strong>\uff1a\u5f00\u6e90\u6743\u91cd\u7684 K3 \u628a\u524d\u6cbf agent \u80fd\u529b\u6253\u5230\u95ed\u6e90\u65d7\u8230 1\/3~1\/10 \u7684\u4ef7\u683c\uff0c\u81ea\u90e8\u7f72\u7684\u7ecf\u6d4e\u8d26\u8d8a\u6765\u8d8a\u597d\u7b97\u4e86<\/li>\n<\/ol>\n

            \u4e00\u53e5\u8bdd\u603b\u7ed3\uff1aK3 \u6ca1\u6709\u53d1\u660e\u65b0\u79d1\u5b66\uff0c\u4f46\u5b83\u628a\u7ebf\u6027\u6ce8\u610f\u529b\u3001MoE\u3001RL\u3001\u7cfb\u7edf\u5de5\u7a0b\u56db\u6761\u7ebf\u7684\u5de5\u7a0b\u7ec4\u5408\u505a\u5230\u4e86\u5f00\u6e90\u4e16\u754c\u7684\u5929\u82b1\u677f<\/strong>\u2014\u2014\u7136\u540e\u7528 2.5\u00d7 \u7684\u8bad\u7ec3\u6548\u7387\u548c 1\/10 \u7684\u63a8\u7406\u6210\u672c\u544a\u8bc9\u4f60\uff1a\u524d\u6cbf\u667a\u80fd\u7684\u5165\u573a\u5238\u6b63\u5728\u53d8\u4fbf\u5b9c\u3002<\/p>\n

            \u672c\u6587\u4e3a Moonshot\u300aKimi K3: Open Frontier Intelligence\u300b\u6280\u672f\u62a5\u544a\uff082026 \u5e74 7 \u6708\uff0c47 \u9875\uff09\u7684\u4e2a\u4eba\u89e3\u8bfb\uff0c\u6570\u636e\u5747\u5f15\u81ea\u62a5\u544a\u539f\u6587\uff0c\u8dd1\u5206\u4ee5\u5b98\u65b9\u53e3\u5f84\u4e3a\u51c6\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

            30 \u79d2\u7248\u672c\uff1aKimi K3 \u662f Moonshot \u6700\u65b0\u5f00\u6e90\u7684\u65d7\u8230\u6a21\u578b\u2014\u20142.8 \u4e07\u4ebf\u53c2\u6570\u7684 MoE\uff0c\u6bcf\u4e2a t […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[12],"tags":[13,15,14,16],"class_list":["post-688","post","type-post","status-publish","format-standard","hentry","category-ji-shu","tag-AI","tag-Kimi","tag-da-mo-xing","tag-ji-shu-jie-du"],"_links":{"self":[{"href":"https:\/\/aitimes.link\/index.php\/wp-json\/wp\/v2\/posts\/688","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/aitimes.link\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/aitimes.link\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/aitimes.link\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/aitimes.link\/index.php\/wp-json\/wp\/v2\/comments?post=688"}],"version-history":[{"count":1,"href":"https:\/\/aitimes.link\/index.php\/wp-json\/wp\/v2\/posts\/688\/revisions"}],"predecessor-version":[{"id":689,"href":"https:\/\/aitimes.link\/index.php\/wp-json\/wp\/v2\/posts\/688\/revisions\/689"}],"wp:attachment":[{"href":"https:\/\/aitimes.link\/index.php\/wp-json\/wp\/v2\/media?parent=688"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/aitimes.link\/index.php\/wp-json\/wp\/v2\/categories?post=688"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/aitimes.link\/index.php\/wp-json\/wp\/v2\/tags?post=688"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}