{"id":769083,"date":"2026-04-25T14:31:01","date_gmt":"2026-04-25T06:31:01","guid":{"rendered":"https:\/\/www.alphaseek.cn\/?p=769083"},"modified":"2026-04-25T14:31:01","modified_gmt":"2026-04-25T06:31:01","slug":"deepseek-v4%e6%8a%a5%e5%91%8a%e4%ba%ae%e4%ba%86%ef%bc%81v4%e5%8f%91%e5%b8%83%e5%bb%b6%e8%bf%9f%e7%9a%84%e7%a7%98%e5%af%86%ef%bc%8c%e7%bb%88%e4%ba%8e%e6%9b%9d%e5%85%89%e4%ba%86","status":"publish","type":"post","link":"https:\/\/www.alphaseek.cn\/?p=769083","title":{"rendered":"DeepSeek-V4\u62a5\u544a\u4eae\u4e86\uff01V4\u53d1\u5e03\u5ef6\u8fdf\u7684\u79d8\u5bc6\uff0c\u7ec8\u4e8e\u66dd\u5149\u4e86"},"content":{"rendered":"

\"DeepSeek-V4\u62a5\u544a\u4eae\u4e86\uff01V4\u53d1\u5e03\u5ef6\u8fdf\u7684\u79d8\u5bc6\uff0c\u7ec8\u4e8e\u66dd\u5149\u4e86\"<\/p>\n

DeepSeek-V4\u7684\u6280\u672f\u62a5\u544a\uff0c\u7b80\u76f4\u8bda\u5b9e\u5f97\u4ee4\u4eba\u9707\u60ca\u3002V4\u53d1\u5e03\u5ef6\u8fdf\u7684\u79d8\u5bc6\uff0c\u88ab\u6b63\u5f0f\u900f\u9732\u4e86\uff01\u8fd9\u9897\u5927\u96f7\u7684\u80cc\u540e\uff0c\u7a76\u7adf\u662f\u6307\u8c01\uff1f\u7814\u7a76\u8005\u4eec\u5df2\u7ecf\u7eb7\u7eb7\u5c55\u5f00\u4e86\u731c\u6d4b\u3002\u5e76\u4e14\uff0c\u8bba\u6587\u4e2d\u7528\u786c\u6838\u5de5\u7a0b\u66b4\u529b\u91cd\u6784Agent\u7684\u64cd\u4f5c\uff0c\u4e5f\u8ba9\u793e\u533a\u76f4\u547c\uff1a\u56fd\u4ea7\u4e4b\u5149\uff0c\u5b9e\u81f3\u540d\u5f52\u3002<\/strong><\/p>\n

\u6628\u5929\uff0c\u662f\u540d\u526f\u5176\u5b9e\u7684AI\u5708\u300c\u6625\u665a\u300d\u3002<\/p>\n

DeepSeek-V4\u7684\u6280\u672f\u62a5\u544a\u4e00\u51fa\uff0c\u8fd160\u9875\u7684\u7bc7\u5e45\uff0c\u4ece\u67b6\u6784\u5230\u8bad\u7ec3\u5230\u540e\u8bad\u7ec3\u5168\u90e8\u644a\u5f00\u3002<\/p>\n

484\u5929\uff0c\u5bf9\u8fd9\u4e2a\u56e2\u961f\u6765\u8bf4\u4e0d\u5bfb\u5e38\u3002V3\u4eceV2\u5230\u53d1\u5e03\u53ea\u7528\u4e86\u4e0d\u52308\u4e2a\u6708\u3002V4\u4e3a\u4ec0\u4e48\u591a\u82b1\u4e86\u5c06\u8fd1\u4e00\u500d\u7684\u65f6\u95f4\uff1f<\/p>\n

\u8ba4\u771f\u7814\u8bfb\u5b8c\u8fd9\u7bc7\u62a5\u544a\uff0c\u6211\u4eec\u53d1\u73b0\u4e86\u80cc\u540e\u53ef\u80fd\u7684\u539f\u56e0\uff0c\u4ee5\u53ca\u8fd9\u5bb6\u300c\u56fd\u4ea7\u4e4b\u5149\u300d\u4ee4\u4eba\u9707\u64bc\u7684\u5de5\u7a0b\u5e95\u8272\u5b9e\u3002<\/p>\n

\u53ef\u4ee5\u8bf4\uff0cDeepSeek-V4\u771f\u6b63\u4ee4\u4eba\u6df1\u601d\u7684\uff0c\u4e0d\u662f\u5b83\u5806\u4e86\u591a\u5c11\u7b97\u529b\uff0c\u800c\u662f\u5b83\u5728Agent\u8bad\u7ec3\u3001\u5de5\u7a0b\u5e95\u5ea7\u3001\u4ee5\u53ca\u5904\u7406\u300c\u8bad\u7ec3\u9707\u8361\u300d\u65f6\u7684\u90a3\u79cd\u8fd1\u4e4e\u6b8b\u9177\u7684\u7406\u6027\u548c\u900f\u660e\u3002<\/p>\n

\u4eca\u5929\uff0c\u6211\u4eec\u76f4\u63a5\u62c6\u5f00V4\u7684\u5f15\u64ce\u76d6\uff0c\u770b\u770b\u91cc\u9762\u85cf\u7740\u54ea\u4e9b\u4e0d\u4e3a\u4eba\u77e5\u7684\u786c\u6838\u7ec6\u8282\u3002<\/p>\n

33T Token + \u4e07\u4ebf\u53c2\u6570<\/strong><\/p>\n

\u96be\u5ea6\u76f4\u63a5\u62c9\u6ee1<\/strong><\/p>\n

\u8ddd\u79bbV3\u53d1\u5e03\u6574\u6574484\u5929\uff0cV4\u624d\u4ee5\u300cpreview version\u300d\u7684\u59ff\u6001\u4e0a\u7ebf\u3002<\/p>\n

\u8bba\u6587\u91cc\u867d\u7136\u6ca1\u6709\u89e3\u91ca\u8fd9\u4e2a\u65f6\u95f4\u8de8\u5ea6\uff0c\u4f46\u6709\u4e00\u6bb5\u5185\u5bb9\u6216\u8bb8\u80fd\u63d0\u4f9b\u7ebf\u7d22\u3002<\/p>\n

\"DeepSeek-V4\u62a5\u544a\u4eae\u4e86\uff01V4\u53d1\u5e03\u5ef6\u8fdf\u7684\u79d8\u5bc6\uff0c\u7ec8\u4e8e\u66dd\u5149\u4e86\"<\/p>\n

V3\u7528\u4e8614.8T token\u505a\u9884\u8bad\u7ec3\uff0cV4\u76f4\u63a5\u7ffb\u500d\uff0cV4-Flash\u8bad\u4e8632T\uff0cV4-Pro\u8bad\u4e8633T\u3002\u53c2\u6570\u91cf\u540c\u6837\u5927\u5e45\u6269\u5f20\uff0cV4-Pro\u603b\u53c2\u65701.6T\uff0cV4-Flash\u4e5f\u6709284B\u3002<\/p>\n

\u6570\u636e\u7ffb\u500d\u3001\u53c2\u6570\u7ffb\u500d\uff0c\u8bad\u7ec3\u7a33\u5b9a\u6027\u7684\u96be\u5ea6\u4e5f\u8ddf\u7740\u4e0a\u4e86\u4e00\u4e2a\u91cf\u7ea7\u3002<\/p>\n

\u62a5\u544a\u91cc\u975e\u5e38\u8bda\u5b9e\uff1aDeepSeek\u660e\u786e\u70b9\u540d\u4e86\u300c\u8bad\u7ec3\u7a33\u5b9a\u6027\u6311\u6218\u300d\u3002<\/p>\n

\"DeepSeek-V4\u62a5\u544a\u4eae\u4e86\uff01V4\u53d1\u5e03\u5ef6\u8fdf\u7684\u79d8\u5bc6\uff0c\u7ec8\u4e8e\u66dd\u5149\u4e86\"<\/p>\n

\u8c37\u6b4cDeepMind\u7814\u7a76\u8005Susan Zhang\u8868\u626c\u8bf4\uff1a\u8fd9\u79cd\u900f\u660e\u7684\u505a\u6cd5\u503c\u5f97\u79f0\u8d5e\u3002\u8fd9\u4e2a\u8bf4\u6cd5\u8fd8\u5f97\u5230\u4e86\u9f99\u867e\u4e4b\u7236\u7684\u8f6c\u53d1<\/p>\n

\u5728\u8d85\u5927\u89c4\u6a21\u96c6\u7fa4\u4e0a\uff0c\u5f53\u53c2\u6570\u91cf\u548c\u8bad\u7ec3\u6570\u636e\u8fbe\u5230\u67d0\u4e2a\u4e34\u754c\u70b9\u65f6\uff0c\u786c\u4ef6\u7684\u7ec6\u5fae\u8bef\u5dee\u4f1a\u88ab\u65e0\u9650\u653e\u5927\u3002<\/p>\n

\u8bba\u6587\u91cc\uff0c\u300cstability\u300d\u8fd9\u4e2a\u8bcd\u51fa\u73b0\u4e86\u5341\u4f59\u6b21\u3002<\/p>\n

\u653e\u5728\u4e00\u7bc7\u6280\u672f\u62a5\u544a\u91cc\uff0c\u8fd9\u4e2a\u9891\u7387\u672c\u8eab\u5c31\u662f\u4fe1\u53f7\u3002\u6b63\u5e38\u60c5\u51b5\u4e0b\uff0c\u7a33\u5b9a\u6027\u662f\u9ed8\u8ba4\u524d\u63d0\uff0c\u4e0d\u503c\u5f97\u53cd\u590d\u63d0\u3002\u53cd\u590d\u63d0\uff0c\u8bf4\u660e\u5b83\u786e\u5b9e\u662f\u4e2a\u95ee\u9898\u3002<\/p>\n

\"DeepSeek-V4\u62a5\u544a\u4eae\u4e86\uff01V4\u53d1\u5e03\u5ef6\u8fdf\u7684\u79d8\u5bc6\uff0c\u7ec8\u4e8e\u66dd\u5149\u4e86\"<\/p>\n

\u5177\u4f53\u6765\u770b\uff0cDeepSeek\u53d1\u73b0MoE\u5c42\u4e2d\u7684\u6570\u503c\u5f02\u5e38\u503c\uff08outlier\uff09\u4f1a\u901a\u8fc7\u8def\u7531\u673a\u5236\u4e0d\u65ad\u653e\u5927\uff0c\u5f62\u6210\u6076\u6027\u5faa\u73af\uff0c\u6700\u7ec8\u89e6\u53d1loss spike\uff0c\u8bad\u7ec3\u66f2\u7ebf\u7a81\u7136\u98d9\u5347\u3002<\/p>\n

\u56e2\u961f\u796d\u51fa\u7684\u4e3b\u8981\u8865\u6551\u63aa\u65bd\u662f\u4e24\u62db\u3002<\/p>\n

\u7b2c\u4e00\u62db\u53ebAnticipatory Routing<\/strong>\u3002\u5b83\u672c\u8d28\u4e0a\u5c31\u662f\u5728\u8def\u7531\u9636\u6bb5\u4f7f\u7528\u7a0d\u65e9\u7248\u672c\u7684\u53c2\u6570\uff0c\u628a\u9aa8\u5e72\u7f51\u7edc\u548c\u8def\u7531\u7f51\u7edc\u7684\u66f4\u65b0\u89e3\u8026\uff0c\u6253\u7834\u4e24\u8005\u4e4b\u95f4\u7684\u6076\u6027\u5faa\u73af\u3002<\/p>\n

\u7b2c\u4e8c\u62db\u662fSwiGLU Clamping<\/strong>\u3002\u5b83\u76f4\u63a5\u628aSwiGLU\u7684\u6570\u503c\u8303\u56f4\u94b3\u5236\u5728[-10, 10]\u4ee5\u5185\uff0c\u4ece\u6e90\u5934\u538b\u5236\u5f02\u5e38\u503c\uff0c\u867d\u7136\u66b4\u529b\u4f46\u5f88\u6709\u6548\u3002<\/p>\n

\"DeepSeek-V4\u62a5\u544a\u4eae\u4e86\uff01V4\u53d1\u5e03\u5ef6\u8fdf\u7684\u79d8\u5bc6\uff0c\u7ec8\u4e8e\u66dd\u5149\u4e86\"<\/p>\n

\u5f53\u524d\u5927\u6a21\u578b\u8bad\u7ec3\u5df2\u8fdb\u5165\u786c\u4ef6\u5e95\u5c42\u3001\u7f16\u8bd1\u5668\u6808\u3001\u4ee5\u53ca\u6570\u5b66\u67b6\u6784<\/strong>\u4e09\u4f4d\u4e00\u4f53\u7684\u65e0\u4eba\u533a<\/p>\n

\u8bba\u6587\u91cc\u6709\u4e2a\u7ec6\u8282\u5f88\u8010\u7422\u78e8\u3002<\/p>\n

Anticipatory Routing\u548cSwiGLU Clamping\uff0cDeepSeek\u786e\u8ba4\u300c\u663e\u8457\u6709\u6548\u300d\uff0c\u4f46\u7d27\u8ddf\u4e00\u53e5\u300c\u5e95\u5c42\u673a\u7406\u4ecd\u662fopen question\u300d\u3002<\/p>\n

\u8fdeQ\/KV\u5f52\u4e00\u5316\u8fd9\u79cd\u5df2\u7ecf\u88ab\u5e7f\u6cdb\u9a8c\u8bc1\u7684\u57fa\u7840\u64cd\u4f5c\uff0c\u8bba\u6587\u7684\u63aa\u8f9e\u90fd\u53ea\u6562\u5199\u300cmay improve training stability\u300d\u3002<\/p>\n

\u4e00\u4e2a\u300cmay\u300d\u5b57\uff0c\u8db3\u4ee5\u8bf4\u660e\u5728\u4e07\u4ebf\u53c2\u6570MoE\u7684\u8bad\u7ec3\u91cc\uff0c\u6ca1\u6709\u4ec0\u4e48\u662f\u767e\u5206\u767e\u9760\u5f97\u4f4f\u7684\u3002<\/p>\n

\"DeepSeek-V4\u62a5\u544a\u4eae\u4e86\uff01V4\u53d1\u5e03\u5ef6\u8fdf\u7684\u79d8\u5bc6\uff0c\u7ec8\u4e8e\u66dd\u5149\u4e86\"<\/p>\n

\u4ece15T\u523033T\uff0c\u6570\u636e\u91cf\u7ffb\u500d\u5e26\u6765\u7684\u4e0d\u662f\u7ebf\u6027\u589e\u957f\u7684\u56f0\u96be\uff0c\u800c\u662f\u6307\u6570\u7ea7\u653e\u5927\u7684\u7cfb\u7edf\u6027\u98ce\u9669<\/a>\u3002<\/p>\n

\u6bcf\u4e00\u5c42\u7f51\u7edc\u3001\u6bcf\u4e00\u4e2a\u68af\u5ea6\u66f4\u65b0\u3001\u6bcf\u4e00\u6b21\u901a\u4fe1\u540c\u6b65\uff0c\u90fd\u5728\u66f4\u5927\u7684\u89c4\u6a21\u4e0b\u88ab\u653e\u5927\u6210\u6f5c\u5728\u7684\u5d29\u6e83\u70b9\u3002<\/p>\n

\u800cDeepSeek\u9009\u62e9\u628a\u8fd9\u4e9b\u5168\u5199\u8fdb\u8bba\u6587\u91cc\uff0c\u8fd9\u5728\u4e1a\u5185\u51e0\u4e4e\u6ca1\u6709\u5148\u4f8b\u3002<\/p>\n

\u786c\u4ef6\u7684\u9505\uff0c\u8fd8\u662f\u8f6f\u4ef6\u7684\u9505\uff1f<\/strong><\/p>\n

\u6240\u4ee5\uff0c\u6280\u672f\u62a5\u544a\u4e2d\u660e\u786e\u63d0\u51fa\u7684\u300c\u8bad\u7ec3\u7a33\u5b9a\u6027\u6311\u6218\u300d\uff0c\u6307\u7684\u5230\u5e95\u662f\u8c01\u5bb6\u7684\u786c\u4ef6\uff1f<\/p>\n

\u867d\u7136\u8bba\u6587\u91cc\u6ca1\u6709\u660e\u786e\u70b9\u540d\u4efb\u4f55\u786c\u4ef6\u5e73\u53f0\uff0c\u4f46\u5df2\u7ecf\u6709\u55c5\u89c9\u654f\u9510\u7684\u4eba\u5f00\u59cb\u731c\u6d4b\u4e86\u3002<\/p>\n

\u6709\u89c2\u70b9\u76f4\u63a5\u6307\u51fa\uff1a\u6240\u8c13\u300c\u8bad\u7ec3\u7a33\u5b9a\u6027\u6311\u6218\u300d\uff0c\u5f88\u53ef\u80fd\u5c31\u662f\u7b97\u529b\u5e73\u53f0\u7684\u95ee\u9898\u3002\u800c\u4e14\u4e0d\u53ea\u662fDeepSeek\u4e00\u5bb6\u8e29\u5751\uff0c\u5404\u5927\u5382\u5546\u90fd\u9047\u5230\u8fc7\u3002<\/p>\n

xAI\u5728\u4e00\u6b21\u53d1\u5e03\u4f1a\u4e0a\uff0cMacrohard\u9879\u76ee\u7684\u8d1f\u8d23\u4eba\u66fe\u9690\u6666\u63d0\u5230\uff0c\u82f1\u4f1f\u8fbe\u6700\u65b0\u7684\u82af\u7247<\/a>\u7ed9\u4ed6\u4eec\u9020\u6210\u4e86\u300c\u4e0d\u5c0f\u7684\u9ebb\u70e6\u300d\uff0c\u4e0d\u5f97\u4e0d\u91cd\u65b0\u5f00\u53d1\u786c\u4ef6\u9002\u914d\u7a0b\u5e8f\u3002\u8fd9\u6216\u8bb8\u4e5f\u89e3\u91ca\u4e86xAI\u8fdb\u5ea6\u7a81\u7136\u653e\u7f13\u7684\u539f\u56e0\u4e4b\u4e00\u3002<\/p>\n

\"DeepSeek-V4\u62a5\u544a\u4eae\u4e86\uff01V4\u53d1\u5e03\u5ef6\u8fdf\u7684\u79d8\u5bc6\uff0c\u7ec8\u4e8e\u66dd\u5149\u4e86\"<\/p>\n

\u4e0d\u8fc7\uff0c\u8fd9\u4ef6\u4e8b\u5f53\u7136\u6ca1\u90a3\u4e48\u7b80\u5355\u3002<\/p>\n

\u5927\u578b\u7b97\u529b\u96c6\u7fa4\u6d89\u53ca\u7684\u53d8\u91cf\u592a\u591a\uff1a\u82af\u7247\u672c\u8eab\u3001\u4e92\u8fde\u67b6\u6784\u3001\u6563\u70ed\u7cfb\u7edf\u3001\u7535\u529b\u4f9b\u5e94\u3001\u9a71\u52a8\u7248\u672c\u3001\u7f16\u8bd1\u6808\u9002\u914d\u3002\u8bad\u7ec3\u4e0d\u7a33\u5b9a\u672a\u5fc5\u7b49\u4e8e\u82af\u7247\u7ea7\u7f3a\u9677\uff0c\u4e5f\u53ef\u80fd\u662f\u7cfb\u7edf\u96c6\u6210\u5c42\u7684\u95ee\u9898\u3002<\/p>\n

\u4e0d\u8fc7\uff0c\u76ee\u524d\u8fd8\u6ca1\u6709\u4efb\u4f55\u5b98\u65b9\u6587\u4ef6\u7ed9\u51fa\u7b54\u6848\u3002<\/p>\n

\u4e00\u5207\u90fd\u8fd8\u5728\u731c\u6d4b\u4e4b\u4e2d\u3002<\/p>\n

\"DeepSeek-V4\u62a5\u544a\u4eae\u4e86\uff01V4\u53d1\u5e03\u5ef6\u8fdf\u7684\u79d8\u5bc6\uff0c\u7ec8\u4e8e\u66dd\u5149\u4e86\"<\/p>\n

Agent\u8bad\u7ec3\u4f53\u7cfb<\/strong><\/p>\n

\u5de5\u7a0b\u80fd\u529b\u8ba9\u4eba\u8083\u7136\u8d77\u656c<\/strong><\/p>\n

\u5982\u679c\u8bf4V4\u7684\u9884\u8bad\u7ec3\u662f\u5728\u548c\u786c\u4ef6\u535a\u5f08\uff0c\u90a3\u4e48\u5b83\u7684Post-training<\/strong>\u5219\u5c55\u73b0\u4e86\u6559\u79d1\u4e66\u7ea7\u522b\u7684\u5de5\u7a0b\u5ba1\u7f8e\u3002<\/p>\n

\u53ef\u4ee5\u8bf4\uff0cAgent\u80fd\u529b\u7684\u5de5\u7a0b\u5316\u8def\u5f84\uff0c\u662fV4\u8bba\u6587\u91cc\u6700\u503c\u5f97\u7ec6\u8bfb\u7684\u90e8\u5206\u3002<\/p>\n

\u4ee5\u5f80\u6211\u4eec\u8ba4\u4e3aAgent\u80fd\u529b\u662f\u300c\u6559\u300d\u51fa\u6765\u7684\uff0c\u4f46DeepSeek\u8ba4\u4e3a\uff0cAgent\u80fd\u529b\u5e94\u8be5\u662f\u300c\u957f\u300d\u51fa\u6765\u7684\u3002<\/p>\n

\"DeepSeek-V4\u62a5\u544a\u4eae\u4e86\uff01V4\u53d1\u5e03\u5ef6\u8fdf\u7684\u79d8\u5bc6\uff0c\u7ec8\u4e8e\u66dd\u5149\u4e86\"<\/p>\n

\u62d2\u7edd\u300c\u786c\u8fc1\u79fb\u300d\uff0c\u9884\u8bad\u7ec3\u9636\u6bb5\u7684\u300c\u8840\u8109\u6ce8\u5165\u300d<\/strong><\/p>\n

\u884c\u4e1a\u5185\u5927\u90e8\u5206\u7684\u505a\u6cd5\u662f\uff0c\u5148\u8bad\u4e00\u4e2a\u5bf9\u8bdd\u6a21\u578b\uff0c\u518d\u786c\u8fc1\u79fb\u6210Agent\u3002DeepSeek\u770b\u6765\uff0c\u8fd9\u592a\u4f4e\u6548\u4e86\u3002<\/p>\n

\u5728V4\u7684mid-training\u9636\u6bb5\uff0c\u4ed6\u4eec\u5c31\u6ce8\u5165\u4e86\u6d77\u91cf\u7684Agentic Data\u3002<\/p>\n

\u8fd9\u610f\u5473\u7740\uff0c\u6a21\u578b\u5728\u57fa\u7840\u5b66\u4e60\u9636\u6bb5\uff0c\u5c31\u5df2\u7ecf\u89c1\u8fc7\u957f\u4efb\u52a1\u94fe\u3001\u73af\u5883\u53cd\u9988\u548c\u6587\u4ef6\u4fee\u6539\u6a21\u5f0f\u3002\u5b83\u8fd8\u6ca1\u5b66\u4f1a\u5199\u8bd7\uff0c\u5c31\u5df2\u7ecf\u89c1\u8fc7\u4e86Linux\u547d\u4ee4\u884c\u7684\u62a5\u9519\u3002<\/p>\n

\u8fd9\u5c31\u662f\u4e00\u79cd\u5730\u57fa\u5c42\u9762\u7684\u8bbe\u8ba1\u3002<\/p>\n

\u72ec\u521b\u7684Specialist Training\uff08\u4e13\u5bb6\u7279\u8bad\u6cd5\uff09<\/strong><\/p>\n

\u53e6\u4e00\u5927\u4eae\u70b9\uff0c\u5c31\u662fDeepSeek\u72ec\u521b\u7684\u4e13\u5bb6\u7279\u8bad\u6cd5\u3002<\/p>\n

V4\u6ca1\u6709\u76f4\u63a5\u7ec3\u4e00\u4e2a\u5168\u80fd\u6218\u58eb\uff0c\u800c\u662f\u5148\u7ec3\u51fa\u4e86\u6570\u5b66\u4e13\u5bb6\u3001\u4ee3\u7801\u4e13\u5bb6\u3001Agent\u4e13\u5bb6\u3001\u6307\u4ee4\u8ddf\u968f\u4e13\u5bb6\u3002<\/p>\n

\u8fd9\u79cd\u5206\u9636\u6bb5\u7684Specialist Training\u4fdd\u8bc1\u4e86\u6bcf\u4e2a\u9886\u57df\u7684\u4e0a\u9650\u88ab\u62c9\u5230\u6700\u9ad8\u3002<\/p>\n

\u6700\u540e\uff0c\u518d\u901a\u8fc7OPD\uff08Multi-teacher On-Policy Distillation\uff0c\u591a\u6559\u5e08\u5728\u7ebf\u7b56\u7565<\/a>\u84b8\u998f\uff09\uff0c\u5c06\u8fd9\u4e9b\u4e13\u5bb6\u7684\u7075\u9b42\u805a\u5408\u6210\u4e00\u4e2a\u7edf\u4e00\u7684\u6a21\u578b\u3002<\/p>\n

\u8fd9\u91cc\u5de5\u7a0b\u4e0a\u7684\u96be\u5ea6\u5728\u4e8e\uff0c\u540c\u65f6\u52a0\u8f7d\u5341\u591a\u4e2a\u4e07\u4ebf\u53c2\u6570\u7ea7\u7684\u6559\u5e08\u6a21\u578b\u505a\u5728\u7ebf\u63a8\u7406\u4e0d\u73b0\u5b9e\u3002<\/p>\n

V4\u7684\u65b9\u6848\u662f\u4e0d\u7f13\u5b58\u6559\u5e08\u7684logits\uff08\u663e\u5b58\u88c5\u4e0d\u4e0b\uff09\uff0c\u53ea\u7f13\u5b58\u6559\u5e08\u6700\u540e\u4e00\u5c42\u7684\u9690\u85cf\u72b6\u6001\uff0c\u8bad\u7ec3\u65f6\u6309\u9700\u901a\u8fc7prediction head\u91cd\u5efalogits\u3002<\/p>\n

\u7136\u540e\uff0c\u6309\u6559\u5e08\u7d22\u5f15\u6392\u5e8f\u8bad\u7ec3\u6837\u672c\uff0c\u786e\u4fdd\u6bcf\u4e2a\u6559\u5e08\u7684prediction head\u53ea\u52a0\u8f7d\u4e00\u6b21\u3002KL\u6563\u5ea6\u8ba1\u7b97\u5219\u7528TileLang\u7f16\u5199\u7684\u4e13\u7528kernel\u52a0\u901f\u3002<\/p>\n

\u544a\u522b\u4f20\u7edfReward Model<\/strong><\/p>\n

\u53e6\u5916\uff0c\u5bf9\u4e8e\u300c\u96be\u4ee5\u9a8c\u8bc1\uff08hard-to-verify\uff09\u300d\u7684\u4efb\u52a1\uff0c\u4f20\u7edf\u7684\u6807\u91cf\u5956\u52b1\u6a21\u578b\uff08Scalar Reward Model\uff09\u5df2\u7ecf\u529b\u4e0d\u4ece\u5fc3\u3002<\/p>\n

\u5bf9\u6b64\uff0cDeepSeek\u9009\u62e9\u5f15\u5165\u4e86Generative Reward Model (GRM)\u3002<\/p>\n

\u5b83\u4e0d\u518d\u7b80\u5355\u5730\u7ed9\u4e00\u4e2a0\u52301\u7684\u5206\u6570\uff0c\u800c\u662f\u6839\u636e\u9884\u8bbe\u7684Rubric\uff08\u8bc4\u4f30\u51c6\u5219\uff09\u751f\u6210\u8be6\u7ec6\u7684\u8bc4\u4f30\u62a5\u544a\u3002<\/p>\n

\u66f4\u5173\u952e\u7684\u662f\uff0cDeepSeek\u5bf9GRM\u672c\u8eab\u4e5f\u505a\u4e86RL\u4f18\u5316\uff0c\u8ba9actor\u7f51\u7edc\u540c\u65f6\u5145\u5f53\u751f\u6210\u5f0f\u5956\u52b1\u6a21\u578b\uff0c\u8bc4\u5224\u80fd\u529b\u548c\u751f\u6210\u80fd\u529b\u5728\u540c\u4e00\u4e2a\u6a21\u578b\u4e2d\u8054\u5408\u4f18\u5316\u3002<\/p>\n

\"DeepSeek-V4\u62a5\u544a\u4eae\u4e86\uff01V4\u53d1\u5e03\u5ef6\u8fdf\u7684\u79d8\u5bc6\uff0c\u7ec8\u4e8e\u66dd\u5149\u4e86\"<\/strong><\/p>\n

\u628aAgent\u505a\u6210\u4e00\u5957\u5206\u5e03\u5f0f\u7cfb\u7edf<\/strong><\/p>\n

\u4e0d\u4ec5\u5982\u6b64\uff0cDeepSeek\u8fd8\u4e3aV4\u4e13\u95e8\u81ea\u7814\u4e86\u4e00\u5957\u5e95\u5ea7\u3002<\/p>\n

DSec\uff1a\u751f\u4ea7\u7ea7\u6c99\u7bb1\u96c6\u7fa4<\/strong><\/p>\n

\u4e3a\u4e86\u8bad\u7ec3Agent\u7684\u5b9e\u64cd\u80fd\u529b\uff0cDeepSeek\u642d\u5efa\u4e86\u4e00\u4e2a\u540d\u4e3aDSec\u7684\u5e73\u53f0\u3002<\/p>\n

3FS\u5206\u5e03\u5f0f\u6587\u4ef6\u7cfb\u7edf\uff0c\u786e\u4fdd\u4e86\u6570\u636e\u7684\u6781\u901f\u5b58\u53d6\uff1b\u6570\u5341\u4e07\u5e76\u53d1Sandbox\u5b9e\u4f8b\uff0c\u5219\u610f\u5473\u7740V4\u5728\u8bad\u7ec3\u65f6\uff0c\u540c\u65f6\u6709\u51e0\u5341\u4e07\u53f0\u300c\u865a\u62df\u7535\u8111\u300d\u5728\u8dd1\u4ee3\u7801\u3001\u6d4bBug\u3002<\/p>\n

MegaMoE\uff1a\u901a\u4fe1\u8ba1\u7b97\u4e00\u4f53\u5316<\/strong><\/p>\n

\u5728MoE\u5c42\uff0cDeepSeek\u628a\u901a\u4fe1\u548c\u8ba1\u7b97\u878d\u5408\u8fdb\u5355\u4e2apipeline kernel\uff0c\u4e13\u5bb6\u6309wave\u8c03\u5ea6\uff0c\u901a\u4fe1\u5ef6\u8fdf\u5b8c\u5168\u9690\u85cf\u5728\u8ba1\u7b97\u4e4b\u4e0b\u3002<\/p>\n

\u7ed3\u679c\u5c31\u662f\uff0c\u901a\u7528\u573a\u666f\u52a0\u901f1.5\u52301.73\u500d\uff0cRL rollout\u7b49\u5ef6\u8fdf\u654f\u611f\u573a\u666f\u6700\u9ad81.96\u500d\u3002<\/p>\n

\u81ea\u7814DSML\uff1a\u62d2\u7edd\u8f6c\u4e49\u5931\u8d25<\/strong><\/p>\n

\u5de5\u5177\u8c03\u7528\u65b9\u9762\uff0cDeepSeek\u5e72\u8106\u81ea\u5df1\u8bbe\u8ba1\u4e86\u4e00\u5957\u7c7b\u4f3cXML\u7684DSL\uff08\u9886\u57df\u7279\u5b9a\u8bed\u8a00\uff09\u3002<\/p>\n

\u8fd9\u5957\u534f\u8bae\u7b80\u5355\u9ad8\u6548\uff0c\u76f4\u63a5\u628a\u5de5\u5177\u8c03\u7528\u7684\u6210\u529f\u7387\u4ece\u300c\u770b\u8fd0\u6c14\u300d\u63d0\u5347\u5230\u4e86\u300c\u5de5\u4e1a\u7ea7\u7a33\u5065\u300d\u3002<\/p>\n

\"DeepSeek-V4\u62a5\u544a\u4eae\u4e86\uff01V4\u53d1\u5e03\u5ef6\u8fdf\u7684\u79d8\u5bc6\uff0c\u7ec8\u4e8e\u66dd\u5149\u4e86\"<\/p>\n

Reasoning Effort\u5206\u6a21\u5f0f\u8bad\u7ec3<\/strong><\/p>\n

\u8fd8\u6709\u4e00\u4e2a\u7cbe\u7ec6\u7684\u8bbe\u8ba1\uff0c\u5c31\u662fV4\u652f\u6301\u4e0d\u540c\u7684\u601d\u8003\u6a21\u5f0f\u3002<\/p>\n

Non-think\u6a21\u5f0f\u662f\u7b80\u5355\u7684\u5de5\u5177\u9009\u62e9\uff0c\u79d2\u56de\u3002High\/Max\u5219\u9488\u5bf9\u957f\u6587\u6863\u3001\u91cd\u6784\u3001\u590d\u6742Bug\uff0c\u62c9\u6ee1\u63a8\u7406\u7b97\u529b\u3002<\/p>\n

\u8fd9\u79cd\u300c\u80fd\u7701\u5219\u7701\uff0c\u8be5\u72e0\u5219\u72e0\u300d\u7684\u7b56\u7565\uff0c\u4e5f\u662fV4\u6210\u672c\u80fd\u505a\u5230Claude 1\/4\u7684\u5173\u952e\u3002<\/p>\n

\u793e\u533a\u7684\u5f88\u591a\u7814\u7a76\u8005\u8bfb\u5b8c\u8fd9\u90e8\u5206\u540e\uff0c\u819c\u62dc\u5f97\u4e94\u4f53\u6295\u5730\uff1a\u300cDeepSeek\u7684\u5de5\u7a0b\u80fd\u529b\uff0c\u4f9d\u65e7\u624e\u5b9e\u5f97\u8ba9\u4eba\u6ca1\u8bdd\u8bf4\u300d\u3002<\/p>\n

\"DeepSeek-V4\u62a5\u544a\u4eae\u4e86\uff01V4\u53d1\u5e03\u5ef6\u8fdf\u7684\u79d8\u5bc6\uff0c\u7ec8\u4e8e\u66dd\u5149\u4e86\"<\/p>\n

Interleaved Thinking<\/strong>\u5347\u7ea7<\/strong><\/p>\n

V3.2\u5728\u6bcf\u4e2a\u65b0\u7528\u6237\u6d88\u606f\u5230\u6765\u65f6\u4f1a\u4e22\u5f03\u4e4b\u524d\u7684\u601d\u8003\u75d5\u8ff9\uff0cV4\u5728Tool-Calling\u573a\u666f\u4e0b\u4fdd\u7559\u4e86\u5b8c\u6574\u7684\u8de8\u8f6e\u6b21\u63a8\u7406\u5386\u53f2\uff0c\u8ba9Agent\u5728\u957f\u65f6\u7a0b\u4efb\u52a1\u4e2d\u7ef4\u6301\u8fde\u8d2f\u7684\u63a8\u7406\u94fe\u3002<\/p>\n

\u666e\u901a\u5bf9\u8bdd\u573a\u666f\u4ecd\u6bcf\u8f6e\u6e05\u7a7a\uff0c\u4fdd\u6301\u4e0a\u4e0b\u6587\u7cbe\u7b80\u3002<\/p>\n

\u786c\u5e01\u7684\u53e6\u4e00\u9762\uff0c\u662f94%\u7684\u5e7b\u89c9\u7387<\/strong><\/p>\n

Artificial Analysis\u7684\u5b9e\u6d4b\u7ed9\u51fa\u4e86\u4e00\u4e2a\u66f4\u7acb\u4f53\u7684\u753b\u9762\u3002<\/p>\n

\u8dd1\u5b8cIntelligence Index\u7684\u5168\u91cf\u57fa\u51c6\u6d4b\u8bd5\uff0cV4 Pro\u53ea\u82b1\u4e861071\u7f8e\u91d1\uff0c\u6bd4Claude Opus 4.7\u76844811\u7f8e\u91d1\u4fbf\u5b9c\u4e86\u56db\u500d\u591a\u3002<\/p>\n

Agent\u80fd\u529b\u65b9\u9762\uff0cV4 Pro Max\u5728GDPval-AA\u5b9e\u6d4b\uff08\u9762\u5411\u771f\u5b9e\u5de5\u4f5c\u4efb\u52a1\u7684Agent\u57fa\u51c6\uff09\u4e2d\u62ff\u5230\u4e861554\u5206\uff0c\u5168\u9762\u9886\u5148\u4e00\u4f17\u5f00\u6e90\u6a21\u578b\u3002<\/p>\n

\"DeepSeek-V4\u62a5\u544a\u4eae\u4e86\uff01V4\u53d1\u5e03\u5ef6\u8fdf\u7684\u79d8\u5bc6\uff0c\u7ec8\u4e8e\u66dd\u5149\u4e86\"<\/p>\n

\"DeepSeek-V4\u62a5\u544a\u4eae\u4e86\uff01V4\u53d1\u5e03\u5ef6\u8fdf\u7684\u79d8\u5bc6\uff0c\u7ec8\u4e8e\u66dd\u5149\u4e86\"<\/p>\n

\u7136\u800c\uff0c\u5929\u4e0b\u6ca1\u6709\u514d\u8d39\u7684\u5348\u9910\u3002<\/p>\n

Aritificial Analysis\u7684\u62a5\u544a\u91cc\u4e5f\u975e\u5e38\u5766\u8bda\u5730\u6307\u51fa\u4e86\u8fd9\u79cd\u505a\u6cd5\u7684\u4ee3\u4ef7\uff1aV4 pro\u5728AA-Ominiscience\u4e0a\u7684\u5e7b\u89c9\u7387\u9ad8\u8fbe94%\u3002<\/p>\n

\"DeepSeek-V4\u62a5\u544a\u4eae\u4e86\uff01V4\u53d1\u5e03\u5ef6\u8fdf\u7684\u79d8\u5bc6\uff0c\u7ec8\u4e8e\u66dd\u5149\u4e86\"<\/p>\n

\u8fd9\u63ed\u793a\u4e86\u4e00\u4e2a\u7ed3\u6784\u6027\u56f0\u5883\uff1a\u8981\u5728\u6709\u9650\u7b97\u529b\u9884\u7b97\u4e0b\u903c\u8fd1\u9876\u7ea7\u6027\u80fd\uff0c\u5c31\u4e0d\u5f97\u4e0d\u5728\u67d0\u4e9b\u7ef4\u5ea6\u4e0a\u505a\u53d6\u820d\u3002<\/p>\n

DeepSeek\u9009\u62e9\u628a\u7b79\u7801\u5168\u538b\u5728\u63a8\u7406\u548cAgent\u80fd\u529b\u4e0a\uff0c\u4ee3\u4ef7\uff0c\u5c31\u662f\u77e5\u8bc6\u90fd\u51c6\u786e\u6027\u3002<\/p>\n<\/p>\n

\u4e3a\u4ec0\u4e48\u6211\u4eec\u4f9d\u7136\u5bf9DeepSeek\u5145\u6ee1\u656c\u610f\uff1f<\/strong><\/p>\n

\u5728\u8fd9\u6b21V4\u7684\u62a5\u544a\u4e2d\uff0c\u6709\u4eba\u770b\u5230\u4e86\u300c\u8bad\u7ec3\u4e0d\u7a33\u300d\u7684\u5c34\u5c2c\uff0c\u6709\u4eba\u770b\u5230\u4e86\u300c\u5e7b\u89c9\u4e25\u91cd\u300d\u7684\u77ed\u677f\u3002<\/p>\n

\u4f46\u5728\u6211\u4eec\u770b\u6765\uff0c\u8fd9\u4efd\u62a5\u544a\u6700\u52a8\u4eba\u7684\u5730\u65b9\u5728\u4e8e\u900f\u660e\u3002<\/p>\n

\u4ed6\u4eec\u6562\u4e8e\u627f\u8ba4\u786c\u4ef6\u9002\u914d\u9635\u75db\uff0c\u6562\u4e8e\u62ab\u9732\u90a3\u4e9b\u770b\u4f3c\u300c\u8865\u4e01\u300d\u7684\u89e3\u51b3\u65b9\u6848\uff0c\u66f4\u6562\u4e8e\u5c55\u793a\u81ea\u5df1\u5982\u4f55\u7528\u6700\u786c\u6838\u7684\u5de5\u7a0b\u80fd\u529b\uff0c\u5728\u51e0\u5341\u4e07\u4e2a\u6c99\u7bb1\u91cc\u4e00\u70b9\u70b9\u78e8\u51faAgent\u7684\u7075\u9b42\u3002<\/p>\n

\u4eceV3\u7684Multi-head Latent Attention\u5230V4\u7684OPD\u84b8\u998f\u548cDSec\u6c99\u7bb1\uff0cDeepSeek\u6b63\u5728\u7528\u4e00\u79cd\u8fd1\u4e4e\u504f\u6267\u7684\u300c\u5de5\u7a0b\u4e3b\u4e49\u300d\uff0c\u63a2\u7d22\u7740\u5927\u6a21\u578b\u901a\u5f80AGI\u7684\u53e6\u4e00\u6761\u8def\u5f84\u2014\u2014<\/p>\n

\u5982\u679c\u67b6\u6784\u8fd8\u6ca1\u5b8c\u7f8e\uff0c\u90a3\u5c31\u7528\u5de5\u7a0b\u628a\u5899\u780c\u539a\uff1b\u5982\u679c\u7b97\u529b\u4e0d\u591f\u4fbf\u5b9c\uff0c\u90a3\u5c31\u7528\u7b97\u6cd5\u628a\u6548\u7387\u69a8\u5e72\u3002<\/p>\n

DeepSeek-V4\u4e5f\u8bb8\u4e0d\u662f\u6700\u5b8c\u7f8e\u7684\u7ec8\u5c40\uff0c\u4f46\u5b83\u7edd\u5bf9\u662f\u76ee\u524d\u6700\u771f\u5b9e\u7684\u3001\u6700\u5145\u6ee1\u6d3b\u529b\u7684\u300c\u4e2d\u56fdAI\u73b0\u573a\u300d\u3002<\/p>\n

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

DeepSeek-V4\u7684\u6280\u672f\u62a5\u544a\uff0c\u7b80\u76f4\u8bda\u5b9e\u5f97\u4ee4\u4eba\u9707\u60ca\u3002V4\u53d1\u5e03\u5ef6\u8fdf\u7684\u79d8\u5bc6\uff0c\u88ab\u6b63\u5f0f\u900f\u9732\u4e86\uff01\u8fd9\u9897\u5927\u96f7\u7684\u80cc\u540e\uff0c\u7a76\u7adf\u662f […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":[],"categories":[6],"tags":[71,88,339,192],"views":12,"_links":{"self":[{"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=\/wp\/v2\/posts\/769083"}],"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=769083"}],"version-history":[{"count":0,"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=\/wp\/v2\/posts\/769083\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=769083"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=769083"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=769083"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}