{"id":1409,"date":"2026-04-28T20:56:00","date_gmt":"2026-04-28T12:56:00","guid":{"rendered":"https:\/\/cdjax.com\/?p=1409"},"modified":"2026-04-28T11:58:29","modified_gmt":"2026-04-28T03:58:29","slug":"deepseek-v4-%e6%8a%80%e6%9c%af%e8%a7%a3%e6%9e%90%ef%bc%9a%e6%9e%b6%e6%9e%84%e9%9d%a9%e6%96%b0%e4%b8%8e-coding-agent-%e5%90%8e%e8%ae%ad%e7%bb%83%e4%bc%98%e5%8c%96","status":"publish","type":"post","link":"https:\/\/cdjax.com\/?p=1409","title":{"rendered":"DeepSeek-V4 \u6280\u672f\u89e3\u6790\uff1a\u67b6\u6784\u9769\u65b0\u4e0e Coding Agent \u540e\u8bad\u7ec3\u4f18\u5316"},"content":{"rendered":"\n

\u672c\u6587\u9996\u53d1\u4e8e\u516c\u4f17\u53f7\uff1a\u672a\u6765\u79d1\u6280<\/p>\n\n\n\n

DeepSeek \u56e2\u961f\u8fd1\u671f\u53d1\u5e03\u7684 DeepSeek-V4 \u7cfb\u5217\uff08\u5305\u62ec 1.6T \u53c2\u6570\u7684 Pro \u7248\u4e0e 284B \u53c2\u6570\u7684 Flash \u7248\uff09\u5c06\u5f00\u6e90\u5927\u6a21\u578b\u7684\u80fd\u529b\u8fb9\u754c\u63a8\u5411\u4e86\u767e\u4e07 token \u4e0a\u4e0b\u6587\u65f6\u4ee3\u3002\u8fd9\u4efd\u6280\u672f\u62a5\u544a\u4e2d\u6700\u503c\u5f97\u6df1\u5165\u63a2\u8ba8\u7684\uff0c\u4e00\u662f\u5176\u5728\u6a21\u578b\u67b6\u6784\u5c42\u9762\u5bf9\u957f\u4e0a\u4e0b\u6587\u6548\u7387\u74f6\u9888\u7684\u7cfb\u7edf\u6027\u7a81\u7834\uff0c\u4e8c\u662f\u5176\u5728\u540e\u8bad\u7ec3\u9636\u6bb5\u4e3a Coding Agent \u573a\u666f\u6240\u505a\u7684\u9488\u5bf9\u6027\u4f18\u5316\u3002\u672c\u6587\u56f4\u7ed5\u8fd9\u4e24\u6761\u4e3b\u7ebf\u5c55\u5f00\u3002<\/p>\n\n\n\n

\n\n\n\n

\u4e00\u3001\u6a21\u578b\u67b6\u6784\u7684\u521b\u65b0<\/h2>\n\n\n\n

DeepSeek-V4 \u5728\u4fdd\u7559 DeepSeek-V3 \u7684 MoE \u6846\u67b6\u4e0e MTP\uff08Multi-Token Prediction\uff09\u7684\u57fa\u7840\u4e0a\uff0c\u5f15\u5165\u4e86\u4e09\u9879\u5173\u952e\u7684\u67b6\u6784\u5347\u7ea7\uff1a\u6df7\u5408\u6ce8\u610f\u529b\u673a\u5236\uff08CSA + HCA\uff09\u3001\u6d41\u5f62\u7ea6\u675f\u8d85\u8fde\u63a5\uff08mHC\uff09\u4ee5\u53ca Muon \u4f18\u5316\u5668\u3002\u8fd9\u4e09\u8005\u5171\u540c\u6784\u6210\u4e86 V4 \u7cfb\u5217\u5728\u6548\u7387\u4e0e\u7a33\u5b9a\u6027\u4e0a\u7684\u62a4\u57ce\u6cb3\u3002<\/p>\n\n\n\n

1.1 \u6df7\u5408\u6ce8\u610f\u529b\uff1aCSA \u4e0e HCA \u7684\u534f\u540c<\/h3>\n\n\n\n

\u957f\u4e0a\u4e0b\u6587\u573a\u666f\u4e0b\uff0c\u6ce8\u610f\u529b\u673a\u5236\u7684\u4e8c\u6b21\u590d\u6742\u5ea6\u662f\u7ed5\u4e0d\u5f00\u7684\u74f6\u9888\u3002V4 \u6ca1\u6709\u9009\u62e9\u5355\u4e00\u8def\u7ebf\uff0c\u800c\u662f\u8bbe\u8ba1\u4e86\u4e24\u79cd\u538b\u7f29\u6ce8\u610f\u529b\u673a\u5236\u5e76\u4ee5\u4ea4\u9519\u65b9\u5f0f\u90e8\u7f72\u3002<\/p>\n\n\n\n

Compressed Sparse Attention\uff08CSA\uff09<\/strong> \u8d70\u7684\u662f”\u5148\u538b\u7f29\u3001\u518d\u7a00\u758f”\u7684\u8def\u7ebf\u3002\u5b83\u9996\u5148\u7528\u4e00\u4e2a token \u7ea7\u538b\u7f29\u5668\u5c06\u6bcf  \u4e2a token \u7684 KV \u7f13\u5b58\u5408\u5e76\u4e3a\u4e00\u6761 entry\uff08V4 \u4e2d \uff09\uff0c\u968f\u540e\u901a\u8fc7\u4e00\u4e2a\u8f7b\u91cf\u7ea7\u7684 Lightning Indexer \u8ba1\u7b97\u538b\u7f29 KV \u5757\u4e0e query \u7684\u76f8\u5173\u6027\u5206\u6570\uff0c\u5e76\u901a\u8fc7 top-k \u9009\u62e9\u5668\uff08Pro \u7248\u9009 1024 \u6761\uff09\u4fdd\u7559\u6700\u76f8\u5173\u7684\u538b\u7f29\u5757\u8fdb\u884c\u6838\u5fc3\u6ce8\u610f\u529b\u8ba1\u7b97\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0cCSA \u4f7f\u7528\u7684\u662f Multi-Query Attention \u5f62\u5f0f\uff0c\u4e14 query \u7684 latent \u5411\u91cf\u5728 indexer \u4e0e\u4e3b\u6ce8\u610f\u529b\u4e4b\u95f4\u5171\u4eab\uff0c\u8fdb\u4e00\u6b65\u538b\u7f29\u4e86\u8ba1\u7b97\u91cf\u3002\u6b64\u5916\uff0c\u7531\u4e8e query \u7684\u8f93\u51fa\u7ef4\u5ea6  \u8f83\u5927\uff0cV4 \u8fd8\u5f15\u5165\u4e86 grouped output projection\u2014\u2014\u5148\u5c06  \u4e2a\u5934\u5206\u6210  \u7ec4\uff0c\u6bcf\u7ec4\u5148\u6295\u5f71\u5230\u4e2d\u95f4\u7ef4\u5ea6 \uff0c\u518d\u5408\u5e76\u4e3a\u6700\u7ec8\u8f93\u51fa\uff0c\u907f\u514d\u4e86\u51fa\u53e3\u6295\u5f71\u6210\u4e3a\u65b0\u7684\u74f6\u9888\u3002<\/p>\n\n\n\n

Heavily Compressed Attention\uff08HCA\uff09<\/strong> \u5219\u8d70\u6781\u81f4\u538b\u7f29\u8def\u7ebf\uff0c\u5c06\u6bcf  \u4e2a token \u5408\u5e76\u4e3a\u4e00\u6761 entry\uff0c\u4f46\u653e\u5f03\u7a00\u758f\u9009\u62e9\uff0c\u5bf9\u6240\u6709\u538b\u7f29\u540e\u7684 entry \u6267\u884c dense attention\u3002\u7531\u4e8e\u538b\u7f29\u6bd4\u8db3\u591f\u5927\uff0cdense \u5f62\u5f0f\u7684\u5f00\u9500\u4e5f\u5df2\u88ab\u644a\u8584\u3002<\/p>\n\n\n\n

CSA \u4e0e HCA \u5728 Transformer \u5c42\u4e4b\u95f4\u4ea4\u9519\u90e8\u7f72\uff1aCSA \u4fdd\u7559\u4e86\u5c40\u90e8\u7cbe\u7ec6\u5ea6\u4f46\u6709 top-k \u4e0a\u9650\uff0cHCA \u63d0\u4f9b\u4e86\u5168\u5c40\u89c6\u91ce\u4f46\u5206\u8fa8\u7387\u8f83\u7c97\uff0c\u4e24\u8005\u5f62\u6210\u4e92\u8865\u3002\u518d\u53e0\u52a0\u51e0\u4e2a\u5de5\u7a0b\u7ec6\u8282\u2014\u2014RoPE \u4ec5\u4f5c\u7528\u4e8e\u6700\u540e 64 \u7ef4\u5e76\u5bf9\u6838\u5fc3\u6ce8\u610f\u529b\u8f93\u51fa\u65bd\u52a0  \u4f4d\u7f6e\u7684\u9006\u5411 RoPE \u4ee5\u6062\u590d\u76f8\u5bf9\u4f4d\u7f6e\u5173\u7cfb\uff1b\u989d\u5916\u7684\u6ed1\u52a8\u7a97\u53e3\u5206\u652f\uff08\uff09\u8865\u507f\u538b\u7f29\u5e26\u6765\u7684\u5c40\u90e8\u4fe1\u606f\u635f\u5931\uff1b\u53ef\u5b66\u4e60\u7684 attention sink logit \u5141\u8bb8\u67e5\u8be2\u5934\u5c06\u603b\u6ce8\u610f\u529b\u5206\u6570\u538b\u4f4e\u81f3\u63a5\u8fd1 0\u3002<\/p>\n\n\n\n

\u6548\u7387\u6536\u76ca\u662f\u60ca\u4eba\u7684\u3002\u5728 1M token \u4e0a\u4e0b\u6587\u4e0b\uff0cV4-Pro \u7684\u5355 token \u63a8\u7406 FLOPs \u4ec5\u4e3a V3.2 \u7684 27%\u3001KV \u7f13\u5b58\u4ec5\u4e3a 10%\uff1bV4-Flash \u66f4\u8fdb\u4e00\u6b65\u964d\u81f3 10% \u4e0e 7%\u3002\u914d\u5408 RoPE \u7ef4\u5ea6\u7528 BF16\u3001\u5176\u4f59\u7ef4\u5ea6\u7528 FP8 \u7684\u6df7\u5408\u5b58\u50a8\u683c\u5f0f\uff0c\u4ee5\u53ca Lightning Indexer \u5185\u90e8\u7684 FP4 \u8ba1\u7b97\uff0cV4 \u7684 KV \u7f13\u5b58\u5728 1M \u573a\u666f\u4e0b\u88ab\u538b\u5230\u4e86 BF16 GQA8 \u57fa\u7ebf\u7684\u7ea6 2%\u3002<\/p>\n\n\n\n

1.2 Manifold-Constrained Hyper-Connections\uff08mHC\uff09<\/h3>\n\n\n\n

\u6b8b\u5dee\u8fde\u63a5\u662f Transformer \u4fe1\u53f7\u4f20\u9012\u7684”\u9ad8\u901f\u516c\u8def”\u3002Hyper-Connections\uff08HC\uff09\u901a\u8fc7\u5c06\u6b8b\u5dee\u6d41\u7684\u5bbd\u5ea6\u4ece  \u6269\u5c55\u5230 \uff08V4 \u4e2d \uff09\u63d0\u4f9b\u4e86\u4e00\u4e2a\u72ec\u7acb\u4e8e hidden size \u7684\u6269\u5c55\u8f74\uff0c\u4f46\u5b9e\u6d4b\u4e2d\u5806\u53e0\u591a\u5c42 HC \u4f1a\u51fa\u73b0\u6570\u503c\u4e0d\u7a33\u5b9a\u3002<\/p>\n\n\n\n

V4 \u63d0\u51fa\u7684 mHC \u7684\u6838\u5fc3\u601d\u60f3\u662f\u5c06\u6b8b\u5dee\u53d8\u6362\u77e9\u9635  \u7ea6\u675f\u5728\u53cc\u968f\u673a\u77e9\u9635\u6d41\u5f62\uff08Birkhoff polytope\uff09\u4e0a<\/strong>\uff0c\u5373\u6ee1\u8db3\u884c\u548c\u5217\u4e4b\u548c\u5747\u4e3a 1 \u4e14\u5143\u7d20\u975e\u8d1f\u3002\u8fd9\u4e00\u7ea6\u675f\u4fdd\u8bc1\u4e86 \uff0c\u4f7f\u6b8b\u5dee\u53d8\u6362\u662f\u975e\u6269\u5f20\u7684\uff08non-expansive\uff09\uff0c\u524d\u5411\u4e0e\u53cd\u5411\u4f20\u64ad\u7684\u6570\u503c\u7a33\u5b9a\u6027\u90fd\u5f97\u5230\u52a0\u5f3a\uff1b\u540c\u65f6\u53cc\u968f\u673a\u77e9\u9635\u96c6\u5408\u5728\u4e58\u6cd5\u4e0b\u5c01\u95ed\uff0c\u6df1\u5c42\u5806\u53e0\u4f9d\u7136\u7a33\u5b9a\u3002\u5177\u4f53\u5b9e\u73b0\u4e0a\uff0c\u672a\u7ea6\u675f\u7684\u539f\u59cb\u77e9\u9635  \u5148\u7ecf\u8fc7\u6307\u6570\u53d8\u6362\u786e\u4fdd\u6b63\u6027\uff0c\u7136\u540e\u7528 Sinkhorn-Knopp \u7b97\u6cd5\u505a 20 \u6b21\u884c\u5217\u4ea4\u66ff\u5f52\u4e00\u5316\u6536\u655b\u5230\u6d41\u5f62\u4e0a\uff1b\u8f93\u5165\u4e0e\u8f93\u51fa\u6620\u5c04 \u3001 \u5219\u7528 Sigmoid \u9650\u5236\u975e\u8d1f\u6709\u754c\u3002<\/p>\n\n\n\n

\u8fd9\u79cd\u7ea6\u675f\u540c\u65f6\u89e3\u51b3\u4e86 HC \u7684\u7a33\u5b9a\u6027\u95ee\u9898\uff0c\u53c8\u4fdd\u7559\u4e86\u5b83\u72ec\u7acb\u4e8e hidden size \u63d0\u4f9b\u989d\u5916\u5bb9\u91cf\u7684\u4f18\u52bf\u3002<\/p>\n\n\n\n

1.3 Muon \u4f18\u5316\u5668<\/h3>\n\n\n\n

V4 \u5728\u5927\u90e8\u5206\u6a21\u5757\u4e0a\u5c06 AdamW \u66ff\u6362\u4e3a Muon\u2014\u2014\u8fd9\u662f Muon \u7b2c\u4e00\u6b21\u88ab\u7528\u5728\u4e07\u4ebf\u53c2\u6570 MoE \u6a21\u578b\u7684\u5168\u89c4\u6a21\u8bad\u7ec3\u4e0a\u3002Muon \u7684\u5173\u952e\u6b65\u9aa4\u662f\u5bf9\u52a8\u91cf\u77e9\u9635\u505a\u8fd1\u4f3c\u6b63\u4ea4\u5316\uff08\u901a\u8fc7 Newton-Schulz \u8fed\u4ee3\uff09\uff0c\u5176\u66f4\u65b0\u65b9\u5411\u63a5\u8fd1  \u7684\u5f62\u5f0f\uff0c\u76f8\u5f53\u4e8e\u628a\u5947\u5f02\u503c\u90fd\u62c9\u5230 1 \u9644\u8fd1\u3002V4 \u91c7\u7528\u4e86\u4e24\u9636\u6bb5\u6df7\u5408 Newton-Schulz \u8fed\u4ee3<\/strong>\uff1a\u524d 8 \u6b65\u7528\u6fc0\u8fdb\u7cfb\u6570  \u5feb\u901f\u903c\u8fd1\uff0c\u540e 2 \u6b65\u7528\u7a33\u5065\u7cfb\u6570  \u7cbe\u786e\u6536\u655b\u5230 1\u3002AdamW \u4ec5\u4fdd\u7559\u7ed9 embedding\u3001prediction head\u3001RMSNorm \u6743\u91cd\u4ee5\u53ca mHC \u7684\u9759\u6001\u504f\u7f6e\u548c\u95e8\u63a7\u56e0\u5b50\u3002<\/p>\n\n\n\n

Muon \u4e0e ZeRO \u7684\u517c\u5bb9\u6027\u662f\u5de5\u7a0b\u96be\u70b9\u2014\u2014ZeRO \u5047\u8bbe\u5143\u7d20\u7ea7\u4f18\u5316\u5668\u5141\u8bb8\u53c2\u6570\u77e9\u9635\u5207\u7247\u66f4\u65b0\uff0c\u800c Muon \u9700\u8981\u5b8c\u6574\u68af\u5ea6\u77e9\u9635\u505a\u6b63\u4ea4\u5316\u3002V4 \u7684\u89e3\u6cd5\u662f dense \u53c2\u6570\u7528\u80cc\u5305\u7b97\u6cd5\u505a\u77e9\u9635\u7ea7\u5206\u6876\u5206\u914d\uff0cMoE \u53c2\u6570\u5219\u6309\u4e13\u5bb6\u7c92\u5ea6\u5c55\u5e73\u3001\u8de8 rank \u5747\u5300\u5206\u914d\uff1bMoE \u68af\u5ea6\u540c\u6b65\u8fd8\u7528\u4e86\u968f\u673a\u820d\u5165\u5230 BF16 + \u4e24\u9636\u6bb5 all-to-all + \u672c\u5730 FP32 \u6c42\u548c\u7684\u65b9\u6848\uff0c\u628a\u901a\u4fe1\u91cf\u51cf\u534a\u7684\u540c\u65f6\u7ef4\u6301\u4e86\u6570\u503c\u7a33\u5065\u6027\u3002<\/p>\n\n\n\n

1.4 \u8bad\u7ec3\u7a33\u5b9a\u6027\u7684\u5b9e\u6218\u7ecf\u9a8c<\/h3>\n\n\n\n

\u4e07\u4ebf\u53c2\u6570 MoE \u8bad\u7ec3\u7684\u4e0d\u7a33\u5b9a\u6027\u51e0\u4e4e\u662f\u666e\u904d\u73b0\u8c61\u3002V4 \u62a5\u544a\u4e86\u4e24\u4e2a\u6709\u6548\u4f46\u7406\u8bba\u5c1a\u672a\u5b8c\u5168\u6e05\u6670\u7684\u6280\u5de7\uff1a<\/p>\n\n\n\n

    \n
  • Anticipatory Routing\uff08\u524d\u77bb\u8def\u7531\uff09<\/strong>\uff1a\u5728\u7b2c\u00a0\u00a0\u6b65\u7528\u5f53\u524d\u53c2\u6570\u00a0\u00a0\u505a\u7279\u5f81\u8ba1\u7b97\uff0c\u4f46\u8def\u7531\u7d22\u5f15\u7528\u00a0\u00a0\u7684\u5386\u53f2\u53c2\u6570\u63d0\u524d\u7b97\u597d\u3002\u8fd9\u4e00\u89e3\u8026\u6253\u7834\u4e86”\u8def\u7531\u5f02\u5e38 \u2192 \u4e13\u5bb6\u5931\u8861 \u2192 outlier \u653e\u5927 \u2192 \u8def\u7531\u66f4\u5f02\u5e38”\u7684\u6076\u6027\u5faa\u73af\u3002\u4ec5\u5728\u68c0\u6d4b\u5230 loss spike \u65f6\u89e6\u53d1\uff0c\u5e73\u65f6\u5173\u95ed\uff0c\u6574\u4f53\u5f00\u9500\u63a7\u5236\u5728\u7ea6 20%\u3002<\/li>\n\n\n\n
  • SwiGLU Clamping<\/strong>\uff1a\u5c06 SwiGLU \u7ebf\u6027\u5206\u91cf\u622a\u65ad\u5230\u00a0\u3001\u95e8\u63a7\u5206\u91cf\u4e0a\u754c\u8bbe\u4e3a 10\u3002\u7b80\u5355\u76f4\u63a5\uff0c\u4f46\u5b9e\u6d4b\u80fd\u6d88\u9664\u5927\u91cf outlier\u3002<\/li>\n<\/ul>\n\n\n\n
    \n\n\n\n

    \u4e8c\u3001Post-Training \u9636\u6bb5\u9488\u5bf9 Coding Agent \u7684\u4f18\u5316<\/h2>\n\n\n\n

    V4 \u7684\u540e\u8bad\u7ec3\u6d41\u6c34\u7ebf\u6574\u4f53\u91c7\u7528”\u4e13\u5bb6\u57f9\u517b \u2192 \u7edf\u4e00\u84b8\u998f”\u4e24\u9636\u6bb5\u8303\u5f0f\uff1a\u5148\u9488\u5bf9\u6570\u5b66\u3001\u4ee3\u7801\u3001agent\u3001\u6307\u4ee4\u8ddf\u968f\u7b49\u4e0d\u540c\u9886\u57df\u5206\u522b\u8bad\u7ec3\u4e13\u5bb6\u6a21\u578b\uff08SFT + GRPO \u5f3a\u5316\u5b66\u4e60\uff09\uff0c\u518d\u901a\u8fc7\u591a\u6559\u5e08 On-Policy Distillation\uff08OPD\uff09\u5c06\u80fd\u529b\u5408\u5e76\u5230\u7edf\u4e00\u6a21\u578b\u4e2d\u3002\u8fd9\u4e00\u8282\u805a\u7126\u5176\u4e2d\u4e0e Coding Agent \u76f4\u63a5\u76f8\u5173\u7684\u4f18\u5316\u3002<\/p>\n\n\n\n

    2.1 \u4e09\u6863\u63a8\u7406\u9884\u7b97\u4e0e Coding \u573a\u666f\u7684\u9002\u914d<\/h3>\n\n\n\n

    V4 \u540c\u65f6\u63d0\u4f9b Non-think\u3001Think High\u3001Think Max \u4e09\u79cd\u63a8\u7406\u6a21\u5f0f\uff0c\u6bcf\u79cd\u6a21\u5f0f\u5728 RL \u8bad\u7ec3\u65f6\u4f7f\u7528\u4e0d\u540c\u7684\u957f\u5ea6\u60e9\u7f5a\u548c\u4e0a\u4e0b\u6587\u7a97\u53e3\uff0c\u4f7f\u5f97\u540c\u4e00\u6a21\u578b\u53ef\u4ee5\u6839\u636e\u4efb\u52a1\u96be\u5ea6\u9009\u62e9”\u601d\u8003\u9884\u7b97”\u3002\u5bf9\u4e8e coding agent \u8fd9\u79cd\u672c\u8d28\u4e0a\u9700\u8981\u957f\u94fe\u8def\u63a8\u7406\u4e0e\u591a\u6b65\u5de5\u5177\u8c03\u7528\u7684\u573a\u666f\uff0cThink Max \u6a21\u5f0f\u989d\u5916\u6ce8\u5165\u4e00\u6bb5\u7cfb\u7edf\u63d0\u793a\uff0c\u660e\u786e\u8981\u6c42\u6a21\u578b\u7a77\u5c3d\u63a8\u7406\u3001\u538b\u529b\u6d4b\u8bd5\u903b\u8f91\u3001\u663e\u5f0f\u8bb0\u5f55\u6240\u6709\u4e2d\u95f4\u6b65\u9aa4\u4e0e\u88ab\u5426\u5b9a\u7684\u5047\u8bbe\u3002\u4ece\u8bc4\u6d4b\u7ed3\u679c\u770b\uff0cTerminal Bench 2.0 \u4e0a\u4ece None \u6a21\u5f0f\u7684 59.1 \u63d0\u5347\u5230 Max \u6a21\u5f0f\u7684 67.9\uff0c\u8bf4\u660e\u8fd9\u79cd\u63a8\u7406\u9884\u7b97\u7684\u5f39\u6027\u5bf9 agent \u4efb\u52a1\u6709\u663e\u8457\u6536\u76ca\u3002<\/p>\n\n\n\n

    2.2 \u65b0\u7684 Tool-Call Schema \u4e0e\u4ea4\u9519\u601d\u7ef4\u7ba1\u7406<\/h3>\n\n\n\n

    V4 \u7528\u4e00\u5957\u57fa\u4e8e <|DSML|><\/code> \u7279\u6b8a token \u7684 XML \u98ce\u683c\u8c03\u7528\u683c\u5f0f\u66ff\u4ee3\u4e86\u4f20\u7edf\u7684 JSON \u8c03\u7528\u7ea6\u5b9a\u3002\u6587\u4e2d\u660e\u786e\u6307\u51fa\uff0cXML \u683c\u5f0f\u53ef\u4ee5\u6709\u6548\u7f13\u89e3\u8f6c\u4e49\u5931\u8d25\u3001\u51cf\u5c11\u5de5\u5177\u8c03\u7528\u9519\u8bef<\/strong>\u2014\u2014\u8fd9\u5728\u957f\u94fe\u8def coding agent \u4e2d\u5c24\u5176\u5173\u952e\uff0c\u56e0\u4e3a\u5355\u4e00\u8c03\u7528\u9519\u8bef\u4f1a\u8ba9\u6574\u6761\u8f68\u8ff9\u5931\u8d25\u3002<\/p>\n\n\n\n

    \u66f4\u91cd\u8981\u7684\u662f Interleaved Thinking<\/strong> \u7684\u7ba1\u7406\u7b56\u7565\u5347\u7ea7\u3002V3.2 \u5728\u6bcf\u4e2a\u65b0\u7528\u6237\u56de\u5408\u5230\u6765\u65f6\u4f1a\u4e22\u5f03\u4e4b\u524d\u6240\u6709\u7684\u601d\u8003\u94fe\uff0c\u6bcf\u6b21\u90fd\u8981\u4ece\u5934\u91cd\u5efa\u4e0a\u4e0b\u6587\u3002V4 \u5229\u7528\u767e\u4e07 token \u4e0a\u4e0b\u6587\u7a97\u53e3\uff0c\u5728 Tool-Calling \u573a\u666f\u4e0b\u5b8c\u6574\u4fdd\u7559\u6240\u6709\u56de\u5408\uff08\u5305\u62ec\u8de8\u7528\u6237\u6d88\u606f\u7684\uff09\u601d\u8003\u5185\u5bb9<\/strong>\uff0c\u8ba9\u6a21\u578b\u5728\u957f horizon \u7684 agentic \u4efb\u52a1\u4e2d\u7ef4\u6301\u7d2f\u79ef\u7684\u3001\u8fde\u8d2f\u7684\u63a8\u7406\u8109\u7edc\uff1b\u53ea\u5728\u4e0d\u6d89\u53ca\u5de5\u5177\u7684\u7eaf\u5bf9\u8bdd\u573a\u666f\u6cbf\u7528\u65e7\u7684\u4e22\u5f03\u7b56\u7565\u3002\u5bf9 coding agent \u6765\u8bf4\uff0c\u8fd9\u610f\u5473\u7740\u6a21\u578b\u5728\u8c03\u8bd5\u4e00\u4e2a\u957f\u8fbe\u6570\u5341\u4e2a\u6b65\u9aa4\u7684\u4fee\u590d\u6d41\u7a0b\u65f6\uff0c\u4e0d\u5fc5\u6bcf\u6b21\u91cd\u65b0\u7406\u89e3\u9879\u76ee\u7ed3\u6784\u4e0e\u5df2\u5c1d\u8bd5\u7684\u65b9\u6848\u3002<\/p>\n\n\n\n

    2.3 \u5168\u8bcd\u8868 OPD\uff1a\u8ba9\u4ee3\u7801\u4e13\u5bb6\u7684\u80fd\u529b\u65e0\u635f\u8fc1\u79fb<\/h3>\n\n\n\n

    OPD \u7684\u76ee\u6807\u662f\u8ba9\u5b66\u751f\u6a21\u578b  \u5728\u81ea\u5df1\u751f\u6210\u7684\u8f68\u8ff9\u4e0a\u5b66\u4e60\u591a\u4e2a\u6559\u5e08\u4e13\u5bb6\uff08\u5305\u62ec coding \u4e13\u5bb6\uff09\u7684\u8f93\u51fa\u5206\u5e03\uff0c\u76ee\u6807\u51fd\u6570\u662f\u52a0\u6743\u53cd\u5411 KL\uff1a<\/p>\n\n\n\n

    \u4ee5\u5f80\u5de5\u4f5c\u901a\u5e38\u5c06\u8fd9\u4e2a KL \u9000\u5316\u4e3a token \u7ea7\u4f30\u8ba1\u5e76\u585e\u8fdb RL \u6846\u67b6\uff0c\u4f46\u8fd9\u4f1a\u5e26\u6765\u9ad8\u65b9\u5dee\u548c\u8bad\u7ec3\u4e0d\u7a33\u5b9a\u3002V4 \u575a\u6301\u5168\u8bcd\u8868 logit \u84b8\u998f<\/strong>\u2014\u2014\u4fdd\u7559\u5b8c\u6574 logit \u5206\u5e03\u505a\u53cd\u5411 KL\uff0c\u68af\u5ea6\u4f30\u8ba1\u66f4\u7a33\uff0c\u6559\u5e08\u77e5\u8bc6\u4fdd\u7559\u66f4\u5fe0\u5b9e\u3002<\/p>\n\n\n\n

    \u8fd9\u80cc\u540e\u7684\u5de5\u7a0b\u652f\u6491\u76f8\u5f53\u624e\u5b9e\uff1a\u8d85\u8fc7\u5341\u4e2a\u4e07\u4ebf\u7ea7\u6559\u5e08\u6a21\u578b\u7684\u6743\u91cd\u88ab offload \u5230\u5206\u5e03\u5f0f\u5b58\u50a8\u6309\u9700 ZeRO \u5f0f\u52a0\u8f7d\uff1b\u4e3a\u907f\u514d  \u8bcd\u8868\u7684 logit \u7269\u5316\u7206\u663e\u5b58\uff0cV4 \u53ea\u7f13\u5b58\u6559\u5e08\u6700\u540e\u4e00\u5c42 hidden state \u5230\u4e2d\u592e buffer\uff0c\u8bad\u7ec3\u65f6\u6309\u9700\u901a\u8fc7 prediction head \u91cd\u5efa logits\uff1b\u8bad\u7ec3\u6837\u672c\u6309\u6559\u5e08\u7d22\u5f15\u6392\u5e8f\uff0c\u4fdd\u8bc1\u6bcf\u4e2a mini-batch \u81f3\u591a\u4e00\u4e2a prediction head \u9a7b\u7559 GPU\u3002\u6700\u7ec8 KL \u6563\u5ea6\u7531\u4e13\u95e8\u7684 TileLang kernel \u8ba1\u7b97\u3002<\/p>\n\n\n\n

    \u5bf9 coding agent \u800c\u8a00\uff0c\u8fd9\u610f\u5473\u7740\u4ee3\u7801\u4e13\u5bb6\u901a\u8fc7\u5f3a\u5316\u5b66\u4e60\u83b7\u5f97\u7684\u7ec6\u7c92\u5ea6\u80fd\u529b\uff08\u9519\u8bef\u5904\u7406\u3001API \u8c03\u7528\u4e60\u60ef\u3001\u91cd\u6784\u5224\u65ad\uff09\u80fd\u591f\u4ee5\u63a5\u8fd1\u65e0\u635f\u7684\u5f62\u5f0f\u878d\u5408\u8fdb\u7edf\u4e00\u6a21\u578b\uff0c\u800c\u4e0d\u662f\u88ab token \u7ea7\u8fd1\u4f3c\u7a00\u91ca\u6389\u3002<\/p>\n\n\n\n

    2.4 RL Rollout \u7684\u957f\u4e0a\u4e0b\u6587\u4e0e\u5bb9\u9519\u652f\u6491<\/h3>\n\n\n\n

    Coding agent \u7684\u8bad\u7ec3 rollout \u7ecf\u5e38\u6d89\u53ca 50 \u4e07 token \u4ee5\u4e0a\u7684\u957f\u8f68\u8ff9\uff0c\u4e14\u5355\u6b21 rollout \u53ef\u80fd\u9700\u8981\u6570\u767e\u6b21\u5de5\u5177\u8c03\u7528\u3002V4 \u5728\u57fa\u7840\u8bbe\u65bd\u4e0a\u505a\u4e86\u51e0\u9879\u5173\u952e\u4f18\u5316\uff1a<\/p>\n\n\n\n

    FP4 \u91cf\u5316\u96c6\u6210\u5230 rollout<\/strong>\uff1arollout \u4e0e\u6240\u6709 inference-only \u524d\u5411\uff08\u5305\u62ec\u6559\u5e08\u4e0e\u53c2\u8003\u6a21\u578b\uff09\u76f4\u63a5\u4f7f\u7528\u539f\u751f FP4 \u6743\u91cd\uff0c\u800c\u8bad\u7ec3\u6b65\u9aa4\u7528 FP4\u2192FP8 \u65e0\u635f\u53cd\u91cf\u5316\u6a21\u62df\uff0c\u590d\u7528\u73b0\u6709 FP8 mixed-precision \u6846\u67b6\u3002\u8fd9\u663e\u8457\u964d\u4f4e\u4e86 rollout \u9636\u6bb5\u7684\u663e\u5b58\u5360\u7528\u4e0e\u91c7\u6837\u5ef6\u8fdf\u3002<\/p>\n\n\n\n

    Token \u7c92\u5ea6 Write-Ahead Log \u5bb9\u9519<\/strong>\uff1a\u6bcf\u751f\u6210\u4e00\u4e2a token \u7acb\u5373\u8ffd\u52a0\u5230\u8be5\u8bf7\u6c42\u7684 WAL\u3002\u88ab\u62a2\u5360\u65f6\u6682\u505c\u63a8\u7406\u5f15\u64ce\u3001\u4fdd\u5b58 KV \u7f13\u5b58\uff1b\u6062\u590d\u65f6\u57fa\u4e8e\u6301\u4e45\u5316\u7684 WAL \u4e0e KV \u7f13\u5b58\u7ee7\u7eed\u89e3\u7801\uff1b\u5373\u4f7f\u786c\u4ef6\u9519\u8bef\u4e5f\u80fd\u901a\u8fc7 WAL \u91cd\u5efa KV \u7f13\u5b58\u7ee7\u7eed\u3002\u4ece\u5934\u91cd\u751f\u6210\u4f1a\u5f15\u5165\u957f\u5ea6\u504f\u5dee\uff08\u77ed\u54cd\u5e94\u66f4\u5bb9\u6613\u5728\u4e2d\u65ad\u4e2d\u5b58\u6d3b\uff0c\u5bfc\u81f4\u6a21\u578b\u504f\u5411\u77ed\u8f93\u51fa\uff09\uff0cWAL \u673a\u5236\u4ece\u6570\u5b66\u4e0a\u4fdd\u8bc1\u4e86\u6b63\u786e\u6027\u3002<\/p>\n\n\n\n

    DSec \u6c99\u7bb1\u5e73\u53f0<\/strong>\uff1a\u4e3a coding agent \u7684\u6267\u884c\u73af\u5883\u63d0\u4f9b\u4e86\u56db\u79cd\u6267\u884c\u5e95\u5ea7\uff08Function Call\u3001Container\u3001microVM\u3001fullVM\uff09\u7684\u7edf\u4e00\u63a5\u53e3\u3002Container \u7528 EROFS \u6309\u9700\u52a0\u8f7d\u57fa\u7840\u955c\u50cf\uff0cmicroVM \u7528 overlaybd \u5b9e\u73b0\u5feb\u7167\u94fe\uff0c\u5355\u96c6\u7fa4\u53ef\u7ba1\u7406\u6570\u5341\u4e07\u5e76\u53d1\u6c99\u7bb1\u5b9e\u4f8b\u3002\u8f68\u8ff9\u65e5\u5fd7\u652f\u6301\u62a2\u5360\u6062\u590d\u65f6\u7684\u5ba2\u6237\u7aef\u5feb\u8fdb\u2014\u2014\u5df2\u5b8c\u6210\u7684\u547d\u4ee4\u76f4\u63a5\u91cd\u653e\u7f13\u5b58\u7ed3\u679c\uff0c\u907f\u514d\u975e\u5e42\u7b49\u64cd\u4f5c\u7684\u91cd\u590d\u6267\u884c\uff08\u8fd9\u5728\u6d89\u53ca\u6587\u4ef6\u7cfb\u7edf\u3001\u7f51\u7edc\u8bf7\u6c42\u7684 coding agent \u4e2d\u81f3\u5173\u91cd\u8981\uff09\u3002<\/p>\n\n\n\n

    2.5 Coding \u8bc4\u6d4b\uff1a\u4ece\u57fa\u51c6\u5230\u771f\u5b9e\u7814\u53d1\u4efb\u52a1<\/h3>\n\n\n\n

    V4-Pro \u5728 Codeforces \u4e0a\u62ff\u5230 3206 Elo\uff08\u4eba\u7c7b\u6392\u540d\u7ea6\u7b2c 23 \u4f4d\uff09\uff0cLiveCodeBench Pass@1 \u8fbe 93.5\uff0c\u662f\u9996\u4e2a\u5728\u7f16\u7a0b\u7ade\u8d5b\u4e0a\u5bf9\u6807 GPT-5.4 \u7684\u5f00\u6e90\u6a21\u578b\u3002SWE Verified 80.6\u3001SWE Multilingual 76.2\u3001Terminal Bench 2.0 67.9\uff08\u5176 Verified \u5b50\u96c6\u7ea6 72.0\uff09\uff0c\u4e0e K2.6\u3001GLM-5.1 \u5904\u4e8e\u540c\u4e00\u68af\u961f\u3002<\/p>\n\n\n\n

    \u4f46\u66f4\u503c\u5f97\u5173\u6ce8\u7684\u662f DeepSeek \u56e2\u961f\u81ea\u5efa\u7684 R&D Coding Benchmark\u2014\u2014\u4ece 50+ \u5185\u90e8\u5de5\u7a0b\u5e08\u6536\u96c6 200 \u4e2a\u771f\u5b9e\u4efb\u52a1\uff0c\u8986\u76d6 PyTorch\u3001CUDA\u3001Rust\u3001C++ \u8de8\u6280\u672f\u6808\u7684\u529f\u80fd\u5f00\u53d1\u3001bug \u4fee\u590d\u3001\u91cd\u6784\u3001\u8bca\u65ad\uff0c\u7ecf\u8fc7\u8d28\u91cf\u7b5b\u9009\u4fdd\u7559 30 \u4e2a\u8bc4\u6d4b\u4efb\u52a1\u3002\u5728\u8fd9\u4e2a\u66f4\u63a5\u8fd1\u771f\u5b9e\u5de5\u7a0b\u7684\u57fa\u51c6\u4e0a\uff1a<\/p>\n\n\n\n

    \u6a21\u578b<\/th>Pass Rate<\/th><\/tr><\/thead>
    Haiku 4.5<\/td>13%<\/td><\/tr>
    Sonnet 4.5<\/td>47%<\/td><\/tr>
    DeepSeek-V4-Pro-Max<\/strong><\/td>67%<\/strong><\/td><\/tr>
    Opus 4.5<\/td>70%<\/td><\/tr>
    Opus 4.5 Thinking<\/td>73%<\/td><\/tr>
    Opus 4.6 Thinking<\/td>80%<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n

    V4-Pro \u663e\u8457\u8d85\u8d8a Sonnet 4.5\u3001\u63a5\u8fd1 Opus 4.5\u3002\u5728\u5bf9 85 \u540d DeepSeek \u5185\u90e8\u7814\u7a76\u5458\u4e0e\u5de5\u7a0b\u5e08\u7684\u8c03\u7814\u4e2d\uff0c52% \u8868\u793a V4-Pro \u5df2\u53ef\u4f5c\u4e3a\u65e5\u5e38\u9ed8\u8ba4\u7f16\u7801\u6a21\u578b\uff0c39% \u503e\u5411\u4e8e\u662f\uff0c\u53cd\u5bf9\u8005\u4e0d\u5230 9%\u3002\u88ab\u53cd\u590d\u63d0\u53ca\u7684\u4e0d\u8db3\u662f\u5076\u53d1\u7684\u4f4e\u7ea7\u9519\u8bef\u3001\u5bf9\u6a21\u7cca\u6307\u4ee4\u7684\u8bef\u89e3\u4e0e\u8fc7\u5ea6\u601d\u8003\u2014\u2014\u8fd9\u6070\u597d\u6307\u5411\u540e\u7eed\u8fed\u4ee3\u7684\u65b9\u5411\u3002<\/p>\n\n\n\n

    \n\n\n\n

    \u4e09\u3001\u5c0f\u7ed3<\/h2>\n\n\n\n

    DeepSeek-V4 \u7684\u67b6\u6784\u521b\u65b0\u53ef\u4ee5\u6982\u62ec\u4e3a\u7528\u7ed3\u6784\u5316\u538b\u7f29\u6362\u53d6\u957f\u4e0a\u4e0b\u6587\u6548\u7387\uff0c\u7528\u6d41\u5f62\u7ea6\u675f\u6362\u53d6\u6df1\u5c42\u7a33\u5b9a\u6027\uff0c\u7528\u77e9\u9635\u7ea7\u4f18\u5316\u6362\u53d6\u6536\u655b\u901f\u5ea6<\/strong>\u3002CSA + HCA \u7684\u6df7\u5408\u6ce8\u610f\u529b\u662f\u5f53\u524d\u5f00\u6e90\u793e\u533a\u5728\u767e\u4e07 token \u4e0a\u4e0b\u6587\u65b9\u5411\u6700\u7cfb\u7edf\u7684\u5de5\u7a0b\u5316\u65b9\u6848\uff1bmHC \u7528 Birkhoff polytope \u8fd9\u4e00\u6f02\u4eae\u7684\u6570\u5b66\u5de5\u5177\u89e3\u51b3\u4e86 Hyper-Connections \u7684\u5b9e\u6218\u75db\u70b9\uff1bMuon \u5728\u4e07\u4ebf\u53c2\u6570\u89c4\u6a21\u7684\u6210\u529f\u90e8\u7f72\u4e5f\u4e3a\u540e\u7eed\u6a21\u578b\u63d0\u4f9b\u4e86\u91cd\u8981\u53c2\u8003\u3002<\/p>\n\n\n\n

    \u540e\u8bad\u7ec3\u4fa7\uff0cV4 \u5e76\u6ca1\u6709\u62bc\u6ce8\u5355\u4e00\u7684 RL \u7b97\u6cd5\u5947\u8ff9\uff0c\u800c\u662f\u7528\u4e13\u5bb6\u57f9\u517b + \u5168\u8bcd\u8868 OPD \u7684\u4e24\u9636\u6bb5\u8303\u5f0f\uff0c\u914d\u5408 XML \u5de5\u5177\u8c03\u7528\u3001\u8de8\u56de\u5408\u601d\u7ef4\u4fdd\u7559\u3001token \u7ea7 WAL\u3001DSec \u6c99\u7bb1\u7b49\u7ec4\u5408\u62f3\uff0c\u628a Coding Agent \u8fd9\u4e00\u957f\u94fe\u8def\u3001\u5f3a\u5de5\u7a0b\u4f9d\u8d56\u7684\u80fd\u529b\u505a\u624e\u5b9e\u3002R&D \u5185\u90e8\u57fa\u51c6 67% \u901a\u8fc7\u7387\u4e0e\u5f00\u53d1\u8005 91% \u7684\u6b63\u5411\u53cd\u9988\uff0c\u6bd4\u4efb\u4f55\u516c\u5f00 benchmark \u90fd\u66f4\u80fd\u8bf4\u660e\u95ee\u9898\u3002<\/p>\n\n\n\n

    \u5bf9\u4ece\u4e1a\u8005\u800c\u8a00\uff0cV4 \u7ed9\u51fa\u7684\u6700\u5927\u542f\u793a\u6216\u8bb8\u662f\uff1a\u957f\u4e0a\u4e0b\u6587\u4e0e agent \u80fd\u529b\u4e0d\u662f\u72ec\u7acb\u8ffd\u6c42\u7684\u4e24\u4ef6\u4e8b<\/strong>\u2014\u2014\u53ea\u6709\u5f53\u6ce8\u610f\u529b\u6548\u7387\u8db3\u591f\u9ad8\u3001\u63a8\u7406\u72b6\u6001\u80fd\u591f\u6301\u7eed\u4fdd\u7559\u3001\u8bad\u7ec3\u57fa\u7840\u8bbe\u65bd\u80fd\u5bb9\u5fcd\u957f rollout \u4e0e\u9891\u7e41\u62a2\u5360\u65f6\uff0cagentic AI \u624d\u771f\u6b63\u5177\u5907\u89c4\u6a21\u5316\u8bad\u7ec3\u7684\u53ef\u80fd\u3002V4 \u628a\u8fd9\u6761\u8def\u5f84\u5b8c\u6574\u5730\u8d70\u901a\u4e86\u4e00\u6b21\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

    \u672c\u6587\u9996\u53d1\u4e8e\u516c\u4f17\u53f7\uff1a\u672a\u6765\u79d1\u6280 DeepSeek \u56e2\u961f\u8fd1\u671f\u53d1\u5e03\u7684 DeepSeek-V4 \u7cfb\u5217\uff08\u5305\u62ec 1.6T \u53c2 […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"jetpack_post_was_ever_published":false,"_jetpack_newsletter_access":"","_jetpack_dont_email_post_to_subs":false,"_jetpack_newsletter_tier_id":0,"_jetpack_memberships_contains_paywalled_content":false,"_jetpack_memberships_contains_paid_content":false,"footnotes":"","jetpack_publicize_message":"","jetpack_publicize_feature_enabled":true,"jetpack_social_post_already_shared":true,"jetpack_social_options":{"image_generator_settings":{"template":"highway","default_image_id":0,"font":"","enabled":false},"version":2}},"categories":[5],"tags":[],"class_list":["post-1409","post","type-post","status-publish","format-standard","hentry","category-product"],"jetpack_publicize_connections":[],"jetpack_featured_media_url":"","jetpack_sharing_enabled":true,"jetpack_shortlink":"https:\/\/wp.me\/p8LQPD-mJ","_links":{"self":[{"href":"https:\/\/cdjax.com\/index.php?rest_route=\/wp\/v2\/posts\/1409","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/cdjax.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/cdjax.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/cdjax.com\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/cdjax.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=1409"}],"version-history":[{"count":1,"href":"https:\/\/cdjax.com\/index.php?rest_route=\/wp\/v2\/posts\/1409\/revisions"}],"predecessor-version":[{"id":1410,"href":"https:\/\/cdjax.com\/index.php?rest_route=\/wp\/v2\/posts\/1409\/revisions\/1410"}],"wp:attachment":[{"href":"https:\/\/cdjax.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=1409"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/cdjax.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=1409"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/cdjax.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=1409"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}