{"id":44,"date":"2026-04-24T11:56:20","date_gmt":"2026-04-24T03:56:20","guid":{"rendered":"https:\/\/blog.uika.cn\/?p=44"},"modified":"2026-04-24T11:56:21","modified_gmt":"2026-04-24T03:56:21","slug":"deepseek-v4%e4%b8%ad%e6%96%87%e7%bf%bb%e8%af%91%ef%bc%88%e7%bf%bb%e8%af%91%e6%a8%a1%e5%9e%8b%ef%bc%9aopus4-7%ef%bc%89","status":"publish","type":"post","link":"https:\/\/blog.uika.cn\/?p=44","title":{"rendered":"DeepSeek-V4\u4e2d\u6587\u7ffb\u8bd1\uff08\u7ffb\u8bd1\u6a21\u578b\uff1aOpus4.7\uff09"},"content":{"rendered":"\n<h1 class=\"wp-block-heading\" id=\"deepseek-v4\uff1a\u8fc8\u5411\u9ad8\u6548\u7684\u767e\u4e07-token-\u957f\u4e0a\u4e0b\u6587\u667a\u80fd\">DeepSeek-V4\uff1a\u8fc8\u5411\u9ad8\u6548\u7684\u767e\u4e07 Token \u957f\u4e0a\u4e0b\u6587\u667a\u80fd<\/h1>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u4f5c\u8005<\/strong>\uff1aDeepSeek-AI\n<strong>\u8054\u7cfb\u65b9\u5f0f<\/strong>\uff1a<a href=\"mailto:research@deepseek.com\" target=\"_blank\"  rel=\"nofollow\" >research@deepseek.com<\/a>\n<strong>\u6a21\u578b\u6743\u91cd<\/strong>\uff1a<a href=\"https:\/\/huggingface.co\/collections\/deepseek-ai\/deepseek-v4\" target=\"_blank\"  rel=\"nofollow\" >https:\/\/huggingface.co\/collections\/deepseek-ai\/deepseek-v4<\/a><\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">\u672c\u6587\u6863\u4e3a DeepSeek-V4 \u6280\u672f\u62a5\u544a\u7684\u4e2d\u6587\u7ffb\u8bd1\u7248\u672c\u3002\u539f\u6587\u4e3a\u82f1\u6587\uff0c\u672c\u7ffb\u8bd1\u5728\u4fdd\u7559\u6280\u672f\u672f\u8bed\u82f1\u6587\u5f62\u5f0f\u7684\u57fa\u7840\u4e0a\u8fdb\u884c\u610f\u8bd1\uff0c\u4ee5\u4fbf\u4e2d\u6587\u8bfb\u8005\u7406\u89e3\u3002\u6587\u4e2d\u6240\u6709\u4e0e\u5176\u4ed6\u6a21\u578b\u7684\u5bf9\u6bd4\u3001\u4e0a\u4ee3\u6a21\u578b\u7684\u63d0\u5347\u7b49\u5173\u952e\u4fe1\u606f\u5747\u5df2\u6807\u6ce8\u548c\u7a81\u51fa\u3002<\/p>\n<\/blockquote>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"\ud83d\udccb-\u6458\u8981\">\ud83d\udccb \u6458\u8981<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">DeepSeek-V4 \u7cfb\u5217\u662f\u672c\u6b21\u53d1\u5e03\u7684\u9884\u89c8\u7248\u672c\uff0c\u5305\u542b\u4e24\u6b3e\u5f3a\u5927\u7684 Mixture-of-Experts\uff08MoE\uff09\u8bed\u8a00\u6a21\u578b\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>DeepSeek-V4-Pro<\/strong>\uff1a\u603b\u53c2\u6570 <strong>1.6T<\/strong>\uff0c\u6fc0\u6d3b\u53c2\u6570 <strong>49B<\/strong><\/li>\n\n\n\n<li><strong>DeepSeek-V4-Flash<\/strong>\uff1a\u603b\u53c2\u6570 <strong>284B<\/strong>\uff0c\u6fc0\u6d3b\u53c2\u6570 <strong>13B<\/strong><\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e24\u6b3e\u6a21\u578b\u5747\u539f\u751f\u652f\u6301 <strong>1M\uff08\u4e00\u767e\u4e07\uff09token<\/strong> \u4e0a\u4e0b\u6587\u957f\u5ea6\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"\u6838\u5fc3\u67b6\u6784\u4e0e\u4f18\u5316\u5347\u7ea7\">\u6838\u5fc3\u67b6\u6784\u4e0e\u4f18\u5316\u5347\u7ea7<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">DeepSeek-V4 \u7cfb\u5217\u5728\u67b6\u6784\u548c\u4f18\u5316\u4e0a\u5f15\u5165\u4e86\u4e09\u9879\u5173\u952e\u5347\u7ea7\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u6df7\u5408\u6ce8\u610f\u529b\u67b6\u6784<\/strong>\uff1a\u7ed3\u5408 Compressed Sparse Attention\uff08CSA\uff0c\u538b\u7f29\u7a00\u758f\u6ce8\u610f\u529b\uff09\u548c Heavily Compressed Attention\uff08HCA\uff0c\u91cd\u538b\u7f29\u6ce8\u610f\u529b\uff09\uff0c\u5927\u5e45\u63d0\u5347\u957f\u4e0a\u4e0b\u6587\u6548\u7387\u3002<\/li>\n\n\n\n<li><strong>Manifold-Constrained Hyper-Connections\uff08mHC\uff0c\u6d41\u5f62\u7ea6\u675f\u8d85\u8fde\u63a5\uff09<\/strong>\uff1a\u5f3a\u5316\u4f20\u7edf\u6b8b\u5dee\u8fde\u63a5\u3002<\/li>\n\n\n\n<li><strong>Muon \u4f18\u5316\u5668<\/strong>\uff1a\u6536\u655b\u66f4\u5feb\uff0c\u8bad\u7ec3\u7a33\u5b9a\u6027\u66f4\u5f3a\u3002<\/li>\n<\/ol>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"\u8bad\u7ec3\u89c4\u6a21\">\u8bad\u7ec3\u89c4\u6a21<\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u4e24\u6b3e\u6a21\u578b\u9884\u8bad\u7ec3\u6570\u636e\u91cf\u5747\u8d85\u8fc7 <strong>32T tokens<\/strong>\uff08DeepSeek-V4-Flash \u4e3a 32T\uff0cDeepSeek-V4-Pro \u4e3a 33T\uff09\u3002<\/li>\n\n\n\n<li>\u968f\u540e\u7ecf\u8fc7\u5b8c\u6574\u7684\u540e\u8bad\u7ec3\u7ba1\u7ebf\u89e3\u9501\u5e76\u8fdb\u4e00\u6b65\u589e\u5f3a\u6a21\u578b\u80fd\u529b\u3002<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"\ud83c\udfaf-\u4e0e-deepseek-v32-\u7684\u6548\u7387\u5bf9\u6bd4\uff081m-token-\u4e0a\u4e0b\u6587\uff09\">\ud83c\udfaf \u4e0e DeepSeek-V3.2 \u7684\u6548\u7387\u5bf9\u6bd4\uff081M-token \u4e0a\u4e0b\u6587\uff09<\/h3>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>\u6307\u6807<\/th><th>DeepSeek-V4-Pro vs V3.2<\/th><th>DeepSeek-V4-Flash vs V3.2<\/th><\/tr><\/thead><tbody><tr><td>\u5355 token \u63a8\u7406 FLOPs<\/td><td>\u4ec5\u4e3a **27%**\uff083.7\u00d7 \u964d\u4f4e\uff09<\/td><td>\u4ec5\u4e3a **10%**\uff089.8\u00d7 \u964d\u4f4e\uff09<\/td><\/tr><tr><td>KV Cache \u5927\u5c0f<\/td><td>\u4ec5\u4e3a **10%**\uff089.5\u00d7 \u7f29\u5c0f\uff09<\/td><td>\u4ec5\u4e3a **7%**\uff0813.7\u00d7 \u7f29\u5c0f\uff09<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4f7f\u5f97 DeepSeek-V4 \u80fd\u591f<strong>\u5e38\u89c4\u652f\u6301\u767e\u4e07 token \u7ea7\u4e0a\u4e0b\u6587<\/strong>\uff0c\u8ba9\u957f\u7a0b\u4efb\u52a1\u548c\u8fdb\u4e00\u6b65\u7684 test-time scaling \u66f4\u4e3a\u53ef\u884c\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"\ud83c\udfc6-\u6838\u5fc3\u6027\u80fd\u5bf9\u6bd4\uff08deepseek-v4-pro-max\uff09\">\ud83c\udfc6 \u6838\u5fc3\u6027\u80fd\u5bf9\u6bd4\uff08DeepSeek-V4-Pro-Max\uff09<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728 DeepSeek-V4-Pro \u7684\u6700\u5927\u63a8\u7406\u52aa\u529b\u6a21\u5f0f\uff08Pro-Max\uff09\u4e0b\uff0c\u4e0e\u9876\u7ea7\u6a21\u578b\u7684\u5173\u952e\u5bf9\u6bd4\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>\u57fa\u51c6\u6d4b\u8bd5<\/th><th>DS-V4-Pro-Max<\/th><th>Claude-Opus-4.6-Max<\/th><th>GPT-5.4-xHigh<\/th><th>Gemini-3.1-Pro-High<\/th><\/tr><\/thead><tbody><tr><td>SimpleQA Verified<\/td><td><strong>57.9<\/strong><\/td><td>46.2<\/td><td>45.3<\/td><td><strong>75.6<\/strong> \ud83e\udd47<\/td><\/tr><tr><td>HLE<\/td><td>37.7<\/td><td>40.0<\/td><td>39.8<\/td><td><strong>44.4<\/strong> \ud83e\udd47<\/td><\/tr><tr><td>Apex Shortlist<\/td><td><strong>90.2<\/strong> \ud83e\udd47<\/td><td>85.9<\/td><td>78.1<\/td><td>89.1<\/td><\/tr><tr><td>Codeforces (Rating)<\/td><td><strong>3206<\/strong> \ud83e\udd47<\/td><td>\u2014<\/td><td>3168<\/td><td>3052<\/td><\/tr><tr><td>SWE Verified (Resolved)<\/td><td>80.6<\/td><td><strong>80.8<\/strong> \ud83e\udd47<\/td><td>\u2014<\/td><td>80.6<\/td><\/tr><tr><td>Terminal-Bench 2.0<\/td><td>67.9<\/td><td>65.4<\/td><td><strong>75.1<\/strong> \ud83e\udd47<\/td><td>68.5<\/td><\/tr><tr><td>Toolathlon<\/td><td>51.8<\/td><td>47.2<\/td><td><strong>54.6<\/strong> \ud83e\udd47<\/td><td>48.8<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"1-\u5f15\u8a00\">1. \u5f15\u8a00<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u63a8\u7406\u6a21\u578b\uff08DeepSeek-R1, OpenAI o \u7cfb\u5217\uff09\u7684\u51fa\u73b0\u5efa\u7acb\u4e86 test-time scaling \u7684\u65b0\u8303\u5f0f\uff0c\u663e\u8457\u63d0\u5347\u4e86\u5927\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u7684\u6027\u80fd\u3002\u7136\u800c\uff0c\u8fd9\u4e00\u6269\u5c55\u8303\u5f0f\u4ece\u6839\u672c\u4e0a\u53d7\u5236\u4e8e vanilla attention \u673a\u5236\u7684<strong>\u4e8c\u6b21\u8ba1\u7b97\u590d\u6742\u5ea6<\/strong>\uff0c\u4f7f\u8d85\u957f\u4e0a\u4e0b\u6587\u548c\u63a8\u7406\u8fc7\u7a0b\u6210\u4e3a\u96be\u4ee5\u903e\u8d8a\u7684\u74f6\u9888\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e0e\u6b64\u540c\u65f6\uff0c\u957f\u7a0b\u4efb\u52a1\uff08\u4ece\u590d\u6742 agent \u5de5\u4f5c\u6d41\u5230\u5927\u89c4\u6a21\u8de8\u6587\u6863\u5206\u6790\uff09\u7684\u5174\u8d77\u4e5f\u8ba9\u5bf9\u8d85\u957f\u4e0a\u4e0b\u6587\u7684\u9ad8\u6548\u652f\u6301\u6210\u4e3a\u672a\u6765\u53d1\u5c55\u7684\u5173\u952e\u3002\u5c3d\u7ba1\u8fd1\u671f\u5f00\u6e90\u5de5\u4f5c\uff08Kimi\u3001DeepSeek-V3\u3001MiniMax\u3001Qwen \u7b49\uff09\u5728\u901a\u7528\u80fd\u529b\u4e0a\u6709\u6240\u63a8\u8fdb\uff0c\u4f46\u5728\u5904\u7406\u8d85\u957f\u5e8f\u5217\u65b9\u9762\u7684\u6838\u5fc3\u67b6\u6784\u6548\u7387\u95ee\u9898\u4ecd\u662f\u4e3b\u8981\u969c\u788d\uff0c\u9650\u5236\u4e86 test-time scaling \u7684\u8fdb\u4e00\u6b65\u6536\u76ca\uff0c\u5e76\u963b\u788d\u4e86\u5bf9\u957f\u7a0b\u4efb\u52a1\u7684\u6df1\u5165\u63a2\u7d22\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"11-deepseek-v4-\u7cfb\u5217\u7684\u67b6\u6784\u521b\u65b0\">1.1 DeepSeek-V4 \u7cfb\u5217\u7684\u67b6\u6784\u521b\u65b0<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e3a\u6253\u7834\u8d85\u957f\u4e0a\u4e0b\u6587\u7684\u6548\u7387\u74f6\u9888\uff0c\u6211\u4eec\u5f00\u53d1\u4e86 DeepSeek-V4 \u7cfb\u5217\u3002\u4e0e DeepSeek-V3 \u76f8\u6bd4\uff0cV4 \u4fdd\u7559\u4e86\u4ee5\u4e0b\u8bbe\u8ba1\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>DeepSeekMoE<\/strong> \u6846\u67b6<\/li>\n\n\n\n<li><strong>Multi-Token Prediction\uff08MTP\uff09<\/strong> \u7b56\u7565<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u540c\u65f6\u5f15\u5165\u4e86\u4ee5\u4e0b\u5173\u952e\u521b\u65b0\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u6df7\u5408\u6ce8\u610f\u529b\u673a\u5236\uff08CSA + HCA\uff09<\/strong>\uff1aCSA \u5728\u5e8f\u5217\u7ef4\u5ea6\u4e0a\u538b\u7f29 KV cache\uff0c\u518d\u6267\u884c DeepSeek Sparse Attention\uff08DSA\uff09\uff1bHCA \u5219\u5bf9 KV cache \u8fdb\u884c\u66f4\u6fc0\u8fdb\u7684\u538b\u7f29\uff0c\u4f46\u4fdd\u6301 dense attention\u3002<\/li>\n\n\n\n<li><strong>Manifold-Constrained Hyper-Connections\uff08mHC\uff09<\/strong>\uff1a\u5347\u7ea7\u4f20\u7edf\u7684\u6b8b\u5dee\u8fde\u63a5\u3002<\/li>\n\n\n\n<li><strong>Muon \u4f18\u5316\u5668<\/strong>\uff1a\u63d0\u5347\u8bad\u7ec3\u6536\u655b\u901f\u5ea6\u548c\u7a33\u5b9a\u6027\u3002<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"12-\u57fa\u7840\u8bbe\u65bd\u4f18\u5316\">1.2 \u57fa\u7840\u8bbe\u65bd\u4f18\u5316<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e3a\u652f\u6301 V4 \u7684\u9ad8\u6548\u8bad\u7ec3\u3001\u63a8\u7406\u548c\u4ea7\u54c1\u5316\u5f00\u53d1\uff0c\u6211\u4eec\u5f15\u5165\u4e86\u591a\u9879\u57fa\u7840\u8bbe\u65bd\u4f18\u5316\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>MoE \u6a21\u5757\u5355\u878d\u5408\u5185\u6838<\/strong>\uff1a\u5728\u5355\u4e2a kernel \u4e2d\u5b8c\u5168\u91cd\u53e0\u8ba1\u7b97\u3001\u901a\u4fe1\u548c\u5185\u5b58\u8bbf\u95ee\u3002<\/li>\n\n\n\n<li><strong>TileLang<\/strong>\uff1a\u4f7f\u7528\u9886\u57df\u4e13\u7528\u8bed\u8a00\uff08DSL\uff09\u5e73\u8861\u5f00\u53d1\u6548\u7387\u548c\u8fd0\u884c\u65f6\u6027\u80fd\u3002<\/li>\n\n\n\n<li><strong>\u6279\u6b21\u4e0d\u53d8\u3001\u786e\u5b9a\u6027\u5185\u6838\u5e93<\/strong>\uff1a\u786e\u4fdd\u8bad\u7ec3\u548c\u63a8\u7406\u4e4b\u95f4\u7684 bitwise \u53ef\u590d\u73b0\u6027\u3002<\/li>\n\n\n\n<li><strong>FP4 \u91cf\u5316\u611f\u77e5\u8bad\u7ec3\uff08QAT\uff09<\/strong>\uff1a\u9488\u5bf9 MoE \u4e13\u5bb6\u6743\u91cd\u548c indexer QK \u8def\u5f84\u3002<\/li>\n\n\n\n<li><strong>\u8bad\u7ec3\u6846\u67b6\u6269\u5c55<\/strong>\uff1a\u5305\u62ec Muon \u4f18\u5316\u5668\u7684\u6df7\u5408 ZeRO \u7b56\u7565\u3001mHC \u7684 cost-effective \u5b9e\u73b0\u3001\u4e24\u9636\u6bb5\u4e0a\u4e0b\u6587\u5e76\u884c\u3002<\/li>\n\n\n\n<li><strong>\u63a8\u7406\u6846\u67b6<\/strong>\uff1a\u5f02\u6784 KV cache \u7ed3\u6784\u4e0e\u78c1\u76d8\u5b58\u50a8\u7b56\u7565\uff0c\u5b9e\u73b0\u9ad8\u6548\u7684 shared-prefix \u590d\u7528\u3002<\/li>\n<\/ol>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"13-\u6838\u5fc3\u8bc4\u4f30\u7ed3\u679c\u603b\u7ed3\">1.3 \u6838\u5fc3\u8bc4\u4f30\u7ed3\u679c\u603b\u7ed3<\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u77e5\u8bc6\u80fd\u529b<\/strong>\uff1aDeepSeek-V4-Pro-Max \u5728 SimpleQA \u548c Chinese-SimpleQA \u4e0a\u663e\u8457\u8d85\u8d8a\u9886\u5148\u5f00\u6e90\u6a21\u578b\u3002\u5728 MMLU-Pro\u3001HLE\u3001GPQA \u7b49\u6559\u80b2\u77e5\u8bc6\u7c7b\u57fa\u51c6\u4e0a\u76f8\u5bf9\u5f00\u6e90\u5bf9\u624b\u7565\u6709\u9886\u5148\u3002\u867d\u4ecd\u843d\u540e\u4e8e Gemini-3.1-Pro\uff0c\u4f46\u5dee\u8ddd\u5df2\u5927\u5e45\u7f29\u5c0f\u3002<\/li>\n\n\n\n<li><strong>\u63a8\u7406\u80fd\u529b<\/strong>\uff1a\u901a\u8fc7\u6269\u5c55\u63a8\u7406 token\uff0cV4-Pro-Max \u5728\u6807\u51c6\u63a8\u7406\u57fa\u51c6\u4e0a\u8d85\u8d8a GPT-5.2 \u548c Gemini-3.0-Pro\uff0c\u4f46\u7565\u900a\u4e8e GPT-5.4 \u548c Gemini-3.1-Pro\uff0c\u6574\u4f53\u53d1\u5c55\u8f68\u8ff9\u6ede\u540e\u524d\u6cbf\u6a21\u578b\u7ea6 3-6 \u4e2a\u6708\u3002V4-Flash-Max \u5728\u590d\u6742\u63a8\u7406\u4efb\u52a1\u4e0a\u8fbe\u5230\u4e86\u4e0e GPT-5.2 \u548c Gemini-3.0-Pro \u76f8\u5f53\u7684\u6027\u80fd\uff0c\u662f\u4e00\u6b3e\u9ad8\u6027\u4ef7\u6bd4\u7684\u67b6\u6784\u3002<\/li>\n\n\n\n<li><strong>Agent \u80fd\u529b<\/strong>\uff1aV4-Pro-Max \u5728\u516c\u5f00\u57fa\u51c6\u4e0a\u4e0e Kimi-K2.6\u3001GLM-5.1 \u7b49\u9886\u5148\u5f00\u6e90\u6a21\u578b\u6301\u5e73\uff0c\u4f46\u7565\u900a\u4e8e\u524d\u6cbf\u95ed\u6e90\u6a21\u578b\u3002<strong>\u5728\u5185\u90e8\u8bc4\u6d4b\u4e2d\uff0cV4-Pro-Max \u8d85\u8d8a\u4e86 Claude Sonnet 4.5\uff0c\u63a5\u8fd1 Opus 4.5 \u6c34\u5e73<\/strong>\u3002<\/li>\n\n\n\n<li><strong>\u957f\u4e0a\u4e0b\u6587<\/strong>\uff1aV4-Pro-Max \u5728 1M \u4e0a\u4e0b\u6587\u7a97\u53e3\u7684\u5408\u6210\u548c\u771f\u5b9e\u7528\u4f8b\u4e2d\u8868\u73b0\u5f3a\u52b2\uff0c\u5728\u5b66\u672f\u57fa\u51c6\u4e0a<strong>\u751a\u81f3\u8d85\u8d8a\u4e86 Gemini-3.1-Pro<\/strong>\u3002<\/li>\n\n\n\n<li><strong>Pro vs Flash<\/strong>\uff1aFlash-Max \u56e0\u53c2\u6570\u89c4\u6a21\u8f83\u5c0f\uff0c\u5728\u77e5\u8bc6\u8bc4\u4f30\u4e2d\u8868\u73b0\u7a0d\u5f31\uff1b\u4f46\u7ed9\u4e88\u66f4\u5927\u601d\u8003\u9884\u7b97\u65f6\uff0c\u5728\u63a8\u7406\u4efb\u52a1\u4e0a\u53ef\u8fbe\u5230\u76f8\u5f53\u6c34\u5e73\u3002\u5728 agent \u8bc4\u6d4b\u4e2d\uff0cFlash-Max \u5728\u90e8\u5206\u57fa\u51c6\u4e0a\u4e0e Pro-Max \u76f8\u5f53\uff0c\u4f46\u5728\u9ad8\u96be\u5ea6\u4efb\u52a1\u4e0a\u4ecd\u843d\u540e\u3002<\/li>\n<\/ul>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"2-\u67b6\u6784\">2. \u67b6\u6784<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">DeepSeek-V4 \u7cfb\u5217\u6574\u4f53\u4e0a\u4fdd\u7559\u4e86 Transformer \u67b6\u6784\u548c MTP \u6a21\u5757\u3002\u76f8\u6bd4 DeepSeek-V3\uff0c\u5173\u952e\u5347\u7ea7\u6709\u4e09\u70b9\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li>\u5f15\u5165 <strong>mHC<\/strong>\uff08Manifold-Constrained Hyper-Connections\uff09\u5f3a\u5316\u6b8b\u5dee\u8fde\u63a5\uff1b<\/li>\n\n\n\n<li>\u8bbe\u8ba1<strong>\u6df7\u5408\u6ce8\u610f\u529b\u67b6\u6784<\/strong>\uff08CSA + HCA\uff09\uff0c\u5927\u5e45\u63d0\u5347\u957f\u4e0a\u4e0b\u6587\u6548\u7387\uff1b<\/li>\n\n\n\n<li>\u91c7\u7528 <strong>Muon<\/strong> \u4f18\u5316\u5668\u3002<\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">MoE \u7ec4\u4ef6\u4ecd\u6cbf\u7528 DeepSeekMoE \u67b6\u6784\uff08\u4ec5\u5bf9 V3 \u505a\u4e86\u5c0f\u5e45\u8c03\u6574\uff09\u3002MTP \u914d\u7f6e\u4e0e V3 \u76f8\u540c\u3002\u5176\u4ed6\u672a\u660e\u786e\u8bf4\u660e\u7684\u7ec6\u8282\u5747\u9075\u5faa V3 \u7684\u8bbe\u7f6e\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"21-\u7ee7\u627f\u81ea-deepseek-v3-\u7684\u8bbe\u8ba1\">2.1 \u7ee7\u627f\u81ea DeepSeek-V3 \u7684\u8bbe\u8ba1<\/h3>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"mixture-of-experts\uff08moe\uff09\">Mixture-of-Experts\uff08MoE\uff09<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">V4 \u7cfb\u5217\u91c7\u7528 DeepSeekMoE \u8303\u5f0f\uff0c\u8bbe\u7f6e<strong>\u7ec6\u7c92\u5ea6\u8def\u7531\u4e13\u5bb6 + \u5171\u4eab\u4e13\u5bb6<\/strong>\u3002\u4e0e V3 \u7684\u5dee\u5f02\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u6fc0\u6d3b\u51fd\u6570\u53d8\u66f4<\/strong>\uff1a\u4ece <code>Sigmoid(\u00b7)<\/code> \u6539\u4e3a <code>Sqrt(Softplus(\u00b7))<\/code> \u6765\u8ba1\u7b97\u4eb2\u548c\u5ea6\u5f97\u5206\u3002<\/li>\n\n\n\n<li><strong>\u8d1f\u8f7d\u5747\u8861<\/strong>\uff1a\u91c7\u7528 auxiliary-loss-free \u7b56\u7565\uff0c\u8f85\u4ee5\u5e8f\u5217\u7ea7\u5e73\u8861\u635f\u5931\uff0c\u9632\u6b62\u5355\u5e8f\u5217\u5185\u6781\u7aef\u4e0d\u5747\u8861\u3002<\/li>\n\n\n\n<li><strong>\u8def\u7531\u8282\u70b9\u9650\u5236\u79fb\u9664<\/strong>\uff1aV4 \u79fb\u9664\u4e86\u8def\u7531\u76ee\u6807\u8282\u70b9\u6570\u91cf\u9650\u5236\uff0c\u5e76\u91cd\u65b0\u8bbe\u8ba1\u4e86\u5e76\u884c\u7b56\u7565\u4ee5\u7ef4\u6301\u8bad\u7ec3\u6548\u7387\u3002<\/li>\n\n\n\n<li><strong>Hash \u8def\u7531<\/strong>\uff1a\u521d\u59cb\u51e0\u4e2a Transformer block \u4e2d\u7684 dense FFN \u5c42\u88ab\u66ff\u6362\u4e3a\u91c7\u7528 Hash \u8def\u7531\u7684 MoE \u5c42\uff0c\u8be5\u7b56\u7565\u6839\u636e token ID \u7684\u54c8\u5e0c\u51fd\u6570\u51b3\u5b9a\u76ee\u6807\u4e13\u5bb6\u3002<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"multi-token-prediction\uff08mtp\uff09\">Multi-Token Prediction\uff08MTP\uff09<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">\u6cbf\u7528 V3 \u7b56\u7565\uff0c\u672a\u505a\u4fee\u6539\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"22-manifold-constrained-hyper-connections\uff08mhc\uff09\">2.2 Manifold-Constrained Hyper-Connections\uff08mHC\uff09<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">mHC \u662f\u5bf9\u4f20\u7edf\u6b8b\u5dee\u8fde\u63a5\u7684\u5f3a\u5316\u3002\u76f8\u6bd4\u6734\u7d20 Hyper-Connections\uff08HC\uff09\uff0c\u5176<strong>\u6838\u5fc3\u601d\u60f3\u662f\u5c06\u6b8b\u5dee\u6620\u5c04\u7ea6\u675f\u5728\u7279\u5b9a\u6d41\u5f62\u4e0a<\/strong>\uff0c\u4ece\u800c\u5728\u4fdd\u6301\u6a21\u578b\u8868\u8fbe\u80fd\u529b\u7684\u540c\u65f6\u63d0\u5347\u8de8\u5c42\u4fe1\u53f7\u4f20\u64ad\u7684\u7a33\u5b9a\u6027\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"\u6807\u51c6-hyper-connections\uff08hc\uff09\">\u6807\u51c6 Hyper-Connections\uff08HC\uff09<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">HC \u5c06\u6b8b\u5dee\u6d41\u7684\u5bbd\u5ea6\u6269\u5c55 $n_{hc}$ \u500d\uff0c\u4ece $\\mathbb{R}^d$ \u6269\u5c55\u4e3a $\\mathbb{R}^{n_{hc} \\times d}$\u3002\u5b83\u5f15\u5165\u4e09\u4e2a\u7ebf\u6027\u6620\u5c04\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u8f93\u5165\u6620\u5c04 $A_l \\in \\mathbb{R}^{1 \\times n_{hc}}$<\/li>\n\n\n\n<li>\u6b8b\u5dee\u53d8\u6362 $B_l \\in \\mathbb{R}^{n_{hc} \\times n_{hc}}$<\/li>\n\n\n\n<li>\u8f93\u51fa\u6620\u5c04 $C_l \\in \\mathbb{R}^{n_{hc} \\times 1}$<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u6b8b\u5dee\u72b6\u6001\u66f4\u65b0\u516c\u5f0f\uff1a<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">$$X_{l+1} = B_l X_l + C_l F_l(A_l X_l)$$<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5176\u4e2d $F_l$ \u8868\u793a\u7b2c $l$ \u5c42\uff08\u5982 MoE \u5c42\uff09\u3002HC \u89e3\u8026\u4e86\u6b8b\u5dee\u5bbd\u5ea6\u4e0e\u5b9e\u9645 hidden size\uff0c\u63d0\u4f9b\u4e86\u4e00\u4e2a\u4e92\u8865\u7684\u6269\u5c55\u7ef4\u5ea6\uff0c\u8ba1\u7b97\u5f00\u9500\u5f88\u5c0f\u3002\u4f46\u5806\u53e0\u591a\u5c42\u540e\uff0c\u8bad\u7ec3\u5e38\u51fa\u73b0\u6570\u503c\u4e0d\u7a33\u5b9a\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"\u6d41\u5f62\u7ea6\u675f\u6b8b\u5dee\u6620\u5c04\">\u6d41\u5f62\u7ea6\u675f\u6b8b\u5dee\u6620\u5c04<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">mHC \u7684\u6838\u5fc3\u521b\u65b0\u662f\u5c06\u6b8b\u5dee\u6620\u5c04\u77e9\u9635 $B_l$ \u7ea6\u675f\u5728<strong>\u53cc\u968f\u673a\u77e9\u9635\u6d41\u5f62<\/strong>\uff08Birkhoff \u591a\u9762\u4f53\uff09\u4e0a\uff1a<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">$$B_l \\in \\mathcal{M} := {M \\in \\mathbb{R}^{n \\times n} | M\\mathbf{1}_n = \\mathbf{1}_n, \\mathbf{1}_n^T M = \\mathbf{1}_n^T, M \\geq 0}$$<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u5173\u952e\u6027\u8d28<\/strong>\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u8c31\u8303\u6570 $|B_l|_2$ \u88ab\u7ea6\u675f \u2264 1\uff0c\u4fdd\u8bc1\u6b8b\u5dee\u53d8\u6362\u975e\u6269\u5f20\uff08non-expansive\uff09\uff0c\u63d0\u5347\u524d\u5411\u4f20\u64ad\u548c\u53cd\u5411\u4f20\u64ad\u7684\u6570\u503c\u7a33\u5b9a\u6027\u3002<\/li>\n\n\n\n<li>\u96c6\u5408 $\\mathcal{M}$ \u5728\u4e58\u6cd5\u4e0b\u5c01\u95ed\uff0c\u4fdd\u8bc1\u6df1\u5c42\u5806\u53e0\u7684\u7a33\u5b9a\u6027\u3002<\/li>\n\n\n\n<li>\u8f93\u5165\u53d8\u6362 $A_l$ \u548c\u8f93\u51fa\u53d8\u6362 $C_l$ \u901a\u8fc7 Sigmoid \u7ea6\u675f\u4e3a<strong>\u975e\u8d1f\u4e14\u6709\u754c<\/strong>\uff0c\u907f\u514d\u4fe1\u53f7\u62b5\u6d88\u98ce\u9669\u3002<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"\u52a8\u6001\u53c2\u6570\u5316\">\u52a8\u6001\u53c2\u6570\u5316<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e09\u4e2a\u7ebf\u6027\u6620\u5c04\u7684\u53c2\u6570\u52a8\u6001\u751f\u6210\uff0c\u5206\u89e3\u4e3a<strong>\u52a8\u6001\uff08\u4f9d\u8d56\u8f93\u5165\uff09+ \u9759\u6001\uff08\u4e0d\u4f9d\u8d56\u8f93\u5165\uff09<\/strong> \u4e24\u4e2a\u5206\u91cf\u3002\u7ed9\u5b9a\u8f93\u5165 $X_l$\uff0c\u5148\u6241\u5e73\u5316\u5e76\u5f52\u4e00\u5316\uff1a$\\hat{X}_l = \\text{RMSNorm}(\\text{vec}(X_l))$\u3002\u7136\u540e\u751f\u6210\u672a\u7ea6\u675f\u7684\u539f\u59cb\u53c2\u6570\uff1a<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">$$\\tilde{A}_l = \\alpha_l^{pre} \\cdot (\\hat{X}_l W_l^{pre}) + S_l^{pre}$$\n$$\\tilde{B}_l = \\alpha_l^{res} \\cdot \\text{Mat}(\\hat{X}_l W_l^{res}) + S_l^{res}$$\n$$\\tilde{C}_l = \\alpha_l^{post} \\cdot (\\hat{X}_l W_l^{post})^T + S_l^{post}$$<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5176\u4e2d $\\alpha_l^{pre}, \\alpha_l^{res}, \\alpha_l^{post}$ \u662f\u521d\u59cb\u5316\u4e3a\u5c0f\u503c\u7684\u53ef\u5b66\u4e60\u95e8\u63a7\u56e0\u5b50\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"\u53c2\u6570\u7ea6\u675f\u5e94\u7528\">\u53c2\u6570\u7ea6\u675f\u5e94\u7528<\/h4>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u8f93\u5165\/\u8f93\u51fa\u6620\u5c04\uff1a\u4f7f\u7528 Sigmoid \u4fdd\u8bc1\u975e\u8d1f\u6709\u754c\uff1a$A_l = \\sigma(\\tilde{A}_l)$\uff0c$C_l = 2\\sigma(\\tilde{C}_l)$\u3002<\/li>\n\n\n\n<li>\u6b8b\u5dee\u6620\u5c04\uff1a\u901a\u8fc7 <strong>Sinkhorn-Knopp \u7b97\u6cd5<\/strong>\u6295\u5f71\u5230\u53cc\u968f\u673a\u77e9\u9635\u6d41\u5f62\u3002\u5148\u5e94\u7528\u6307\u6570\u51fd\u6570\u4fdd\u8bc1\u6b63\u6027 $M^{(0)} = \\exp(\\tilde{B}_l)$\uff0c\u7136\u540e\u8fed\u4ee3\u6267\u884c\u884c\u5217\u5f52\u4e00\u5316\uff1a<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">$$M^{(t)} = T_r(T_c(M^{(t-1)}))$$<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5b9e\u8df5\u4e2d\u9009\u62e9 $t_{\\max} = 20$\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"23-\u6df7\u5408\u6ce8\u610f\u529b\uff1acsa-\u4e0e-hca\">2.3 \u6df7\u5408\u6ce8\u610f\u529b\uff1aCSA \u4e0e HCA<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5f53\u4e0a\u4e0b\u6587\u957f\u5ea6\u8fbe\u5230\u6781\u7aef\u89c4\u6a21\u65f6\uff0c\u6ce8\u610f\u529b\u673a\u5236\u6210\u4e3a\u4e3b\u8981\u8ba1\u7b97\u74f6\u9888\u3002V4 \u8bbe\u8ba1\u4e86\u4e24\u79cd\u9ad8\u6548\u6ce8\u610f\u529b\u67b6\u6784\u5e76\u4ea4\u9519\u6df7\u5408\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>CSA\uff08Compressed Sparse Attention\uff09<\/strong>\uff1a\u5148\u5c06\u6bcf $m$ \u4e2a token \u7684 KV cache \u538b\u7f29\u4e3a 1 \u4e2a\u6761\u76ee\uff0c\u518d\u5e94\u7528 DSA\uff08DeepSeek Sparse Attention\uff09\uff0c\u8ba9\u6bcf\u4e2a query token \u53ea\u5173\u6ce8 $k$ \u4e2a\u538b\u7f29 KV \u6761\u76ee\u3002<\/li>\n\n\n\n<li><strong>HCA\uff08Heavily Compressed Attention\uff09<\/strong>\uff1a\u66f4\u6fc0\u8fdb\u7684\u538b\u7f29\uff0c\u5c06\u6bcf $m'$\uff08$\\gg m$\uff09\u4e2a token \u7684 KV cache \u5408\u5e76\u4e3a 1 \u4e2a\u6761\u76ee\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u79cd\u6df7\u5408\u67b6\u6784\u663e\u8457\u6539\u5584\u4e86\u957f\u4e0a\u4e0b\u6587\u6548\u7387\uff0c\u4f7f\u5f97 1M token \u4e0a\u4e0b\u6587\u5728\u5b9e\u8df5\u4e2d\u53ef\u884c\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"231-compressed-sparse-attention\uff08csa\uff09\">2.3.1 Compressed Sparse Attention\uff08CSA\uff09<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u538b\u7f29 KV \u6761\u76ee<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7ed9\u5b9a\u8f93\u5165 hidden states $H \\in \\mathbb{R}^{n \\times d}$\uff0cCSA \u9996\u5148\u8ba1\u7b97\u4e24\u7ec4 KV \u6761\u76ee\u548c\u5bf9\u5e94\u7684\u538b\u7f29\u6743\u91cd\uff1a<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">$$C^a = H \\cdot W^{aKV}, \\quad C^b = H \\cdot W^{bKV}$$\n$$Z^a = H \\cdot W^{aZ}, \\quad Z^b = H \\cdot W^{bZ}$$<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7136\u540e\u6bcf $m$ \u4e2a KV \u6761\u76ee\u538b\u7f29\u4e3a\u4e00\u4e2a\uff1a<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">$$[S^a_{mi:m(i+1)-1}; S^b_{m(i-1):mi-1}] = \\text{Softmax}<em>{row}([Z^a<\/em>{mi:m(i+1)-1} + B^a; Z^b_{m(i-1):mi-1} + B^b])$$<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">$$C_i^{\\text{Comp}} = \\sum_{j=mi}^{m(i+1)-1} S^a_j \\odot C^a_j + \\sum_{j=m(i-1)}^{mi-1} S^b_j \\odot C^b_j$$<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6ce8\u610f\uff1a\u6bcf\u4e2a $C_i^{\\text{Comp}}$ \u6d3e\u751f\u81ea $2m$ \u4e2a KV \u6761\u76ee\uff0c\u4f46\u5b58\u5728\u91cd\u53e0\uff0c\u56e0\u6b64 CSA \u5b9e\u9645\u5c06\u5e8f\u5217\u957f\u5ea6\u538b\u7f29\u4e3a $1\/m$\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Lightning Indexer \u7a00\u758f\u9009\u62e9<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u83b7\u5f97\u538b\u7f29 KV \u6761\u76ee\u540e\uff0cCSA \u5e94\u7528 DSA \u7b56\u7565\u9009\u62e9 top-k \u6761\u76ee\u8fdb\u884c\u6838\u5fc3\u6ce8\u610f\u529b\u3002\u5148\u751f\u6210\u538b\u7f29 indexer keys $K^{\\text{IComp}} \\in \\mathbb{R}^{\\frac{n}{m} \\times c^I}$\u3002\u5bf9\u4e8e query token $t$\uff0c\u4ee5\u4f4e\u79e9\u65b9\u5f0f\u751f\u6210 indexer queries\uff1a<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">$$c_t^Q = h_t \\cdot W^{DQ}$$\n$$[q_{t,1}^I; \\ldots; q_{t,n_h^I}^I] = q_t^I = c_t^Q \\cdot W^{IUQ}$$<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7d22\u5f15\u5f97\u5206 $I_{t,s}$ \u5b9a\u4e49\u4e3a\uff1a<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">$$I_{t,s} = \\sum_{h=1}^{n_h^I} w_{t,h}^I \\cdot \\text{ReLU}(q_{t,h}^I \\cdot K_s^{\\text{IComp}})$$<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f7f\u7528 top-k selector \u4fdd\u7559\u7a00\u758f KV \u6761\u76ee\u5b50\u96c6\u7528\u4e8e\u6838\u5fc3\u6ce8\u610f\u529b\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u5171\u4eab KV \u591a\u67e5\u8be2\u6ce8\u610f\u529b\uff08Shared KV MQA\uff09<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u9009\u5b9a\u7a00\u758f KV \u6761\u76ee\u540e\uff0cCSA \u4ee5 MQA \u65b9\u5f0f\u6267\u884c\u6838\u5fc3\u6ce8\u610f\u529b\uff0c\u6bcf\u4e2a\u538b\u7f29 KV \u6761\u76ee\u540c\u65f6\u4f5c\u4e3a key \u548c value\u3002query \u4e0e indexer query \u5171\u4eab\u6f5c\u5728\u5411\u91cf $c_t^Q$\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u5206\u7ec4\u8f93\u51fa\u6295\u5f71\uff08Grouped Output Projection\uff09<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">V4 \u4e2d $c \\cdot n_h$ \u8f83\u5927\uff0c\u76f4\u63a5\u6295\u5f71\u5230 $d$ \u7ef4\u4f1a\u5e26\u6765\u5de8\u5927\u8ba1\u7b97\u5f00\u9500\u3002\u8bbe\u8ba1\u5206\u7ec4\u6295\u5f71\u7b56\u7565\uff1a\u5148\u5c06 $n_h$ \u4e2a\u8f93\u51fa\u5206\u4e3a $g$ \u7ec4\uff0c\u6bcf\u7ec4\u6295\u5f71\u5230 $d_g$ \u7ef4\u4e2d\u95f4\u8f93\u51fa\uff0c\u6700\u540e\u518d\u6295\u5f71\u5230\u6700\u7ec8 $d$ \u7ef4\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"232-heavily-compressed-attention\uff08hca\uff09\">2.3.2 Heavily Compressed Attention\uff08HCA\uff09<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">HCA \u91c7\u7528\u66f4\u5927\u7684\u538b\u7f29\u7387 $m' \\gg m$\uff0c\u4e14<strong>\u4e0d\u6267\u884c\u91cd\u53e0\u538b\u7f29<\/strong>\uff1a<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">$$C = H \\cdot W^{KV}, \\quad Z = H \\cdot W^Z$$\n$$S_{m'i:m'(i+1)-1} = \\text{Softmax}<em>{row}(Z<\/em>{m'i:m'(i+1)-1} + B)$$\n$$C_i^{\\text{Comp}} = \\sum_{j=m'i}^{m'(i+1)-1} S_j \\odot C_j$$<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">HCA \u5c06\u5e8f\u5217\u957f\u5ea6\u538b\u7f29\u4e3a $1\/m'$\u3002\u540c\u6837\u91c7\u7528\u5171\u4eab KV MQA \u548c\u5206\u7ec4\u8f93\u51fa\u6295\u5f71\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"233-\u5176\u4ed6\u5173\u952e\u7ec6\u8282\">2.3.3 \u5176\u4ed6\u5173\u952e\u7ec6\u8282<\/h4>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Query \u4e0e KV \u5f52\u4e00\u5316<\/strong>\uff1a\u5728\u6838\u5fc3\u6ce8\u610f\u529b\u4e4b\u524d\uff0c\u5bf9 queries \u7684\u6bcf\u4e2a head \u548c\u538b\u7f29 KV \u6761\u76ee\u7684\u5355\u4e00 head \u989d\u5916\u6267\u884c RMSNorm\uff0c\u907f\u514d\u6ce8\u610f\u529b logits \u7206\u70b8\uff0c\u63d0\u5347\u8bad\u7ec3\u7a33\u5b9a\u6027\u3002<\/li>\n\n\n\n<li><strong>\u90e8\u5206\u65cb\u8f6c\u4f4d\u7f6e\u7f16\u7801\uff08Partial RoPE\uff09<\/strong>\uff1a\u5bf9 query \u548c KV \u6761\u76ee\u5411\u91cf\u7684<strong>\u6700\u540e 64 \u7ef4<\/strong>\u5e94\u7528 RoPE\u3002\u7531\u4e8e KV \u6761\u76ee\u540c\u65f6\u4f5c\u4e3a key \u548c value\uff0c\u6838\u5fc3\u6ce8\u610f\u529b\u8f93\u51fa\u4f1a\u643a\u5e26\u7edd\u5bf9\u4f4d\u7f6e\u5d4c\u5165\uff08\u6765\u81ea KV \u6761\u76ee\u7684\u52a0\u6743\u548c\uff09\u3002\u4f5c\u4e3a\u5e94\u5bf9\uff0c\u5bf9\u6bcf\u4e2a $o_{t,i}$ \u7684\u6700\u540e 64 \u7ef4\u4e5f\u5e94\u7528\u4f4d\u7f6e\u4e3a $-i$ \u7684 RoPE\uff0c\u4f7f\u8f93\u51fa\u643a\u5e26\u76f8\u5bf9\u4f4d\u7f6e\u4fe1\u606f\u3002<\/li>\n\n\n\n<li><strong>\u6ed1\u52a8\u7a97\u53e3\u6ce8\u610f\u529b\u8865\u5145\u5206\u652f<\/strong>\uff1a\u4e3a\u4e25\u683c\u4fdd\u7559 CSA \u548c HCA \u7684\u56e0\u679c\u6027\uff0c\u6bcf\u4e2a query \u53ea\u5173\u6ce8\u4e4b\u524d\u7684\u538b\u7f29 KV \u5757\uff0c\u56e0\u6b64\u65e0\u6cd5\u8bbf\u95ee\u5176\u81ea\u8eab\u538b\u7f29\u5757\u5185\u7684\u5176\u4ed6 token \u4fe1\u606f\u3002\u800c\u6700\u8fd1\u7684 token \u5bf9 query \u901a\u5e38\u5177\u6709\u66f4\u9ad8\u76f8\u5173\u6027\u3002\u4e3a\u6b64\uff0cCSA \u548c HCA \u5747\u5f15\u5165\u6ed1\u52a8\u7a97\u53e3\u5206\u652f\uff0c\u4e3a\u6bcf\u4e2a query token \u989d\u5916\u4ea7\u751f $n_{\\text{win}}$ \u4e2a\u672a\u538b\u7f29 KV \u6761\u76ee\u3002<\/li>\n\n\n\n<li><strong>Attention Sink<\/strong>\uff1a\u8bbe\u7f6e\u53ef\u5b66\u4e60\u7684 sink logits ${z_1', \\ldots, z_{n_h}'}$\uff0c\u5176\u6307\u6570\u9879\u52a0\u5230\u6ce8\u610f\u529b\u5f97\u5206\u7684\u5206\u6bcd\u4e2d\uff1a<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">$$s_{h,i,j} = \\frac{\\exp(z_{h,i,j})}{\\sum_k \\exp(z_{h,i,k}) + \\exp(z_h')}$$<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5141\u8bb8 query head \u8c03\u6574\u603b\u6ce8\u610f\u529b\u5f97\u5206\uff0c\u751a\u81f3\u63a5\u8fd1 0\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"234-\u6548\u7387\u8ba8\u8bba\">2.3.4 \u6548\u7387\u8ba8\u8bba<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">\u7531\u4e8e\u91c7\u7528\u6df7\u5408 CSA + HCA \u4ee5\u53ca\u4f4e\u7cbe\u5ea6\u8ba1\u7b97\u5b58\u50a8\uff0cV4 \u6ce8\u610f\u529b\u6a21\u5757\u5728\u957f\u4e0a\u4e0b\u6587\u573a\u666f\u4e2d\u5b9e\u73b0\u4e86\u5353\u8d8a\u7684\u6548\u7387\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u6df7\u5408\u5b58\u50a8\u683c\u5f0f<\/strong>\uff1aRoPE \u7ef4\u5ea6\u4f7f\u7528 BF16\uff0c\u5176\u4f59\u7ef4\u5ea6\u4f7f\u7528 FP8\uff0c\u76f8\u6bd4\u7eaf BF16 \u5b58\u50a8\uff0cKV cache \u5927\u5c0f\u51cf\u534a\u3002<\/li>\n\n\n\n<li><strong>Lightning Indexer \u4e2d\u7684 FP4<\/strong>\uff1a\u6ce8\u610f\u529b\u8ba1\u7b97\u5728 FP4 \u7cbe\u5ea6\u4e0b\u8fdb\u884c\uff0c\u52a0\u901f\u8d85\u957f\u4e0a\u4e0b\u6587\u4e0b\u7684\u6ce8\u610f\u529b\u64cd\u4f5c\u3002<\/li>\n\n\n\n<li><strong>\u66f4\u5c0f\u7684 top-k<\/strong>\uff1a\u76f8\u6bd4 DeepSeek-V3.2\uff0cV4 \u4f7f\u7528\u66f4\u5c0f\u7684 top-k\uff0c\u63d0\u5347\u77ed\u4e2d\u957f\u5ea6\u6587\u672c\u7684\u6548\u7387\u3002<\/li>\n\n\n\n<li><strong>\u6700\u91cd\u8981\u7684\u662f<\/strong>\uff1a\u538b\u7f29\u6ce8\u610f\u529b\u548c\u6df7\u5408\u6ce8\u610f\u529b\u6280\u672f\u5927\u5e45\u964d\u4f4e KV cache \u5927\u5c0f\u548c\u8ba1\u7b97 FLOPs\u3002<\/li>\n<\/ol>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\"><strong>\u5173\u952e\u5bf9\u6bd4<\/strong>\uff1a\u4ee5 BF16 GQA8\uff08\u5934\u7ef4\u5ea6 128\uff09\u4f5c\u4e3a\u57fa\u51c6\uff0cV4 \u7684 KV cache \u5927\u5c0f\u5728 1M \u4e0a\u4e0b\u6587\u4e0b\u53ef\u964d\u81f3\u57fa\u51c6\u7684\u7ea6 **2%**\u3002\u5373\u4f7f\u4e0e\u5df2\u7ecf\u9ad8\u6548\u7684 V3.2 \u76f8\u6bd4\uff0cV4 \u4ecd\u6709\u663e\u8457\u4f18\u52bf\uff08\u89c1 Figure 1\uff09\u3002<\/p>\n<\/blockquote>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"24-muon-\u4f18\u5316\u5668\">2.4 Muon \u4f18\u5316\u5668<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">V4 \u7cfb\u5217\u5bf9\u5927\u591a\u6570\u6a21\u5757\u4f7f\u7528 Muon \u4f18\u5316\u5668\uff0c\u56e0\u5176\u6536\u655b\u66f4\u5feb\u3001\u8bad\u7ec3\u66f4\u7a33\u5b9a\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u57fa\u672c\u914d\u7f6e<\/strong>\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>AdamW\uff1a\u7528\u4e8e embedding \u6a21\u5757\u3001\u9884\u6d4b\u5934\u6a21\u5757\u3001mHC \u7684\u9759\u6001\u504f\u7f6e\u548c\u95e8\u63a7\u56e0\u5b50\u3001\u6240\u6709 RMSNorm \u6a21\u5757\u7684\u6743\u91cd\u3002<\/li>\n\n\n\n<li>Muon\uff1a\u66f4\u65b0\u5176\u4ed6\u6240\u6709\u6a21\u5757\u3002\u5e94\u7528\u6743\u91cd\u8870\u51cf\uff0c\u4f7f\u7528 Nesterov \u6280\u5de7\uff0c\u5bf9\u66f4\u65b0\u77e9\u9635\u7684 RMS \u8fdb\u884c rescale \u4ee5\u590d\u7528 AdamW \u8d85\u53c2\u6570\u3002<strong>\u4e0e Liu et al. (2025) \u7684\u4e0d\u540c\u5728\u4e8e\u4f7f\u7528\u6df7\u5408 Newton-Schulz \u8fed\u4ee3\u8fdb\u884c\u6b63\u4ea4\u5316<\/strong>\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u6df7\u5408 Newton-Schulz \u8fed\u4ee3<\/strong>\uff1a<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Newton-Schulz \u8fed\u4ee3\u7684\u76ee\u6807\u662f\u5c06\u77e9\u9635 $M = U\\Sigma V^T$ \u8fd1\u4f3c\u6b63\u4ea4\u5316\u4e3a $UV^T$\u3002V4 \u7684\u6df7\u5408\u65b9\u6848\u6267\u884c <strong>10 \u6b21\u8fed\u4ee3\u3001\u5206\u4e24\u9636\u6bb5<\/strong>\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u524d 8 \u6b65\u4f7f\u7528 $(a, b, c) = (3.4445, -4.7750, 2.0315)$ \u9a71\u52a8\u5feb\u901f\u6536\u655b\u3002<\/li>\n\n\n\n<li>\u540e 2 \u6b65\u4f7f\u7528 $(a, b, c) = (2, -1.5, 0.5)$ \u7cbe\u786e\u7a33\u5b9a\u5947\u5f02\u503c\u5728 1\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u907f\u514d\u6ce8\u610f\u529b logits \u7206\u70b8<\/strong>\uff1aV4 \u7684\u6ce8\u610f\u529b\u67b6\u6784\u5141\u8bb8\u76f4\u63a5\u5bf9 query \u548c KV \u6761\u76ee\u5e94\u7528 RMSNorm\uff0c\u6709\u6548\u9632\u6b62 logits \u7206\u70b8\uff0c\u56e0\u6b64<strong>\u4e0d\u9700\u8981\u4f7f\u7528 QK-Clip \u6280\u672f<\/strong>\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"3-\u901a\u7528\u57fa\u7840\u8bbe\u65bd\">3. \u901a\u7528\u57fa\u7840\u8bbe\u65bd<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"31-\u4e13\u5bb6\u5e76\u884c\u4e2d\u7684\u7ec6\u7c92\u5ea6\u901a\u4fe1-\u8ba1\u7b97\u91cd\u53e0\">3.1 \u4e13\u5bb6\u5e76\u884c\u4e2d\u7684\u7ec6\u7c92\u5ea6\u901a\u4fe1-\u8ba1\u7b97\u91cd\u53e0<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">MoE \u53ef\u901a\u8fc7\u4e13\u5bb6\u5e76\u884c\uff08EP\uff09\u52a0\u901f\uff0c\u4f46 EP \u9700\u8981\u590d\u6742\u7684\u8282\u70b9\u95f4\u901a\u4fe1\uff0c\u5bf9\u4e92\u8054\u5e26\u5bbd\u548c\u5ef6\u8fdf\u6709\u6781\u9ad8\u8981\u6c42\u3002\u4e3a\u7f13\u89e3 EP \u4e2d\u7684\u901a\u4fe1\u74f6\u9888\uff0cV4 \u63d0\u51fa<strong>\u7ec6\u7c92\u5ea6 EP \u65b9\u6848<\/strong>\uff0c\u5c06\u901a\u4fe1\u548c\u8ba1\u7b97\u878d\u5408\u5230\u5355\u4e2a\u6d41\u6c34\u7ebf kernel \u4e2d\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"\u901a\u4fe1\u5ef6\u8fdf\u53ef\u4ee5\u88ab\u9690\u85cf\">\u901a\u4fe1\u5ef6\u8fdf\u53ef\u4ee5\u88ab\u9690\u85cf<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">V4 \u4e2d\uff0c\u6bcf\u4e2a MoE \u5c42\u4e3b\u8981\u53ef\u5206\u4e3a 4 \u4e2a\u9636\u6bb5\uff1a\u901a\u4fe1\u7ed1\u5b9a\u7684 <strong>Dispatch<\/strong> \u548c <strong>Combine<\/strong>\uff0c\u8ba1\u7b97\u7ed1\u5b9a\u7684 <strong>Linear-1<\/strong> \u548c <strong>Linear-2<\/strong>\u3002Profiling \u8868\u660e\uff1a<strong>\u5355\u4e2a MoE \u5c42\u7684\u603b\u901a\u4fe1\u65f6\u95f4\u5c0f\u4e8e\u603b\u8ba1\u7b97\u65f6\u95f4<\/strong>\u3002\u56e0\u6b64\uff0c\u5c06\u901a\u4fe1\u548c\u8ba1\u7b97\u878d\u5408\u5230\u7edf\u4e00\u6d41\u6c34\u7ebf\u540e\uff0c\u8ba1\u7b97\u4ecd\u662f\u4e3b\u5bfc\u74f6\u9888\uff0c\u610f\u5473\u7740\u7cfb\u7edf\u53ef\u5bb9\u5fcd\u66f4\u4f4e\u7684\u4e92\u8054\u5e26\u5bbd\u800c\u4e0d\u4f1a\u964d\u4f4e\u7aef\u5230\u7aef\u6027\u80fd\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"\u7ec6\u7c92\u5ea6-ep-\u65b9\u6848\">\u7ec6\u7c92\u5ea6 EP \u65b9\u6848<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fdb\u4e00\u6b65\u5c06\u4e13\u5bb6\u5206\u6210\u591a\u4e2a <strong>wave<\/strong> \u8c03\u5ea6\u3002\u6bcf\u4e2a wave \u5305\u542b\u5c11\u91cf\u4e13\u5bb6\u3002\u4e00\u65e6 wave \u5185\u6240\u6709\u4e13\u5bb6\u5b8c\u6210\u901a\u4fe1\uff0c\u8ba1\u7b97\u7acb\u5373\u5f00\u59cb\u800c\u65e0\u9700\u7b49\u5f85\u5176\u4ed6\u4e13\u5bb6\u3002\u7a33\u6001\u4e0b\uff0c\u5f53\u524d wave \u7684\u8ba1\u7b97\u3001\u4e0b\u4e00 wave \u7684 token \u4f20\u8f93\u3001\u5df2\u5b8c\u6210\u4e13\u5bb6\u7684\u7ed3\u679c\u53d1\u9001<strong>\u5168\u90e8\u5e76\u53d1\u8fdb\u884c<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u6027\u80fd\u5bf9\u6bd4<\/strong>\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>\u65b9\u6848<\/th><th>\u7406\u8bba\u52a0\u901f\u6bd4<\/th><\/tr><\/thead><tbody><tr><td>Naive Solution<\/td><td>1.0\u00d7\uff08\u57fa\u51c6\uff09<\/td><\/tr><tr><td>Comet<\/td><td>1.42\u00d7<\/td><\/tr><tr><td>V4\uff08\u672c\u5de5\u4f5c\uff09<\/td><td><strong>1.92\u00d7<\/strong> \ud83c\udfc6<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u5b9e\u6d4b\u6570\u636e\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u4e00\u822c\u63a8\u7406\u5de5\u4f5c\u8d1f\u8f7d\uff1a1.50\u20131.73\u00d7 \u52a0\u901f\u3002<\/li>\n\n\n\n<li>\u5ef6\u8fdf\u654f\u611f\u573a\u666f\uff08RL rollout\u3001\u9ad8\u901f agent \u670d\u52a1\uff09\uff1a\u6700\u9ad8 <strong>1.96\u00d7<\/strong> \u52a0\u901f\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">CUDA \u7248 mega-kernel \u540d\u4e3a <strong>MegaMoE<\/strong>\uff0c\u5df2\u5f00\u6e90\u4e8e DeepGEMM\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"\u7ed9\u786c\u4ef6\u5382\u5546\u7684\u89c2\u5bdf\u4e0e\u5efa\u8bae\">\u7ed9\u786c\u4ef6\u5382\u5546\u7684\u89c2\u5bdf\u4e0e\u5efa\u8bae<\/h4>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u8ba1\u7b97-\u901a\u4fe1\u6bd4<\/strong>\uff1a\u5168\u91cd\u53e0\u53d6\u51b3\u4e8e\u8ba1\u7b97-\u901a\u4fe1\u6bd4\uff0c\u800c\u975e\u5355\u7eaf\u5e26\u5bbd\u3002\u5bf9\u4e8e DeepSeek-V4-Pro\uff0c\u6bcf\u4e2a token-expert \u5bf9\u9700 $6hd$ FLOPs\uff08SwiGLU gate\/up\/down \u6295\u5f71\uff09\uff0c\u4f46\u53ea\u9700 $3h$ \u5b57\u8282\u901a\u4fe1\uff08FP8 Dispatch + BF16 Combine\uff09\uff0c\u7b80\u5316\u4e3a\uff1a$C\/B \\leq 2d = 6144$ FLOPs\/Byte\u3002\u5373\u6bcf GBps \u4e92\u8054\u5e26\u5bbd\u8db3\u4ee5\u9690\u85cf 6.1 TFLOP\/s \u8ba1\u7b97\u3002\u8d85\u8fc7\u6b64\u9608\u503c\u540e\uff0c\u7ee7\u7eed\u589e\u52a0\u5e26\u5bbd\u6536\u76ca\u9012\u51cf\u3002<\/li>\n\n\n\n<li><strong>\u529f\u8017\u9884\u7b97<\/strong>\uff1a\u6781\u81f4 kernel \u878d\u5408\u4f1a\u540c\u65f6\u9ad8\u8d1f\u8f7d\u9a71\u52a8\u8ba1\u7b97\u3001\u5185\u5b58\u3001\u7f51\u7edc\uff0c\u529f\u8017\u9650\u6d41\u662f\u5173\u952e\u6027\u80fd\u9650\u5236\u3002<\/li>\n\n\n\n<li><strong>\u901a\u4fe1\u539f\u8bed<\/strong>\uff1aV4 \u91c7\u7528 pull-based \u65b9\u6848\uff0c\u6bcf\u4e2a GPU \u4e3b\u52a8\u4ece\u8fdc\u7a0b GPU \u8bfb\u53d6\u6570\u636e\uff0c\u907f\u514d fine-grained push \u5e26\u6765\u7684\u9ad8\u901a\u77e5\u5ef6\u8fdf\u3002<\/li>\n\n\n\n<li><strong>\u6fc0\u6d3b\u51fd\u6570<\/strong>\uff1a\u5efa\u8bae\u7528<strong>\u4f4e\u6210\u672c\u7684 element-wise \u6fc0\u6d3b\u51fd\u6570<\/strong>\u66ff\u6362 SwiGLU\uff0c\u907f\u514d\u6307\u6570\u548c\u9664\u6cd5\u3002<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"32-\u4f7f\u7528-tilelang-\u7075\u6d3b\u9ad8\u6548\u5730\u5f00\u53d1-kernel\">3.2 \u4f7f\u7528 TileLang \u7075\u6d3b\u9ad8\u6548\u5730\u5f00\u53d1 Kernel<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">V4 \u7cbe\u5de7\u7684\u6a21\u578b\u67b6\u6784\u672c\u4f1a\u4ea7\u751f\u6570\u767e\u4e2a\u7ec6\u7c92\u5ea6 Torch ATen \u7b97\u5b50\u3002V4 \u91c7\u7528 <strong>TileLang<\/strong> \u4f5c\u4e3a DSL \u5f00\u53d1\u4e00\u5957\u878d\u5408 kernel \u66ff\u6362\u5927\u90e8\u5206\u7b97\u5b50\uff0c\u5728\u5b9e\u73b0\u6700\u4f73\u6027\u80fd\u7684\u540c\u65f6\u964d\u4f4e\u5f00\u53d1\u5de5\u4f5c\u91cf\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"\u901a\u8fc7-host-codegen-\u964d\u4f4e\u8c03\u7528\u5f00\u9500\">\u901a\u8fc7 Host Codegen \u964d\u4f4e\u8c03\u7528\u5f00\u9500<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">\u968f\u7740\u52a0\u901f\u5668\u6027\u80fd\u6301\u7eed\u589e\u957f\uff0cCPU \u4fa7\u7f16\u6392\u5f00\u9500\u8d8a\u6765\u8d8a\u7a81\u51fa\u3002V4 \u901a\u8fc7 <strong>Host Codegen<\/strong> \u7f13\u89e3\u6b64\u5f00\u9500\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5728 IR \u7ea7\u522b\u534f\u540c\u751f\u6210\u8bbe\u5907 kernel \u548c\u8f7b\u91cf\u7ea7 host launcher\u3002<\/li>\n\n\n\n<li>\u5d4c\u5165\u5fc5\u8981\u7684\u5143\u6570\u636e\uff08\u6570\u636e\u7c7b\u578b\u3001rank\/shape \u7ea6\u675f\u3001stride\/layout \u5047\u8bbe\uff09\u3002<\/li>\n\n\n\n<li>Launcher \u964d\u7ea7\u4e3a\u57fa\u4e8e TVM-FFI \u7684 host \u6e90\u7801\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u5b9e\u6d4b\u7ed3\u679c\uff1a<strong>CPU \u4fa7\u9a8c\u8bc1\u5f00\u9500\u4ece\u6570\u5341\/\u6570\u767e\u5fae\u79d2\u964d\u81f3\u6bcf\u6b21\u8c03\u7528\u4e0d\u5230 1 \u5fae\u79d2<\/strong>\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"smt-\u6c42\u89e3\u5668\u8f85\u52a9\u7684\u5f62\u5f0f\u5316\u6574\u6570\u5206\u6790\">SMT \u6c42\u89e3\u5668\u8f85\u52a9\u7684\u5f62\u5f0f\u5316\u6574\u6570\u5206\u6790<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">V4 \u5c06 <strong>Z3 SMT \u6c42\u89e3\u5668<\/strong>\u96c6\u6210\u5230 TileLang \u7684\u4ee3\u6570\u7cfb\u7edf\u4e2d\u3002\u5c06 TileLang \u6574\u6570\u8868\u8fbe\u5f0f\u7ffb\u8bd1\u4e3a Z3 \u7684\u65e0\u91cf\u8bcd\u975e\u7ebf\u6027\u6574\u6570\u7b97\u672f\uff08QF_NIA\uff09\u3002\u5728\u5408\u7406\u8d44\u6e90\u9650\u5236\u4e0b\uff0cZ3 \u63d0\u5347\u6574\u4f53\u4f18\u5316\u6027\u80fd\uff0c\u540c\u65f6\u5c06\u7f16\u8bd1\u65f6\u5f00\u9500\u9650\u5236\u5728\u51e0\u79d2\u5185\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"\u6570\u503c\u7cbe\u5ea6\u4e0e-bitwise-\u53ef\u590d\u73b0\u6027\">\u6570\u503c\u7cbe\u5ea6\u4e0e bitwise \u53ef\u590d\u73b0\u6027<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">V4 \u9ed8\u8ba4\u4f18\u5148\u4fdd\u8bc1\u7cbe\u5ea6\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u7f16\u8bd1\u5668\u7ea7\u5173\u95ed fast-math \u4f18\u5316\u3002<\/li>\n\n\n\n<li>\u7cbe\u5ea6\u5f71\u54cd\u7c7b\u8fd1\u4f3c\u4ec5\u4f5c\u4e3a\u663e\u5f0f\u524d\u7aef\u7b97\u5b50\u63d0\u4f9b\uff08\u5982 <code>T.__exp<\/code>\u3001<code>T.__log<\/code>\u3001<code>T.__sin<\/code>\uff09\u3002<\/li>\n\n\n\n<li>\u9700\u4e25\u683c IEEE-754 \u8bed\u4e49\u65f6\uff0c\u63d0\u4f9b\u5177\u6709\u663e\u5f0f rounding \u6a21\u5f0f\u7684 IEEE-compliant intrinsics\u3002<\/li>\n\n\n\n<li>\u901a\u8fc7 <code>T.annotate_layout<\/code> \u7b49\u5e03\u5c40\u6807\u6ce8\uff0c\u53ef\u56fa\u5b9a layout \u76f8\u5173\u7684\u964d\u7ea7\u51b3\u7b56\uff0c\u4fdd\u6301\u8ba1\u7b97\u548c\u7d2f\u52a0\u987a\u5e8f\u4e0e\u53c2\u8003 CUDA \u5b9e\u73b0\u4e00\u81f4\uff0c\u5b9e\u73b0 bit-identical \u8f93\u51fa\u3002<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"33-\u9ad8\u6027\u80fd\u6279\u6b21\u4e0d\u53d8\u6027\u4e0e\u786e\u5b9a\u6027\u5185\u6838\u5e93\">3.3 \u9ad8\u6027\u80fd\u6279\u6b21\u4e0d\u53d8\u6027\u4e0e\u786e\u5b9a\u6027\u5185\u6838\u5e93<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e3a\u652f\u6301\u9ad8\u6548\u8bad\u7ec3\u548c\u63a8\u7406\uff0cV4 \u5f00\u53d1\u4e86\u4e00\u5957\u9ad8\u6027\u80fd\u8ba1\u7b97\u5185\u6838\uff0c<strong>\u9664\u6ee1\u8db3\u57fa\u672c\u529f\u80fd\u548c\u786c\u4ef6\u5229\u7528\u7387\u6700\u5927\u5316\u5916\uff0c\u53e6\u4e00\u5173\u952e\u8bbe\u8ba1\u76ee\u6807\u662f\u4fdd\u8bc1\u8bad\u7ec3\u53ef\u590d\u73b0\u6027\u4ee5\u53ca\u9884\u8bad\u7ec3\u3001\u540e\u8bad\u7ec3\u3001\u63a8\u7406\u7ba1\u7ebf\u4e4b\u95f4\u7684 bitwise \u5bf9\u9f50<\/strong>\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"\u6279\u6b21\u4e0d\u53d8\u6027\uff08batch-invariance\uff09\">\u6279\u6b21\u4e0d\u53d8\u6027\uff08Batch Invariance\uff09<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">\u786e\u4fdd\u4efb\u610f token \u7684\u8f93\u51fa\u5728\u6279\u6b21\u4e2d\u4f4d\u7f6e\u65e0\u5173\u4fdd\u6301 bitwise \u4e00\u81f4\u3002\u4e3b\u8981\u6311\u6218\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u6ce8\u610f\u529b<\/strong>\uff1a\u4e0d\u80fd\u4f7f\u7528 split-KV \u65b9\u6cd5\uff08\u4f1a\u5c06\u5355\u5e8f\u5217\u7684\u6ce8\u610f\u529b\u8ba1\u7b97\u5206\u5e03\u5230\u591a\u4e2a SM \u4e0a\u5e73\u8861\u8d1f\u8f7d\uff09\u3002\u653e\u5f03\u6b64\u6280\u672f\u4f1a\u5bfc\u81f4\u4e25\u91cd\u7684 wave-quantization \u95ee\u9898\u3002V4 \u5f00\u53d1\u4e86<strong>\u53cc kernel \u7b56\u7565<\/strong>\uff1a\u7b2c\u4e00\u4e2a kernel \u5728\u5355\u4e2a SM \u5185\u8ba1\u7b97\u6574\u4e2a\u5e8f\u5217\u7684\u6ce8\u610f\u529b\u8f93\u51fa\uff08\u9ad8\u541e\u5410\uff09\uff1b\u7b2c\u4e8c\u4e2a kernel \u4f7f\u7528\u591a\u4e2a SM \u5904\u7406\u6700\u7ec8\u90e8\u5206\u586b\u5145\u7684 wave\uff08\u964d\u4f4e\u5ef6\u8fdf\uff09\u3002\u4e24\u4e2a kernel \u901a\u8fc7\u7cbe\u5fc3\u8bbe\u8ba1\u8ba1\u7b97\u8def\u5f84\u4fdd\u8bc1\u76f8\u540c\u7d2f\u52a0\u987a\u5e8f\uff0c\u7b2c\u4e8c\u4e2a kernel \u5229\u7528 thread-block cluster \u5185\u7684 <strong>distributed shared memory<\/strong> \u5b9e\u73b0\u8de8 SM \u9ad8\u901f\u6570\u636e\u4ea4\u6362\u3002<\/li>\n\n\n\n<li><strong>\u77e9\u9635\u4e58\u6cd5<\/strong>\uff1a\u4f20\u7edf cuBLAS \u65e0\u6cd5\u5b9e\u73b0\u6279\u6b21\u4e0d\u53d8\u3002V4 \u7aef\u5230\u7aef\u66ff\u6362\u4e3a <strong>DeepGEMM<\/strong>\u3002\u5bf9\u4e8e\u5c0f batch size\uff0c\u4f20\u7edf\u5b9e\u73b0\u5e38\u7528 split-k \u63d0\u5347\u6027\u80fd\uff0c\u4f46 split-k \u65e0\u6cd5\u4fdd\u8bc1\u6279\u6b21\u4e0d\u53d8\u3002V4 \u5728\u5927\u591a\u6570\u573a\u666f\u653e\u5f03 split-k\uff0c\u5e76\u5f15\u5165\u4e00\u7ec4\u4f18\u5316\u4f7f\u77e9\u9635\u4e58\u5b9e\u73b0\u5728\u5927\u591a\u6570\u4e3b\u8981\u573a\u666f\u4e0b\u5339\u914d\u751a\u81f3\u8d85\u8d8a\u6807\u51c6 split-k \u6027\u80fd\u3002<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"\u786e\u5b9a\u6027\uff08determinism\uff09\">\u786e\u5b9a\u6027\uff08Determinism\uff09<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">\u786e\u5b9a\u6027\u8bad\u7ec3\u5bf9\u8c03\u8bd5\u786c\u4ef6\/\u8f6f\u4ef6\u95ee\u9898\u6781\u6709\u5e2e\u52a9\u3002\u5f53\u8bad\u7ec3\u51fa\u73b0 loss spike \u7b49\u5f02\u5e38\u65f6\uff0c\u786e\u5b9a\u6027\u4f7f\u7814\u7a76\u4eba\u5458\u66f4\u6613\u5b9a\u4f4d\u6570\u503c\u539f\u56e0\u3002\u975e\u786e\u5b9a\u6027\u901a\u5e38\u6e90\u4e8e\u975e\u786e\u5b9a\u6027\u7d2f\u52a0\u987a\u5e8f\uff08\u5e38\u7531\u539f\u5b50\u52a0\u6307\u4ee4\u5bfc\u81f4\uff09\uff0c\u4e3b\u8981\u53d1\u751f\u5728\u53cd\u5411\u4f20\u64ad\u7684\u4ee5\u4e0b\u90e8\u5206\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u6ce8\u610f\u529b\u53cd\u5411<\/strong>\uff1a\u4f20\u7edf\u7a00\u758f\u6ce8\u610f\u529b\u53cd\u5411\u4f7f\u7528 atomicAdd \u7d2f\u52a0 KV token \u68af\u5ea6\uff0c\u7531\u4e8e\u6d6e\u70b9\u52a0\u6cd5\u7684\u975e\u7ed3\u5408\u6027\u5f15\u5165\u975e\u786e\u5b9a\u6027\u3002V4 \u4e3a\u6bcf\u4e2a SM \u5206\u914d\u72ec\u7acb\u7d2f\u52a0\u7f13\u51b2\u533a\uff0c\u518d\u8de8\u6240\u6709\u7f13\u51b2\u533a\u505a\u5168\u5c40\u786e\u5b9a\u6027\u6c42\u548c\u3002<\/li>\n\n\n\n<li><strong>MoE \u53cd\u5411<\/strong>\uff1a\u591a\u4e2a\u6765\u81ea\u4e0d\u540c rank \u7684 SM \u5e76\u53d1\u5411\u540c\u4e00\u63a5\u6536 rank \u7684\u7f13\u51b2\u533a\u5199\u5165\u65f6\uff0c\u534f\u5546\u5199\u5165\u4f4d\u7f6e\u4e5f\u5f15\u5165\u975e\u786e\u5b9a\u6027\u3002V4 \u5728\u5355 rank \u5185\u8bbe\u8ba1 token \u987a\u5e8f\u9884\u5904\u7406\u673a\u5236\uff0c\u8de8\u591a rank \u91c7\u7528\u7f13\u51b2\u533a\u9694\u79bb\u3002<\/li>\n\n\n\n<li><strong>mHC \u77e9\u9635\u4e58\u6cd5<\/strong>\uff1amHC \u6d89\u53ca\u8f93\u51fa\u7ef4\u5ea6\u4ec5\u4e3a 24 \u7684\u77e9\u9635\u4e58\u6cd5\uff0c\u5c0f batch \u4e0b\u5fc5\u987b\u4f7f\u7528 split-k\u3002V4 \u8ba9\u6bcf\u4e2a split \u90e8\u5206\u5355\u72ec\u8f93\u51fa\uff0c\u540e\u7eed kernel \u8fdb\u884c\u786e\u5b9a\u6027 reduction\uff0c\u540c\u65f6\u4fdd\u7559\u6027\u80fd\u548c\u786e\u5b9a\u6027\u3002<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"34-fp4-\u91cf\u5316\u611f\u77e5\u8bad\u7ec3\">3.4 FP4 \u91cf\u5316\u611f\u77e5\u8bad\u7ec3<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e3a\u5728\u90e8\u7f72\u65f6\u52a0\u901f\u63a8\u7406\u5e76\u8282\u7701\u5185\u5b58\uff0cV4 \u5728<strong>\u540e\u8bad\u7ec3\u9636\u6bb5<\/strong>\u5f15\u5165 <strong>Quantization-Aware Training\uff08QAT\uff09<\/strong>\u3002\u5e94\u7528 <strong>FP4\uff08MXFP4\uff09<\/strong> \u91cf\u5316\u5230\u4e24\u4e2a\u7ec4\u4ef6\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>MoE \u4e13\u5bb6\u6743\u91cd<\/strong>\uff08GPU \u5185\u5b58\u5360\u7528\u7684\u4e3b\u8981\u6765\u6e90\uff09<\/li>\n\n\n\n<li><strong>CSA indexer \u7684 QK \u8def\u5f84<\/strong>\uff08QK \u6fc0\u6d3b\u5b8c\u5168\u4ee5 FP4 \u7f13\u5b58\u3001\u52a0\u8f7d\u3001\u76f8\u4e58\uff0c\u52a0\u901f\u8d85\u957f\u4e0a\u4e0b\u6587\u7684\u6ce8\u610f\u529b\u5f97\u5206\u8ba1\u7b97\uff09<\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">\u6b64\u5916\uff0c\u8fdb\u4e00\u6b65\u5c06 index scores \u4ece FP32 \u91cf\u5316\u5230 BF16\u3002\u6b64\u4f18\u5316<strong>\u4e3a top-k selector \u5e26\u6765 2\u00d7 \u52a0\u901f\uff0c\u540c\u65f6\u4fdd\u6301 KV \u6761\u76ee 99.7% \u7684\u53ec\u56de\u7387<\/strong>\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"moe-\u4e13\u5bb6\u6743\u91cd\">MoE \u4e13\u5bb6\u6743\u91cd<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">\u9075\u5faa QAT \u5e38\u89c4\u505a\u6cd5\uff0cFP32 master weight \u5148\u91cf\u5316\u4e3a FP4\uff0c\u518d\u53cd\u91cf\u5316\u4e3a FP8 \u4f9b\u8ba1\u7b97\u4f7f\u7528\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c<strong>FP4 \u5230 FP8 \u7684\u53cd\u91cf\u5316\u662f\u65e0\u635f\u7684<\/strong>\uff08\u56e0\u4e3a FP8 E4M3 \u6bd4 FP4 E2M1 \u591a 2 \u4e2a\u6307\u6570\u4f4d\uff0c\u52a8\u6001\u8303\u56f4\u66f4\u5927\uff09\u3002\u53ea\u8981 FP4 \u5b50\u5757\uff081\u00d732 tile\uff09\u5728 FP8 \u91cf\u5316\u5757\uff08128\u00d7128 tile\uff09\u5185\u7684\u6700\u5927\/\u6700\u5c0f scale factor \u6bd4\u503c\u4e0d\u8d85\u8fc7\u67d0\u4e2a\u9608\u503c\uff0c\u7ec6\u7c92\u5ea6 scale \u4fe1\u606f\u53ef\u5b8c\u5168\u88ab FP8 \u52a8\u6001\u8303\u56f4\u5438\u6536\u3002\u8fd9\u4f7f\u6574\u4e2a QAT \u7ba1\u7ebf<strong>\u5b8c\u5168\u590d\u7528\u73b0\u6709 FP8 \u8bad\u7ec3\u6846\u67b6<\/strong>\uff0c\u65e0\u9700\u4efb\u4f55\u4fee\u6539\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u53cd\u5411\u4f20\u64ad\u4e2d\uff0c\u68af\u5ea6\u76f8\u5bf9\u4e8e\u524d\u5411 FP8 \u6743\u91cd\u8ba1\u7b97\uff0c\u76f4\u63a5\u4f20\u9012\u56de FP32 master \u6743\u91cd\u2014\u2014\u7b49\u6548\u4e8e\u901a\u8fc7\u91cf\u5316\u64cd\u4f5c\u5e94\u7528\u76f4\u901a\u4f30\u8ba1\u5668\uff08STE\uff09\uff0c\u4e5f\u907f\u514d\u4e86\u91cd\u65b0\u91cf\u5316\u8f6c\u7f6e\u6743\u91cd\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>RL \u8bad\u7ec3\u7684\u63a8\u7406\u548c rollout \u9636\u6bb5<\/strong>\uff08\u4e0d\u6d89\u53ca\u53cd\u5411\uff09\u76f4\u63a5\u4f7f\u7528\u771f\u5b9e FP4 \u91cf\u5316\u6743\u91cd\u800c\u975e\u6a21\u62df\u91cf\u5316\uff0c\u786e\u4fdd\u91c7\u6837\u884c\u4e3a\u4e0e\u5728\u7ebf\u90e8\u7f72\u5b8c\u5168\u4e00\u81f4\uff0c\u540c\u65f6\u51cf\u5c11 kernel \u5185\u5b58\u52a0\u8f7d\u3001\u663e\u8457\u964d\u4f4e\u5185\u5b58\u6d88\u8017\u3002CSA indexer \u7684 QK \u8def\u5f84\u4e5f\u7c7b\u4f3c\u5904\u7406\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"35-\u8bad\u7ec3\u6846\u67b6\">3.5 \u8bad\u7ec3\u6846\u67b6<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u8bad\u7ec3\u6846\u67b6\u57fa\u4e8e V3 \u7684\u53ef\u6269\u5c55\u9ad8\u6548\u57fa\u7840\u8bbe\u65bd\u3002\u4e3a\u9002\u914d V4 \u7684\u65b0\u67b6\u6784\u7ec4\u4ef6\uff08Muon \u4f18\u5316\u5668\u3001mHC\u3001\u6df7\u5408\u6ce8\u610f\u529b\uff09\uff0c\u5f15\u5165\u4e86\u82e5\u5e72\u5173\u952e\u521b\u65b0\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"351-muon-\u7684\u9ad8\u6548\u5b9e\u73b0\">3.5.1 Muon \u7684\u9ad8\u6548\u5b9e\u73b0<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">Muon \u9700\u8981<strong>\u5b8c\u6574\u7684\u68af\u5ea6\u77e9\u9635<\/strong>\u6765\u8ba1\u7b97\u53c2\u6570\u66f4\u65b0\uff0c\u4e0e ZeRO \u5b58\u5728\u51b2\u7a81\uff08\u4f20\u7edf ZeRO \u4e3a AdamW \u7b49 element-wise \u4f18\u5316\u5668\u8bbe\u8ba1\uff09\u3002V4 \u8bbe\u8ba1<strong>\u6df7\u5408 ZeRO bucket \u5206\u914d\u7b56\u7565<\/strong>\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Dense \u53c2\u6570<\/strong>\uff1a\u9650\u5236 ZeRO \u5e76\u884c\u6700\u5927\u5c3a\u5bf8\uff0c\u4f7f\u7528\u80cc\u5305\u7b97\u6cd5\u5c06\u53c2\u6570\u77e9\u9635\u5206\u914d\u7ed9 rank\uff0c\u786e\u4fdd\u8d1f\u8f7d\u5927\u81f4\u5747\u8861\u3002\u6bcf\u4e2a rank \u4e0a\u7684 bucket padding \u5230\u4e0e\u6700\u5927 bucket \u4e00\u81f4\u3002\u5728\u901a\u5e38\u6bcf\u4e2a rank \u7ba1\u7406\u4e0d\u8d85\u8fc7 5 \u4e2a\u53c2\u6570\u77e9\u9635\u7684\u60c5\u51b5\u4e0b\uff0cpadding \u5f00\u9500 &lt; 10%\u3002<\/li>\n\n\n\n<li><strong>MoE \u53c2\u6570<\/strong>\uff1a\u6bcf\u4e2a\u4e13\u5bb6\u72ec\u7acb\u4f18\u5316\u3002\u5148\u6241\u5e73\u5316\u6240\u6709\u4e13\u5bb6\u6240\u6709\u5c42\u7684 SwiGLU down \u6295\u5f71\u77e9\u9635\uff0c\u7136\u540e\u662f\u6241\u5e73\u5316\u7684 up \u6295\u5f71\u77e9\u9635\u548c gate \u77e9\u9635\uff0cpadding \u540e\u5747\u5300\u5206\u5e03\u5230\u6240\u6709 rank\u3002\u7531\u4e8e\u4e13\u5bb6\u6570\u91cf\u5927\uff0cMoE \u53c2\u6570\u4e0d\u8bbe ZeRO \u5e76\u884c\u9650\u5236\uff0cpadding \u5f00\u9500\u53ef\u5ffd\u7565\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u989d\u5916\u4f18\u5316<\/strong>\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u540c rank \u4e0a\u5f62\u72b6\u76f8\u540c\u7684\u8fde\u7eed\u53c2\u6570\u81ea\u52a8\u5408\u5e76\uff0c\u4f7f Newton-Schulz \u8fed\u4ee3\u53ef\u6279\u91cf\u6267\u884c\uff0c\u63d0\u5347\u786c\u4ef6\u5229\u7528\u7387\u3002<\/li>\n\n\n\n<li>Newton-Schulz \u8fed\u4ee3\u7528 BF16 \u77e9\u9635\u4e58\u4ecd\u7a33\u5b9a\uff0c\u56e0\u6b64 V4 \u4ee5\u968f\u673a\u820d\u5165\u65b9\u5f0f\u5c06\u8de8 DP rank \u540c\u6b65\u7684 MoE \u68af\u5ea6\u91cf\u5316\u5230 BF16\uff0c<strong>\u901a\u4fe1\u91cf\u51cf\u534a<\/strong>\u3002<\/li>\n\n\n\n<li>\u4e3a\u907f\u514d\u4f4e\u7cbe\u5ea6\u52a0\u6cd5\u5668\u5f15\u5165\u7684\u7d2f\u79ef\u8bef\u5dee\uff0c\u7528<strong>\u4e24\u9636\u6bb5\u65b9\u6cd5<\/strong>\u66ff\u6362\u4f20\u7edf tree\/ring reduce-scatter\uff1a\u5148 all-to-all \u4ea4\u6362\u672c\u5730\u68af\u5ea6\uff0c\u7136\u540e\u6bcf\u4e2a rank \u5728 FP32 \u4e2d\u672c\u5730\u6c42\u548c\u3002<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"352-mhc-\u7684\u7ecf\u6d4e\u9ad8\u6548\u3001\u5185\u5b58\u9ad8\u6548\u5b9e\u73b0\">3.5.2 mHC \u7684\u7ecf\u6d4e\u9ad8\u6548\u3001\u5185\u5b58\u9ad8\u6548\u5b9e\u73b0<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">mHC \u76f8\u6bd4\u4f20\u7edf\u6b8b\u5dee\u8fde\u63a5\u589e\u52a0\u4e86\u6fc0\u6d3b\u5185\u5b58\u6d88\u8017\u548c\u6d41\u6c34\u7ebf\u9636\u6bb5\u95f4\u901a\u4fe1\u91cf\u3002V4 \u5b9e\u65bd\u591a\u9879\u4f18\u5316\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u878d\u5408 kernel<\/strong>\uff1a\u4e3a\u8bad\u7ec3\u548c\u63a8\u7406\u7cbe\u5fc3\u8bbe\u8ba1\u5e76\u5b9e\u73b0 mHC \u878d\u5408 kernel\u3002<\/li>\n\n\n\n<li><strong>\u91cd\u8ba1\u7b97\u7b56\u7565<\/strong>\uff1a\u9009\u62e9\u6027 checkpoint \u4e2d\u95f4\u5f20\u91cf\uff0c\u91cd\u8ba1\u7b97\u5927\u591a\u6570\u5c42\u95f4 hidden states \u548c\u6240\u6709\u5f52\u4e00\u5316\u5c42\u8f93\u5165\uff0c\u907f\u514d\u91cd\u8ba1\u7b97\u8ba1\u7b97\u5bc6\u96c6\u578b\u64cd\u4f5c\u3002<\/li>\n\n\n\n<li><strong>DualPipe 1F1B \u8c03\u6574<\/strong>\uff1a\u9002\u914d\u589e\u52a0\u7684\u6d41\u6c34\u7ebf\u901a\u4fe1\uff0c\u5141\u8bb8 mHC \u4e2d\u90e8\u5206\u64cd\u4f5c\u5e76\u53d1\u6267\u884c\u3002<\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e0a\u8ff0\u4f18\u5316\u5c06 mHC \u7684 wall-time \u5f00\u9500\u9650\u5236\u5728**\u91cd\u53e0 1F1B \u6d41\u6c34\u7ebf\u9636\u6bb5\u7684 6.7%**\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"353-\u957f\u4e0a\u4e0b\u6587\u6ce8\u610f\u529b\u7684-contextual-parallelism\">3.5.3 \u957f\u4e0a\u4e0b\u6587\u6ce8\u610f\u529b\u7684 Contextual Parallelism<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f20\u7edf Context Parallelism\uff08CP\uff09\u5728\u5e8f\u5217\u7ef4\u5ea6\u5207\u5206\uff0c\u6bcf\u4e2a rank \u7ef4\u62a4\u8fde\u7eed $s$ \u4e2a token\u3002\u8fd9\u5bf9 V4 \u7684\u538b\u7f29\u6ce8\u610f\u529b\uff08CSA\u3001HCA\uff09\u5f15\u5165\u4e24\u4e2a\u6311\u6218\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li>\u8bad\u7ec3\u6837\u672c\u7531\u591a\u5e8f\u5217\u6253\u5305\uff0c\u6bcf\u5e8f\u5217\u4ee5\u56e0\u5b50 $m$ \u72ec\u7acb\u538b\u7f29\uff0c\u5c3e\u90e8\u5c11\u4e8e $m$ \u7684 token \u88ab\u4e22\u5f03\u3002\u56e0\u6b64\u538b\u7f29\u540e KV \u957f\u5ea6\u901a\u5e38\u5c11\u4e8e $s\/m$\uff0c\u4e14\u5728 rank \u95f4\u4e0d\u540c\u3002<\/li>\n\n\n\n<li>\u538b\u7f29\u9700\u8981 $m$ \u4e2a\u8fde\u7eed KV \u6761\u76ee\uff0c\u53ef\u80fd\u8de8\u8d8a CP rank \u8fb9\u754c\u3002<\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">V4 \u8bbe\u8ba1<strong>\u4e24\u9636\u6bb5\u901a\u4fe1\u65b9\u6848<\/strong>\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u7b2c\u4e00\u9636\u6bb5<\/strong>\uff1a\u6bcf\u4e2a rank $i$ \u5c06\u6700\u540e $m$ \u4e2a\u672a\u538b\u7f29 KV \u6761\u76ee\u53d1\u9001\u7ed9 rank $i+1$\u3002rank $i+1$ \u538b\u7f29\u90e8\u5206\u63a5\u6536\u6761\u76ee\u4e0e\u5176\u672c\u5730 $s$ \u4e2a\u672a\u538b\u7f29 KV \u6761\u76ee\uff0c\u4ea7\u751f\u56fa\u5b9a\u957f\u5ea6 $s\/m + 1$ \u4e2a\u538b\u7f29\u6761\u76ee\uff08\u542b\u90e8\u5206 padding\uff09\u3002<\/li>\n\n\n\n<li><strong>\u7b2c\u4e8c\u9636\u6bb5<\/strong>\uff1aall-gather \u6536\u96c6\u6240\u6709 CP rank \u7684\u672c\u5730\u538b\u7f29 KV \u6761\u76ee\u3002\u878d\u5408\u7684 select-and-pad \u7b97\u5b50\u5c06\u5176\u91cd\u7ec4\u4e3a\u603b\u957f\u5ea6 $\\text{cp_size} \\cdot s\/m$ \u7684\u5b8c\u6574\u538b\u7f29 KV \u6761\u76ee\u96c6\uff0cpadding \u6761\u76ee\u7f6e\u4e8e\u5c3e\u90e8\u3002<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"354-\u6269\u5c55\u81ea\u52a8\u5fae\u5206\u4ee5\u7075\u6d3b\u6fc0\u6d3b\u68c0\u67e5\u70b9\">3.5.4 \u6269\u5c55\u81ea\u52a8\u5fae\u5206\u4ee5\u7075\u6d3b\u6fc0\u6d3b\u68c0\u67e5\u70b9<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f20\u7edf\u6fc0\u6d3b checkpoint \u5b9e\u73b0\u4ee5<strong>\u6574\u4e2a\u6a21\u5757\u4e3a\u7c92\u5ea6<\/strong>\uff0c\u5bfc\u81f4\u91cd\u8ba1\u7b97\u6210\u672c\u548c\u6fc0\u6d3b\u5185\u5b58\u4e4b\u95f4\u7684\u6b21\u4f18\u6743\u8861\u3002\u53e6\u4e00\u79cd\u65b9\u6848\u662f\u624b\u52a8\u5b9e\u73b0\u6574\u5c42\u524d\u5411\u540e\u5411\u903b\u8f91\uff0c\u867d\u7cbe\u7ec6\u4f46\u5931\u53bb autograd \u4fbf\u5229\u6027\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">V4 \u5b9e\u73b0<strong>\u5f20\u91cf\u7ea7\u6fc0\u6d3b checkpoint \u673a\u5236\u5e76\u652f\u6301\u81ea\u52a8\u5fae\u5206<\/strong>\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5f00\u53d1\u8005\u53ea\u9700\u5b9e\u73b0\u524d\u5411\u4f20\u64ad\uff0c\u5e76\u9009\u62e9\u6027\u6807\u6ce8\u9700\u8981\u81ea\u52a8 checkpoint \u548c\u91cd\u8ba1\u7b97\u7684\u5f20\u91cf\u3002<\/li>\n\n\n\n<li>\u6846\u67b6\u5229\u7528 <strong>TorchFX<\/strong> \u8ffd\u8e2a\u5b8c\u6574\u8ba1\u7b97\u56fe\uff0c\u5bf9\u6bcf\u4e2a\u6807\u6ce8\u5f20\u91cf\u53cd\u5411\u904d\u5386\u8bc6\u522b\u6700\u5c0f\u91cd\u8ba1\u7b97\u5b50\u56fe\uff0c\u5b9a\u4e49\u4e3a<strong>\u91cd\u8ba1\u7b97\u56fe<\/strong>\u5e76\u63d2\u5165\u5230\u53cd\u5411\u903b\u8f91\u4e2d\u76f8\u5e94\u68af\u5ea6\u8ba1\u7b97\u4e4b\u524d\u3002<\/li>\n\n\n\n<li>\u91cd\u8ba1\u7b97\u901a\u8fc7\u76f4\u63a5\u91ca\u653e\u6807\u6ce8\u5f20\u91cf\u7684 GPU \u5185\u5b58\u5e76\u590d\u7528\u91cd\u8ba1\u7b97\u5f20\u91cf\u7684\u5b58\u50a8\u6307\u9488\u5b9e\u73b0\uff0c<strong>\u65e0 GPU \u5185\u5b58\u62f7\u8d1d<\/strong>\u3002<\/li>\n\n\n\n<li>\u7531\u4e8e\u56fe\u8ffd\u8e2a\u5177\u4f53\u6267\u884c\u6a21\u578b\uff0c\u53ef\u8ffd\u8e2a\u6bcf\u4e2a\u5f20\u91cf\u7684\u5e95\u5c42\u5b58\u50a8\u6307\u9488\uff0c\u5b9e\u73b0\u5171\u4eab\u5b58\u50a8\u5f20\u91cf\uff08\u5982 reshape \u7684\u8f93\u5165\u8f93\u51fa\uff09\u91cd\u8ba1\u7b97\u7684\u81ea\u52a8\u53bb\u91cd\u3002<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"36-\u63a8\u7406\u6846\u67b6\">3.6 \u63a8\u7406\u6846\u67b6<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u63a8\u7406\u6846\u67b6\u57fa\u672c\u7ee7\u627f\u81ea V3\uff0c\u4e3b\u8981\u5dee\u5f02\u5728 <strong>KV Cache \u7ba1\u7406<\/strong>\u4e0a\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"361-kv-cache-\u7ed3\u6784\u4e0e\u7ba1\u7406\">3.6.1 KV Cache \u7ed3\u6784\u4e0e\u7ba1\u7406<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>V4 \u4e2d\u7684\u5f02\u6784 KV \u6761\u76ee<\/strong>\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Lightning Indexer \u5f15\u5165\u7684\u7ef4\u5ea6\uff08embedding size \u4e0e\u4e3b\u6ce8\u610f\u529b\u4e0d\u540c\uff09<\/li>\n\n\n\n<li>CSA \u548c HCA \u7684\u538b\u7f29\u6280\u672f\u5c06\u5e8f\u5217\u957f\u5ea6\u5206\u522b\u7f29\u51cf\u4e3a $1\/m$ \u548c $1\/m'$\uff0c\u5bfc\u81f4\u5404\u5c42 KV cache \u5927\u5c0f\u4e0d\u540c<\/li>\n\n\n\n<li>SWA\uff08Sliding Window Attention\uff09\u5c42\u6709\u72ec\u7acb KV cache \u5927\u5c0f\u3001\u547d\u4e2d\u548c\u9a71\u9010\u7b56\u7565<\/li>\n\n\n\n<li>\u538b\u7f29\u5206\u652f\u4e2d\uff0c\u6bcf $m$ \u4e2a token \u751f\u6210\u4e00\u4e2a KV \u6761\u76ee\uff1b\u5269\u4f59\u4e0d\u8db3\u7684 token \u5fc5\u987b\u4fdd\u7559\u5728 buffer \u4e2d\u76f4\u5230\u53ef\u6267\u884c\u538b\u7f29\u64cd\u4f5c<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u6311\u6218<\/strong>\uff1a\u6df7\u5408\u6ce8\u610f\u529b\u8fdd\u53cd PagedAttention \u53ca\u5176\u53d8\u4f53\u7684\u57fa\u672c\u5047\u8bbe\u3002\u5373\u4f7f\u6700\u8fd1\u7684 Jenga\u3001Hymba \u7b49\u9488\u5bf9\u4e00\u822c\u6df7\u5408\u6ce8\u610f\u529b\u6a21\u578b\u6216\u7279\u5b9a\u7ed3\u6784\uff0c\u4ecd\u5b58\u5728\u4e24\u4e2a\u4e3b\u8981\u969c\u788d\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u591a\u6837\u7684\u7f13\u5b58\u7b56\u7565\uff08\u5982 SWA\uff09<\/li>\n\n\n\n<li>\u9ad8\u6027\u80fd\u6ce8\u610f\u529b kernel \u7684\u7ea6\u675f\uff08\u5982\u5bf9\u9f50\u8981\u6c42\uff09<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>V4 \u7684\u89e3\u51b3\u65b9\u6848<\/strong>\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>SWA \u4e0e\u672a\u538b\u7f29\u5c3e token \u7684\u72b6\u6001\u7f13\u5b58<\/strong>\uff1a\u5c06 SWA \u548c\u538b\u7f29\u5206\u652f\u7684\u672a\u538b\u7f29\u5c3e token \u89c6\u4e3a<strong>\u72b6\u6001\u7a7a\u95f4\u6a21\u578b<\/strong>\uff0c\u5176 KV cache \u662f\u4f9d\u8d56\u5f53\u524d\u4f4d\u7f6e\u7684\u5e8f\u5217\u7279\u5b9a\u72b6\u6001\u3002\u9884\u5206\u914d\u56fa\u5b9a\u5927\u5c0f\u7684\u72b6\u6001\u7f13\u5b58\u6c60\uff0c\u52a8\u6001\u5206\u914d\u7ed9\u6bcf\u4e2a\u5e8f\u5217\u3002<\/li>\n\n\n\n<li><strong>\u7a00\u758f\u6ce8\u610f\u529b\u5185\u6838\u534f\u540c\u8bbe\u8ba1<\/strong>\uff1a\u901a\u8fc7\u9ad8\u6027\u80fd\u7a00\u758f\u6ce8\u610f\u529b kernel\uff0c\u4e0d\u540c\u5c42\u53ef\u652f\u6301\u6bcf block \u53ef\u53d8 token \u6570\u800c\u4e0d\u635f\u5931\u6027\u80fd\u3002\u5bf9\u4e8e\u538b\u7f29\u7387 $m$ \u7684 CSA \u548c $m'$ \u7684 HCA\uff0c\u6bcf block \u539f\u59cb token \u6570\u53ef\u4ee5\u662f $\\text{lcm}(m, m')$ \u7684\u4efb\u610f\u500d\u6570\u3002<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"362-\u78c1\u76d8-kv-cache-\u5b58\u50a8\">3.6.2 \u78c1\u76d8 KV Cache \u5b58\u50a8<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">\u670d\u52a1 V4 \u65f6\uff0c\u5229\u7528<strong>\u78c1\u76d8 KV cache \u5b58\u50a8<\/strong>\u6d88\u9664 shared-prefix \u8bf7\u6c42\u7684\u91cd\u590d prefill\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>CSA\/HCA \u538b\u7f29 KV \u6761\u76ee<\/strong>\uff1a\u76f4\u63a5\u5b58\u50a8\u6240\u6709\u538b\u7f29 KV \u6761\u76ee\u3002\u547d\u4e2d\u65f6\u8bfb\u53d6\u5e76\u590d\u7528\u5bf9\u5e94\u524d\u7f00\u7684\u538b\u7f29 KV \u6761\u76ee\uff0c\u76f4\u5230\u6700\u540e\u4e00\u4e2a\u5b8c\u6574\u538b\u7f29\u5757\u3002\u5c3e\u90e8\u4e0d\u5b8c\u6574\u5757\u7684\u524d\u7f00 token \u4ecd\u9700\u91cd\u8ba1\u7b97\u4ee5\u6062\u590d\u672a\u538b\u7f29 KV \u6761\u76ee\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>SWA KV \u6761\u76ee<\/strong>\uff1a\u7531\u4e8e\u672a\u538b\u7f29\u4e14\u5b58\u5728\u4e8e\u6bcf\u4e00\u5c42\uff0c<strong>\u4f53\u79ef\u7ea6\u6bd4\u538b\u7f29 CSA\/HCA KV \u6761\u76ee\u5927 8 \u500d<\/strong>\u3002V4 \u8bbe\u8ba1\u4e09\u79cd\u7b56\u7565\u6743\u8861\u5b58\u50a8\u5f00\u9500\u548c\u8ba1\u7b97\u5197\u4f59\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>\u7b56\u7565<\/th><th>\u5b58\u50a8<\/th><th>\u8ba1\u7b97<\/th><th>\u8bf4\u660e<\/th><\/tr><\/thead><tbody><tr><td><strong>Full SWA Caching<\/strong><\/td><td>\u5b58\u50a8\u6240\u6709 token \u7684 SWA KV<\/td><td>\u96f6\u5197\u4f59\u8ba1\u7b97<\/td><td>\u5bf9\u73b0\u4ee3 SSD \u5b58\u50a8\u4f4e\u6548\uff0c\u5199\u5bc6\u96c6\u8bbf\u95ee\u4e0d\u5747\u8861<\/td><\/tr><tr><td><strong>Periodic Checkpointing<\/strong><\/td><td>\u6bcf $p$ \u4e2a token \u68c0\u67e5\u4e00\u6b21 SWA KV<\/td><td>\u52a0\u8f7d+\u5c3e\u90e8\u91cd\u8ba1\u7b97<\/td><td>\u901a\u8fc7 $p$ \u53ef\u8c03\u8282\u5b58\u50a8\u8ba1\u7b97\u6743\u8861<\/td><\/tr><tr><td><strong>Zero SWA Caching<\/strong><\/td><td>\u4e0d\u5b58\u50a8 SWA KV<\/td><td>\u6700\u5927\u91cd\u8ba1\u7b97\uff08\u6700\u540e $n_{\\text{win}} \\cdot L$ \u4e2a token\uff09<\/td><td>\u5229\u7528\u7f13\u5b58\u7684 CSA\/HCA KV \u6062\u590d SWA KV<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u6839\u636e\u5177\u4f53\u90e8\u7f72\u573a\u666f\u9009\u62e9\u6700\u5408\u9002\u7b56\u7565\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"4-\u9884\u8bad\u7ec3\">4. \u9884\u8bad\u7ec3<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"41-\u6570\u636e\u6784\u5efa\">4.1 \u6570\u636e\u6784\u5efa<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728 V3 \u9884\u8bad\u7ec3\u6570\u636e\u7684\u57fa\u7840\u4e0a\uff0cV4 \u6784\u5efa\u4e86\u66f4\u591a\u6837\u5316\u3001\u66f4\u9ad8\u8d28\u91cf\u3001<strong>\u6709\u6548\u4e0a\u4e0b\u6587\u66f4\u957f<\/strong>\u7684\u8bad\u7ec3\u8bed\u6599\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u7f51\u7edc\u6570\u636e<\/strong>\uff1a\u5b9e\u65bd\u8fc7\u6ee4\u7b56\u7565\u79fb\u9664\u6279\u91cf\u81ea\u52a8\u751f\u6210\u548c\u6a21\u677f\u5316\u5185\u5bb9\uff0c\u964d\u4f4e\u6a21\u578b\u5d29\u584c\uff08model collapse\uff09\u98ce\u9669\u3002<\/li>\n\n\n\n<li><strong>\u6570\u5b66\u548c\u7f16\u7a0b\u8bed\u6599<\/strong>\uff1a\u6838\u5fc3\u7ec4\u6210\u90e8\u5206\uff0c\u5e76\u901a\u8fc7\u5728 mid-training \u9636\u6bb5\u5f15\u5165 <strong>agentic \u6570\u636e<\/strong>\u8fdb\u4e00\u6b65\u589e\u5f3a\u7f16\u7801\u80fd\u529b\u3002<\/li>\n\n\n\n<li><strong>\u591a\u8bed\u8a00\u6570\u636e<\/strong>\uff1a\u4e3a V4 \u6784\u5efa\u66f4\u5927\u8bed\u6599\u5e93\uff0c\u6539\u5584\u5bf9\u4e0d\u540c\u6587\u5316\u957f\u5c3e\u77e5\u8bc6\u7684\u6355\u83b7\u3002<\/li>\n\n\n\n<li><strong>\u957f\u6587\u6863\u6570\u636e<\/strong>\uff1aV4 \u7279\u522b\u5f3a\u8c03\u957f\u6587\u6863\u6570\u636e\u7b56\u5212\uff0c\u4f18\u5148\u9009\u62e9\u79d1\u5b66\u8bba\u6587\u3001\u6280\u672f\u62a5\u544a\u7b49\u53cd\u6620\u72ec\u7279\u5b66\u672f\u4ef7\u503c\u7684\u6750\u6599\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u603b\u89c4\u6a21<\/strong>\uff1a\u9884\u8bad\u7ec3\u8bed\u6599\u8d85\u8fc7 <strong>32T tokens<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Tokenizer<\/strong>\uff1a\u5728 V3 tokenizer \u57fa\u7840\u4e0a\u5f15\u5165\u5c11\u91cf\u7279\u6b8a token \u7528\u4e8e\u4e0a\u4e0b\u6587\u6784\u9020\uff0c<strong>\u8bcd\u8868\u5927\u5c0f\u4ecd\u4e3a 128K<\/strong>\u3002\u7ee7\u627f V3 \u7684 token-splitting \u548c Fill-in-Middle\uff08FIM\uff09\u7b56\u7565\u3002<strong>\u4e0e V3 \u4e0d\u540c\u7684\u662f\uff0cV4 \u5728\u9884\u8bad\u7ec3\u4e2d\u91c7\u7528\u6837\u672c\u7ea7\u6ce8\u610f\u529b\u63a9\u7801<\/strong>\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"42-\u9884\u8bad\u7ec3\u8bbe\u7f6e\">4.2 \u9884\u8bad\u7ec3\u8bbe\u7f6e<\/h3>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"421-\u6a21\u578b\u8bbe\u7f6e\">4.2.1 \u6a21\u578b\u8bbe\u7f6e<\/h4>\n\n\n\n<h5 class=\"wp-block-heading\" id=\"deepseek-v4-flash\">DeepSeek-V4-Flash<\/h5>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>\u9879<\/th><th>\u503c<\/th><\/tr><\/thead><tbody><tr><td>Transformer \u5c42\u6570<\/td><td>43<\/td><\/tr><tr><td>Hidden \u7ef4\u5ea6 $d$<\/td><td>4096<\/td><\/tr><tr><td>\u524d 2 \u5c42<\/td><td>\u7eaf\u6ed1\u52a8\u7a97\u53e3\u6ce8\u610f\u529b<\/td><\/tr><tr><td>\u540e\u7eed\u5c42<\/td><td>CSA \u4e0e HCA \u4ea4\u9519<\/td><\/tr><tr><td>CSA \u538b\u7f29\u7387 $m$<\/td><td>4<\/td><\/tr><tr><td>CSA indexer query head \u6570 $n_h^I$<\/td><td>64<\/td><\/tr><tr><td>CSA indexer head \u7ef4\u5ea6 $c^I$<\/td><td>128<\/td><\/tr><tr><td>CSA attention top-k<\/td><td><strong>512<\/strong><\/td><\/tr><tr><td>HCA \u538b\u7f29\u7387 $m'$<\/td><td>128<\/td><\/tr><tr><td>Query head \u6570 $n_h$<\/td><td>64<\/td><\/tr><tr><td>Head \u7ef4\u5ea6 $c$<\/td><td>512<\/td><\/tr><tr><td>Query \u538b\u7f29\u7ef4\u5ea6 $d_c$<\/td><td>1024<\/td><\/tr><tr><td>\u8f93\u51fa\u6295\u5f71\u5206\u7ec4\u6570 $g$<\/td><td>8<\/td><\/tr><tr><td>\u4e2d\u95f4\u6ce8\u610f\u529b\u8f93\u51fa\u7ef4\u5ea6 $d_g$<\/td><td>1024<\/td><\/tr><tr><td>\u6ed1\u52a8\u7a97\u53e3\u5927\u5c0f $n_{\\text{win}}$<\/td><td>128<\/td><\/tr><tr><td>\u5171\u4eab\u4e13\u5bb6\u6570<\/td><td>1<\/td><\/tr><tr><td>\u8def\u7531\u4e13\u5bb6\u6570<\/td><td>256<\/td><\/tr><tr><td>\u6bcf\u4e2a\u4e13\u5bb6\u4e2d\u95f4 hidden \u7ef4\u5ea6<\/td><td>2048<\/td><\/tr><tr><td>\u6bcf token \u6fc0\u6d3b\u4e13\u5bb6\u6570<\/td><td>6<\/td><\/tr><tr><td>MTP \u6df1\u5ea6<\/td><td>1<\/td><\/tr><tr><td>\u524d 3 \u4e2a MoE \u5c42<\/td><td>Hash \u8def\u7531<\/td><\/tr><tr><td>mHC \u6269\u5c55\u56e0\u5b50 $n_{hc}$<\/td><td>4<\/td><\/tr><tr><td>Sinkhorn-Knopp \u8fed\u4ee3\u6570<\/td><td>20<\/td><\/tr><tr><td><strong>\u603b\u53c2\u6570 \/ \u6fc0\u6d3b\u53c2\u6570<\/strong><\/td><td><strong>284B \/ 13B<\/strong><\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<h5 class=\"wp-block-heading\" id=\"deepseek-v4-pro\">DeepSeek-V4-Pro<\/h5>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>\u9879<\/th><th>\u503c<\/th><\/tr><\/thead><tbody><tr><td>Transformer \u5c42\u6570<\/td><td>61<\/td><\/tr><tr><td>Hidden \u7ef4\u5ea6 $d$<\/td><td>7168<\/td><\/tr><tr><td>\u524d 2 \u5c42<\/td><td>HCA<\/td><\/tr><tr><td>\u540e\u7eed\u5c42<\/td><td>CSA \u4e0e HCA \u4ea4\u9519<\/td><\/tr><tr><td>CSA \u538b\u7f29\u7387 $m$<\/td><td>4<\/td><\/tr><tr><td>CSA indexer query head \u6570 $n_h^I$<\/td><td>64<\/td><\/tr><tr><td>CSA indexer head \u7ef4\u5ea6 $c^I$<\/td><td>128<\/td><\/tr><tr><td>CSA attention top-k<\/td><td><strong>1024<\/strong><\/td><\/tr><tr><td>HCA \u538b\u7f29\u7387 $m'$<\/td><td>128<\/td><\/tr><tr><td>Query head \u6570 $n_h$<\/td><td>128<\/td><\/tr><tr><td>Head \u7ef4\u5ea6 $c$<\/td><td>512<\/td><\/tr><tr><td>Query \u538b\u7f29\u7ef4\u5ea6 $d_c$<\/td><td>1536<\/td><\/tr><tr><td>\u8f93\u51fa\u6295\u5f71\u5206\u7ec4\u6570 $g$<\/td><td>16<\/td><\/tr><tr><td>\u4e2d\u95f4\u6ce8\u610f\u529b\u8f93\u51fa\u7ef4\u5ea6 $d_g$<\/td><td>1024<\/td><\/tr><tr><td>\u6ed1\u52a8\u7a97\u53e3\u5927\u5c0f $n_{\\text{win}}$<\/td><td>128<\/td><\/tr><tr><td>\u5171\u4eab\u4e13\u5bb6\u6570<\/td><td>1<\/td><\/tr><tr><td>\u8def\u7531\u4e13\u5bb6\u6570<\/td><td>384<\/td><\/tr><tr><td>\u6bcf\u4e2a\u4e13\u5bb6\u4e2d\u95f4 hidden \u7ef4\u5ea6<\/td><td>3072<\/td><\/tr><tr><td>\u6bcf token \u6fc0\u6d3b\u4e13\u5bb6\u6570<\/td><td>6<\/td><\/tr><tr><td>MTP \u6df1\u5ea6<\/td><td>1<\/td><\/tr><tr><td>\u524d 3 \u4e2a MoE \u5c42<\/td><td>Hash \u8def\u7531<\/td><\/tr><tr><td>mHC \u6269\u5c55\u56e0\u5b50 $n_{hc}$<\/td><td>4<\/td><\/tr><tr><td>Sinkhorn-Knopp \u8fed\u4ee3\u6570<\/td><td>20<\/td><\/tr><tr><td><strong>\u603b\u53c2\u6570 \/ \u6fc0\u6d3b\u53c2\u6570<\/strong><\/td><td><strong>1.6T \/ 49B<\/strong><\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"422-\u8bad\u7ec3\u8bbe\u7f6e\">4.2.2 \u8bad\u7ec3\u8bbe\u7f6e<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>DeepSeek-V4-Flash<\/strong>\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u4f18\u5316\u5668\uff1a\u5927\u90e8\u5206\u53c2\u6570\u7528 Muon\uff0cembedding\/prediction head\/RMSNorm \u6743\u91cd\u7528 AdamW\u3002<\/li>\n\n\n\n<li>AdamW \u8d85\u53c2\uff1a$\\beta_1 = 0.9, \\beta_2 = 0.95, \\varepsilon = 10^{-20}$, weight_decay = 0.1\u3002<\/li>\n\n\n\n<li>Muon\uff1a\u52a8\u91cf 0.95\uff0c\u6743\u91cd\u8870\u51cf 0.1\uff0crescale \u6bcf\u4e2a\u66f4\u65b0\u77e9\u9635\u7684 RMS \u4e3a 0.18\u3002<\/li>\n\n\n\n<li>\u8bad\u7ec3\u6570\u636e\uff1a<strong>32T tokens<\/strong>\u3002<\/li>\n\n\n\n<li>Batch size\uff1a\u4ece\u5c0f\u503c\u8c03\u5ea6\u5230 75.5M tokens\uff0c\u5728\u5927\u90e8\u5206\u8bad\u7ec3\u4e2d\u7ef4\u6301\u3002<\/li>\n\n\n\n<li>\u5b66\u4e60\u7387\uff1a\u524d 2000 \u6b65\u7ebf\u6027 warmup\uff0c\u5927\u90e8\u5206\u8bad\u7ec3\u7ef4\u6301\u5728 $2.7 \\times 10^{-4}$\uff0c\u6700\u540e\u6309 cosine schedule \u8870\u51cf\u5230 $2.7 \\times 10^{-5}$\u3002<\/li>\n\n\n\n<li>\u5e8f\u5217\u957f\u5ea6\uff1a\u4ece 4K \u9010\u6b65\u6269\u5c55\u5230 16K\u300164K\u30011M\u3002<\/li>\n\n\n\n<li>\u7a00\u758f\u6ce8\u610f\u529b\u7b56\u7565\uff1a\u524d 1T tokens \u4f7f\u7528 dense \u6ce8\u610f\u529b warmup\uff0c\u5728 64K \u5e8f\u5217\u957f\u5ea6\u5f15\u5165\u7a00\u758f\u6ce8\u610f\u529b\u5e76\u4fdd\u6301\u5230\u8bad\u7ec3\u7ed3\u675f\u3002<\/li>\n\n\n\n<li>\u8f85\u52a9\u65e0\u635f\u8d1f\u8f7d\u5747\u8861 bias \u66f4\u65b0\u901f\u5ea6\uff1a0.001\u3002<\/li>\n\n\n\n<li>\u5e73\u8861\u635f\u5931\u6743\u91cd\uff1a0.0001\u3002<\/li>\n\n\n\n<li>MTP \u635f\u5931\u6743\u91cd\uff1a\u5927\u90e8\u5206\u8bad\u7ec3\u4e3a 0.3\uff0c\u5b66\u4e60\u7387\u8870\u51cf\u5f00\u59cb\u65f6\u8bbe\u4e3a 0.1\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>DeepSeek-V4-Pro<\/strong>\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u8bbe\u7f6e\u4e0e Flash \u57fa\u672c\u4e00\u81f4\uff0c\u5dee\u5f02\u9879\uff1a\n<ul class=\"wp-block-list\">\n<li>\u8bad\u7ec3\u6570\u636e\uff1a<strong>33T tokens<\/strong><\/li>\n\n\n\n<li>\u6700\u5927 batch size\uff1a<strong>94.4M tokens<\/strong><\/li>\n\n\n\n<li>\u5cf0\u503c\u5b66\u4e60\u7387\uff1a$2.0 \\times 10^{-4}$\uff0c\u7ed3\u675f\u5b66\u4e60\u7387\uff1a$2.0 \\times 10^{-5}$<\/li>\n\n\n\n<li>dense \u6ce8\u610f\u529b\u9636\u6bb5\u66f4\u957f<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>\u5176\u4f59\u4e0e Flash \u76f8\u540c\uff08\u8d77\u59cb\u5e8f\u5217 4K \u2192 16K \u2192 64K \u2192 1M\uff09\u3002<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"423-\u7f13\u89e3\u8bad\u7ec3\u4e0d\u7a33\u5b9a\u6027\">4.2.3 \u7f13\u89e3\u8bad\u7ec3\u4e0d\u7a33\u5b9a\u6027<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">\u8bad\u7ec3\u4e07\u4ebf\u53c2\u6570 MoE \u6a21\u578b\u9762\u4e34\u663e\u8457\u7a33\u5b9a\u6027\u6311\u6218\uff0cV4 \u4e5f\u4e0d\u4f8b\u5916\u3002<strong>\u7ecf\u9a8c\u6027\u5730\u53d1\u73b0 spike \u7684\u53d1\u751f\u59cb\u7ec8\u4e0e MoE \u5c42\u4e2d\u7684 outlier \u76f8\u5173\uff0c\u8def\u7531\u673a\u5236\u672c\u8eab\u4f3c\u4e4e\u52a0\u5267\u8fd9\u4e9b outlier \u7684\u51fa\u73b0<\/strong>\u3002\u56e0\u6b64 V4 \u4ece\u4e24\u4e2a\u7ef4\u5ea6\u5e94\u5bf9\uff1a\u6253\u7834\u8def\u7531\u5f15\u8d77\u7684\u6076\u6027\u5faa\u73af\uff0c\u76f4\u63a5\u6291\u5236\u5f02\u5e38\u503c\u3002<\/p>\n\n\n\n<h5 class=\"wp-block-heading\" id=\"anticipatory-routing\uff08\u9884\u671f\u8def\u7531\uff09\">Anticipatory Routing\uff08\u9884\u671f\u8def\u7531\uff09<\/h5>\n\n\n\n<p class=\"wp-block-paragraph\">\u89e3\u8026\u9aa8\u5e72\u7f51\u7edc\u548c\u8def\u7531\u7f51\u7edc\u7684\u540c\u6b65\u66f4\u65b0\u663e\u8457\u63d0\u5347\u8bad\u7ec3\u7a33\u5b9a\u6027\u3002\u5728\u7b2c $t$ \u6b65\uff0c\u4f7f\u7528\u5f53\u524d\u7f51\u7edc\u53c2\u6570 $\\theta_t$ \u8fdb\u884c\u7279\u5f81\u8ba1\u7b97\uff0c\u4f46<strong>\u8def\u7531\u7d22\u5f15\u4f7f\u7528\u5386\u53f2\u7f51\u7edc\u53c2\u6570 $\\theta_{t-\\Delta t}$ \u8ba1\u7b97\u548c\u5e94\u7528<\/strong>\u3002\u4e3a\u907f\u514d\u4e24\u6b21\u52a0\u8f7d\u6a21\u578b\u53c2\u6570\u7684\u5f00\u9500\uff0c\u5728\u7b2c $t - \\Delta t$ \u6b65\u63d0\u524d\u83b7\u53d6\u7b2c $t$ \u6b65\u7684\u6570\u636e\uff0c\"\u9884\u671f\u6027\" \u5730\u8ba1\u7b97\u5e76\u7f13\u5b58\u8def\u7531\u7d22\u5f15\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u57fa\u7840\u8bbe\u65bd\u7ea7\u4f18\u5316<\/strong>\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u4ed4\u7ec6\u7f16\u6392\u6d41\u6c34\u7ebf\u6267\u884c\u548c\u8ba1\u7b97\u4e0e EP \u901a\u4fe1\u7684\u91cd\u53e0\uff0cAnticipatory Routing \u7684**\u989d\u5916 wall-clock \u5f00\u9500\u7ea6\u4e3a 20%**\u3002<\/li>\n\n\n\n<li>\u5f15\u5165\u81ea\u52a8\u68c0\u6d4b\u673a\u5236\uff1a\u4ec5\u5728 loss spike \u65f6\u89e6\u53d1\u77ed\u6682\u56de\u6eda\u5e76\u6fc0\u6d3b Anticipatory Routing\uff0c\u8fd0\u884c\u4e00\u6bb5\u65f6\u95f4\u540e\u6062\u590d\u6807\u51c6\u8bad\u7ec3\u3002\u52a8\u6001\u5e94\u7528\u4f7f\u5f97\u53ef\u4ee5<strong>\u4ee5\u53ef\u5ffd\u7565\u7684\u989d\u5916\u8bad\u7ec3\u5f00\u9500\u907f\u514d loss spike\uff0c\u4e14\u4e0d\u635f\u5bb3\u6a21\u578b\u6027\u80fd<\/strong>\u3002<\/li>\n<\/ul>\n\n\n\n<h5 class=\"wp-block-heading\" id=\"swiglu-clamping\">SwiGLU Clamping<\/h5>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728 V4-Flash \u548c V4-Pro \u8bad\u7ec3\u4e2d\uff0c\u5c06 <strong>SwiGLU \u7684\u7ebf\u6027\u5206\u91cf clamp \u5230 [-10, 10]\uff0cgate \u5206\u91cf\u4e0a\u9650\u5c01\u9876\u4e3a 10<\/strong>\u3002\u7ecf\u9a8c\u8bc1\u660e\u8fd9\u6709\u6548\u6d88\u9664 outlier\uff0c\u663e\u8457\u7a33\u5b9a\u8bad\u7ec3\u8fc7\u7a0b\uff0c\u4e14\u4e0d\u635f\u5bb3\u6027\u80fd\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"43-\u8bc4\u4f30\">4.3 \u8bc4\u4f30<\/h3>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"431-\u8bc4\u4f30\u57fa\u51c6\">4.3.1 \u8bc4\u4f30\u57fa\u51c6<\/h4>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u4e16\u754c\u77e5\u8bc6<\/strong>\uff1aAGIEval\u3001C-Eval\u3001CMMLU\u3001MMLU\u3001MMLU-Redux\u3001MMLU-Pro\u3001MMMLU\u3001MultiLoKo\u3001Simple-QA verified\u3001SuperGPQA\u3001FACTS Parametric\u3001TriviaQA<\/li>\n\n\n\n<li><strong>\u8bed\u8a00\u7406\u89e3\u4e0e\u63a8\u7406<\/strong>\uff1aBBH\u3001DROP\u3001HellaSwag\u3001CLUEWSC\u3001WinoGrande<\/li>\n\n\n\n<li><strong>\u7f16\u7a0b\u4e0e\u6570\u5b66<\/strong>\uff1aBigCodeBench\u3001HumanEval\u3001GSM8K\u3001MATH\u3001MGSM\u3001CMath<\/li>\n\n\n\n<li><strong>\u957f\u4e0a\u4e0b\u6587<\/strong>\uff1aLongBench-V2<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"432-\u8bc4\u4f30\u7ed3\u679c-\ud83c\udfc6\">4.3.2 \u8bc4\u4f30\u7ed3\u679c \ud83c\udfc6<\/h4>\n\n\n\n<h5 class=\"wp-block-heading\" id=\"table-1\uff1abase-\u6a21\u578b\u5bf9\u6bd4\">Table 1\uff1aBase \u6a21\u578b\u5bf9\u6bd4<\/h5>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\"><strong>\u5dee\u8ddd\u5728 0.3 \u4ee5\u5185\u89c6\u4e3a\u540c\u6c34\u5e73<\/strong>\u3002<strong>\u7c97\u4f53\u4e3a\u6700\u9ad8\u5206\uff0c\u4e0b\u5212\u7ebf\u4e3a\u7b2c\u4e8c<\/strong>\u3002<\/p>\n<\/blockquote>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>\u57fa\u51c6\uff08\u6307\u6807\uff09<\/th><th># Shots<\/th><th><strong>V3.2 Base<\/strong> (671B \/ 37B)<\/th><th><strong>V4-Flash Base<\/strong> (284B \/ 13B)<\/th><th><strong>V4-Pro Base<\/strong> (1.6T \/ 49B)<\/th><\/tr><\/thead><tbody><tr><td><strong>\ud83c\udf0d \u4e16\u754c\u77e5\u8bc6<\/strong><\/td><td><\/td><td><\/td><td><\/td><td><\/td><\/tr><tr><td>AGIEval (EM)<\/td><td>0-shot<\/td><td>80.1<\/td><td>82.6<\/td><td><strong>83.1<\/strong><\/td><\/tr><tr><td>MMLU (EM)<\/td><td>5-shot<\/td><td>87.8<\/td><td>88.7<\/td><td><strong>90.1<\/strong><\/td><\/tr><tr><td>MMLU-Redux (EM)<\/td><td>5-shot<\/td><td>87.5<\/td><td>89.4<\/td><td><strong>90.8<\/strong><\/td><\/tr><tr><td>MMLU-Pro (EM)<\/td><td>5-shot<\/td><td>65.5<\/td><td>68.3<\/td><td><strong>73.5<\/strong><\/td><\/tr><tr><td>MMMLU (EM)<\/td><td>5-shot<\/td><td>87.9<\/td><td>88.8<\/td><td><strong>90.3<\/strong><\/td><\/tr><tr><td>C-Eval (EM)<\/td><td>5-shot<\/td><td>90.4<\/td><td>92.1<\/td><td><strong>93.1<\/strong><\/td><\/tr><tr><td>CMMLU (EM)<\/td><td>5-shot<\/td><td>88.9<\/td><td>90.4<\/td><td><strong>90.8<\/strong><\/td><\/tr><tr><td>MultiLoKo (EM)<\/td><td>5-shot<\/td><td>38.7<\/td><td>42.2<\/td><td><strong>51.1<\/strong><\/td><\/tr><tr><td>Simple-QA verified (EM)<\/td><td>25-shot<\/td><td>28.3<\/td><td>30.1<\/td><td><strong>55.2<\/strong><\/td><\/tr><tr><td>SuperGPQA (EM)<\/td><td>5-shot<\/td><td>45.0<\/td><td>46.5<\/td><td><strong>53.9<\/strong><\/td><\/tr><tr><td>FACTS Parametric (EM)<\/td><td>25-shot<\/td><td>27.1<\/td><td>33.9<\/td><td><strong>62.6<\/strong><\/td><\/tr><tr><td>TriviaQA (EM)<\/td><td>5-shot<\/td><td>83.3<\/td><td>82.8<\/td><td><strong>85.6<\/strong><\/td><\/tr><tr><td><strong>\ud83d\udd24 \u8bed\u8a00\u7406\u89e3\u4e0e\u63a8\u7406<\/strong><\/td><td><\/td><td><\/td><td><\/td><td><\/td><\/tr><tr><td>BBH (EM)<\/td><td>3-shot<\/td><td><strong>87.6<\/strong><\/td><td>86.9<\/td><td>87.5<\/td><\/tr><tr><td>DROP (F1)<\/td><td>1-shot<\/td><td>88.2<\/td><td>88.6<\/td><td><strong>88.7<\/strong><\/td><\/tr><tr><td>HellaSwag (EM)<\/td><td>0-shot<\/td><td>86.4<\/td><td>85.7<\/td><td><strong>88.0<\/strong><\/td><\/tr><tr><td>WinoGrande (EM)<\/td><td>0-shot<\/td><td>78.9<\/td><td>79.5<\/td><td><strong>81.5<\/strong><\/td><\/tr><tr><td>CLUEWSC (EM)<\/td><td>5-shot<\/td><td>83.5<\/td><td>82.2<\/td><td><strong>85.2<\/strong><\/td><\/tr><tr><td><strong>\ud83d\udcbb \u7f16\u7a0b\u4e0e\u6570\u5b66<\/strong><\/td><td><\/td><td><\/td><td><\/td><td><\/td><\/tr><tr><td>BigCodeBench (Pass@1)<\/td><td>3-shot<\/td><td><strong>63.9<\/strong><\/td><td>56.8<\/td><td>59.2<\/td><\/tr><tr><td>HumanEval (Pass@1)<\/td><td>0-shot<\/td><td>62.8<\/td><td>69.5<\/td><td><strong>76.8<\/strong><\/td><\/tr><tr><td>GSM8K (EM)<\/td><td>8-shot<\/td><td>91.1<\/td><td>90.8<\/td><td><strong>92.6<\/strong><\/td><\/tr><tr><td>MATH (EM)<\/td><td>4-shot<\/td><td>60.5<\/td><td>57.4<\/td><td><strong>64.5<\/strong><\/td><\/tr><tr><td>MGSM (EM)<\/td><td>8-shot<\/td><td>81.3<\/td><td><strong>85.7<\/strong><\/td><td>84.4<\/td><\/tr><tr><td>CMath (EM)<\/td><td>3-shot<\/td><td>92.6<\/td><td><strong>93.6<\/strong><\/td><td>90.9<\/td><\/tr><tr><td><strong>\ud83d\udcdc \u957f\u4e0a\u4e0b\u6587<\/strong><\/td><td><\/td><td><\/td><td><\/td><td><\/td><\/tr><tr><td>LongBench-V2 (EM)<\/td><td>1-shot<\/td><td>40.2<\/td><td>44.7<\/td><td><strong>51.5<\/strong><\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<h5 class=\"wp-block-heading\" id=\"\ud83d\udd0d-\u5173\u952e\u5bf9\u6bd4\u7ed3\u8bba\">\ud83d\udd0d \u5173\u952e\u5bf9\u6bd4\u7ed3\u8bba<\/h5>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>V4-Flash-Base vs V3.2-Base<\/strong>\uff08\u6fc0\u6d3b\u548c\u603b\u53c2\u6570\u90fd\u663e\u8457\u66f4\u5c11\uff09\uff1a<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">\u2b50 \u5c3d\u7ba1\u53c2\u6570\u89c4\u6a21\u5927\u5e45\u7f29\u51cf\uff08\u6fc0\u6d3b 13B vs 37B\uff0c\u603b\u8ba1 284B vs 671B\uff09\uff0cV4-Flash-Base \u5728<strong>\u5927\u591a\u6570\u57fa\u51c6<\/strong>\u4e0a\u4ecd\u8d85\u8d8a V3.2-Base\uff0c\u5728<strong>\u4e16\u754c\u77e5\u8bc6\u548c\u957f\u4e0a\u4e0b\u6587<\/strong>\u573a\u666f\u4e2d\u4f18\u52bf\u5c24\u4e3a\u660e\u663e\u3002\u67b6\u6784\u6539\u8fdb\u3001\u6570\u636e\u8d28\u91cf\u63d0\u5347\u548c\u8bad\u7ec3\u4f18\u5316\u8ba9\u66f4\u7cbe\u7b80\u7684\u53c2\u6570\u9884\u7b97\u4e0b\u4ea4\u4ed8\u4e86\u66f4\u4f18\u6027\u80fd\u3002<\/p>\n<\/blockquote>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>V4-Pro-Base vs V3.2-Base \/ V4-Flash-Base<\/strong>\uff1a<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">\ud83c\udfc6 <strong>V4-Pro-Base \u51e0\u4e4e\u5168\u65b9\u4f4d\u9886\u5148<\/strong>\uff0c\u5728\u51e0\u4e4e\u6240\u6709\u7c7b\u522b\u4e2d\u8868\u73b0\u51fa\u663e\u8457\u63d0\u5347\uff0c\u5728\u6700\u5177\u6311\u6218\u6027\u7684\u57fa\u51c6\u4e0a\u8fbe\u5230 DeepSeek base \u6a21\u578b\u7684<strong>\u65b0\u9ad8<\/strong>\u3002\u5728\u77e5\u8bc6\u5bc6\u96c6\u578b\u8bc4\u4f30\u4e2d\u63d0\u5347\u5c24\u4e3a\u663e\u8457\uff0c\u5728\u957f\u4e0a\u4e0b\u6587\u7406\u89e3\u65b9\u9762\u4e5f\u6709\u5b9e\u8d28\u6027\u8fdb\u6b65\u3002\u5728\u5927\u591a\u6570\u63a8\u7406\u548c\u4ee3\u7801\u57fa\u51c6\u4e0a\u4e5f\u8d85\u8d8a\u524d\u4e24\u8005\u3002\u8fd9\u8bc1\u5b9e V4-Pro-Base \u662f DeepSeek \u7cfb\u5217<strong>\u6700\u5f3a\u7684\u57fa\u7840\u6a21\u578b<\/strong>\u3002<\/p>\n<\/blockquote>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"5-\u540e\u8bad\u7ec3\">5. \u540e\u8bad\u7ec3<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"51-\u540e\u8bad\u7ec3\u7ba1\u7ebf\">5.1 \u540e\u8bad\u7ec3\u7ba1\u7ebf<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u9884\u8bad\u7ec3\u4e4b\u540e\u8fdb\u884c\u540e\u8bad\u7ec3\u4ee5\u4ea7\u51fa V4 \u7cfb\u5217\u6700\u7ec8\u6a21\u578b\u3002\u8bad\u7ec3\u7ba1\u7ebf\u5927\u4f53\u6cbf\u7528 V3.2\uff0c<strong>\u4f46\u505a\u4e86\u5173\u952e\u65b9\u6cd5\u66ff\u6362\uff1a\u6df7\u5408 RL \u9636\u6bb5\u88ab\u5b8c\u5168\u66ff\u6362\u4e3a On-Policy Distillation\uff08OPD\uff0c\u7b56\u7565\u5185\u84b8\u998f\uff09<\/strong>\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"511-\u4e13\u5bb6\u8bad\u7ec3\uff08specialist-training\uff09\">5.1.1 \u4e13\u5bb6\u8bad\u7ec3\uff08Specialist Training\uff09<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">\u6bcf\u4e2a\u9886\u57df\u6a21\u578b\u901a\u8fc7\u521d\u59cb fine-tuning + RL \u987a\u5e8f\u4f18\u5316\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>RL \u9636\u6bb5\u4f7f\u7528 <strong>GRPO\uff08Group Relative Policy Optimization\uff09\u7b97\u6cd5<\/strong>\u3002<\/li>\n\n\n\n<li>\u8d85\u53c2\u4e0e DeepSeek-R1\u3001V3 \u7cfb\u5217\u4fdd\u6301\u4e00\u81f4\u3002<\/li>\n<\/ul>\n\n\n\n<h5 class=\"wp-block-heading\" id=\"\u63a8\u7406\u52aa\u529b\uff08reasoning-efforts\uff09\">\u63a8\u7406\u52aa\u529b\uff08Reasoning Efforts\uff09<\/h5>\n\n\n\n<p class=\"wp-block-paragraph\">V4-Pro \u548c V4-Flash \u5747\u652f\u6301\u4e09\u79cd\u63a8\u7406\u52aa\u529b\u6a21\u5f0f\u3002\u6bcf\u79cd\u6a21\u5f0f\u5728 RL \u8bad\u7ec3\u4e2d\u65bd\u52a0\u4e0d\u540c\u7684\u957f\u5ea6\u60e9\u7f5a\u548c\u4e0a\u4e0b\u6587\u7a97\u53e3\uff0c\u4ea7\u751f\u4e0d\u540c\u957f\u5ea6\u7684\u63a8\u7406\u8f93\u51fa\u3002\u4f7f\u7528 <code>&lt;think&gt;<\/code> \u548c <code>&lt;\/think&gt;<\/code> \u6807\u7b7e\u5212\u5206\u63a8\u7406\u5185\u5bb9\u3002<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>\u63a8\u7406\u6a21\u5f0f<\/th><th>\u7279\u5f81<\/th><th>\u5178\u578b\u573a\u666f<\/th><th>\u54cd\u5e94\u683c\u5f0f<\/th><\/tr><\/thead><tbody><tr><td><strong>Non-think<\/strong><\/td><td>\u57fa\u4e8e\u4e60\u60ef\u6216\u7b80\u5355\u89c4\u5219\u7684\u5feb\u901f\u76f4\u89c9\u54cd\u5e94<\/td><td>\u65e5\u5e38\u4efb\u52a1\u3001\u7d27\u6025\u53cd\u5e94\u3001\u4f4e\u98ce\u9669\u51b3\u7b56<\/td><td><code>&lt;\/think&gt; summary<\/code><\/td><\/tr><tr><td><strong>Think High<\/strong><\/td><td>\u6709\u610f\u8bc6\u7684\u903b\u8f91\u5206\u6790\uff0c\u8f83\u6162\u4f46\u66f4\u51c6\u786e<\/td><td>\u590d\u6742\u95ee\u9898\u89e3\u51b3\u3001\u89c4\u5212\u3001\u4e2d\u7b49\u98ce\u9669\u51b3\u7b56<\/td><td><code>&lt;think&gt; thinking tokens &lt;\/think&gt; summary<\/code><\/td><\/tr><tr><td><strong>Think Max<\/strong><\/td><td>\u5c06\u63a8\u7406\u63a8\u5230\u6781\u81f4\uff0c\u6162\u4f46\u5f3a\u5927<\/td><td>\u63a2\u7d22\u6a21\u578b\u63a8\u7406\u80fd\u529b\u8fb9\u754c<\/td><td>1. \u7cfb\u7edf\u63d0\u793a\u5f00\u5934\u7684\u7279\u6b8a\u6307\u4ee4\uff1b2. <code>&lt;think&gt; thinking tokens &lt;\/think&gt; summary<\/code><\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Think Max \u7684\u7cfb\u7edf\u63d0\u793a\u6ce8\u5165\u6307\u4ee4<\/strong>\uff08\u7ffb\u8bd1\u6458\u8981\uff09\uff1a<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">\u63a8\u7406\u52aa\u529b\uff1a\u7edd\u5bf9\u6700\u5927\uff0c\u4e0d\u5141\u8bb8\u8d70\u6377\u5f84\u3002\u4f60\u5fc5\u987b\u975e\u5e38\u5f7b\u5e95\u5730\u601d\u8003\u5e76\u5168\u9762\u5206\u89e3\u95ee\u9898\u4ee5\u89e3\u51b3\u6839\u672c\u539f\u56e0\uff0c\u9488\u5bf9\u6240\u6709\u6f5c\u5728\u8def\u5f84\u3001\u8fb9\u7f18\u60c5\u51b5\u548c\u5bf9\u6297\u573a\u666f\u4e25\u683c\u538b\u529b\u6d4b\u8bd5\u4f60\u7684\u903b\u8f91\u3002\u660e\u786e\u5199\u51fa\u6574\u4e2a\u63a8\u7406\u8fc7\u7a0b\uff0c\u8bb0\u5f55\u6bcf\u4e2a\u4e2d\u95f4\u6b65\u9aa4\u3001\u8003\u8651\u7684\u5907\u9009\u65b9\u6848\u548c\u88ab\u62d2\u7edd\u7684\u5047\u8bbe\uff0c\u786e\u4fdd\u6ca1\u6709\u4efb\u4f55\u5047\u8bbe\u672a\u7ecf\u68c0\u9a8c\u3002<\/p>\n<\/blockquote>\n\n\n\n<h5 class=\"wp-block-heading\" id=\"\u751f\u6210\u5f0f\u5956\u52b1\u6a21\u578b\uff08generative-reward-model-grm\uff09\">\u751f\u6210\u5f0f\u5956\u52b1\u6a21\u578b\uff08Generative Reward Model, GRM\uff09<\/h5>\n\n\n\n<p class=\"wp-block-paragraph\">\u5bf9\u4e8e<strong>\u96be\u4ee5\u9a8c\u8bc1\u7684\u4efb\u52a1<\/strong>\uff0cV4 \u6452\u5f03\u4f20\u7edf\u7684\u6807\u91cf\u5956\u52b1\u6a21\u578b\uff0c\u8f6c\u800c\u7b56\u5212 <strong>rubric-guided RL \u6570\u636e<\/strong>\u5e76\u4f7f\u7528 GRM \u8bc4\u4f30\u7b56\u7565\u8f68\u8ff9\u3002<strong>\u5173\u952e\u521b\u65b0\uff1aRL \u4f18\u5316\u76f4\u63a5\u5e94\u7528\u4e8e GRM \u672c\u8eab<\/strong>\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Actor \u7f51\u7edc\u539f\u751f\u5145\u5f53 GRM\uff0c\u8054\u5408\u4f18\u5316\u6a21\u578b\u7684\u8bc4\u4f30\uff08\u5224\u65ad\uff09\u80fd\u529b\u4e0e\u6807\u51c6\u751f\u6210\u80fd\u529b\u3002<\/li>\n\n\n\n<li>\u7edf\u4e00\u8fd9\u4e24\u4e2a\u89d2\u8272\u540e\uff0c\u6a21\u578b\u5185\u90e8\u63a8\u7406\u80fd\u529b\u81ea\u7136\u878d\u5165\u8bc4\u4f30\u8fc7\u7a0b\uff0c\u5f97\u5230\u9ad8\u5ea6\u7a33\u5065\u7684\u8bc4\u5206\u3002<\/li>\n\n\n\n<li>\u6b64\u65b9\u6cd5\u53ea\u9700<strong>\u6781\u5c11\u91cf\u591a\u6837\u5316\u7684\u4eba\u5de5\u6807\u6ce8<\/strong>\u5373\u53ef\u5b9e\u73b0\u4f18\u5f02\u6027\u80fd\u3002<\/li>\n<\/ul>\n\n\n\n<h5 class=\"wp-block-heading\" id=\"\u5de5\u5177\u8c03\u7528-schema-\u4e0e\u7279\u6b8a-token\">\u5de5\u5177\u8c03\u7528 Schema \u4e0e\u7279\u6b8a Token<\/h5>\n\n\n\n<p class=\"wp-block-paragraph\">V4 \u4f7f\u7528\u4e13\u7528 <code>&lt;think&gt;&lt;\/think&gt;<\/code> \u6807\u7b7e\u5212\u5206\u63a8\u7406\u8def\u5f84\u3002V4 \u5f15\u5165\u4e86<strong>\u65b0\u7684\u5de5\u5177\u8c03\u7528 schema<\/strong>\uff0c\u91c7\u7528 <code>|DSML|<\/code> \u7279\u6b8a token \u548c\u57fa\u4e8e XML \u7684\u683c\u5f0f\u3002<strong>\u5b9e\u9a8c\u8868\u660e XML \u683c\u5f0f\u6709\u6548\u51cf\u5c11\u8f6c\u4e49\u5931\u8d25\u5e76\u964d\u4f4e\u5de5\u5177\u8c03\u7528\u9519\u8bef<\/strong>\uff0c\u4e3a\u6a21\u578b-\u5de5\u5177\u4ea4\u4e92\u63d0\u4f9b\u66f4\u5065\u58ee\u7684\u63a5\u53e3\u3002<\/p>\n\n\n\n<h5 class=\"wp-block-heading\" id=\"\u4ea4\u9519\u601d\u8003\uff08interleaved-thinking\uff09\">\u4ea4\u9519\u601d\u8003\uff08Interleaved Thinking\uff09<\/h5>\n\n\n\n<p class=\"wp-block-paragraph\">V3.2 \u5f15\u5165\u4e86\u8de8\u5de5\u5177\u7ed3\u679c\u8f6e\u6b21\u4fdd\u7559\u63a8\u7406\u8ffd\u8e2a\u4f46\u5728\u65b0\u7528\u6237\u6d88\u606f\u5230\u6765\u65f6\u4e22\u5f03\u7684\u4e0a\u4e0b\u6587\u7ba1\u7406\u7b56\u7565\u3002\u8fd9\u5728\u590d\u6742 agent \u5de5\u4f5c\u6d41\u4e2d\u4ecd\u4f1a\u9020\u6210\u4e0d\u5fc5\u8981\u7684 token \u6d6a\u8d39\u2014\u2014\u6bcf\u6b21\u65b0\u7528\u6237\u8f6e\u6b21\u90fd\u4f1a\u6e05\u7a7a\u6240\u6709\u7d2f\u79ef\u63a8\u7406\u5185\u5bb9\uff0c\u5f3a\u8feb\u6a21\u578b\u91cd\u5efa\u95ee\u9898\u89e3\u51b3\u72b6\u6001\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>V4 \u5229\u7528 1M token \u4e0a\u4e0b\u6587\u7a97\u53e3\u7684\u6269\u5c55\uff0c\u8fdb\u4e00\u6b65\u4f18\u5316\u6b64\u673a\u5236<\/strong>\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u5de5\u5177\u8c03\u7528\u573a\u666f<\/strong>\uff1a\u6240\u6709\u63a8\u7406\u5185\u5bb9\u5728<strong>\u6574\u4e2a\u5bf9\u8bdd\u4e2d\u5b8c\u6574\u4fdd\u7559<\/strong>\uff0c\u5305\u62ec\u8de8\u7528\u6237\u6d88\u606f\u8fb9\u754c\u3002\u8fd9\u8ba9\u6a21\u578b\u5728\u957f\u7a0b agent \u4efb\u52a1\u4e2d\u7ef4\u6301\u8fde\u8d2f\u3001\u7d2f\u79ef\u7684\u601d\u7ef4\u94fe\u3002<\/li>\n\n\n\n<li><strong>\u4e00\u822c\u5bf9\u8bdd\u573a\u666f<\/strong>\uff1a\u4fdd\u7559\u539f\u7b56\u7565\uff0c\u65b0\u7528\u6237\u6d88\u606f\u5230\u6765\u65f6\u4e22\u5f03\u4e4b\u524d\u8f6e\u6b21\u7684\u63a8\u7406\u5185\u5bb9\uff0c\u4e0a\u4e0b\u6587\u4fdd\u6301\u7b80\u6d01\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u26a0\ufe0f \u5bf9\u4e8e\u901a\u8fc7\u7528\u6237\u6d88\u606f\u6a21\u62df\u5de5\u5177\u4ea4\u4e92\u7684 agent \u6846\u67b6\uff08\u5982 Terminus\uff09\uff0c\u53ef\u80fd\u4e0d\u89e6\u53d1\u5de5\u5177\u8c03\u7528\u4e0a\u4e0b\u6587\u8def\u5f84\uff0c\u65e0\u6cd5\u53d7\u76ca\u4e8e\u589e\u5f3a\u7684\u63a8\u7406\u6301\u4e45\u6027\u3002V4 \u5efa\u8bae\u6b64\u7c7b\u67b6\u6784\u4f7f\u7528 non-think \u6a21\u578b\u3002<\/p>\n\n\n\n<h5 class=\"wp-block-heading\" id=\"quick-instruction\uff08\u5feb\u901f\u6307\u4ee4\uff09\">Quick Instruction\uff08\u5feb\u901f\u6307\u4ee4\uff09<\/h5>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728 chatbot \u573a\u666f\u4e2d\uff0c\u8bb8\u591a\u8f85\u52a9\u4efb\u52a1\uff08\u5224\u65ad\u662f\u5426\u89e6\u53d1 Web \u641c\u7d22\u3001\u610f\u56fe\u8bc6\u522b\u7b49\uff09\u5fc5\u987b\u5728\u751f\u6210\u54cd\u5e94\u524d\u6267\u884c\uff0c\u4f20\u7edf\u7531\u72ec\u7acb\u5c0f\u6a21\u578b\u5904\u7406\uff0c\u5b58\u5728\u5197\u4f59 prefill\u3002V4 \u5f15\u5165 Quick Instruction\uff1a<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u76f4\u63a5\u5c06\u4e00\u7ec4\u4e13\u7528\u7279\u6b8a token \u9644\u52a0\u5230\u8f93\u5165\u5e8f\u5217\uff0c\u6bcf\u4e2a token \u5bf9\u5e94\u7279\u5b9a\u8f85\u52a9\u4efb\u52a1\u3002\u901a\u8fc7\u76f4\u63a5\u590d\u7528\u5df2\u8ba1\u7b97\u7684 KV cache\uff0c\u5b8c\u5168\u907f\u514d\u5197\u4f59 prefill<\/strong>\uff0c\u5141\u8bb8\u67d0\u4e9b\u4efb\u52a1\uff08\u5982\u751f\u6210\u641c\u7d22\u67e5\u8be2\u3001\u5224\u65ad\u6743\u5a01\u6027\u548c\u9886\u57df\uff09<strong>\u5e76\u884c\u6267\u884c<\/strong>\u3002\u8fd9\u663e\u8457\u964d\u4f4e\u7528\u6237\u611f\u77e5\u7684 TTFT\uff08time-to-first-token\uff09\uff0c\u5e76\u6d88\u9664\u7ef4\u62a4\u989d\u5916\u5c0f\u6a21\u578b\u7684\u5de5\u7a0b\u5f00\u9500\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u652f\u6301\u7684 Quick Instruction \u7279\u6b8a token<\/strong>\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>\u7279\u6b8a Token<\/th><th>\u8bf4\u660e<\/th><\/tr><\/thead><tbody><tr><td>`&lt;<\/td><td>action<\/td><\/tr><tr><td>`&lt;<\/td><td>title<\/td><\/tr><tr><td>`&lt;<\/td><td>query<\/td><\/tr><tr><td>`&lt;<\/td><td>authority<\/td><\/tr><tr><td>`&lt;<\/td><td>domain<\/td><\/tr><tr><td>`&lt;<\/td><td>extracted_url<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"512-on-policy-distillation\uff08opd\uff09\">5.1.2 On-Policy Distillation\uff08OPD\uff09<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">\u901a\u8fc7\u4e13\u95e8 fine-tuning \u548c RL \u8bad\u7ec3\u51fa\u591a\u4e2a\u9886\u57df\u4e13\u5bb6\u540e\uff0cV4 \u91c7\u7528<strong>\u591a\u6559\u5e08 OPD<\/strong> \u4f5c\u4e3a\u5c06\u4e13\u5bb6\u80fd\u529b\u5408\u5e76\u5230\u6700\u7ec8\u6a21\u578b\u7684\u4e3b\u8981\u6280\u672f\u3002<strong>\u5b66\u751f\u901a\u8fc7\u5728\u81ea\u5df1\u751f\u6210\u7684\u8f68\u8ff9\u4e0a\u5b66\u4e60\u6559\u5e08\u6a21\u578b\u7684\u8f93\u51fa\u5206\u5e03\u5b9e\u73b0<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7ed9\u5b9a $N$ \u4e2a\u4e13\u5bb6\u6a21\u578b\u96c6\u5408 ${\\pi_{E_1}, \\ldots, \\pi_{E_N}}$\uff0cOPD \u76ee\u6807\u51fd\u6570\u5b9a\u4e49\u4e3a\uff1a<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">$$\\mathcal{L}<em>{\\text{OPD}}(\\theta) = \\sum<\/em>{i=1}^{N} w_i \\cdot D_{\\text{KL}}(\\pi_\\theta | \\pi_{E_i})$$<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5176\u4e2d $w_i$ \u662f\u6bcf\u4e2a\u4e13\u5bb6\u7684\u6743\u91cd\uff08\u901a\u5e38\u7531\u76f8\u5bf9\u91cd\u8981\u6027\u51b3\u5b9a\uff09\u3002\u8ba1\u7b97\u53cd\u5411 KL \u635f\u5931\u9700\u8981\u4ece\u5b66\u751f $\\pi_\\theta$ \u91c7\u6837\u8bad\u7ec3\u8f68\u8ff9\u4ee5\u7ef4\u6301 on-policy \u5b66\u4e60\u3002\u7edf\u4e00\u7b56\u7565 $\\pi_\\theta$ \u4f1a<strong>\u9009\u62e9\u6027\u5730\u4ece\u4e0e\u5f53\u524d\u4efb\u52a1\u4e0a\u4e0b\u6587\u76f8\u5173\u7684\u4e13\u5bb6\u5b66\u4e60<\/strong>\uff08\u5982\u6570\u5b66\u63a8\u7406\u4efb\u52a1\u5bf9\u9f50\u6570\u5b66\u4e13\u5bb6\uff0c\u7f16\u7a0b\u4efb\u52a1\u5bf9\u9f50\u7f16\u7a0b\u4e13\u5bb6\uff09\u3002\u901a\u8fc7\u6b64\u673a\u5236\uff0c\u6765\u81ea\u7269\u7406\u4e0a\u4e0d\u540c\u4e13\u5bb6\u6743\u91cd\u7684\u77e5\u8bc6\u901a\u8fc7 logits \u7ea7\u5bf9\u9f50\u6574\u5408\u5230\u7edf\u4e00\u53c2\u6570\u7a7a\u95f4\u4e2d\uff0c<strong>\u5b9e\u9645\u89c4\u907f\u4e86\u4f20\u7edf\u6743\u91cd\u5408\u5e76\u6216\u6df7\u5408 RL \u6280\u672f\u4e2d\u5e38\u9047\u5230\u7684\u6027\u80fd\u9000\u5316<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u6b64\u9636\u6bb5\u4f7f\u7528\u8d85\u8fc7 10 \u4e2a\u8986\u76d6\u5404\u9886\u57df\u7684\u6559\u5e08\u6a21\u578b<\/strong>\u84b8\u998f\u5355\u4e2a\u5b66\u751f\u6a21\u578b\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e0e\u5148\u524d\u5de5\u4f5c\u5c06\u5b8c\u6574\u8bcd\u8868 KL \u635f\u5931\u7b80\u5316\u4e3a token \u7ea7 KL \u4f30\u8ba1\u7684\u505a\u6cd5\u4e0d\u540c\uff0cV4 \u91c7\u7528 <strong>\u5b8c\u6574\u8bcd\u8868 logit \u84b8\u998f<\/strong>\u3002\u4fdd\u7559\u5b8c\u6574 logit \u5206\u5e03\u8ba1\u7b97\u53cd\u5411 KL \u635f\u5931\u53ef\u4ee5\u4ea7\u751f\u66f4\u7a33\u5b9a\u7684\u68af\u5ea6\u4f30\u8ba1\uff0c\u5e76\u5fe0\u5b9e\u84b8\u998f\u6559\u5e08\u77e5\u8bc6\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"52-rl-\u548c-opd-\u57fa\u7840\u8bbe\u65bd\">5.2 RL \u548c OPD \u57fa\u7840\u8bbe\u65bd<\/h3>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"521-fp4-\u91cf\u5316\u96c6\u6210\">5.2.1 FP4 \u91cf\u5316\u96c6\u6210<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">\u5c06 FP4\uff08MXFP4\uff09\u91cf\u5316\u5e94\u7528\u4e8e\u52a0\u901f rollout \u548c\u6240\u6709\u4ec5\u63a8\u7406\u7684\u524d\u5411\uff08\u5305\u62ec\u6559\u5e08\u548c\u53c2\u8003\u6a21\u578b\uff09\uff0c\u51cf\u5c11\u5185\u5b58\u6d41\u91cf\u548c\u91c7\u6837\u5ef6\u8fdf\u3002rollout \u548c\u63a8\u7406\u9636\u6bb5\u76f4\u63a5\u4f7f\u7528\u539f\u751f FP4 \u6743\u91cd\uff1b\u8bad\u7ec3\u6b65\u9aa4\u901a\u8fc7\u65e0\u635f FP4 \u5230 FP8 \u53cd\u91cf\u5316\u6a21\u62df FP4 \u91cf\u5316\uff0c\u65e0\u9700\u4fee\u6539\u53cd\u5411\u7ba1\u7ebf\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"522-\u5b8c\u6574\u8bcd\u8868-opd-\u7684\u9ad8\u6548\u6559\u5e08\u8c03\u5ea6\">5.2.2 \u5b8c\u6574\u8bcd\u8868 OPD \u7684\u9ad8\u6548\u6559\u5e08\u8c03\u5ea6<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">V4 \u6846\u67b6\u652f\u6301\u5b8c\u6574\u8bcd\u8868 OPD\uff0c<strong>\u7406\u8bba\u4e0a\u6559\u5e08\u6570\u91cf\u4e0d\u53d7\u9650\u5236<\/strong>\uff0c\u6bcf\u4e2a\u6559\u5e08\u53ef\u80fd\u6709\u4e07\u4ebf\u53c2\u6570\u3002\u6280\u672f\u7ec6\u8282\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u6743\u91cd\u5378\u8f7d<\/strong>\uff1a\u6240\u6709\u6559\u5e08\u6743\u91cd\u5378\u8f7d\u5230\u96c6\u4e2d\u5f0f\u5206\u5e03\u5f0f\u5b58\u50a8\uff0c\u5728\u6559\u5e08\u524d\u5411\u65f6\u6309\u9700\u52a0\u8f7d\uff0c\u91c7\u7528 ZeRO \u6837\u53c2\u6570\u5206\u7247\u51cf\u8f7b I\/O \u548c DRAM \u538b\u529b\u3002<\/li>\n\n\n\n<li><strong>Hidden state \u7f13\u5b58<\/strong>\uff1a\u7531\u4e8e\u8bcd\u8868 $|V| > 100k$\uff0c\u540c\u65f6\u4e3a\u6240\u6709\u6559\u5e08\u7269\u5316 logits \u6210\u672c\u8fc7\u9ad8\uff08\u5373\u4f7f spool \u5230\u78c1\u76d8\u4e5f\u4e0d\u884c\uff09\u3002V4 \u524d\u5411\u65f6<strong>\u4ec5\u5728\u96c6\u4e2d\u5f0f\u7f13\u51b2\u533a\u7f13\u5b58\u6700\u540e\u4e00\u5c42\u6559\u5e08 hidden states<\/strong>\uff0c\u8bad\u7ec3\u65f6\u901a\u8fc7\u5bf9\u5e94 prediction head \u6a21\u5757\u5b9e\u65f6\u91cd\u6784\u5b8c\u6574 logits\u3002\u91cd\u8ba1\u7b97\u5f00\u9500\u53ef\u5ffd\u7565\uff0c\u5b8c\u5168\u89c4\u907f\u663e\u5f0f logits \u7269\u5316\u7684\u5185\u5b58\u8d1f\u62c5\u3002<\/li>\n\n\n\n<li><strong>\u6559\u5e08\u8c03\u5ea6<\/strong>\uff1a\u6309\u6559\u5e08\u7d22\u5f15\u6392\u5e8f\u8bad\u7ec3\u6837\u672c\u5206\u53d1\uff0c\u786e\u4fdd\u6bcf\u4e2a\u4e0d\u540c\u6559\u5e08 head \u6bcf\u4e2a mini-batch \u53ea\u52a0\u8f7d\u4e00\u6b21\uff0c\u4efb\u4f55\u65f6\u523b\u81f3\u591a\u4e00\u4e2a\u6559\u5e08 head \u9a7b\u7559\u8bbe\u5907\u5185\u5b58\u3002<\/li>\n\n\n\n<li><strong>\u5f02\u6b65 I\/O<\/strong>\uff1a\u6240\u6709\u53c2\u6570\u548c hidden state \u52a0\u8f7d\/\u5378\u8f7d\u64cd\u4f5c\u5728\u540e\u53f0\u5f02\u6b65\u8fdb\u884c\uff0c\u4e0d\u963b\u585e\u5173\u952e\u8def\u5f84\u8ba1\u7b97\u3002<\/li>\n\n\n\n<li><strong>KL \u8ba1\u7b97<\/strong>\uff1a\u4f7f\u7528\u4e13\u7528 TileLang kernel \u8ba1\u7b97\u6559\u5e08-\u5b66\u751f logits \u95f4\u7684\u7cbe\u786e KL \u6563\u5ea6\uff0c\u52a0\u901f\u8ba1\u7b97\u5e76\u51cf\u5c11\u52a8\u6001\u5185\u5b58\u5206\u914d\u3002<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"523-\u53ef\u62a2\u5360\u4e14\u5bb9\u9519\u7684-rollout-\u670d\u52a1\">5.2.3 \u53ef\u62a2\u5360\u4e14\u5bb9\u9519\u7684 Rollout \u670d\u52a1<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">GPU \u96c6\u7fa4\u91c7\u7528\u96c6\u7fa4\u8303\u56f4\u62a2\u5360\u5f0f\u4efb\u52a1\u8c03\u5ea6\u5668\uff0c\u4efb\u4f55\u8fd0\u884c\u4efb\u52a1\u53ef\u968f\u65f6\u88ab\u62a2\u5360\u3002V4 \u5b9e\u73b0 <strong>token \u7c92\u5ea6\u7684 Write-Ahead Log\uff08WAL\uff09<\/strong>\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u6bcf\u6b21\u4e3a\u8bf7\u6c42\u751f\u6210\u65b0 token \u7acb\u5373\u8ffd\u52a0\u5230\u8be5\u8bf7\u6c42\u7684 WAL\u3002<\/li>\n\n\n\n<li>\u62a2\u5360\u65f6\u6682\u505c\u63a8\u7406\u5f15\u64ce\uff0c\u4fdd\u5b58\u672a\u5b8c\u6210\u8bf7\u6c42\u7684 KV cache\u3002<\/li>\n\n\n\n<li>\u6062\u590d\u65f6\u4f7f\u7528\u6301\u4e45\u5316\u7684 WAL \u548c\u4fdd\u5b58\u7684 KV cache \u7ee7\u7eed\u89e3\u7801\u3002<\/li>\n\n\n\n<li>\u53d1\u751f\u81f4\u547d\u786c\u4ef6\u9519\u8bef\u65f6\uff0c\u53ef\u4f7f\u7528 WAL \u4e2d\u6301\u4e45\u5316\u7684 token \u91cd\u65b0\u8fd0\u884c prefill \u9636\u6bb5\u91cd\u5efa KV cache\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u91cd\u8981\u539f\u7406<\/strong>\uff1a\u4ece\u5934\u91cd\u65b0\u751f\u6210\u672a\u5b8c\u6210\u8bf7\u6c42\u5728\u6570\u5b66\u4e0a\u4e0d\u6b63\u786e\uff08\u4f1a\u5f15\u5165\u957f\u5ea6\u504f\u5dee\u2014\u2014\u8f83\u77ed\u54cd\u5e94\u66f4\u53ef\u80fd\u5728\u4e2d\u65ad\u540e\u5e78\u5b58\uff0c\u4f7f\u6a21\u578b\u5728\u4e2d\u65ad\u53d1\u751f\u65f6\u66f4\u503e\u5411\u4ea7\u751f\u8f83\u77ed\u5e8f\u5217\uff09\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"524-\u767e\u4e07-token-\u4e0a\u4e0b\u6587\u7684-rl-\u6846\u67b6\u6269\u5c55\">5.2.4 \u767e\u4e07 Token \u4e0a\u4e0b\u6587\u7684 RL \u6846\u67b6\u6269\u5c55<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">\u9488\u5bf9\u767e\u4e07 token \u5e8f\u5217\u7684\u9ad8\u6548 RL \u548c OPD \u76ee\u6807\u4f18\u5316\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>rollout \u9636\u6bb5\u91c7\u7528\u53ef\u62a2\u5360\u5bb9\u9519\u670d\u52a1\uff08\u89c1 5.2.3\uff09\u3002<\/li>\n\n\n\n<li>\u63a8\u7406\u8bad\u7ec3\u9636\u6bb5\uff0c\u5c06 rollout \u6570\u636e\u683c\u5f0f\u5206\u89e3\u4e3a<strong>\u8f7b\u91cf\u7ea7\u5143\u6570\u636e\u548c\u91cd\u578b\u6bcf token \u5b57\u6bb5<\/strong>\u3002\u6570\u636e\u5206\u53d1\u65f6\uff0c\u52a0\u8f7d\u6574\u4e2a rollout \u6570\u636e\u7684\u5143\u6570\u636e\u6267\u884c\u5168\u5c40 shuffle \u548c packing layout \u8ba1\u7b97\u3002<\/li>\n\n\n\n<li>\u91cd\u578b\u6bcf token \u5b57\u6bb5\u901a\u8fc7<strong>\u5171\u4eab\u5185\u5b58\u6570\u636e\u52a0\u8f7d\u5668<\/strong>\u52a0\u8f7d\uff0c\u6d88\u9664\u8282\u70b9\u5185\u6570\u636e\u5197\u4f59\uff0c\u5728 mini-batch \u7c92\u5ea6\u6d88\u8d39\u65f6\u7acb\u5373\u91ca\u653e\uff0c\u5927\u5e45\u964d\u4f4e CPU \u548c GPU \u5185\u5b58\u538b\u529b\u3002<\/li>\n\n\n\n<li>\u8bbe\u5907\u4e0a mini-batch \u6570\u91cf\u6839\u636e\u5de5\u4f5c\u8d1f\u8f7d\u52a8\u6001\u786e\u5b9a\uff0c\u5728\u8ba1\u7b97\u541e\u5410\u548c I\/O \u91cd\u53e0\u95f4\u9ad8\u6548\u6743\u8861\u3002<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"525-agentic-ai-\u7684\u6c99\u7bb1\u57fa\u7840\u8bbe\u65bd\">5.2.5 Agentic AI \u7684\u6c99\u7bb1\u57fa\u7840\u8bbe\u65bd<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e3a\u6ee1\u8db3\u540e\u8bad\u7ec3\u548c\u8bc4\u4f30\u4e2d agentic AI \u7684\u591a\u6837\u5316\u6267\u884c\u9700\u6c42\uff0cV4 \u6784\u5efa\u4e86\u751f\u4ea7\u7ea7\u6c99\u7bb1\u5e73\u53f0 <strong>DSec\uff08DeepSeek Elastic Compute\uff09<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">DSec \u7531\u4e09\u4e2a Rust \u7ec4\u4ef6\u6784\u6210\uff1a<strong>Apiserver<\/strong>\uff08API \u7f51\u5173\uff09\u3001<strong>Edge<\/strong>\uff08\u6bcf\u4e3b\u673a\u4ee3\u7406\uff09\u3001<strong>Watcher<\/strong>\uff08\u96c6\u7fa4\u76d1\u63a7\u5668\uff09\u3002\u901a\u8fc7\u81ea\u5b9a\u4e49 RPC \u534f\u8bae\u4e92\u8fde\uff0c\u5728 3FS \u5206\u5e03\u5f0f\u6587\u4ef6\u7cfb\u7edf\u4e0a\u6c34\u5e73\u6269\u5c55\u3002<strong>\u751f\u4ea7\u4e2d\u5355\u4e2a DSec \u96c6\u7fa4\u7ba1\u7406\u6570\u5341\u4e07\u5e76\u53d1\u6c99\u7bb1\u5b9e\u4f8b<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">DSec \u7684 4 \u9879\u6838\u5fc3\u8bbe\u8ba1\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><p><strong>\u4e00\u4e2a\u7edf\u4e00\u63a5\u53e3\u80cc\u540e\u7684\u56db\u4e2a\u6267\u884c\u57fa\u5e95<\/strong>\uff1aPython SDK <code>libdsec<\/code> \u62bd\u8c61 4 \u79cd\u57fa\u5e95\uff1a<\/p>\n<ul class=\"wp-block-list\">\n<li><strong>Function Call<\/strong>\uff1a\u6d3e\u53d1\u65e0\u72b6\u6001\u8c03\u7528\u5230\u9884\u70ed\u5bb9\u5668\u6c60\uff0c\u6d88\u9664\u51b7\u542f\u52a8\u5f00\u9500\u3002<\/li>\n\n\n\n<li><strong>Container<\/strong>\uff1a\u5b8c\u5168 Docker \u517c\u5bb9\uff0c\u5229\u7528 EROFS \u6309\u9700\u52a0\u8f7d\u5b9e\u73b0\u9ad8\u6548\u955c\u50cf\u7ec4\u88c5\u3002<\/li>\n\n\n\n<li><strong>microVM<\/strong>\uff1a\u57fa\u4e8e Firecracker\uff0c\u4e3a\u5b89\u5168\u654f\u611f\u9ad8\u5bc6\u5ea6\u90e8\u7f72\u6dfb\u52a0 VM \u7ea7\u9694\u79bb\u3002<\/li>\n\n\n\n<li><strong>fullVM<\/strong>\uff1a\u57fa\u4e8e QEMU\uff0c\u652f\u6301\u4efb\u610f\u5ba2\u6237\u64cd\u4f5c\u7cfb\u7edf\u3002<\/li>\n\n\n\n<li>\u56db\u8005\u5171\u4eab\u901a\u7528 API surface\uff08\u547d\u4ee4\u6267\u884c\u3001\u6587\u4ef6\u4f20\u8f93\u3001TTY \u8bbf\u95ee\uff09\uff0c\u5207\u6362\u53ea\u9700\u53c2\u6570\u66f4\u6539\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><p><strong>\u901a\u8fc7\u5206\u5c42\u5b58\u50a8\u5b9e\u73b0\u5feb\u901f\u955c\u50cf\u52a0\u8f7d<\/strong>\uff1a<\/p>\n<ul class=\"wp-block-list\">\n<li><strong>\u5bb9\u5668<\/strong>\uff1a\u57fa\u7840\u955c\u50cf\u548c\u6587\u4ef6\u7cfb\u7edf\u63d0\u4ea4\u5b58\u50a8\u4e3a 3FS \u652f\u6301\u7684\u53ea\u8bfb EROFS \u5c42\uff0c\u76f4\u63a5\u6302\u8f7d\u5230 overlay lowerdirs\u3002\u6587\u4ef6\u5143\u6570\u636e\u5728\u6302\u8f7d\u65f6\u5728\u672c\u5730\u78c1\u76d8\u7acb\u5373\u53ef\u7528\uff1b\u6570\u636e\u5757\u6309\u9700\u4ece 3FS \u83b7\u53d6\u3002<\/li>\n\n\n\n<li><strong>microVM<\/strong>\uff1a\u4f7f\u7528 overlaybd \u78c1\u76d8\u683c\u5f0f\u3002\u53ea\u8bfb\u57fa\u7840\u5c42\u9a7b\u7559\u5728 3FS \u4e0a\u4f9b\u8de8\u5b9e\u4f8b\u5171\u4eab\uff0c\u5199\u5165\u672c\u5730 copy-on-write \u5c42\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><p><strong>\u5927\u89c4\u6a21\u5e76\u53d1\u4e0b\u7684\u5bc6\u5ea6\u4f18\u5316<\/strong>\uff1a<\/p>\n<ul class=\"wp-block-list\">\n<li>\u7f13\u89e3\u865a\u62df\u5316\u73af\u5883\u4e2d\u91cd\u590d\u7684 page-cache \u5360\u7528\uff0c\u5e76\u5e94\u7528\u5185\u5b58\u56de\u6536\u5b9e\u73b0\u5b89\u5168\u8d85\u5206\u914d\u3002<\/li>\n\n\n\n<li>\u7f13\u89e3\u5bb9\u5668\u8fd0\u884c\u65f6\u4e2d\u7684 spinlock \u7ade\u4e89\uff0c\u964d\u4f4e\u6bcf\u6c99\u7bb1 CPU \u5f00\u9500\uff0c\u663e\u8457\u63d0\u9ad8\u6bcf\u4e3b\u673a\u6253\u5305\u5bc6\u5ea6\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><p><strong>\u8f68\u8ff9\u65e5\u5fd7\u548c\u62a2\u5360\u5b89\u5168\u6062\u590d<\/strong>\uff1a\u6bcf\u4e2a\u6c99\u7bb1\u7ef4\u62a4\u5168\u5c40\u6709\u5e8f\u8f68\u8ff9\u65e5\u5fd7\uff0c\u6301\u4e45\u8bb0\u5f55\u6bcf\u6b21\u547d\u4ee4\u8c03\u7528\u53ca\u7ed3\u679c\u3002\u4e09\u4e2a\u7528\u9014\uff1a<\/p>\n<ul class=\"wp-block-list\">\n<li><strong>\u5ba2\u6237\u7aef\u5feb\u901f\u524d\u8fdb<\/strong>\uff1a\u8bad\u7ec3\u4efb\u52a1\u88ab\u62a2\u5360\u65f6\u6c99\u7bb1\u8d44\u6e90\u4fdd\u7559\uff0c\u6062\u590d\u65f6 DSec \u91cd\u653e\u5148\u524d\u5b8c\u6210\u547d\u4ee4\u7684\u7f13\u5b58\u7ed3\u679c\u52a0\u901f\u6062\u590d\u3002<\/li>\n\n\n\n<li><strong>\u7ec6\u7c92\u5ea6\u6eaf\u6e90<\/strong>\uff1a\u6bcf\u4e2a\u72b6\u6001\u53d8\u5316\u7684\u6765\u6e90\u548c\u5bf9\u5e94\u7ed3\u679c\u53ef\u8ffd\u6eaf\u3002<\/li>\n\n\n\n<li><strong>\u786e\u5b9a\u6027\u56de\u653e<\/strong>\uff1a\u4efb\u4f55\u5386\u53f2\u4f1a\u8bdd\u53ef\u4ece\u5176\u8f68\u8ff9\u5fe0\u5b9e\u590d\u73b0\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ol>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"53-\u6807\u51c6\u57fa\u51c6\u8bc4\u4f30\">5.3 \u6807\u51c6\u57fa\u51c6\u8bc4\u4f30<\/h3>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"531-\u8bc4\u4f30\u8bbe\u7f6e\">5.3.1 \u8bc4\u4f30\u8bbe\u7f6e<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u77e5\u8bc6\u4e0e\u63a8\u7406\u6570\u636e\u96c6<\/strong>\uff1aMMLU-Pro\u3001GPQA\u3001Human Last Exam\uff08HLE\uff09\u3001Simple-QA Verified\u3001Chinese-SimpleQA\u3001LiveCodeBench-v6\u3001CodeForces\uff08\u5185\u90e8\u57fa\u51c6\uff09\u3001HMMT 2026 Feb\u3001Apex\u3001Apex Shortlist\u3001IMOAnswerBench\u3001PutnamBench\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u4ee3\u7801 Codeforces<\/strong>\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u6536\u96c6 2025 \u5e74 5 \u6708\u81f3 11 \u6708\u7684 14 \u573a Codeforces Division 1 \u6bd4\u8d5b\uff0c\u5171 <strong>114 \u9053\u9898<\/strong>\u3002<\/li>\n\n\n\n<li>\u6bcf\u9898\u751f\u6210 <strong>32 \u4e2a\u5019\u9009\u65b9\u6848<\/strong>\u3002<\/li>\n\n\n\n<li>\u6bcf\u9898\u72ec\u7acb\u91c7\u6837 10 \u4e2a\u65b9\u6848\uff08\u4e0d\u653e\u56de\uff09\uff0c\u968f\u673a\u6392\u5e8f\u5f62\u6210\u63d0\u4ea4\u5e8f\u5217\uff0c\u7531\u9886\u57df\u4e13\u5bb6\u6784\u5efa\u7684\u6d4b\u8bd5\u5957\u4ef6\u5224\u9898\u3002<\/li>\n\n\n\n<li>\u91c7\u7528 OpenAI \u7684\u60e9\u7f5a\u65b9\u6848\uff1a\u89e3\u9898\u5f97\u5206\u4e3a\u540c\u6837\u5931\u8d25\u6b21\u6570\u524d\u63d0\u4e0b\u901a\u8fc7\u540c\u9898\u76ee\u7684\u4eba\u7c7b\u53c2\u4e0e\u8005\u7684\u4e2d\u4f4d\u6570\u5f97\u5206\u3002<\/li>\n\n\n\n<li>\u901a\u8fc7 Codeforces \u6807\u51c6\u8bc4\u7ea7\u7cfb\u7edf\u8f6c\u6362\u4e3a\u8bc4\u7ea7\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u63a8\u7406\u4e0e\u77e5\u8bc6\u4efb\u52a1\u8bbe\u7f6e<\/strong>\uff1a\u6e29\u5ea6 1.0\uff0c\u4e0a\u4e0b\u6587\u7a97\u53e3\u5206\u522b\u8bbe\u4e3a 8K\uff08Non-think\uff09\u3001128K\uff08High\uff09\u3001384K\uff08Max\uff09\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u6570\u5b66\u4efb\u52a1<\/strong>\uff08HMMT\u3001IMOAnswerBench\u3001Apex\u3001HLE\uff09\u4f7f\u7528\u6a21\u677f\uff1a<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">\"{question}\\nPlease reason step by step, and put your final answer within \\boxed{}.\"<\/p>\n<\/blockquote>\n\n\n\n<p class=\"wp-block-paragraph\">\u5bf9 V4-Pro-Max \u6570\u5b66\u4efb\u52a1\u4f7f\u7528\u66f4\u6df1\u63a8\u7406\u7684\u6a21\u677f\uff0c\u8981\u6c42\u4e25\u683c\u8bc1\u660e\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u5f62\u5f0f\u6570\u5b66\u4efb\u52a1<\/strong>\uff1a\u5728 Lean v4.28.0-rc1 \u7684 agentic \u8bbe\u7f6e\u4e0b\u8bc4\u4f30\uff0c\u8bbf\u95ee Lean \u7f16\u8bd1\u5668\u548c\u8bed\u4e49\u6218\u672f\u641c\u7d22\u5f15\u64ce\uff0c\u6700\u591a 500 \u6b21\u5de5\u5177\u8c03\u7528\uff0c\u6700\u5927\u63a8\u7406\u52aa\u529b\u3002\u9644\u52a0\u8ba1\u7b97\u5bc6\u96c6\u578b\u7ba1\u7ebf\uff1a\u5019\u9009\u81ea\u7136\u8bed\u8a00\u65b9\u6848\u5148\u7531\u81ea\u9a8c\u8bc1\u751f\u6210\u548c\u8fc7\u6ee4\uff0c\u4fdd\u7559\u7684\u65b9\u6848\u4f5c\u4e3a\u6307\u5bfc\u63d0\u4f9b\u7ed9\u5f62\u5f0f agent \u8bc1\u660e\u5bf9\u5e94 Lean \u9648\u8ff0\u3002<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">\u26a0\ufe0f \u6ce8\uff1aK2.6 \u548c GLM-5.1 \u7684\u67d0\u4e9b\u6761\u76ee\u7559\u7a7a\uff0c\u56e0\u5176 API \u592a\u5fd9\u672a\u54cd\u5e94\u67e5\u8be2\u3002<\/p>\n<\/blockquote>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>1M Token \u4e0a\u4e0b\u6587\u8bc4\u4f30<\/strong>\uff1a\u4f7f\u7528 OpenAI MRCR \u548c CorpusQA \u4f5c\u4e3a\u57fa\u51c6\u3002<strong>\u5728\u76f8\u540c\u914d\u7f6e\u4e0b\u91cd\u65b0\u8bc4\u4f30\u4e86 Claude Opus 4.6 \u548c Gemini 3.1 Pro<\/strong>\u3002GPT-5.4 \u56e0 API \u5bf9\u5927\u91cf\u67e5\u8be2\u65e0\u54cd\u5e94\u672a\u8bc4\u4f30\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Agent \u6570\u636e\u96c6<\/strong>\uff1aTerminal Bench 2.0\u3001SWE-Verified\u3001SWE Multilingual\u3001SWE-Pro\u3001BrowseComp\u3001MCPAtlas \u516c\u5f00\u8bc4\u4f30\u96c6\u3001GDPval-AA\u3001Tool-Decathlon\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u4ee3\u7801 agent \u4efb\u52a1<\/strong>\uff1a\u4f7f\u7528\u5185\u90e8\u5f00\u53d1\u6846\u67b6\uff0c\u4ec5\u63d0\u4f9b\u6700\u5c0f\u5de5\u5177\u96c6\uff08bash \u5de5\u5177 + \u6587\u4ef6\u7f16\u8f91\u5de5\u5177\uff09\uff0c\u6700\u5927\u4ea4\u4e92\u6b65\u6570 500\uff0c\u6700\u5927\u4e0a\u4e0b\u6587 512K tokens\u3002\u5728 Terminal-Bench 2.0 Verified \u5b50\u96c6\u4e0a\uff0cV4-Pro \u5f97\u5206\u7ea6 <strong>72.0<\/strong>\u3002<\/li>\n\n\n\n<li><strong>\u641c\u7d22 agent \u4efb\u52a1<\/strong>\uff1a\u4f7f\u7528\u5185\u90e8 harness\uff0cwebsearch + Python \u5de5\u5177\uff0c\u6700\u5927\u4ea4\u4e92\u6b65\u6570 500\uff0c\u6700\u5927\u4e0a\u4e0b\u6587 512K\u3002BrowseComp \u4f7f\u7528\u4e0e V3.2 \u76f8\u540c\u7684 discard-all \u4e0a\u4e0b\u6587\u7ba1\u7406\u7b56\u7565\u3002<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"532-\u8bc4\u4f30\u7ed3\u679c-\ud83c\udfc6\">5.3.2 \u8bc4\u4f30\u7ed3\u679c \ud83c\udfc6<\/h4>\n\n\n\n<h5 class=\"wp-block-heading\" id=\"table-6\uff1adeepseek-v4-pro-max-vs-\u95ed\u6e90\u5f00\u6e90\u9886\u5148\u6a21\u578b\">Table 6\uff1aDeepSeek-V4-Pro-Max vs \u95ed\u6e90\/\u5f00\u6e90\u9886\u5148\u6a21\u578b<\/h5>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">\ud83e\udd47 <strong>\u7c97\u4f53\u4e3a\u6700\u4f73\uff0c\u4e0b\u5212\u7ebf\u4e3a\u7b2c\u4e8c<\/strong>\u3002\"Max\"\u3001\"xHigh\"\u3001\"High\" \u8868\u793a\u63a8\u7406\u52aa\u529b\u7b49\u7ea7\u3002<\/p>\n<\/blockquote>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>\u57fa\u51c6 (\u6307\u6807)<\/th><th><strong>Opus-4.6 Max<\/strong><\/th><th><strong>GPT-5.4 xHigh<\/strong><\/th><th><strong>Gemini-3.1-Pro High<\/strong><\/th><th><strong>K2.6 Thinking<\/strong><\/th><th><strong>GLM-5.1 Thinking<\/strong><\/th><th><strong>DS-V4-Pro Max<\/strong><\/th><\/tr><\/thead><tbody><tr><td><strong>\ud83e\udde0 \u77e5\u8bc6\u4e0e\u63a8\u7406<\/strong><\/td><td><\/td><td><\/td><td><\/td><td><\/td><td><\/td><td><\/td><\/tr><tr><td>MMLU-Pro (EM)<\/td><td>89.1<\/td><td>87.5<\/td><td><strong>91.0<\/strong> \ud83e\udd47<\/td><td>87.1<\/td><td>86.0<\/td><td>87.5<\/td><\/tr><tr><td>SimpleQA-Verified (Pass@1)<\/td><td>46.2<\/td><td>45.3<\/td><td><strong>75.6<\/strong> \ud83e\udd47<\/td><td>36.9<\/td><td>38.1<\/td><td>57.9<\/td><\/tr><tr><td>Chinese-SimpleQA (Pass@1)<\/td><td>76.4<\/td><td>76.8<\/td><td><strong>85.9<\/strong> \ud83e\udd47<\/td><td>75.9<\/td><td>75.0<\/td><td>84.4<\/td><\/tr><tr><td>GPQA Diamond (Pass@1)<\/td><td>91.3<\/td><td>93.0<\/td><td><strong>94.3<\/strong> \ud83e\udd47<\/td><td>90.5<\/td><td>86.2<\/td><td>90.1<\/td><\/tr><tr><td>HLE (Pass@1)<\/td><td>40.0<\/td><td>39.8<\/td><td><strong>44.4<\/strong> \ud83e\udd47<\/td><td>36.4<\/td><td>34.7<\/td><td>37.7<\/td><\/tr><tr><td>LiveCodeBench (Pass@1)<\/td><td>88.8<\/td><td>\u2014<\/td><td>91.7<\/td><td>89.6<\/td><td>\u2014<\/td><td><strong>93.5<\/strong> \ud83e\udd47<\/td><\/tr><tr><td>Codeforces (Rating)<\/td><td>\u2014<\/td><td>3168<\/td><td>3052<\/td><td>\u2014<\/td><td>\u2014<\/td><td><strong>3206<\/strong> \ud83e\udd47<\/td><\/tr><tr><td>HMMT 2026 Feb (Pass@1)<\/td><td>96.2<\/td><td><strong>97.7<\/strong> \ud83e\udd47<\/td><td>94.7<\/td><td>92.7<\/td><td>89.4<\/td><td>95.2<\/td><\/tr><tr><td>IMOAnswerBench (Pass@1)<\/td><td>75.3<\/td><td><strong>91.4<\/strong> \ud83e\udd47<\/td><td>81.0<\/td><td>86.0<\/td><td>83.8<\/td><td>89.8<\/td><\/tr><tr><td>Apex (Pass@1)<\/td><td>34.5<\/td><td>54.1<\/td><td><strong>60.9<\/strong> \ud83e\udd47<\/td><td>24.0<\/td><td>11.5<\/td><td>38.3<\/td><\/tr><tr><td>Apex Shortlist (Pass@1)<\/td><td>85.9<\/td><td>78.1<\/td><td>89.1<\/td><td>75.5<\/td><td>72.4<\/td><td><strong>90.2<\/strong> \ud83e\udd47<\/td><\/tr><tr><td>LongMRCR 1M (MMR)<\/td><td><strong>92.9<\/strong> \ud83e\udd47<\/td><td>\u2014<\/td><td>76.3<\/td><td>\u2014<\/td><td>\u2014<\/td><td>83.5<\/td><\/tr><tr><td>CorpusQA 1M (ACC)<\/td><td><strong>71.7<\/strong> \ud83e\udd47<\/td><td>\u2014<\/td><td>53.8<\/td><td>\u2014<\/td><td>\u2014<\/td><td>62.0<\/td><\/tr><tr><td><strong>\ud83e\udd16 Agentic<\/strong><\/td><td><\/td><td><\/td><td><\/td><td><\/td><td><\/td><td><\/td><\/tr><tr><td>Terminal Bench 2.0 (Acc)<\/td><td>65.4<\/td><td><strong>75.1<\/strong> \ud83e\udd47<\/td><td>68.5<\/td><td>66.7<\/td><td>63.5<\/td><td>67.9<\/td><\/tr><tr><td>SWE Verified (Resolved)<\/td><td><strong>80.8<\/strong> \ud83e\udd47<\/td><td>\u2014<\/td><td>80.6<\/td><td>80.2<\/td><td>\u2014<\/td><td>80.6<\/td><\/tr><tr><td>SWE Pro (Resolved)<\/td><td>57.3<\/td><td>57.7<\/td><td>54.2<\/td><td><strong>58.6<\/strong> \ud83e\udd47<\/td><td>58.4<\/td><td>55.4<\/td><\/tr><tr><td>SWE Multilingual (Resolved)<\/td><td><strong>77.5<\/strong> \ud83e\udd47<\/td><td>\u2014<\/td><td>\u2014<\/td><td>76.7<\/td><td>73.3<\/td><td>76.2<\/td><\/tr><tr><td>BrowseComp (Pass@1)<\/td><td>83.7<\/td><td>82.7<\/td><td><strong>85.9<\/strong> \ud83e\udd47<\/td><td>83.2<\/td><td>79.3<\/td><td>83.4<\/td><\/tr><tr><td>HLE w\/ tools (Pass@1)<\/td><td>53.1<\/td><td>52.0<\/td><td>51.6<\/td><td><strong>54.0<\/strong> \ud83e\udd47<\/td><td>50.4<\/td><td>48.2<\/td><\/tr><tr><td>GDPval-AA (Elo)<\/td><td>1619<\/td><td><strong>1674<\/strong> \ud83e\udd47<\/td><td>1314<\/td><td>1482<\/td><td>1535<\/td><td>1554<\/td><\/tr><tr><td>MCPAtlas Public (Pass@1)<\/td><td><strong>73.8<\/strong> \ud83e\udd47<\/td><td>67.2<\/td><td>69.2<\/td><td>66.6<\/td><td>71.8<\/td><td>73.6<\/td><\/tr><tr><td>Toolathlon (Pass@1)<\/td><td>47.2<\/td><td><strong>54.6<\/strong> \ud83e\udd47<\/td><td>48.8<\/td><td>50.0<\/td><td>40.7<\/td><td>51.8<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<h5 class=\"wp-block-heading\" id=\"table-7\uff1av4-flash-\u548c-v4-pro-\u4e0d\u540c\u6a21\u5f0f\u5bf9\u6bd4\">Table 7\uff1aV4-Flash \u548c V4-Pro \u4e0d\u540c\u6a21\u5f0f\u5bf9\u6bd4<\/h5>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>\u57fa\u51c6\uff08\u6307\u6807\uff09<\/th><th><strong>V4-Flash<\/strong><br>Non-Think<\/th><th>V4-Flash<br>High<\/th><th>V4-Flash<br>Max<\/th><th><strong>V4-Pro<\/strong><br>Non-Think<\/th><th>V4-Pro<br>High<\/th><th>V4-Pro<br>Max<\/th><\/tr><\/thead><tbody><tr><td><strong>\ud83e\udde0 \u77e5\u8bc6\u4e0e\u63a8\u7406<\/strong><\/td><td><\/td><td><\/td><td><\/td><td><\/td><td><\/td><td><\/td><\/tr><tr><td>MMLU-Pro (EM)<\/td><td>83.0<\/td><td>86.4<\/td><td>86.2<\/td><td>82.9<\/td><td>87.1<\/td><td><strong>87.5<\/strong><\/td><\/tr><tr><td>SimpleQA-Verified (Pass@1)<\/td><td>23.1<\/td><td>28.9<\/td><td>34.1<\/td><td>45.0<\/td><td>46.2<\/td><td><strong>57.9<\/strong><\/td><\/tr><tr><td>Chinese-SimpleQA (Pass@1)<\/td><td>71.5<\/td><td>73.2<\/td><td>78.9<\/td><td>75.8<\/td><td>77.7<\/td><td><strong>84.4<\/strong><\/td><\/tr><tr><td>GPQA Diamond (Pass@1)<\/td><td>71.2<\/td><td>87.4<\/td><td>88.1<\/td><td>72.9<\/td><td>89.1<\/td><td><strong>90.1<\/strong><\/td><\/tr><tr><td>HLE (Pass@1)<\/td><td>8.1<\/td><td>29.4<\/td><td>34.8<\/td><td>7.7<\/td><td>34.5<\/td><td><strong>37.7<\/strong><\/td><\/tr><tr><td>LiveCodeBench (Pass@1-COT)<\/td><td>55.2<\/td><td>88.4<\/td><td>91.6<\/td><td>56.8<\/td><td>89.8<\/td><td><strong>93.5<\/strong><\/td><\/tr><tr><td>Codeforces (Rating)<\/td><td>\u2014<\/td><td>2816<\/td><td>3052<\/td><td>\u2014<\/td><td>2919<\/td><td><strong>3206<\/strong><\/td><\/tr><tr><td>HMMT 2026 Feb (Pass@1)<\/td><td>40.8<\/td><td>91.9<\/td><td>94.8<\/td><td>31.7<\/td><td>94.0<\/td><td><strong>95.2<\/strong><\/td><\/tr><tr><td>IMOAnswerBench (Pass@1)<\/td><td>41.9<\/td><td>85.1<\/td><td>88.4<\/td><td>35.3<\/td><td>88.0<\/td><td><strong>89.8<\/strong><\/td><\/tr><tr><td>Apex (Pass@1)<\/td><td>1.0<\/td><td>19.1<\/td><td>33.0<\/td><td>0.4<\/td><td>27.4<\/td><td><strong>38.3<\/strong><\/td><\/tr><tr><td>Apex Shortlist (Pass@1)<\/td><td>9.3<\/td><td>72.1<\/td><td>85.7<\/td><td>9.2<\/td><td>85.5<\/td><td><strong>90.2<\/strong><\/td><\/tr><tr><td><strong>\ud83d\udcdc \u957f\u4e0a\u4e0b\u6587<\/strong><\/td><td><\/td><td><\/td><td><\/td><td><\/td><td><\/td><td><\/td><\/tr><tr><td>MRCR 1M (MMR)<\/td><td>37.5<\/td><td>76.9<\/td><td>78.7<\/td><td>44.7<\/td><td>83.3<\/td><td><strong>83.5<\/strong><\/td><\/tr><tr><td>CorpusQA 1M (ACC)<\/td><td>15.5<\/td><td>59.3<\/td><td>60.5<\/td><td>35.6<\/td><td>56.5<\/td><td><strong>62.0<\/strong><\/td><\/tr><tr><td><strong>\ud83e\udd16 Agentic<\/strong><\/td><td><\/td><td><\/td><td><\/td><td><\/td><td><\/td><td><\/td><\/tr><tr><td>Terminal Bench 2.0 (Acc)<\/td><td>49.1<\/td><td>56.6<\/td><td>56.9<\/td><td>59.1<\/td><td>63.3<\/td><td><strong>67.9<\/strong><\/td><\/tr><tr><td>SWE Verified (Resolved)<\/td><td>73.7<\/td><td>78.6<\/td><td>79.0<\/td><td>73.6<\/td><td>79.4<\/td><td><strong>80.6<\/strong><\/td><\/tr><tr><td>SWE Pro (Resolved)<\/td><td>49.1<\/td><td>52.3<\/td><td>52.6<\/td><td>52.1<\/td><td>54.4<\/td><td><strong>55.4<\/strong><\/td><\/tr><tr><td>SWE Multilingual (Resolved)<\/td><td>69.7<\/td><td>70.2<\/td><td>73.3<\/td><td>69.8<\/td><td>74.1<\/td><td><strong>76.2<\/strong><\/td><\/tr><tr><td>BrowseComp (Pass@1)<\/td><td>\u2014<\/td><td>53.5<\/td><td>73.2<\/td><td>\u2014<\/td><td>80.4<\/td><td><strong>83.4<\/strong><\/td><\/tr><tr><td>HLE w\/ tools (Pass@1)<\/td><td>\u2014<\/td><td>40.3<\/td><td>45.1<\/td><td>\u2014<\/td><td>44.7<\/td><td><strong>48.2<\/strong><\/td><\/tr><tr><td>MCPAtlas Public (Pass@1)<\/td><td>64.0<\/td><td>67.4<\/td><td>69.0<\/td><td>69.4<\/td><td><strong>74.2<\/strong><\/td><td>73.6<\/td><\/tr><tr><td>GDPval-AA (Elo)<\/td><td>\u2014<\/td><td>\u2014<\/td><td>1395<\/td><td>\u2014<\/td><td>\u2014<\/td><td><strong>1554<\/strong><\/td><\/tr><tr><td>Toolathlon (Pass@1)<\/td><td>40.7<\/td><td>43.5<\/td><td>47.8<\/td><td>46.3<\/td><td>49.0<\/td><td><strong>51.8<\/strong><\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<h5 class=\"wp-block-heading\" id=\"\ud83d\udd0d-\u7ed3\u679c\u6df1\u5ea6\u89e3\u8bfb\">\ud83d\udd0d \u7ed3\u679c\u6df1\u5ea6\u89e3\u8bfb<\/h5>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\ud83c\udf0d \u77e5\u8bc6\u8bc4\u4f30<\/strong>\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>V4-Pro-Max \u5728 <strong>SimpleQA-Verified<\/strong> \u4e0a<strong>\u9886\u5148\u6240\u6709\u5f00\u6e90\u57fa\u7ebf 20 \u7edd\u5bf9\u767e\u5206\u70b9<\/strong>\uff0c\u4f46\u4ecd\u843d\u540e\u9886\u5148\u95ed\u6e90\u6a21\u578b Gemini-3.1-Pro\u3002<\/li>\n\n\n\n<li>\u5728\u6559\u80b2\u77e5\u8bc6\u548c\u63a8\u7406\uff08MMLU-Pro\u3001GPQA\u3001HLE\uff09\u4e0a\uff0cV4-Pro-Max \u7565\u8d85 Kimi \u548c GLM\uff0c\u4f46\u843d\u540e\u9876\u7ea7\u95ed\u6e90\u6a21\u578b\u3002<\/li>\n\n\n\n<li>V4-Pro-Max \u6807\u5fd7\u7740\u5f00\u6e90\u6a21\u578b\u4e16\u754c\u77e5\u8bc6\u80fd\u529b\u7684<strong>\u91cd\u5927\u91cc\u7a0b\u7891<\/strong>\u3002<\/li>\n\n\n\n<li>V4-Flash \u5728\u77e5\u8bc6\u4efb\u52a1\u4e0a\u4e0e V4-Pro \u5b58\u5728\u660e\u663e\u5dee\u8ddd\uff08\u7b26\u5408\u9884\u671f\uff0c\u53c2\u6570\u8d8a\u5927\u77e5\u8bc6\u4fdd\u7559\u8d8a\u591a\uff09\uff1b\u66f4\u9ad8\u63a8\u7406\u52aa\u529b\u53ef\u63d0\u5347\u4e24\u8005\u5728\u77e5\u8bc6\u57fa\u51c6\u7684\u8868\u73b0\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\ud83e\uddee \u63a8\u7406\u80fd\u529b<\/strong>\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>V4-Pro-Max \u5728\u63a8\u7406\u57fa\u51c6\u4e0a<strong>\u8d85\u8d8a\u6240\u6709\u5148\u524d\u5f00\u6e90\u6a21\u578b<\/strong>\uff0c\u5728\u8bb8\u591a\u6307\u6807\u4e0a\u5339\u914d SOTA \u95ed\u6e90\u6a21\u578b\u3002<\/li>\n\n\n\n<li>V4-Flash-Max \u5728\u4ee3\u7801\u548c\u6570\u5b66\u63a8\u7406\u4efb\u52a1\u4e0a<strong>\u8d85\u8d8a\u4e86\u4e4b\u524d\u6700\u4f73\u5f00\u6e90\u6a21\u578b K2.6-Thinking<\/strong>\u3002<\/li>\n\n\n\n<li>\u5728\u7f16\u7a0b\u7ade\u8d5b\u4e2d\uff0cV4-Pro \u548c V4-Flash \u8868\u73b0\u51fa\u8272\u3002\u6027\u80fd\u4e0e GPT-5.4 \u76f8\u5f53\u2014\u2014<strong>\u8fd9\u662f\u5f00\u6e90\u6a21\u578b\u9996\u6b21\u5728\u6b64\u4efb\u52a1\u4e0a\u5339\u914d\u95ed\u6e90\u6a21\u578b<\/strong>\u3002<\/li>\n\n\n\n<li><strong>V4-Pro-Max \u5728 Codeforces \u6392\u884c\u699c\u4e0a\u5f53\u524d\u6392\u540d\u4eba\u7c7b\u9009\u624b\u4e2d\u7b2c 23 \u4f4d<\/strong>\u3002<\/li>\n\n\n\n<li>V4 \u5728\u5f62\u5f0f\u6570\u5b66\u4efb\u52a1\u4e0a\u4e5f\u8868\u73b0\u5f3a\u52b2\uff0c\u5728 agentic \u8bbe\u7f6e\u4e0b SOTA\uff0c\u8d85\u8d8a Seed Prover \u7b49\u5148\u524d\u6a21\u578b\uff1b\u5728\u8ba1\u7b97\u5bc6\u96c6\u578b\u7ba1\u7ebf\u4e0b\u8fdb\u4e00\u6b65\u63d0\u5347\uff0c\u8d85\u8d8a Aristotle \u7b49\u7cfb\u7edf\u3002<\/li>\n<\/ul>\n\n\n\n<h5 class=\"wp-block-heading\" id=\"figure-8\uff1a\u5f62\u5f0f\u63a8\u7406\u5bf9\u6bd4\">Figure 8\uff1a\u5f62\u5f0f\u63a8\u7406\u5bf9\u6bd4<\/h5>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u5b9e\u7528\u4f53\u5236\uff08Putnam-200 Pass@8\uff09<\/strong>\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>\u6a21\u578b<\/th><th>\u5f97\u5206<\/th><\/tr><\/thead><tbody><tr><td>Seed-1.5-Prover<\/td><td>26.50<\/td><\/tr><tr><td>Gemini-3-Pro<\/td><td>26.50<\/td><\/tr><tr><td>Seed-2.0-Pro<\/td><td>35.50<\/td><\/tr><tr><td><strong>DeepSeek-V4-Flash-Max<\/strong><\/td><td><strong>81.00<\/strong> \ud83c\udfc6<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u524d\u6cbf\u4f53\u5236\uff08Putnam-2025\uff0c\u6df7\u5408\u5f62\u5f0f-\u975e\u5f62\u5f0f\u63a8\u7406\uff09<\/strong>\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>\u6a21\u578b<\/th><th>\u5f97\u5206<\/th><\/tr><\/thead><tbody><tr><td>Aristotle<\/td><td>100\/120<\/td><\/tr><tr><td>Seed-1.5-Prover<\/td><td>110\/120<\/td><\/tr><tr><td>Axiom<\/td><td>120\/120<\/td><\/tr><tr><td><strong>DeepSeek-V4<\/strong><\/td><td><strong>120\/120<\/strong> \ud83c\udfc6<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\ud83e\udd16 Agent \u80fd\u529b<\/strong>\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>V4-Pro \u5728\u4ee3\u7801 agent \u4efb\u52a1\u4e0a\u4e0e K2.6 \u548c GLM-5.1 \u76f8\u5f53\uff0c\u4f46\u6240\u6709\u5f00\u6e90\u6a21\u578b\u4ecd\u843d\u540e\u95ed\u6e90\u5bf9\u624b\u3002<\/li>\n\n\n\n<li>V4-Flash \u5728\u7f16\u7a0b\u4efb\u52a1\u4e0a\u5f31\u4e8e V4-Pro\uff0c\u7279\u522b\u662f Terminal Bench 2.0\u3002<\/li>\n\n\n\n<li>V4-Pro \u5728 <strong>MCPAtlas \u548c Toolathlon<\/strong>\uff08\u5305\u542b\u5927\u91cf\u5de5\u5177\u548c MCP \u670d\u52a1\u7684\u8bc4\u4f30\u6d4b\u8bd5\u96c6\uff09\u4e0a\u8868\u73b0\u826f\u597d\uff0c\u8868\u660e <strong>V4 \u5177\u6709\u5353\u8d8a\u7684\u6cdb\u5316\u80fd\u529b\uff0c\u4e0d\u53ea\u5728\u5185\u90e8\u6846\u67b6\u4e0a\u8868\u73b0\u597d<\/strong>\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\ud83d\udcdc 1M Token \u4e0a\u4e0b\u6587<\/strong>\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>V4-Pro \u5728 MRCR \u4efb\u52a1\uff08\u6d4b\u91cf\u4e0a\u4e0b\u6587\u68c0\u7d22\uff09\u4e0a<strong>\u8d85\u8d8a Gemini-3.1-Pro<\/strong>\uff0c\u4f46\u4ecd\u843d\u540e Claude Opus 4.6\u3002<\/li>\n\n\n\n<li>128K \u4ee5\u5185\u4e0a\u4e0b\u6587\u68c0\u7d22\u6027\u80fd<strong>\u9ad8\u5ea6\u7a33\u5b9a<\/strong>\u3002\u8d85\u8fc7 128K \u540e\u51fa\u73b0\u6027\u80fd\u4e0b\u964d\uff0c\u4f46 1M token \u4e0b\u7684\u68c0\u7d22\u80fd\u529b\u4e0e\u95ed\u6e90\u548c\u5f00\u6e90\u5bf9\u624b\u76f8\u6bd4<strong>\u4ecd\u975e\u5e38\u5f3a<\/strong>\u3002<\/li>\n\n\n\n<li>\u4e0e MRCR \u4e0d\u540c\uff0cCorpusQA \u66f4\u63a5\u8fd1\u771f\u5b9e\u573a\u666f\u3002<strong>\u5728\u8be5\u57fa\u51c6\u4e0a V4-Pro \u4e5f\u4f18\u4e8e Gemini-3.1-Pro<\/strong>\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u26a1 \u63a8\u7406\u52aa\u529b<\/strong>\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Max \u6a21\u5f0f\uff08\u4f7f\u7528\u66f4\u957f\u4e0a\u4e0b\u6587\u548c RL \u4e2d\u66f4\u4f4e\u957f\u5ea6\u60e9\u7f5a\uff09\u5728\u6700\u5177\u6311\u6218\u6027\u4efb\u52a1\u4e0a\u8d85\u8d8a High \u6a21\u5f0f\u3002<\/li>\n\n\n\n<li>\u901a\u8fc7\u6269\u5c55 test-time \u8ba1\u7b97\uff0cV4 \u76f8\u5bf9\u524d\u4ee3\u5b9e\u73b0\u663e\u8457\u63d0\u5347\u3002<\/li>\n\n\n\n<li>\u5728 HLE \u7b49\u63a8\u7406\u4efb\u52a1\u4e0a\uff0c<strong>V4-Pro \u5c55\u73b0\u51fa\u6bd4 V3.2 \u66f4\u9ad8\u7684 token \u6548\u7387<\/strong>\u3002<\/li>\n<\/ul>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"54-\u771f\u5b9e\u4e16\u754c\u4efb\u52a1\u6027\u80fd\">5.4 \u771f\u5b9e\u4e16\u754c\u4efb\u52a1\u6027\u80fd<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u6807\u51c6\u5316\u57fa\u51c6\u5e38\u96be\u4ee5\u6355\u83b7\u591a\u6837\u771f\u5b9e\u4efb\u52a1\u7684\u590d\u6742\u6027\uff0c\u9020\u6210\u6d4b\u8bd5\u7ed3\u679c\u4e0e\u5b9e\u9645\u7528\u6237\u4f53\u9a8c\u7684\u5dee\u8ddd\u3002V4 \u5f00\u53d1\u4e86\u4e13\u6709\u5185\u90e8\u6307\u6807\uff0c\u4f18\u5148\u5173\u6ce8\u771f\u5b9e\u4f7f\u7528\u6a21\u5f0f\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"541-\u4e2d\u6587\u5199\u4f5c\">5.4.1 \u4e2d\u6587\u5199\u4f5c<\/h4>\n\n\n\n<h5 class=\"wp-block-heading\" id=\"\u529f\u80fd\u5199\u4f5c\uff08table-12-\u6458\u8981\uff09\">\u529f\u80fd\u5199\u4f5c\uff08Table 12 \u6458\u8981\uff09<\/h5>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>V4-Pro vs Gemini-3.1-Pro<\/strong>\uff08Gemini-3.1-Pro \u662f V4 \u8bc4\u4f30\u4e2d\u5916\u90e8\u6a21\u578b\u4e2d\u6587\u5199\u4f5c\u8868\u73b0\u6700\u4f73\u7684\uff09\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>\u5927\u7c7b<\/th><th>\u6837\u672c\u6570<\/th><th>DS \u80dc\u7387<\/th><th>Gemini \u80dc\u7387<\/th><th>\u5e73\u5c40<\/th><\/tr><\/thead><tbody><tr><td>\u529e\u516c\u6587\u672c<\/td><td>1349<\/td><td>65.16%<\/td><td>32.32%<\/td><td>2.52%<\/td><\/tr><tr><td>\u5a92\u4f53\u6587\u672c<\/td><td>666<\/td><td>57.96%<\/td><td>38.44%<\/td><td>3.60%<\/td><\/tr><tr><td>\u751f\u6d3b\u6587\u672c<\/td><td>390<\/td><td>69.49%<\/td><td>25.90%<\/td><td>4.62%<\/td><\/tr><tr><td>\u53e3\u5934\u6587\u672c<\/td><td>319<\/td><td>58.62%<\/td><td>37.62%<\/td><td>3.76%<\/td><\/tr><tr><td>\u516c\u6587\u6587\u672c<\/td><td>230<\/td><td>54.78%<\/td><td>42.17%<\/td><td>3.04%<\/td><\/tr><tr><td>\u5b66\u672f\u6587\u672c<\/td><td>216<\/td><td>63.43%<\/td><td>32.87%<\/td><td>3.70%<\/td><\/tr><tr><td><strong>\u603b\u8ba1<\/strong><\/td><td><strong>3170<\/strong><\/td><td><strong>62.65%<\/strong> \ud83c\udfc6<\/td><td><strong>34.10%<\/strong><\/td><td><strong>3.25%<\/strong><\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">\ud83d\udcdd <strong>V4-Pro \u4ee5 62.65% \u603b\u4f53\u80dc\u7387\u51fb\u8d25 Gemini-3.1-Pro<\/strong>\u3002\u4e3b\u8981\u56e0\u4e3a Gemini \u5728\u4e2d\u6587\u5199\u4f5c\u573a\u666f\u4e0b\u6709\u65f6\u4f1a\u8ba9\u5176\u56fa\u6709\u7684\u98ce\u683c\u504f\u597d\u51cc\u9a7e\u4e8e\u7528\u6237\u7684\u660e\u786e\u9700\u6c42\u4e4b\u4e0a\u3002<\/p>\n<\/blockquote>\n\n\n\n<h5 class=\"wp-block-heading\" id=\"\u521b\u610f\u5199\u4f5c\uff08table-13-\u6458\u8981\uff09\">\u521b\u610f\u5199\u4f5c\uff08Table 13 \u6458\u8981\uff09<\/h5>\n\n\n\n<p class=\"wp-block-paragraph\">\u6cbf\u4e24\u4e2a\u7ef4\u5ea6\u8bc4\u4f30\uff1a<strong>\u6307\u4ee4\u9075\u5faa<\/strong>\u548c<strong>\u5199\u4f5c\u8d28\u91cf<\/strong>\u3002<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>\u7ef4\u5ea6<\/th><th>V4-Pro \u80dc\u7387<\/th><th>Gemini \u80dc\u7387<\/th><th>\u5e73\u5c40<\/th><\/tr><\/thead><tbody><tr><td>\u6307\u4ee4\u9075\u5faa<\/td><td><strong>60.03%<\/strong><\/td><td>39.44%<\/td><td>0.53%<\/td><\/tr><tr><td>\u5199\u4f5c\u8d28\u91cf<\/td><td><strong>77.48%<\/strong> \ud83c\udfc6<\/td><td>22.35%<\/td><td>0.18%<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">\u2728 V4-Pro \u5728\u6307\u4ee4\u9075\u5faa\u4e0a\u7565\u6709\u63d0\u5347\uff0c\u5728<strong>\u5199\u4f5c\u8d28\u91cf\u4e0a\u5927\u5e45\u9886\u5148<\/strong>\u3002<\/p>\n<\/blockquote>\n\n\n\n<h5 class=\"wp-block-heading\" id=\"\u6311\u6218\u6027\u63d0\u793a\uff08table-14\uff09\">\u6311\u6218\u6027\u63d0\u793a\uff08Table 14\uff09<\/h5>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>V4-Pro vs Claude-Opus-4.5<\/strong>\uff08\u590d\u6742\u6307\u4ee4\u8ddf\u968f\u548c\u591a\u8f6e\u5199\u4f5c\uff09\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>\u7c7b\u522b<\/th><th>#<\/th><th>V4-Pro \u80dc<\/th><th>Opus \u80dc<\/th><th>\u5e73\u5c40<\/th><th>V4-Pro%<\/th><th>Opus%<\/th><th>\u5e73\u5c40%<\/th><\/tr><\/thead><tbody><tr><td>\u590d\u6742\u6307\u4ee4\u8ddf\u968f<\/td><td>49<\/td><td>23<\/td><td>26<\/td><td>0<\/td><td>46.9%<\/td><td><strong>53.1%<\/strong><\/td><td>0.0%<\/td><\/tr><tr><td>\u591a\u8f6e\u5199\u4f5c<\/td><td>147<\/td><td>67<\/td><td>76<\/td><td>4<\/td><td>45.6%<\/td><td><strong>51.7%<\/strong><\/td><td>2.7%<\/td><\/tr><tr><td><strong>\u603b\u8ba1<\/strong><\/td><td>196<\/td><td>90<\/td><td>102<\/td><td>4<\/td><td>45.9%<\/td><td><strong>52.0%<\/strong><\/td><td>2.0%<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">\u26a0\ufe0f <strong>\u5728\u9ad8\u590d\u6742\u5ea6\u7ea6\u675f\u6216\u591a\u8f6e\u573a\u666f\u7684\u6700\u5177\u6311\u6218\u6027\u63d0\u793a\u4e0b\uff0cClaude Opus 4.5 \u4ecd\u5bf9 V4-Pro \u4fdd\u6709\u6027\u80fd\u4f18\u52bf<\/strong>\uff0852.0% vs 45.9%\uff09\u3002<\/p>\n<\/blockquote>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"542-\u641c\u7d22\">5.4.2 \u641c\u7d22<\/h4>\n\n\n\n<h5 class=\"wp-block-heading\" id=\"\u68c0\u7d22\u589e\u5f3a\u641c\u7d22\uff08rag\uff09\">\u68c0\u7d22\u589e\u5f3a\u641c\u7d22\uff08RAG\uff09<\/h5>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>V4-Pro vs V3.2 \u5728\u641c\u7d22 Q&amp;A \u4efb\u52a1\u4e0a\u7684\u6bd4\u8f83<\/strong>\uff08Table 11 \u6458\u8981\uff09\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>\u7c7b\u522b<\/th><th>\u5b50\u7c7b\u522b<\/th><th>#<\/th><th>V4-Pro%<\/th><th>V3.2%<\/th><th>\u5e73\u5c40%<\/th><\/tr><\/thead><tbody><tr><td>\u5ba2\u89c2\u95ee\u7b54<\/td><td>\u5355\u503c\u4fe1\u606f\u67e5\u627e\uff08\u5355\u503c\u641c\u7d22\uff09<\/td><td>95<\/td><td><strong>37.9%<\/strong><\/td><td>10.5%<\/td><td>51.6%<\/td><\/tr><tr><td>\u5ba2\u89c2\u95ee\u7b54<\/td><td>\u5b9e\u4f53\u4fe1\u606f\u67e5\u627e<\/td><td>99<\/td><td><strong>24.2%<\/strong><\/td><td>7.1%<\/td><td>68.7%<\/td><\/tr><tr><td>\u5ba2\u89c2\u95ee\u7b54<\/td><td>\u679a\u4e3e\u578b\u4fe1\u606f\u67e5\u627e<\/td><td>95<\/td><td><strong>20.0%<\/strong><\/td><td>8.4%<\/td><td>71.6%<\/td><\/tr><tr><td>\u4e3b\u89c2\u95ee\u7b54<\/td><td>\u539f\u56e0\u5206\u6790<\/td><td>100<\/td><td><strong>28.0%<\/strong><\/td><td>5.0%<\/td><td>67.0%<\/td><\/tr><tr><td>\u4e3b\u89c2\u95ee\u7b54<\/td><td>\u5bf9\u6bd4<\/td><td>96<\/td><td><strong>29.2%<\/strong><\/td><td>20.8%<\/td><td>50.0%<\/td><\/tr><tr><td>\u4e3b\u89c2\u95ee\u7b54<\/td><td>\u5bfb\u6c42\u5efa\u8bae<\/td><td>92<\/td><td><strong>25.0%<\/strong><\/td><td>8.7%<\/td><td>66.3%<\/td><\/tr><tr><td>\u4e3b\u89c2\u95ee\u7b54<\/td><td>\u63a8\u8350<\/td><td>95<\/td><td><strong>27.4%<\/strong><\/td><td>20.0%<\/td><td>52.6%<\/td><\/tr><tr><td>\u4e3b\u89c2\u95ee\u7b54<\/td><td>\u653b\u7565\u8ba1\u5212<\/td><td>92<\/td><td><strong>34.8%<\/strong><\/td><td>12.0%<\/td><td>53.3%<\/td><\/tr><tr><td>\u4e3b\u89c2\u95ee\u7b54<\/td><td>\u8bc4\u4ef7\u770b\u6cd5<\/td><td>96<\/td><td><strong>31.2%<\/strong><\/td><td>8.3%<\/td><td>60.4%<\/td><\/tr><tr><td>\u4e3b\u89c2\u95ee\u7b54<\/td><td>\u8d8b\u52bf\u5206\u6790<\/td><td>96<\/td><td><strong>24.0%<\/strong><\/td><td>3.1%<\/td><td>72.9%<\/td><\/tr><tr><td><strong>\u603b\u8ba1<\/strong><\/td><td><\/td><td><strong>956<\/strong><\/td><td><strong>28.1%<\/strong><\/td><td><strong>10.4%<\/strong><\/td><td><strong>61.5%<\/strong><\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">\ud83d\udd0e <strong>V4-Pro \u663e\u8457\u8d85\u8d8a V3.2<\/strong>\uff0c\u5728<strong>\u5355\u503c\u641c\u7d22<\/strong>\u548c<strong>\u653b\u7565\u8ba1\u5212<\/strong>\u4efb\u52a1\u4e0a\u63d0\u5347\u6700\u663e\u8457\uff0c\u8868\u660e V4-Pro \u64c5\u957f\u5b9a\u4f4d\u7cbe\u786e\u4e8b\u5b9e\u7b54\u6848\u548c\u4ece\u68c0\u7d22\u4e0a\u4e0b\u6587\u4e2d\u5408\u6210\u7ed3\u6784\u5316\u8ba1\u5212\u3002<strong>V3.2 \u5728\u5bf9\u6bd4\u548c\u63a8\u8350\u4efb\u52a1\u4e0a\u4ecd\u76f8\u5bf9\u6709\u7ade\u4e89\u529b<\/strong>\uff0c\u6697\u793a V4-Pro \u5728\u9700\u8981\u5e73\u8861\u3001\u591a\u89c6\u89d2\u63a8\u7406\u641c\u7d22\u7ed3\u679c\u7684\u573a\u666f\u4e0a\u4ecd\u6709\u6539\u8fdb\u7a7a\u95f4\u3002<\/p>\n<\/blockquote>\n\n\n\n<h5 class=\"wp-block-heading\" id=\"agentic-\u641c\u7d22\">Agentic \u641c\u7d22<\/h5>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e0d\u540c\u4e8e\u6807\u51c6 RAG\uff0cagentic \u641c\u7d22\u5141\u8bb8\u6a21\u578b\u6309\u67e5\u8be2\u8fed\u4ee3\u8c03\u7528\u641c\u7d22\u548c\u6293\u53d6\u5de5\u5177\uff0c\u663e\u8457\u63d0\u5347\u6574\u4f53\u641c\u7d22\u6027\u80fd\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Agentic Search vs RAG \u6210\u672c\u5bf9\u6bd4<\/strong>\uff08V4-Pro\uff09\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>\u7248\u672c<\/th><th>\u5de5\u5177\u8c03\u7528\u6570<\/th><th>Prefill (tokens)<\/th><th>\u8f93\u51fa (tokens)<\/th><\/tr><\/thead><tbody><tr><td>V4 Agentic Search<\/td><td>16.2<\/td><td>13649<\/td><td>1526<\/td><\/tr><tr><td>V4 RAG<\/td><td>\u2014<\/td><td>10453<\/td><td>1308<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">\ud83d\udca1 <strong>Agentic \u641c\u7d22\u7684\u6210\u672c\u4ec5\u7565\u9ad8\u4e8e\u6807\u51c6 RAG<\/strong>\uff08\u5927\u90e8\u5206\u5de5\u5177\u8c03\u7528\u5e76\u884c\u6267\u884c\uff09\u3002<\/p>\n<\/blockquote>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Agentic Search vs RAG \u6027\u80fd\u5bf9\u6bd4<\/strong>\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>\u96be\u5ea6<\/th><th>\u7c7b\u522b<\/th><th>#<\/th><th>Agent%<\/th><th>RAG%<\/th><th>\u5e73\u5c40%<\/th><\/tr><\/thead><tbody><tr><td>Easy<\/td><td>\u5ba2\u89c2\u95ee\u7b54<\/td><td>196<\/td><td><strong>56.1%<\/strong><\/td><td>21.9%<\/td><td>21.9%<\/td><\/tr><tr><td>Easy<\/td><td>\u4e3b\u89c2\u95ee\u7b54<\/td><td>321<\/td><td><strong>61.7%<\/strong><\/td><td>17.4%<\/td><td>20.9%<\/td><\/tr><tr><td>Hard<\/td><td>\u5ba2\u89c2\u95ee\u7b54<\/td><td>168<\/td><td><strong>60.7%<\/strong><\/td><td>19.6%<\/td><td>19.6%<\/td><\/tr><tr><td>Hard<\/td><td>\u4e3b\u89c2\u95ee\u7b54<\/td><td>184<\/td><td><strong>68.5%<\/strong><\/td><td>14.7%<\/td><td>16.8%<\/td><\/tr><tr><td><strong>\u603b\u8ba1<\/strong><\/td><td><\/td><td>869<\/td><td><strong>61.7%<\/strong> \ud83c\udfc6<\/td><td>18.3%<\/td><td>20.0%<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">\u2b50 Agentic \u641c\u7d22\u5728\u590d\u6742\u4efb\u52a1\u4e0a\u5c24\u5176\u660e\u663e\u5730\u4f18\u4e8e RAG\u3002<\/p>\n<\/blockquote>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"543-\u767d\u9886\u4efb\u52a1\">5.4.3 \u767d\u9886\u4efb\u52a1<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e3a\u4e25\u683c\u8bc4\u4f30\u6a21\u578b\u5728\u590d\u6742\u4f01\u4e1a\u751f\u4ea7\u529b\u573a\u666f\u4e2d\u7684\u5b9e\u7528\u6027\uff0cV4 \u6784\u5efa\u4e86 <strong>30 \u4e2a\u9ad8\u7ea7\u4e2d\u6587\u4e13\u4e1a\u4efb\u52a1<\/strong>\uff0c\u6db5\u76d6\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>13 \u4e2a\u5173\u952e\u884c\u4e1a<\/strong>\uff08\u91d1\u878d\u3001\u6559\u80b2\u3001\u6cd5\u5f8b\u3001\u79d1\u6280\u7b49\uff09<\/li>\n\n\n\n<li>\u9ad8\u5c42\u6b21\u8ba4\u77e5\u9700\u6c42\uff1a\u6df1\u5165\u4fe1\u606f\u5206\u6790\u3001\u5168\u9762\u6587\u6863\u751f\u6210\u3001\u7ec6\u81f4\u6587\u6863\u7f16\u8f91<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f7f\u7528\u914d\u5907\u57fa\u672c\u5de5\u5177\uff08Bash + Web \u641c\u7d22\uff09\u7684\u5185\u90e8 agent harness \u8fdb\u884c\u8bc4\u4f30\u3002\u7531\u4e8e\u4efb\u52a1\u5f00\u653e\u6027\uff0c\u91c7\u7528<strong>\u4eba\u5de5\u76f2\u8bc4<\/strong>\u5bf9\u6bd4 V4-Pro-Max \u548c Opus-4.6-Max\uff0c\u56db\u4e2a\u7ef4\u5ea6\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u4efb\u52a1\u5b8c\u6210\u5ea6<\/strong>\uff1a\u662f\u5426\u6210\u529f\u89e3\u51b3\u6838\u5fc3\u95ee\u9898\u3002<\/li>\n\n\n\n<li><strong>\u6307\u4ee4\u9075\u5faa<\/strong>\uff1a\u9075\u5b88\u7279\u5b9a\u7ea6\u675f\u548c\u6307\u4ee4\u7684\u7a0b\u5ea6\u3002<\/li>\n\n\n\n<li><strong>\u5185\u5bb9\u8d28\u91cf<\/strong>\uff1a\u4e8b\u5b9e\u51c6\u786e\u6027\u3001\u903b\u8f91\u8fde\u8d2f\u6027\u3001\u4e13\u4e1a\u57fa\u8c03\u3002<\/li>\n\n\n\n<li><strong>\u683c\u5f0f\u7f8e\u611f<\/strong>\uff1a\u5e03\u5c40\u53ef\u8bfb\u6027\u548c\u89c6\u89c9\u5448\u73b0\u3002<\/li>\n<\/ol>\n\n\n\n<h5 class=\"wp-block-heading\" id=\"\u5206\u7c7b\u80dc\u7387\u5bf9\u6bd4\uff08figure-11\uff09\">\u5206\u7c7b\u80dc\u7387\u5bf9\u6bd4\uff08Figure 11\uff09<\/h5>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>\u7c7b\u522b<\/th><th>V4-Pro-Max \u80dc\u7387<\/th><th>\u5e73\u5c40<\/th><th>Opus-4.6-Max \u80dc\u7387<\/th><\/tr><\/thead><tbody><tr><td>\u5206\u6790<\/td><td><strong>55.0%<\/strong><\/td><td>8.0%<\/td><td>37.0%<\/td><\/tr><tr><td>\u751f\u6210<\/td><td><strong>52.0%<\/strong><\/td><td>10.0%<\/td><td>38.0%<\/td><\/tr><tr><td>\u7f16\u8f91<\/td><td><strong>47.0%<\/strong><\/td><td>18.0%<\/td><td>35.0%<\/td><\/tr><tr><td><strong>\u603b\u4f53<\/strong><\/td><td><strong>53.0%<\/strong> \ud83c\udfc6<\/td><td>10.0%<\/td><td>37.0%<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">\ud83c\udfaf **V4-Pro-Max \u6574\u4f53\u975e\u8f93\u7387\u8fbe 63%**\uff0853% \u80dc + 10% \u5e73\u5c40\uff09\uff0c\u5728\u5206\u6790\u3001\u751f\u6210\u3001\u7f16\u8f91\u4efb\u52a1\u4e0a\u5747\u6709\u4e00\u81f4\u4f18\u52bf\u3002<\/p>\n<\/blockquote>\n\n\n\n<h5 class=\"wp-block-heading\" id=\"\u8be6\u7ec6\u7ef4\u5ea6\u5f97\u5206\uff08figure-12\uff09\">\u8be6\u7ec6\u7ef4\u5ea6\u5f97\u5206\uff08Figure 12\uff09<\/h5>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>\u7ef4\u5ea6<\/th><th>V4-Pro-Max<\/th><th>Opus-4.6-Max<\/th><\/tr><\/thead><tbody><tr><td>\u4efb\u52a1\u5b8c\u6210\u5ea6<\/td><td><strong>98.32<\/strong> \ud83c\udfc6<\/td><td>96.68<\/td><\/tr><tr><td>\u6307\u4ee4\u9075\u5faa<\/td><td>87.76<\/td><td><strong>88.88<\/strong> \ud83e\udd47<\/td><\/tr><tr><td>\u5185\u5bb9\u8d28\u91cf<\/td><td><strong>83.32<\/strong> \ud83c\udfc6<\/td><td>78.00<\/td><\/tr><tr><td>\u683c\u5f0f\u7f8e\u611f<\/td><td><strong>76.68<\/strong> \ud83c\udfc6<\/td><td>72.68<\/td><\/tr><tr><td><strong>\u603b\u4f53<\/strong><\/td><td><strong>86.52<\/strong> \ud83c\udfc6<\/td><td>84.06<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">\ud83d\udd0d <strong>V4-Pro-Max \u7684\u4e3b\u8981\u4f18\u52bf<\/strong>\u5728\u4efb\u52a1\u5b8c\u6210\u5ea6\u548c\u5185\u5bb9\u8d28\u91cf\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u4e3b\u52a8\u9884\u671f\u9690\u542b\u7528\u6237\u610f\u56fe<\/strong>\uff0c\u9891\u7e41\u63d0\u4f9b\u8865\u5145\u89c1\u89e3\u548c\u81ea\u9a8c\u8bc1\u6b65\u9aa4\u3002<\/li>\n\n\n\n<li>\u64c5\u957f\u957f\u7bc7\u751f\u6210\uff0c\u4ea4\u4ed8\u6709\u6df1\u5ea6\u3001\u8fde\u8d2f\u7684\u53d9\u8ff0\uff0c\u800c\u975e Opus-4.6-Max \u5e38\u4ea7\u51fa\u7684\u8fc7\u4e8e\u7b80\u5316\u7684\u9879\u76ee\u5217\u8868\u3002<\/li>\n\n\n\n<li>\u4e25\u683c\u9075\u5faa\u6b63\u5f0f\u4e13\u4e1a\u89c4\u8303\uff0c\u5982\u6807\u51c6\u5316\u4e2d\u6587\u5c42\u7ea7\u7f16\u53f7\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u4e0d\u8db3<\/strong>\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u6307\u4ee4\u9075\u5faa\u65b9\u9762\u5076\u5c14\u5ffd\u7565\u7279\u5b9a\u683c\u5f0f\u7ea6\u675f\uff0c\u7565\u900a Opus\u3002<\/li>\n\n\n\n<li>\u5bf9\u5197\u957f\u6587\u672c\u8f93\u5165\u7684\u7b80\u6d01\u603b\u7ed3\u80fd\u529b\u8f83\u5f31\u3002<\/li>\n\n\n\n<li>\u683c\u5f0f\u7f8e\u611f\u65b9\u9762\uff0c\u6f14\u793a\u6587\u7a3f\u7684\u6574\u4f53\u89c6\u89c9\u8bbe\u8ba1\u4ecd\u6709\u660e\u663e\u6539\u8fdb\u7a7a\u95f4\u3002<\/li>\n<\/ul>\n<\/blockquote>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"544-\u4ee3\u7801-agent\">5.4.4 \u4ee3\u7801 Agent<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e3a\u57fa\u51c6\u6d4b\u8bd5\u7f16\u7801 agent \u80fd\u529b\uff0cV4 \u4ece<strong>\u771f\u5b9e\u5185\u90e8 R&amp;D \u5de5\u4f5c\u8d1f\u8f7d<\/strong>\u4e2d\u7b56\u5212\u4efb\u52a1\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u4ece 50+ \u5185\u90e8\u5de5\u7a0b\u5e08\u6536\u96c6 <strong>\u7ea6 200 \u4e2a\u6311\u6218\u6027\u4efb\u52a1<\/strong><\/li>\n\n\n\n<li>\u6db5\u76d6 feature \u5f00\u53d1\u3001bug \u4fee\u590d\u3001\u91cd\u6784\u3001\u8bca\u65ad<\/li>\n\n\n\n<li>\u591a\u6837\u6280\u672f\u6808\uff1aPyTorch\u3001CUDA\u3001Rust\u3001C++<\/li>\n\n\n\n<li>\u6bcf\u4e2a\u4efb\u52a1\u9644\u5e26\u539f\u59cb\u4ed3\u5e93\u3001\u5bf9\u5e94\u6267\u884c\u73af\u5883\u3001\u4eba\u5de5\u6807\u6ce8\u8bc4\u5206 rubrics<\/li>\n\n\n\n<li>\u7ecf\u8fc7\u4e25\u683c\u8d28\u91cf\u8fc7\u6ee4\u540e\u4fdd\u7559 <strong>30 \u4e2a\u4efb\u52a1<\/strong>\u4f5c\u4e3a\u8bc4\u4f30\u96c6<\/li>\n<\/ul>\n\n\n\n<h5 class=\"wp-block-heading\" id=\"table-8\uff1ard-\u7f16\u7a0b\u57fa\u51c6\u5bf9\u6bd4\">Table 8\uff1aR&amp;D \u7f16\u7a0b\u57fa\u51c6\u5bf9\u6bd4<\/h5>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>\u6a21\u578b<\/th><th>Pass Rate (%)<\/th><\/tr><\/thead><tbody><tr><td>Haiku 4.5<\/td><td>13<\/td><\/tr><tr><td>Sonnet 4.5<\/td><td>47<\/td><\/tr><tr><td><strong>DeepSeek-V4-Pro-Max<\/strong><\/td><td><strong>67<\/strong><\/td><\/tr><tr><td>Opus 4.5<\/td><td>70<\/td><\/tr><tr><td>Opus 4.5 Thinking<\/td><td>73<\/td><\/tr><tr><td>Opus 4.6 Thinking<\/td><td><strong>80<\/strong> \ud83e\udd47<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">\ud83c\udfaf <strong>V4-Pro \u663e\u8457\u8d85\u8d8a Claude Sonnet 4.5\uff0c\u63a5\u8fd1 Claude Opus 4.5 \u6c34\u5e73<\/strong>\u3002<\/p>\n<\/blockquote>\n\n\n\n<h5 class=\"wp-block-heading\" id=\"\u5f00\u53d1\u8005\u8c03\u7814\uff08n85\uff09\">\u5f00\u53d1\u8005\u8c03\u7814\uff08N=85\uff09<\/h5>\n\n\n\n<p class=\"wp-block-paragraph\">\u8be2\u95ee DeepSeek \u5f00\u53d1\u8005\u548c\u7814\u7a76\u5458\uff08\u5747\u6709\u65e5\u5e38 agentic \u7f16\u7801\u4f7f\u7528 V4-Pro \u7ecf\u9a8c\uff09V4-Pro \u662f\u5426\u5df2\u51c6\u5907\u597d\u4f5c\u4e3a\u4e0e\u5176\u4ed6\u524d\u6cbf\u6a21\u578b\u76f8\u6bd4\u7684\u9ed8\u8ba4\u4e3b\u7f16\u7a0b\u6a21\u578b\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>52% \u80af\u5b9a<\/strong><\/li>\n\n\n\n<li><strong>39% \u503e\u5411\u80af\u5b9a<\/strong><\/li>\n\n\n\n<li><strong>&lt; 9% \u5426\u5b9a<\/strong><\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u53d7\u8bbf\u8005\u53cd\u9988\uff1aV4-Pro \u5728\u5927\u591a\u6570\u4efb\u52a1\u4e0a\u4ea4\u4ed8\u6ee1\u610f\u7ed3\u679c\uff0c\u4f46\u6ce8\u610f\u5230<strong>\u4e00\u4e9b\u7ec6\u8282\u9519\u8bef<\/strong>\u3001<strong>\u5bf9\u6a21\u7cca\u63d0\u793a\u7684\u8bef\u89e3<\/strong>\u3001<strong>\u5076\u5c14\u8fc7\u5ea6\u601d\u8003<\/strong>\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"6-\u7ed3\u8bba\u3001\u5c40\u9650\u6027\u4e0e\u672a\u6765\u65b9\u5411\">6. \u7ed3\u8bba\u3001\u5c40\u9650\u6027\u4e0e\u672a\u6765\u65b9\u5411<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u672c\u6587\u63d0\u51fa <strong>DeepSeek-V4<\/strong> \u7cfb\u5217\u7684\u9884\u89c8\u7248\u672c\uff0c\u65e8\u5728\u6253\u9020\u65b0\u4e00\u4ee3\u7a81\u7834\u8d85\u957f\u4e0a\u4e0b\u6587\u5904\u7406\u6548\u7387\u74f6\u9888\u7684\u5927\u8bed\u8a00\u6a21\u578b\u3002\u901a\u8fc7\u7ed3\u5408 <strong>CSA + HCA \u7684\u6df7\u5408\u6ce8\u610f\u529b\u67b6\u6784<\/strong>\uff0cV4 \u5728\u957f\u5e8f\u5217\u6548\u7387\u4e0a\u5b9e\u73b0\u98de\u8dc3\u3002\u67b6\u6784\u521b\u65b0\u4e0e\u5e7f\u6cdb\u7684\u57fa\u7840\u8bbe\u65bd\u4f18\u5316\u5171\u540c\u652f\u6301<strong>\u767e\u4e07 token \u4e0a\u4e0b\u6587\u7684\u539f\u751f\u9ad8\u6548\u652f\u6301<\/strong>\uff0c\u4e3a\u672a\u6765\u7684 test-time scaling\u3001\u957f\u7a0b\u4efb\u52a1\u4ee5\u53ca\u5728\u7ebf\u5b66\u4e60\u7b49\u65b0\u5174\u8303\u5f0f\u5960\u5b9a\u4e86\u5fc5\u8981\u57fa\u7840\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"\u6838\u5fc3\u6210\u5c31\u603b\u7ed3\">\u6838\u5fc3\u6210\u5c31\u603b\u7ed3<\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\ud83c\udfc6 <strong>V4-Pro-Max \u91cd\u65b0\u5b9a\u4e49\u5f00\u6e90\u6a21\u578b SOTA<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li>\u5728\u77e5\u8bc6\u57fa\u51c6\u4e0a<strong>\u663e\u8457\u8d85\u8d8a\u5148\u524d\u5f00\u6e90\u6a21\u578b<\/strong><\/li>\n\n\n\n<li>\u63a5\u8fd1\u524d\u6cbf\u95ed\u6e90\u6a21\u578b\u7684\u63a8\u7406\u6027\u80fd<\/li>\n\n\n\n<li>\u4ea4\u4ed8\u5177\u6709\u7ade\u4e89\u529b\u7684 agent \u80fd\u529b<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>\ud83c\udfc6 <strong>V4-Flash-Max \u5b9e\u73b0\u4e0e\u9886\u5148\u95ed\u6e90\u6a21\u578b\u76f8\u5f53\u7684\u63a8\u7406\u6027\u80fd<\/strong>\uff0c\u540c\u65f6\u4fdd\u6301\u9ad8\u6027\u4ef7\u6bd4\u67b6\u6784<\/li>\n\n\n\n<li>\ud83c\udfc6 \u4e3a<strong>\u5f00\u6e90\u6a21\u578b<\/strong>\u5f00\u542f<strong>\u767e\u4e07\u7ea7\u4e0a\u4e0b\u6587<\/strong>\u7684\u65b0\u65f6\u4ee3<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"\u5c40\u9650\u6027\u4e0e\u672a\u6765\u5de5\u4f5c\u65b9\u5411\">\u5c40\u9650\u6027\u4e0e\u672a\u6765\u5de5\u4f5c\u65b9\u5411<\/h3>\n\n\n\n<ol class=\"wp-block-list\">\n<li><p><strong>\u67b6\u6784\u590d\u6742\u6027<\/strong>\uff1a\u4e3a\u8ffd\u6c42\u6781\u81f4\u957f\u4e0a\u4e0b\u6587\u6548\u7387\uff0cV4 \u91c7\u53d6\u4e86\u5927\u80c6\u67b6\u6784\u8bbe\u8ba1\u3002\u4e3a\u6700\u5c0f\u5316\u98ce\u9669\uff0c\u4fdd\u7559\u4e86\u8bb8\u591a\u521d\u6b65\u9a8c\u8bc1\u8fc7\u7684\u7ec4\u4ef6\u548c\u6280\u5de7\uff0c\u8fd9\u867d\u7136\u6709\u6548\u4f46\u4e5f\u4f7f\u67b6\u6784\u76f8\u5bf9\u590d\u6742\u3002\u672a\u6765\u5c06\u8fdb\u884c\u66f4\u5168\u9762\u3001\u6709\u539f\u5219\u7684\u7814\u7a76\uff0c<strong>\u5c06\u67b6\u6784\u63d0\u70bc\u5230\u6700\u672c\u8d28\u7684\u8bbe\u8ba1<\/strong>\uff0c\u4f7f\u5176\u66f4\u4f18\u96c5\u800c\u4e0d\u635f\u5931\u6027\u80fd\u3002<\/p><br><\/li>\n\n\n\n<li><p><strong>\u8bad\u7ec3\u7a33\u5b9a\u6027\u7684\u7406\u8bba\u7406\u89e3<\/strong>\uff1a\u5c3d\u7ba1 <strong>Anticipatory Routing<\/strong> \u548c <strong>SwiGLU Clamping<\/strong> \u5df2\u88ab\u8bc1\u660e\u6709\u6548\u7f13\u89e3\u8bad\u7ec3\u4e0d\u7a33\u5b9a\uff0c\u4f46\u5176<strong>\u5e95\u5c42\u539f\u7406\u4ecd\u7406\u89e3\u4e0d\u8db3<\/strong>\u3002\u672a\u6765\u5c06\u79ef\u6781\u7814\u7a76\u8bad\u7ec3\u7a33\u5b9a\u6027\u7684\u57fa\u7840\u95ee\u9898\u5e76\u52a0\u5f3a\u5185\u90e8\u6307\u6807\u76d1\u63a7\uff0c\u8ffd\u6c42\u66f4\u6709\u539f\u5219\u3001\u66f4\u5177\u9884\u6d4b\u6027\u7684\u5927\u89c4\u6a21\u7a33\u5b9a\u8bad\u7ec3\u65b9\u6cd5\u3002<\/p><br><\/li>\n\n\n\n<li><p><strong>\u7a00\u758f\u6027\u7684\u65b0\u7ef4\u5ea6<\/strong>\uff1a\u9664 MoE \u548c\u7a00\u758f\u6ce8\u610f\u529b\u5916\uff0cV4 \u8fd8\u5c06<strong>\u4e3b\u52a8\u63a2\u7d22\u6a21\u578b\u7a00\u758f\u6027\u7684\u65b0\u7ef4\u5ea6<\/strong>\u2014\u2014\u5982\u66f4\u7a00\u758f\u7684 embedding \u6a21\u5757\uff08Cheng et al., 2026\uff09\u2014\u2014\u8fdb\u4e00\u6b65\u63d0\u5347\u8ba1\u7b97\u548c\u5185\u5b58\u6548\u7387\u800c\u4e0d\u635f\u5bb3\u80fd\u529b\u3002<\/p><br><\/li>\n\n\n\n<li><p><strong>\u4f4e\u5ef6\u8fdf\u67b6\u6784<\/strong>\uff1a\u5c06\u6301\u7eed\u7814\u7a76\u4f4e\u5ef6\u8fdf\u67b6\u6784\u548c\u7cfb\u7edf\u6280\u672f\uff0c\u4f7f\u957f\u4e0a\u4e0b\u6587\u90e8\u7f72\u548c\u4ea4\u4e92\u66f4\u5177\u54cd\u5e94\u6027\u3002<\/p><br><\/li>\n\n\n\n<li><p><strong>\u957f\u7a0b\u3001\u591a\u8f6e agent \u4efb\u52a1<\/strong>\uff1a\u8ba4\u8bc6\u5230\u957f\u7a0b\u591a\u8f6e agent \u4efb\u52a1\u7684\u91cd\u8981\u6027\u548c\u5b9e\u9645\u4ef7\u503c\uff0c\u5c06\u7ee7\u7eed\u5728\u8be5\u65b9\u5411\u8fed\u4ee3\u63a2\u7d22\u3002<\/p><br><\/li>\n\n\n\n<li><p><strong>\u591a\u6a21\u6001\u80fd\u529b<\/strong>\uff1a\u6b63\u5728\u4e3a\u6a21\u578b\u52a0\u5165\u591a\u6a21\u6001\u80fd\u529b\u3002<\/p><br><\/li>\n\n\n\n<li><p><strong>\u6570\u636e\u7b56\u7565<\/strong>\uff1a\u81f4\u529b\u4e8e\u5f00\u53d1\u66f4\u597d\u7684\u6570\u636e\u7b56\u5212\u548c\u5408\u6210\u7b56\u7565\uff0c\u5728\u65e5\u76ca\u5e7f\u6cdb\u7684\u573a\u666f\u548c\u4efb\u52a1\u4e2d\u6301\u7eed\u589e\u5f3a\u6a21\u578b\u667a\u80fd\u3001\u7a33\u5065\u6027\u548c\u5b9e\u7528\u6027\u3002<\/p><br><\/li>\n<\/ol>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"\ud83d\udcce-\u9644\u5f55\">\ud83d\udcce \u9644\u5f55<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"\u4f5c\u8005\u5217\u8868\">\u4f5c\u8005\u5217\u8868<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f5c\u8005\u6309\u540d\u5b57\u5b57\u6bcd\u987a\u5e8f\u6392\u5217\u3002\u5e26 * \u6807\u8bb0\u7684\u540d\u5b57\u8868\u793a\u5df2\u4ece\u56e2\u961f\u79bb\u804c\u7684\u6210\u5458\u3002\uff08\u5b8c\u6574\u5217\u8868\u89c1\u539f\u6587\uff09<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u7814\u7a76\u4e0e\u5de5\u7a0b<\/strong>\uff1aAnyi Xu\u3001Bangcai Lin\u3001Bing Xue\u3001Bingxuan Wang<em>\u3001Bingzheng Xu\u3001Bochao Wu\u3001Bowei Zhang\u3001Chaofan Lin\u3001Chen Dong\u3001Chengda Lu\u3001Chenggang Zhao\u3001Chengqi Deng\u3001Chenhao Xu\u3001Chenze Shao\u3001Chong Ruan<\/em> \u7b49\u8fd1 300 \u4eba\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u5546\u4e1a\u4e0e\u5408\u89c4<\/strong>\uff1aChenchen Ling\u3001Chengyu Hou\u3001Dongjie Ji\u3001Fang Wei \u7b49\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"\u81f4\u8c22\">\u81f4\u8c22<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u611f\u8c22 Dolly Deng \u53ca\u5176\u4ed6\u6d4b\u8bd5\u4eba\u5458\u5bf9 DeepSeek-V4 \u7cfb\u5217\u6a21\u578b\u80fd\u529b\u63d0\u4f9b\u7684\u5b9d\u8d35\u5efa\u8bae\u548c\u53cd\u9988\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"\u53c2\u8003\u6587\u732e\">\u53c2\u8003\u6587\u732e<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5b8c\u6574\u53c2\u8003\u6587\u732e\u5217\u8868\u8bf7\u53c2\u9605\u539f\u59cb\u82f1\u6587\u8bba\u6587\u7b2c 45-53 \u9875\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"\ud83d\uddc2\ufe0f-\u7ffb\u8bd1\u8bf4\u660e\">\ud83d\uddc2\ufe0f \u7ffb\u8bd1\u8bf4\u660e<\/h2>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u7ffb\u8bd1\u65e5\u671f<\/strong>\uff1a2026 \u5e74 4 \u6708<\/li>\n\n\n\n<li><strong>\u7ffb\u8bd1\u7248\u672c<\/strong>\uff1a\u57fa\u4e8e DeepSeek-V4 \u9884\u89c8\u7248\u6280\u672f\u62a5\u544a\u82f1\u6587\u539f\u6587<\/li>\n\n\n\n<li><strong>\u7ffb\u8bd1\u98ce\u683c<\/strong>\uff1a\u6280\u672f\u672f\u8bed\u4fdd\u7559\u82f1\u6587\u5f62\u5f0f\uff08\u5982 attention\u3001MoE\u3001KV cache \u7b49\uff09\uff0c\u4e0a\u4e0b\u6587\u5173\u952e\u8bcd\u6ce8\u660e\u4e2d\u82f1\u5bf9\u7167<\/li>\n\n\n\n<li><strong>\u91cd\u70b9\u7a81\u51fa<\/strong>\uff1a\u4e0e\u5176\u4ed6\u6a21\u578b\u7684\u5bf9\u6bd4\u3001\u76f8\u5bf9\u4e0a\u4ee3\u6a21\u578b\u7684\u63d0\u5347\u3001\u5404\u9879\u57fa\u51c6\u5206\u6570\u5dee\u5f02\u5747\u4ee5\u8868\u683c\u6216\u5f15\u7528\u5757\u65b9\u5f0f\u6807\u6ce8\uff0c\u4fbf\u4e8e\u5feb\u901f\u67e5\u9605<\/li>\n\n\n\n<li><strong>\u4fdd\u7559\u539f\u6587\u4fe1\u606f<\/strong>\uff1a\u6240\u6709\u6570\u5b57\u3001\u8868\u683c\u3001\u516c\u5f0f\u3001\u6a21\u578b\u914d\u7f6e\u5747\u5b8c\u6574\u4fdd\u7559<\/li>\n\n\n\n<li><strong>\u516c\u5f0f<\/strong>\uff1a\u4fdd\u7559 LaTeX \u683c\u5f0f\u4ee5\u4fbf\u6e32\u67d3<\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>DeepSeek-V4\uff1a\u8fc8\u5411\u9ad8\u6548\u7684\u767e\u4e07 Token \u957f\u4e0a\u4e0b\u6587\u667a\u80fd \u4f5c\u8005\uff1aDeepSeek-AI \u8054\u7cfb\u65b9\u5f0f\uff1aresearch@dee &#8230;<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_import_markdown_pro_load_document_selector":0,"_import_markdown_pro_submit_text_textarea":"","emotion":"","emotion_color":"","title_style":"","license":"","footnotes":""},"categories":[1],"tags":[],"class_list":["post-44","post","type-post","status-publish","format-standard","hentry","category-uncategorized"],"_links":{"self":[{"href":"https:\/\/blog.uika.cn\/index.php?rest_route=\/wp\/v2\/posts\/44","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/blog.uika.cn\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/blog.uika.cn\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/blog.uika.cn\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/blog.uika.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=44"}],"version-history":[{"count":1,"href":"https:\/\/blog.uika.cn\/index.php?rest_route=\/wp\/v2\/posts\/44\/revisions"}],"predecessor-version":[{"id":45,"href":"https:\/\/blog.uika.cn\/index.php?rest_route=\/wp\/v2\/posts\/44\/revisions\/45"}],"wp:attachment":[{"href":"https:\/\/blog.uika.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=44"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/blog.uika.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=44"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/blog.uika.cn\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=44"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}