{"id":2212,"date":"2025-01-27T07:09:29","date_gmt":"2025-01-27T07:09:29","guid":{"rendered":"https:\/\/oimi.me\/?p=2212"},"modified":"2025-01-27T07:09:29","modified_gmt":"2025-01-27T07:09:29","slug":"deepseek-r1-%e5%8f%91%e5%b8%83%ef%bc%8c%e6%80%a7%e8%83%bd%e5%af%b9%e6%a0%87-openai-o1-%e6%ad%a3%e5%bc%8f%e7%89%88","status":"publish","type":"post","link":"https:\/\/myya.net\/index.php\/2025\/01\/27\/deepseek-r1-%e5%8f%91%e5%b8%83%ef%bc%8c%e6%80%a7%e8%83%bd%e5%af%b9%e6%a0%87-openai-o1-%e6%ad%a3%e5%bc%8f%e7%89%88\/","title":{"rendered":"DeepSeek-R1 \u53d1\u5e03\uff0c\u6027\u80fd\u5bf9\u6807 OpenAI o1 \u6b63\u5f0f\u7248"},"content":{"rendered":"<p dir=\"auto\">DeepSeek-V3\uff0c\u8fd9\u662f\u4e00\u4e2a\u5f3a\u5927\u7684\u4e13\u5bb6\u6df7\u5408 \uff08MoE\uff09 \u8bed\u8a00\u6a21\u578b\uff0c\u603b\u5171\u6709 671B \u4e2a\u53c2\u6570\uff0c\u6bcf\u4e2a\u4ee4\u724c\u6fc0\u6d3b\u4e86 37B\u3002\u4e3a\u4e86\u5b9e\u73b0\u9ad8\u6548\u7684\u63a8\u7406\u548c\u5177\u6709\u6210\u672c\u6548\u76ca\u7684\u8bad\u7ec3\uff0cDeepSeek-V3 \u91c7\u7528\u4e86\u591a\u5934\u6f5c\u5728\u6ce8\u610f\u529b \uff08MLA\uff09 \u548c DeepSeekMoE \u67b6\u6784\uff0c\u8fd9\u4e9b\u67b6\u6784\u5728 DeepSeek-V2 \u4e2d\u5f97\u5230\u4e86\u5168\u9762\u9a8c\u8bc1\u3002\u6b64\u5916\uff0cDeepSeek-V3 \u5f00\u521b\u4e86\u4e00\u79cd\u7528\u4e8e\u8d1f\u8f7d\u5e73\u8861\u7684\u8f85\u52a9\u65e0\u635f\u5931\u7b56\u7565\uff0c\u5e76\u8bbe\u5b9a\u4e86\u591a\u4ee4\u724c\u9884\u6d4b\u8bad\u7ec3\u76ee\u6807\u4ee5\u83b7\u5f97\u66f4\u5f3a\u7684\u6027\u80fd\u3002\u6211\u4eec\u5728 14.8 \u4e07\u4ebf\u4e2a\u591a\u6837\u5316\u548c\u9ad8\u8d28\u91cf\u4ee4\u724c\u4e0a\u5bf9 DeepSeek-V3 \u8fdb\u884c\u4e86\u9884\u8bad\u7ec3\uff0c\u7136\u540e\u662f\u76d1\u7763\u5fae\u8c03\u548c\u5f3a\u5316\u5b66\u4e60\u9636\u6bb5\uff0c\u4ee5\u5145\u5206\u5229\u7528\u5176\u529f\u80fd\u3002\u7efc\u5408\u8bc4\u4f30\u8868\u660e\uff0cDeepSeek-V3 \u7684\u6027\u80fd\u4f18\u4e8e\u5176\u4ed6\u5f00\u6e90\u6a21\u578b\uff0c\u5e76\u5b9e\u73b0\u4e86\u4e0e\u9886\u5148\u7684\u95ed\u6e90\u6a21\u578b\u76f8\u5f53\u7684\u6027\u80fd\u3002\u5c3d\u7ba1\u6027\u80fd\u51fa\u8272\uff0c\u4f46 DeepSeek-V3 \u7684\u5b8c\u6574\u8bad\u7ec3\u53ea\u9700\u8981 2.788M H800 GPU \u5c0f\u65f6\u3002\u6b64\u5916\uff0c\u5b83\u7684\u8bad\u7ec3\u8fc7\u7a0b\u975e\u5e38\u7a33\u5b9a\u3002\u5728\u6574\u4e2a\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0c\u6211\u4eec\u6ca1\u6709\u9047\u5230\u4efb\u4f55\u4e0d\u53ef\u6062\u590d\u7684\u635f\u5931\u5cf0\u503c\u6216\u6267\u884c\u4efb\u4f55\u56de\u6eda\u3002<!--more--><\/p>\n<p dir=\"auto\" align=\"center\"><a href=\"https:\/\/github.com\/deepseek-ai\/DeepSeek-V3\/blob\/main\/figures\/benchmark.png\" target=\"_blank\" rel=\"noopener noreferrer\"><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter size-full wp-image-2214\" title=\"benchmark\" src=\"https:\/\/oimi.me\/wp-content\/uploads\/2025\/01\/benchmark.png\" alt=\"benchmark\" width=\"1702\" height=\"994\" \/><\/a><\/p>\n<div class=\"markdown-heading\" dir=\"auto\">\n<h2 class=\"heading-element\" dir=\"auto\" tabindex=\"-1\">2. \u6a21\u578b\u6458\u8981<\/h2>\n<p><a id=\"user-content-2-model-summary\" class=\"anchor\" href=\"https:\/\/github.com\/deepseek-ai\/DeepSeek-V3#2-model-summary\" aria-label=\"\u6c38\u4e45\u94fe\u63a5\uff1a2. \u6a21\u578b\u603b\u7ed3\"><\/a><\/div>\n<hr \/>\n<p dir=\"auto\"><strong>\u67b6\u6784\uff1a\u521b\u65b0\u7684\u8d1f\u8f7d\u5747\u8861\u7b56\u7565\u548c\u8bad\u7ec3\u76ee\u6807<\/strong><\/p>\n<ul dir=\"auto\">\n<li>\u9664\u4e86 DeepSeek-V2 \u7684\u9ad8\u6548\u67b6\u6784\u4e4b\u5916\uff0c\u6211\u4eec\u8fd8\u5f00\u521b\u4e86\u4e00\u79cd\u7528\u4e8e\u8d1f\u8f7d\u5747\u8861\u7684\u8f85\u52a9\u65e0\u635f\u7b56\u7565\uff0c\u8be5\u7b56\u7565\u53ef\u4ee5\u6700\u5927\u9650\u5ea6\u5730\u51cf\u5c11\u56e0\u9f13\u52b1\u8d1f\u8f7d\u5747\u8861\u800c\u5f15\u8d77\u7684\u6027\u80fd\u4e0b\u964d\u3002<\/li>\n<li>\u6211\u4eec\u7814\u7a76\u4e86\u591a\u4ee4\u724c\u9884\u6d4b \uff08MTP\uff09 \u76ee\u6807\uff0c\u5e76\u8bc1\u660e\u5b83\u5bf9\u6a21\u578b\u6027\u80fd\u6709\u76ca\u3002\u5b83\u8fd8\u53ef\u7528\u4e8e\u63a8\u7406\u52a0\u901f\u7684\u63a8\u6d4b\u89e3\u7801\u3002<\/li>\n<\/ul>\n<hr \/>\n<p dir=\"auto\"><strong>\u8bad\u7ec3\u524d\uff1a\u8fc8\u5411\u7ec8\u6781\u8bad\u7ec3\u6548\u7387<\/strong><\/p>\n<ul dir=\"auto\">\n<li>\u6211\u4eec\u8bbe\u8ba1\u4e86\u4e00\u4e2a FP8 \u6df7\u5408\u7cbe\u5ea6\u8bad\u7ec3\u6846\u67b6\uff0c\u5e76\u9996\u6b21\u5728\u8d85\u5927\u89c4\u6a21\u6a21\u578b\u4e0a\u9a8c\u8bc1\u4e86 FP8 \u8bad\u7ec3\u7684\u53ef\u884c\u6027\u548c\u6709\u6548\u6027\u3002<\/li>\n<li>\u901a\u8fc7\u7b97\u6cd5\u3001\u6846\u67b6\u548c\u786c\u4ef6\u7684\u534f\u540c\u8bbe\u8ba1\uff0c\u6211\u4eec\u514b\u670d\u4e86\u8de8\u8282\u70b9 MoE \u8bad\u7ec3\u4e2d\u7684\u901a\u4fe1\u74f6\u9888\uff0c\u51e0\u4e4e\u5b9e\u73b0\u4e86\u5b8c\u5168\u7684\u8ba1\u7b97-\u901a\u4fe1\u91cd\u53e0\u3002<br \/>\n\u8fd9\u663e\u8457\u63d0\u9ad8\u4e86\u6211\u4eec\u7684\u8bad\u7ec3\u6548\u7387\uff0c\u964d\u4f4e\u4e86\u8bad\u7ec3\u6210\u672c\uff0c\u4f7f\u6211\u4eec\u80fd\u591f\u5728\u6ca1\u6709\u989d\u5916\u5f00\u9500\u7684\u60c5\u51b5\u4e0b\u8fdb\u4e00\u6b65\u6269\u5927\u6a21\u578b\u89c4\u6a21\u3002<\/li>\n<li>\u6211\u4eec\u4ee5\u4ec5 2.664M H800 GPU \u5c0f\u65f6\u7684\u7ecf\u6d4e\u6210\u672c\uff0c\u5728 14.8T \u4ee4\u724c\u4e0a\u5b8c\u6210\u4e86 DeepSeek-V3 \u7684\u9884\u8bad\u7ec3\uff0c\u4ea7\u751f\u4e86\u76ee\u524d\u6700\u5f3a\u7684\u5f00\u6e90\u57fa\u7840\u6a21\u578b\u3002\u9884\u8bad\u7ec3\u540e\u7684\u540e\u7eed\u8bad\u7ec3\u9636\u6bb5\u53ea\u9700\u8981 0.1M GPU \u5c0f\u65f6\u3002<\/li>\n<\/ul>\n<hr \/>\n<p dir=\"auto\"><strong>\u57f9\u8bad\u540e\uff1aDeepSeek-R1 \u7684\u77e5\u8bc6\u63d0\u70bc<\/strong><\/p>\n<ul dir=\"auto\">\n<li>\u6211\u4eec\u5f15\u5165\u4e86\u4e00\u79cd\u521b\u65b0\u65b9\u6cd5\uff0c\u5c06\u957f\u94fe\u601d\u7ef4 \uff08CoT\uff09 \u6a21\u578b\u4e2d\u7684\u63a8\u7406\u80fd\u529b\u63d0\u70bc\u51fa\u6765\uff0c\u7279\u522b\u662f\u4ece DeepSeek R1 \u7cfb\u5217\u6a21\u578b\u4e4b\u4e00\u4e2d\u63d0\u70bc\u5230\u6807\u51c6 LLM \u4e2d\uff0c\u7279\u522b\u662f DeepSeek-V3\u3002\u6211\u4eec\u7684\u7ba1\u9053\u5c06 R1 \u7684\u9a8c\u8bc1\u548c\u53cd\u5c04\u6a21\u5f0f\u4f18\u96c5\u5730\u6574\u5408\u5230 DeepSeek-V3 \u4e2d\uff0c\u5e76\u663e\u8457\u63d0\u9ad8\u4e86\u5176\u63a8\u7406\u6027\u80fd\u3002\u540c\u65f6\uff0c\u6211\u4eec\u8fd8\u4fdd\u6301\u4e86\u5bf9 DeepSeek-V3 \u7684\u8f93\u51fa\u6837\u5f0f\u548c\u957f\u5ea6\u7684\u63a7\u5236\u3002<\/li>\n<\/ul>\n<hr \/>\n<div class=\"markdown-heading\" dir=\"auto\">\n<h2 class=\"heading-element\" dir=\"auto\" tabindex=\"-1\">3. \u6a21\u578b\u4e0b\u8f7d<\/h2>\n<p><a id=\"user-content-3-model-downloads\" class=\"anchor\" href=\"https:\/\/github.com\/deepseek-ai\/DeepSeek-V3#3-model-downloads\" aria-label=\"\u6c38\u4e45\u94fe\u63a5\uff1a3. \u6a21\u578b\u4e0b\u8f7d\"><\/a><\/div>\n<div dir=\"auto\" align=\"center\">\n<table>\n<thead>\n<tr>\n<th align=\"center\"><strong>\u578b<\/strong><\/th>\n<th align=\"center\"><strong>#Total Params<\/strong><\/th>\n<th align=\"center\"><strong>#Activated Params<\/strong><\/th>\n<th align=\"center\"><strong>\u4e0a\u4e0b\u6587\u957f\u5ea6<\/strong><\/th>\n<th align=\"center\"><strong>\u4e0b\u8f7d<\/strong><\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td align=\"center\">DeepSeek-V3-\u57fa\u7840<\/td>\n<td align=\"center\">671B \u7cfb\u5217<\/td>\n<td align=\"center\">\u7f16\u53f7 37B<\/td>\n<td align=\"center\">128K<\/td>\n<td align=\"center\"><a href=\"https:\/\/huggingface.co\/deepseek-ai\/DeepSeek-V3-Base\" rel=\"nofollow\">&#x1f917; \u62e5\u62b1\u8138<\/a><\/td>\n<\/tr>\n<tr>\n<td align=\"center\">\u6df1\u5ea6\u641c\u7d22-V3<\/td>\n<td align=\"center\">671B \u7cfb\u5217<\/td>\n<td align=\"center\">\u7f16\u53f7 37B<\/td>\n<td align=\"center\">128K<\/td>\n<td align=\"center\"><a href=\"https:\/\/huggingface.co\/deepseek-ai\/DeepSeek-V3\" rel=\"nofollow\">&#x1f917; \u62e5\u62b1\u8138<\/a><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<div class=\"markdown-alert markdown-alert-note\" dir=\"auto\">\n<p class=\"markdown-alert-title\" dir=\"auto\"><span>\u6ce8\u610f<\/span><\/p>\n<p dir=\"auto\">Hugging Face \u4e0a\u7684 DeepSeek-V3 \u6a21\u578b\u7684\u603b\u5927\u5c0f\u4e3a 685B\uff0c\u5176\u4e2d\u5305\u62ec 671B \u7684\u4e3b\u6a21\u578b\u6743\u91cd\u548c 14B \u7684\u591a\u6807\u8bb0\u9884\u6d4b \uff08MTP\uff09 \u6a21\u5757\u6743\u91cd\u3002<\/p>\n<\/div>\n<p dir=\"auto\">\u4e3a\u4e86\u786e\u4fdd\u6700\u4f73\u6027\u80fd\u548c\u7075\u6d3b\u6027\uff0c\u6211\u4eec\u4e0e\u5f00\u6e90\u793e\u533a\u548c\u786c\u4ef6\u4f9b\u5e94\u5546\u5408\u4f5c\uff0c\u63d0\u4f9b\u4e86\u591a\u79cd\u5728\u672c\u5730\u8fd0\u884c\u6a21\u578b\u7684\u65b9\u6cd5\u3002\u6709\u5173\u5206\u6b65\u6307\u5357\uff0c\u8bf7\u67e5\u770b\u7b2c 6 \u8282\uff1a<a href=\"https:\/\/github.com\/deepseek-ai\/DeepSeek-V3#6-how-to-run-locally\">How_to Run_Locally<\/a>\u3002<\/p>\n<p dir=\"auto\">\u5bf9\u4e8e\u5e0c\u671b\u6df1\u5165\u4e86\u89e3\u7684\u5f00\u53d1\u4eba\u5458\uff0c\u6211\u4eec\u5efa\u8bae\u63a2\u7d22\u00a0<a href=\"https:\/\/github.com\/deepseek-ai\/DeepSeek-V3\/blob\/main\/README_WEIGHTS.md\">README_WEIGHTS.md<\/a>\u00a0\u4ee5\u4e86\u89e3\u6709\u5173\u4e3b\u6a21\u578b\u6743\u91cd\u548c\u591a\u4ee4\u724c\u9884\u6d4b \uff08MTP\uff09 \u6a21\u5757\u7684\u8be6\u7ec6\u4fe1\u606f\u3002\u8bf7\u6ce8\u610f\uff0cMTP \u652f\u6301\u76ee\u524d\u6b63\u5728\u793e\u533a\u5185\u79ef\u6781\u5f00\u53d1\u4e2d\uff0c\u6211\u4eec\u6b22\u8fce\u60a8\u7684\u8d21\u732e\u548c\u53cd\u9988\u3002<\/p>\n<div class=\"markdown-heading\" dir=\"auto\">\n<h2 class=\"heading-element\" dir=\"auto\" tabindex=\"-1\">4. \u8bc4\u4f30\u7ed3\u679c<\/h2>\n<p><a id=\"user-content-4-evaluation-results\" class=\"anchor\" href=\"https:\/\/github.com\/deepseek-ai\/DeepSeek-V3#4-evaluation-results\" aria-label=\"\u6c38\u4e45\u94fe\u63a5\uff1a 4. \u8bc4\u4f30\u7ed3\u679c\"><\/a><\/div>\n<div class=\"markdown-heading\" dir=\"auto\">\n<h3 class=\"heading-element\" dir=\"auto\" tabindex=\"-1\">\u57fa\u672c\u6a21\u578b<\/h3>\n<p><a id=\"user-content-base-model\" class=\"anchor\" href=\"https:\/\/github.com\/deepseek-ai\/DeepSeek-V3#base-model\" aria-label=\"\u6c38\u4e45\u94fe\u63a5\uff1a\u57fa\u672c\u6a21\u578b\"><\/a><\/div>\n<div class=\"markdown-heading\" dir=\"auto\">\n<h4 class=\"heading-element\" dir=\"auto\" tabindex=\"-1\">\u6807\u51c6\u57fa\u51c6<\/h4>\n<p><a id=\"user-content-standard-benchmarks\" class=\"anchor\" href=\"https:\/\/github.com\/deepseek-ai\/DeepSeek-V3#standard-benchmarks\" aria-label=\"\u6c38\u4e45\u94fe\u63a5\uff1a\u6807\u51c6\u57fa\u51c6\"><\/a><\/div>\n<div dir=\"auto\" align=\"center\">\n<table>\n<thead>\n<tr>\n<th><\/th>\n<th>\u57fa\u51c6 \uff08\u5ea6\u91cf\uff09<\/th>\n<th># \u955c\u5934<\/th>\n<th>\u6df1\u5ea6\u641c\u7d22-V2<\/th>\n<th>Qwen2.5 72B<\/th>\n<th>LLaMA3.1 405B<\/th>\n<th>\u6df1\u5ea6\u641c\u7d22-V3<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td><\/td>\n<td>\u5efa\u7b51<\/td>\n<td>&#8211;<\/td>\n<td>\u6559\u80b2\u90e8<\/td>\n<td>\u7a20<\/td>\n<td>\u7a20<\/td>\n<td>\u6559\u80b2\u90e8<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td># \u6fc0\u6d3b\u7684 Params<\/td>\n<td>&#8211;<\/td>\n<td>21B<\/td>\n<td>72B \u7cfb\u5217<\/td>\n<td>405B \u7cfb\u5217<\/td>\n<td>\u7f16\u53f7 37B<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td># \u603b\u53c2\u6570<\/td>\n<td>&#8211;<\/td>\n<td>\u7f16\u53f7 236B<\/td>\n<td>72B \u7cfb\u5217<\/td>\n<td>405B \u7cfb\u5217<\/td>\n<td>671B \u7cfb\u5217<\/td>\n<\/tr>\n<tr>\n<td>\u82f1\u8bed<\/td>\n<td>\u6869\u6d4b\u8bd5 \uff08BPB\uff09<\/td>\n<td>&#8211;<\/td>\n<td>0.606<\/td>\n<td>0.638<\/td>\n<td><strong>0.542<\/strong><\/td>\n<td>0.548<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>BBH \uff08EM\uff09<\/td>\n<td>3 \u955c\u5934<\/td>\n<td>78.8<\/td>\n<td>79.8<\/td>\n<td>82.9<\/td>\n<td><strong>87.5<\/strong><\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>MMLU \uff08Acc.\uff09<\/td>\n<td>5 \u955c\u5934<\/td>\n<td>78.4<\/td>\n<td>85.0<\/td>\n<td>84.4<\/td>\n<td><strong>87.1<\/strong><\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>MMLU-Redux \uff08Acc.\uff09<\/td>\n<td>5 \u955c\u5934<\/td>\n<td>75.6<\/td>\n<td>83.2<\/td>\n<td>81.3<\/td>\n<td><strong>86.2<\/strong><\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>MMLU-Pro \uff08\u9644\u4ef6\uff09<\/td>\n<td>5 \u955c\u5934<\/td>\n<td>51.4<\/td>\n<td>58.3<\/td>\n<td>52.8<\/td>\n<td><strong>64.4<\/strong><\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>DROP \uff08F1\uff09<\/td>\n<td>3 \u955c\u5934<\/td>\n<td>80.4<\/td>\n<td>80.6<\/td>\n<td>86.0<\/td>\n<td><strong>89.0<\/strong><\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>ARC-Easy \uff08\u7d2f\u79ef\uff09<\/td>\n<td>25 \u53d1<\/td>\n<td>97.6<\/td>\n<td>98.4<\/td>\n<td>98.4<\/td>\n<td><strong>98.9<\/strong><\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>ARC-Challenge \uff08\u7d2f\u79ef\uff09<\/td>\n<td>25 \u53d1<\/td>\n<td>92.2<\/td>\n<td>94.5<\/td>\n<td><strong>95.3<\/strong><\/td>\n<td><strong>95.3<\/strong><\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>HellaSwag \uff08\u7d2f\u79ef\uff09<\/td>\n<td>10 \u6b21\u62cd\u6444<\/td>\n<td>87.1<\/td>\n<td>84.8<\/td>\n<td><strong>89.2<\/strong><\/td>\n<td>88.9<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>PIQA \uff08\u7d2f\u79ef\uff09<\/td>\n<td>0 \u6b21\u5c04\u51fb<\/td>\n<td>83.9<\/td>\n<td>82.6<\/td>\n<td><strong>85.9<\/strong><\/td>\n<td>84.7<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>WinoGrande \uff08Acc.\uff09<\/td>\n<td>5 \u955c\u5934<\/td>\n<td><strong>86.3<\/strong><\/td>\n<td>82.3<\/td>\n<td>85.2<\/td>\n<td>84.9<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>RACE-\u4e2d \uff08Acc.\uff09<\/td>\n<td>5 \u955c\u5934<\/td>\n<td>73.1<\/td>\n<td>68.1<\/td>\n<td><strong>74.2<\/strong><\/td>\n<td>67.1<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>RACE-High \uff08\u7d2f\u79ef\uff09<\/td>\n<td>5 \u955c\u5934<\/td>\n<td>52.6<\/td>\n<td>50.3<\/td>\n<td><strong>56.8<\/strong><\/td>\n<td>51.3<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>\u82b1\u7d6eQA \uff08EM\uff09<\/td>\n<td>5 \u955c\u5934<\/td>\n<td>80.0<\/td>\n<td>71.9<\/td>\n<td><strong>82.7<\/strong><\/td>\n<td><strong>82.9<\/strong><\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>\u81ea\u7136\u95ee\u9898 \uff08EM\uff09<\/td>\n<td>5 \u955c\u5934<\/td>\n<td>38.6<\/td>\n<td>33.2<\/td>\n<td><strong>41.5<\/strong><\/td>\n<td>40.0<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>AGIEval \uff08Acc.\uff09<\/td>\n<td>0 \u6b21\u5c04\u51fb<\/td>\n<td>57.5<\/td>\n<td>75.8<\/td>\n<td>60.6<\/td>\n<td><strong>79.6<\/strong><\/td>\n<\/tr>\n<tr>\n<td>\u6cd5\u5178<\/td>\n<td>HumanEval \uff08Pass@1\uff09<\/td>\n<td>0 \u6b21\u5c04\u51fb<\/td>\n<td>43.3<\/td>\n<td>53.0<\/td>\n<td>54.9<\/td>\n<td><strong>65.2<\/strong><\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>MBPP \uff08Pass@1\uff09<\/td>\n<td>3 \u955c\u5934<\/td>\n<td>65.0<\/td>\n<td>72.6<\/td>\n<td>68.4<\/td>\n<td><strong>75.4<\/strong><\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>LiveCodeBench-Base \uff08Pass@1\uff09<\/td>\n<td>3 \u955c\u5934<\/td>\n<td>11.6<\/td>\n<td>12.9<\/td>\n<td>15.5<\/td>\n<td><strong>19.4<\/strong><\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>CRUXEval-I \uff08Acc.\uff09<\/td>\n<td>2 \u6b21\u5c04\u51fb<\/td>\n<td>52.5<\/td>\n<td>59.1<\/td>\n<td>58.5<\/td>\n<td><strong>67.3<\/strong><\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>CRUXEval-O \uff08Acc.\uff09<\/td>\n<td>2 \u6b21\u5c04\u51fb<\/td>\n<td>49.8<\/td>\n<td>59.9<\/td>\n<td>59.9<\/td>\n<td><strong>69.8<\/strong><\/td>\n<\/tr>\n<tr>\n<td>\u6570\u5b66<\/td>\n<td>GSM8K \uff08\u7535\u78c1\uff09<\/td>\n<td>8 \u955c\u5934<\/td>\n<td>81.6<\/td>\n<td>88.3<\/td>\n<td>83.5<\/td>\n<td><strong>89.3<\/strong><\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>\u6570\u5b66 \uff08EM\uff09<\/td>\n<td>4 \u53d1<\/td>\n<td>43.4<\/td>\n<td>54.4<\/td>\n<td>49.0<\/td>\n<td><strong>61.6<\/strong><\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>MGSM \uff08EM\uff09<\/td>\n<td>8 \u955c\u5934<\/td>\n<td>63.6<\/td>\n<td>76.2<\/td>\n<td>69.9<\/td>\n<td><strong>79.8<\/strong><\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>CMath \uff08EM\uff09<\/td>\n<td>3 \u955c\u5934<\/td>\n<td>78.7<\/td>\n<td>84.5<\/td>\n<td>77.3<\/td>\n<td><strong>90.7<\/strong><\/td>\n<\/tr>\n<tr>\n<td>\u4e2d\u6587<\/td>\n<td>CLUEWSC \uff08EM\uff09<\/td>\n<td>5 \u955c\u5934<\/td>\n<td>82.0<\/td>\n<td>82.5<\/td>\n<td><strong>83.0<\/strong><\/td>\n<td>82.7<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>C-Eval \uff08Acc.\uff09<\/td>\n<td>5 \u955c\u5934<\/td>\n<td>81.4<\/td>\n<td>89.2<\/td>\n<td>72.5<\/td>\n<td><strong>90.1<\/strong><\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>CMMLU \uff08Acc.\uff09<\/td>\n<td>5 \u955c\u5934<\/td>\n<td>84.0<\/td>\n<td><strong>89.5<\/strong><\/td>\n<td>73.7<\/td>\n<td>88.8<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>CMRC \uff08EM\uff09<\/td>\n<td>1 \u6b21<\/td>\n<td><strong>77.4<\/strong><\/td>\n<td>75.8<\/td>\n<td>76.0<\/td>\n<td>76.3<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>C3 \uff08\u7d2f\u79ef\uff09<\/td>\n<td>0 \u6b21\u5c04\u51fb<\/td>\n<td>77.4<\/td>\n<td>76.7<\/td>\n<td><strong>79.7<\/strong><\/td>\n<td>78.6<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>CCPM \uff08\u7d2f\u79ef\uff09<\/td>\n<td>0 \u6b21\u5c04\u51fb<\/td>\n<td><strong>93.0<\/strong><\/td>\n<td>88.5<\/td>\n<td>78.6<\/td>\n<td>92.0<\/td>\n<\/tr>\n<tr>\n<td>\u591a\u79cd\u8bed\u8a00<\/td>\n<td>MMMLU-non-English \uff08Acc.\uff09<\/td>\n<td>5 \u955c\u5934<\/td>\n<td>64.0<\/td>\n<td>74.8<\/td>\n<td>73.8<\/td>\n<td><strong>79.4<\/strong><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<div class=\"markdown-alert markdown-alert-note\" dir=\"auto\">\n<p class=\"markdown-alert-title\" dir=\"auto\"><span>\u6ce8\u610f<\/span><\/p>\n<p dir=\"auto\">\u6700\u4f73\u7ed3\u679c\u4ee5\u7c97\u4f53\u663e\u793a\u3002\u5dee\u8ddd\u4e0d\u8d85\u8fc7 0.3 \u7684\u5206\u6570\u88ab\u89c6\u4e3a\u5904\u4e8e\u540c\u4e00\u6c34\u5e73\u3002DeepSeek-V3 \u5728\u5927\u591a\u6570\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u90fd\u53d6\u5f97\u4e86\u6700\u4f73\u6027\u80fd\uff0c\u5c24\u5176\u662f\u5728\u6570\u5b66\u548c\u4ee3\u7801\u4efb\u52a1\u4e0a\u3002\u6709\u5173\u66f4\u591a\u8bc4\u4f30\u8be6\u60c5\uff0c\u8bf7\u67e5\u770b\u6211\u4eec\u7684\u8bba\u6587\u3002<\/p>\n<\/div>\n<div class=\"markdown-heading\" dir=\"auto\">\n<h4 class=\"heading-element\" dir=\"auto\" tabindex=\"-1\">\u4e0a\u4e0b\u6587\u7a97\u53e3<\/h4>\n<p><a id=\"user-content-context-window\" class=\"anchor\" href=\"https:\/\/github.com\/deepseek-ai\/DeepSeek-V3#context-window\" aria-label=\"\u6c38\u4e45\u94fe\u63a5\uff1a\u4e0a\u4e0b\u6587\u7a97\u53e3\"><\/a><\/div>\n<p dir=\"auto\" align=\"center\"><a href=\"https:\/\/github.com\/deepseek-ai\/DeepSeek-V3\/blob\/main\/figures\/niah.png\" target=\"_blank\" rel=\"noopener noreferrer\"><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter size-full wp-image-2213\" title=\"niah\" src=\"https:\/\/oimi.me\/wp-content\/uploads\/2025\/01\/niah.png\" alt=\"niah\" width=\"2310\" height=\"1201\" \/><\/a><\/p>\n<p dir=\"auto\"><span>\uff08NIAH\uff09 \u6d4b\u8bd5\u7684\u8bc4\u4f30\u7ed3\u679c\u3002DeepSeek-V3 \u5728\u9ad8\u8fbe\u00a0<strong>128K<\/strong>\u00a0\u7684\u6240\u6709\u4e0a\u4e0b\u6587\u7a97\u53e3\u957f\u5ea6\u4e0a\u90fd\u8868\u73b0\u826f\u597d\u3002<\/span><code>Needle In A Haystack<\/code><\/p>\n<div class=\"markdown-heading\" dir=\"auto\">\n<h3 class=\"heading-element\" dir=\"auto\" tabindex=\"-1\">\u804a\u5929\u6a21\u578b<\/h3>\n<p><a id=\"user-content-chat-model\" class=\"anchor\" href=\"https:\/\/github.com\/deepseek-ai\/DeepSeek-V3#chat-model\" aria-label=\"\u6c38\u4e45\u94fe\u63a5\uff1a\u804a\u5929\u6a21\u578b\"><\/a><\/div>\n<div class=\"markdown-heading\" dir=\"auto\">\n<h4 class=\"heading-element\" dir=\"auto\" tabindex=\"-1\">\u6807\u51c6\u57fa\u51c6\uff08\u5927\u4e8e 67B \u7684\u578b\u53f7\uff09<\/h4>\n<p><a id=\"user-content-standard-benchmarks-models-larger-than-67b\" class=\"anchor\" href=\"https:\/\/github.com\/deepseek-ai\/DeepSeek-V3#standard-benchmarks-models-larger-than-67b\" aria-label=\"\u6c38\u4e45\u94fe\u63a5\uff1a\u6807\u51c6\u57fa\u51c6\u6d4b\u8bd5\uff08\u5927\u4e8e 67B \u7684\u6a21\u578b\uff09\"><\/a><\/div>\n<div dir=\"auto\" align=\"center\">\n<table>\n<thead>\n<tr>\n<th><\/th>\n<th><strong>\u57fa\u51c6 \uff08\u5ea6\u91cf\uff09<\/strong><\/th>\n<th><strong>\u6df1\u5ea6\u5bfb\u89c5 V2-0506<\/strong><\/th>\n<th><strong>\u6df1\u5ea6seek V2.5-0905<\/strong><\/th>\n<th><strong>Qwen2.5 72B-\u7814\u7a76\u6240<\/strong><\/th>\n<th><strong>\u9a86\u9a7c3.1 405B-\u7814\u7a76\u6240<\/strong><\/th>\n<th><strong>\u514b\u52b3\u5fb7-3.5-\u5341\u56db\u884c\u8bd7-1022<\/strong><\/th>\n<th><strong>GPT-4o 0513\u7684<\/strong><\/th>\n<th><strong>\u6df1\u5ea6seek V3<\/strong><\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td><\/td>\n<td>\u5efa\u7b51<\/td>\n<td>\u6559\u80b2\u90e8<\/td>\n<td>\u6559\u80b2\u90e8<\/td>\n<td>\u7a20<\/td>\n<td>\u7a20<\/td>\n<td>&#8211;<\/td>\n<td>&#8211;<\/td>\n<td>\u6559\u80b2\u90e8<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td># \u6fc0\u6d3b\u7684 Params<\/td>\n<td>21B<\/td>\n<td>21B<\/td>\n<td>72B \u7cfb\u5217<\/td>\n<td>405B \u7cfb\u5217<\/td>\n<td>&#8211;<\/td>\n<td>&#8211;<\/td>\n<td>\u7f16\u53f7 37B<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td># \u603b\u53c2\u6570<\/td>\n<td>\u7f16\u53f7 236B<\/td>\n<td>\u7f16\u53f7 236B<\/td>\n<td>72B \u7cfb\u5217<\/td>\n<td>405B \u7cfb\u5217<\/td>\n<td>&#8211;<\/td>\n<td>&#8211;<\/td>\n<td>671B \u7cfb\u5217<\/td>\n<\/tr>\n<tr>\n<td>\u82f1\u8bed<\/td>\n<td>MMLU \uff08EM\uff09<\/td>\n<td>78.2<\/td>\n<td>80.6<\/td>\n<td>85.3<\/td>\n<td><strong>88.6<\/strong><\/td>\n<td><strong>88.3<\/strong><\/td>\n<td>87.2<\/td>\n<td><strong>88.5<\/strong><\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>MMLU-Redux \uff08EM\uff09<\/td>\n<td>77.9<\/td>\n<td>80.3<\/td>\n<td>85.6<\/td>\n<td>86.2<\/td>\n<td><strong>88.9<\/strong><\/td>\n<td>88.0<\/td>\n<td><strong>89.1<\/strong><\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>MMLU-Pro \uff08EM\uff09<\/td>\n<td>58.5<\/td>\n<td>66.2<\/td>\n<td>71.6<\/td>\n<td>73.3<\/td>\n<td><strong>78.0<\/strong><\/td>\n<td>72.6<\/td>\n<td>75.9<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>DROP \uff083 \u6b21 F1\uff09<\/td>\n<td>83.0<\/td>\n<td>87.8<\/td>\n<td>76.7<\/td>\n<td>88.7<\/td>\n<td>88.3<\/td>\n<td>83.7<\/td>\n<td><strong>91.6<\/strong><\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>IF-Eval (Prompt Strict)<\/td>\n<td>57.7<\/td>\n<td>80.6<\/td>\n<td>84.1<\/td>\n<td>86.0<\/td>\n<td><strong>86.5<\/strong><\/td>\n<td>84.3<\/td>\n<td>86.1<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>GPQA \u94bb\u77f3 \uff08Pass@1\uff09<\/td>\n<td>35.3<\/td>\n<td>41.3<\/td>\n<td>49.0<\/td>\n<td>51.1<\/td>\n<td><strong>65.0<\/strong><\/td>\n<td>49.9<\/td>\n<td>59.1<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>SimpleQA\uff08\u6b63\u786e\uff09<\/td>\n<td>9.0<\/td>\n<td>10.2<\/td>\n<td>9.1<\/td>\n<td>17.1<\/td>\n<td>28.4<\/td>\n<td><strong>38.2<\/strong><\/td>\n<td>24.9<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>FRAMES \uff08Acc.\uff09<\/td>\n<td>66.9<\/td>\n<td>65.4<\/td>\n<td>69.8<\/td>\n<td>70.0<\/td>\n<td>72.5<\/td>\n<td><strong>80.5<\/strong><\/td>\n<td>73.3<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>LongBench v2 \uff08Acc.\uff09<\/td>\n<td>31.6<\/td>\n<td>35.4<\/td>\n<td>39.4<\/td>\n<td>36.1<\/td>\n<td>41.0<\/td>\n<td>48.1<\/td>\n<td><strong>48.7<\/strong><\/td>\n<\/tr>\n<tr>\n<td>\u6cd5\u5178<\/td>\n<td>HumanEval-Mul \uff08Pass@1\uff09<\/td>\n<td>69.3<\/td>\n<td>77.4<\/td>\n<td>77.3<\/td>\n<td>77.2<\/td>\n<td>81.7<\/td>\n<td>80.5<\/td>\n<td><strong>82.6<\/strong><\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>LiveCodeBench \uff08Pass@1-COT\uff09<\/td>\n<td>18.8<\/td>\n<td>29.2<\/td>\n<td>31.1<\/td>\n<td>28.4<\/td>\n<td>36.3<\/td>\n<td>33.4<\/td>\n<td><strong>40.5<\/strong><\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>LiveCodeBench \uff08Pass@1\uff09<\/td>\n<td>20.3<\/td>\n<td>28.4<\/td>\n<td>28.7<\/td>\n<td>30.1<\/td>\n<td>32.8<\/td>\n<td>34.2<\/td>\n<td><strong>37.6<\/strong><\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>Codeforces \uff08\u767e\u5206\u4f4d\u6570\uff09<\/td>\n<td>17.5<\/td>\n<td>35.6<\/td>\n<td>24.8<\/td>\n<td>25.3<\/td>\n<td>20.3<\/td>\n<td>23.6<\/td>\n<td><strong>51.6<\/strong><\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>SWE \u9a8c\u8bc1 \uff08\u5df2\u89e3\u51b3\uff09<\/td>\n<td>&#8211;<\/td>\n<td>22.6<\/td>\n<td>23.8<\/td>\n<td>24.5<\/td>\n<td><strong>50.8<\/strong><\/td>\n<td>38.8<\/td>\n<td>42.0<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>Aider-Edit \uff08Acc.\uff09<\/td>\n<td>60.3<\/td>\n<td>71.6<\/td>\n<td>65.4<\/td>\n<td>63.9<\/td>\n<td><strong>84.2<\/strong><\/td>\n<td>72.9<\/td>\n<td>79.7<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>Aider-Polyglot \uff08Acc.\uff09<\/td>\n<td>&#8211;<\/td>\n<td>18.2<\/td>\n<td>7.6<\/td>\n<td>5.8<\/td>\n<td>45.3<\/td>\n<td>16.0<\/td>\n<td><strong>49.6<\/strong><\/td>\n<\/tr>\n<tr>\n<td>\u6570\u5b66<\/td>\n<td>AIME 2024 \uff08Pass@1\uff09<\/td>\n<td>4.6<\/td>\n<td>16.7<\/td>\n<td>23.3<\/td>\n<td>23.3<\/td>\n<td>16.0<\/td>\n<td>9.3<\/td>\n<td><strong>39.2<\/strong><\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>\u6570\u5b66 500 \uff08EM\uff09<\/td>\n<td>56.3<\/td>\n<td>74.7<\/td>\n<td>80.0<\/td>\n<td>73.8<\/td>\n<td>78.3<\/td>\n<td>74.6<\/td>\n<td><strong>90.2<\/strong><\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>CNMO 2024 \uff08Pass@1\uff09<\/td>\n<td>2.8<\/td>\n<td>10.8<\/td>\n<td>15.9<\/td>\n<td>6.8<\/td>\n<td>13.1<\/td>\n<td>10.8<\/td>\n<td><strong>43.2<\/strong><\/td>\n<\/tr>\n<tr>\n<td>\u4e2d\u6587<\/td>\n<td>CLUEWSC \uff08EM\uff09<\/td>\n<td>89.9<\/td>\n<td>90.4<\/td>\n<td><strong>91.4<\/strong><\/td>\n<td>84.7<\/td>\n<td>85.4<\/td>\n<td>87.9<\/td>\n<td>90.9<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>C-Eval \uff08EM\uff09<\/td>\n<td>78.6<\/td>\n<td>79.5<\/td>\n<td>86.1<\/td>\n<td>61.5<\/td>\n<td>76.7<\/td>\n<td>76.0<\/td>\n<td><strong>86.5<\/strong><\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>C-SimpleQA\uff08\u6b63\u786e\uff09<\/td>\n<td>48.5<\/td>\n<td>54.1<\/td>\n<td>48.4<\/td>\n<td>50.4<\/td>\n<td>51.3<\/td>\n<td>59.3<\/td>\n<td><strong>64.8<\/strong><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<div class=\"markdown-alert markdown-alert-note\" dir=\"auto\">\n<p class=\"markdown-alert-title\" dir=\"auto\"><span>\u6ce8\u610f<\/span><\/p>\n<p dir=\"auto\">\u6240\u6709\u6a21\u578b\u90fd\u5728\u5c06\u8f93\u51fa\u957f\u5ea6\u9650\u5236\u4e3a 8K \u7684\u914d\u7f6e\u4e2d\u8fdb\u884c\u8bc4\u4f30\u3002\u5305\u542b\u5c11\u4e8e 1000 \u4e2a\u6837\u672c\u7684\u57fa\u51c6\u6d4b\u8bd5\u4f7f\u7528\u4e0d\u540c\u7684\u6e29\u5ea6\u8bbe\u7f6e\u8fdb\u884c\u591a\u6b21\u6d4b\u8bd5\uff0c\u4ee5\u83b7\u5f97\u53ef\u9760\u7684\u6700\u7ec8\u7ed3\u679c\u3002DeepSeek-V3 \u662f\u6027\u80fd\u6700\u4f73\u7684\u5f00\u6e90\u6a21\u578b\uff0c\u5e76\u4e14\u4e0e\u524d\u6cbf\u7684\u95ed\u6e90\u6a21\u578b\u76f8\u6bd4\u4e5f\u8868\u73b0\u51fa\u6709\u7ade\u4e89\u529b\u7684\u6027\u80fd\u3002<\/p>\n<\/div>\n<div class=\"markdown-heading\" dir=\"auto\">\n<h4 class=\"heading-element\" dir=\"auto\" tabindex=\"-1\">\u5f00\u653e\u5f0f\u751f\u6210\u8bc4\u4f30<\/h4>\n<p><a id=\"user-content-open-ended-generation-evaluation\" class=\"anchor\" href=\"https:\/\/github.com\/deepseek-ai\/DeepSeek-V3#open-ended-generation-evaluation\" aria-label=\"\u6c38\u4e45\u94fe\u63a5\uff1a\u5f00\u653e\u5f0f\u751f\u6210\u8bc4\u4f30\"><\/a><\/div>\n<div dir=\"auto\" align=\"center\">\n<table>\n<thead>\n<tr>\n<th>\u578b<\/th>\n<th>\u7ade\u6280\u573a-\u56f0\u96be<\/th>\n<th>\u7f8a\u9a7c\u8bc4\u4f30 2.0<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>\u6df1\u5ea6\u641c\u7d22-V2.5-0905<\/td>\n<td>76.2<\/td>\n<td>50.5<\/td>\n<\/tr>\n<tr>\n<td>Qwen2.5-72B-\u6307\u4ee4<\/td>\n<td>81.2<\/td>\n<td>49.1<\/td>\n<\/tr>\n<tr>\n<td>LLaMA-3.1 405B<\/td>\n<td>69.3<\/td>\n<td>40.5<\/td>\n<\/tr>\n<tr>\n<td>GPT-4o-0513<\/td>\n<td>80.4<\/td>\n<td>51.1<\/td>\n<\/tr>\n<tr>\n<td>\u514b\u52b3\u5fb7-\u5341\u56db\u884c\u8bd7-3.5-1022<\/td>\n<td>85.2<\/td>\n<td>52.0<\/td>\n<\/tr>\n<tr>\n<td>\u6df1\u5ea6\u641c\u7d22-V3<\/td>\n<td><strong>85.5<\/strong><\/td>\n<td><strong>70.0<\/strong><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<div class=\"markdown-alert markdown-alert-note\" dir=\"auto\">\n<p class=\"markdown-alert-title\" dir=\"auto\"><span>\u6ce8\u610f<\/span><\/p>\n<p dir=\"auto\">\u82f1\u8bed\u5f00\u653e\u5f0f\u5bf9\u8bdd\u8bc4\u4f30\u3002\u5bf9\u4e8e AlpacaEval 2.0\uff0c\u6211\u4eec\u4f7f\u7528\u957f\u5ea6\u63a7\u5236\u7684\u80dc\u7387\u4f5c\u4e3a\u6307\u6807\u3002<\/p>\n<\/div>\n<div class=\"markdown-heading\" dir=\"auto\">\n<h2 class=\"heading-element\" dir=\"auto\" tabindex=\"-1\">5. \u804a\u5929\u7f51\u7ad9\u548cAPI\u5e73\u53f0<\/h2>\n<p><a id=\"user-content-5-chat-website--api-platform\" class=\"anchor\" href=\"https:\/\/github.com\/deepseek-ai\/DeepSeek-V3#5-chat-website--api-platform\" aria-label=\"\u6c38\u4e45\u94fe\u63a5\uff1a5. \u804a\u5929\u7f51\u7ad9\u548cAPI\u5e73\u53f0\"><\/a><\/div>\n<p dir=\"auto\">\u60a8\u53ef\u4ee5\u5728 DeepSeek \u7684\u5b98\u65b9\u7f51\u7ad9\u4e0a\u4e0e DeepSeek-V3 \u804a\u5929\uff1a<a href=\"https:\/\/chat.deepseek.com\/sign_in\" rel=\"nofollow\">chat.deepseek.com<\/a><\/p>\n<p dir=\"auto\">\u6211\u4eec\u8fd8\u5728 DeepSeek \u5e73\u53f0\u4e0a\u63d0\u4f9b\u517c\u5bb9 OpenAI \u7684 API\uff1a<a href=\"https:\/\/platform.deepseek.com\/\" rel=\"nofollow\">platform.deepseek.com<\/a><\/p>\n<div class=\"markdown-heading\" dir=\"auto\">\n<h2 class=\"heading-element\" dir=\"auto\" tabindex=\"-1\">6. \u5982\u4f55\u5728\u672c\u5730\u8fd0\u884c<\/h2>\n<p><a id=\"user-content-6-how-to-run-locally\" class=\"anchor\" href=\"https:\/\/github.com\/deepseek-ai\/DeepSeek-V3#6-how-to-run-locally\" aria-label=\"\u6c38\u4e45\u94fe\u63a5\uff1a6. \u5982\u4f55\u5728\u672c\u5730\u8fd0\u884c\"><\/a><\/div>\n<p dir=\"auto\">DeepSeek-V3 \u53ef\u4ee5\u4f7f\u7528\u4ee5\u4e0b\u786c\u4ef6\u548c\u5f00\u6e90\u793e\u533a\u8f6f\u4ef6\u5728\u672c\u5730\u90e8\u7f72\uff1a<\/p>\n<ol dir=\"auto\">\n<li><strong>DeepSeek-Infer Demo<\/strong>\uff1a\u6211\u4eec\u4e3a FP8 \u548c BF16 \u63a8\u7406\u63d0\u4f9b\u4e86\u4e00\u4e2a\u7b80\u5355\u8f7b\u91cf\u7ea7\u7684 Demo\u3002<\/li>\n<li><strong>SGLang<\/strong>\uff1a\u5728 BF16 \u548c FP8 \u63a8\u7406\u6a21\u5f0f\u4e0b\u5b8c\u5168\u652f\u6301 DeepSeek-V3 \u6a21\u578b\uff0c<a href=\"https:\/\/github.com\/sgl-project\/sglang\/issues\/2591\" data-hovercard-type=\"issue\" data-hovercard-url=\"\/sgl-project\/sglang\/issues\/2591\/hovercard\">\u5373\u5c06\u63a8\u51fa<\/a>\u591a\u6807\u8bb0\u9884\u6d4b\u3002<\/li>\n<li><strong>LMDeploy<\/strong>\uff1a\u4e3a\u672c\u5730\u548c\u4e91\u90e8\u7f72\u63d0\u4f9b\u9ad8\u6548\u7684 FP8 \u548c BF16 \u63a8\u7406\u3002<\/li>\n<li><strong>TensorRT-LLM<\/strong>\u00a0\uff1a\u76ee\u524d\u652f\u6301 BF16 \u63a8\u7406\u548c INT4\/8 \u91cf\u5316\uff0c\u5373\u5c06\u63a8\u51fa FP8 \u652f\u6301\u3002<\/li>\n<li><strong>vLLM<\/strong>\uff1a\u652f\u6301 DeepSeek-V3 \u6a21\u578b\uff0c\u5177\u6709 FP8 \u548c BF16 \u6a21\u5f0f\uff0c\u7528\u4e8e\u5f20\u91cf\u5e76\u884c\u548c\u7ba1\u9053\u5e76\u884c\u3002<\/li>\n<li><strong>AMD GPU<\/strong>\uff1a\u652f\u6301\u901a\u8fc7 SGLang \u5728 BF16 \u548c FP8 \u6a21\u5f0f\u4e0b\u5728 AMD GPU \u4e0a\u8fd0\u884c DeepSeek-V3 \u6a21\u578b\u3002<\/li>\n<li><strong>\u534e\u4e3a\u6607\u817e NPU<\/strong>\uff1a\u652f\u6301\u5728\u534e\u4e3a\u6607\u817e\u8bbe\u5907\u4e0a\u8fd0\u884c DeepSeek-V3\u3002<\/li>\n<\/ol>\n<p dir=\"auto\">\u7531\u4e8e FP8 \u8bad\u7ec3\u5728\u6211\u4eec\u7684\u6846\u67b6\u4e2d\u662f\u539f\u751f\u91c7\u7528\u7684\uff0c\u56e0\u6b64\u6211\u4eec\u53ea\u63d0\u4f9b FP8 \u6743\u91cd\u3002\u5982\u679c\u60a8\u9700\u8981 BF16 \u6743\u91cd\u8fdb\u884c\u5b9e\u9a8c\uff0c\u5219\u53ef\u4ee5\u4f7f\u7528\u63d0\u4f9b\u7684\u8f6c\u6362\u811a\u672c\u6765\u6267\u884c\u8f6c\u6362\u3002<\/p>\n<p dir=\"auto\">\u4ee5\u4e0b\u662f\u5c06 FP8 \u6743\u91cd\u8f6c\u6362\u4e3a BF16 \u7684\u793a\u4f8b\uff1a<\/p>\n<div class=\"highlight highlight-source-shell notranslate position-relative overflow-auto\" dir=\"auto\">\n<pre><span class=\"pl-c1\">cd<\/span> inference\npython fp8_cast_bf16.py --input-fp8-hf-path \/path\/to\/fp8_weights --output-bf16-hf-path \/path\/to\/bf16_weights<\/pre>\n<div class=\"zeroclipboard-container\">\u5b98\u7f51<\/div>\n<\/div>\n<div class=\"markdown-alert markdown-alert-note\" dir=\"auto\">\n<p class=\"markdown-alert-title\" dir=\"auto\"><a href=\"https:\/\/www.deepseek.com\/\">https:\/\/www.deepseek.com\/<\/a><\/p>\n<\/div>\n","protected":false},"excerpt":{"rendered":"<p>DeepSeek-V3\uff0c\u8fd9\u662f\u4e00\u4e2a\u5f3a\u5927\u7684\u4e13\u5bb6\u6df7\u5408 \uff08MoE\uff09 \u8bed\u8a00\u6a21\u578b\uff0c\u603b\u5171\u6709 671B \u4e2a\u53c2\u6570\uff0c\u6bcf\u4e2a\u4ee4\u724c\u6fc0\u6d3b\u4e86 [&hellip;]<\/p>\n","protected":false},"author":2,"featured_media":2214,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[3,5],"tags":[],"class_list":["post-2212","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ai","category-lab"],"_links":{"self":[{"href":"https:\/\/myya.net\/index.php\/wp-json\/wp\/v2\/posts\/2212","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/myya.net\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/myya.net\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/myya.net\/index.php\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/myya.net\/index.php\/wp-json\/wp\/v2\/comments?post=2212"}],"version-history":[{"count":0,"href":"https:\/\/myya.net\/index.php\/wp-json\/wp\/v2\/posts\/2212\/revisions"}],"wp:attachment":[{"href":"https:\/\/myya.net\/index.php\/wp-json\/wp\/v2\/media?parent=2212"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/myya.net\/index.php\/wp-json\/wp\/v2\/categories?post=2212"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/myya.net\/index.php\/wp-json\/wp\/v2\/tags?post=2212"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}