{"id":113607,"date":"2026-10-06T23:49:17","date_gmt":"2026-10-06T15:49:17","guid":{"rendered":"https:\/\/www.wsisp.com\/helps\/113607.html"},"modified":"2026-10-06T23:49:17","modified_gmt":"2026-10-06T15:49:17","slug":"%e4%b8%80%e6%96%87%e8%af%bb%e6%87%82-deepseek-v4-1-flash%ef%bc%9a%e9%99%8d%e4%bb%b7%e7%9a%84%e4%b8%8d%e6%98%af%e8%be%93%e5%87%ba%e4%bb%b7%ef%bc%8c%e6%98%af%e4%bd%a0%e4%bb%8e%e6%b2%a1%e7%9c%8b%e8%bf%87","status":"publish","type":"post","link":"https:\/\/www.wsisp.com\/helps\/113607.html","title":{"rendered":"\u4e00\u6587\u8bfb\u61c2 DeepSeek V4.1 Flash\uff1a\u964d\u4ef7\u7684\u4e0d\u662f\u8f93\u51fa\u4ef7\uff0c\u662f\u4f60\u4ece\u6ca1\u770b\u8fc7\u7684\u90a3\u4e00\u884c"},"content":{"rendered":"

DeepSeek V4.1-Flash \u00b7 <\/span>\u67b6\u6784\u4e0e\u6210\u672c\u6df1\u5ea6\u89e3\u6790<\/span><\/p>\n

\u67b6\u6784<\/span> \u2192 <\/span>\u7f13\u5b58<\/span> \u2192 <\/span>\u5b9a\u4ef7<\/span> \u2192 <\/span>\u8d26\u5355：\u4e00\u6761\u56e0\u679c\u94fe\u8d70\u5230\u5e95<\/span><\/p>\n

\u6570\u636e\u622a\u6b62<\/span> 2026 <\/span>\u5e74<\/span> 9 <\/span>\u6708<\/span> 30 <\/span>\u65e5<\/span> ｜<\/span> \u5168\u6587\u7ea6<\/span> 2.0 <\/span>\u4e07\u5b57\u7b26\u3001<\/span>3 <\/span>\u5f20\u6587\u672c\u793a\u610f\u4e0e<\/span> 2 <\/span>\u5f20\u56fe\u8868<\/span> ｜<\/span> \u5b98\u65b9\u4e00\u624b\u6570\u636e<\/span> + <\/span>\u672c\u673a\u5b9e\u8dd1\u811a\u672c，\u4ef7\u683c\u53e3\u5f84\u5168\u90e8\u6ce8\u660e<\/span><\/p>\n

\u76ee\u5f55<\/span><\/p>\n

1. \u5f15\u5b50：\u4e00\u4e2a\u540d\u4e0d\u526f\u5b9e\u7684\u300cFlash\u300d<\/span><\/p>\n

2. \u5148\u770b\u4ef7\u683c\u8868：\u964d\u5f97\u6700\u72e0\u7684\u662f\u6700\u4e0d\u8d77\u773c\u7684\u90a3\u4e00\u884c<\/span><\/p>\n

3. \u8d26\u5355\u7684\u4e24\u534a：98.6% \u7684 token，45% \u7684\u94b1<\/span><\/p>\n

4. 890 bytes\/token：\u4e00\u6761\u56db\u4ee3\u66f2\u7ebf<\/span><\/p>\n

5. \u67b6\u6784\u4e00：CED \u4e0d\u5bf9\u79f0\u2014\u2014\u8bfb\u7528 8B，\u5199\u7528 16B<\/span><\/p>\n

6. \u67b6\u6784\u4e8c：CSA2 \u7684\u4e09\u79cd\u6ce8\u610f\u529b\u6a21\u5f0f<\/span><\/p>\n

7. \u67b6\u6784\u4e09：FP4 \u91cf\u5316\u4e0e\u7f13\u5b58\u7684\u751f\u547d\u5468\u671f<\/span><\/p>\n

8. \u4e09\u4e2a\u624b\u6bb5\u5408\u8d77\u6765\u7701\u4e86\u591a\u5c11<\/span><\/p>\n

9. \u80fd\u529b\u8d26：\u5b83\u8d62\u4e86\u54ea\u91cc，\u53c8\u8f93\u5728\u54ea\u91cc<\/span><\/p>\n

10. \u53e3\u5f84\u8b66\u544a：\u540c\u4e00\u4e2a\u6a21\u578b，\u6362\u4e2a\u5916\u58f3\u5dee 8.7 \u5206<\/span><\/p>\n

11. reasoning effort：\u4ece 25 \u62c9\u5230 100，\u4e70\u5230\u4ec0\u4e48<\/span><\/p>\n

12. \u4e09\u6761\u8bda\u5b9e\u8fb9\u754c<\/span><\/p>\n

13. \u516d\u6761\u4f7f\u7528\u89c4\u5219<\/span><\/p>\n

14. \u5e38\u89c1\u95ee\u9898<\/span><\/p>\n

15. \u6570\u636e\u65f6\u6548\u4e0e\u53c2\u8003\u8d44\u6599<\/span><\/p>\n

16. \u9644\u5f55：\u672c\u673a\u5b9e\u9a8c\u53f0<\/span><\/p>\n\n\n\n\n
\n

\u6458\u8981<\/span><\/p>\n<\/td>\n<\/tr>\n

\n

\u6458\u8981：<\/span> 9 <\/span>\u6708<\/span> 10 <\/span>\u65e5，<\/span>DeepSeek <\/span>\u53d1\u5e03<\/span> V4.1-Flash<\/span>\u3002\u5b98\u65b9\u7ed9\u7684\u6700\u9192\u76ee\u7684\u4e24\u4e2a\u6570\u5b57\u662f\u300c<\/span>552B <\/span>\u53c2\u6570\u300d\u548c\u300c\u8f93\u51fa\u4ef7\u964d<\/span> 11%<\/span>\u300d<\/span>\u2014\u2014<\/span>\u51e0\u4e4e\u6240\u6709\u62a5\u9053\u90fd\u4f1a\u5f15\u7528\u5b83\u4eec\u3002\u4f46\u5982\u679c\u4f60\u628a\u4e00\u4e2a\u771f\u5b9e<\/span> Agent <\/span>\u7684\u8d26\u5355\u6309\u65b0\u4ef7\u683c\u8868\u91cd\u7b97\u4e00\u904d，\u4f1a\u53d1\u73b0\u8fd9\u4efd\u8d26\u5355\u5b9e\u9645\u964d\u4e86<\/span> 46%<\/span>\u3002\u5dee\u989d\u4e0d\u5728\u8f93\u51fa\u90a3\u4e00\u884c，\u5728\u6807\u4ef7\u8868\u6700\u4e0b\u9762\u90a3\u4e00\u884c：<\/span>\u7f13\u5b58\u547d\u4e2d<\/span>\u3002\u672c\u6587\u6cbf\u7740\u300c\u67b6\u6784<\/span> \u2192 <\/span>\u7f13\u5b58<\/span> \u2192 <\/span>\u5b9a\u4ef7<\/span> \u2192 <\/span>\u8d26\u5355\u300d\u8fd9\u6761\u56e0\u679c\u94fe\u8d70\u4e00\u904d：<\/span>CED <\/span>\u4e0d\u5bf9\u79f0\u67b6\u6784\u600e\u4e48\u628a<\/span> 100 <\/span>\u4e07<\/span> token <\/span>\u4e0a\u4e0b\u6587\u7684<\/span> KV cache <\/span>\u538b\u5230<\/span> 890 <\/span>\u5b57\u8282<\/span>\/token<\/span>，<\/span>CSA2 <\/span>\u7684\u4e09\u79cd\u6ce8\u610f\u529b\u6a21\u5f0f\u600e\u4e48\u5728\u5c42\u4e0e\u5c42\u4e4b\u95f4\u590d\u7528\u8bb0\u5fc6，\u4ee5\u53ca\u4e3a\u4ec0\u4e48\u8fd9\u4e09\u4ef6\u770b\u8d77\u6765\u6781\u5176\u5e95\u5c42\u7684\u4e8b，\u6700\u540e\u4f1a\u53d8\u6210\u4f60\u8d26\u5355\u4e0a\u7684\u4e00\u884c\u6570\u5b57\u3002\u4e00\u53e5\u8bdd\u7ed3\u8bba：<\/span>\u8fd9\u8f6e\u964d\u4ef7\u7684\u771f\u8eab，\u662f\u300c\u8ba9 Agent \u7684\u91cd\u8bfb\u53d8\u4fbf\u5b9c\u300d，\u800c\u4e0d\u662f\u300c\u8ba9\u751f\u6210\u53d8\u4fbf\u5b9c\u300d\u3002<\/span><\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

\u5173\u4e8e\u672c\u6587\u6570\u5b57\u7684\u4e09\u79cd\u53e3\u5f84<\/span>（\u6cbf\u7528\u672c\u7cfb\u5217\u60ef\u4f8b）：<\/span><\/p>\n

  • \u5b98\u65b9\u4e00\u624b<\/span>：\u6765\u81ea<\/span> DeepSeek <\/span>\u5b98\u65b9\u65b0\u95fb\u9875\u3001\u5b98\u65b9\u5b9a\u4ef7\u9875\u3001\u5b98\u65b9\u6280\u672f\u62a5\u544a，\u6587\u5185\u76f4\u63a5\u6ce8\u660e\u3002<\/span><\/li>\n
  • \u672c\u673a\u63a8\u7b97<\/span>：\u57fa\u4e8e\u5b98\u65b9\u8d39\u7387\u4e0e\u672c\u673a\u5b9e\u8dd1\u811a\u672c（\u9644\u5f55）\u7b97\u51fa\u7684\u6570\u5b57，\u6807\u4e3a\u300c\u672c\u673a\u63a8\u7b97\u300d\u3002<\/span><\/li>\n
  • \u4e09\u65b9\u53e3\u5f84<\/span>：\u6765\u81ea\u516c\u5f00\u5a92\u4f53\u62a5\u9053\u4e0e\u7b2c\u4e09\u65b9\u63a8\u7406\u5e73\u53f0\u5b9a\u4ef7，\u6ce8\u660e\u6765\u6e90；\u8fd9\u7c7b\u6570\u5b57\u4f1a\u53d8，\u53ea\u4f5c\u53c2\u8003\u3002<\/span><\/li>\n

    \u65f6\u6548\u58f0\u660e<\/span>：<\/span>V4.1-Flash <\/span>\u4e8e<\/span> 2026 <\/span>\u5e74<\/span> 9 <\/span>\u6708<\/span> 10 <\/span>\u65e5\u53d1\u5e03，\u672c\u6587\u5199\u4f5c\u65f6\u8ddd\u53d1\u5e03<\/span> 20 <\/span>\u5929\u3002\u5b98\u65b9\u5df2\u660e\u786e<\/span> deepseek-v4-pro<\/span> \u81ea<\/span> 9 <\/span>\u6708<\/span> 14 <\/span>\u65e5\u8d77\u8def\u7531\u81f3<\/span> Flash<\/span>，\u76f4\u5230<\/span> V4.1-Pro <\/span>\u53d1\u5e03<\/span>\u2014\u2014<\/span>\u8fd9\u610f\u5473\u7740\u8fd9\u4e2a\u6a21\u578b\u6240\u5904\u7684\u7248\u672c\u7a97\u53e3\u4ecd\u5728\u53d8\u52a8\u4e2d<\/span>\u3002\u5f15\u7528\u4ef7\u683c\u524d\u8bf7\u6838\u5bf9\u5b98\u65b9\u5b9a\u4ef7\u9875\u3002<\/span><\/p>\n

    1. \u5f15\u5b50：\u4e00\u4e2a\u540d\u4e0d\u526f\u5b9e\u7684\u300cFlash\u300d<\/span><\/h3>\n

    \u5148\u770b\u4e09\u7ec4\u6570\u5b57\u3002<\/span><\/p>\n\n\n\n\n\n\n
    \n

    \u9879\u76ee<\/span><\/p>\n<\/td>\n

    		\n

    \u5b83\u8981\u53d6\u4ee3\u7684<\/span> V4-Flash<\/span><\/p>\n<\/td>\n

    		\n

    V4.1-Flash<\/span><\/p>\n<\/td>\n

    		\n

    \u53d8\u5316<\/span><\/p>\n<\/td>\n<\/tr>\n

    \n

    \u4e3b\u5e72\u53c2\u6570<\/span><\/p>\n<\/td>\n

    		\n

    284B<\/span><\/p>\n<\/td>\n

    		\n

    552B<\/span><\/p>\n<\/td>\n

    		\n

    \u5927 94%<\/span><\/p>\n<\/td>\n<\/tr>\n

    \n

    \u6bcf<\/span> token <\/span> \u6fc0\u6d3b\u53c2\u6570<\/span><\/p>\n<\/td>\n

    		\n

    13B<\/span><\/p>\n<\/td>\n

    		\n

    8B<\/span> （\u8bfb）<\/span> \/ 16B<\/span> （\u5199）<\/span><\/p>\n<\/td>\n

    		\n

    \u8bfb\u66f4\u5c11<\/span><\/p>\n<\/td>\n<\/tr>\n

    \n

    \u540d\u5b57<\/span><\/p>\n<\/td>\n

    		\n

    Flash<\/span> （\u5feb）<\/span><\/p>\n<\/td>\n

    		\n

    Flash<\/span> （\u5feb）<\/span><\/p>\n<\/td>\n

    		\n

    \u4e00\u6837<\/span><\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

    \u4e00\u4e2a\u53eb\u300c<\/span>Flash<\/span>\u300d\u7684\u6a21\u578b，\u4e3b\u5e72\u6bd4\u5b83\u53d6\u4ee3\u7684\u90a3\u4e2a\u5927\u4e86\u4e00\u500d<\/span>\u2014\u2014<\/span>\u8fd9\u4ef6\u4e8b\u672c\u8eab\u5c31\u503c\u5f97\u505c\u4e0b\u6765\u60f3\u4e00\u60f3\u3002<\/span><\/p>\n

    \u518d\u770b\u5b83\u7684\u6210\u7ee9：\u5b98\u65b9\u5bf9\u6bd4\u8868\u91cc，<\/span>V4.1-Flash <\/span>\u5728<\/span>\u4e24\u8fb9\u90fd\u6709\u5206\u7684 14 \u884c\u57fa\u51c6\u91cc\u8d62\u4e86 12 \u884c<\/span>，\u5bf9\u624b\u662f\u81ea\u5bb6\u7684<\/span> V4-Pro\u2014\u2014<\/span>\u4e00\u4e2a<\/span> 1.6T <\/span>\u4e3b\u5e72\u3001<\/span>49B <\/span>\u6fc0\u6d3b\u7684\u65d7\u8230\u3002\u800c\u5b83\u7528\u7684\u662f<\/span>\u4e09\u5206\u4e4b\u4e00\u7684\u4e3b\u5e72\u53c2\u6570\u3001\u516d\u5206\u4e4b\u4e00\u7684\u6fc0\u6d3b\u53c2\u6570<\/span>\u3002<\/span><\/p>\n

    \u6700\u540e\u770b\u4ef7\u683c：\u5b83\u6bd4\u4e0a\u4e00\u4ee3<\/span>\u66f4\u4fbf\u5b9c<\/span>\u3002<\/span><\/p>\n

    \u5927\u4e86\u4e00\u500d\u3001\u5f3a\u4e86\u4e00\u4e2a\u6863\u3001\u8fd8\u66f4\u4fbf\u5b9c\u3002<\/span> \u8fd9\u4e09\u4ef6\u4e8b\u540c\u65f6\u6210\u7acb，\u5c31\u610f\u5473\u7740\u65e7\u7684\u90a3\u4e2a\u300c\u53c2\u6570\u91cf<\/span> = <\/span>\u6210\u672c\u300d\u7684\u76f4\u89c9，\u5728\u8fd9\u4e2a\u6a21\u578b\u4e0a\u5931\u6548\u4e86\u3002\u5931\u6548\u7684\u539f\u56e0，\u662f\u8fd9\u8f6e\u53d1\u5e03\u91cc\u771f\u6b63\u91cd\u8981\u7684\u4e1c\u897f<\/span>\u2014\u2014<\/span>\u4e0d\u662f\u53c2\u6570\u91cf，\u662f<\/span>\u67b6\u6784\u628a\u300c\u8bfb\u300d\u548c\u300c\u5199\u300d\u62c6\u5f00\u4e86<\/span>\u3002<\/span><\/p>\n

    \u5148\u8bf4\u4e00\u6761\u65f6\u95f4\u7ebf，\u56e0\u4e3a\u5b83\u51b3\u5b9a\u4e86\u4f60\u6b64\u523b\u770b\u5230\u7684\u4ef7\u683c\u662f\u4e34\u65f6\u7684\u8fd8\u662f\u957f\u671f\u7684：<\/span><\/p>\n\n\n\n\n\n\n\n
    \n

    \u65f6\u95f4<\/span><\/p>\n<\/td>\n

    		\n

    \u4e8b\u4ef6<\/span><\/p>\n<\/td>\n<\/tr>\n

    \n

    2026-08-13<\/span><\/p>\n<\/td>\n

    		\n

    V4-Pro <\/span> \u6b63\u5f0f\u7248（<\/span> 1.6T <\/span> \u4e3b\u5e72<\/span> \/ 49B <\/span> \u6fc0\u6d3b）<\/span><\/p>\n<\/td>\n<\/tr>\n

    \n

    2026-09-10<\/span><\/p>\n<\/td>\n

    		\n

    V4.1-Flash <\/span> \u53d1\u5e03；\u65b0\u4ef7\u4e8e<\/span> 04:00 UTC <\/span> \u751f\u6548<\/span><\/p>\n<\/td>\n<\/tr>\n

    \n

    2026-09-14<\/span><\/p>\n<\/td>\n

    		\n

    deepseek-v4-pro<\/span> \u7684\u8bf7\u6c42\u5f00\u59cb\u8def\u7531\u81f3<\/span> Flash<\/span><\/p>\n<\/td>\n<\/tr>\n

    \n

    \u5f85\u5b9a<\/span><\/p>\n<\/td>\n

    		\n

    V4.1-Pro <\/span> \u53d1\u5e03，<\/span> Pro <\/span> \u7684\u8def\u7531\u968f\u4e4b\u7ec8\u6b62<\/span><\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

    \u672c\u6587\u8981\u56de\u7b54\u4e09\u4e2a\u95ee\u9898：<\/span><\/p>\n

  • \u8fd9\u8f6e\u964d\u4ef7，\u7a76\u7adf\u964d\u7684\u662f\u54ea\u4e00\u884c？<\/span><\/li>\n
  • \u4e3a\u4ec0\u4e48\u67b6\u6784\u80fd\u51b3\u5b9a\u8d26\u5355？<\/span><\/li>\n
  • \u5b83\u8d62\u4e86\u4ec0\u4e48\u3001\u53c8\u8f93\u5728\u54ea\u91cc，\u4ec0\u4e48\u4efb\u52a1\u4e0d\u8be5\u7528\u5b83？<\/span><\/li>\n

    2. \u5148\u770b\u4ef7\u683c\u8868：\u964d\u5f97\u6700\u72e0\u7684\u662f\u6700\u4e0d\u8d77\u773c\u7684\u90a3\u4e00\u884c<\/span><\/h3>\n

    \u5b98\u65b9\u5b9a\u4ef7\u9875\u7ed9\u51fa\u7684\u5b8c\u6574\u4ef7\u76ee\u8868\u5982\u4e0b（\u5355\u4f4d：\u7f8e\u5143<\/span> \/ <\/span>\u6bcf\u767e\u4e07<\/span> token<\/span>）：<\/span><\/p>\n\n\n\n\n\n\n
    \n

    \u8ba1\u8d39\u9879<\/span><\/p>\n<\/td>\n

    		\n

    Flash <\/span> \u95f2\u65f6<\/span><\/p>\n<\/td>\n

    		\n

    Flash <\/span> \u9ad8\u5cf0<\/span><\/p>\n<\/td>\n

    		\n

    V4-Pro <\/span> \u95f2\u65f6<\/span><\/p>\n<\/td>\n

    		\n

    V4-Pro <\/span> \u9ad8\u5cf0<\/span><\/p>\n<\/td>\n<\/tr>\n

    \n

    \u8f93\u5165<\/span> \u00b7<\/span> \u7f13\u5b58\u547d\u4e2d<\/span><\/p>\n<\/td>\n

    		\n

    0.003<\/span><\/p>\n<\/td>\n

    		\n

    0.006<\/span><\/p>\n<\/td>\n

    		\n

    0.022<\/span><\/p>\n<\/td>\n

    		\n

    0.044<\/span><\/p>\n<\/td>\n<\/tr>\n

    \n

    \u8f93\u5165<\/span> \u00b7<\/span> \u7f13\u5b58\u672a\u547d\u4e2d<\/span><\/p>\n<\/td>\n

    		\n

    0.15<\/span><\/p>\n<\/td>\n

    		\n

    0.30<\/span><\/p>\n<\/td>\n

    		\n

    0.66<\/span><\/p>\n<\/td>\n

    		\n

    1.32<\/span><\/p>\n<\/td>\n<\/tr>\n

    \n

    \u8f93\u51fa<\/span><\/p>\n<\/td>\n

    		\n

    0.60<\/span><\/p>\n<\/td>\n

    		\n

    1.20<\/span><\/p>\n<\/td>\n

    		\n

    1.98<\/span><\/p>\n<\/td>\n

    		\n

    3.96<\/span><\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

    \u4e24\u6761\u89c4\u5219\u8981\u8bb0\u4f4f：<\/span><\/p>\n

      \n
    • \u95f2\u65f6\u4ef7 = \u9ad8\u5cf0\u4ef7\u7684\u4e00\u534a\u3002<\/span> \u9ad8\u5cf0\u65f6\u6bb5\u662f\u5468\u4e00\u81f3\u5468\u4e94<\/span> 01:00\u201304:00 <\/span>\u4e0e<\/span> 06:00\u201310:00 UTC<\/span>（\u5317\u4eac\u65f6\u95f4<\/span> 09:00\u201312:00 <\/span>\u4e0e<\/span> 14:00\u201318:00<\/span>），<\/span>\u4e2d\u56fd\u6cd5\u5b9a\u8282\u5047\u65e5\u4e0d\u7b97\u9ad8\u5cf0<\/span>\u3002\u4e00\u5468<\/span> 168 <\/span>\u5c0f\u65f6\u91cc，\u9ad8\u5cf0\u53ea\u5360<\/span> 35 <\/span>\u5c0f\u65f6\u3002<\/span><\/li>\n
    • \u7f13\u5b58\u547d\u4e2d\u7684\u5355\u4ef7，\u662f\u672a\u547d\u4e2d\u7684 1\/50\u3002<\/span> \u95f2\u65f6<\/span> 0.003 <\/span>\u5bf9<\/span> 0.15<\/span>，\u6b63\u597d<\/span> 50 <\/span>\u500d\u3002<\/span><\/li>\n<\/ul>\n

      \u73b0\u5728\u770b\u964d\u5e45\u3002\u628a<\/span> V4-Flash <\/span>\u7684\u65e7\u4ef7\u548c<\/span> V4.1-Flash <\/span>\u7684\u65b0\u4ef7\u5e76\u6392（\u4eba\u6c11\u5e01\u53e3\u5f84，\u95f2\u65f6）：<\/span><\/p>\n\n\n\n\n\n\n
      \n

      \u8ba1\u8d39\u9879<\/span><\/p>\n<\/td>\n

      		\n

      V4-Flash<\/span><\/p>\n<\/td>\n

      		\n

      V4.1-Flash<\/span><\/p>\n<\/td>\n

      		\n

      \u964d\u5e45<\/span><\/p>\n<\/td>\n<\/tr>\n

      \n

      \u7f13\u5b58\u547d\u4e2d<\/span><\/p>\n<\/td>\n

      		\n

      0.05 <\/span> \u5143<\/span><\/p>\n<\/td>\n

      		\n

      0.02 <\/span> \u5143<\/span><\/p>\n<\/td>\n

      		\n

      \u221260.0%<\/span><\/p>\n<\/td>\n<\/tr>\n

      \n

      \u7f13\u5b58\u672a\u547d\u4e2d<\/span><\/p>\n<\/td>\n

      		\n

      1.50 <\/span> \u5143<\/span><\/p>\n<\/td>\n

      		\n

      1.00 <\/span> \u5143<\/span><\/p>\n<\/td>\n

      		\n

      \u221233.3%<\/span><\/p>\n<\/td>\n<\/tr>\n

      \n

      \u8f93\u51fa<\/span><\/p>\n<\/td>\n

      		\n

      4.50 <\/span> \u5143<\/span><\/p>\n<\/td>\n

      		\n

      4.00 <\/span> \u5143<\/span><\/p>\n<\/td>\n

      		\n

      \u221211.1%<\/span><\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

      \u770b\u51fa\u95ee\u9898\u4e86\u5417？<\/span><\/p>\n

      \u4e09\u6761\u7ebf\u91cc，\u964d\u5e45\u6700\u5927\u7684\u662f\u7f13\u5b58\u547d\u4e2d（\u221260%），\u6700\u5c0f\u7684\u662f\u8f93\u51fa（\u221211.1%）\u3002<\/span> \u800c\u7edd\u5927\u591a\u6570\u62a5\u9053\u4f1a\u5f15\u7528\u7684，\u6070\u6070\u662f\u90a3\u4e2a\u6700\u5c0f\u7684\u6570\u5b57\u3002<\/span><\/p>\n

      \u00a0\u00a0\u8fd9\u4e0d\u662f\u5a92\u4f53\u5077\u61d2，\u662f\u300c\u770b\u5f97\u89c1\u7684\u4e1c\u897f\u300d\u548c\u300c\u91cd\u8981\u7684\u4e1c\u897f\u300d\u7ecf\u5e38\u4e0d\u662f\u540c\u4e00\u4e2a\u3002<\/span> \u8f93\u51fa\u4ef7\u662f\u6700\u597d\u7406\u89e3\u7684<\/span>\u2014\u2014<\/span>\u4f60\u751f\u6210\u591a\u5c11\u5b57，\u5c31\u4ed8\u591a\u5c11\u94b1\u3002\u7f13\u5b58\u547d\u4e2d\u4ef7\u5219\u9700\u8981\u4f60\u5148\u7406\u89e3\u300c\u7f13\u5b58\u300d\u662f\u4ec0\u4e48，\u624d\u80fd\u77e5\u9053\u5b83\u4e3a\u4ec0\u4e48\u91cd\u8981\u3002<\/span><\/p>\n

      \u90a3\u5b83\u5230\u5e95\u6709\u591a\u91cd\u8981？\u770b\u4e0b\u4e00\u8282\u3002<\/span><\/p>\n

      3. \u8d26\u5355\u7684\u4e24\u534a：98.6% \u7684 token，45% \u7684\u94b1<\/span><\/h3>\n

      \u8fd9\u4e00\u8282\u662f\u5168\u6587\u7684\u6838\u5fc3，\u4e5f\u662f\u6700\u5bb9\u6613\u88ab\u8bf4\u9519\u7684\u4e00\u8282\u3002<\/span><\/p>\n

      \u6709\u4e00\u4efd\u7b2c\u4e09\u65b9\u5bf9<\/span>\u771f\u5b9e Agent \u6d41\u91cf<\/span>\u7684<\/span> 30 <\/span>\u5929\u89c2\u6d4b\u6570\u636e，\u7ed3\u6784\u662f\u8fd9\u6837\u7684：<\/span><\/p>\n\n\n\n\n
      \n

      \u793a\u610f\u56fe<\/span> \u00b7 <\/span> \u6587\u672c\u7ed8\u5236<\/span><\/p>\n<\/td>\n<\/tr>\n

      \n

      \u00a0\u00a0\u65b0\u8f93\u5165<\/span> 26.7M token<\/span><\/p>\n

      \u00a0\u00a0\u8f93\u51fa<\/span> 5.0M token<\/span><\/p>\n

      \u00a0\u00a0\u7f13\u5b58\u91cd\u8bfb<\/span> 1,920.0M token <- 19.2 <\/span>\u4ebf<\/span><\/p>\n

      \u00a0\u00a0————————————<\/span><\/p>\n

      \u00a0\u00a0\u5408\u8ba1<\/span> 1,951.7M token<\/span><\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

      \u8fd9\u91cc\u9762\u6709\u4e2a\u975e\u5e38\u53cd\u76f4\u89c9\u7684\u4e8b\u5b9e：<\/span>\u8f93\u51fa\u53ea\u6709 500 \u4e07 token，\u7f13\u5b58\u91cd\u8bfb\u6709 19.2 \u4ebf\u3002<\/span><\/p>\n

      \u4e3a\u4ec0\u4e48？\u56e0\u4e3a<\/span> Agent <\/span>\u7684\u5de5\u4f5c\u65b9\u5f0f\u5c31\u662f\u53cd\u590d\u91cd\u8bfb\u81ea\u5df1：<\/span><\/p>\n

        \n
      • \u6bcf\u4e00\u8f6e\u90fd\u8981\u5e26\u4e0a<\/span> System Prompt<\/span><\/li>\n
      • \u6bcf\u4e00\u8f6e\u90fd\u8981\u5e26\u4e0a\u9879\u76ee\u8bf4\u660e<\/span><\/li>\n
      • \u6bcf\u4e00\u8f6e\u90fd\u8981\u5e26\u4e0a\u5de5\u5177\u5b9a\u4e49<\/span><\/li>\n
      • \u6bcf\u4e00\u8f6e\u90fd\u8981\u5e26\u4e0a\u524d\u9762\u51e0\u5341\u8f6e\u7684\u5386\u53f2<\/span><\/li>\n
      • \u6bcf\u4e00\u8f6e\u90fd\u8981\u5e26\u4e0a\u4e0a\u4e00\u6b65\u7684\u6267\u884c\u7ed3\u679c<\/span><\/li>\n<\/ul>\n

        \u540c\u4e00\u4e2a\u524d\u7f00，\u5b83\u8bfb\u4e86 100 \u904d\u3002<\/span><\/p>\n

        \u00a0\u00a0\u628a\u5b83\u60f3\u6210\u4e00\u672c\u83dc\u8c31\u3002<\/span> \u4f60\u6bcf\u505a\u4e00\u9053\u83dc\u90fd\u4e0d\u91cd\u5370\u6574\u672c\u4e66，\u53ea\u662f\u628a\u4e66\u644a\u5728\u684c\u4e0a，\u7ffb\u5230\u540c\u4e00\u9875\u518d\u770b\u4e00\u904d\u3002<\/span>Agent <\/span>\u4e5f\u662f\u8fd9\u6837<\/span>\u2014\u2014<\/span>\u5b83\u4e0d\u91cd\u65b0<\/span>"<\/span>\u60f3<\/span>"<\/span>\u8fd9\u4e9b\u5185\u5bb9，\u5b83\u91cd\u65b0<\/span>"<\/span>\u770b<\/span>"<\/span>\u8fd9\u4e9b\u5185\u5bb9\u3002\u800c<\/span>"<\/span>\u91cd\u65b0\u770b<\/span>"<\/span>\u662f\u8981\u4ed8\u8d39\u7684\u3002<\/span><\/p>\n

        \u73b0\u5728\u5173\u952e\u7684\u90e8\u5206\u6765\u4e86\u3002\u8fd9<\/span> 19.2 <\/span>\u4ebf\u7f13\u5b58\u91cd\u8bfb，<\/span>\u5360 token \u603b\u91cf\u7684 98.4%<\/span>（\u5982\u679c\u53ea\u7b97\u8f93\u5165\u4fa7，\u662f<\/span> 98.6%\u2014\u2014<\/span>\u5a92\u4f53\u5f15\u7528\u7684\u5c31\u662f\u8fd9\u4e2a\u53e3\u5f84）\u3002<\/span><\/p>\n

        \u4f46<\/span>\u5728\u94b1\u4e0a，\u5b83\u662f\u591a\u5c11？<\/span><\/p>\n\n\n\n\n\n\n\n
        \n

        \u6210\u672c\u9879<\/span><\/p>\n<\/td>\n

        		\n

        \u91d1\u989d<\/span><\/p>\n<\/td>\n

        		\n

        \u5360\u6bd4<\/span><\/p>\n<\/td>\n<\/tr>\n

        \n

        \u7f13\u5b58\u91cd\u8bfb<\/span><\/p>\n<\/td>\n

        		\n

        $5.76<\/span><\/p>\n<\/td>\n

        		\n

        45.1%<\/span><\/p>\n<\/td>\n<\/tr>\n

        \n

        \u65b0\u8f93\u5165<\/span><\/p>\n<\/td>\n

        		\n

        $4.00<\/span><\/p>\n<\/td>\n

        		\n

        31.4%<\/span><\/p>\n<\/td>\n<\/tr>\n

        \n

        \u8f93\u51fa<\/span><\/p>\n<\/td>\n

        		\n

        $3.00<\/span><\/p>\n<\/td>\n

        		\n

        23.5%<\/span><\/p>\n<\/td>\n<\/tr>\n

        \n

        \u5408\u8ba1<\/span><\/p>\n<\/td>\n

        		\n

        $12.77<\/span><\/p>\n<\/td>\n

        		\n

        100%<\/span><\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n\n\n\n
        \n

        token \u4e0a\u5360 98.6% \u7684\u90a3\u4e00\u5757，\u6210\u672c\u4e0a\u53ea\u5360 45.1%\u3002<\/span>\u56e0\u4e3a\u5b83\u7684\u5355\u4ef7\u53ea\u6709\u672a\u547d\u4e2d\u7684<\/span> 1\/50<\/span>\u3002<\/span><\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

        \u8fd9\u4e0d\u662f\u77db\u76fe，\u8fd9\u662f<\/span>\u540c\u4e00\u4e2a\u4e8b\u5b9e\u7684\u4e24\u4e2a\u9762<\/span>\u3002\u800c\u4e14\u4e24\u4e2a\u9762\u90fd\u5f88\u6709\u7528：<\/span><\/p>\n

          \n
        • \u770b<\/span> token <\/span>\u5360\u6bd4<\/span> \u2192 <\/span>\u4f60\u77e5\u9053<\/span>\u4f18\u5316\u65b9\u5411\u5728\u54ea<\/span>（\u91cd\u8bfb）<\/span><\/li>\n
        • \u770b\u6210\u672c\u5360\u6bd4<\/span> \u2192 <\/span>\u4f60\u77e5\u9053<\/span>\u4f18\u5316\u7a7a\u95f4\u6709\u591a\u5927<\/span>（<\/span>45%<\/span>）<\/span><\/li>\n<\/ul>\n

          \u00a0\u00a0\u300c\u5360\u6bd4\u300d\u8fd9\u4ef6\u4e8b，\u770b token \u548c\u770b\u94b1\u662f\u4e24\u4e2a\u7b54\u6848\u3002<\/span> \u5982\u679c\u4f60\u770b\u5230\u4e00\u7bc7\u6587\u7ae0\u8bf4\u300c\u7f13\u5b58\u5360<\/span> Agent <\/span>\u8d26\u5355\u7684<\/span> 98.6%<\/span>\u300d，\u90a3\u4e0d\u662f\u7b97\u9519\u4e86，\u5c31\u662f\u6df7\u6dc6\u4e86\u53e3\u5f84<\/span>\u2014\u2014<\/span>\u5b83\u8bf4\u7684\u662f<\/span> token，\u4e0d\u662f\u94b1\u3002<\/span>\"\"\u8d26\u5355\u7684\u4e24\u5f20\u8138：token <\/span>\u5360\u6bd4<\/span> 98.4% <\/span>\u5bf9\u6210\u672c\u5360\u6bd4<\/span> 45.1%<\/span>，\u4ee5\u53ca\u56db\u5bb6\u7684\u7f13\u5b58\u8bfb\u53d6\u4ef7<\/span><\/p>\n

          \u672c\u673a\u63a8\u7b97\u7684 <\/span>$12.77<\/span>，\u4e0e\u7b2c\u4e09\u65b9\u516c\u5e03\u7684 <\/span>$12.78<\/span> \u543b\u5408\u5230\u5206（\u53e3\u5f84\u4e00\u81f4），\u53ef\u4f5c\u4e3a\u4ea4\u53c9\u6821\u9a8c\u3002<\/span><\/p>\n

          \u90a3\u300c\u964d 60%\u300d\u5230\u5e95\u503c\u591a\u5c11？<\/span><\/h4>\n

          \u628a\u540c\u4e00\u4efd\u6d41\u91cf，\u5206\u522b\u6309\u4e24\u4ee3\u4ef7\u683c\u7b97（\u4eba\u6c11\u5e01，\u95f2\u65f6，\u672c\u673a\u63a8\u7b97）：<\/span><\/p>\n\n\n\n\n\n\n
          \n

          \n<\/td>\n

          		\n

          \u91d1\u989d<\/span><\/p>\n<\/td>\n<\/tr>\n

          \n

          \u6309<\/span> V4-Flash <\/span> \u65e7\u4ef7<\/span><\/p>\n<\/td>\n

          		\n

          \u00a5158.55<\/span><\/p>\n<\/td>\n<\/tr>\n

          \n

          \u6309<\/span> V4.1-Flash <\/span> \u65b0\u4ef7<\/span><\/p>\n<\/td>\n

          		\n

          \u00a585.10<\/span><\/p>\n<\/td>\n<\/tr>\n

          \n

          \u8fd9\u4efd\u8d26\u5355\u7684\u5b9e\u9645\u964d\u5e45<\/span><\/p>\n<\/td>\n

          		\n

          \u221246.3%<\/span><\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

          \u5bf9\u7167\u4e00\u4e0b：<\/span>\u8f93\u51fa\u5355\u4ef7\u53ea\u964d\u4e86 11.1%\u3002<\/span><\/p>\n

          \u00a0\u00a0\u53ea\u76ef\u8f93\u51fa\u5355\u4ef7\u7684\u4eba，\u4f1a\u628a\u8fd9\u6b21\u964d\u4ef7\u4f4e\u4f30\u8fd1\u56db\u500d\u3002<\/span> \u56e0\u4e3a\u5728<\/span> Agent <\/span>\u573a\u666f\u91cc，\u8f93\u51fa\u672c\u6765\u5c31\u4e0d\u662f\u8d26\u5355\u7684\u4e3b\u4f53<\/span>\u2014\u2014<\/span>\u91cd\u8bfb\u624d\u662f\u3002<\/span><\/p>\n

          \u8fd8\u6709\u66f4\u6781\u7aef\u7684\u4e00\u7ec4\u6570：\u5982\u679c\u8fd9<\/span> 19.2 <\/span>\u4ebf<\/span> token <\/span>\u5168\u90fd\u6309\u300c\u672a\u547d\u4e2d\u300d\u8ba1\u8d39<\/span>（\u76f8\u5f53\u4e8e\u6839\u672c\u6ca1\u6709\u7f13\u5b58\u673a\u5236），\u8d26\u5355\u662f <\/span>$295.00<\/span>，<\/span>\u662f\u73b0\u4ef7\u7684 23.1 \u500d<\/span>\u3002<\/span><\/p>\n\n\n\n
          \n

          \u7f13\u5b58\u673a\u5236\u672c\u8eab\u7701\u4e0b\u7684\u94b1，\u6bd4\u8fd9\u6b21\u964d\u4ef7\u7701\u4e0b\u7684\u94b1\u591a\u5f97\u591a\u3002<\/span> \u964d\u4ef7\u662f\u5728\u4e00\u4e2a\u5df2\u7ecf\u5f88\u7701\u94b1\u7684\u5730\u57fa\u4e0a\u518d\u964d\u4e00\u70b9\u3002<\/span><\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

          \u8fd9\u4e5f\u89e3\u91ca\u4e86\u4e3a\u4ec0\u4e48\u8fd9\u4e00\u8282\u8981\u82b1\u8fd9\u4e48\u591a\u7bc7\u5e45：<\/span>\u5982\u679c\u4f60\u60f3\u7701\u94b1，\u7b2c\u4e00\u4e2a\u8be5\u76ef\u7684\u6570\u4e0d\u662f\u5355\u4ef7，\u662f\u7f13\u5b58\u547d\u4e2d\u7387\u3002<\/span><\/p>\n

          \u7f13\u5b58\u547d\u4e2d\u7387：\u6700\u8be5\u76ef\u7684\u4e00\u4e2a\u6570<\/span><\/h4>\n

          \u672c\u673a\u63a8\u7b97：\u6bcf\u6708<\/span> 1 <\/span>\u4ebf\u8f93\u5165<\/span> + 2000 <\/span>\u4e07\u8f93\u51fa，\u5168\u90e8\u95f2\u65f6\u8fd0\u884c（\u4eba\u6c11\u5e01）：<\/span><\/p>\n\n\n\n\n\n\n\n\n
          \n

          \u7f13\u5b58\u547d\u4e2d\u7387<\/span><\/p>\n<\/td>\n

          		\n

          \u8f93\u5165\u6210\u672c<\/span><\/p>\n<\/td>\n

          		\n

          \u8f93\u51fa\u6210\u672c<\/span><\/p>\n<\/td>\n

          		\n

          \u6708\u5408\u8ba1<\/span><\/p>\n<\/td>\n<\/tr>\n

          \n

          0%<\/span><\/p>\n<\/td>\n

          		\n

          100.0<\/span><\/p>\n<\/td>\n

          		\n

          80.0<\/span><\/p>\n<\/td>\n

          		\n

          180.0<\/span><\/p>\n<\/td>\n<\/tr>\n

          \n

          40%<\/span><\/p>\n<\/td>\n

          		\n

          60.8<\/span><\/p>\n<\/td>\n

          		\n

          80.0<\/span><\/p>\n<\/td>\n

          		\n

          140.8<\/span><\/p>\n<\/td>\n<\/tr>\n

          \n

          80%<\/span><\/p>\n<\/td>\n

          		\n

          21.6<\/span><\/p>\n<\/td>\n

          		\n

          80.0<\/span><\/p>\n<\/td>\n

          		\n

          101.6<\/span><\/p>\n<\/td>\n<\/tr>\n

          \n

          95%<\/span><\/p>\n<\/td>\n

          		\n

          6.9<\/span><\/p>\n<\/td>\n

          		\n

          80.0<\/span><\/p>\n<\/td>\n

          		\n

          86.9<\/span><\/p>\n<\/td>\n<\/tr>\n

          \n

          100%<\/span><\/p>\n<\/td>\n

          		\n

          2.0<\/span><\/p>\n<\/td>\n

          		\n

          80.0<\/span><\/p>\n<\/td>\n

          		\n

          82.0<\/span><\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

          \u547d\u4e2d\u7387\u4ece<\/span> 0% <\/span>\u62c9\u5230<\/span> 100%<\/span>，\u6708\u6210\u672c\u4ece<\/span> 180 <\/span>\u5143\u964d\u5230<\/span> 82 <\/span>\u5143\u3002<\/span>\u4f46\u6ce8\u610f：\u8f93\u51fa\u90a3\u5757 80 \u5143\u4e00\u52a8\u4e0d\u52a8\u3002<\/span><\/p>\n

          \u00a0\u00a0**<\/span>\u5728<\/span> prompt <\/span>\u91cc\u62a0\u5b57，\u7701\u7684\u662f<\/span>"<\/span>\u672a\u547d\u4e2d<\/span>"<\/span>\u90a3\u4e00\u5757；\u628a\u524d\u7f00\u7a33\u4f4f\u3001\u8ba9\u7f13\u5b58\u80fd\u547d\u4e2d，\u7701\u7684\u662f\u6574\u5757\u3002\u524d\u8005\u662f\u7701\u94b1，\u540e\u8005\u662f\u6539\u547d\u3002<\/span><\/p>\n

          \u90a3\u5177\u4f53\u600e\u4e48\u505a？\u6838\u5fc3\u53ea\u6709\u4e00\u53e5\u8bdd：<\/span>\u628a"\u4f1a\u91cd\u590d\u7684\u90e8\u5206"\u53d8\u6210"\u7a33\u5b9a\u7684\u524d\u7f00"\u3002<\/span> \u51e0\u6761\u80fd\u76f4\u63a5\u843d\u5730\u7684：<\/span><\/p>\n

        • System Prompt \u91cc\u653e\u6613\u53d8\u5185\u5bb9，\u662f\u7f13\u5b58\u7b2c\u4e00\u6740\u624b\u3002<\/span> \u65f6\u95f4\u6233\u3001\u7528\u6237<\/span> ID<\/span>\u3001\u968f\u673a\u6570\u3001\u5f53\u524d\u4f59\u989d<\/span>\u2014\u2014<\/span>\u8fd9\u7c7b\u6bcf\u6b21\u90fd\u53d8\u7684\u4e1c\u897f\u4e00\u5f8b\u632a\u5230\u6d88\u606f\u672b\u5c3e\u3002\u524d\u7f00\u91cc\u53ea\u8981\u6709\u4e00\u4e2a\u5b57\u7b26\u53d8\u4e86，\u540e\u9762\u6574\u6bb5\u7f13\u5b58\u5168\u90e8\u5931\u6548\u3002<\/span><\/li>\n
        • \u5de5\u5177\u5b9a\u4e49\u7684\u5b57\u6bb5\u987a\u5e8f\u8981\u56fa\u5b9a\u3002<\/span> \u5982\u679c\u6bcf\u6b21\u5e8f\u5217\u5316\u90fd\u968f\u673a\u6392\u5e8f，\u7f13\u5b58\u5c31\u6c38\u8fdc\u547d\u4e2d\u4e0d\u4e86\u3002<\/span><\/li>\n
        • \u62fc\u63a5\u987a\u5e8f：\u5148\u7a33\u5b9a，\u540e\u6613\u53d8\u3002<\/span> \u89c4\u8303\u3001\u63a5\u53e3\u5b9a\u4e49\u3001\u9886\u57df\u77e5\u8bc6\u653e\u524d\u9762；\u5f53\u524d\u4efb\u52a1\u63cf\u8ff0\u3001\u4e34\u65f6\u6570\u636e\u653e\u540e\u9762\u3002<\/span><\/li>\n
        • \u957f\u4f1a\u8bdd\u4fdd\u6301\u540c\u4e00\u6761\u4f1a\u8bdd\u94fe<\/span>，\u4e0d\u8981\u6bcf\u8f6e\u65b0\u5efa<\/span> session<\/span>\u3002<\/span><\/li>\n
        • \u56fa\u5b9a\u6a21\u677f\u63aa\u8f9e\u3002<\/span> "<\/span>\u8bf7\u56de\u7b54<\/span>"<\/span>\u548c<\/span>"<\/span>\u8bf7\u4f5c\u7b54<\/span>"<\/span>\u8bed\u4e49\u76f8\u540c，\u4f46\u5bf9\u7f13\u5b58\u6765\u8bf4\u662f\u4e24\u4e2a\u5b8c\u5168\u4e0d\u540c\u7684\u524d\u7f00\u3002<\/span><\/li>\n

          \u8fd9\u4e94\u6761\u7684\u5171\u540c\u70b9：<\/span>\u90fd\u4e0d\u9700\u8981\u6539\u6a21\u578b，\u53ea\u9700\u8981\u6539\u4f60\u62fc prompt \u7684\u65b9\u5f0f\u3002<\/span> \u800c\u5b83\u4eec\u5bf9\u8d26\u5355\u7684\u5f71\u54cd，\u6bd4\u6362\u6a21\u578b\u66f4\u5927\u3002<\/span>**<\/span><\/p>\n

          4. 890 bytes\/token：\u4e00\u6761\u56db\u4ee3\u66f2\u7ebf<\/span><\/h3>\n

          \u7f13\u5b58\u4e3a\u4ec0\u4e48\u80fd\u4fbf\u5b9c？\u56e0\u4e3a<\/span> DeepSeek <\/span>\u82b1\u4e86\u56db\u4ee3\u65f6\u95f4，\u628a\u6a21\u578b<\/span>"<\/span>\u8bb0\u4e8b<\/span>"<\/span>\u7684\u5360\u7528\u538b\u4e86\u4e0b\u53bb\u3002<\/span><\/p>\n

          KV cache <\/span>\u662f\u4ec0\u4e48？\u4e00\u53e5\u8bdd：<\/span>\u6a21\u578b\u6bcf\u591a\u8bfb\u4e00\u4e2a token，\u5c31\u8981\u5728\u663e\u5b58\u91cc\u591a\u8bb0\u4e00\u4efd"\u8fd9\u4e2a token \u662f\u4ec0\u4e48\u610f\u601d"\u7684\u7b14\u8bb0\u3002<\/span> \u4e0a\u4e0b\u6587\u8d8a\u957f，\u7b14\u8bb0\u8d8a\u539a\u3002\u8fd9\u662f\u957f\u4e0a\u4e0b\u6587\u63a8\u7406\u7684\u7b2c\u4e00\u5927\u6210\u672c\u3002<\/span><\/p>\n\n\n\n\n\n\n\n
          \n

          \u6a21\u578b<\/span><\/p>\n<\/td>\n

          		\n

          \u5168\u5c40<\/span> KV cache<\/span> （\u5b57\u8282<\/span> \/token<\/span> ）<\/span><\/p>\n<\/td>\n

          		\n

          \u76f8\u5bf9<\/span> V4.1-Flash<\/span><\/p>\n<\/td>\n

          		\n

          \u6bd4\u4e0a\u4e00\u4ee3\u7f29\u5c0f<\/span><\/p>\n<\/td>\n<\/tr>\n

          \n

          DeepSeek-V1<\/span><\/p>\n<\/td>\n

          		\n

          389,120<\/span><\/p>\n<\/td>\n

          		\n

          437.2x<\/span><\/p>\n<\/td>\n

          		\n

          \u2014<\/span><\/p>\n<\/td>\n<\/tr>\n

          \n

          DeepSeek-V3.2<\/span><\/p>\n<\/td>\n

          		\n

          48,068<\/span><\/p>\n<\/td>\n

          		\n

          54.0x<\/span><\/p>\n<\/td>\n

          		\n

          8.10x<\/span><\/p>\n<\/td>\n<\/tr>\n

          \n

          DeepSeek-V4-Flash<\/span><\/p>\n<\/td>\n

          		\n

          3,514<\/span><\/p>\n<\/td>\n

          		\n

          3.9x<\/span><\/p>\n<\/td>\n

          		\n

          13.68x<\/span><\/p>\n<\/td>\n<\/tr>\n

          \n

          DeepSeek-V4.1-Flash<\/span><\/p>\n<\/td>\n

          		\n

          890<\/span><\/p>\n<\/td>\n

          		\n

          1.0x<\/span><\/p>\n<\/td>\n

          		\n

          3.95x<\/span><\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

          \"\"KV cache \u56db\u4ee3\u66f2\u7ebf：<\/span>890 <\/span>\u5b57\u8282<\/span>\/token<\/span>，\u4ee5\u53ca\u4e24\u4e2a\u300c\u964d\u4e86\u591a\u5c11\u300d<\/span><\/p>\n

          \u6ce8\u610f\u8fd9\u91cc\u6709\u4e2a\u53e3\u5f84\u5751，\u4e2d\u6587\u62a5\u9053\u91cc\u975e\u5e38\u666e\u904d\u3002<\/span><\/p>\n

          \u5f88\u591a\u4e2d\u6587\u8d44\u8baf\u5199\u300c<\/span>V4.1-Flash <\/span>\u7684<\/span> KV Cache <\/span>\u538b\u7f29\u81f3\u539f\u6765\u7684<\/span> 1\/437<\/span>\u300d\u3002\u8fd9\u4e2a\u6570\u5b57\u662f\u5bf9\u7684，\u4f46<\/span>\u5bf9\u7167\u7684\u662f\u56db\u4ee3\u4e4b\u524d\u7684 DeepSeek-V1<\/span>\u2014\u2014<\/span>\u62ec\u53f7\u91cc\u90a3\u53e5<\/span>"<\/span>\u539f\u6765\u7684<\/span>"<\/span>\u5f88\u5bb9\u6613\u8ba9\u4eba\u4ee5\u4e3a\u8bf4\u7684\u662f\u4e0a\u4e00\u4ee3\u3002<\/span><\/p>\n

          \u5b9e\u9645\u60c5\u51b5\u662f：<\/span><\/p>\n

            \n
          • \u76f8\u5bf9 <\/span>V1<\/span>：<\/span>437 <\/span>\u500d<\/span> \u2705<\/span>（\u771f\u5b9e\u6570\u5b57）<\/span><\/li>\n
          • \u76f8\u5bf9 <\/span>\u4e0a\u4e00\u4ee3 V4-Flash<\/span>：<\/span>\u53ea\u6709 3.95 \u500d<\/span><\/li>\n<\/ul>\n\n\n\n
            \n

            1\/437 \u548c 1\/4，\u5dee\u6574\u6574\u4e24\u4e2a\u91cf\u7ea7\u7684\u8bed\u4e49\u3002<\/span>\u524d\u8005\u542c\u8d77\u6765\u50cf<\/span>"<\/span>\u8de8\u65f6\u4ee3\u7684\u9769\u547d<\/span>"<\/span>，\u540e\u8005\u542c\u8d77\u6765\u50cf<\/span>"<\/span>\u4e00\u6b21\u6b63\u5e38\u7684\u8fed\u4ee3<\/span>"\u2014\u2014<\/span>\u800c\u4e24\u8005\u8bf4\u7684\u662f\u540c\u4e00\u4e2a\u6a21\u578b\u3002\u770b\u5230\u8fd9\u79cd\u6570\u5b57，\u7b2c\u4e00\u4ef6\u4e8b\u662f\u95ee\u4e00\u53e5：<\/span>\u5206\u6bcd\u662f\u8c01？<\/span><\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

            \u987a\u5e26\u7b97\u4e2a\u5177\u4f53\u7684：\u6309<\/span> 100 <\/span>\u4e07<\/span> token <\/span>\u4e0a\u4e0b\u6587\u4f30\u7b97\u5168\u5c40\u7f13\u5b58\u5360\u7528（\u672c\u673a\u63a8\u7b97）：<\/span><\/p>\n\n\n\n\n\n
            \n

            \n<\/td>\n

            		\n

            1M <\/span> \u4e0a\u4e0b\u6587\u7684<\/span> KV <\/span> \u5360\u7528<\/span><\/p>\n<\/td>\n<\/tr>\n

            \n

            V4.1-Flash<\/span><\/p>\n<\/td>\n

            		\n

            0.87 GB<\/span><\/p>\n<\/td>\n<\/tr>\n

            \n

            V4-Flash<\/span><\/p>\n<\/td>\n

            		\n

            3.43 GB<\/span><\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

            0.87 GB \u662f\u4ec0\u4e48\u6982\u5ff5？<\/span> \u4e00\u5f20<\/span> 40GB <\/span>\u7ea7\u522b\u7684\u5361，\u6263\u6389\u6743\u91cd\u548c\u6fc0\u6d3b\u4e4b\u540e，\u8fd8\u88c5\u5f97\u4e0b\u767e\u4e07<\/span> token <\/span>\u7684\u7f13\u5b58<\/span>\u2014\u2014<\/span>\u8fd9\u5728\u4e0a\u4e00\u4ee3\u662f\u5f88\u96be\u60f3\u8c61\u7684\u3002\u800c\u8fd9\u6b63\u662f\u300c<\/span>890 bytes\/token<\/span>\u300d\u8fd9\u4e2a\u6570\u5b57\u7684\u5de5\u7a0b\u610f\u4e49：<\/span>\u5b83\u51b3\u5b9a\u4e86\u957f\u4e0a\u4e0b\u6587\u80fd\u4e0d\u80fd\u8dd1\u5728\u4fbf\u5b9c\u786c\u4ef6\u4e0a\u3002<\/span><\/p>\n

            5. \u67b6\u6784\u4e00：CED \u4e0d\u5bf9\u79f0\u2014\u2014\u8bfb\u7528 8B，\u5199\u7528 16B<\/span><\/h3>\n

            \u73b0\u5728\u8fdb\u5165\u67b6\u6784\u3002\u7b2c\u4e00\u4e2a，\u4e5f\u662f\u6700\u6838\u5fc3\u7684\u4e00\u4e2a\u3002<\/span><\/p>\n

            V4.1-Flash <\/span>\u7528\u7684\u662f<\/span> Causal Encoder-Decoder（CED，\u56e0\u679c\u7f16\u7801\u5668-\u89e3\u7801\u5668）<\/span>\u3002<\/span>40 <\/span>\u5c42<\/span> Transformer <\/span>\u88ab\u5207\u6210\u4e24\u534a：<\/span><\/p>\n

              \n
            • 20 \u5c42\u7f16\u7801\u5668（encoder）<\/span>：\u8d1f\u8d23<\/span>"<\/span>\u8bfb<\/span>"<\/span><\/li>\n
            • 20 \u5c42\u89e3\u7801\u5668（decoder）<\/span>：\u8d1f\u8d23<\/span>"<\/span>\u5199<\/span>"<\/span><\/li>\n<\/ul>\n

              \u5173\u952e\u5728\u4e8e\u4e24\u8fb9<\/span>\u6fc0\u6d3b\u7684\u53c2\u6570\u4e0d\u4e00\u6837<\/span>：<\/span><\/p>\n\n\n\n\n
              \n

              \u793a\u610f\u56fe<\/span> \u00b7 <\/span> \u6587\u672c\u7ed8\u5236<\/span><\/p>\n<\/td>\n<\/tr>\n

              \n

              \u00a0\u00a0CED <\/span>\u4e0d\u5bf9\u79f0\u67b6\u6784<\/span>: "<\/span>\u8bfb<\/span>" <\/span>\u548c<\/span> "<\/span>\u5199<\/span>" <\/span>\u62c6\u5f00\u8d70\u4e0d\u540c\u7684\u5c42<\/span><\/p>\n

              \u200b<\/span><\/p>\n

              \u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u8f93\u5165<\/span> prompt <\/span>\u8f93\u51fa<\/span> generation<\/span><\/p>\n

              \u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0(<\/span>\u8f93\u5165\u91cf\u662f\u8f93\u51fa\u7684<\/span> 389 <\/span>\u500d<\/span>) (<\/span>\u771f\u6b63\u8981\u751f\u6210\u7684<\/span>)<\/span><\/p>\n

              \u200b<\/span><\/p>\n

              \u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0| ^<\/span><\/p>\n

              \u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0v |<\/span><\/p>\n

              \u00a0\u00a0+———————+ +———————+<\/span><\/p>\n

              \u00a0\u00a0| 20 encoder layers | —> | 20 decoder layers |<\/span><\/p>\n

              \u00a0\u00a0| active = 8B | | active = 16B |<\/span><\/p>\n

              \u00a0\u00a0+———————+ +———————+<\/span><\/p>\n

              \u00a0\u00a0\u00a0\u00a0\u00a0"<\/span>\u8bfb<\/span>" <\/span>\u53ea\u8d70\u8fd9\u4e00\u534a<\/span> "<\/span>\u5199<\/span>" <\/span>\u624d\u7528\u8fd9\u4e00\u534a<\/span><\/p>\n

              \u00a0\u00a0\u00a0\u00a0\u00a0prompt <\/span>\u4e0d\u7ecf\u8fc7<\/span> decoder<\/span><\/p>\n

              \u200b<\/span><\/p>\n

              \u00a0\u00a0\u9aa8\u67b6<\/span> 552B, <\/span>\u4f46\u6bcf\u4e2a<\/span> token <\/span>\u53ea\u6fc0\u6d3b\u5176\u4e2d\u4e00\u5c0f\u90e8\u5206<\/span><\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

              \u90a3\u4e48，\u8fd9\u4ef6\u4e8b\u7701\u5728\u54ea\u91cc？<\/span><\/p>\n

              prompt \u9636\u6bb5\u7684 token，\u6839\u672c\u4e0d\u7ecf\u8fc7 decoder\u3002<\/span> decoder <\/span>\u7684\u5168\u5c40<\/span> KV <\/span>\u4e0d\u662f\u4e00\u5c42\u5c42\u7b97\u51fa\u6765\u7684，\u800c\u662f\u4ece<\/span> encoder <\/span>\u7684\u6700\u7ec8\u72b6\u6001<\/span>\u6295\u5f71<\/span>\u51fa\u6765\u7684（\u6bcf\u5c42\u6709\u81ea\u5df1\u7684\u6295\u5f71\u6743\u91cd）\u3002\u8fd9\u5e26\u6765\u4e24\u4e2a\u76f4\u63a5\u7ed3\u679c：<\/span><\/p>\n

            • prefill \u8ba1\u7b97\u91cf\u5927\u7ea6\u51cf\u534a<\/span>\u2014\u2014<\/span>\u56e0\u4e3a\u53ea\u8d70\u4e86\u4e00\u534a\u7684\u5c42\u3002<\/span><\/li>\n
            • \u8f93\u5165\u4fa7\u6bcf token \u53ea\u6fc0\u6d3b 8B \u53c2\u6570<\/span>，\u53ea\u6709\u751f\u6210\u65f6\u624d\u7528\u4e0a<\/span> 16B<\/span>\u3002<\/span><\/li>\n\n\n\n\n
              \n

              \u91cd\u70b9\u63d0\u793a<\/span><\/p>\n<\/td>\n<\/tr>\n

              \n

              \u8fd9\u4e2a\u8bbe\u8ba1\u5bf9\u5e94\u4e00\u4e2a\u5f88\u5177\u4f53\u7684\u73b0\u5b9e：Agent \u7684\u8f93\u5165\u548c\u8f93\u51fa，\u91cf\u7ea7\u5b8c\u5168\u4e0d\u5bf9\u7b49\u3002<\/span> \u524d\u9762\u90a3\u7ec4\u6570\u636e\u91cc，\u8f93\u5165（\u542b\u91cd\u8bfb）<\/span>1,946M<\/span>，\u8f93\u51fa<\/span> 5M\u2014\u2014<\/span>\u8f93\u5165\u662f\u8f93\u51fa\u7684 389 \u500d<\/span>\u3002<\/span><\/p>\n

              \u4e00\u4e2a\u4e3a\u804a\u5929\u8bbe\u8ba1\u7684\u6a21\u578b，\u8f93\u5165\u8f93\u51fa\u5927\u81f4\u76f8\u5f53，\u5bf9\u79f0\u67b6\u6784\u662f\u5408\u7406\u7684\u3002\u4f46\u4e00\u4e2a\u4e3a<\/span> Agent <\/span>\u8bbe\u8ba1\u7684\u6a21\u578b，\u5982\u679c\u8fd8\u8ba9<\/span> 16B <\/span>\u7684\u53c2\u6570\u53bb\u5904\u7406\u90a3<\/span> 389 <\/span>\u500d\u7684\u91cd\u8bfb，\u5c31\u662f\u628a\u94b1\u82b1\u5728\u6700\u4e0d\u9700\u8981\u7684\u5730\u65b9\u3002<\/span><\/p>\n

              CED \u7684"\u4e0d\u5bf9\u79f0"，\u8bf4\u767d\u4e86\u5c31\u662f：\u8bfb\u6587\u4ef6\u7528\u8f7b\u4fbf\u7684\u9605\u8bfb\u5668，\u5199\u62a5\u544a\u624d\u53eb\u4e0a\u4e3b\u7b14\u3002<\/span><\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

              \u6548\u679c：<\/span>\u5f53\u4e0a\u4e0b\u6587\u4ece 4K \u6da8\u5230 1M，\u5355 token \u7684 decode FLOPs \u53ea\u589e\u52a0\u7ea6 1.25 \u500d<\/span>（\u5b98\u65b9\u6280\u672f\u62a5\u544a\u53e3\u5f84）\u3002<\/span><\/p>\n

              \u00a0\u00a0\u6362\u6210\u4eba\u8bdd：<\/span>\u4ece\u8bfb\u4e00\u9875\u7eb8\u5230\u8bfb\u4e00\u672c\u4e66，\u6bcf\u5199\u4e00\u4e2a\u5b57\u7684\u4ee3\u4ef7\u53ea\u591a 25%\u3002<\/span> \u800c\u4e0a\u4e0b\u6587\u957f\u5ea6\u6da8\u4e86<\/span> 250 <\/span>\u500d\u3002<\/span><\/p>\n

              6. \u67b6\u6784\u4e8c：CSA2 \u7684\u4e09\u79cd\u6ce8\u610f\u529b\u6a21\u5f0f<\/span><\/h3>\n

              \u7b2c\u4e8c\u4e2a\u624b\u6bb5，\u89e3\u51b3\u7684\u662f<\/span>"<\/span>\u90a3<\/span> 890 <\/span>\u5b57\u8282\u8fd8\u80fd\u4e0d\u80fd\u518d\u538b<\/span>"<\/span>\u3002<\/span><\/p>\n

              \u5148\u8bf4\u80cc\u666f：\u4f20\u7edf\u6ce8\u610f\u529b\u91cc，<\/span>\u6bcf\u4e00\u5c42\u90fd\u81ea\u5df1\u7b97\u4e00\u4efd KV cache<\/span>\u3002<\/span>40 <\/span>\u5c42\u5c31\u662f<\/span> 40 <\/span>\u4efd，\u5c42\u4e0e\u5c42\u4e4b\u95f4\u660e\u660e\u5728\u770b\u540c\u4e00\u6bb5\u6587\u672c，\u5374\u5404\u8bb0\u5404\u7684\u7b14\u8bb0\u3002<\/span><\/p>\n

              CSA2（Compressed Sparse Attention 2，\u7b2c\u4e8c\u4ee3\u538b\u7f29\u7a00\u758f\u6ce8\u610f\u529b）<\/span> \u5e72\u7684\u4e8b\u5c31\u662f：<\/span>\u8ba9\u5c42\u4e4b\u95f4\u53ef\u4ee5\u6284\u7b14\u8bb0\u3002<\/span><\/p>\n

              \u5b83\u7ed9\u6bcf\u4e00\u5c42\u9759\u6001\u6307\u6d3e\u4e09\u79cd\u8eab\u4efd\u4e4b\u4e00：<\/span><\/p>\n\n\n\n\n
              \n

              \u793a\u610f\u56fe<\/span> \u00b7 <\/span> \u6587\u672c\u7ed8\u5236<\/span><\/p>\n<\/td>\n<\/tr>\n

              \n

              \u00a0\u00a040 <\/span>\u5c42\u4e0d\u518d\u5404\u8bb0\u4e00\u4efd\u7b14\u8bb0<\/span>, <\/span>\u800c\u662f\u5c42\u4e0e\u5c42\u4e4b\u95f4<\/span>"<\/span>\u6284\u4f5c\u4e1a<\/span>"<\/span><\/p>\n

              \u200b<\/span><\/p>\n

              \u00a0\u00a0Full <\/span>\u81ea\u5df1\u7b97<\/span> KV, <\/span>\u81ea\u5df1\u9009<\/span> Top-512 <\/span>\u7d22\u5f15<\/span><\/p>\n

              \u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u7ec4\u5185\u7684<\/span>"<\/span>\u6e90\u5934\u5c42<\/span>", <\/span>\u6bcf\u7ec4\u53ea\u6709\u4e00\u4e2a<\/span><\/p>\n

              \u200b<\/span><\/p>\n

              \u00a0\u00a0Reindex <\/span>\u590d\u7528\u4e0a\u4e00\u4e2a<\/span> Full <\/span>\u5c42\u7684<\/span> KV, <\/span>\u7528\u81ea\u5df1\u7684<\/span> Q <\/span>\u91cd\u65b0\u6253\u5206<\/span><\/p>\n

              \u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u4e0d\u91cd\u7b97\u4e3b<\/span> KV, <\/span>\u53ea\u91cd\u6392\u7d22\u5f15<\/span><\/p>\n

              \u200b<\/span><\/p>\n

              \u00a0\u00a0Reuse <\/span>\u76f4\u63a5\u590d\u7528\u6700\u8fd1\u7684<\/span> Top-K <\/span>\u7d22\u5f15<\/span>, <\/span>\u8fde\u7d22\u5f15\u5668\u90fd\u8df3\u8fc7<\/span><\/p>\n

              \u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u6700\u7701<\/span>, <\/span>\u7ec4\u5185\u5360\u5927\u591a\u6570<\/span><\/p>\n

              \u200b<\/span><\/p>\n

              \u00a0\u00a0———————————————————-<\/span><\/p>\n

              \u00a0\u00a0encoder <\/span>\u4fa7<\/span> 18 <\/span>\u5c42<\/span> ratio 2-in-3 3 <\/span>\u7ec4<\/span> x 6 <\/span>\u5c42<\/span><\/p>\n

              \u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u6bcf\u7ec4<\/span> = 1 Full + 5 Reuse<\/span><\/p>\n

              \u200b<\/span><\/p>\n

              \u00a0\u00a0decoder <\/span>\u4fa7<\/span> 20 <\/span>\u5c42<\/span> ratio 1-in-4 5 <\/span>\u7ec4<\/span> x 4 <\/span>\u5c42<\/span><\/p>\n

              \u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u7b2c<\/span> 1 <\/span>\u7ec4<\/span> = 1 Full + 3 Reuse<\/span><\/p>\n

              \u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u5176\u4f59<\/span> = 1 Reindex + 3 Reuse<\/span><\/p>\n

              \u200b<\/span><\/p>\n

              \u00a0\u00a0\u5c42\u7ea7\u7a00\u758f\u7d22\u5f15\u5668<\/span>: <\/span>\u5019\u9009\u6c60\u6700\u591a<\/span> 16,384 <\/span>\u4e2a\u4f4d\u7f6e<\/span><\/p>\n

              \u00a0\u00a0-> <\/span>\u627e\u7d22\u5f15\u7684\u4ee3\u4ef7\u4e0d\u518d\u968f\u4e0a\u4e0b\u6587\u957f\u5ea6\u589e\u957f<\/span><\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

              \u8bfb\u6cd5<\/span>：<\/span>40 <\/span>\u5c42\u91cc\u771f\u6b63<\/span>"<\/span>\u4ece\u5934\u7b97\u8bb0\u5fc6<\/span>"<\/span>\u7684\u53ea\u6709\u5c11\u6570\u51e0\u5c42，\u5176\u4f59\u90fd\u9760\u590d\u7528\u3002\u8fd9\u5c31\u662f<\/span> 890 bytes\/token <\/span>\u91cc<\/span>"4 <\/span>\u500d\u538b\u7f29<\/span>"<\/span>\u7684\u6765\u6e90<\/span>\u2014\u2014<\/span>\u4e0d\u662f\u628a\u7b14\u8bb0\u5199\u5f97\u66f4\u77ed，\u662f\u8ba9\u66f4\u591a\u5c42\u5171\u7528\u540c\u4e00\u4efd\u7b14\u8bb0\u3002<\/span><\/p>\n

              \u53e6\u5916\u8fd8\u6709\u4e00\u4e2a<\/span>\u5c42\u7ea7\u7a00\u758f\u7d22\u5f15\u5668<\/span>：\u7b2c\u4e00\u4e2a<\/span> Full <\/span>\u5c42\u4f1a\u5efa\u4e00\u4e2a\u6700\u591a<\/span> 16,384 \u4e2a\u4f4d\u7f6e<\/span>（<\/span>2,048 <\/span>\u5757<\/span> \u00d7 8<\/span>）\u7684\u5019\u9009\u6c60，\u540e\u7eed<\/span> Reindex <\/span>\u5c42\u5728\u8fd9\u4e2a\u6709\u754c\u96c6\u5408\u91cc\u6253\u5206，\u800c\u4e0d\u662f\u5728\u5168\u90e8\u4e0a\u4e0b\u6587\u91cc\u6253\u5206\u3002<\/span><\/p>\n

              \u00a0\u00a0\u8fd9\u8ba9"\u627e\u76f8\u5173\u7247\u6bb5"\u7684\u4ee3\u4ef7\u4e0d\u518d\u968f\u4e0a\u4e0b\u6587\u957f\u5ea6\u589e\u957f\u3002<\/span> \u8bfb<\/span> 1 <\/span>\u4e07<\/span> token <\/span>\u548c\u8bfb<\/span> 100 <\/span>\u4e07<\/span> token<\/span>，\u627e\u7d22\u5f15\u7684\u5f00\u9500\u662f\u4e00\u6837\u7684<\/span>\u2014\u2014<\/span>\u8fd9\u662f\u5b83\u80fd\u625b\u4f4f<\/span> 1M <\/span>\u7684\u53e6\u4e00\u4e2a\u524d\u63d0\u3002<\/span><\/p>\n

              \u00a0\u00a0\u7528\u4e00\u4e2a\u6bd4\u55bb\u6536\u4e00\u4e0b\u8fd9\u4e24\u8282：<\/span> CED <\/span>\u89e3\u51b3\u7684\u662f<\/span>"<\/span>\u8c01\u6765\u8bfb<\/span>"<\/span>，<\/span>CSA2 <\/span>\u89e3\u51b3\u7684\u662f<\/span>"<\/span>\u8bb0\u591a\u5c11<\/span>"<\/span>\u3002\u524d\u8005\u8ba9\u4fbf\u5b9c\u7684\u53c2\u6570\u53bb\u5e72\u91cd\u8bfb\u7684\u6d3b，\u540e\u8005\u8ba9<\/span> 40 <\/span>\u5c42\u5171\u4eab\u4e00\u4efd\u7b14\u8bb0\u800c\u4e0d\u662f\u5404\u8bb0<\/span> 40 <\/span>\u4efd\u3002<\/span><\/p>\n

              7. \u67b6\u6784\u4e09：FP4 \u91cf\u5316\u4e0e\u7f13\u5b58\u7684\u751f\u547d\u5468\u671f<\/span><\/h3>\n

              \u7b2c\u4e09\u4e2a\u624b\u6bb5，\u662f\u628a\u6bcf\u4e00\u4efd\u7b14\u8bb0\u672c\u8eab\u5199\u5f97\u66f4\u5c0f，\u5e76\u4e14\u7ba1\u597d\u5b83\u7684<\/span>"<\/span>\u4fdd\u8d28\u671f<\/span>"<\/span>\u3002<\/span><\/p>\n

              \u2460 FP4 \u91cf\u5316\u3002<\/span> \u4e3b<\/span> KV cache <\/span>\u7528<\/span> E2M1<\/span> \u683c\u5f0f\u5b58\u50a8，\u6bcf<\/span> 16 <\/span>\u4e2a\u901a\u9053\u914d\u4e00\u4e2a<\/span> E4M3<\/span> \u7f29\u653e\u56e0\u5b50（\u5373<\/span> NVFP4 <\/span>\u65b9\u6848\u53bb\u6389\u5168\u5c40<\/span> scale <\/span>\u7684\u505a\u6cd5）\u3002\u8fd9\u5957\u91cf\u5316\u662f\u5728\u540e\u8bad\u7ec3\u9636\u6bb5\u901a\u8fc7\u91cf\u5316\u611f\u77e5\u8bad\u7ec3\u5f15\u5165\u7684，<\/span>\u76f8\u6bd4 V4 \u7684 FP8 \u5927\u7ea6\u518d\u7701\u4e00\u534a\u7a7a\u95f4\u3002<\/span><\/p>\n

              \u2461 SWA Bounded Replay（\u6ed1\u52a8\u7a97\u53e3\u6ce8\u610f\u529b\u7684\u6709\u754c\u91cd\u653e）\u3002<\/span> decoder <\/span>\u7684\u6ed1\u52a8\u7a97\u53e3\u72b6\u6001\u4e0d\u91cd\u8dd1\u5168\u90e8\u5c42，<\/span>\u53ea\u91cd\u653e\u6700\u540e 128 \u4e2a prompt token<\/span> \u5c31\u80fd\u91cd\u5efa\u3002<\/span><\/p>\n

              \u2462 \u7f13\u5b58\u7684\u751f\u547d\u5468\u671f\u7ba1\u7406\u3002<\/span> \u8fd9\u662f\u6700\u5bb9\u6613\u88ab\u5ffd\u7565\u3001\u4f46\u5bf9\u751f\u4ea7\u90e8\u7f72\u6700\u5173\u952e\u7684\u4e00\u6761：<\/span><\/p>\n\n\n\n\n\n
              \n

              \u7f13\u5b58\u7c7b\u578b<\/span><\/p>\n<\/td>\n

              		\n

              \u5b58\u653e\u4f4d\u7f6e<\/span><\/p>\n<\/td>\n

              		\n

              \u751f\u547d\u5468\u671f<\/span><\/p>\n<\/td>\n<\/tr>\n

              \n

              \u5168\u5c40<\/span> KV<\/span><\/p>\n<\/td>\n

              		\n

              \u6301\u4e45\u7f13\u5b58<\/span><\/p>\n<\/td>\n

              		\n

              \u4fdd\u8bc1 \u2265 72 \u5c0f\u65f6<\/span><\/p>\n<\/td>\n<\/tr>\n

              \n

              SWA KV<\/span><\/p>\n<\/td>\n

              		\n

              \u4ece <\/span>10% \u4e3b\u673a DRAM<\/span> \u5212\u51fa\u7684\u5206\u5e03\u5f0f\u6c60， <\/span>\u4e0d\u843d SSD<\/span><\/p>\n<\/td>\n

              		\n

              \u5206\u949f\u7ea7<\/span><\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n\n\n\n
              \n

              \u8fd9\u5e26\u6765\u4e00\u4e2a\u53cd\u5e38\u8bc6\u7684\u7ed3\u8bba：\u771f\u6b63\u7684\u74f6\u9888，\u5df2\u7ecf\u4e0d\u662f\u663e\u5b58，\u800c\u662f"\u7f13\u5b58\u57fa\u7840\u8bbe\u65bd\u7b56\u7565"\u3002<\/span><\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

              \u60f3\u60f3\u8fd9\u610f\u5473\u7740\u4ec0\u4e48：\u4e00\u4e2a\u591a\u79df\u6237\u7684\u957f\u4e0a\u4e0b\u6587\u670d\u52a1，\u80fd\u4e0d\u80fd\u625b\u4f4f，\u53d6\u51b3\u4e8e\u4f60\u600e\u4e48\u8bbe\u8ba1<\/span> 72 <\/span>\u5c0f\u65f6\u548c\u5206\u949f\u7ea7\u8fd9\u4e24\u6863\u7684\u6dd8\u6c70\u7b56\u7565<\/span>\u2014\u2014<\/span>\u8fd9\u66f4\u63a5\u8fd1\u4e00\u4e2a\u5b58\u50a8\u7cfb\u7edf\u95ee\u9898，\u4e0d\u662f\u6a21\u578b\u95ee\u9898\u3002<\/span><\/p>\n

              \u00a0\u00a0\u6362\u4e2a\u8bf4\u6cd5：<\/span> \u524d\u4e24\u8282\u5728\u538b\u7f29<\/span>"<\/span>\u8bb0\u5fc6\u7684\u539a\u5ea6<\/span>"<\/span>，\u8fd9\u4e00\u8282\u5728\u7ba1\u7406<\/span>"<\/span>\u8bb0\u5fc6\u7684\u5bff\u547d<\/span>"<\/span>\u3002<\/span>\u8bb0\u5fc6\u8d8a\u4fbf\u5b9c，\u5c31\u8d8a\u6709\u8d44\u683c\u88ab\u7559\u4e0b\u6765；\u80fd\u7559\u4e0b\u6765，\u624d\u8c08\u5f97\u4e0a\u590d\u7528\u3002<\/span><\/p>\n

              8. \u4e09\u4e2a\u624b\u6bb5\u5408\u8d77\u6765\u7701\u4e86\u591a\u5c11<\/span><\/h3>\n

              \u628a\u4e09\u4e2a\u67b6\u6784\u624b\u6bb5\u548c\u5b83\u4eec\u7684\u6548\u679c\u644a\u5728\u4e00\u5f20\u8868\u4e0a：<\/span><\/p>\n\n\n\n\n\n\n\n\n
              \n

              \u624b\u6bb5<\/span><\/p>\n<\/td>\n

              		\n

              \u4f5c\u7528\u5bf9\u8c61<\/span><\/p>\n<\/td>\n

              		\n

              \u5b98\u65b9\u53e3\u5f84\u6548\u679c<\/span><\/p>\n<\/td>\n<\/tr>\n

              \n

              CED <\/span> \u4e0d\u5bf9\u79f0（<\/span> 8B <\/span> \u8bfb<\/span> \/ 16B <\/span> \u5199）<\/span><\/p>\n<\/td>\n

              		\n

              \u8ba1\u7b97\u91cf<\/span><\/p>\n<\/td>\n

              		\n

              prefill <\/span> \u7ea6\u51cf\u534a；<\/span> 4K\u21921M <\/span> \u5355<\/span> token decode FLOPs <\/span> \u53ea\u6da8<\/span> 1.25x<\/span><\/p>\n<\/td>\n<\/tr>\n

              \n

              CSA2<\/span> （<\/span> Full \/ Reindex \/ Reuse<\/span> ）<\/span><\/p>\n<\/td>\n

              		\n

              \u7f13\u5b58\u4f53\u79ef<\/span><\/p>\n<\/td>\n

              		\n

              \u5c42\u95f4\u590d\u7528，\u538b\u7f29\u6bd4<\/span> 2-in-3<\/span> （<\/span> encoder<\/span> ）<\/span> \/ 1-in-4<\/span> （<\/span> decoder<\/span> ）<\/span><\/p>\n<\/td>\n<\/tr>\n

              \n

              FP4 KV <\/span> \u91cf\u5316<\/span><\/p>\n<\/td>\n

              		\n

              \u7f13\u5b58\u4f53\u79ef<\/span><\/p>\n<\/td>\n

              		\n

              \u76f8\u6bd4<\/span> FP8 <\/span> \u7ea6\u518d\u7701\u4e00\u534a<\/span><\/p>\n<\/td>\n<\/tr>\n

              \n

              SWA Bounded Replay<\/span><\/p>\n<\/td>\n

              		\n

              \u6301\u4e45\u5316\u5360\u7528<\/span><\/p>\n<\/td>\n

              		\n

              SSD \/ <\/span> \u4e3b\u673a\u5185\u5b58\u5360\u7528\u964d\u81f3\u7ea6<\/span> 1\/8<\/span><\/p>\n<\/td>\n<\/tr>\n

              \n

              \u5408\u8ba1<\/span><\/p>\n<\/td>\n

              		\n

              \u5168\u5c40<\/span> KV cache<\/span><\/p>\n<\/td>\n

              		\n

              890 bytes\/token \u2248 \u4e0a\u4e00\u4ee3\u7684 1\/4<\/span><\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

              \u6ce8\u610f\u6700\u540e\u4e00\u884c\u548c\u7b2c 4 \u8282\u90a3\u6761\u66f2\u7ebf\u7684\u5173\u7cfb<\/span>：<\/span><\/p>\n

                \n
              • \u300c<\/span>1\/4<\/span>\u300d\u662f<\/span>\u8fd9\u4e00\u4ee3\u67b6\u6784<\/span>\u7684\u8d21\u732e（\u4e0a\u9762\u8fd9\u4e94\u4e2a\u624b\u6bb5）<\/span><\/li>\n
              • \u300c<\/span>1\/437<\/span>\u300d\u662f<\/span>\u56db\u4ee3\u7d2f\u79ef<\/span>\u7684\u8d21\u732e（\u56de\u770b\u7b2c<\/span> 4 <\/span>\u8282\u7684\u8868）<\/span><\/li>\n<\/ul>\n

                \u00a0\u00a0\u8fd9\u4e24\u4e2a\u6570\u5b57\u4e0d\u51b2\u7a81，\u5b83\u4eec\u56de\u7b54\u7684\u662f\u4e0d\u540c\u7684\u95ee\u9898\u3002<\/span> \u524d\u8005\u56de\u7b54<\/span>"<\/span>\u8fd9\u6b21\u6539\u4e86\u4ec0\u4e48<\/span>"<\/span>，\u540e\u8005\u56de\u7b54<\/span>"<\/span>\u8fd9\u6761\u8def\u8d70\u4e86\u591a\u8fdc<\/span>"<\/span>\u3002\u5199\u6587\u7ae0\u65f6\u5982\u679c\u53ea\u5f15\u4e00\u4e2a，\u8bfb\u8005\u5c31\u4f1a\u5f97\u5230\u4e00\u5f20\u5931\u771f\u7684\u56fe\u3002<\/span><\/p>\n

                \u800c\u8fd9\u4e00\u5207\u7684\u7ec8\u70b9，\u5c31\u662f\u4f60\u770b\u5230\u7684\u90a3\u5f20\u4ef7\u683c\u8868<\/span>\u2014\u2014<\/span>\u7f13\u5b58\u547d\u4e2d\u964d\u5230 0.003 \u7f8e\u5143\u3002<\/span> \u67b6\u6784\u4e0a\u7684<\/span> 890<\/span>，\u548c\u8d26\u5355\u4e0a\u7684<\/span> 0.003<\/span>，\u662f\u540c\u4e00\u4ef6\u4e8b\u7684\u4e24\u7aef\u3002<\/span><\/p>\n

                9. \u80fd\u529b\u8d26：\u5b83\u8d62\u4e86\u54ea\u91cc，\u53c8\u8f93\u5728\u54ea\u91cc<\/span><\/h3>\n

                \u73b0\u5728\u8bf4\u80fd\u529b\u3002\u5b98\u65b9\u5bf9\u6bd4\u8868（\u6700\u5927\u63a8\u7406\u52aa\u529b\u8bbe\u7f6e，\u5373<\/span> reasoning_effort=100<\/span>）\u7684\u51e0\u4e2a\u5173\u952e\u884c：<\/span><\/p>\n\n\n\n\n\n\n\n\n\n\n
                \n

                \u57fa\u51c6<\/span><\/p>\n<\/td>\n

                		\n

                V4.1-Flash<\/span><\/p>\n<\/td>\n

                		\n

                V4-Pro 0813<\/span><\/p>\n<\/td>\n

                		\n

                GLM-5.3<\/span><\/p>\n<\/td>\n

                		\n

                Kimi-K3<\/span><\/p>\n<\/td>\n

                		\n

                GPT-5.6 Sol<\/span><\/p>\n<\/td>\n

                		\n

                Claude Opus 5<\/span><\/p>\n<\/td>\n<\/tr>\n

                \n

                Terminal-Bench 2.1<\/span><\/p>\n<\/td>\n

                		\n

                90.6<\/span><\/p>\n<\/td>\n

                		\n

                87.9<\/span><\/p>\n<\/td>\n

                		\n

                88.2<\/span><\/p>\n<\/td>\n

                		\n

                88.3<\/span><\/p>\n<\/td>\n

                		\n

                88.8<\/span><\/p>\n<\/td>\n

                		\n

                89.1<\/span><\/p>\n<\/td>\n<\/tr>\n

                \n

                DeepSWE v1.1<\/span><\/p>\n<\/td>\n

                		\n

                74.2<\/span><\/p>\n<\/td>\n

                		\n

                62.7<\/span><\/p>\n<\/td>\n

                		\n

                66.9<\/span><\/p>\n<\/td>\n

                		\n

                67.5<\/span><\/p>\n<\/td>\n

                		\n

                73.0<\/span><\/p>\n<\/td>\n

                		\n

                74.0<\/span><\/p>\n<\/td>\n<\/tr>\n

                \n

                CyberGym<\/span><\/p>\n<\/td>\n

                		\n

                88.1<\/span><\/p>\n<\/td>\n

                		\n

                83.3<\/span><\/p>\n<\/td>\n

                		\n

                84.5<\/span><\/p>\n<\/td>\n

                		\n

                80.0<\/span><\/p>\n<\/td>\n

                		\n

                84.5<\/span><\/p>\n<\/td>\n

                		\n

                \u2014<\/span><\/p>\n<\/td>\n<\/tr>\n

                \n

                Automation-Bench<\/span><\/p>\n<\/td>\n

                		\n

                54.8<\/span><\/p>\n<\/td>\n

                		\n

                43.2<\/span><\/p>\n<\/td>\n

                		\n

                48.8<\/span><\/p>\n<\/td>\n

                		\n

                46.7<\/span><\/p>\n<\/td>\n

                		\n

                45.8<\/span><\/p>\n<\/td>\n

                		\n

                50.3<\/span><\/p>\n<\/td>\n<\/tr>\n

                \n

                Codeforces <\/span> \u8bc4\u5206<\/span><\/p>\n<\/td>\n

                		\n

                3471<\/span><\/p>\n<\/td>\n

                		\n

                3348<\/span><\/p>\n<\/td>\n

                		\n

                \u2014<\/span><\/p>\n<\/td>\n

                		\n

                \u2014<\/span><\/p>\n<\/td>\n

                		\n

                \u2014<\/span><\/p>\n<\/td>\n

                		\n

                \u2014<\/span><\/p>\n<\/td>\n<\/tr>\n

                \n

                Agents' Last Exam<\/span><\/p>\n<\/td>\n

                		\n

                31.8<\/span><\/p>\n<\/td>\n

                		\n

                25.7<\/span><\/p>\n<\/td>\n

                		\n

                28.5<\/span><\/p>\n<\/td>\n

                		\n

                27.6<\/span><\/p>\n<\/td>\n

                		\n

                26.7<\/span><\/p>\n<\/td>\n

                		\n

                28.6<\/span><\/p>\n<\/td>\n<\/tr>\n

                \n

                HLE<\/span> （\u5e26\u5de5\u5177）<\/span><\/p>\n<\/td>\n

                		\n

                63.9<\/span><\/p>\n<\/td>\n

                		\n

                60.0<\/span><\/p>\n<\/td>\n

                		\n

                62.5<\/span><\/p>\n<\/td>\n

                		\n

                59.8<\/span><\/p>\n<\/td>\n

                		\n

                \u2014<\/span><\/p>\n<\/td>\n

                		\n

                63.6<\/span><\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

                \u8d62\u7684\u5730\u65b9\u5f88\u96c6\u4e2d：Agent \u4e0e\u5de5\u7a0b\u7c7b\u4efb\u52a1\u3002<\/span> Terminal-Bench<\/span>\u3001<\/span>DeepSWE<\/span>\u3001<\/span>CyberGym<\/span>\u3001<\/span>Automation-Bench<\/span>\u3001<\/span>Codeforces\u2014\u2014<\/span>\u5b83\u57fa\u672c\u90fd\u62ff\u4e86\u8fd9\u5f20\u8868\u7684\u7b2c\u4e00\u3002<\/span><\/p>\n

                \u4f46\u8f93\u7684\u5730\u65b9，\u5b98\u65b9\u4e00\u70b9\u6ca1\u85cf：<\/span><\/p>\n\n\n\n\n\n\n\n
                \n

                \u57fa\u51c6<\/span><\/p>\n<\/td>\n

                		\n

                V4.1-Flash<\/span><\/p>\n<\/td>\n

                		\n

                V4-Pro 0813<\/span><\/p>\n<\/td>\n

                		\n

                Claude Opus 5<\/span><\/p>\n<\/td>\n<\/tr>\n

                \n

                GPQA Diamond<\/span><\/p>\n<\/td>\n

                		\n

                90.9<\/span><\/p>\n<\/td>\n

                		\n

                92.4<\/span><\/p>\n<\/td>\n

                		\n

                93.4<\/span><\/p>\n<\/td>\n<\/tr>\n

                \n

                HLE<\/span> （\u4e0d\u7528\u5de5\u5177）<\/span><\/p>\n<\/td>\n

                		\n

                36.8<\/span><\/p>\n<\/td>\n

                		\n

                42.7<\/span><\/p>\n<\/td>\n

                		\n

                56.3<\/span><\/p>\n<\/td>\n<\/tr>\n

                \n

                Terminal-Bench 3.0<\/span><\/p>\n<\/td>\n

                		\n

                30.0<\/span><\/p>\n<\/td>\n

                		\n

                11.8<\/span><\/p>\n<\/td>\n

                		\n

                43.3<\/span><\/p>\n<\/td>\n<\/tr>\n

                \n

                Terminal-Bench 4.0<\/span><\/p>\n<\/td>\n

                		\n

                31.2<\/span><\/p>\n<\/td>\n

                		\n

                12.4<\/span><\/p>\n<\/td>\n

                		\n

                51.8<\/span><\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n\n\n\n
                \n

                \u8fd9\u5f20\u8868\u8bf4\u6e05\u4e86\u4e00\u4ef6\u5f88\u91cd\u8981\u7684\u4e8b：\u5b83\u8d62\u7684\u662f"\u5f00\u5377"，\u4e0d\u662f"\u95ed\u5377"\u3002<\/span><\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

                \u600e\u4e48\u7406\u89e3？<\/span>\u5e26\u5de5\u5177\u7684 HLE \u662f 63.9，\u4e0d\u5e26\u5de5\u5177\u53ea\u6709 36.8\u3002<\/span> \u540c\u4e00\u4e2a\u6a21\u578b，\u540c\u4e00\u4e2a\u57fa\u51c6，\u5dee\u522b\u53ea\u5728\u4e8e<\/span>"<\/span>\u80fd\u4e0d\u80fd\u67e5\u8d44\u6599<\/span>"<\/span>\u3002<\/span><\/p>\n

                \u00a0\u00a0\u5b83\u662f\u4e00\u4e2a\u975e\u5e38\u4f1a"\u7528\u5de5\u5177\u529e\u4e8b"\u7684\u6a21\u578b，\u4e0d\u662f\u4e00\u4e2a\u65b0\u7684"\u63a8\u7406\u4e4b\u738b"\u3002<\/span> \u5982\u679c\u4f60\u8981\u62ff\u5b83\u505a\u95ed\u5377\u77e5\u8bc6\u63a8\u7406\u3001\u786c\u6838\u7406\u8bba\u63a8\u5bfc，\u5b83\u7684\u8868\u73b0\u4f1a\u4f4e\u4e8e\u4f60\u7684\u9884\u671f<\/span>\u2014\u2014<\/span>\u5c24\u5176\u5728<\/span> Terminal-Bench 3.0 \/ 4.0 <\/span>\u8fd9\u4e24\u4e2a\u66f4\u65b0\u3001\u66f4\u96be\u7684\u7248\u672c\u4e0a，\u548c<\/span> Opus 5 <\/span>\u7684\u5dee\u8ddd\u662f<\/span> 13 <\/span>\u5230<\/span> 20 <\/span>\u5206，<\/span>\u8fd9\u4e0d\u662f\u5c0f\u5dee\u8ddd，\u662f\u65ad\u5c42\u3002<\/span><\/p>\n

                \u5173\u4e8e\u300c\u8d62<\/span> 12\/14 <\/span>\u884c\u300d\u8fd9\u4e2a\u8bf4\u6cd5\u4e5f\u8981\u8bf4\u6e05\u695a：<\/span>\u4e24\u8fb9\u90fd\u6709\u5206\u6570\u7684 14 \u884c\u91cc\u5b83\u8d62 12 \u884c<\/span>，\u4e0d\u662f<\/span>"19 <\/span>\u9879\u6d4b\u8bc4\u91cc\u8d62<\/span> 14 <\/span>\u9879<\/span>"<\/span>\u3002\u5206\u6bcd\u4e0d\u540c，\u7ed3\u8bba\u5c31\u4e0d\u540c<\/span>\u2014\u2014<\/span>\u8fd9\u53c8\u662f\u4e00\u6b21\u53e3\u5f84\u68c0\u67e5\u3002<\/span><\/p>\n

                10. \u53e3\u5f84\u8b66\u544a：\u540c\u4e00\u4e2a\u6a21\u578b，\u6362\u4e2a\u5916\u58f3\u5dee 8.7 \u5206<\/span><\/h3>\n

                \u8fd9\u4e00\u8282\u662f\u672c\u7cfb\u5217\u7684\u8001\u670b\u53cb，\u4f46\u5728<\/span> V4.1-Flash <\/span>\u4e0a\u53c8\u51fa\u73b0\u4e86\u4e00\u6b21，\u800c\u4e14\u66f4\u660e\u663e\u3002<\/span><\/p>\n

                \u5b98\u65b9\u516c\u5e03\u4e86\u540c\u4e00\u4e2a\u6a21\u578b\u5728<\/span>\u516b\u79cd\u4e0d\u540c Agent \u5916\u58f3（scaffold）<\/span>\u4e0b\u7684\u8bc4\u5206：<\/span><\/p>\n\n\n\n\n\n
                \n

                \u57fa\u51c6<\/span><\/p>\n<\/td>\n

                		\n

                \u6700\u597d<\/span><\/p>\n<\/td>\n

                		\n

                \u6700\u5dee<\/span><\/p>\n<\/td>\n

                		\n

                \u5dee\u8ddd<\/span><\/p>\n<\/td>\n<\/tr>\n

                \n

                DeepSWE v1.1<\/span><\/p>\n<\/td>\n

                		\n

                74.2<\/span> （<\/span> mini-SWE<\/span> ）<\/span><\/p>\n<\/td>\n

                		\n

                65.5<\/span> （<\/span> OpenCode<\/span> ）<\/span><\/p>\n<\/td>\n

                		\n

                8.7 \u5206<\/span><\/p>\n<\/td>\n<\/tr>\n

                \n

                Terminal-Bench 2.1<\/span><\/p>\n<\/td>\n

                		\n

                90.6<\/span> （<\/span> Minimal<\/span> ）<\/span><\/p>\n<\/td>\n

                		\n

                84.1<\/span> （<\/span> Codex<\/span> ）<\/span><\/p>\n<\/td>\n

                		\n

                6.5 <\/span> \u5206<\/span><\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

                \u540c\u4e00\u4e2a\u6a21\u578b\u3001\u540c\u4e00\u4efd\u6743\u91cd\u3001\u540c\u4e00\u4e2a\u57fa\u51c6<\/span>，\u53ea\u56e0\u4e3a\u5916\u9762\u5957\u7684\u58f3\u4e0d\u4e00\u6837，\u5206\u6570\u5dee\u4e86<\/span> 8.7 <\/span>\u5206<\/span>\u2014\u2014<\/span>\u6bd4\u5b83\u548c Opus 5 \u5728 DeepSWE \u4e0a\u7684\u5dee\u8ddd（0.2 \u5206）\u5927 40 \u591a\u500d\u3002<\/span><\/p>\n\n\n\n\n
                \n

                \u91cd\u70b9\u63d0\u793a<\/span><\/p>\n<\/td>\n<\/tr>\n

                \n

                \u8fd9\u610f\u5473\u7740：\u4f60\u770b\u5230\u7684\u6bcf\u4e00\u4e2a Agent \u57fa\u51c6\u5206\u6570，\u90fd\u662f"\u6a21\u578b \u00d7 \u5916\u58f3"\u7684\u8054\u5408\u6210\u7ee9，\u4e0d\u662f\u6a21\u578b\u7684\u6210\u7ee9\u3002<\/span><\/p>\n

                \u5f53\u4f60\u770b\u5230<\/span>"A <\/span>\u6a21\u578b<\/span> 74.2 <\/span>\u5206，\u8d85\u8fc7<\/span> B <\/span>\u6a21\u578b\u7684<\/span> 74.0 <\/span>\u5206<\/span>"<\/span>，\u8bf7\u7acb\u523b\u95ee：<\/span>\u5b83\u4fe9\u7528\u7684\u662f\u540c\u4e00\u4e2a\u5916\u58f3\u5417？<\/span> \u5982\u679c<\/span> A <\/span>\u7528\u7684\u662f<\/span> mini-SWE<\/span>\u3001<\/span>B <\/span>\u7528\u7684\u662f<\/span> Codex<\/span>，\u90a3\u8fd9\u4e2a<\/span> 0.2 <\/span>\u5206\u7684<\/span>"<\/span>\u9886\u5148<\/span>"<\/span>\u5b8c\u5168\u53ef\u80fd\u662f\u5916\u58f3\u5e26\u6765\u7684，\u8ddf\u6a21\u578b\u65e0\u5173\u3002<\/span><\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

                \u800c\u4e14\u8981\u6ce8\u610f\u5b98\u65b9\u8bc4\u6d4b\u7684\u9ed8\u8ba4\u8bbe\u7f6e：\u4ee3\u7801\u7c7b<\/span> Agent <\/span>\u57fa\u51c6\u8dd1\u5728<\/span> DeepSeek Harness \u7684 Minimal \u6a21\u5f0f + 1M \u4e0a\u4e0b\u6587\u7a97\u53e3<\/span>\u4e0b\u3002<\/span>**<\/span>\u6362\u6389\u8fd9\u4e2a\u524d\u63d0，\u8868\u91cc\u7684\u6570\u5b57\u5c31\u4e0d\u6210\u7acb\u3002<\/span><\/p>\n

                \u6240\u4ee5\u770b\u4efb\u4f55\u4e00\u7bc7<\/span> Agent <\/span>\u8bc4\u6d4b，\u5148\u95ee\u4e09\u4e2a\u95ee\u9898：<\/span><\/p>\n\n\n\n\n\n\n
                \n

                \u8be5\u95ee\u7684<\/span><\/p>\n<\/td>\n

                		\n

                \u4e3a\u4ec0\u4e48\u95ee<\/span><\/p>\n<\/td>\n<\/tr>\n

                \n

                \u7528\u7684\u54ea\u4e2a\u5916\u58f3？<\/span><\/p>\n<\/td>\n

                		\n

                \u540c\u4e00\u4e2a\u6a21\u578b\u6362\u5916\u58f3\u80fd\u5dee<\/span> 8.7 <\/span> \u5206<\/span><\/p>\n<\/td>\n<\/tr>\n

                \n

                \u4e0a\u4e0b\u6587\u7a97\u53e3\u8bbe\u4e86\u591a\u5927？<\/span><\/p>\n<\/td>\n

                		\n

                \u5b98\u65b9\u7528<\/span> 1M<\/span> ；\u4f60\u7684\u573a\u666f\u53ef\u80fd\u53ea\u6709<\/span> 128K<\/span><\/p>\n<\/td>\n<\/tr>\n

                \n

                effort <\/span> \u8bbe\u4e86\u591a\u5c11？<\/span><\/p>\n<\/td>\n

                		\n

                25 <\/span> \u548c<\/span> 100 <\/span> \u5dee<\/span> 8.2 <\/span> \u5206，\u6210\u672c\u8fd8\u5dee<\/span> 2.5 <\/span> \u500d<\/span><\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

                \u4e09\u4e2a\u91cc\u6709\u4e00\u4e2a\u6ca1\u5199\u6e05，\u8fd9\u4e2a\u5206\u6570\u5c31\u4e0d\u80fd\u8de8\u6587\u7ae0\u6bd4\u8f83\u3002<\/span>**<\/span><\/p>\n

                11. reasoning effort：\u4ece 25 \u62c9\u5230 100，\u4e70\u5230\u4ec0\u4e48<\/span><\/h3>\n

                V4.1-Flash <\/span>\u652f\u6301<\/span>\u8fde\u7eed\u53ef\u63a7\u7684\u63a8\u7406\u52aa\u529b<\/span>：\u4e00\u4e2a<\/span> 1 <\/span>\u5230<\/span> 100 <\/span>\u7684\u6574\u6570（\u5b98\u65b9\u7ed9\u4e86<\/span> low \/ high \/ max <\/span>\u4e09\u6863\u6620\u5c04，\u5bf9\u5e94<\/span> 50 \/ 75 \/ 100<\/span>）\u3002\u8fd9\u662f\u8fd9\u6b21\u53d1\u5e03\u91cc\u5bf9\u6210\u672c\u5f71\u54cd\u6700\u76f4\u63a5\u7684\u4e00\u4e2a\u65cb\u94ae\u3002<\/span><\/p>\n

                \u5b98\u65b9\u7ed9\u4e86\u4e00\u7ec4\u5f88\u5e72\u51c0\u7684\u5bf9\u7167：<\/span><\/p>\n\n\n\n\n\n
                \n

                \u57fa\u51c6<\/span><\/p>\n<\/td>\n

                		\n

                effort=25<\/span><\/p>\n<\/td>\n

                		\n

                effort=100<\/span><\/p>\n<\/td>\n

                		\n

                \u6da8\u5206<\/span><\/p>\n<\/td>\n

                		\n

                \u6da8\u5e45<\/span><\/p>\n<\/td>\n<\/tr>\n

                \n

                DeepSWE v1.1<\/span><\/p>\n<\/td>\n

                		\n

                66.0<\/span><\/p>\n<\/td>\n

                		\n

                74.2<\/span><\/p>\n<\/td>\n

                		\n

                +8.2<\/span><\/p>\n<\/td>\n

                		\n

                12.4%<\/span><\/p>\n<\/td>\n<\/tr>\n

                \n

                Terminal-Bench 2.1<\/span><\/p>\n<\/td>\n

                		\n

                82.4<\/span><\/p>\n<\/td>\n

                		\n

                90.6<\/span><\/p>\n<\/td>\n

                		\n

                +8.2<\/span><\/p>\n<\/td>\n

                		\n

                10.0%<\/span><\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

                \u4ee3\u4ef7\u662f：\u8f93\u51fa token \u7ea6 2.5 \u500d\u3002<\/span>（\u5b98\u65b9\u53e3\u5f84）<\/span><\/p>\n

                \u672c\u673a\u63a8\u7b97：\u6309\u95f2\u65f6\u8f93\u51fa\u4ef7<\/span> 4.0 <\/span>\u5143<\/span>\/<\/span>\u767e\u4e07<\/span> token<\/span>，\u5355\u4f4d\u4efb\u52a1\u7684\u8f93\u51fa\u6210\u672c\u4ece<\/span> 4.0 <\/span>\u5143\u6da8\u5230<\/span> 10.0 <\/span>\u5143\u3002<\/span><\/p>\n

                \u00a0\u00a0\u82b1 2.5 \u500d\u7684\u94b1，\u4e70 8.2 \u4e2a\u70b9\u3002<\/span><\/p>\n

                \u8fd9\u4e2a\u4ea4\u6362\u5212\u4e0d\u5212\u7b97，\u53d6\u51b3\u4e8e\u4efb\u52a1：<\/span><\/p>\n

                  \n
                • \u503c\u5f97<\/span>：\u9519\u7684\u4ee3\u4ef7\u9ad8\u3001\u9700\u8981\u591a\u6b65\u63a8\u7406\u3001\u7b54\u6848\u6709\u660e\u786e\u5bf9\u9519（\u5199\u4ee3\u7801\u3001\u8dd1\u7ec8\u7aef\u547d\u4ee4\u3001\u505a\u9898）<\/span><\/li>\n
                • \u4e0d\u503c\u5f97<\/span>：\u4efb\u52a1\u672c\u8eab\u5f88\u7b80\u5355，\u53ea\u662f\u91cf\u5927（\u5206\u7c7b\u3001\u62bd\u53d6\u3001\u683c\u5f0f\u5316\u3001\u7b80\u5355\u95ee\u7b54）<\/span><\/li>\n<\/ul>\n

                  \u00a0\u00a0\u6b63\u786e\u7528\u6cd5\u662f"\u5206\u6863"，\u4e0d\u662f"\u5168\u5c40\u62c9\u6ee1"\u3002<\/span> \u628a<\/span> effort <\/span>\u5f53\u6210\u4e00\u4e2a\u6309\u4efb\u52a1\u5206\u7ea7\u7684\u65cb\u94ae<\/span>\u2014\u2014<\/span>\u7b80\u5355\u4efb\u52a1\u7528\u4f4e\u6863\u8dd1\u91cf，\u96be\u4efb\u52a1\u624d\u62c9\u6ee1\u3002\u8fd9\u4e00\u53e5\u8bdd\u80fd\u7701\u4e0b\u7684\u94b1，\u901a\u5e38\u6bd4\u6362\u6a21\u578b\u66f4\u591a\u3002<\/span><\/p>\n

                  12. \u4e09\u6761\u8bda\u5b9e\u8fb9\u754c<\/span><\/h3>\n

                  \u524d\u9762\u8bb2\u4e86\u5b83\u7684\u597d\u3002\u8fd9\u4e00\u8282\u8bb2\u4e09\u4e2a\u5fc5\u987b\u77e5\u9053\u7684\u574f\u3002<\/span><\/p>\n

                  \u8fb9\u754c\u4e00：\u5f00\u6e90 \u2260 \u80fd\u8dd1<\/span><\/h4>\n

                  V4.1-Flash <\/span>\u6743\u91cd\u4ee5<\/span> MIT \u8bb8\u53ef<\/span>\u5f00\u6e90，<\/span>Hugging Face <\/span>\u4e0a\u6709\u5b8c\u6574<\/span> MoE <\/span>\u6743\u91cd\u548c\u6280\u672f\u62a5\u544a\u3002\u4f46<\/span>\u300c\u5f00\u6e90\u300d\u548c\u300c\u4f60\u80fd\u8dd1\u300d\u662f\u4e24\u4ef6\u4e8b<\/span>\u3002<\/span><\/p>\n

                  \u793e\u533a\u5728\u53d1\u5e03\u540e<\/span> 30 <\/span>\u5c0f\u65f6\u5185\u5728<\/span> Hugging Face <\/span>\u4e0a\u51fa\u73b0\u4e86<\/span> 44 <\/span>\u4e2a\u4ed3\u5e93，\u5b83\u4eec\u7684\u6a21\u578b\u5361\u5199\u6e05\u4e86\u672c\u5730\u73b0\u5b9e：<\/span><\/p>\n\n\n\n\n\n\n
                  \n

                  \u91cf\u5316\u7248\u672c<\/span><\/p>\n<\/td>\n

                  		\n

                  \u4f53\u79ef<\/span><\/p>\n<\/td>\n

                  		\n

                  \u76ee\u6807\u673a\u5668<\/span><\/p>\n<\/td>\n

                  		\n

                  \u901f\u5ea6<\/span><\/p>\n<\/td>\n<\/tr>\n

                  \n

                  4-bit<\/span> （<\/span> Apple silicon<\/span> ）<\/span><\/p>\n<\/td>\n

                  		\n

                  427.6 GB<\/span><\/p>\n<\/td>\n

                  		\n

                  512 GB <\/span> \u673a\u5668<\/span><\/p>\n<\/td>\n

                  		\n

                  ~14 tok\/s<\/span><\/p>\n<\/td>\n<\/tr>\n

                  \n

                  4-bit<\/span> （\u53e6\u4e00\u7248）<\/span><\/p>\n<\/td>\n

                  		\n

                  463 GB<\/span><\/p>\n<\/td>\n

                  		\n

                  512 GB <\/span> \u673a\u5668<\/span><\/p>\n<\/td>\n

                  		\n

                  ~14 tok\/s<\/span><\/p>\n<\/td>\n<\/tr>\n

                  \n

                  2-bit<\/span><\/p>\n<\/td>\n

                  		\n

                  238.8 GB<\/span><\/p>\n<\/td>\n

                  		\n

                  256 GB Mac Studio<\/span><\/p>\n<\/td>\n

                  		\n

                  ~9.5 tok\/s<\/span><\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

                  \u800c\u4e14\u90a3\u4e2a\u80fd\u585e\u8fdb<\/span> 256GB <\/span>\u673a\u5668\u7684<\/span> 2-bit <\/span>\u7248\u672c，\u4f5c\u8005\u81ea\u5df1\u5199\u4e86\u8b66\u544a：<\/span>\u300c\u91cd\u590d\u3001\u4e0d\u8fde\u8d2f\u548c\u4efb\u52a1\u5931\u8d25\u4ecd\u53ef\u80fd\u51fa\u73b0\u3002\u300d<\/span> \u8fd8\u6709\u4e00\u4e2a<\/span> 4-bit <\/span>\u7248\u672c\u7684\u6a21\u578b\u5361\u76f4\u63a5\u8bf4\u660e：<\/span>\u8fd9\u4e2a\u67b6\u6784\u300c\u5728\u4efb\u4f55\u8fd0\u884c\u65f6\u91cc\u90fd\u4e0d\u5b58\u5728\u300d<\/span>\u2014\u2014transformers<\/span>\u3001<\/span>mlx-lm<\/span>\u3001<\/span>mlx-vlm <\/span>\u90fd\u4e0d\u652f\u6301，\u5b83\u81ea\u5e26\u4e86\u4e00\u4efd\u79fb\u690d\u3002<\/span><\/p>\n

                  \u00a0\u00a0\u6ca1\u6709\u4efb\u4f55\u516c\u5f00\u6784\u5efa\u80fd\u88c5\u8fdb 128GB \u6216\u6d88\u8d39\u7ea7 GPU\u3002<\/span><\/p>\n

                  \u6240\u4ee5\u5173\u4e8e<\/span>"<\/span>\u5f00\u6e90\u81ea\u90e8\u7f72<\/span>"<\/span>，\u52a1\u5b9e\u7684\u6001\u5ea6\u662f：<\/span><\/p>\n

                    \n
                  • \u60f3\u7528\u5b83 \u2192 \u8d70 API\u3002<\/span> \u8fd9\u662f\u7edd\u5927\u591a\u6570\u4eba\u7684\u6b63\u786e\u7b54\u6848\u3002<\/span><\/li>\n
                  • \u60f3\u81ea\u90e8\u7f72 \u2192 \u5148\u7b97\u673a\u5668\u3002<\/span> 510 GB <\/span>\u7ea7\u522b\u662f\u8d77\u70b9，\u4e0d\u662f\u7ec8\u70b9\u3002<\/span><\/li>\n
                  • \u770b\u5230"MIT \u5f00\u6e90"\u5c31\u4ee5\u4e3a\u514d\u8d39 \u2192 \u4f1a\u88ab\u786c\u4ef6\u6210\u672c\u6559\u80b2\u3002<\/span><\/li>\n<\/ul>\n

                    \u8fb9\u754c\u4e8c：\u5b98\u65b9\u81ea\u5df1\u627f\u8ba4\u7684\u9000\u5316\u98ce\u9669<\/span><\/h4>\n

                    \u8fd9\u4e00\u6761\u5f88\u5c11\u88ab\u62a5\u9053，\u4f46\u5b83\u5199\u5728\u5b98\u65b9\u6750\u6599\u91cc，\u503c\u5f97\u5355\u72ec\u62ff\u51fa\u6765\u3002<\/span><\/p>\n

                    DeepSeek <\/span>\u660e\u786e\u8bf4\u660e，\u65b0\u67b6\u6784\u5f15\u5165\u4e86<\/span>\u5c1a\u672a\u5145\u5206\u6d4b\u8bd5\u7684\u9c81\u68d2\u6027\u8fb9\u754c<\/span>：<\/span><\/p>\n

                      \n
                    • \u7a00\u758f\u9009\u62e9\u8bef\u5dee<\/span>：\u6311\u76f8\u5173\u7247\u6bb5\u65f6\u53ef\u80fd\u6311\u9519<\/span><\/li>\n
                    • \u8fd1\u4f3c\u72b6\u6001\u91cd\u5efa<\/span>：\u7f13\u5b58\u590d\u7528\u662f<\/span>"<\/span>\u6284\u7b14\u8bb0<\/span>"<\/span>，\u6284\u7684\u8fc7\u7a0b\u6709\u8fd1\u4f3c<\/span><\/li>\n<\/ul>\n

                      \u8fd9\u4e24\u4ef6\u4e8b\u5728<\/span>\u4e24\u4e2a\u5177\u4f53\u573a\u666f<\/span>\u4e0b\u53ef\u80fd\u5bfc\u81f4\u80fd\u529b\u9000\u5316：<\/span><\/p>\n

                    • \u8d85\u957f\u4e0a\u4e0b\u6587\u4e0a\u7684\u7a00\u758f\u68c0\u7d22<\/span><\/li>\n
                    • \u7f13\u5b58\u6062\u590d\u7684\u8fb9\u754c\u5904<\/span>（\u5373\u8de8\u4f1a\u8bdd\u91cd\u7528\u7f13\u5b58\u7684\u90a3\u4e2a<\/span>"<\/span>\u63a5\u7f1d<\/span>"<\/span>）<\/span><\/li>\n

                      \u5b98\u65b9\u8868\u793a\u8ba1\u5212\u505a\u989d\u5916\u538b\u529b\u6d4b\u8bd5\u3002<\/span><\/p>\n

                      \u00a0\u00a0\u8fd9\u6761\u8981\u8fd9\u4e48\u7528：<\/span> \u5982\u679c\u4f60\u7684\u573a\u666f\u6b63\u597d\u5361\u5728\u8fd9\u4e24\u5904<\/span>\u2014\u2014<\/span>\u6bd4\u5982<\/span>"<\/span>\u767e\u4e07<\/span> token <\/span>\u4e0a\u4e0b\u6587\u91cc\u505a\u7cbe\u786e\u68c0\u7d22<\/span>"<\/span>，\u6216\u8005<\/span>"<\/span>\u8de8\u4f1a\u8bdd\u53cd\u590d\u590d\u7528\u957f\u7f13\u5b58<\/span>"\u2014\u2014<\/span>\u4e0a\u7ebf\u524d\u5fc5\u987b\u81ea\u5df1\u538b\u6d4b，\u4e0d\u80fd\u5047\u8bbe\u5b83\u548c\u77ed\u4e0a\u4e0b\u6587\u4e00\u6837\u7a33\u3002<\/span> \u5b98\u65b9\u6ca1\u6d4b\u7684\u5730\u65b9，\u5c31\u662f\u98ce\u9669\u6700\u96c6\u4e2d\u7684\u5730\u65b9\u3002<\/span><\/p>\n

                      \u8fb9\u754c\u4e09：\u5b83\u7684\u5b9a\u4f4d\u662f"\u505a\u4e8b"，\u4e0d\u662f"\u60f3\u4e8b"<\/span><\/h4>\n

                      \u7b2c<\/span> 9 <\/span>\u8282\u5df2\u7ecf\u7528\u6570\u636e\u8bf4\u8fc7\u4e86，\u8fd9\u91cc\u53ea\u505a\u4e2a\u6536\u53e3：<\/span><\/p>\n\n\n\n\n\n\n\n
                      \n

                      \u5b83\u64c5\u957f<\/span><\/p>\n<\/td>\n

                      		\n

                      \u5b83\u4e0d\u64c5\u957f<\/span><\/p>\n<\/td>\n<\/tr>\n

                      \n

                      \u5f00\u5377（\u5e26\u5de5\u5177）\u4efb\u52a1<\/span><\/p>\n<\/td>\n

                      		\n

                      \u95ed\u5377\u63a8\u7406<\/span><\/p>\n<\/td>\n<\/tr>\n

                      \n

                      \u591a\u6b65\u5de5\u7a0b\u6d41\u7a0b<\/span><\/p>\n<\/td>\n

                      		\n

                      \u786c\u6838\u7406\u8bba\u63a8\u5bfc<\/span><\/p>\n<\/td>\n<\/tr>\n

                      \n

                      \u957f\u4e0a\u4e0b\u6587<\/span> + <\/span> \u5927\u91cf\u91cd\u8bfb<\/span><\/p>\n<\/td>\n

                      		\n

                      \u5355\u6b21\u8d85\u957f\u751f\u6210<\/span><\/p>\n<\/td>\n<\/tr>\n

                      \n

                      \u6210\u672c\u654f\u611f\u7684\u5927\u6279\u91cf\u8c03\u7528<\/span><\/p>\n<\/td>\n

                      		\n

                      \u9700\u8981<\/span> "<\/span> \u6700\u5f3a\u5355\u70b9\u667a\u529b<\/span> "<\/span> \u7684\u573a\u666f<\/span><\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

                      \u00a0\u00a0\u9009\u6a21\u578b\u4e0d\u662f\u9009"\u8c01\u6700\u5f3a"，\u662f\u9009"\u8c01\u6700\u4e0d\u6d6a\u8d39"\u3002<\/span> \u4e00\u4e2a\u5f00\u5377\u8003\u8bd5\u51a0\u519b\u53bb\u8003\u95ed\u5377，\u662f\u8d44\u6e90\u9519\u914d；\u53cd\u8fc7\u6765\u4e5f\u4e00\u6837\u3002<\/span><\/p>\n

                      13. \u516d\u6761\u4f7f\u7528\u89c4\u5219<\/span><\/h3>\n

                      \u628a\u524d\u9762\u6240\u6709\u5185\u5bb9\u538b\u6210\u516d\u6761\u80fd\u76f4\u63a5\u7167\u7740\u505a\u7684\u89c4\u5219：<\/span><\/p>\n\n\n\n\n
                      \n

                      \u793a\u610f\u56fe<\/span> \u00b7 <\/span> \u6587\u672c\u7ed8\u5236<\/span><\/p>\n<\/td>\n<\/tr>\n

                      \n

                      \u00a0\u00a0\u516d\u6761\u89c4\u5219<\/span>, <\/span>\u6309<\/span>"<\/span>\u6760\u6746\u4ece\u5927\u5230\u5c0f<\/span>"<\/span>\u6392<\/span><\/p>\n

                      \u200b<\/span><\/p>\n

                      \u00a0\u00a0#1 <\/span>\u5148\u7b97\u7f13\u5b58\u547d\u4e2d\u7387<\/span>, <\/span>\u518d\u8c08\u5355\u4ef7<\/span><\/p>\n

                      \u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u547d\u4e2d\u7387<\/span> 0% -> 100%, <\/span>\u6708\u8d26\u5dee<\/span> 2.2 <\/span>\u500d<\/span><\/p>\n

                      \u200b<\/span><\/p>\n

                      \u00a0\u00a0#2 <\/span>\u80fd\u7b49\u7684\u4efb\u52a1\u6392\u5230\u95f2\u65f6\u65f6\u6bb5<\/span><\/p>\n

                      \u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u4e00\u5468\u53ea\u6709<\/span> 35 \/ 168 <\/span>\u5c0f\u65f6\u662f\u9ad8\u5cf0<\/span><\/p>\n

                      \u200b<\/span><\/p>\n

                      \u00a0\u00a0#3 effort <\/span>\u6309\u4efb\u52a1\u5206\u6863<\/span>, <\/span>\u4e0d\u8981\u5168\u5c40\u62c9\u6ee1<\/span><\/p>\n

                      \u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u62c9\u6ee1<\/span> = 2.5 <\/span>\u500d\u8f93\u51fa<\/span> token<\/span><\/p>\n

                      \u200b<\/span><\/p>\n

                      \u00a0\u00a0#4 <\/span>\u957f\u4e0a\u4e0b\u6587\u8981\u770b<\/span>"<\/span>\u7f13\u5b58\u5bff\u547d<\/span>"<\/span><\/p>\n

                      \u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u5168\u5c40<\/span> KV <\/span>\u4fdd\u8bc1<\/span> 72h, SWA <\/span>\u53ea\u6709\u5206\u949f\u7ea7<\/span><\/p>\n

                      \u200b<\/span><\/p>\n

                      \u00a0\u00a0#5 <\/span>\u770b<\/span> Agent <\/span>\u5206\u6570<\/span>, <\/span>\u5148\u95ee\u7528\u7684\u54ea\u4e2a\u5916\u58f3<\/span><\/p>\n

                      \u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u540c\u4e00\u4e2a\u6a21\u578b\u6362\u58f3\u80fd\u5dee<\/span> 8.7 <\/span>\u5206<\/span><\/p>\n

                      \u200b<\/span><\/p>\n

                      \u00a0\u00a0#6 <\/span>\u5f15\u6570\u5b57\u5148\u95ee\u5206\u6bcd<\/span><\/p>\n

                      \u00a0\u00a0\u00a0\u00a0\u00a0\u00a0"1\/437" <\/span>\u662f\u5bf9<\/span> V1, "1\/4" <\/span>\u624d\u662f\u5bf9\u4e0a\u4e00\u4ee3<\/span><\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

                      \u5c55\u5f00\u8bf4：<\/span><\/p>\n

                      \u2460 \u5148\u7b97\u7f13\u5b58\u547d\u4e2d\u7387，\u518d\u8c08\u5355\u4ef7\u3002<\/span> \u547d\u4e2d\u7387\u4ece<\/span> 0% <\/span>\u5230<\/span> 100%<\/span>，\u6708\u8d26\u5355\u80fd\u5dee<\/span> 2.2 <\/span>\u500d（\u7b2c<\/span> 3 <\/span>\u8282）\u3002\u628a<\/span> System Prompt<\/span>\u3001\u9879\u76ee\u8bf4\u660e\u3001\u5de5\u5177\u5b9a\u4e49\u3001\u957f\u671f\u4e0a\u4e0b\u6587<\/span>\u56fa\u5b9a\u6210\u7a33\u5b9a\u524d\u7f00<\/span>，\u662f\u6700\u9ad8\u6760\u6746\u7684\u52a8\u4f5c\u3002<\/span><\/p>\n

                      \u2461 \u80fd\u7b49\u5c31\u6392\u5230\u95f2\u65f6\u3002<\/span> \u95f2\u65f6\u4ef7\u662f\u9ad8\u5cf0\u7684\u4e00\u534a\u3002\u4e00\u5468\u53ea\u6709<\/span> 35 <\/span>\u5c0f\u65f6\u662f\u9ad8\u5cf0\u3002<\/span>\u79bb\u7ebf\u6279\u5904\u7406\u3001\u65e5\u5fd7\u5206\u6790\u3001\u6279\u91cf\u751f\u6210<\/span>\u8fd9\u7c7b\u4efb\u52a1\u5168\u90e8\u6392\u5230\u95f2\u65f6，\u76f4\u63a5\u4e94\u6298\u3002<\/span><\/p>\n

                      \u2462 effort \u6309\u4efb\u52a1\u5206\u6863，\u4e0d\u8981\u5168\u5c40\u62c9\u6ee1\u3002<\/span> \u62c9\u6ee1<\/span> = 2.5 <\/span>\u500d\u8f93\u51fa<\/span> token<\/span>\u3002\u7b80\u5355\u4efb\u52a1\u7528\u4f4e\u6863，\u96be\u4efb\u52a1\u624d\u62c9\u6ee1\u3002<\/span><\/p>\n

                      \u2463 \u957f\u4e0a\u4e0b\u6587\u8981\u770b\u6e05"\u8bb0\u5fc6\u7684\u5bff\u547d"\u3002<\/span> \u5168\u5c40<\/span> KV <\/span>\u4fdd\u8bc1<\/span> 72 <\/span>\u5c0f\u65f6，<\/span>SWA <\/span>\u53ea\u6709\u5206\u949f\u7ea7\u3002<\/span>\u8de8\u4f1a\u8bdd\u590d\u7528\u957f\u7f13\u5b58<\/span>\u7684\u8bbe\u8ba1\u5fc5\u987b\u8003\u8651\u8fd9\u4e2a\u8fb9\u754c，\u5426\u5219\u4f1a\u51fa\u73b0<\/span>"<\/span>\u4e0a\u5348\u5f88\u5feb\u3001\u4e0b\u5348\u53d8\u6162<\/span>"<\/span>\u7684\u8be1\u5f02\u73b0\u8c61\u3002<\/span><\/p>\n

                      \u2464 \u770b Agent \u57fa\u51c6\u5206\u6570，\u5148\u95ee\u5916\u58f3\u3002<\/span> \u540c\u4e00\u4e2a\u6a21\u578b\u6362\u5916\u58f3\u5dee<\/span> 8.7 <\/span>\u5206（\u7b2c<\/span> 10 <\/span>\u8282）\u3002<\/span>\u53ea\u6bd4\u81ea\u5df1\u7528\u540c\u4e00\u5916\u58f3\u8dd1\u51fa\u6765\u7684\u6570<\/span>，\u522b\u8de8\u6587\u7ae0\u6bd4\u3002<\/span><\/p>\n

                      \u2465 \u522b\u7528\u300c1\/437\u300d\u4e0b\u5224\u65ad，\u7528\u300c1\/4\u300d\u3002<\/span> \u5f15\u6570\u5b57\u5148\u95ee\u5206\u6bcd（\u7b2c<\/span> 4 <\/span>\u8282）\u3002\u8fd9\u4e2a\u4e60\u60ef\u4e0d\u53ea\u80fd\u7528\u5728\u7f13\u5b58\u4e0a\u3002<\/span><\/p>\n

                      14. \u5e38\u89c1\u95ee\u9898<\/span><\/h3>\n

                      Q1：\u5b83\u548c V4-Pro \u5230\u5e95\u4ec0\u4e48\u5173\u7cfb？V4-Pro \u8fd8\u80fd\u7528\u5417？<\/span><\/p>\n

                      \u5b98\u65b9\u53e3\u5f84：\u81ea<\/span> 2026 <\/span>\u5e74<\/span> 9 <\/span>\u6708<\/span> 14 <\/span>\u65e5<\/span> 04:00 UTC <\/span>\u8d77，\u6240\u6709<\/span> deepseek-v4-pro<\/span> \u8bf7\u6c42<\/span>\u8def\u7531\u5230 V4.1-Flash，\u5e76\u6309 Flash \u7684\u4ef7\u683c\u8ba1\u8d39<\/span>，\u76f4\u5230<\/span> V4.1-Pro <\/span>\u53d1\u5e03\u3002\u6240\u4ee5\u5982\u679c\u4f60\u7684\u4ee3\u7801\u91cc\u786c\u7f16\u7801\u4e86<\/span> deepseek-v4-pro<\/span>，<\/span>\u5b83\u4e0d\u4f1a\u62a5\u9519，\u4f46\u6a21\u578b\u548c\u884c\u4e3a\u90fd\u53d8\u4e86<\/span>\u2014\u2014<\/span>\u8fd9\u7c7b<\/span>"<\/span>\u9759\u9ed8\u66ff\u6362<\/span>"<\/span>\u662f\u4e0a\u7ebf\u524d\u6700\u5bb9\u6613\u6f0f\u6389\u7684\u98ce\u9669\u3002<\/span><\/p>\n

                      \u5efa\u8bae：<\/span>\u73b0\u5728\u5c31\u7528\u540c\u4e00\u5957\u6d41\u7a0b\u5206\u522b\u8dd1 <\/span>deepseek-flash<\/span> \u548c <\/span>deepseek-v4-pro<\/span>，\u56fa\u5b9a\u4e0a\u4e0b\u6587\u3001\u5de5\u5177\u3001\u5ba1\u6279\u4e0e\u6210\u529f\u6807\u51c6，\u53ea\u6362\u6a21\u578b，\u7136\u540e\u5bf9\u6bd4\u5b8c\u6210\u7387\u3001\u91cd\u8bd5\u7387\u3001\u547d\u4e2d\u7387\u3001\u8f93\u51fa<\/span> token <\/span>\u548c\u5899\u4e0a\u65f6\u95f4\u3002<\/span><\/p>\n

                      Q2：\u65e7\u6a21\u578b\u540d\u8fd8\u80fd\u7528\u5417？<\/span><\/p>\n

                      \u80fd\u3002<\/span>deepseek-v4-flash<\/span> \u548c <\/span>deepseek-v4-flash-vision-exp<\/span> \u4ecd\u7136\u63a5\u53d7<\/span>，\u4f46\u5b9e\u9645\u7531<\/span> V4.1-Flash <\/span>\u670d\u52a1，\u6309<\/span> Flash <\/span>\u4ef7\u8ba1\u8d39\u3002<\/span>\u540d\u5b57\u8fd8\u5728，\u6a21\u578b\u4e0d\u662f\u90a3\u4e2a\u4e86\u3002<\/span><\/p>\n

                      Q3：\u591a\u6a21\u6001\u8981\u4e0d\u8981\u5355\u72ec\u8c03\u7528\u53e6\u4e00\u4e2a\u6a21\u578b？<\/span><\/p>\n

                      \u4e0d\u7528\u3002<\/span>\u89c6\u89c9\u7406\u89e3\u5df2\u7ecf\u505a\u8fdb\u6a21\u578b\u5185\u90e8<\/span>，<\/span>deepseek-v4-flash-vision-exp<\/span> \u5df2\u7ecf\u9000\u5f79\u3002\u4e00\u4e2a\u7aef\u70b9\u540c\u65f6\u5904\u7406\u6587\u672c\u548c\u56fe\u7247\u3002\u89c6\u89c9\u57fa\u51c6（\u57fa\u7840\u6a21\u578b\u53e3\u5f84）：<\/span>DocVQA 95.6<\/span>\u3001<\/span>RefCOCO-avg 86.0<\/span>\u3002<\/span><\/p>\n

                      Q4：\u8f93\u51fa\u957f\u5ea6\u4e0a\u9650\u662f\u591a\u5c11？<\/span><\/p>\n

                      \u5b98\u65b9\u5b9a\u4ef7\u9875\u5199<\/span>\u6700\u5927\u8f93\u51fa 384K<\/span>，\u4e0a\u4e0b\u6587<\/span> 1M<\/span>（\u7cbe\u786e\u503c<\/span> 1,048,576<\/span>，\u53ef\u5728\u6a21\u578b<\/span> config.json <\/span>\u91cc\u6838\u5bf9）\u3002\u4f46\u6ce8\u610f：<\/span>\u7b2c\u4e09\u65b9\u63a8\u7406\u5e73\u53f0\u901a\u5e38\u6709\u66f4\u5c0f\u7684\u4e0a\u9650<\/span>（\u4f8b\u5982<\/span> 16,384<\/span>），\u8fd9\u662f\u5e73\u53f0\u9650\u5236\u4e0d\u662f\u6a21\u578b\u9650\u5236\u3002\u8981\u957f\u8f93\u51fa，\u5148\u786e\u8ba4\u4f60\u8d70\u7684\u662f\u54ea\u6761\u8def\u3002<\/span><\/p>\n

                      Q5：\u5e76\u53d1\u591f\u7528\u5417？<\/span><\/p>\n

                      \u5b98\u65b9\u5e76\u53d1\u4e0a\u9650 <\/span>2,500<\/span>（<\/span>V4-Pro <\/span>\u662f<\/span> 500<\/span>）\u3002\u4ece<\/span> Pro <\/span>\u8fc1\u8fc7\u6765\u7684\u4eba\u4f1a\u62ff\u5230\u66f4\u9ad8\u7684\u9650\u989d\u3002<\/span><\/p>\n

                      Q6：\u73b0\u5728\u5165\u624b\u5408\u9002\u5417，\u8fd8\u662f\u7b49 V4.1-Pro？<\/span><\/p>\n

                      \u4e24\u4e2a\u8003\u8651：<\/span><\/p>\n

                        \n
                      • \u4ef7\u683c<\/span>：\u73b0\u5728\u5904\u4e8e<\/span>"Pro <\/span>\u8def\u7531\u5230<\/span> Flash"<\/span>\u7684\u7a97\u53e3\u671f，\u7528<\/span> Flash <\/span>\u7684\u4ef7\u62ff<\/span> Flash <\/span>\u7684\u80fd\u529b\u3002<\/span><\/li>\n
                      • \u53d8\u52a8<\/span>：<\/span>V4.1-Pro <\/span>\u53d1\u5e03\u540e，<\/span>Pro <\/span>\u7684\u8def\u7531\u4f1a\u53d8\u56de\u53bb，\u5c4a\u65f6\u4ef7\u683c\u4f53\u7cfb\u53ef\u80fd\u518d\u8c03\u3002<\/span><\/li>\n<\/ul>\n

                        \u52a1\u5b9e\u505a\u6cd5<\/span>：\u5982\u679c\u73b0\u5728\u5c31\u6709\u6210\u672c\u654f\u611f\u7684\u5927\u6279\u91cf\u4efb\u52a1，\u73b0\u5728\u5c31\u53ef\u4ee5\u4e0a；\u4f46\u8981<\/span>\u628a\u6a21\u578b\u7248\u672c\u8bb0\u8fdb\u65e5\u5fd7<\/span>，\u5e76\u5728<\/span> V4.1-Pro <\/span>\u53d1\u5e03\u540e\u91cd\u8dd1\u4e00\u6b21\u56de\u5f52\u3002<\/span><\/p>\n

                        15. \u6570\u636e\u65f6\u6548\u4e0e\u53c2\u8003\u8d44\u6599<\/span><\/h3>\n

                        \u65f6\u6548\u58f0\u660e<\/span>：\u672c\u6587\u5199\u4e8e <\/span>2026 \u5e74 9 \u6708 30 \u65e5<\/span>，\u8ddd<\/span> V4.1-Flash <\/span>\u53d1\u5e03（<\/span>9 <\/span>\u6708<\/span> 10 <\/span>\u65e5）<\/span>20 <\/span>\u5929\u3002\u4ee5\u4e0b\u5185\u5bb9\u53d8\u52a8\u98ce\u9669\u8f83\u9ad8，\u5f15\u7528\u524d\u8bf7\u6838\u5bf9\u5b98\u65b9\u6765\u6e90：<\/span><\/p>\n\n\n\n\n\n\n\n
                        \n

                        \u6613\u53d8\u9879<\/span><\/p>\n<\/td>\n

                        		\n

                        \u5f53\u524d\u503c<\/span><\/p>\n<\/td>\n

                        		\n

                        \u8bf4\u660e<\/span><\/p>\n<\/td>\n<\/tr>\n

                        \n

                        API <\/span> \u4ef7\u683c<\/span><\/p>\n<\/td>\n

                        		\n

                        \u89c1\u8868<\/span><\/p>\n<\/td>\n

                        		\n

                        \u5b98\u65b9\u4ef7\u683c\u9875\u660e\u786e\u4fdd\u7559\u8c03\u4ef7\u6743<\/span><\/p>\n<\/td>\n<\/tr>\n

                        \n

                        V4-Pro <\/span> \u8def\u7531<\/span><\/p>\n<\/td>\n

                        		\n

                        9\/14 <\/span> \u8d77\u8def\u7531\u81f3<\/span> Flash<\/span><\/p>\n<\/td>\n

                        		\n

                        \u76f4\u5230<\/span> V4.1-Pro <\/span> \u53d1\u5e03，\u5c4a\u65f6\u4f1a\u53d8<\/span><\/p>\n<\/td>\n<\/tr>\n

                        \n

                        \u5e76\u53d1\u4e0a\u9650<\/span><\/p>\n<\/td>\n

                        		\n

                        2,500<\/span><\/p>\n<\/td>\n

                        		\n

                        \u53ef\u80fd\u968f\u5bb9\u91cf\u8c03\u6574<\/span><\/p>\n<\/td>\n<\/tr>\n

                        \n

                        \u7b2c\u4e09\u65b9\u5e73\u53f0\u4ef7\u683c<\/span><\/p>\n<\/td>\n

                        		\n

                        \u5404\u5bb6\u4e0d\u540c<\/span><\/p>\n<\/td>\n

                        		\n

                        \u89c1\u4e0b，\u4ec5\u4f5c\u53c2\u8003<\/span><\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

                        \u5b98\u65b9\u4e00\u624b\u6765\u6e90<\/span><\/p>\n

                          \n
                        • DeepSeek <\/span>\u5b98\u65b9\u65b0\u95fb\u9875：<\/span>https:\/\/www.deepseek.com\/en\/news\/deepseek-v4-1-flash<\/span>（\u53d1\u5e03\u65e5<\/span> 2026-09-10<\/span>）<\/span><\/li>\n
                        • DeepSeek <\/span>\u5b98\u65b9\u5b9a\u4ef7\u9875：<\/span>https:\/\/api-docs.deepseek.com\/quick_start\/pricing<\/span><\/li>\n
                        • \u6a21\u578b\u6743\u91cd\u4e0e\u6280\u672f\u62a5\u544a：<\/span>Hugging Face <\/span>deepseek-ai\/DeepSeek-V4.1-Flash<\/span><\/li>\n<\/ul>\n

                          \u4e09\u65b9\u53e3\u5f84\u6765\u6e90<\/span>（\u4ec5\u4f9b\u53c2\u8003，\u4f1a\u53d8）<\/span><\/p>\n

                            \n
                          • \u72ec\u7acb\u8bc4\u6d4b：<\/span>Neowin <\/span>\u5b8c\u6574\u57fa\u51c6\u8868\u3001<\/span>Artificial Analysis<\/span>（<\/span>190 tok\/s<\/span>\u3001<\/span>Index 40<\/span>）<\/span><\/li>\n
                          • \u63a8\u7406\u5e73\u53f0\u5b9a\u4ef7：<\/span>DeepInfra<\/span>（<\/span>Standard $0.20 <\/span>\u8f93\u5165<\/span> \/ $0.60 <\/span>\u8f93\u51fa<\/span> \/ $0.006 <\/span>\u7f13\u5b58\u8bfb\u53d6，\u53e6\u6709<\/span> Priority 1.5x <\/span>\u4e0e<\/span> Flex 0.8x <\/span>\u4e24\u6863）<\/span><\/li>\n
                          • \u771f\u5b9e<\/span> Agent <\/span>\u6d41\u91cf\u89c2\u6d4b：\u7b2c\u4e09\u65b9<\/span> 30 <\/span>\u5929\u6d41\u91cf\u6837\u672c（<\/span>26.7M <\/span>\u65b0\u8f93\u5165<\/span> \/ 5.0M <\/span>\u8f93\u51fa<\/span> \/ 19.2 <\/span>\u4ebf\u7f13\u5b58\u91cd\u8bfb）<\/span><\/li>\n
                          • \u7ade\u54c1\u7f13\u5b58\u8bfb\u53d6\u4ef7：<\/span>GPT-5.6 Sol $0.40<\/span>\u3001<\/span>Claude Opus 5 $0.50<\/span>\u3001<\/span>Kimi K3 $0.30<\/span>（\u5404\u81ea\u516c\u5f00\u5b9a\u4ef7）<\/span><\/li>\n<\/ul>\n

                            \u53e3\u5f84\u5bf9\u7167\u901f\u67e5<\/span><\/p>\n\n\n\n\n\n\n\n
                            \n

                            \u8bf4\u6cd5<\/span><\/p>\n<\/td>\n

                            		\n

                            \u6b63\u786e\u7406\u89e3<\/span><\/p>\n<\/td>\n<\/tr>\n

                            \n

                            \u300c\u7f13\u5b58\u5360\u8d26\u5355<\/span> 98.6%<\/span> \u300d<\/span><\/p>\n<\/td>\n

                            		\n

                            \u662f <\/span>token \u5360\u6bd4<\/span> ；\u6210\u672c\u5360\u6bd4\u662f <\/span>45.1%<\/span><\/p>\n<\/td>\n<\/tr>\n

                            \n

                            \u300c<\/span> KV cache <\/span> \u964d\u5230<\/span> 1\/437<\/span> \u300d<\/span><\/p>\n<\/td>\n

                            		\n

                            \u662f\u76f8\u5bf9 <\/span>V1<\/span> ；\u76f8\u5bf9\u4e0a\u4e00\u4ee3\u662f <\/span>1\/4<\/span><\/p>\n<\/td>\n<\/tr>\n

                            \n

                            \u300c<\/span> 19 <\/span> \u9879\u6d4b\u8bc4\u8d62<\/span> 14 <\/span> \u9879\u300d<\/span><\/p>\n<\/td>\n

                            		\n

                            \u5b98\u65b9\u5b8c\u6574\u8868<\/span> 19 <\/span> \u884c；<\/span> \u4e24\u8fb9\u90fd\u6709\u5206\u7684 14 \u884c\u91cc\u8d62 12 \u884c<\/span><\/p>\n<\/td>\n<\/tr>\n

                            \n

                            \u300c\u5f00\u6e90\u53ef\u81ea\u90e8\u7f72\u300d<\/span><\/p>\n<\/td>\n

                            		\n

                            \u6743\u91cd<\/span> MIT <\/span> \u5f00\u6e90；<\/span> \u6700\u5c0f\u53ef\u7528\u6784\u5efa\u7ea6 239 GB<\/span> ，\u4e14<\/span> 2-bit <\/span> \u7248\u6709\u8d28\u91cf\u8b66\u544a<\/span><\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

                            \n

                            16. \u9644\u5f55：\u672c\u673a\u5b9e\u9a8c\u53f0<\/span><\/h3>\n

                            \u672c\u6587\u6240\u6709\u6807\u6ce8\u300c\u672c\u673a\u63a8\u7b97\u300d\u7684\u6570\u5b57，\u90fd\u7531\u4e0b\u9762\u8fd9\u4e2a\u811a\u672c\u4ea7\u51fa\u3002\u5b83<\/span>\u53ea\u7528 Python \u6807\u51c6\u5e93<\/span>，\u4e0d\u9700\u8981\u4efb\u4f55\u7b2c\u4e09\u65b9\u5305，\u628a\u4ef7\u683c\u5e38\u91cf\u6539\u6389\u5c31\u80fd\u590d\u7b97\u4f60\u81ea\u5df1\u7684\u573a\u666f：<\/span><\/p>\n\n\n\n\n
                            \n

                            Python<\/span><\/p>\n<\/td>\n<\/tr>\n

                            \n

                            # -*- coding: utf-8 -*-<\/span><\/p>\n

                            """DeepSeek V4.1-Flash <\/span>\u8d26\u5355\u5b9e\u9a8c\u53f0<\/span> \u2014\u2014 <\/span>\u7eaf\u6807\u51c6\u5e93<\/span>, <\/span>\u96f6\u4f9d\u8d56<\/span>"""<\/span><\/p>\n

                            \u200b<\/span><\/p>\n

                            M = 1_000_000.0<\/span><\/p>\n

                            # <\/span>\u6ce8：\u8f93\u51fa\u91cc\u7684<\/span> Y <\/span>\u4ee3\u8868\u4eba\u6c11\u5e01\u7b26\u53f7（\u907f\u5f00\u90e8\u5206\u7ec8\u7aef\u7684\u7f16\u7801\u95ee\u9898）<\/span><\/p>\n

                            \u200b<\/span><\/p>\n

                            # <\/span>\u5b98\u65b9\u4ef7\u683c（\u7f8e\u5143<\/span> \/ <\/span>\u6bcf\u767e\u4e07<\/span> token<\/span>）：<\/span>(<\/span>\u7f13\u5b58\u547d\u4e2d<\/span>, <\/span>\u7f13\u5b58\u672a\u547d\u4e2d<\/span>, <\/span>\u8f93\u51fa<\/span>)<\/span><\/p>\n

                            USD_OFF = (0.003, 0.15, 0.60) # V4.1-Flash <\/span>\u95f2\u65f6<\/span><\/p>\n

                            USD_PEAK = (0.006, 0.30, 1.20) # V4.1-Flash <\/span>\u9ad8\u5cf0<\/span><\/p>\n

                            # <\/span>\u4eba\u6c11\u5e01<\/span> \/ <\/span>\u6bcf\u767e\u4e07<\/span> token<\/span>（\u95f2\u65f6）<\/span><\/p>\n

                            CNY_OFF = (0.02, 1.0, 4.0)<\/span><\/p>\n

                            CNY_V4F = (0.05, 1.5, 4.5) # <\/span>\u4e0a\u4e00\u4ee3<\/span> V4-Flash <\/span>\u65e7\u4ef7<\/span><\/p>\n

                            \u200b<\/span><\/p>\n

                            # —- <\/span>\u4e00\u6b21\u771f\u5b9e<\/span> agent <\/span>\u6d41\u91cf\u7684<\/span> 30 <\/span>\u5929\u7ed3\u6784（\u7b2c\u4e09\u65b9\u53e3\u5f84）<\/span>—-<\/span><\/p>\n

                            FRESH_IN = 26.7 * M # <\/span>\u65b0\u8f93\u5165<\/span> token<\/span><\/p>\n

                            OUT_TOK = 5.0 * M # <\/span>\u8f93\u51fa<\/span> token<\/span><\/p>\n

                            CACHE_IN = 1_920.0 * M # <\/span>\u7f13\u5b58\u91cd\u8bfb<\/span> token<\/span><\/p>\n

                            \u200b<\/span><\/p>\n

                            print("=== <\/span>\u4e00\u3001<\/span>token <\/span>\u7ed3\u6784<\/span> ===")<\/span><\/p>\n

                            total = FRESH_IN + OUT_TOK + CACHE_IN<\/span><\/p>\n

                            print(f"<\/span>\u7f13\u5b58\u91cd\u8bfb<\/span> \/ <\/span>\u5168\u90e8<\/span> token = {CACHE_IN\/total*100:.1f}%")<\/span><\/p>\n

                            print(f"<\/span>\u7f13\u5b58\u91cd\u8bfb<\/span> \/ <\/span>\u8f93\u5165<\/span> token = {CACHE_IN\/(FRESH_IN+CACHE_IN)*100:.1f}% <- 98.6% <\/span>\u53e3\u5f84<\/span>")<\/span><\/p>\n

                            \u200b<\/span><\/p>\n

                            print()<\/span><\/p>\n

                            print("=== <\/span>\u4e8c\u3001\u6210\u672c\u7ed3\u6784（<\/span>V4.1-Flash <\/span>\u95f2\u65f6）<\/span>===")<\/span><\/p>\n

                            c_cache = CACHE_IN \/ M * USD_OFF[0]<\/span><\/p>\n

                            c_fresh = FRESH_IN \/ M * USD_OFF[1]<\/span><\/p>\n

                            c_out = OUT_TOK \/ M * USD_OFF[2]<\/span><\/p>\n

                            c_tot = c_cache + c_fresh + c_out<\/span><\/p>\n

                            for name, v in (("<\/span>\u7f13\u5b58\u91cd\u8bfb<\/span>", c_cache), ("<\/span>\u65b0\u8f93\u5165<\/span>", c_fresh), ("<\/span>\u8f93\u51fa<\/span>", c_out)):<\/span><\/p>\n

                            \u00a0\u00a0\u00a0\u00a0print(f" {name} | ${v:6.2f} | {v\/c_tot*100:5.1f}%")<\/span><\/p>\n

                            print(f" <\/span>\u5408\u8ba1<\/span> | ${c_tot:6.2f} | 100.0%")<\/span><\/p>\n

                            \u200b<\/span><\/p>\n

                            print()<\/span><\/p>\n

                            print("=== <\/span>\u4e09\u3001\u82e5\u5b8c\u5168\u6ca1\u6709\u7f13\u5b58\u673a\u5236（\u5168\u6309\u672a\u547d\u4e2d）<\/span>===")<\/span><\/p>\n

                            nocache = CACHE_IN \/ M * USD_OFF[1] + c_fresh + c_out<\/span><\/p>\n

                            print(f"${nocache:.2f} = <\/span>\u73b0\u4ef7\u7684<\/span> {nocache\/c_tot:.1f} <\/span>\u500d<\/span>")<\/span><\/p>\n

                            \u200b<\/span><\/p>\n

                            print()<\/span><\/p>\n

                            print("=== <\/span>\u56db\u3001\u540c\u4e00\u4efd\u6d41\u91cf<\/span>, <\/span>\u4e24\u4ee3\u4ef7\u683c（\u4eba\u6c11\u5e01）<\/span>===")<\/span><\/p>\n

                            old = sum(x\/M*y for x, y in zip((CACHE_IN, FRESH_IN, OUT_TOK), CNY_V4F))<\/span><\/p>\n

                            new = sum(x\/M*y for x, y in zip((CACHE_IN, FRESH_IN, OUT_TOK), CNY_OFF))<\/span><\/p>\n

                            print(f" V4-Flash <\/span>\u65e7\u4ef7<\/span> | Y{old:8.2f}")<\/span><\/p>\n

                            print(f" V4.1-Flash <\/span>\u65b0\u4ef7<\/span> | Y{new:8.2f}")<\/span><\/p>\n

                            print(f"<\/span>\u8d26\u5355\u964d\u5e45<\/span> {(new\/old-1)*100:6.1f}% (<\/span>\u5bf9\u7167<\/span>: <\/span>\u8f93\u51fa\u5355\u4ef7\u964d<\/span> "<\/span><\/p>\n

                            \u00a0\u00a0\u00a0\u00a0\u00a0\u00a0f"{(CNY_OFF[2]\/CNY_V4F[2]-1)*100:.1f}%)")<\/span><\/p>\n

                            \u200b<\/span><\/p>\n

                            print()<\/span><\/p>\n

                            print("=== <\/span>\u4e94\u3001\u7f13\u5b58\u547d\u4e2d\u7387\u626b\u63cf（\u6708<\/span> 100M <\/span>\u8f93\u5165<\/span> + 20M <\/span>\u8f93\u51fa<\/span>, <\/span>\u95f2\u65f6）<\/span>===")<\/span><\/p>\n

                            IN_M, OUT_M = 100.0, 20.0<\/span><\/p>\n

                            for hit in (0, 40, 80, 95, 100):<\/span><\/p>\n

                            \u00a0\u00a0\u00a0\u00a0r = hit \/ 100<\/span><\/p>\n

                            \u00a0\u00a0\u00a0\u00a0cin = IN_M*r*CNY_OFF[0] + IN_M*(1-r)*CNY_OFF[1]<\/span><\/p>\n

                            \u00a0\u00a0\u00a0\u00a0cout = OUT_M*CNY_OFF[2]<\/span><\/p>\n

                            \u00a0\u00a0\u00a0\u00a0print(f" <\/span>\u547d\u4e2d\u7387<\/span> {hit:3d}% | <\/span>\u8f93\u5165<\/span> Y{cin:7.1f} | <\/span>\u8f93\u51fa<\/span> Y{cout:5.1f} | <\/span>\u5408\u8ba1<\/span> Y{cin+cout:7.1f}")<\/span><\/p>\n

                            \u200b<\/span><\/p>\n

                            print()<\/span><\/p>\n

                            print("=== <\/span>\u516d\u3001<\/span>KV cache <\/span>\u5386\u53f2\u66f2\u7ebf（\u5b57\u8282<\/span>\/token<\/span>）<\/span>===")<\/span><\/p>\n

                            for name, b in (("V1", 389_120), ("V3.2", 48_068),<\/span><\/p>\n

                            \u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0("V4-Flash", 3_514), ("V4.1-Flash", 890)):<\/span><\/p>\n

                            \u00a0\u00a0\u00a0\u00a0print(f" {name:<11} {b:>9,} | <\/span>\u76f8\u5bf9<\/span> V4.1 = {b\/890:6.1f}x")<\/span><\/p>\n

                            print(f"V1 -> V4.1 : {389_120\/890:.0f}x (<\/span>\u5a92\u4f53\u5e38\u5f15\u7528\u7684<\/span> 1\/437)")<\/span><\/p>\n

                            print(f"<\/span>\u4e0a\u4e00\u4ee3<\/span> -> <\/span>\u65b0<\/span>: {3_514\/890:.2f}x (<\/span>\u76f8\u5bf9\u4e0a\u4e00\u4ee3\u5176\u5b9e\u53ea\u6709\u8fd9\u4e00\u6863<\/span>)")<\/span><\/p>\n

                            print(f"1M <\/span>\u4e0a\u4e0b\u6587\u5168\u5c40\u7f13\u5b58<\/span>: {890*1_048_576\/1024**3:.2f} GB")<\/span><\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

                            \u672c\u673a\u5b9e\u8dd1\u8f93\u51fa（\u5b8c\u6574）：<\/span><\/p>\n\n\n\n\n
                            \n

                            \u8fd0\u884c\u7ed3\u679c<\/span><\/p>\n<\/td>\n<\/tr>\n

                            \n

                            === <\/span>\u4e00\u3001<\/span>token <\/span>\u7ed3\u6784<\/span> ===<\/span><\/p>\n

                            \u7f13\u5b58\u91cd\u8bfb<\/span> \/ <\/span>\u5168\u90e8<\/span> token = 98.4%<\/span><\/p>\n

                            \u7f13\u5b58\u91cd\u8bfb<\/span> \/ <\/span>\u8f93\u5165<\/span> token = 98.6% <- 98.6% <\/span>\u53e3\u5f84<\/span><\/p>\n

                            \u200b<\/span><\/p>\n

                            === <\/span>\u4e8c\u3001\u6210\u672c\u7ed3\u6784（<\/span>V4.1-Flash <\/span>\u95f2\u65f6）<\/span>===<\/span><\/p>\n

                            \u00a0\u00a0\u7f13\u5b58\u91cd\u8bfb<\/span> | $ 5.76 | 45.1%<\/span><\/p>\n

                            \u00a0\u00a0\u65b0\u8f93\u5165<\/span> | $ 4.00 | 31.4%<\/span><\/p>\n

                            \u00a0\u00a0\u8f93\u51fa<\/span> | $ 3.00 | 23.5%<\/span><\/p>\n

                            \u00a0\u00a0\u5408\u8ba1<\/span> | $ 12.77 | 100.0%<\/span><\/p>\n

                            \u200b<\/span><\/p>\n

                            === <\/span>\u4e09\u3001\u82e5\u5b8c\u5168\u6ca1\u6709\u7f13\u5b58\u673a\u5236（\u5168\u6309\u672a\u547d\u4e2d）<\/span>===<\/span><\/p>\n

                            $295.00 = <\/span>\u73b0\u4ef7\u7684<\/span> 23.1 <\/span>\u500d<\/span><\/p>\n

                            \u200b<\/span><\/p>\n

                            === <\/span>\u56db\u3001\u540c\u4e00\u4efd\u6d41\u91cf<\/span>, <\/span>\u4e24\u4ee3\u4ef7\u683c（\u4eba\u6c11\u5e01）<\/span>===<\/span><\/p>\n

                            \u00a0\u00a0V4-Flash <\/span>\u65e7\u4ef7<\/span> | Y 158.55<\/span><\/p>\n

                            \u00a0\u00a0V4.1-Flash <\/span>\u65b0\u4ef7<\/span> | Y 85.10<\/span><\/p>\n

                            \u8d26\u5355\u964d\u5e45<\/span> -46.3% (<\/span>\u5bf9\u7167<\/span>: <\/span>\u8f93\u51fa\u5355\u4ef7\u964d<\/span> -11.1%)<\/span><\/p>\n

                            \u200b<\/span><\/p>\n

                            === <\/span>\u4e94\u3001\u7f13\u5b58\u547d\u4e2d\u7387\u626b\u63cf（\u6708<\/span> 100M <\/span>\u8f93\u5165<\/span> + 20M <\/span>\u8f93\u51fa<\/span>, <\/span>\u95f2\u65f6）<\/span>===<\/span><\/p>\n

                            \u00a0\u00a0\u547d\u4e2d\u7387<\/span> 0% | <\/span>\u8f93\u5165<\/span> Y 100.0 | <\/span>\u8f93\u51fa<\/span> Y 80.0 | <\/span>\u5408\u8ba1<\/span> Y 180.0<\/span><\/p>\n

                            \u00a0\u00a0\u547d\u4e2d\u7387<\/span> 40% | <\/span>\u8f93\u5165<\/span> Y 60.8 | <\/span>\u8f93\u51fa<\/span> Y 80.0 | <\/span>\u5408\u8ba1<\/span> Y 140.8<\/span><\/p>\n

                            \u00a0\u00a0\u547d\u4e2d\u7387<\/span> 80% | <\/span>\u8f93\u5165<\/span> Y 21.6 | <\/span>\u8f93\u51fa<\/span> Y 80.0 | <\/span>\u5408\u8ba1<\/span> Y 101.6<\/span><\/p>\n

                            \u00a0\u00a0\u547d\u4e2d\u7387<\/span> 95% | <\/span>\u8f93\u5165<\/span> Y 6.9 | <\/span>\u8f93\u51fa<\/span> Y 80.0 | <\/span>\u5408\u8ba1<\/span> Y 86.9<\/span><\/p>\n

                            \u00a0\u00a0\u547d\u4e2d\u7387<\/span> 100% | <\/span>\u8f93\u5165<\/span> Y 2.0 | <\/span>\u8f93\u51fa<\/span> Y 80.0 | <\/span>\u5408\u8ba1<\/span> Y 82.0<\/span><\/p>\n

                            \u200b<\/span><\/p>\n

                            === <\/span>\u516d\u3001<\/span>KV cache <\/span>\u5386\u53f2\u66f2\u7ebf（\u5b57\u8282<\/span>\/token<\/span>）<\/span>===<\/span><\/p>\n

                            \u00a0\u00a0V1 389,120 | <\/span>\u76f8\u5bf9<\/span> V4.1 = 437.2x<\/span><\/p>\n

                            \u00a0\u00a0V3.2 48,068 | <\/span>\u76f8\u5bf9<\/span> V4.1 = 54.0x<\/span><\/p>\n

                            \u00a0\u00a0V4-Flash 3,514 | <\/span>\u76f8\u5bf9<\/span> V4.1 = 3.9x<\/span><\/p>\n

                            \u00a0\u00a0V4.1-Flash 890 | <\/span>\u76f8\u5bf9<\/span> V4.1 = 1.0x<\/span><\/p>\n

                            V1 -> V4.1 : 437x (<\/span>\u5a92\u4f53\u5e38\u5f15\u7528\u7684<\/span> 1\/437)<\/span><\/p>\n

                            \u4e0a\u4e00\u4ee3<\/span> -> <\/span>\u65b0<\/span>: 3.95x (<\/span>\u76f8\u5bf9\u4e0a\u4e00\u4ee3\u5176\u5b9e\u53ea\u6709\u8fd9\u4e00\u6863<\/span>)<\/span><\/p>\n

                            1M <\/span>\u4e0a\u4e0b\u6587\u5168\u5c40\u7f13\u5b58<\/span>: 0.87 GB<\/span><\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

                            \u00a0\u00a0\u6700\u540e\u8bf4\u4e00\u53e5\u4e3a\u4ec0\u4e48\u8981\u628a\u811a\u672c\u8d34\u51fa\u6765\u3002<\/span> \u672c\u6587\u6240\u6709\u5173\u952e\u7ed3\u8bba<\/span>\u2014\u2014<\/span>\u300c\u964d<\/span> 46% <\/span>\u4e0d\u662f<\/span> 11%<\/span>\u300d\u300c<\/span>98.6% <\/span>\u662f<\/span> token <\/span>\u4e0d\u662f\u94b1\u300d\u300c\u7f13\u5b58\u7701\u4e0b<\/span> 23 <\/span>\u500d\u300d<\/span>\u2014\u2014<\/span>\u90fd\u53ef\u4ee5\u88ab\u4e0a\u9762\u8fd9\u6bb5\u4ee3\u7801\u63a8\u7ffb\u6216\u8bc1\u5b9e\u3002<\/span> \u4ef7\u683c\u8868\u662f\u516c\u5f00\u7684，\u7b97\u5f0f\u662f\u7b80\u5355\u7684，\u4f60\u4e0d\u9700\u8981\u76f8\u4fe1\u4f5c\u8005\u7684\u5224\u65ad，\u53ea\u9700\u8981\u628a\u5e38\u91cf\u6362\u6210\u4f60\u7684\u7528\u91cf\u8dd1\u4e00\u904d\u3002<\/span><\/p>\n

                            \u4e00\u4e2a\u6a21\u578b\u503c\u4e0d\u503c\u5f97\u7528，\u6700\u7ec8\u4e0d\u662f\u770b\u5b83\u7684\u53c2\u6570，\u662f\u770b\u5b83\u5728\u4f60\u7684\u8d26\u5355\u7ed3\u6784\u91cc，\u843d\u5728\u54ea\u4e00\u884c\u3002<\/span><\/p>\n","protected":false},"excerpt":{"rendered":"

                            DeepSeek V4.1-Flash \u00b7 \u67b6\u6784\u4e0e\u6210\u672c\u6df1\u5ea6\u89e3\u6790\u67b6\u6784 \u2192 \u7f13\u5b58 \u2192 \u5b9a\u4ef7 \u2192 \u8d26\u5355：\u4e00\u6761\u56e0\u679c\u94fe\u8d70\u5230\u5e95\u6570\u636e\u622a\u6b62 2026 \u5e74 9 \u6708 30 \u65e5 ｜ \u5168\u6587\u7ea6 2.0 \u4e07\u5b57\u7b26\u30013 \u5f20\u6587\u672c\u793a\u610f\u4e0e 2 \u5f20\u56fe\u8868 ｜ \u5b98\u65b9\u4e00\u624b\u6570\u636e \u672c\u673a\u5b9e\u8dd1\u811a\u672c，\u4ef7\u683c\u53e3\u5f84\u5168\u90e8\u6ce8\u660e\u76ee\u5f551. \u5f15\u5b50：\u4e00\u4e2a\u540d\u4e0d\u526f\u5b9e\u7684\u300cFlash\u300d2. \u5148\u770b\u4ef7\u683c\u8868：\u964d\u5f97\u6700\u72e0\u7684\u662f\u6700\u4e0d\u8d77\u773c\u7684\u90a3\u4e00\u884c3. \u8d26\u5355\u7684\u4e24\u534a：98.6%<\/p>\n","protected":false},"author":2,"featured_media":113605,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[66,99,101,50,190],"topic":[],"class_list":["post-113607","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-server","tag-ai","tag-java","tag-spring","tag-50","tag-190"],"yoast_head":"\n\u4e00\u6587\u8bfb\u61c2 DeepSeek V4.1 Flash\uff1a\u964d\u4ef7\u7684\u4e0d\u662f\u8f93\u51fa\u4ef7\uff0c\u662f\u4f60\u4ece\u6ca1\u770b\u8fc7\u7684\u90a3\u4e00\u884c - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.wsisp.com\/helps\/113607.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"\u4e00\u6587\u8bfb\u61c2 DeepSeek V4.1 Flash\uff1a\u964d\u4ef7\u7684\u4e0d\u662f\u8f93\u51fa\u4ef7\uff0c\u662f\u4f60\u4ece\u6ca1\u770b\u8fc7\u7684\u90a3\u4e00\u884c - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"DeepSeek V4.1-Flash \u00b7 \u67b6\u6784\u4e0e\u6210\u672c\u6df1\u5ea6\u89e3\u6790\u67b6\u6784 \u2192 \u7f13\u5b58 \u2192 \u5b9a\u4ef7 \u2192 \u8d26\u5355：\u4e00\u6761\u56e0\u679c\u94fe\u8d70\u5230\u5e95\u6570\u636e\u622a\u6b62 2026 \u5e74 9 \u6708 30 \u65e5 ｜ \u5168\u6587\u7ea6 2.0 \u4e07\u5b57\u7b26\u30013 \u5f20\u6587\u672c\u793a\u610f\u4e0e 2 \u5f20\u56fe\u8868 ｜ \u5b98\u65b9\u4e00\u624b\u6570\u636e \u672c\u673a\u5b9e\u8dd1\u811a\u672c，\u4ef7\u683c\u53e3\u5f84\u5168\u90e8\u6ce8\u660e\u76ee\u5f551. \u5f15\u5b50：\u4e00\u4e2a\u540d\u4e0d\u526f\u5b9e\u7684\u300cFlash\u300d2. \u5148\u770b\u4ef7\u683c\u8868：\u964d\u5f97\u6700\u72e0\u7684\u662f\u6700\u4e0d\u8d77\u773c\u7684\u90a3\u4e00\u884c3. \u8d26\u5355\u7684\u4e24\u534a：98.6%\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/113607.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-10-06T15:49:17+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/10\/20261006154913-6ac518795b548.png\" \/>\n<meta name=\"author\" content=\"admin\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"\u4f5c\u8005\" \/>\n\t<meta name=\"twitter:data1\" content=\"admin\" \/>\n\t<meta name=\"twitter:label2\" content=\"\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4\" \/>\n\t<meta name=\"twitter:data2\" content=\"10 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/113607.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/113607.html\",\"name\":\"\u4e00\u6587\u8bfb\u61c2 DeepSeek V4.1 Flash\uff1a\u964d\u4ef7\u7684\u4e0d\u662f\u8f93\u51fa\u4ef7\uff0c\u662f\u4f60\u4ece\u6ca1\u770b\u8fc7\u7684\u90a3\u4e00\u884c - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-10-06T15:49:17+00:00\",\"dateModified\":\"2026-10-06T15:49:17+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/113607.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/113607.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/113607.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u4e00\u6587\u8bfb\u61c2 DeepSeek V4.1 Flash\uff1a\u964d\u4ef7\u7684\u4e0d\u662f\u8f93\u51fa\u4ef7\uff0c\u662f\u4f60\u4ece\u6ca1\u770b\u8fc7\u7684\u90a3\u4e00\u884c\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\",\"url\":\"https:\/\/www.wsisp.com\/helps\/\",\"name\":\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"description\":\"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}\"},\"query-input\":\"required name=search_term_string\"}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\",\"name\":\"admin\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"contentUrl\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"caption\":\"admin\"},\"sameAs\":[\"http:\/\/wp.wsisp.com\"],\"url\":\"https:\/\/www.wsisp.com\/helps\/author\/admin\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"\u4e00\u6587\u8bfb\u61c2 DeepSeek V4.1 Flash\uff1a\u964d\u4ef7\u7684\u4e0d\u662f\u8f93\u51fa\u4ef7\uff0c\u662f\u4f60\u4ece\u6ca1\u770b\u8fc7\u7684\u90a3\u4e00\u884c - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.wsisp.com\/helps\/113607.html","og_locale":"zh_CN","og_type":"article","og_title":"\u4e00\u6587\u8bfb\u61c2 DeepSeek V4.1 Flash\uff1a\u964d\u4ef7\u7684\u4e0d\u662f\u8f93\u51fa\u4ef7\uff0c\u662f\u4f60\u4ece\u6ca1\u770b\u8fc7\u7684\u90a3\u4e00\u884c - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"DeepSeek V4.1-Flash \u00b7 \u67b6\u6784\u4e0e\u6210\u672c\u6df1\u5ea6\u89e3\u6790\u67b6\u6784 \u2192 \u7f13\u5b58 \u2192 \u5b9a\u4ef7 \u2192 \u8d26\u5355：\u4e00\u6761\u56e0\u679c\u94fe\u8d70\u5230\u5e95\u6570\u636e\u622a\u6b62 2026 \u5e74 9 \u6708 30 \u65e5 ｜ \u5168\u6587\u7ea6 2.0 \u4e07\u5b57\u7b26\u30013 \u5f20\u6587\u672c\u793a\u610f\u4e0e 2 \u5f20\u56fe\u8868 ｜ \u5b98\u65b9\u4e00\u624b\u6570\u636e \u672c\u673a\u5b9e\u8dd1\u811a\u672c，\u4ef7\u683c\u53e3\u5f84\u5168\u90e8\u6ce8\u660e\u76ee\u5f551. \u5f15\u5b50：\u4e00\u4e2a\u540d\u4e0d\u526f\u5b9e\u7684\u300cFlash\u300d2. \u5148\u770b\u4ef7\u683c\u8868：\u964d\u5f97\u6700\u72e0\u7684\u662f\u6700\u4e0d\u8d77\u773c\u7684\u90a3\u4e00\u884c3. \u8d26\u5355\u7684\u4e24\u534a：98.6%","og_url":"https:\/\/www.wsisp.com\/helps\/113607.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-10-06T15:49:17+00:00","og_image":[{"url":"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/10\/20261006154913-6ac518795b548.png"}],"author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"10 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/113607.html","url":"https:\/\/www.wsisp.com\/helps\/113607.html","name":"\u4e00\u6587\u8bfb\u61c2 DeepSeek V4.1 Flash\uff1a\u964d\u4ef7\u7684\u4e0d\u662f\u8f93\u51fa\u4ef7\uff0c\u662f\u4f60\u4ece\u6ca1\u770b\u8fc7\u7684\u90a3\u4e00\u884c - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-10-06T15:49:17+00:00","dateModified":"2026-10-06T15:49:17+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/113607.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/113607.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/113607.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"\u4e00\u6587\u8bfb\u61c2 DeepSeek V4.1 Flash\uff1a\u964d\u4ef7\u7684\u4e0d\u662f\u8f93\u51fa\u4ef7\uff0c\u662f\u4f60\u4ece\u6ca1\u770b\u8fc7\u7684\u90a3\u4e00\u884c"}]},{"@type":"WebSite","@id":"https:\/\/www.wsisp.com\/helps\/#website","url":"https:\/\/www.wsisp.com\/helps\/","name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","description":"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}"},"query-input":"required name=search_term_string"}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41","name":"admin","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/","url":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","contentUrl":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","caption":"admin"},"sameAs":["http:\/\/wp.wsisp.com"],"url":"https:\/\/www.wsisp.com\/helps\/author\/admin"}]}},"_links":{"self":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/113607","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/comments?post=113607"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/113607\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media\/113605"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=113607"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=113607"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=113607"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=113607"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}