{"id":3643,"date":"2026-08-09T23:46:44","date_gmt":"2026-08-09T15:46:44","guid":{"rendered":"https:\/\/blog.zengqq.com.cn\/?p=3643"},"modified":"2026-08-10T00:16:43","modified_gmt":"2026-08-09T16:16:43","slug":"%e5%af%b9%e4%ba%8e%ef%bc%8c%e4%b8%ba%e4%bb%80%e4%b9%88%e7%9c%8b%e5%88%b0%e7%9a%84%e6%af%94%e6%84%9f%e8%a7%89%e7%9a%84%e6%85%a2","status":"publish","type":"post","link":"https:\/\/blog.zengqq.com.cn\/?p=3643","title":{"rendered":"\u5bf9\u4e8e\u81ea\u5df1\u67b6\u8bbe\u7684GPT\uff0c\u4e3a\u4ec0\u4e48\u56de\u7b54\u95ee\u9898\u65f6\u770b\u5230\u7684\u6bd4\u611f\u89c9\u7684\u6162"},"content":{"rendered":"\n<article class=\"tech-post\">\n  <p class=\"lead\">\n    \u5728\u672c\u5730\u4f7f\u7528 Ollama \u8fd0\u884c Qwen3.5:27b \u8fd9\u7c7b\u5927\u53c2\u6570\u6a21\u578b\u65f6\uff0c\u5f88\u591a\u4eba\u4f1a\u9047\u5230\u4e00\u4e2a\u770b\u8d77\u6765\u975e\u5e38\u77db\u76fe\u7684\u73b0\u8c61\uff1aAgent \u6267\u884c\u7ec6\u8282\u91cc\u6bcf\u4e00\u6b65\u5de5\u5177\u8c03\u7528\u90fd\u53ea\u663e\u793a\u51e0\u79d2\u3001\u5341\u51e0\u79d2\uff0c\u4f46\u5b8c\u6574\u56de\u7b54\u5374\u9700\u8981\u5341\u51e0\u5206\u949f\u3002\u8fd9\u7bc7\u6587\u7ae0\u4ece\u4e00\u4e2a\u5b9e\u9645\u6848\u4f8b\u51fa\u53d1\uff0c\u62c6\u89e3\u201c\u601d\u8003\u7528\u65f6\u201d\u5230\u5e95\u8868\u793a\u4ec0\u4e48\uff0c\u4ee5\u53ca\u5982\u4f55\u5224\u65ad\u81ea\u5df1\u4e0b\u8f7d\u7684\u6a21\u578b\u662f\u5426\u7ecf\u8fc7\u91cf\u5316\u3002\n  <\/p>\n\n  <blockquote style=\"border-left:4px solid #4a90d9; padding:12px 16px; margin:24px 0; background:#f6f8fa;\">\n    <p style=\"margin:0;\">\n      \u5982\u679c\u4f60\u770b\u5230\u7c7b\u4f3c <code>list_knowledge<\/code>\u3001<code>query_knowledge_files<\/code>\u3001<code>view_file<\/code>\u3001<code>grep_knowledge_files<\/code> \u7684\u6267\u884c\u8017\u65f6\u53ea\u6709\u51e0\u79d2\uff0c\u4f46\u603b\u56de\u7b54\u65f6\u95f4\u4ecd\u7136\u5f88\u957f\uff0c\u901a\u5e38\u8bf4\u660e\u6162\u7684\u4e0d\u662f\u5de5\u5177\u8c03\u7528\uff0c\u800c\u662f\u6a21\u578b\u672c\u4f53\u751f\u6210\u6587\u672c\u7684\u8fc7\u7a0b\u3002\n    <\/p>\n  <\/blockquote>\n\n  <h2>\u4e00\u3001\u95ee\u9898\u73b0\u8c61\uff1a\u5de5\u5177\u8c03\u7528\u51e0\u79d2\uff0c\u603b\u56de\u7b54\u5374\u8981 15 \u5206\u949f<\/h2>\n\n  <p>\n    \u5728\u672c\u5730\u8fd0\u884c Ollama \u65f6\uff0c\u5982\u679c\u914d\u5408 Agent\u3001RAG \u6216\u6587\u4ef6\u68c0\u7d22\u7c7b\u5de5\u5177\u94fe\uff0c\u4f60\u53ef\u80fd\u4f1a\u770b\u5230\u7c7b\u4f3c\u4e0b\u9762\u7684\u6267\u884c\u6458\u8981\uff1a\n  <\/p>\n\n  <pre style=\"tech-code\"><code>\u5df2\u5206\u6790 list_knowledge, query_knowledge_files, 4 view_file, grep_knowledge_files\n\n\u601d\u8003\u7528\u65f6 3 \u79d2\nView Result from list_knowledge\n\n\u601d\u8003\u7528\u65f6 7 \u79d2\nView Result from query_knowledge_files\n\n\u601d\u8003\u7528\u65f6 17 \u79d2\nView Result from view_file\nView Result from view_file\n\n\u601d\u8003\u7528\u65f6 3 \u79d2\nView Result from grep_knowledge_files\nView Result from view_file\nView Result from view_file<\/code><\/pre>\n\n  <p>\n    \u4ece\u8fd9\u4e9b\u65e5\u5fd7\u770b\uff0c\u6bcf\u4e00\u6b65\u5de5\u5177\u6267\u884c\u4f3c\u4e4e\u90fd\u5f88\u5feb\uff1a3 \u79d2\u30017 \u79d2\u300117 \u79d2\u30013 \u79d2\u3002\u4f46\u7528\u6237\u5b9e\u9645\u7b49\u5f85\u7684\u5b8c\u6574\u56de\u7b54\u65f6\u95f4\u5374\u53ef\u80fd\u8fbe\u5230 10 \u5206\u949f\u300115 \u5206\u949f\uff0c\u751a\u81f3\u66f4\u4e45\u3002\n  <\/p>\n\n  <p>\n    \u8fd9\u91cc\u7684\u5173\u952e\u95ee\u9898\u662f\uff1a<strong>\u8fd9\u4e9b\u201c\u601d\u8003\u7528\u65f6\u201d\u5230\u5e95\u4ee3\u8868\u4ec0\u4e48\uff1f\u5b83\u4eec\u662f\u5426\u7b49\u4e8e\u6a21\u578b\u751f\u6210\u56de\u7b54\u7684\u603b\u65f6\u95f4\uff1f<\/strong>\n  <\/p>\n\n  <h2>\u4e8c\u3001\u4f60\u770b\u5230\u7684\u201c\u601d\u8003\u7528\u65f6\u201d\u5927\u6982\u7387\u4e0d\u662f LLM \u63a8\u7406\u603b\u8017\u65f6<\/h2>\n\n  <p>\n    \u5728 Agent \u6216 RAG \u7cfb\u7edf\u4e2d\uff0c\u4e00\u6b21\u5b8c\u6574\u56de\u7b54\u901a\u5e38\u5305\u542b\u591a\u4e2a\u9636\u6bb5\uff1a\n  <\/p>\n\n  <ul>\n    <li>\u6a21\u578b\u51b3\u5b9a\u4e0b\u4e00\u6b65\u8c03\u7528\u54ea\u4e2a\u5de5\u5177\uff1b<\/li>\n    <li>\u5de5\u5177\u5b9e\u9645\u6267\u884c\uff0c\u4f8b\u5982\u5217\u76ee\u5f55\u3001\u67e5\u8be2\u77e5\u8bc6\u5e93\u3001\u8bfb\u53d6\u6587\u4ef6\u3001\u641c\u7d22\u5185\u5bb9\uff1b<\/li>\n    <li>\u5de5\u5177\u7ed3\u679c\u8fd4\u56de\u7ed9\u6a21\u578b\uff1b<\/li>\n    <li>\u6a21\u578b\u7ee7\u7eed\u8bfb\u53d6\u4e0a\u4e0b\u6587\u3001\u63a8\u7406\u5e76\u7ec4\u7ec7\u6700\u7ec8\u56de\u7b54\uff1b<\/li>\n    <li>\u6a21\u578b\u9010 token \u751f\u6210\u5b8c\u6574\u6587\u672c\u3002<\/li>\n  <\/ul>\n\n  <p>\n    \u5f88\u591a\u754c\u9762\u91cc\u663e\u793a\u7684\u201c\u601d\u8003\u7528\u65f6 X \u79d2\u201d\uff0c\u5f80\u5f80\u53ea\u662f\u67d0\u4e00\u6b21\u5de5\u5177\u8c03\u7528\u6216\u67d0\u4e00\u5c0f\u6bb5\u5904\u7406\u8fc7\u7a0b\u7684\u8017\u65f6\uff0c\u800c\u4e0d\u662f\u6574\u4e2a\u6a21\u578b\u751f\u6210\u56de\u7b54\u7684\u8017\u65f6\u3002\n  <\/p>\n\n  <table style=\"width:100%; border-collapse:collapse; margin:24px 0;\">\n    <thead>\n      <tr style=\"background:#f0f3f7;\">\n        <th style=\"border:1px solid #d7dde5; padding:10px; text-align:left;\">\u9636\u6bb5<\/th>\n        <th style=\"border:1px solid #d7dde5; padding:10px; text-align:left;\">\u53ef\u80fd\u8017\u65f6<\/th>\n        <th style=\"border:1px solid #d7dde5; padding:10px; text-align:left;\">\u8bf4\u660e<\/th>\n      <\/tr>\n    <\/thead>\n    <tbody>\n      <tr>\n        <td style=\"border:1px solid #d7dde5; padding:10px;\">\u5de5\u5177\u8c03\u7528<\/td>\n        <td style=\"border:1px solid #d7dde5; padding:10px;\">\u51e0\u79d2\u5230\u5341\u51e0\u79d2<\/td>\n        <td style=\"border:1px solid #d7dde5; padding:10px;\">\u5982 list_knowledge\u3001view_file\u3001grep \u7b49\uff0c\u901a\u5e38\u662f\u672c\u5730\u6587\u4ef6\u64cd\u4f5c\u6216\u68c0\u7d22\u64cd\u4f5c\u3002<\/td>\n      <\/tr>\n      <tr>\n        <td style=\"border:1px solid #d7dde5; padding:10px;\">\u4e0a\u4e0b\u6587\u7ec4\u88c5<\/td>\n        <td style=\"border:1px solid #d7dde5; padding:10px;\">\u8f83\u77ed<\/td>\n        <td style=\"border:1px solid #d7dde5; padding:10px;\">\u5c06\u5de5\u5177\u7ed3\u679c\u62fc\u63a5\u8fdb Prompt\uff0c\u901a\u5e38\u4e0d\u662f\u4e3b\u8981\u74f6\u9888\u3002<\/td>\n      <\/tr>\n      <tr>\n        <td style=\"border:1px solid #d7dde5; padding:10px;\"><strong>LLM \u63a8\u7406\u751f\u6210<\/strong><\/td>\n        <td style=\"border:1px solid #d7dde5; padding:10px;\"><strong>\u6570\u5206\u949f\u5230\u5341\u51e0\u5206\u949f<\/strong><\/td>\n        <td style=\"border:1px solid #d7dde5; padding:10px;\">\u6a21\u578b\u9010 token \u751f\u6210\u6700\u7ec8\u56de\u7b54\uff0c\u672c\u5730\u8fd0\u884c 27B \u6a21\u578b\u65f6\u5c24\u5176\u8017\u65f6\u3002<\/td>\n      <\/tr>\n    <\/tbody>\n  <\/table>\n\n  <p>\n    \u56e0\u6b64\uff0c\u5982\u679c\u65e5\u5fd7\u4e2d\u663e\u793a\u201c\u601d\u8003\u7528\u65f6 3 \u79d2\u201d\u201c\u601d\u8003\u7528\u65f6 7 \u79d2\u201d\uff0c\u8fd9\u4e9b\u6570\u5b57\u672c\u8eab\u53ef\u80fd\u662f\u771f\u5b9e\u7684\u3002\u4f46\u5b83\u4eec\u5e76\u4e0d\u4e00\u5b9a\u4ee3\u8868\u6a21\u578b\u5df2\u7ecf\u5b8c\u6210\u4e86\u5168\u90e8\u63a8\u7406\u3002\u771f\u6b63\u6162\u7684\u90e8\u5206\uff0c\u5f80\u5f80\u662f\u6a21\u578b\u8bfb\u53d6\u5927\u91cf\u4e0a\u4e0b\u6587\u4e4b\u540e\uff0c\u5f00\u59cb\u751f\u6210\u6700\u7ec8\u56de\u7b54\u7684\u8fc7\u7a0b\u3002\n  <\/p>\n\n  <h2>\u4e09\u3001\u672c\u5730\u8fd0\u884c 27B \u6a21\u578b\u4e3a\u4ec0\u4e48\u8fd9\u4e48\u6162\uff1f<\/h2>\n\n  <p>\n    27B \u8868\u793a\u6a21\u578b\u5927\u7ea6\u6709 270 \u4ebf\u53c2\u6570\u3002\u5373\u4f7f\u7ecf\u8fc7\u91cf\u5316\uff0c\u5b83\u4f9d\u7136\u662f\u4e00\u4e2a\u975e\u5e38\u5927\u7684\u6a21\u578b\u3002\u672c\u5730\u63a8\u7406\u6162\u901a\u5e38\u4e0d\u662f\u5355\u4e00\u539f\u56e0\uff0c\u800c\u662f\u591a\u4e2a\u56e0\u7d20\u53e0\u52a0\u3002\n  <\/p>\n\n  <h3>1. \u663e\u5b58\u4e0d\u8db3\u5bfc\u81f4\u6a21\u578b\u90e8\u5206\u8dd1\u5728 CPU \u4e0a<\/h3>\n\n  <p>\n    \u5927\u6a21\u578b\u6700\u9002\u5408\u653e\u5728 GPU \u663e\u5b58\u4e2d\u8fd0\u884c\u3002\u5982\u679c\u663e\u5b58\u4e0d\u8db3\u4ee5\u5bb9\u7eb3\u6a21\u578b\u6743\u91cd\u548c\u4e0a\u4e0b\u6587 KV Cache\uff0cOllama \u53ef\u80fd\u4f1a\u628a\u4e00\u90e8\u5206\u8ba1\u7b97\u653e\u5230 CPU \u548c\u5185\u5b58\u4e2d\u3002\n  <\/p>\n\n  <p>\n    \u4e00\u65e6\u6a21\u578b\u5927\u91cf\u4f9d\u8d56 CPU \u63a8\u7406\uff0ctoken \u751f\u6210\u901f\u5ea6\u901a\u5e38\u4f1a\u663e\u8457\u4e0b\u964d\u3002\u5bf9\u4e8e 27B \u7ea7\u522b\u6a21\u578b\uff0c\u5728\u666e\u901a CPU \u4e0a\u6bcf\u79d2\u53ea\u80fd\u751f\u6210\u51e0\u4e2a token \u5e76\u4e0d\u7f55\u89c1\u3002\n  <\/p>\n\n  <h3>2. \u6a21\u578b\u6587\u4ef6\u867d\u7136\u91cf\u5316\u4e86\uff0c\u4f46\u4ecd\u7136\u5f88\u5927<\/h3>\n\n  <p>\n    27B \u6a21\u578b\u5373\u4f7f\u4f7f\u7528 4-bit \u91cf\u5316\uff0c\u4f53\u79ef\u901a\u5e38\u4e5f\u5728 16GB \u5de6\u53f3\u3002\u8fd9\u610f\u5473\u7740\u5b83\u4ecd\u7136\u9700\u8981\u8f83\u5927\u7684\u5185\u5b58\u6216\u663e\u5b58\u7a7a\u95f4\u3002\n  <\/p>\n\n  <p>\n    \u91cf\u5316\u53ef\u4ee5\u964d\u4f4e\u786c\u4ef6\u95e8\u69db\uff0c\u4f46\u5e76\u4e0d\u80fd\u8ba9\u5927\u6a21\u578b\u5728\u6240\u6709\u8bbe\u5907\u4e0a\u90fd\u53d8\u5feb\u3002\u5982\u679c\u786c\u4ef6\u672c\u8eab\u5e26\u5bbd\u4e0d\u8db3\u3001\u7b97\u529b\u4e0d\u8db3\uff0c\u91cf\u5316\u540e\u7684\u6a21\u578b\u4f9d\u7136\u53ef\u80fd\u5f88\u6162\u3002\n  <\/p>\n\n  <h3>3. \u957f\u4e0a\u4e0b\u6587\u4f1a\u8fdb\u4e00\u6b65\u653e\u5927\u8017\u65f6<\/h3>\n\n  <p>\n    Agent \u6a21\u5f0f\u4e0b\uff0c\u6a21\u578b\u4e0d\u4ec5\u8981\u5904\u7406\u7528\u6237\u95ee\u9898\uff0c\u8fd8\u8981\u8bfb\u53d6\u5de5\u5177\u8fd4\u56de\u7ed3\u679c\u3001\u5386\u53f2\u5bf9\u8bdd\u3001\u6587\u4ef6\u5185\u5bb9\u3001\u77e5\u8bc6\u5e93\u7247\u6bb5\u7b49\u3002\u4e0a\u4e0b\u6587\u8d8a\u957f\uff0c\u6a21\u578b\u5904\u7406\u8d8a\u6162\u3002\n  <\/p>\n\n  <p>\n    \u5c24\u5176\u662f\u5f53\u591a\u4e2a <code>view_file<\/code> \u6216 <code>grep_knowledge_files<\/code> \u8fd4\u56de\u5927\u91cf\u6587\u672c\u65f6\uff0c\u6a21\u578b\u9700\u8981\u6d88\u5316\u8fd9\u4e9b\u5185\u5bb9\uff0c\u518d\u751f\u6210\u5b8c\u6574\u56de\u7b54\uff0c\u65f6\u95f4\u6210\u672c\u4f1a\u660e\u663e\u589e\u52a0\u3002\n  <\/p>\n\n  <h2>\u56db\u3001\u5982\u4f55\u5224\u65ad Qwen3.5:27b \u662f\u5426\u91cf\u5316\u8fc7\uff1f<\/h2>\n\n  <p>\n    \u5728 Ollama \u4e2d\uff0c\u5982\u679c\u4f60\u76f4\u63a5\u8fd0\u884c\u7c7b\u4f3c\u4e0b\u9762\u7684\u547d\u4ee4\uff1a\n  <\/p>\n\n  <pre style=\"background:#1e1e1e; color:#e6e6e6; padding:16px; overflow:auto; border-radius:6px;\"><code>ollama run qwen3.5:27b<\/code><\/pre>\n\n  <p>\n    \u901a\u5e38\u62c9\u53d6\u5230\u7684\u662f Ollama \u5b98\u65b9\u4ed3\u5e93\u4e2d\u7684\u9ed8\u8ba4\u6807\u7b7e\u3002\u5bf9\u4e8e\u5927\u53c2\u6570\u6a21\u578b\uff0c\u5b98\u65b9\u5e93\u4e3a\u4e86\u517c\u5bb9\u66f4\u591a\u786c\u4ef6\uff0c\u9ed8\u8ba4\u6807\u7b7e\u5927\u6982\u7387\u5df2\u7ecf\u662f\u91cf\u5316\u7248\u672c\uff0c\u4f8b\u5982 <code>Q4_K_M<\/code>\u3001<code>Q4_0<\/code> \u6216\u7c7b\u4f3c\u7b49\u7ea7\u3002\n  <\/p>\n\n  <p>\n    \u4f60\u53ef\u4ee5\u901a\u8fc7\u4e24\u79cd\u65b9\u5f0f\u786e\u8ba4\u3002\n  <\/p>\n\n  <h3>\u65b9\u6cd5\u4e00\uff1a\u4f7f\u7528 ollama show \u67e5\u770b\u6a21\u578b\u4fe1\u606f<\/h3>\n\n  <p>\u5728\u7ec8\u7aef\u6267\u884c\uff1a<\/p>\n\n  <pre style=\"background:#1e1e1e; color:#e6e6e6; padding:16px; overflow:auto; border-radius:6px;\"><code>ollama show qwen3.5:27b<\/code><\/pre>\n\n  <p>\n    \u5982\u679c\u4f60\u7684\u6a21\u578b\u540d\u79f0\u5b9e\u9645\u4e0a\u662f <code>qwen2.5:27b<\/code> \u6216 <code>qwen3:27b<\/code>\uff0c\u8bf7\u5c06\u547d\u4ee4\u4e2d\u7684\u6a21\u578b\u540d\u66ff\u6362\u4e3a\u5b9e\u9645\u540d\u79f0\u3002\n  <\/p>\n\n  <p>\u8f93\u51fa\u4e2d\u901a\u5e38\u4f1a\u6709\u7c7b\u4f3c\u5b57\u6bb5\uff1a<\/p>\n\n  <pre style=\"background:#1e1e1e; color:#e6e6e6; padding:16px; overflow:auto; border-radius:6px;\"><code>Information:\n  architecture: qwen\n  parameters: 27.0B\n  quantization: Q4_K_M\n  context length: 32768<\/code><\/pre>\n\n  <p>\n    \u5176\u4e2d <code>quantization<\/code> \u5b57\u6bb5\u5c31\u662f\u5173\u952e\u3002\n  <\/p>\n\n  <table style=\"width:100%; border-collapse:collapse; margin:24px 0;\">\n    <thead>\n      <tr style=\"background:#f0f3f7;\">\n        <th style=\"border:1px solid #d7dde5; padding:10px; text-align:left;\">quantization \u503c<\/th>\n        <th style=\"border:1px solid #d7dde5; padding:10px; text-align:left;\">\u542b\u4e49<\/th>\n      <\/tr>\n    <\/thead>\n    <tbody>\n      <tr>\n        <td style=\"border:1px solid #d7dde5; padding:10px;\"><code>Q4_K_M<\/code><\/td>\n        <td style=\"border:1px solid #d7dde5; padding:10px;\">\u5e38\u89c1\u7684 4-bit \u91cf\u5316\u7248\u672c\uff0c\u517c\u987e\u4f53\u79ef\u3001\u901f\u5ea6\u548c\u8d28\u91cf\u3002<\/td>\n      <\/tr>\n      <tr>\n        <td style=\"border:1px solid #d7dde5; padding:10px;\"><code>Q5_K_M<\/code><\/td>\n        <td style=\"border:1px solid #d7dde5; padding:10px;\">5-bit \u91cf\u5316\uff0c\u8d28\u91cf\u7565\u9ad8\uff0c\u4f53\u79ef\u548c\u901f\u5ea6\u8981\u6c42\u4e5f\u66f4\u9ad8\u3002<\/td>\n      <\/tr>\n      <tr>\n        <td style=\"border:1px solid #d7dde5; padding:10px;\"><code>Q6_K<\/code><\/td>\n        <td style=\"border:1px solid #d7dde5; padding:10px;\">6-bit \u91cf\u5316\uff0c\u66f4\u63a5\u8fd1\u539f\u59cb\u7cbe\u5ea6\uff0c\u4f46\u786c\u4ef6\u538b\u529b\u66f4\u5927\u3002<\/td>\n      <\/tr>\n      <tr>\n        <td style=\"border:1px solid #d7dde5; padding:10px;\"><code>Q8_0<\/code><\/td>\n        <td style=\"border:1px solid #d7dde5; padding:10px;\">8-bit \u91cf\u5316\uff0c\u7cbe\u5ea6\u9ad8\uff0c\u6a21\u578b\u6587\u4ef6\u4e5f\u66f4\u5927\u3002<\/td>\n      <\/tr>\n      <tr>\n        <td style=\"border:1px solid #d7dde5; padding:10px;\"><code>F16<\/code><\/td>\n        <td style=\"border:1px solid #d7dde5; padding:10px;\">\u534a\u7cbe\u5ea6\u539f\u59cb\u6a21\u578b\uff0c\u4f53\u79ef\u901a\u5e38\u975e\u5e38\u5927\u3002<\/td>\n      <\/tr>\n    <\/tbody>\n  <\/table>\n\n  <h3>\u65b9\u6cd5\u4e8c\uff1a\u4f7f\u7528 ollama list \u67e5\u770b\u6a21\u578b\u5927\u5c0f<\/h3>\n\n  <p>\u6267\u884c\u4ee5\u4e0b\u547d\u4ee4\uff1a<\/p>\n\n  <pre style=\"background:#1e1e1e; color:#e6e6e6; padding:16px; overflow:auto; border-radius:6px;\"><code>ollama list<\/code><\/pre>\n\n  <p>\n    \u8f93\u51fa\u4f1a\u5217\u51fa\u672c\u5730\u5df2\u6709\u6a21\u578b\u53ca\u5176\u5360\u7528\u7a7a\u95f4\u3002\u901a\u8fc7\u6587\u4ef6\u5927\u5c0f\u4e5f\u80fd\u5927\u81f4\u5224\u65ad\u6a21\u578b\u662f\u5426\u91cf\u5316\uff0c\u4ee5\u53ca\u91cf\u5316\u7a0b\u5ea6\u3002\n  <\/p>\n\n  <p>\n    \u4ee5 27B \u7ea7\u522b\u6a21\u578b\u4e3a\u4f8b\uff0c\u4e0d\u540c\u7cbe\u5ea6\u4e0b\u7684\u5927\u81f4\u4f53\u79ef\u53ef\u4ee5\u8fd9\u6837\u7406\u89e3\uff1a\n  <\/p>\n\n  <table style=\"width:100%; border-collapse:collapse; margin:24px 0;\">\n    <thead>\n      <tr style=\"background:#f0f3f7;\">\n        <th style=\"border:1px solid #d7dde5; padding:10px; text-align:left;\">\u6a21\u578b\u5927\u5c0f<\/th>\n        <th style=\"border:1px solid #d7dde5; padding:10px; text-align:left;\">\u53ef\u80fd\u7cbe\u5ea6<\/th>\n        <th style=\"border:1px solid #d7dde5; padding:10px; text-align:left;\">\u8bf4\u660e<\/th>\n      <\/tr>\n    <\/thead>\n    <tbody>\n      <tr>\n        <td style=\"border:1px solid #d7dde5; padding:10px;\">\u7ea6 16GB<\/td>\n        <td style=\"border:1px solid #d7dde5; padding:10px;\">Q4_K_M \/ Q4_0<\/td>\n        <td style=\"border:1px solid #d7dde5; padding:10px;\">Ollama \u9ed8\u8ba4\u5e38\u89c1\u6863\u4f4d\uff0c\u4f53\u79ef\u548c\u6548\u679c\u8f83\u5e73\u8861\u3002<\/td>\n      <\/tr>\n      <tr>\n        <td style=\"border:1px solid #d7dde5; padding:10px;\">\u7ea6 19GB<\/td>\n        <td style=\"border:1px solid #d7dde5; padding:10px;\">Q5_K_M<\/td>\n        <td style=\"border:1px solid #d7dde5; padding:10px;\">\u8d28\u91cf\u7a0d\u597d\uff0c\u786c\u4ef6\u8981\u6c42\u66f4\u9ad8\u3002<\/td>\n      <\/tr>\n      <tr>\n        <td style=\"border:1px solid #d7dde5; padding:10px;\">\u7ea6 22GB<\/td>\n        <td style=\"border:1px solid #d7dde5; padding:10px;\">Q6_K<\/td>\n        <td style=\"border:1px solid #d7dde5; padding:10px;\">\u63a5\u8fd1\u9ad8\u7cbe\u5ea6\uff0c\u4f46\u4f53\u79ef\u660e\u663e\u589e\u52a0\u3002<\/td>\n      <\/tr>\n      <tr>\n        <td style=\"border:1px solid #d7dde5; padding:10px;\">\u7ea6 29GB<\/td>\n        <td style=\"border:1px solid #d7dde5; padding:10px;\">Q8_0<\/td>\n        <td style=\"border:1px solid #d7dde5; padding:10px;\">\u7cbe\u5ea6\u8f83\u9ad8\uff0c\u9002\u5408\u663e\u5b58\u5145\u8db3\u7684\u7528\u6237\u3002<\/td>\n      <\/tr>\n      <tr>\n        <td style=\"border:1px solid #d7dde5; padding:10px;\">\u7ea6 54GB<\/td>\n        <td style=\"border:1px solid #d7dde5; padding:10px;\">F16<\/td>\n        <td style=\"border:1px solid #d7dde5; padding:10px;\">\u672a\u91cf\u5316\u534a\u7cbe\u5ea6\u6a21\u578b\uff0c\u666e\u901a\u8bbe\u5907\u5f88\u96be\u6d41\u7545\u8fd0\u884c\u3002<\/td>\n      <\/tr>\n    <\/tbody>\n  <\/table>\n\n  <p>\n    \u5982\u679c\u4f60\u770b\u5230 <code>qwen3.5:27b<\/code> \u7684\u4f53\u79ef\u5728 16GB \u5230 20GB \u5de6\u53f3\uff0c\u57fa\u672c\u53ef\u4ee5\u5224\u65ad\u5b83\u662f\u91cf\u5316\u6a21\u578b\uff0c\u800c\u4e0d\u662f\u539f\u59cb FP16 \u6a21\u578b\u3002\n  <\/p>\n\n  <h2>\u4e94\u3001\u91cf\u5316\u6a21\u578b\u4e3a\u4ec0\u4e48\u8fd8\u662f\u6162\uff1f<\/h2>\n\n  <p>\n    \u5f88\u591a\u4eba\u4f1a\u8bef\u4ee5\u4e3a\u201c\u91cf\u5316 = \u5feb\u201d\u3002\u5b9e\u9645\u4e0a\uff0c\u91cf\u5316\u7684\u4e3b\u8981\u4ef7\u503c\u9996\u5148\u662f\u964d\u4f4e\u5185\u5b58\/\u663e\u5b58\u5360\u7528\uff0c\u5176\u6b21\u624d\u53ef\u80fd\u63d0\u5347\u901f\u5ea6\u3002\n  <\/p>\n\n  <p>\n    \u5bf9\u4e8e\u672c\u5730\u5927\u6a21\u578b\u6765\u8bf4\uff0c\u901f\u5ea6\u74f6\u9888\u5f80\u5f80\u4e0d\u662f\u6a21\u578b\u6709\u6ca1\u6709\u91cf\u5316\uff0c\u800c\u662f\u4ee5\u4e0b\u51e0\u4e2a\u56e0\u7d20\uff1a\n  <\/p>\n\n  <ul>\n    <li><strong>\u662f\u5426\u5b8c\u5168\u8fd0\u884c\u5728 GPU \u4e0a<\/strong>\uff1a\u5982\u679c\u90e8\u5206\u5c42\u6216\u5168\u90e8\u8ba1\u7b97\u843d\u5728 CPU \u4e0a\uff0c\u901f\u5ea6\u4f1a\u660e\u663e\u4e0b\u964d\u3002<\/li>\n    <li><strong>\u663e\u5b58\u662f\u5426\u8db3\u591f<\/strong>\uff1a\u6a21\u578b\u6743\u91cd\u3001\u4e0a\u4e0b\u6587 KV Cache\u3001\u5de5\u5177\u8fd4\u56de\u5185\u5bb9\u90fd\u4f1a\u5360\u7528\u663e\u5b58\u3002<\/li>\n    <li><strong>\u5185\u5b58\u5e26\u5bbd\u662f\u5426\u8db3\u591f<\/strong>\uff1a\u672c\u5730\u63a8\u7406\u5c24\u5176\u662f CPU \u63a8\u7406\uff0c\u7ecf\u5e38\u53d7\u5185\u5b58\u5e26\u5bbd\u9650\u5236\u3002<\/li>\n    <li><strong>\u4e0a\u4e0b\u6587\u957f\u5ea6\u662f\u5426\u8fc7\u957f<\/strong>\uff1a\u957f\u6587\u672c\u8f93\u5165\u4f1a\u663e\u8457\u589e\u52a0\u63a8\u7406\u538b\u529b\u3002<\/li>\n    <li><strong>\u751f\u6210 token \u6570\u91cf\u662f\u5426\u8fc7\u591a<\/strong>\uff1a\u56de\u7b54\u8d8a\u957f\uff0c\u7b49\u5f85\u65f6\u95f4\u8d8a\u4e45\u3002<\/li>\n  <\/ul>\n\n  <p>\n    \u6362\u53e5\u8bdd\u8bf4\uff0c\u4e00\u4e2a 27B \u7684 Q4_K_M \u6a21\u578b\u867d\u7136\u6bd4 FP16 \u5c0f\u5f88\u591a\uff0c\u4f46\u5b83\u4ecd\u7136\u662f\u4e00\u4e2a\u5927\u578b\u6a21\u578b\u3002\u5982\u679c\u4f60\u7684\u8bbe\u5907\u663e\u5b58\u4e0d\u8db3\uff0c\u6216\u8005\u4e3b\u8981\u4f9d\u8d56 CPU \u63a8\u7406\uff0c\u5b83\u4f9d\u7136\u53ef\u80fd\u9700\u8981\u5341\u51e0\u5206\u949f\u624d\u80fd\u751f\u6210\u4e00\u6bb5\u8f83\u957f\u56de\u7b54\u3002\n  <\/p>\n\n  <h2>\u516d\u3001\u5982\u4f55\u9a8c\u8bc1 Ollama \u7684\u771f\u5b9e\u63a8\u7406\u901f\u5ea6\uff1f<\/h2>\n\n  <p>\n    \u4f60\u53ef\u4ee5\u4f7f\u7528 verbose \u6a21\u5f0f\u8fd0\u884c\u6a21\u578b\uff1a\n  <\/p>\n\n  <pre style=\"background:#1e1e1e; color:#e6e6e6; padding:16px; overflow:auto; border-radius:6px;\"><code>ollama run qwen3.5:27b --verbose<\/code><\/pre>\n\n  <p>\n    \u5728\u56de\u7b54\u5b8c\u6210\u540e\uff0c\u7ec8\u7aef\u4e2d\u901a\u5e38\u4f1a\u663e\u793a\u4e00\u4e9b\u7edf\u8ba1\u4fe1\u606f\uff0c\u4f8b\u5982\u751f\u6210 token \u6570\u91cf\u3001\u603b\u8017\u65f6\u3001token\/s \u7b49\u3002\n  <\/p>\n\n  <p>\n    \u5982\u679c\u4f60\u770b\u5230\u7c7b\u4f3c <code>eval rate: 3 tokens\/s<\/code> \u8fd9\u6837\u7684\u8f93\u51fa\uff0c\u5c31\u610f\u5473\u7740\u6a21\u578b\u6bcf\u79d2\u5927\u7ea6\u53ea\u80fd\u751f\u6210 3 \u4e2a token\u3002\u6b64\u65f6\u5982\u679c\u6700\u7ec8\u56de\u7b54\u6709 2000 \u4e2a token\uff0c\u751f\u6210\u65f6\u95f4\u63a5\u8fd1 10 \u5206\u949f\u5c31\u5f88\u6b63\u5e38\u3002\n  <\/p>\n\n  <p>\n    \u53e6\u5916\uff0c\u53ef\u4ee5\u4f7f\u7528\u4ee5\u4e0b\u547d\u4ee4\u67e5\u770b\u5f53\u524d\u6a21\u578b\u8fd0\u884c\u72b6\u6001\uff1a\n  <\/p>\n\n  <pre style=\"background:#1e1e1e; color:#e6e6e6; padding:16px; overflow:auto; border-radius:6px;\"><code>ollama ps<\/code><\/pre>\n\n  <p>\n    \u5982\u679c\u8f93\u51fa\u4e2d\u663e\u793a\u6a21\u578b\u5927\u91cf\u4f7f\u7528 CPU\uff0c\u6216\u8005 GPU\/CPU \u6df7\u5408\u8fd0\u884c\uff0c\u90a3\u4e48\u901f\u5ea6\u6162\u7684\u539f\u56e0\u901a\u5e38\u5c31\u6bd4\u8f83\u660e\u786e\u4e86\u3002\n  <\/p>\n\n  <h2>\u4e03\u3001\u4f18\u5316\u5efa\u8bae\uff1a\u5982\u4f55\u8ba9\u56de\u7b54\u66f4\u5feb\uff1f<\/h2>\n\n  <h3>1. \u6362\u7528\u66f4\u5c0f\u53c2\u6570\u6a21\u578b<\/h3>\n\n  <p>\n    \u5982\u679c\u4f60\u7684\u76ee\u6807\u662f\u6d41\u7545\u4f53\u9a8c\uff0c\u800c\u4e0d\u662f\u6781\u9650\u6a21\u578b\u80fd\u529b\uff0c\u53ef\u4ee5\u5c1d\u8bd5 7B \u6216 14B \u7248\u672c\uff1a\n  <\/p>\n\n  <pre style=\"background:#1e1e1e; color:#e6e6e6; padding:16px; overflow:auto; border-radius:6px;\"><code>ollama run qwen3.5:7b\nollama run qwen3.5:14b<\/code><\/pre>\n\n  <p>\n    \u5bf9\u4e8e\u666e\u901a\u7b14\u8bb0\u672c\u6216\u4e2d\u7aef\u53f0\u5f0f\u673a\uff0c7B \u548c 14B \u6a21\u578b\u901a\u5e38\u6bd4 27B \u6a21\u578b\u66f4\u5bb9\u6613\u8fbe\u5230\u53ef\u7528\u901f\u5ea6\u3002\n  <\/p>\n\n  <h3>2. \u68c0\u67e5\u662f\u5426\u5b8c\u5168\u4f7f\u7528 GPU<\/h3>\n\n  <p>\n    \u5982\u679c\u4f60\u6709\u5927\u663e\u5b58 GPU\uff0c\u786e\u4fdd Ollama \u80fd\u591f\u8bc6\u522b\u5e76\u4f7f\u7528 GPU\u3002\u53ef\u4ee5\u901a\u8fc7 <code>ollama ps<\/code> \u67e5\u770b\u6a21\u578b\u662f\u5426\u4e3b\u8981\u8fd0\u884c\u5728 GPU \u4e0a\u3002\n  <\/p>\n\n  <h3>3. \u63a7\u5236\u4e0a\u4e0b\u6587\u957f\u5ea6<\/h3>\n\n  <p>\n    \u5728 Agent\/RAG \u573a\u666f\u4e2d\uff0c\u5c3d\u91cf\u51cf\u5c11\u4e0d\u5fc5\u8981\u7684\u6587\u4ef6\u5185\u5bb9\u548c\u68c0\u7d22\u7ed3\u679c\u3002\u4e0d\u8981\u628a\u6240\u6709\u76f8\u5173\u6587\u4ef6\u90fd\u585e\u8fdb\u4e0a\u4e0b\u6587\uff0c\u800c\u5e94\u8ba9\u5de5\u5177\u8fd4\u56de\u66f4\u7cbe\u786e\u7684\u7247\u6bb5\u3002\n  <\/p>\n\n  <h3>4. \u964d\u4f4e\u6700\u5927\u8f93\u51fa\u957f\u5ea6<\/h3>\n\n  <p>\n    \u5982\u679c\u6a21\u578b\u7ecf\u5e38\u751f\u6210\u975e\u5e38\u957f\u7684\u56de\u7b54\uff0c\u53ef\u4ee5\u9002\u5f53\u9650\u5236\u6700\u5927\u8f93\u51fa token \u6570\u3002\u8f83\u77ed\u7684\u56de\u7b54\u901a\u5e38\u4f1a\u66f4\u5feb\u5b8c\u6210\u3002\n  <\/p>\n\n  <h3>5. \u663e\u5b58\u4e0d\u8db3\u65f6\u4e0d\u8981\u5f3a\u884c\u4e0a 27B<\/h3>\n\n  <p>\n    27B \u6a21\u578b\u5bf9\u786c\u4ef6\u8981\u6c42\u8f83\u9ad8\u3002\u5982\u679c\u4f60\u7684\u663e\u5b58\u53ea\u6709 8GB\u300112GB \u6216 16GB\uff0c\u8fd0\u884c 27B \u6a21\u578b\u65f6\u5f88\u5bb9\u6613\u9047\u5230 CPU \u56de\u9000\u3001\u5185\u5b58\u4ea4\u6362\u548c\u901f\u5ea6\u9aa4\u964d\u7684\u95ee\u9898\u3002\n  <\/p>\n\n  <h2>\u516b\u3001\u7ed3\u8bba<\/h2>\n\n  <p>\n    \u4f60\u770b\u5230\u7684\u201c\u601d\u8003\u7528\u65f6 3 \u79d2\u201d\u201c\u601d\u8003\u7528\u65f6 7 \u79d2\u201d\u7b49\u6570\u5b57\uff0c\u5f88\u53ef\u80fd\u53ea\u662f\u5de5\u5177\u8c03\u7528\u6216\u5c40\u90e8\u6b65\u9aa4\u7684\u8017\u65f6\uff0c\u800c\u4e0d\u662f\u6a21\u578b\u751f\u6210\u5b8c\u6574\u56de\u7b54\u7684\u603b\u65f6\u95f4\u3002\u771f\u6b63\u6d88\u8017\u65f6\u95f4\u7684\uff0c\u5f80\u5f80\u662f\u6a21\u578b\u8bfb\u53d6\u4e0a\u4e0b\u6587\u540e\u8fdb\u884c token \u751f\u6210\u7684\u9636\u6bb5\u3002\n  <\/p>\n\n  <p>\n    \u5728 Ollama \u4e0a\u76f4\u63a5\u4e0b\u8f7d\u7684 <code>qwen3.5:27b<\/code> \u5927\u6982\u7387\u5df2\u7ecf\u662f\u91cf\u5316\u7248\u672c\uff0c\u5e38\u89c1\u53ef\u80fd\u662f <code>Q4_K_M<\/code> \u6216\u7c7b\u4f3c\u7b49\u7ea7\u3002\u4f60\u53ef\u4ee5\u901a\u8fc7\uff1a\n  <\/p>\n\n  <pre style=\"background:#1e1e1e; color:#e6e6e6; padding:16px; overflow:auto; border-radius:6px;\"><code>ollama show qwen3.5:27b\nollama list<\/code><\/pre>\n\n  <p>\n    \u6765\u786e\u8ba4\u6a21\u578b\u7684\u91cf\u5316\u7c7b\u578b\u548c\u6587\u4ef6\u5927\u5c0f\u3002\n  <\/p>\n\n  <p>\n    \u5982\u679c\u6a21\u578b\u5df2\u7ecf\u91cf\u5316\u4f46\u4f9d\u7136\u9700\u8981 15 \u5206\u949f\u624d\u56de\u7b54\uff0c\u8bf4\u660e\u74f6\u9888\u5927\u6982\u7387\u4e0d\u5728\u5de5\u5177\u8c03\u7528\uff0c\u800c\u5728\u672c\u5730\u63a8\u7406\u786c\u4ef6\u3001\u663e\u5b58\u5360\u7528\u3001\u4e0a\u4e0b\u6587\u957f\u5ea6\u6216\u6a21\u578b\u89c4\u6a21\u672c\u8eab\u3002\u5bf9\u4e8e\u5927\u591a\u6570\u666e\u901a\u8bbe\u5907\u6765\u8bf4\uff0c27B \u6a21\u578b\u5e76\u4e0d\u662f\u8f7b\u677e\u53ef\u4ee5\u201c\u79d2\u56de\u201d\u7684\u89c4\u6a21\u3002\u5982\u679c\u8ffd\u6c42\u6d41\u7545\u4f53\u9a8c\uff0c\u5efa\u8bae\u4f18\u5148\u8003\u8651 7B \u6216 14B \u7248\u672c\uff0c\u6216\u8005\u5347\u7ea7\u5230\u663e\u5b58\u66f4\u5145\u8db3\u7684 GPU \u73af\u5883\u3002\n  <\/p>\n<\/article>\n","protected":false},"excerpt":{"rendered":"<p>\u5728\u672c\u5730\u4f7f\u7528 Ollama \u8fd0\u884c Qwen3.5:2 [&#8230;]<\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[89],"tags":[],"class_list":["post-3643","post","type-post","status-publish","format-standard","hentry","category-live"],"_links":{"self":[{"href":"https:\/\/blog.zengqq.com.cn\/index.php?rest_route=\/wp\/v2\/posts\/3643","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/blog.zengqq.com.cn\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/blog.zengqq.com.cn\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/blog.zengqq.com.cn\/index.php?rest_route=\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/blog.zengqq.com.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=3643"}],"version-history":[{"count":5,"href":"https:\/\/blog.zengqq.com.cn\/index.php?rest_route=\/wp\/v2\/posts\/3643\/revisions"}],"predecessor-version":[{"id":3651,"href":"https:\/\/blog.zengqq.com.cn\/index.php?rest_route=\/wp\/v2\/posts\/3643\/revisions\/3651"}],"wp:attachment":[{"href":"https:\/\/blog.zengqq.com.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=3643"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/blog.zengqq.com.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=3643"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/blog.zengqq.com.cn\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=3643"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}