{"id":574,"date":"2026-07-30T20:04:20","date_gmt":"2026-07-30T12:04:20","guid":{"rendered":"http:\/\/www.tantan.ink\/index.php\/2026\/07\/30\/mac-%e4%b8%8a%e8%bf%98%e6%9c%89%e6%af%94-ollama-%e6%9b%b4%e5%bf%ab%e7%9a%84%e6%9c%ac%e5%9c%b0%e6%8e%a8%e7%90%86%e5%bc%95%e6%93%8e\/"},"modified":"2026-07-30T20:04:20","modified_gmt":"2026-07-30T12:04:20","slug":"mac-%e4%b8%8a%e8%bf%98%e6%9c%89%e6%af%94-ollama-%e6%9b%b4%e5%bf%ab%e7%9a%84%e6%9c%ac%e5%9c%b0%e6%8e%a8%e7%90%86%e5%bc%95%e6%93%8e","status":"publish","type":"post","link":"https:\/\/www.tantan.ink\/index.php\/2026\/07\/30\/mac-%e4%b8%8a%e8%bf%98%e6%9c%89%e6%af%94-ollama-%e6%9b%b4%e5%bf%ab%e7%9a%84%e6%9c%ac%e5%9c%b0%e6%8e%a8%e7%90%86%e5%bc%95%e6%93%8e\/","title":{"rendered":"Mac \u4e0a\u8fd8\u6709\u6bd4 Ollama \u66f4\u5feb\u7684\u672c\u5730\u63a8\u7406\u5f15\u64ce"},"content":{"rendered":"

\u6211\u684c\u9762\u4e0a\u4e00\u76f4\u8dd1\u7740 Ollama\u2014\u2014\u5b83\u7b80\u5355\u3001\u8de8\u5e73\u53f0\u3001\u6a21\u578b\u4ed3\u5e93\u5b8c\u5584\uff0c\u662f\u672c\u5730\u6a21\u578b\u7684\u4e8b\u5b9e\u5165\u53e3\u3002\u4f46\u7528\u4e86\u51e0\u5e74\uff0c\u5fc3\u91cc\u4e00\u76f4\u6709\u4e2a\u5c0f\u7599\u7629\uff1a\u5728 Apple Silicon \u4e0a\uff0c\u5b83\u603b\u611f\u89c9\u5dee\u90a3\u4e48\u4e00\u53e3\u6c14\u3002\u6253\u5f00\u6d3b\u52a8\u76d1\u89c6\u5668\u770b\uff0cCPU \u5360\u7528\u4e0d\u9ad8\uff0c\u4f46\u7b2c\u4e00\u4e2a token \u51fa\u6765\u8981\u7b49\u4e00\u4e24\u79d2\uff0c\u957f\u4e00\u70b9\u7684\u56de\u7b54\u8089\u773c\u80fd\u770b\u51fa”\u4e00\u6bb5\u4e00\u6bb5\u8e66”\u3002<\/p>\n

\u524d\u6bb5\u65f6\u95f4\u7ffb GitHub \u7ffb\u5230\u4e00\u4e2a\u9879\u76ee\u53eb Rapid-MLX<\/strong>\u3002\u4f5c\u8005 Raullen Chai \u628a\u5b83\u5b9a\u4f4d\u6210”Apple Silicon \u4e0a\u6700\u5feb\u7684\u672c\u5730 AI \u5f15\u64ce”\uff0c\u4ed3\u5e93\u63cf\u8ff0\u76f4\u63a5\u5199”4.2x faster than Ollama”\u3002\u5148\u770b\u51e0\u4e2a\u786c\u6307\u6807\uff1a3.4k stars\uff08\u6700\u8fd1\u4e00\u5468 +125\uff09\uff0c1,269 \u4e2a commits\uff0c185 \u4e2a tags\uff0cApache 2.0 \u534f\u8bae\u3002\u8fd9\u4e2a\u6708\u8fdb\u4e86 Homebrew core\uff0c\u6700\u8fd1\u4e00\u6b21\u63d0\u4ea4\u5c31\u5728\u4eca\u5929\u3002\u6d3b\u8dc3\u5ea6\u548c\u4f53\u91cf\u90fd\u591f\u5f97\u4e0a”\u53ef\u4ee5\u8ba4\u771f\u770b”\u3002<\/p>\n

\"github\"<\/p>\n

\u6211\u6bd4\u8f83\u611f\u5174\u8da3\u7684\u662f\u5b83\u7684\u5b9e\u73b0\u601d\u8def\u3002Ollama \u548c llama.cpp \u672c\u8d28\u4e0a\u662f C\/C++ \u5199\u7684\u901a\u7528\u5f15\u64ce\uff0c\u5230\u4e86 Apple Silicon \u8d70\u4e00\u5c42 Metal \u9002\u914d\u3002Rapid-MLX \u53cd\u8fc7\u6765\u2014\u2014\u76f4\u63a5\u7528 Apple \u5b98\u65b9\u7684 MLX \u6846\u67b6\uff0cpure MLX kernels\uff0c\u6ca1\u6709 llama.cpp fallback\uff0c\u4e5f\u6ca1\u6709 Metal shim\u3002\u5bf9\u7edf\u4e00\u5185\u5b58\u67b6\u6784\u7684 Mac \u6765\u8bf4\uff0c\u5c11\u4e00\u5c42\u62bd\u8c61\u610f\u5473\u7740\u5c11\u4e00\u6b21\u5185\u5b58\u62f7\u8d1d\u548c\u683c\u5f0f\u8f6c\u6362\uff0c\u8fd9\u662f\u5b83”\u5feb”\u7684\u7269\u7406\u57fa\u7840\u3002<\/p>\n

\u88c5\u5b83\u5f88\u7b80\u5355\u3002Homebrew \u5df2\u7ecf\u6536\u5f55\uff1a<\/p>\n

brew install rapid-mlx<\/code><\/pre>\n
\u4e0d\u559c\u6b22 brew \u7684\uff0c\u5b98\u65b9\u8fd8\u6709 curl -fsSL https:\/\/rapidmlx.com\/install.sh | bash<\/code> \u7684\u4e00\u952e\u811a\u672c\uff0c\u81ea\u52a8\u5efa venv \u5728 ~\/.rapid-mlx\/<\/code>\u3001\u68c0\u6d4b\u5185\u5b58\u9009\u8d77\u6b65\u6a21\u578b\uff0c\u8fde sudo \u90fd\u4e0d\u9700\u8981\u3002\u6574\u4e2a\u5b89\u88c5\u5305\u4e0d\u5230 500MB\u3002<\/p>\n
\u88c5\u5b8c\u76f4\u63a5 serve\uff1a<\/p>\n
rapid-mlx serve qwen3.5-4b-4bit<\/code><\/pre>\n
\u7b2c\u4e00\u6b21\u542f\u52a8\u4f1a\u4e0b\u8f7d\u6a21\u578b\u6743\u91cd\uff084B \u7684 4bit \u5927\u7ea6 2.5GB\uff09\uff0c\u8fdb\u5ea6\u6761\u8dd1\u5b8c\u5c31\u8d77\u6765\u4e86\u3002\u76d1\u542c http:\/\/localhost:8000<\/code>\uff0cOpenAI \u517c\u5bb9\u534f\u8bae\uff0ccurl\u3001Python \u5ba2\u6237\u7aef\u3001\u5404\u79cd UI \u5de5\u5177\u968f\u4fbf\u63a5\u3002<\/p>\n
\"homepage\"<\/p>\n
\u4f53\u611f\u5dee\u5f02\u662f\u660e\u786e\u7684\u3002\u540c\u6837\u7684 4B \u6a21\u578b\uff0cOllama \u4e4b\u524d\u5728\u6211\u8fd9\u53f0\u673a\u5668\u8dd1\u5927\u6982 50 \u591a tok\/s\uff0cRapid-MLX \u5e72\u5230 130+ tok\/s\uff0c\u5dee\u4e86\u4e00\u500d\u591a\u3002\u5b98\u65b9\u7684\u5bf9\u6bd4\u56fe\u6211\u653e\u4e0b\u9762\u2014\u2014\u8fd9\u662f M3 Ultra 256GB \u4e0a\u7684\u6570\u636e\uff0c\u6211\u624b\u91cc\u8fd9\u53f0 16GB \u8dd1\u4e0d\u5230\u90a3\u4e48\u9ad8\uff0c\u4f46\u540c\u6863\u5bf9\u6bd4\u7684\u76f8\u5bf9\u5173\u7cfb\u662f\u4e00\u81f4\u7684\uff1a<\/p>\n
\"benchmark\"<\/p>\n
Phi-4 Mini 14B \u5728 M3 Ultra \u4e0a\u662f 3.2x\uff0cQwen3.5-9B \u662f 2.6x\uff0cGemma 4 12B \u662f 1.5x\u3002\u5e73\u5747\u4e0b\u6765”2 \u5230 4 \u500d\u66f4\u5feb”\u8fd9\u4e2a\u8bf4\u6cd5\u7ad9\u5f97\u4f4f\u3002<\/p>\n
\u6709\u51e0\u4e2a\u5de5\u7a0b\u7ec6\u8282\u8ba9”Ollama \u5076\u5c14\u8ba9\u6211\u76b1\u7709\u7684\u5730\u65b9”\u88ab\u9010\u4e2a\u5904\u7406\u4e86\uff1a<\/p>\n
    \n
  • tool calling \u771f\u7684\u7a33\u4e86<\/strong>\uff1a\u5185\u7f6e 17 \u79cd parser\uff0c\u91cf\u5316\u6a21\u578b\u8f93\u51fa\u683c\u5f0f\u98d8\u4e86\u4f1a\u81ea\u52a8\u6062\u590d\u6210\u7ed3\u6784\u5316\u3002\u672c\u5730\u8dd1 agent \u6700\u6015\u7684\u5c31\u662f\u5de5\u5177\u8c03\u7528\u683c\u5f0f\u574f\u6389\u3001\u6574\u4e2a\u6d41\u7a0b\u91cd\u6765\u4e00\u904d\uff0c\u8fd9\u70b9\u5bf9\u5de5\u7a0b\u4f53\u9a8c\u6539\u5584\u5f88\u5927\u3002<\/li>\n
  • prompt cache \u662f\u53cc\u5c42\u7684<\/strong>\uff1aradix \u6811\u524d\u7f00\u53bb\u91cd + DeltaNet RNN \u72b6\u6001\u5feb\u7167\u3002Qwen3.5 \u8fd9\u79cd\u6df7\u5408\u67b6\u6784\uff08\u90e8\u5206\u5c42\u662f RNN\uff09\u4e5f\u80fd\u547d\u4e2d\u7f13\u5b58\uff0c\u7f13\u5b58\u547d\u4e2d\u540e\u9996 token \u5ef6\u8fdf 0.08 \u79d2\u3002\u591a\u8f6e\u5bf9\u8bdd\u590d\u7528\u540c\u4e00\u4e2a system prompt \u65f6\u4f53\u611f\u5dee\u5f02\u660e\u663e\u3002<\/li>\n
  • \u91cf\u5316 KV \u7f13\u5b58<\/strong>\uff1aTurboQuant K8V4\uff08K \u7528 8bit\uff0cV \u7528 4bit\uff09\uff0c\u8fde\u7eed\u6279\u5904\u7406\u65f6\u8fdb\u4e00\u6b65\u7701\u5185\u5b58\u3002\u591a\u7528\u6237\u5e76\u53d1\u573a\u666f\u4e0b\u541e\u5410\u80fd\u6491\u4f4f\u3002<\/li>\n<\/ul>\n
    \u751f\u6001\u63a5\u5f97\u5f88\u5bbd\u3002191 \u4e2a\u6a21\u578b\u522b\u540d\u300115 \u4e2a\u5bb6\u65cf<\/strong>\uff0c\u89c6\u89c9\u3001\u97f3\u9891\uff08TTS + STT\uff09\u3001embedding \u90fd\u6709\uff0c\u89c6\u9891\u751f\u6210\uff08LTX-2.3\uff09\u4e5f\u4e0a\u7ebf\u4e86\u300211 \u4e2a agent CLI \u548c 3 \u4e2a Python \u6846\u67b6\u5b98\u65b9\u7ef4\u62a4”\u7ebf\u7f06\u517c\u5bb9”\u2014\u2014\u5b9a\u671f\u8dd1\u771f\u5b9e\u6743\u91cd\u7684\u7aef\u5230\u7aef\u9a8c\u8bc1\u3002\u6211\u7528\u672c\u5730\u51e0\u4e2a OpenAI \u534f\u8bae\u7684\u5ba2\u6237\u7aef\u63a5\u8fc7\u53bb\uff0c\u96f6\u4fee\u6539\u76f4\u63a5\u7528\uff0c\u6ca1\u6709\u4e00\u4e2a\u51fa\u73b0\u534f\u8bae\u4e0d\u517c\u5bb9\u7684\u95ee\u9898\u3002<\/p>\n
    \u4f46\u6709\u4ef6\u4e8b\u5f97\u8bf4\u6e05\u695a\uff1a16GB \u5185\u5b58\u5c31\u662f 16GB \u7684\u5929\u82b1\u677f<\/strong>\u3002<\/p>\n
    \u6211\u65e5\u5e38\u5e38\u7528\u7684\u90a3\u4e2a 35B \u7ea7 MoE \u6a21\u578b\uff0cRapid-MLX \u5b98\u65b9\u5efa\u8bae\u662f 48GB \u8d77\u6b65\uff0c\u6211\u6839\u672c\u88c5\u4e0d\u4e0a\u53bb\u3002\u8981\u5728\u8fd9\u79cd\u5185\u5b58\u4e0b\u8dd1\u90a3\u79cd\u89c4\u6a21\u7684\u6a21\u578b\uff0c\u8fd8\u5f97\u56de\u5230 llama.cpp \u8def\u7ebf\uff0c\u9760 1.58bit\u30012-bit \u8fd9\u79cd\u6781\u7aef\u91cf\u5316\u628a\u6a21\u578b\u538b\u8fdb 16GB\uff0c\u4ed8\u51fa\u660e\u663e\u7684\u901f\u5ea6\u635f\u5931\u3002\u8fd9\u662f\u7269\u7406\u9650\u5236\uff0c\u4e0d\u662f\u5f15\u64ce\u7684\u95ee\u9898\u3002<\/p>\n
    \"models\"<\/p>\n
    \u5b98\u65b9\u90a3\u5f20”\u54ea\u4e2a Mac \u8dd1\u54ea\u4e2a\u6a21\u578b”\u7684\u5bf9\u7167\u8868\u5f88\u5b9e\u5728\uff1a16GB \u2192 4B \u2192 147 tok\/s\uff0c24GB \u2192 9B \u2192 101 tok\/s\uff0c48GB \u2192 35B-A3B 8bit \u2192 80 tok\/s\uff0c96GB \u2192 122B \u2192 43 tok\/s\u3002\u68af\u5ea6\u62c9\u5f97\u5f88\u5f00\uff0c\u6bcf\u4e00\u6863\u90fd\u6709\u5408\u9002\u7684\u6a21\u578b\u3002<\/p>\n
    \u6240\u4ee5\u73b0\u5728\u6211\u684c\u9762\u662f\u4e24\u4e2a\u8fdb\u7a0b\u5e76\u5b58\uff1a\u5199\u4ee3\u7801\u3001\u4e34\u65f6\u95ee\u95ee\u9898\u4e22\u7ed9 Rapid-MLX \u7684 4B \u6a21\u578b\uff0c\u79d2\u56de\uff1b\u78b0\u5230\u9700\u8981\u590d\u6742\u63a8\u7406\u3001\u957f\u4e0a\u4e0b\u6587\uff0c\u4e34\u65f6\u8d77\u4e00\u4e2a llama-server \u8dd1\u91cf\u5316\u540e\u7684 35B\uff0c\u5404\u53d6\u6240\u957f\u3002\u4e24\u8005\u63a5\u7684\u90fd\u662f OpenAI \u534f\u8bae\uff0c\u5207\u6362\u5ba2\u6237\u7aef\u914d\u7f6e\u4e00\u884c\u7684\u4e8b\u3002<\/p>\n
    \u8fd9\u4e0d\u662f Ollama \u7684\u66ff\u4ee3\u54c1\uff0c\u662f Apple Silicon \u4e0a\u4e13\u95e8\u7684\u4e00\u628a\u5feb\u5200\u3002Ollama \u5728\u8de8\u5e73\u53f0\u3001\u901a\u7528\u751f\u6001\u4e0a\u8fd8\u662f\u4e8b\u5b9e\u6807\u51c6\uff1b\u5982\u679c\u4f60\u7684\u4e3b\u6218\u573a\u662f Mac\u3001\u53c8\u5acc Ollama \u5dee\u53e3\u6c14\uff0cbrew install \u4e00\u53e5\u5c31\u88c5\u4e0a\u4e86\uff0c\u635f\u5931\u4e0d\u4e86\u4ec0\u4e48\u3002\u771f\u4e0d\u6ee1\u610f\uff0crm -rf ~\/.rapid-mlx\/<\/code> \u5c31\u6e05\u7406\u5e72\u51c0\uff0c\u4e0d\u5728\u7cfb\u7edf\u4e0a\u7559\u4efb\u4f55\u75d5\u8ff9\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"
    Apple Silicon \u4e0a\u6709\u4e2a\u53eb Rapid-MLX \u7684\u7eaf MLX \u672c\u5730\u63a8\u7406\u5f15\u64ce\uff0c\u5b98\u65b9\u8bf4\u6cd5\u662f\u6bd4 Ollama \u5feb 2 \u5230 4 \u500d\u3002\u6211\u5728 16GB Mac \u4e0a\u88c5\u4e86\u4e00\u628a\uff0c\u4f53\u611f\u5dee\u5f02\u660e\u663e\uff0c\u4f46 16GB \u88c5\u4e0d\u4e0b 35B\u2014\u2014\u8fd9\u662f\u4e00\u7bc7\u6298\u817e\u65e5\u5fd7\uff0c\u4e0d\u662f\u5e7f\u544a\u3002<\/p>\n","protected":false},"author":2,"featured_media":570,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[2,176],"tags":[],"class_list":["post-574","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ai","category-hands-on"],"_links":{"self":[{"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/posts\/574","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/comments?post=574"}],"version-history":[{"count":0,"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/posts\/574\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/media\/570"}],"wp:attachment":[{"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/media?parent=574"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/categories?post=574"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/tags?post=574"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}