{"id":975,"date":"2026-08-22T20:06:47","date_gmt":"2026-08-22T12:06:47","guid":{"rendered":"http:\/\/www.tantan.ink\/index.php\/2026\/08\/22\/ollama-%e9%80%89%e4%ba%86-mlx%ef%bc%8c16gb-mac-%e7%9a%84%e6%9c%ac%e5%9c%b0%e6%a8%a1%e5%9e%8b%e4%b8%8b%e4%b8%80%e7%ab%99%e5%9c%a8%e5%93%aa\/"},"modified":"2026-08-22T20:06:47","modified_gmt":"2026-08-22T12:06:47","slug":"ollama-%e9%80%89%e4%ba%86-mlx%ef%bc%8c16gb-mac-%e7%9a%84%e6%9c%ac%e5%9c%b0%e6%a8%a1%e5%9e%8b%e4%b8%8b%e4%b8%80%e7%ab%99%e5%9c%a8%e5%93%aa","status":"publish","type":"post","link":"https:\/\/www.tantan.ink\/index.php\/2026\/08\/22\/ollama-%e9%80%89%e4%ba%86-mlx%ef%bc%8c16gb-mac-%e7%9a%84%e6%9c%ac%e5%9c%b0%e6%a8%a1%e5%9e%8b%e4%b8%8b%e4%b8%80%e7%ab%99%e5%9c%a8%e5%93%aa\/","title":{"rendered":"Ollama \u9009\u4e86 MLX\uff0c16GB Mac \u7684\u672c\u5730\u6a21\u578b\u4e0b\u4e00\u7ad9\u5728\u54ea"},"content":{"rendered":"

\u4e0a\u4e2a\u6708\u6211\u628a Ollama \u5347\u5230 0.19.0\uff0c\u542f\u52a8\u540e\u53f0\u670d\u52a1\u7684\u65f6\u5019\u4e0b\u610f\u8bc6\u626b\u4e86\u4e00\u773c\u5b98\u65b9\u4ed3\u5e93\u7684 release notes\u30023 \u6708 30 \u65e5\u90a3\u6761 commit \u5199\u5f97\u975e\u5e38\u76f4\u767d\uff1a\u300cOllama is now powered by MLX on Apple Silicon in preview\u300d<\/strong>\u3002\u540c\u4e00\u5929\u5237\u51fa\u6765\u4e00\u5806 benchmark\u2014\u2014Qwen3.5-35B-A3B \u8fd9\u4e2a MoE \u6a21\u578b\u5728 Ollama \u8001\u7248\u672c\uff08llama.cpp\uff09\u4e0a\u662f 58 tok\/s \u7684 decode\uff0c\u65b0\u7248\u672c\uff08MLX\uff09\u8dd1\u5230 112\uff0c\u51e0\u4e4e\u7ffb\u500d\uff1bprefill \u4ece 1,154 \u63d0\u5230 1,810 tok\/s\u3002<\/p>\n

\u6570\u5b57\u662f\u633a\u6f02\u4eae\u7684\u3002\u4f46\u5f80\u4e0b\u770b\u5230\u4e00\u884c\uff1a\u300cMLX preview requires a Mac with more than 32 GB unified memory\u300d<\/strong>\u3002<\/p>\n

\u6211\u90a3\u53f0\u57fa\u7840\u6b3e Mac \u662f 16GB \u7edf\u4e00\u5185\u5b58\u3002\u8fd9\u610f\u5473\u7740\u65b0\u5f15\u64ce\u8ddf\u6211\u8fd9\u53f0\u6ca1\u5173\u7cfb\u3002<\/p>\n

\"banner\"<\/strong><\/p>\n

\u51ed\u4ec0\u4e48\u6362\u5f15\u64ce<\/h3>\n

Ollama \u5207 MLX \u4e0d\u662f\u4e3a\u4e86\u51d1\u70ed\u95f9\u3002Apple \u8fd9\u5957 MLX \u6846\u67b6\u662f 2023 \u5e74\u5e95\u5f00\u6e90\u3001\u4e13\u95e8\u4e3a M \u7cfb\u5217\u7edf\u4e00\u5185\u5b58\u67b6\u6784\u5199\u7684\u2014\u2014CPU \u548c GPU \u5171\u4eab\u540c\u4e00\u5757\u7269\u7406\u5185\u5b58\uff0ctensor \u5728\u4e24\u8005\u4e4b\u95f4\u4f20\u9012\u662f\u96f6\u62f7\u8d1d\u7684\uff0c\u4e0d\u8d70 PCIe \u90a3\u6761\u603b\u7ebf\u3002\u8fd9\u4ef6\u4e8b\u5bf9 LLM \u63a8\u7406\u7279\u522b\u91cd\u8981\uff0c\u56e0\u4e3a\u73b0\u5728\u6d88\u8d39\u7ea7\u786c\u4ef6\u8dd1\u672c\u5730\u6a21\u578b\uff0c\u51b3\u5b9a\u901f\u5ea6\u7684\u74f6\u9888\u662f\u5185\u5b58\u5e26\u5bbd\uff0c\u4e0d\u662f\u7b97\u529b\u3002\u5e26\u5bbd\u66f4\u5bbd\u7684\u65b9\u5411\u662f\u76f4\u63a5\u8bfb\u5199\u540c\u4e00\u5757\u5185\u5b58\uff0c\u4e0d\u662f\u5728 CPU\/GPU \u4e4b\u95f4\u642c\u6570\u636e<\/strong>\u3002<\/p>\n

llama.cpp \u662f\u53e6\u4e00\u4e2a\u601d\u8def\uff1a\u5b83\u662f C++ \u5199\u7684\u8de8\u5e73\u53f0\u63a8\u7406\u5f15\u64ce\uff0c\u4e3a\u4e86\u517c\u987e NVIDIA\u3001AMD\u3001Intel\u3001Vulkan\u3001Apple Metal \u4e94\u79cd\u540e\u7aef\uff0c\u5728\u6bcf\u4e2a\u5e73\u53f0\u4e0a\u90fd\u4ed8\u51fa\u4e00\u4e9b”\u7ffb\u8bd1\u7a0e”\u3002Metal \u540e\u7aef\u5728 M3\/M4 \u4e0a\u5bf9\u5e94 MLX \u901a\u5e38\u6162 20\u201340%\uff0c\u539f\u56e0\u662f\u5b83\u628a CUDA \u8ba1\u7b97\u56fe\u7ffb\u8bd1\u6210 Metal shader\uff0c\u6ca1\u6cd5\u5229\u7528 M5 \u91cc\u585e\u8fdb GPU \u6838\u5fc3\u7684 Neural Accelerator\u3002<\/p>\n

\u6240\u4ee5 Ollama \u8fd9\u6b65\u5207\u6362\u672c\u8d28\u4e0a\u662f\u4e00\u53e5\u8bdd\uff1a\u5728 Mac \u4e0a\u4fdd\u7559\u4e00\u6761\u517c\u5bb9\u6027 llama.cpp \u540e\u8def\uff0c\u518d\u5f00\u4e00\u6761\u539f\u751f MLX \u5feb\u901f\u8def<\/strong>\u3002\u4e24\u8def\u5e76\u5b58\uff0c\u6309\u6a21\u578b\u548c\u673a\u5668\u5c5e\u6027\u81ea\u52a8\u9009\u3002\u8fd9\u4e2a\u8bbe\u8ba1\u633a\u5e72\u51c0\u7684\u3002<\/p>\n

\u4f46\u662f\u5b83\u5728 preview \u9636\u6bb5\u753b\u4e86\u4e00\u6761\u786c\u7ebf\uff1a<\/p>\n

    \n
  • Mac \u7edf\u4e00\u5185\u5b58 \u2265 32GB<\/li>\n
  • M4 \u6216 M5 \u7cfb\u5217\u82af\u7247<\/li>\n
  • \u6a21\u578b\u5728 MLX\/safetensors \u683c\u5f0f\u6709\u6b63\u5f0f\u7248<\/li>\n
  • \u6a21\u578b\u89c4\u6a21\u5728 14B \u4ee5\u4e0a\uff0814B \u4ee5\u4e0b MLX \u4f18\u52bf\u4e0d\u660e\u663e\uff09<\/li>\n<\/ul>\n

    \u4e0d\u7b26\u5408\u4efb\u610f\u4e00\u6761\u5c31\u8d70 llama.cpp Metal \u8001\u8def\u3002\u6362\u53e5\u8bdd\u8bf4\uff0c\u6211\u8fd9\u79cd 16GB \u57fa\u7840\u6b3e Mac \u7528\u6237\uff0c\u88ab\u5b98\u65b9\u5199\u8fdb\u4e86 fallback \u4e00\u680f<\/strong>\u3002<\/p>\n

    \u9000\u56de llama.cpp \u4e4b\u540e\u6211\u770b\u5230\u4ec0\u4e48<\/h3>\n

    \u771f\u53bb\u8dd1\u7684\u65f6\u5019\uff0c\u6700\u76f4\u89c2\u7684\u4f53\u611f\u53d8\u5316\u5012\u4e0d\u662f token \u6570\uff0c\u800c\u662f \u542f\u52a8\u65f6\u95f4<\/strong>\u3002\u540c\u4e00\u4e2a 35B-A3B \u6a21\u578b\uff0c\u65b0\u7248\u672c\uff08\u5373\u4f7f\u662f fallback \u5230 llama.cpp \u7684\u65e7\u8def\u5f84\uff09\u51b7\u542f\u5230\u9996 token \u51fa\u6765\u6bd4\u4e4b\u524d\u5feb\u4e86\u96f6\u70b9\u51e0\u79d2\u3002\u8fd9\u4e2a\u6211\u76f8\u4fe1\u662f MLX \u8def\u5f84\u6539\u4e86\u4e00\u90e8\u5206\u8d44\u6e90\u8c03\u5ea6\u540e\uff0c\u6574\u5957 binary \u90fd\u91cd\u7f16\u4e86\u3002<\/p>\n

    \u53e6\u4e00\u4e2a\u4f53\u611f\u53d8\u5316\uff1aGGUF \u6a21\u578b\u4ecd\u7136\u7167\u5e38\u5de5\u4f5c<\/strong>\u3002\u6bd4\u5982\u672c\u5730\u653e\u7740\u7684 7B \u7ea7\u522b\u7684 Qwen2.5-Coder-Q4_K_M\u3001Qwen3-TTS\u3001gpt-oss-20b \u8fd9\u4e9b GGUF \u6587\u4ef6\u90fd\u6ca1\u53d7\u5f71\u54cd\uff0c\u8dd1\u5728 30-60 tok\/s \u4e4b\u95f4\uff0c\u6ca1\u6389\u901f\u4e5f\u6ca1\u5d29\u3002<\/p>\n

    \u8fd9\u8bf4\u660e Ollama \u8fd9\u6b21\u5207\u6362\u662f\u53cc\u8f68\u5236<\/strong>\uff1a\u6a21\u578b\u6587\u4ef6\u662f MLX\/safetensors \u8d70 MLX\uff1b\u6a21\u578b\u6587\u4ef6\u662f GGUF \u8d70 llama.cpp\u3002\u4e0d\u662f\u4e00\u5200\u5207\u66ff\u6362\uff0c\u8001\u751f\u6001\u6ca1\u5d29\u3002<\/p>\n

    Q4_K_M \u4e0d\u662f\u6211\u4ee5\u4e3a\u7684\u8d8a\u5927\u8d8a\u6162<\/h3>\n

    \u65e2\u7136\u88ab\u8e22\u56de llama.cpp Metal\uff0c\u90a3\u5c31\u628a\u76ee\u5149\u8f6c\u56de\u91cf\u5316\u65b9\u6848\u9009\u62e9\u4e0a\u3002\u4e00\u4e2a\u53cd\u76f4\u89c9\u7684\u4e8b\u60c5\uff1a\u5728 5060 Ti \u4e0a\u91cf\u5316\u8d8a\u5c0f\uff08Q2\u3001Q3\uff09\u901a\u5e38\u610f\u5473\u7740\u6a21\u578b\u66f4\u5c0f\u3001token \u6570\u8d8a\u9ad8\uff1b\u4f46\u5728 Apple Silicon Metal \u4e0a\uff0c\u6b63\u597d\u53cd\u8fc7\u6765\u2014\u2014Q4_K_M \u6bd4 Q2_K\u3001Q3_K_M \u90fd\u5feb<\/strong>\u3002<\/p>\n

    \"comparison\"<\/strong><\/p>\n

    \u5177\u4f53\u6570\u5b57\uff08\u6765\u81ea Inventive HQ \u5728 M3 Max 36GB \u4e0a\u8dd1 Qwen2.5-Coder 7B \u7684\u6570\u636e\uff09\uff1a<\/p>\n