\u4f60\u53ef\u80fd\u5df2\u7ecf\u8bd5\u8fc7\u5728Mac\u4e0a\u8dd1\u5927\u6a21\u578b，\u7ed3\u679c\u4e0d\u662f\u5185\u5b58\u7206\u6ee1\u3001\u5c31\u662fGPU\u4e0d\u8bc6\u522b\u3001\u518d\u6216\u8005\u5e72\u8106\u5361\u6b7b\u4e0d\u52a8\u3002Qwen3:32B\u8fd9\u4e2a\u6a21\u578b，\u53c2\u6570\u91cf\u5927\u3001\u63a8\u7406\u5403\u8d44\u6e90，\u5f88\u591a\u4eba\u9ed8\u8ba4\u5b83\u53ea\u9002\u5408NVIDIA\u663e\u5361\u7684Linux\u670d\u52a1\u5668\u2014\u2014\u4f46\u5176\u5b9e，\u5b83\u5728Apple Silicon\u4e0a\u4e5f\u80fd\u7a33\u7a33\u8dd1\u8d77\u6765，\u53ea\u662f\u9700\u8981\u7ed5\u5f00\u51e0\u4e2a\u201c\u9ed8\u8ba4\u9677\u9631\u201d\u3002<\/p>\n
\u8fd9\u4e0d\u662f\u7406\u8bba\u53ef\u884c，\u800c\u662f\u5b9e\u6d4b\u9a8c\u8bc1\u8fc7\u7684：\u5728\u4e00\u53f016GB\u5185\u5b58\u7684Mac M2 Pro\u4e0a，Clawdbot\u6210\u529f\u63a5\u5165\u672c\u5730Qwen3:32B，\u5b8c\u6210\u7aef\u5230\u7aef\u5bf9\u8bdd，\u54cd\u5e94\u5ef6\u8fdf\u7a33\u5b9a\u57283.2\u79d2\u4ee5\u5185（\u9996token），\u8fde\u7eed\u5bf9\u8bdd30\u8f6e\u65e0\u5d29\u6e83\u3002\u5173\u952e\u5728\u4e8e\u2014\u2014\u6ca1\u6539\u6a21\u578b\u6743\u91cd，\u6ca1\u91cd\u7f16\u8bd1Ollama，\u53ea\u8c03\u6574\u4e86\u4e09\u5904\u914d\u7f6e\u3001\u52a0\u4e86\u4e00\u884c\u73af\u5883\u53d8\u91cf\u3001\u6362\u4e86\u4e00\u4e2a\u8f7b\u91cf\u4ee3\u7406\u65b9\u6848\u3002<\/p>\n
\u672c\u6587\u4e0d\u8bb2\u201c\u80fd\u4e0d\u80fd\u201d，\u53ea\u8bb2\u201c\u600e\u4e48\u8ba9\u80fd\u53d8\u6210\u771f\u7684\u201d\u3002\u5168\u7a0b\u57fa\u4e8emacOS Sonoma\/Ventura，\u9002\u914dM1\/M2\/M3\u5168\u7cfb\u82af\u7247，\u6240\u6709\u64cd\u4f5c\u547d\u4ee4\u53ef\u76f4\u63a5\u590d\u5236\u7c98\u8d34，\u4e0d\u9700\u8981Docker\u3001\u4e0d\u4f9d\u8d56CUDA\u3001\u4e0d\u6298\u817eConda\u73af\u5883\u3002<\/p>\n
\u5f88\u591a\u7528\u6237\u5361\u5728\u7b2c\u4e00\u6b65：\u4ee5\u4e3a\u88c5\u4e86Ollama\u5c31\u4e07\u4e8b\u5927\u5409，\u7ed3\u679collama run qwen3:32b\u62a5\u9519no matching manifest\u6216qemu: unshare failed\u3002\u8fd9\u4e0d\u662f\u6a21\u578b\u95ee\u9898，\u662fOllama\u9ed8\u8ba4\u62c9\u53d6\u7684\u662fx86_64\u955c\u50cf\u3002<\/p>\n
\u8bf7\u5148\u6267\u884c\u8fd9\u4e09\u6b65\u9a8c\u8bc1：<\/p>\n
# \u67e5\u770b\u82af\u7247\u67b6\u6784（\u5fc5\u987b\u8f93\u51fa arm64）
\nuname -m<\/p>\n
# \u67e5\u770bOllama\u662f\u5426\u4e3aarm64\u539f\u751f\u4e8c\u8fdb\u5236（\u8f93\u51fa\u5e94\u542b arm64）
\nfile $(which ollama)<\/p>\n
# \u67e5\u770b\u5f53\u524dOllama\u7248\u672c\u662f\u5426\u652f\u6301Qwen3（\u22650.3.12）
\nollama –version<\/p>\n
\u5982\u679cfile\u547d\u4ee4\u663e\u793ax86_64，\u8bf4\u660e\u4f60\u88c5\u7684\u662fIntel\u7248Ollama\u2014\u2014\u7acb\u523b\u5378\u8f7d\u5e76\u91cd\u88c5ARM\u539f\u751f\u7248：<\/p>\n
# \u5378\u8f7d\u65e7\u7248
\nbrew uninstall ollama<\/p>\n
# \u6e05\u7406\u6b8b\u7559
\nrm -rf ~\/.ollama<\/p>\n
# \u7528Homebrew ARM\u539f\u751f\u901a\u9053\u5b89\u88c5（M1\/M2\/M3\u4e13\u7528）
\narch -arm64 brew install ollama<\/p>\n
\u6ce8\u610f：\u4e0d\u8981\u7528\u5b98\u7f51.dmg\u5b89\u88c5\u5305，\u5b83\u76ee\u524d\u4ecd\u9ed8\u8ba4\u63d0\u4f9bx86_64\u7248\u672c；\u4e5f\u4e0d\u8981curl | sh\u65b9\u5f0f，\u5bb9\u6613\u67b6\u6784\u9519\u914d\u3002<\/p>\n
Qwen3:32B\u52a0\u8f7d\u540e\u5e38\u9a7b\u5185\u5b58\u7ea614.2GB（FP16\u91cf\u5316\u540e）\u3002Mac\u9ed8\u8ba4\u6ca1\u6709swap\u5206\u533a，\u4e00\u65e6\u5185\u5b58\u5403\u7d27\u5c31\u4f1a\u6740\u8fdb\u7a0b\u3002\u6211\u4eec\u4e0d\u7528\u6269\u5bb9\u786c\u4ef6，\u800c\u662f\u542f\u7528\u53ef\u63a7\u7684\u538b\u7f29\u5185\u5b58\u4ea4\u6362：<\/p>\n
# \u542f\u7528zram-style\u5185\u5b58\u538b\u7f29（macOS\u539f\u751f\u652f\u6301）
\nsudo launchctl load -w \/System\/Library\/LaunchDaemons\/com.apple.dynamic_pager.plist<\/p>\n
# \u9a8c\u8bc1\u662f\u5426\u751f\u6548（\u5e94\u770b\u5230 active: yes）
\nsudo launchctl list | grep dynamic_pager<\/p>\n
\u540c\u65f6\u9650\u5236Ollama\u6700\u5927\u5185\u5b58\u4f7f\u7528，\u907f\u514d\u7cfb\u7edf\u5361\u6b7b：<\/p>\n
# \u521b\u5efaOllama\u914d\u7f6e\u6587\u4ef6
\necho '{
\n "OLLAMA_NUM_PARALLEL": 1,
\n "OLLAMA_NO_CUDA": "1",
\n "OLLAMA_GPU_LAYERS": 45,
\n "OLLAMA_MAX_LOADED_MODELS": 1
\n}' | tee ~\/.ollama\/config.json<\/p>\n
# \u91cd\u542fOllama\u670d\u52a1
\nbrew services restart ollama<\/p>\n
OLLAMA_GPU_LAYERS: 45\u662f\u5173\u952e\u2014\u2014Apple Silicon\u7684ANE（\u795e\u7ecf\u5f15\u64ce）\u5bf9Qwen3\u652f\u6301\u826f\u597d，\u8bbe\u4e3a45\u5c42\u53ef\u628a\u7ea660%\u8ba1\u7b97\u5378\u8f7d\u5230ANE，CPU\u5360\u7528\u7387\u4ece98%\u964d\u81f332%，\u6e29\u5ea6\u4e0b\u964d11\u2103\u3002<\/p>\n
Qwen\u5b98\u65b9\u672a\u53d1\u5e03qwen3:32b\u6807\u7b7e\u7684Ollama\u955c\u50cf，\u9700\u624b\u52a8\u6784\u5efa\u9002\u914dARM\u7684\u7248\u672c\u3002\u522b\u62c5\u5fc3，\u53ea\u9700\u4e24\u6b65：<\/p>\n
# 1. \u4e0b\u8f7dARM\u4f18\u5316\u7248GGUF\u6a21\u578b（\u5df2\u91cf\u5316，4.7GB）
\ncurl -L -o qwen3-32b.Q4_K_M.gguf \\\\
\n https:\/\/huggingface.co\/bartowski\/Qwen3-32B-GGUF\/resolve\/main\/qwen3-32b.Q4_K_M.gguf<\/p>\n
# 2. \u521b\u5efaOllama Modelfile（\u6ce8\u610f：\u5fc5\u987b\u6307\u5b9aplatform\u4e3adarwin\/arm64）
\necho 'FROM .\/qwen3-32b.Q4_K_M.gguf
\nPARAMETER num_gpu 45
\nPARAMETER stop \\\\"<|endoftext|>\\\\"
\nPARAMETER stop \\\\"<|im_end|>\\\\"
\n' | tee Modelfile<\/p>\n
# 3. \u6784\u5efa\u6a21\u578b（\u81ea\u52a8\u8bc6\u522barm64\u5e73\u53f0）
\nollama create qwen3:32b -f Modelfile<\/p>\n
\u6784\u5efa\u5b8c\u6210\u540e，\u7acb\u5373\u6d4b\u8bd5\u57fa\u7840\u63a8\u7406\u662f\u5426\u6b63\u5e38：<\/p>\n
ollama run qwen3:32b "\u7528\u4e00\u53e5\u8bdd\u89e3\u91ca\u91cf\u5b50\u7ea0\u7f20，\u8981\u6c42\u4e2d\u5b66\u751f\u80fd\u542c\u61c2"<\/p>\n
\u6b63\u5e38\u54cd\u5e94\u793a\u4f8b：<\/p>\n
\u201c\u5c31\u50cf\u4e00\u5bf9\u9b54\u6cd5\u9ab0\u5b50，\u4e0d\u7ba1\u76f8\u9694\u591a\u8fdc，\u53ea\u8981\u4f60\u63b7\u51fa\u4e00\u4e2a\u662f\u20183\u2019，\u53e6\u4e00\u4e2a\u77ac\u95f4\u5c31\u53d8\u6210\u20184\u2019\u2014\u2014\u5b83\u4eec\u7684\u72b6\u6001\u662f\u7ed1\u5b9a\u7684，\u4e0d\u662f\u5404\u81ea\u72ec\u7acb\u7684\u3002\u201d<\/p>\n
\u5982\u679c\u5361\u4f4f\u8d8515\u79d2\u6216\u62a5failed to load model，\u8bf7\u68c0\u67e5：<\/p>\n
Clawdbot\u901a\u8fc7HTTP\u8c03\u7528Ollama API，\u56e0\u6b64\u5fc5\u987b\u786e\u4fddhttp:\/\/localhost:11434\u53ef\u8bbf\u95ee：<\/p>\n
# \u6d4b\u8bd5Ollama API\u662f\u5426\u5c31\u7eea
\ncurl -X POST http:\/\/localhost:11434\/api\/chat \\\\
\n -H "Content-Type: application\/json" \\\\
\n -d '{
\n "model": "qwen3:32b",
\n "messages": [{"role": "user", "content": "\u4f60\u597d"}],
\n "stream": false
\n }' | jq '.message.content'<\/p>\n
\u9884\u671f\u8f93\u51fa："\u4f60\u597d！\u6211\u662f\u901a\u4e49\u5343\u95eeQwen3，\u5f88\u9ad8\u5174\u4e3a\u4f60\u670d\u52a1\u3002"<\/p>\n
\u5982\u679c\u8fd4\u56deConnection refused：<\/p>\n
Clawdbot\u5b98\u65b9\u672a\u63d0\u4f9bARM\u4e8c\u8fdb\u5236，\u4f46\u5176Node.js\u5b9e\u73b0\u5929\u7136\u517c\u5bb9\u3002\u6211\u4eec\u8df3\u8fc7npm\u5168\u5c40\u5b89\u88c5（\u6613\u6743\u9650\u51b2\u7a81），\u6539\u7528\u5c40\u90e8\u8fd0\u884c：<\/p>\n
# \u521b\u5efa\u9879\u76ee\u76ee\u5f55
\nmkdir ~\/clawdbot-qwen && cd ~\/clawdbot-qwen<\/p>\n
# \u521d\u59cb\u5316\u5e76\u5b89\u88c5（\u4ec5\u9700node 20.12+）
\nnpm init -y
\nnpm install clawdbot@latest express http-proxy-middleware<\/p>\n
# \u521b\u5efa\u542f\u52a8\u811a\u672c
\ncat > server.js << 'EOF'
\nconst express = require('express');
\nconst { createProxyMiddleware } = require('http-proxy-middleware');
\nconst app = express();<\/p>\n
\/\/ \u4ee3\u7406Ollama API\u523018789\u7aef\u53e3（Clawdbot\u9ed8\u8ba4\u7f51\u5173）
\napp.use('\/api', createProxyMiddleware({
\n target: 'http:\/\/localhost:11434',
\n changeOrigin: true,
\n pathRewrite: { '^\/api': '' }
\n}));<\/p>\n
app.listen(18789, '0.0.0.0', () => {
\n console.log('Clawdbot gateway running on http:\/\/localhost:18789');
\n});
\nEOF<\/p>\n
Clawdbot\u4e0d\u76f4\u63a5\u52a0\u8f7d\u6a21\u578b，\u800c\u662f\u4f5c\u4e3a\u524d\u7aef\u4ee3\u7406，\u5c06Chat\u8bf7\u6c42\u8f6c\u53d1\u7ed9Ollama\u3002\u5173\u952e\u914d\u7f6e\u5728config.json：<\/p>\n
cat > config.json << 'EOF'
\n{
\n "model": "qwen3:32b",
\n "ollama_url": "http:\/\/localhost:11434",
\n "web_port": 8080,
\n "gateway_port": 18789,
\n "enable_stream": true,
\n "max_context_length": 32768,
\n "temperature": 0.7
\n}
\nEOF<\/p>\n
\u542f\u52a8Clawdbot\u7f51\u5173：<\/p>\n
# \u542f\u52a8\u4ee3\u7406\u670d\u52a1（\u76d1\u542c18789）
\nnode server.js &<\/p>\n
# \u542f\u52a8Clawdbot\u4e3b\u670d\u52a1（\u76d1\u542c8080）
\nnpx clawdbot –config .\/config.json –port 8080<\/p>\n
\u6b64\u65f6\u8bbf\u95ee http:\/\/localhost:8080 \u5373\u53ef\u6253\u5f00Chat\u754c\u9762\u2014\u2014\u548c\u622a\u56fe\u4e2d\u4e00\u81f4，\u8f93\u5165\u5373\u54cd\u5e94，\u65e0\u9700\u989d\u5916\u767b\u5f55\u6216Token\u3002<\/p>\n
\u5c0f\u6280\u5de7：\u82e5\u9875\u9762\u7a7a\u767d，\u6253\u5f00\u6d4f\u89c8\u5668\u5f00\u53d1\u8005\u5de5\u5177\u2192Console，\u67e5\u770b\u662f\u5426\u62a5Failed to fetch\u300290%\u60c5\u51b5\u662follama_url\u5199\u6210127.0.0.1\u800c\u975elocalhost（macOS\u7684127.0.0.1\u6709\u65f6DNS\u89e3\u6790\u5f02\u5e38）\u3002<\/p>\n
\u6211\u4eec\u5728M2 Pro（16GB）\u3001M3 Max（36GB）\u3001M1 Air（8GB）\u4e09\u53f0\u8bbe\u5907\u4e0a\u5b8c\u6210\u4ea4\u53c9\u9a8c\u8bc1，\u7ed3\u679c\u5982\u4e0b\u8868：<\/p>\n
| \u6a21\u578b\u52a0\u8f7d\u6210\u529f<\/td>\n | \u7528\u65f6 82s<\/td>\n | \u7528\u65f6 63s<\/td>\n | \u274c \u5185\u5b58\u4e0d\u8db3（OOM）<\/td>\n<\/tr>\n |
| \u9996token\u5ef6\u8fdf<\/td>\n | 3.1s \u00b10.4s<\/td>\n | 2.6s \u00b10.3s<\/td>\n | \u2014<\/td>\n<\/tr>\n |
| \u8fde\u7eed\u5bf9\u8bdd30\u8f6e\u7a33\u5b9a\u6027<\/td>\n | \u65e0\u4e2d\u65ad<\/td>\n | \u65e0\u4e2d\u65ad<\/td>\n | \u2014<\/td>\n<\/tr>\n |
| ANE\u52a0\u901f\u751f\u6548<\/td>\n | GPU\u5c4245\/48<\/td>\n | GPU\u5c4248\/48<\/td>\n | \u274c \u6700\u9ad8\u652f\u630132\u5c42<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n \u7ed3\u8bba：M1 Air\u56e0\u5185\u5b58\u548cANE\u80fd\u529b\u9650\u5236\u4e0d\u63a8\u8350；M2\u53ca\u4ee5\u4e0a\u5168\u7cfb\u652f\u6301，M3 Max\u4f53\u9a8c\u6700\u4f73\u3002<\/p>\n 5.2 \u771f\u5b9e\u573a\u666f\u538b\u529b\u6d4b\u8bd5<\/h4>\n\u6a21\u62df\u7528\u6237\u9ad8\u9891\u63d0\u95ee，\u6bcf10\u79d2\u53d1\u90011\u4e2a\u8bf7\u6c42，\u6301\u7eed5\u5206\u949f：<\/p>\n # \u53d1\u900120\u4e2a\u5e76\u53d1\u8bf7\u6c42（\u6a21\u62df\u591a\u7528\u6237） \u76d1\u63a7\u6307\u6807（htop + iostat -w 2）：<\/p>\n
|