{"id":490,"date":"2026-07-24T20:07:31","date_gmt":"2026-07-24T12:07:31","guid":{"rendered":"http:\/\/www.tantan.ink\/index.php\/2026\/07\/24\/16gb-mac-%e8%b7%91-35b-%e6%a8%a1%e5%9e%8b%ef%bc%8c1-58bit-%e9%87%8f%e5%8c%96%e7%9c%9f%e7%9a%84%e8%83%bd%e7%94%a8%e5%90%97%ef%bc%9f\/"},"modified":"2026-07-24T20:07:52","modified_gmt":"2026-07-24T12:07:52","slug":"16gb-mac-%e8%b7%91-35b-%e6%a8%a1%e5%9e%8b%ef%bc%8c1-58bit-%e9%87%8f%e5%8c%96%e7%9c%9f%e7%9a%84%e8%83%bd%e7%94%a8%e5%90%97%ef%bc%9f","status":"publish","type":"post","link":"https:\/\/www.tantan.ink\/index.php\/2026\/07\/24\/16gb-mac-%e8%b7%91-35b-%e6%a8%a1%e5%9e%8b%ef%bc%8c1-58bit-%e9%87%8f%e5%8c%96%e7%9c%9f%e7%9a%84%e8%83%bd%e7%94%a8%e5%90%97%ef%bc%9f\/","title":{"rendered":"16GB Mac \u8dd1 35B \u6a21\u578b\uff0c1.58bit \u91cf\u5316\u771f\u7684\u80fd\u7528\u5417\uff1f"},"content":{"rendered":"<p>\u6211\u7684 Mac \u662f 16GB \u7edf\u4e00\u5185\u5b58\u3002\u8dd1 Qwen3.6-35B-A3B \u8fd9\u4ef6\u4e8b\uff0c\u4e4b\u524d\u9760 llama.cpp \u7684 mmap  trick \u52c9\u5f3a\u80fd\u884c\u2014\u2014\u6a21\u578b\u6587\u4ef6 20GB \u5de6\u53f3\uff0c\u7cfb\u7edf\u53ea\u628a\u6d3b\u8dc3\u4e13\u5bb6\u9875\u8fdb\u5185\u5b58\uff0c\u5e73\u65f6\u770b\u7740\u5185\u5b58\u5360\u7528\u4e0d\u9ad8\uff0c\u4f46\u4e0a\u4e0b\u6587\u4e00\u957f\u6216\u8005\u591a\u5f00\u51e0\u4e2a\u5e94\u7528\uff0c\u8be5swap\u7684\u8fd8\u662fswap\uff0c\u98ce\u6247\u8be5\u8f6c\u7684\u8fd8\u662f\u8f6c\u3002<\/p>\n<p>\u524d\u51e0\u5929\u5237\u5230 TurboQuant-MLX \u7684\u4e00\u4e2a\u5b9e\u9a8c\u6027\u6784\u5efa\uff0c\u76f4\u63a5\u628a 35B \u6a21\u578b\u4ece 70GB \u538b\u5230\u4e86 9.4GB\u3002\u7528\u7684\u4e0d\u662f\u5e38\u89c4\u7684 Q4\uff0c\u800c\u662f<strong>1.58bit \u4e09\u503c\u91cf\u5316<\/strong>\u2014\u2014\u6bcf\u4e2a\u6743\u91cd\u53ea\u5b58\u4e09\u79cd\u503c\uff1a-c\u30010\u3001+c\u3002\u8fd9\u4e2a\u538b\u7f29\u7387\u542c\u8d77\u6765\u50cf\u9b54\u6cd5\uff0c\u4f46\u6211\u4ed4\u7ec6\u770b\u4e86\u6a21\u578b\u5361\u548c\u5b9e\u6d4b\u6570\u636e\uff0c\u89c9\u5f97\u6709\u5fc5\u8981\u62ff\u51fa\u6765\u804a\u804a\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.tantan.ink\/wp-content\/uploads\/2026\/07\/hf-model-card.png\" alt=\"hf-model-card.png\" loading=\"lazy\" \/><\/p>\n<p><strong>\u5148\u653e\u7ed3\u8bba\uff1a\u5b83\u80fd\u8dd1\uff0c\u800c\u4e14\u80fd\u5b8c\u6574\u9a7b\u7559\u5728 16GB \u5185\u5b58\u91cc\uff0c\u4e0d\u7528 swap\uff0c\u4e0d\u7528\u5916\u63a5 SSD \u6d41\u5f0f\u52a0\u8f7d\u3002<\/strong>\u4f46\u522b\u6307\u671b\u5b83\u66ff\u4ee3\u4f60\u73b0\u6709\u7684\u4e3b\u529b\u65b9\u6848\u3002<\/p>\n<h2>\u4e3a\u4ec0\u4e48 1.58bit \u6ca1\u53d8\u6210\u4e71\u7801<\/h2>\n<p>\u6309\u5e38\u8bc6\uff0c\u628a\u6743\u91cd\u538b\u5230 1.58bit\uff0c\u6a21\u578b\u8f93\u51fa\u5e94\u8be5\u76f4\u63a5\u5d29\u6e83\u6210\u80e1\u8a00\u4e71\u8bed\u3002\u4f46\u8fd9\u4e2a\u6784\u5efa\u6ca1\u5d29\uff0c\u6838\u5fc3\u539f\u56e0\u5728\u4e8e\u4e09\u70b9\uff1a<\/p>\n<p><strong>\u7b2c\u4e00\uff0c\u5b83\u53ea\u538b\u4e13\u5bb6\u5c42\u3002<\/strong>Qwen3.6-35B-A3B \u662f MoE \u67b6\u6784\uff0c256 \u4e2a\u8def\u7531\u4e13\u5bb6\uff0c\u6bcf token \u53ea\u6fc0\u6d3b 8 \u4e2a\u3002\u4e13\u5bb6\u5c42\u53c2\u6570\u91cf\u6700\u5927\uff0c\u4f46\u5355\u4e2a\u4e13\u5bb6\u53ea\u5904\u7406\u4e00\u5c0f\u90e8\u5206 token\uff0c\u5197\u4f59\u5ea6\u6781\u9ad8\u3002\u6ce8\u610f\u529b\u5c42\u3001GDN \u6295\u5f71\u77e9\u9635\u3001lm_head \u8fd9\u4e9b\u88ab\u6bcf\u4e2a token \u90fd\u8d70\u4e00\u904d\u7684&#8221;\u5173\u952e\u8def\u5f84&#8221;\uff0c\u4fdd\u7559\u5728 3-bit\u3002\u8def\u7531\u5668\u9009\u4e13\u5bb6\u7684 gate \u751a\u81f3\u4fdd\u6301\u5168\u7cbe\u5ea6\u2014\u2014\u9009\u9519\u4e13\u5bb6\u7684\u4ee3\u4ef7\u662f\u4e0d\u53ef\u6062\u590d\u7684\u3002<\/p>\n<p><strong>\u7b2c\u4e8c\uff0cMoE \u7684\u5197\u4f59\u62b5\u6d88\u4e86\u91cf\u5316\u566a\u58f0\u3002<\/strong>8\/256 \u7684\u6fc0\u6d3b\u6bd4\u4f8b\u610f\u5473\u7740\uff0c\u5373\u4f7f\u5355\u4e2a\u4e13\u5bb6\u88ab\u538b\u5f97\u6bd4\u8f83\u72e0\uff0c\u6a21\u578b\u53ef\u4ee5\u901a\u8fc7&#8221;\u6362\u4e00\u6279\u4e13\u5bb6&#8221;\u6765\u5e73\u5747\u6389\u8bef\u5dee\u3002\u6a21\u578b\u5361\u91cc\u7684\u539f\u8bdd\u662f\uff1atop-8 of 256 \u7684\u5197\u4f59\u5ea6\uff0c\u662f 32-expert \u6a21\u578b\u7684 8 \u500d\u2014\u2014\u540e\u8005\u5728 sub-2-bit \u4e0b demonstrably fails\u3002<\/p>\n<p><strong>\u7b2c\u4e09\uff0c\u4e09\u503c\u6bd4\u4e8c\u503c\u597d\u3002<\/strong>\u7eaf 1-bit\uff08\u53ea\u6709 -c \u548c +c\uff09\u7684\u4e13\u5bb6\u5c42\u6211\u8bd5\u8fc7\u5176\u4ed6\u6a21\u578b\uff0c\u8f93\u51fa\u57fa\u672c\u662f\u8bcd\u6c99\u62c9\u3002\u52a0\u5165 0 \u4e4b\u540e\uff0c1.58bit \u7684\u91cd\u5efa\u8d28\u91cf\u843d\u5728\u4e86\u7ea6 2-bit \u7684\u6c34\u5e73\u3002\u7528 base-3 \u7f16\u7801\u628a 20 \u4e2a\u4e09\u503c\u7d22\u5f15\u585e\u8fdb\u4e00\u4e2a 32-bit \u6574\u6570\uff0c\u5b9e\u9645\u4f4d\u5bbd\u7ea6 1.6 bpw\uff0c\u8fd9\u624d\u628a\u603b\u5927\u5c0f\u62c9\u5230\u4e86 10GB \u4ee5\u5185\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.tantan.ink\/wp-content\/uploads\/2026\/07\/performance-chart.png\" alt=\"performance-chart.png\" loading=\"lazy\" \/><\/p>\n<h2>\u5b9e\u6d4b\u6570\u636e\uff1a\u80fd\u8dd1\uff0c\u4f46\u4e0d\u591f\u5feb<\/h2>\n<p>\u6a21\u578b\u5361\u91cc\u7684\u6570\u636e\u662f\u5728 16GB Mac mini \u4e0a\u6d4b\u7684\uff1a<\/p>\n<ul>\n<li>\u89e3\u7801\u901f\u5ea6\uff1a\u7ea6 <strong>10.4 tok\/s<\/strong>\uff08\u5e73\u5766\uff0c512 token \u548c 1800 token \u56de\u7b54\u901f\u5ea6\u4e00\u81f4\uff09<\/li>\n<li>\u5185\u5b58\u5cf0\u503c\uff1a<strong>10.42 GB<\/strong>\uff0c\u4e0d\u968f\u4e0a\u4e0b\u6587\u589e\u957f\uff0830 \u5c42 GatedDeltaNet \u7528\u56fa\u5b9a\u5927\u5c0f\u7684\u5faa\u73af\u72b6\u6001\uff0c\u53ea\u6709 10 \u5c42 full-attention \u7684 KV \u4f1a\u81a8\u80c0\uff09<\/li>\n<li>\u4e0d\u9700\u8981 sudo \u8c03 wired_limit\uff0c\u9ed8\u8ba4 Metal \u5185\u5b58\u4e0a\u9650\u5c31\u80fd\u8dd1<\/li>\n<\/ul>\n<p>\u4f5c\u4e3a\u5bf9\u6bd4\uff0c\u6211\u76ee\u524d\u5728 llama.cpp \u4e0a\u8dd1 Q4_K_M \u7684 35B\uff0cdecode \u5927\u6982 17 tok\/s\u3002TurboQuant \u8fd9\u4e2a 1.58bit \u6784\u5efa\u6162\u4e86\u5c06\u8fd1\u4e00\u534a\uff0c\u4f46\u6362\u6765\u4e86<strong>\u5b8c\u6574\u9a7b\u7559\u5185\u5b58<\/strong>\u2014\u2014\u6ca1\u6709 mmap \u7684\u78c1\u76d8\u6296\u52a8\uff0c\u6ca1\u6709\u4e13\u5bb6\u6362\u5165\u6362\u51fa\u7684\u5ef6\u8fdf\u5c16\u5cf0\u3002<\/p>\n<p>\u8d28\u91cf\u65b9\u9762\uff0cMMLU-Redux \u601d\u8003\u6a21\u5f0f\u4e0b 80.1%\uff0c\u539f\u7248 93.3%\u3002\u5927\u7ea6\u662f\u539f\u7248\u7684 86%\u3002\u975e\u601d\u8003\u6a21\u5f0f 75.2%\u3002\u65e5\u5e38\u804a\u5929\u3001\u5199\u4ee3\u7801\u3001\u603b\u7ed3\u6587\u6863\uff0c\u8fd9\u4e2a\u6c34\u5e73\u591f\u7528\u3002\u4f46 GSM8K \u6570\u5b66 benchmark \u53ea\u6709 67%\uff0c\u8bf4\u660e\u7b97\u672f\u7cbe\u786e\u5ea6\u662f\u660e\u663e\u77ed\u677f\u3002<\/p>\n<h2>\u6700\u81f4\u547d\u7684\u5751\uff1aagent \u5de5\u5177\u5faa\u73af\u76f4\u63a5\u4e0d\u53ef\u7528<\/h2>\n<p>\u6a21\u578b\u5361\u91cc\u6709\u4e00\u6bb5\u8bdd\u8ba9\u6211\u5370\u8c61\u6df1\u523b\uff0c\u800c\u4e14\u662f\u88ab\u5b9e\u6d4b\u9a8c\u8bc1\u8fc7\u7684\uff0c\u4e0d\u662f\u63a8\u6d4b\uff1a<\/p>\n<p>&#8220;Tool-call syntax works&#8230; but tool-use judgment does not survive 1.58-bit.&#8221;<\/p>\n<p>\u5177\u4f53\u6d4b\u8bd5\u662f\uff1a\u7528 Opencode \u8dd1&#8221;\u4fee\u590d\u5931\u8d25\u6d4b\u8bd5&#8221;\u4efb\u52a1\uff0c1.58bit \u6784\u5efa 4\/4 \u6b21\u5931\u8d25\u2014\u2014\u8981\u4e48\u91cd\u590d\u6267\u884c\u540c\u4e00\u4e2a\u9519\u8bef\u547d\u4ee4\u51e0\u767e\u6b21\uff0c\u8981\u4e48\u51ed\u7a7a\u634f\u9020\u6587\u4ef6\u8def\u5f84\u548c URL\u3002\u800c\u540c\u4e00\u4e2a server + harness + prompt \u4e0b\uff0c3-bit \u7684 sibling \u6784\u5efa 33 \u79d2\u5c31\u641e\u5b9a\u4e86\u3002<\/p>\n<p>\u8fd9\u4e0d\u662f\u91c7\u6837\u53c2\u6570\u6ca1\u8c03\u5bf9\uff0c\u800c\u662f<strong>\u4fe1\u606f\u5730\u677f<\/strong>\u7684\u95ee\u9898\u3002\u4e09\u503c\u4e13\u5bb6\u5c42\u4fdd\u7559\u7684\u8bed\u4e49\u7cbe\u5ea6\uff0c\u4e0d\u8db3\u4ee5\u652f\u6491\u591a\u6b65 observe \u2192 diagnose \u2192 edit \u7684\u89c4\u5212\u5faa\u73af\u3002<\/p>\n<p>\u6240\u4ee5\u6a21\u578b\u5361\u7684\u5efa\u8bae\u5f88\u660e\u786e\uff1a\u8fd9\u4e2a\u6784\u5efa\u9002\u5408 chat\u3001drafting\u3001summarization\u3001Q&amp;A\u3001\u4e00\u6b21\u6027\u4ee3\u7801\u751f\u6210\u3002<strong>\u4e0d\u8981\u63a5\u8fdb coding agent\u3002<\/strong><\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.tantan.ink\/wp-content\/uploads\/2026\/07\/github-turboquant.png\" alt=\"github-turboquant.png\" loading=\"lazy\" \/><\/p>\n<h2>\u6211\u7684\u7528\u6cd5<\/h2>\n<p>\u6211\u76ee\u524d\u7684\u8ba1\u5212\u662f\uff1allama.cpp Q4 \u7ee7\u7eed\u5f53\u4e3b\u529b\uff0c\u5904\u7406\u9700\u8981\u5de5\u5177\u8c03\u7528\u548c\u957f\u4e0a\u4e0b\u6587\u63a8\u7406\u7684\u4efb\u52a1\u3002TurboQuant \u7684 1.58bit \u6784\u5efa\u88c5\u4e00\u4e2a\u5907\u7528\uff0c\u4e13\u95e8\u7528\u6765\u8dd1<strong>\u4e0d\u9700\u8981\u601d\u8003\u6a21\u5f0f\u7684\u5feb\u95ee\u5feb\u7b54<\/strong>\u2014\u2014\u67e5\u8d44\u6599\u3001\u6539\u6587\u6848\u3001\u7b80\u5355\u811a\u672c\u300210 tok\/s \u4e0d\u7b97\u5feb\uff0c\u4f46\u6bd4\u7b49\u4e91\u7aef API \u7684\u51b7\u542f\u52a8\u5feb\uff0c\u800c\u4e14\u5b8c\u5168\u4e0d\u5360 swap\u3002<\/p>\n<p>\u5b89\u88c5\u5012\u662f\u7b80\u5355\uff0c\u4e24\u6761\u547d\u4ee4\uff1a<\/p>\n<pre><code>pip install \"turboquant-mlx-full&gt;=0.12.3\" mlx-lm\npython -m turboquant_mlx.generate \n    --model manjunathshiva\/Qwen3.6-35B-A3B-tq3a-tqTe-g64 \n    --prompt \"\u4f60\u7684\u95ee\u9898\" --max-tokens 512 --no-think<\/code><\/pre>\n<p>\u52a0\u4e0a <code>--no-think<\/code> \u4e4b\u540e\u54cd\u5e94\u5feb\u5f88\u591a\uff0c\u6570\u5b66\u548c\u4e25\u683c\u683c\u5f0f\u4efb\u52a1\u4e5f\u66f4\u7a33\u3002\u9700\u8981\u601d\u8003\u6a21\u5f0f\u7684\u65f6\u5019\uff0c\u7ed9\u591f 4K+ \u7684 token \u9884\u7b97\uff0c\u4e0d\u7136\u5bb9\u6613\u5728 <code>&lt;think&gt;<\/code> \u91cc\u7ed5\u5708\u51fa\u4e0d\u6765\u3002<\/p>\n<h2>\u8fd9\u80cc\u540e\u7684\u8d8b\u52bf<\/h2>\n<p>1.58bit \u4e0d\u662f TurboQuant \u4e00\u5bb6\u7684\u73a9\u6cd5\u3002antirez \u7684 ds4 \u7ed9 DeepSeek V4 Flash \u505a\u4e86\u975e\u5bf9\u79f0 2\/8-bit \u91cf\u5316\uff0cMoE \u4e13\u5bb6\u5c42\u538b\u5230 2-bit\uff0c\u6ce8\u610f\u529b\u5c42\u4fdd\u6301 8-bit\u3002\u601d\u8def\u5b8c\u5168\u4e00\u81f4\uff1a<strong>\u8be5\u72e0\u538b\u7684\u5730\u65b9\u72e0\u538b\uff0c\u4e0d\u8be5\u538b\u7684\u5730\u65b9\u5bf8\u6b65\u4e0d\u8ba9\u3002<\/strong><\/p>\n<p>\u8fd9\u79cd&#8221;\u5206\u5c42\u91cf\u5316&#8221;\u6bd4\u4e00\u5200\u5207\u7684 Q4 \u9ad8\u7ea7\u592a\u591a\uff0c\u800c\u4e14\u4f1a\u8d8a\u6765\u8d8a\u4e3b\u6d41\u3002\u56e0\u4e3a\u672c\u5730\u63a8\u7406\u7684\u74f6\u9888\u4ece\u6765\u4e0d\u662f&#8221;\u80fd\u4e0d\u80fd\u8dd1&#8221;\uff0c\u800c\u662f<strong>\u5728\u6709\u9650\u7684\u5185\u5b58\u9884\u7b97\u91cc\uff0c\u628a\u7cbe\u5ea6\u5206\u914d\u5230\u6700\u9700\u8981\u5b83\u7684\u5730\u65b9\u3002<\/strong><\/p>\n<p>\u5bf9 16GB Mac \u7528\u6237\u6765\u8bf4\uff0c1.58bit \u4e09\u503c\u91cf\u5316\u610f\u5473\u7740 35B \u7ea7\u522b\u7684\u6a21\u578b\u7ec8\u4e8e\u53ef\u4ee5\u4ece&#8221;\u52c9\u5f3a\u80fd\u8dd1&#8221;\u53d8\u6210&#8221;\u5e38\u9a7b\u53ef\u7528&#8221;\u3002\u867d\u7136\u901f\u5ea6\u548c\u7cbe\u5ea6\u90fd\u6709\u59a5\u534f\uff0c\u4f46\u8fd9\u59a5\u534f\u662f\u6709\u660e\u786e\u8fb9\u754c\u7684\u2014\u2014\u4f60\u77e5\u9053\u5b83\u64c5\u957f\u4ec0\u4e48\u3001\u4e0d\u64c5\u957f\u4ec0\u4e48\uff0c\u7136\u540e\u6309\u9700\u8c03\u7528\u3002<\/p>\n<p>\u8fd9\u6bd4\u76f2\u76ee\u8ffd\u6c42&#8221;\u4e00\u4e2a\u6a21\u578b\u5e72\u6240\u6709\u4e8b&#8221;\u52a1\u5b9e\u591a\u4e86\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.tantan.ink\/wp-content\/uploads\/2026\/07\/github-qwen36.png\" alt=\"github-qwen36.png\" loading=\"lazy\" \/><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6211\u7684 Mac \u662f 16GB \u7edf\u4e00\u5185\u5b58\u3002\u8dd1 Qwen3.6-35B-A3B \u8fd9\u4ef6\u4e8b\uff0c\u4e4b\u524d\u9760 llama.cpp  &#8230; <a title=\"16GB Mac \u8dd1 35B \u6a21\u578b\uff0c1.58bit \u91cf\u5316\u771f\u7684\u80fd\u7528\u5417\uff1f\" class=\"read-more\" href=\"https:\/\/www.tantan.ink\/index.php\/2026\/07\/24\/16gb-mac-%e8%b7%91-35b-%e6%a8%a1%e5%9e%8b%ef%bc%8c1-58bit-%e9%87%8f%e5%8c%96%e7%9c%9f%e7%9a%84%e8%83%bd%e7%94%a8%e5%90%97%ef%bc%9f\/\" aria-label=\"\u9605\u8bfb 16GB Mac \u8dd1 35B \u6a21\u578b\uff0c1.58bit \u91cf\u5316\u771f\u7684\u80fd\u7528\u5417\uff1f\">\u9605\u8bfb\u66f4\u591a<\/a><\/p>\n","protected":false},"author":2,"featured_media":486,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[2,176],"tags":[],"class_list":["post-490","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\/490","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=490"}],"version-history":[{"count":1,"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/posts\/490\/revisions"}],"predecessor-version":[{"id":491,"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/posts\/490\/revisions\/491"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/media\/486"}],"wp:attachment":[{"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/media?parent=490"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/categories?post=490"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/tags?post=490"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}