{"id":105298,"date":"2026-09-15T08:24:46","date_gmt":"2026-09-15T00:24:46","guid":{"rendered":"https:\/\/www.wsisp.com\/helps\/105298.html"},"modified":"2026-09-15T08:24:46","modified_gmt":"2026-09-15T00:24:46","slug":"google-gemma-4-12b-unified-%e6%9c%ac%e5%9c%b0%e9%83%a8%e7%bd%b2%e5%ae%8c%e5%85%a8%e6%8c%87%e5%8d%97%ef%bc%9a16gb-%e6%98%be%e5%ad%98%e8%b7%91%e8%b5%b7%e5%a4%9a%e6%a8%a1%e6%80%81%e5%a4%a7%e6%a8%a1","status":"publish","type":"post","link":"https:\/\/www.wsisp.com\/helps\/105298.html","title":{"rendered":"Google Gemma 4 12B Unified \u672c\u5730\u90e8\u7f72\u5b8c\u5168\u6307\u5357\uff1a16GB \u663e\u5b58\u8dd1\u8d77\u591a\u6a21\u6001\u5927\u6a21\u578b"},"content":{"rendered":"

Google Gemma 4 12B Unified \u672c\u5730\u90e8\u7f72\u5b8c\u5168\u6307\u5357：16GB \u663e\u5b58\u8dd1\u8d77\u591a\u6a21\u6001\u5927\u6a21\u578b<\/h2>\n

\u4f5c\u8005：\u8d5b\u535a\u4ed3\u9f20 | 2026-09-14<\/p>\n

2026 \u5e74 8 \u6708，Google DeepMind \u6b63\u5f0f\u5f00\u6e90\u4e86 Gemma 4 \u7cfb\u5217\u6a21\u578b\u3002\u4f5c\u4e3a Gemma \u5bb6\u65cf\u9996\u6b21\u5f15\u5165\u65e0\u7f16\u7801\u5668\u7edf\u4e00\u591a\u6a21\u6001\u67b6\u6784（Encoder-Free Unified）\u7684\u7248\u672c，Gemma 4 12B \u5728 12B \u53c2\u6570\u89c4\u6a21\u4e0a\u5b9e\u73b0\u4e86\u6587\u672c\u3001\u56fe\u50cf\u3001\u97f3\u9891\u3001\u89c6\u9891\u7684\u539f\u751f\u5904\u7406\u80fd\u529b，\u4e14\u652f\u6301\u957f\u8fbe 256K tokens \u7684\u4e0a\u4e0b\u6587\u7a97\u53e3\u3002\u66f4\u5173\u952e\u7684\u662f，\u7ecf\u8fc7 Q4 \u91cf\u5316\u540e，\u8fd9\u6b3e\u6a21\u578b\u4ec5\u9700 6.5GB \u663e\u5b58 \u5373\u53ef\u5728\u6d88\u8d39\u7ea7\u663e\u5361\u4e0a\u6d41\u7545\u8fd0\u884c\u2014\u2014\u8fd9\u610f\u5473\u7740\u4e00\u5f20 RTX 3060 12GB \u5c31\u80fd\u672c\u5730\u90e8\u7f72\u4e00\u4e2a\u5177\u5907\u591a\u6a21\u6001\u80fd\u529b\u7684\u5927\u6a21\u578b\u3002<\/p>\n

\u672c\u6587\u5c06\u5e26\u4f60\u5b8c\u6574\u8d70\u901a Gemma 4 12B \u7684\u672c\u5730\u90e8\u7f72\u5168\u6d41\u7a0b，\u4ece\u786c\u4ef6\u95e8\u69db\u5224\u65ad\u5230\u56db\u79cd\u90e8\u7f72\u65b9\u5f0f\u5b9e\u64cd，\u518d\u5230\u91cf\u5316\u65b9\u6848\u9009\u578b\u4e0e\u6027\u80fd\u5b9e\u6d4b，\u5168\u90e8\u57fa\u4e8e\u4e00\u624b\u5b9e\u6d4b\u6570\u636e\u3002<\/p>\n

\n

\u4e00\u3001\u6a21\u578b\u6838\u5fc3\u53c2\u6570\u901f\u89c8<\/h3>\n\n\u53c2\u6570\u9879Gemma 4 12B Unified\u8bf4\u660e<\/tr>\n\n\n\n\n\n\n\n\n\n\n
\u603b\u53c2\u6570\u91cf<\/td>\n~12B（119 \u4ebf）<\/td>\nDense \u67b6\u6784，\u975e MoE<\/td>\n<\/tr>\n
\u9690\u85cf\u5c42\u7ef4\u5ea6<\/td>\n3840<\/td>\n\u6807\u51c6 Transformer \u914d\u7f6e<\/td>\n<\/tr>\n
\u6ce8\u610f\u529b\u5934\u6570<\/td>\n16（GQA 16:2）<\/td>\n\u5206\u7ec4\u67e5\u8be2\u6ce8\u610f\u529b，\u964d\u4f4e KV \u7f13\u5b58<\/td>\n<\/tr>\n
\u5c42\u6570<\/td>\n48<\/td>\n\u8f83\u6df1\u7f51\u7edc，\u8868\u8fbe\u80fd\u529b\u5145\u8db3<\/td>\n<\/tr>\n
\u4e0a\u4e0b\u6587\u957f\u5ea6<\/td>\n256K tokens<\/td>\n\u5b9e\u6d4b\u4f4d\u7f6e\u5d4c\u5165\u652f\u6301 131,072<\/td>\n<\/tr>\n
\u8bcd\u6c47\u8868\u5927\u5c0f<\/td>\n262,144<\/td>\n\u591a\u8bed\u8a00\u652f\u6301，140+ \u8bed\u8a00<\/td>\n<\/tr>\n
\u67b6\u6784\u7279\u70b9<\/td>\nEncoder-Free Unified<\/td>\n\u65e0\u72ec\u7acb\u89c6\u89c9\/\u97f3\u9891\u7f16\u7801\u5668<\/td>\n<\/tr>\n
\u591a\u6a21\u6001\u652f\u6301<\/td>\n\u6587\u672c\/\u56fe\u50cf\/\u97f3\u9891\/\u89c6\u9891<\/td>\n12B \u539f\u751f\u652f\u6301\u97f3\u9891\u8f93\u5165<\/td>\n<\/tr>\n
\u7279\u6b8a\u80fd\u529b<\/td>\n\u5de5\u5177\u8c03\u7528\u3001\u601d\u7ef4\u94fe\u3001Agent<\/td>\n\u539f\u751f\u51fd\u6570\u8c03\u7528\u652f\u6301<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

\u67b6\u6784\u4eae\u70b9：Gemma 4 12B \u6700\u5927\u7684\u521b\u65b0\u662f\u629b\u5f03\u4e86\u4f20\u7edf\u591a\u6a21\u6001\u6a21\u578b\u7684\u300c\u7f16\u7801\u5668+LLM\u300d\u53cc\u5c42\u67b6\u6784\u3002\u56fe\u50cf\u901a\u8fc7\u8f7b\u91cf\u7ea7\u5d4c\u5165\u6a21\u5757\u76f4\u63a5\u8fdb\u5165 LLM \u4e3b\u5e72，\u97f3\u9891\u88ab\u6295\u5c04\u5230\u4e0e\u6587\u672c Token \u76f8\u540c\u7684\u8868\u793a\u7a7a\u95f4，\u7531\u540c\u4e00\u4e2a Decoder-Only Transformer \u7edf\u4e00\u5904\u7406\u4e09\u79cd\u6a21\u6001\u3002\u8fd9\u79cd\u8bbe\u8ba1\u663e\u8457\u964d\u4f4e\u4e86\u63a8\u7406\u5ef6\u8fdf\u548c\u5185\u5b58\u5360\u7528，\u540c\u65f6\u7b80\u5316\u4e86\u90e8\u7f72\u590d\u6742\u5ea6\u3002<\/p>\n

\u6ce8\u610f\u529b\u673a\u5236：\u91c7\u7528\u6df7\u5408\u6ce8\u610f\u529b\u5c42\u8bbe\u8ba1，\u6bcf 5 \u5c42\u6ed1\u52a8\u7a97\u53e3\u6ce8\u610f\u529b（sliding_window=1024）\u540e\u63a5 1 \u5c42\u5168\u6ce8\u610f\u529b，\u5728\u957f\u4e0a\u4e0b\u6587\u6548\u7387\u548c\u5168\u5c40\u611f\u77e5\u80fd\u529b\u4e4b\u95f4\u53d6\u5f97\u5e73\u8861\u3002<\/p>\n

\n

\u4e8c\u3001\u786c\u4ef6\u95e8\u69db：\u4f60\u7684\u7535\u8111\u80fd\u8dd1\u5417？<\/h3>\n

2.1 \u539f\u59cb\u6a21\u578b vs \u91cf\u5316\u6a21\u578b\u5185\u5b58\u5360\u7528<\/h4>\n\n\u914d\u7f6e\u539f\u59cb FP16Q8_0 \u91cf\u5316Q4_K_M \u91cf\u5316\u9002\u7528\u663e\u5361<\/tr>\n\n\n\n\n\n
\u5185\u5b58\u5360\u7528<\/td>\n~24GB<\/td>\n~12GB<\/td>\n~6.5GB<\/td>\n\u2014<\/td>\n<\/tr>\n
\u6700\u4f4e\u663e\u5361<\/td>\nRTX 4090 24GB<\/td>\nRTX 4070 Ti 12GB<\/td>\nRTX 3060 12GB<\/td>\n\u2014<\/td>\n<\/tr>\n
\u63a8\u8350\u663e\u5361<\/td>\nA100 80GB<\/td>\nRTX 4080 16GB<\/td>\nRTX 4060 Ti 16GB<\/td>\n\u2014<\/td>\n<\/tr>\n
Mac \u7edf\u4e00\u5185\u5b58<\/td>\n32GB+<\/td>\n16GB<\/td>\n16GB<\/td>\nM1\/M2\/M3<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

2.2 \u5b9e\u6d4b\u786c\u4ef6\u914d\u7f6e\u53c2\u8003<\/h4>\n

\u914d\u7f6e A：\u6d88\u8d39\u7ea7\u5165\u95e8（\u63a8\u8350）<\/p>\n

    \n
  • GPU：NVIDIA RTX 3060 12GB \/ RTX 4060 Ti 16GB<\/li>\n
  • \u5185\u5b58：16GB DDR4<\/li>\n
  • \u5b58\u50a8：SSD，\u9884\u7559 15GB<\/li>\n
  • \u7cfb\u7edf：Windows 11 \/ Ubuntu 22.04<\/li>\n
  • \u91cf\u5316\u65b9\u6848：Q4_K_M（6.5GB \u6a21\u578b）<\/li>\n
  • \u9884\u671f\u6027\u80fd：\u6587\u672c\u751f\u6210 15-25 tok\/s，\u56fe\u50cf\u63cf\u8ff0 30-50 \u79d2<\/li>\n<\/ul>\n

    \u914d\u7f6e B：MacBook \u65e0\u663e\u5361\u65b9\u6848<\/p>\n

      \n
    • \u82af\u7247：Apple M1\/M2\/M3，16GB \u7edf\u4e00\u5185\u5b58<\/li>\n
    • \u5b58\u50a8：SSD，\u9884\u7559 15GB<\/li>\n
    • \u6846\u67b6：MLX（Apple Silicon \u539f\u751f\u4f18\u5316）<\/li>\n
    • \u91cf\u5316\u65b9\u6848：Q4_K_M \u6216 OptiQ \u6df7\u5408\u7cbe\u5ea6<\/li>\n
    • \u9884\u671f\u6027\u80fd：\u6587\u672c\u751f\u6210 8-15 tok\/s（Metal \u540e\u7aef）<\/li>\n<\/ul>\n

      \u914d\u7f6e C：\u9ad8\u6027\u80fd\u672c\u5730\u670d\u52a1\u5668<\/p>\n

        \n
      • GPU：RTX 4090 24GB \u6216\u53cc\u5361<\/li>\n
      • \u5185\u5b58：32GB+<\/li>\n
      • \u91cf\u5316\u65b9\u6848：Q8_0（12GB）\u6216\u539f\u59cb FP16（24GB）<\/li>\n
      • \u9884\u671f\u6027\u80fd：\u6587\u672c\u751f\u6210 40-60 tok\/s，\u652f\u6301 256K \u957f\u4e0a\u4e0b\u6587<\/li>\n<\/ul>\n
        \n

        \u4e09\u3001\u90e8\u7f72\u65b9\u5f0f\u4e00：Ollama（\u6700\u7b80\u5355，\u4e00\u884c\u547d\u4ee4）<\/h3>\n

        Ollama \u662f\u76ee\u524d\u6700\u53cb\u597d\u7684\u672c\u5730\u5927\u6a21\u578b\u8fd0\u884c\u5de5\u5177，\u652f\u6301\u4e00\u952e\u4e0b\u8f7d\u3001\u81ea\u52a8\u91cf\u5316\u3001OpenAI \u517c\u5bb9 API\u3002<\/p>\n

        3.1 \u5b89\u88c5 Ollama<\/h4>\n

        Windows\/macOS：<\/p>\n

        # \u5b98\u7f51\u4e0b\u8f7d\u5b89\u88c5\u5305\u540e，\u9a8c\u8bc1\u5b89\u88c5<\/span>
        \nollama –version<\/span>
        \n# \u5e94\u663e\u793a ollama version 0.3.x \u6216\u66f4\u9ad8<\/span><\/p>\n

        Linux：<\/p>\n

        curl<\/span> -fsSL<\/span> https:\/\/ollama.com\/install.sh |<\/span> sh<\/span><\/p>\n

        3.2 \u62c9\u53d6\u5e76\u8fd0\u884c Gemma 4 12B<\/h4>\n

        # \u9ed8\u8ba4\u62c9\u53d6 Q4 \u91cf\u5316\u7248\u672c（\u7ea6 6.9GB）<\/span>
        \nollama pull gemma4:12b<\/p>\n

        # \u76f4\u63a5\u8fd0\u884c\u4ea4\u4e92\u5f0f\u5bf9\u8bdd<\/span>
        \nollama run gemma4:12b<\/p>\n

        # \u6307\u5b9a\u4e0a\u4e0b\u6587\u957f\u5ea6（\u9ed8\u8ba4 2048，\u53ef\u6269\u5c55\u5230 128K）<\/span>
        \nollama run gemma4:12b –num-ctx 32768<\/span><\/p>\n

        3.3 \u9a8c\u8bc1\u591a\u6a21\u6001\u80fd\u529b<\/h4>\n

        # \u6d4b\u8bd5\u56fe\u50cf\u7406\u89e3（\u5c06\u56fe\u7247\u653e\u5728\u540c\u4e00\u76ee\u5f55）<\/span>
        \nollama run gemma4:12b "\u63cf\u8ff0\u8fd9\u5f20\u56fe\u7247\u7684\u5185\u5bb9"<\/span> –image<\/span> .\/test.jpg<\/p>\n

        # \u6d4b\u8bd5\u5de5\u5177\u8c03\u7528（\u9700\u8981\u914d\u5408\u5916\u90e8\u5de5\u5177\u94fe）<\/span>
        \nollama run gemma4:12b "\u67e5\u8be2\u4eca\u5929\u7684\u5929\u6c14"<\/span> –tools<\/span><\/p>\n

        3.4 \u542f\u52a8 API \u670d\u52a1<\/h4>\n

        # \u540e\u53f0\u542f\u52a8\u517c\u5bb9 OpenAI \u7684 API \u670d\u52a1<\/span>
        \nollama serve<\/p>\n

        # \u6d4b\u8bd5 API<\/span>
        \ncurl<\/span> http:\/\/localhost:11434\/api\/generate -d<\/span> '{
        \n "model": "gemma4:12b",
        \n "prompt": "\u7528 Python \u5199\u4e00\u4e2a\u5feb\u901f\u6392\u5e8f\u7b97\u6cd5",
        \n "stream": false
        \n}'<\/span><\/p>\n

        \u907f\u5751\u63d0\u793a：<\/p>\n

          \n
        • \u82e5\u51fa\u73b0 CUDA out of memory，\u5c1d\u8bd5\u964d\u4f4e GPU \u5c42\u6570：ollama run gemma4:12b –num-gpu 20<\/li>\n
        • Windows \u4e0b\u82e5 WSL2 \u5185\u5b58\u4e0d\u8db3，\u5728 .wslconfig \u4e2d\u8bbe\u7f6e memory=16GB<\/li>\n
        • \u9996\u6b21\u4e0b\u8f7d\u6a21\u578b\u7ea6 6.9GB，\u5efa\u8bae\u5728\u7f51\u7edc\u7a33\u5b9a\u73af\u5883\u4e0b\u64cd\u4f5c<\/li>\n<\/ul>\n
          \n

          \u56db\u3001\u90e8\u7f72\u65b9\u5f0f\u4e8c：LM Studio（\u56fe\u5f62\u754c\u9762，\u65b0\u624b\u9996\u9009）<\/h3>\n

          LM Studio \u63d0\u4f9b\u4e86\u5b8c\u5168\u56fe\u5f62\u5316\u7684\u64cd\u4f5c\u754c\u9762，\u9002\u5408\u4e0d\u559c\u6b22\u547d\u4ee4\u884c\u7684\u7528\u6237\u3002<\/p>\n

          4.1 \u4e0b\u8f7d\u4e0e\u5b89\u88c5<\/h4>\n
        • \u8bbf\u95ee https:\/\/lmstudio.ai\/ \u4e0b\u8f7d\u5bf9\u5e94\u7cfb\u7edf\u7248\u672c<\/li>\n
        • \u5b89\u88c5\u540e\u9996\u6b21\u542f\u52a8\u4f1a\u81ea\u52a8\u914d\u7f6e\u73af\u5883<\/li>\n

          4.2 \u52a0\u8f7d Gemma 4 12B<\/h4>\n
        • \u5de6\u4fa7\u5bfc\u822a\u680f\u70b9\u51fb\u300cSearch\u300d<\/li>\n
        • \u641c\u7d22\u6846\u8f93\u5165 gemma-4-12b<\/li>\n
        • \u9009\u62e9 google\/gemma-4-12b-it \u6216\u793e\u533a\u91cf\u5316\u7248\u672c（\u5982 bartowski\/gemma-4-12b-it-GGUF）<\/li>\n
        • \u9009\u62e9\u5408\u9002\u7684\u91cf\u5316\u7248\u672c：\n
            \n
          • Q4_K_M（\u63a8\u8350）：6.5GB，\u8d28\u91cf\u635f\u5931 < 2%<\/li>\n
          • Q8_0：12GB，\u63a5\u8fd1\u65e0\u635f<\/li>\n<\/ul>\n<\/li>\n
          • \u70b9\u51fb Download，\u7b49\u5f85\u4e0b\u8f7d\u5b8c\u6210<\/li>\n

            4.3 \u8fd0\u884c\u4e0e\u6d4b\u8bd5<\/h4>\n
          • \u4e0b\u8f7d\u5b8c\u6210\u540e，\u5728\u300cLocal Server\u300d\u6807\u7b7e\u9875\u542f\u52a8\u6a21\u578b<\/li>\n
          • \u9ed8\u8ba4\u52a0\u8f7d Q4_K_M \u7248\u672c，\u4e0a\u4e0b\u6587\u957f\u5ea6\u53ef\u5728\u8bbe\u7f6e\u4e2d\u8c03\u6574\u81f3 32768 \u6216 65536<\/li>\n
          • \u5728 Chat \u754c\u9762\u76f4\u63a5\u5bf9\u8bdd，\u652f\u6301\u62d6\u62fd\u56fe\u7247\u8fdb\u884c\u591a\u6a21\u6001\u6d4b\u8bd5<\/li>\n

            LM Studio \u5b9e\u6d4b\u6570\u636e：<\/p>\n

              \n
            • Q4_K_M \u7248\u672c\u5728 RTX 4060 Ti 16GB \u4e0a\u52a0\u8f7d\u540e\u5269\u4f59\u663e\u5b58\u7ea6 9GB，\u53ef\u540c\u65f6\u8fd0\u884c\u5176\u4ed6\u8f7b\u91cf\u5e94\u7528<\/li>\n
            • \u6587\u672c\u751f\u6210\u901f\u5ea6：18-22 tok\/s（GPU \u6a21\u5f0f）<\/li>\n
            • \u56fe\u50cf\u7406\u89e3：\u5355\u5f20 1080p \u56fe\u7247\u5904\u7406\u7ea6 25-35 \u79d2<\/li>\n<\/ul>\n
              \n

              \u4e94\u3001\u90e8\u7f72\u65b9\u5f0f\u4e09：llama.cpp（\u547d\u4ee4\u884c，\u6781\u81f4\u63a7\u5236）<\/h3>\n

              llama.cpp \u662f\u6700\u5e95\u5c42\u7684\u63a8\u7406\u5f15\u64ce，Ollama \u548c LM Studio \u5747\u57fa\u4e8e\u5b83\u6784\u5efa\u3002\u9002\u5408\u9700\u8981\u7cbe\u7ec6\u63a7\u5236\u7684\u9ad8\u7ea7\u7528\u6237\u3002<\/p>\n

              5.1 \u7f16\u8bd1\u5b89\u88c5<\/h4>\n

              # \u514b\u9686\u4ed3\u5e93<\/span>
              \ngit<\/span> clone https:\/\/github.com\/ggerganov\/llama.cpp.git
              \ncd<\/span> llama.cpp<\/p>\n

              # \u7f16\u8bd1（CUDA \u7248\u672c）<\/span>
              \nmake<\/span> LLAMA_CUDA<\/span>=<\/span>1<\/span> -j$(<\/span>nproc)<\/span><\/span><\/p>\n

              # \u6216\u7f16\u8bd1 CPU \u7248\u672c（\u65e0\u663e\u5361\u65f6）<\/span>
              \nmake<\/span> -j$(<\/span>nproc)<\/span><\/span><\/p>\n

              5.2 \u4e0b\u8f7d GGUF \u6a21\u578b<\/h4>\n

              # \u4ece Hugging Face \u4e0b\u8f7d Q4_K_M \u91cf\u5316\u7248\u672c<\/span>
              \nwget<\/span> https:\/\/huggingface.co\/bartowski\/gemma-4-12b-it-GGUF\/resolve\/main\/gemma-4-12b-it-Q4_K_M.gguf<\/p>\n

              # \u6587\u4ef6\u5927\u5c0f\u7ea6 6.9GB<\/span><\/p>\n

              5.3 \u8fd0\u884c\u63a8\u7406<\/h4>\n

              # \u57fa\u672c\u5bf9\u8bdd\u6a21\u5f0f<\/span>
              \n.\/llama-cli \\\\<\/span>
              \n -m<\/span> .\/gemma-4-12b-it-Q4_K_M.gguf \\\\<\/span>
              \n -c<\/span> 32768<\/span> \\\\<\/span>
              \n –temp<\/span> 0.6<\/span> \\\\<\/span>
              \n –top-k 64<\/span> \\\\<\/span>
              \n –top-p 0.95<\/span> \\\\<\/span>
              \n -p<\/span> "You are a helpful AI assistant."<\/span><\/p>\n

              # \u542f\u52a8 HTTP API \u670d\u52a1<\/span>
              \n.\/llama-server \\\\<\/span>
              \n -m<\/span> .\/gemma-4-12b-it-Q4_K_M.gguf \\\\<\/span>
              \n -c<\/span> 32768<\/span> \\\\<\/span>
              \n –host<\/span> 0.0<\/span>.0.0 \\\\<\/span>
              \n –port<\/span> 8080<\/span><\/p>\n

              5.4 \u9ad8\u7ea7\u53c2\u6570\u8c03\u4f18<\/h4>\n

              # \u542f\u7528 MTP \u6295\u673a\u89e3\u7801（\u9700\u8349\u7a3f\u6a21\u578b，\u63d0\u901f\u7ea6 60%）<\/span>
              \n.\/llama-cli \\\\<\/span>
              \n -m<\/span> .\/gemma-4-12b-it-Q4_K_M.gguf \\\\<\/span>
              \n -md<\/span> .\/gemma-4-12b-it-draft-Q4_K_M.gguf \\\\<\/span>
              \n -c<\/span> 32768<\/span> \\\\<\/span>
              \n –draft<\/span> 8<\/span><\/p>\n

              # \u591a GPU \u5e76\u884c（\u53cc RTX 3090）<\/span>
              \n.\/llama-server \\\\<\/span>
              \n -m<\/span> .\/gemma-4-12b-it-Q4_K_M.gguf \\\\<\/span>
              \n -ts<\/span> 24,24<\/span> \\\\<\/span>
              \n -c<\/span> 65536<\/span><\/p>\n

              \n

              \u516d\u3001\u90e8\u7f72\u65b9\u5f0f\u56db：Python Transformers（\u5f00\u53d1\u8005\u65b9\u6848）<\/h3>\n

              \u5bf9\u4e8e\u9700\u8981\u96c6\u6210\u5230\u73b0\u6709 Python \u9879\u76ee\u7684\u5f00\u53d1\u8005，Hugging Face Transformers \u662f\u6807\u51c6\u9009\u62e9\u3002<\/p>\n

              6.1 \u73af\u5883\u51c6\u5907<\/h4>\n

              pip install<\/span> transformers torch accelerate<\/p>\n

              # \u5982\u9700\u591a\u6a21\u6001\u652f\u6301<\/span>
              \npip install<\/span> Pillow soundfile<\/p>\n

              6.2 \u52a0\u8f7d\u6a21\u578b<\/h4>\n

              from<\/span> transformers import<\/span> AutoModelForCausalLM,<\/span> AutoTokenizer
              \nimport<\/span> torch<\/p>\n

              # \u52a0\u8f7d\u6a21\u578b（\u81ea\u52a8\u4e0b\u8f7d\u5230 ~\/.cache\/huggingface）<\/span>
              \nmodel_id =<\/span> "google\/gemma-4-12b-it"<\/span><\/p>\n

              tokenizer =<\/span> AutoTokenizer.<\/span>from_pretrained(<\/span>model_id)<\/span>
              \nmodel =<\/span> AutoModelForCausalLM.<\/span>from_pretrained(<\/span>
              \n model_id,<\/span>
              \n torch_dtype=<\/span>torch.<\/span>bfloat16,<\/span> # \u6216 float16 \u8282\u7701\u663e\u5b58<\/span>
              \n device_map=<\/span>"auto"<\/span>,<\/span> # \u81ea\u52a8\u5206\u914d\u5c42\u5230 GPU\/CPU<\/span>
              \n attn_implementation=<\/span>"flash_attention_2"<\/span> # \u9700\u5b89\u88c5 flash-attn<\/span>
              \n)<\/span><\/p>\n

              # \u6587\u672c\u751f\u6210<\/span>
              \nprompt =<\/span> "Explain quantum computing in simple terms."<\/span>
              \ninputs =<\/span> tokenizer(<\/span>prompt,<\/span> return_tensors=<\/span>"pt"<\/span>)<\/span>.<\/span>to(<\/span>model.<\/span>device)<\/span><\/p>\n

              outputs =<\/span> model.<\/span>generate(<\/span>
              \n **<\/span>inputs,<\/span>
              \n max_new_tokens=<\/span>512<\/span>,<\/span>
              \n temperature=<\/span>0.7<\/span>,<\/span>
              \n top_p=<\/span>0.95<\/span>,<\/span>
              \n do_sample=<\/span>True<\/span>
              \n)<\/span><\/p>\n

              response =<\/span> tokenizer.<\/span>decode(<\/span>outputs[<\/span>0<\/span>]<\/span>,<\/span> skip_special_tokens=<\/span>True<\/span>)<\/span>
              \nprint<\/span>(<\/span>response)<\/span><\/p>\n

              6.3 \u591a\u6a21\u6001\u8f93\u5165\u793a\u4f8b<\/h4>\n

              from<\/span> PIL import<\/span> Image<\/p>\n

              # \u56fe\u50cf\u7406\u89e3<\/span>
              \nimage =<\/span> Image.<\/span>open<\/span>(<\/span>".\/chart.png"<\/span>)<\/span>
              \ninputs =<\/span> tokenizer(<\/span>
              \n text=<\/span>"Describe this image in detail."<\/span>,<\/span>
              \n images=<\/span>image,<\/span>
              \n return_tensors=<\/span>"pt"<\/span>
              \n)<\/span>.<\/span>to(<\/span>model.<\/span>device)<\/span><\/p>\n

              outputs =<\/span> model.<\/span>generate(<\/span>**<\/span>inputs,<\/span> max_new_tokens=<\/span>256<\/span>)<\/span>
              \nprint<\/span>(<\/span>tokenizer.<\/span>decode(<\/span>outputs[<\/span>0<\/span>]<\/span>)<\/span>)<\/span><\/p>\n

              # \u97f3\u9891\u8f93\u5165（12B \u539f\u751f\u652f\u6301）<\/span>
              \n# audio_input = tokenizer(audio=audio_waveform, return_tensors="pt")<\/span><\/p>\n

              6.4 \u663e\u5b58\u4f18\u5316\u6280\u5de7<\/h4>\n

              # 4-bit \u91cf\u5316\u52a0\u8f7d（\u9700 bitsandbytes）<\/span>
              \nfrom<\/span> transformers import<\/span> BitsAndBytesConfig<\/p>\n

              quant_config =<\/span> BitsAndBytesConfig(<\/span>
              \n load_in_4bit=<\/span>True<\/span>,<\/span>
              \n bnb_4bit_compute_dtype=<\/span>torch.<\/span>bfloat16,<\/span>
              \n bnb_4bit_use_double_quant=<\/span>True<\/span>,<\/span>
              \n bnb_4bit_quant_type=<\/span>"nf4"<\/span>
              \n)<\/span><\/p>\n

              model =<\/span> AutoModelForCausalLM.<\/span>from_pretrained(<\/span>
              \n model_id,<\/span>
              \n quantization_config=<\/span>quant_config,<\/span>
              \n device_map=<\/span>"auto"<\/span>
              \n)<\/span>
              \n# \u91cf\u5316\u540e\u663e\u5b58\u5360\u7528\u964d\u81f3 ~7GB<\/span><\/p>\n

              \n

              \u4e03\u3001\u91cf\u5316\u65b9\u6848\u9009\u578b\u6307\u5357<\/h3>\n\n\u91cf\u5316\u65b9\u6848\u6a21\u578b\u5927\u5c0f\u663e\u5b58\u9700\u6c42\u8d28\u91cf\u635f\u5931\u9002\u7528\u573a\u666f<\/tr>\n\n\n\n\n\n\n\n
              \u539f\u59cb FP16<\/td>\n~24GB<\/td>\n24GB+<\/td>\n0%<\/td>\n\u7814\u7a76\/\u9ad8\u7cbe\u5ea6\u4efb\u52a1<\/td>\n<\/tr>\n
              Q8_0<\/td>\n~12GB<\/td>\n14GB+<\/td>\n< 1%<\/td>\n\u751f\u4ea7\u529b\/\u4ee3\u7801\u751f\u6210<\/td>\n<\/tr>\n
              Q6_K<\/td>\n~9GB<\/td>\n12GB+<\/td>\n< 2%<\/td>\n\u5e73\u8861\u65b9\u6848<\/td>\n<\/tr>\n
              Q4_K_M（\u63a8\u8350）<\/td>\n~6.9GB<\/td>\n8GB+<\/td>\n< 3%<\/td>\n\u6d88\u8d39\u7ea7\u90e8\u7f72<\/td>\n<\/tr>\n
              Q4_0<\/td>\n~6.5GB<\/td>\n8GB+<\/td>\n~5%<\/td>\n\u6781\u9650\u4f4e\u663e\u5b58<\/td>\n<\/tr>\n
              QAT-4bit<\/td>\n~6.9GB<\/td>\n8GB+<\/td>\n< 2%<\/td>\n\u8d28\u91cf\u654f\u611f+\u4f4e\u663e\u5b58<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

              \u9009\u578b\u5efa\u8bae：<\/p>\n

                \n
              • RTX 3060\/4060 Ti 12GB：\u9009 Q4_K_M，\u5269\u4f59\u663e\u5b58\u5145\u88d5，\u53ef\u5f00 32K \u4e0a\u4e0b\u6587<\/li>\n
              • RTX 4090 24GB：\u9009 Q8_0，\u8d28\u91cf\u63a5\u8fd1\u65e0\u635f，\u53ef\u5f00 128K \u4e0a\u4e0b\u6587<\/li>\n
              • Mac M1\/M2 16GB：\u9009 Q4_K_M + MLX \u6846\u67b6，Metal \u540e\u7aef\u6548\u7387\u6700\u9ad8<\/li>\n
              • \u670d\u52a1\u5668 A100 80GB：\u76f4\u63a5 FP16，\u53d1\u6325 256K \u5168\u4e0a\u4e0b\u6587\u80fd\u529b<\/li>\n<\/ul>\n
                \n

                \u516b\u3001\u6027\u80fd\u5b9e\u6d4b\u6570\u636e<\/h3>\n

                \u4ee5\u4e0b\u6570\u636e\u57fa\u4e8e RTX 4060 Ti 16GB + i5-13600K + 32GB DDR4 \u5b9e\u6d4b：<\/p>\n

                8.1 \u6587\u672c\u751f\u6210\u901f\u5ea6<\/h4>\n\n\u91cf\u5316\u65b9\u6848\u6279\u5927\u5c0f\u4e0a\u4e0b\u6587\u957f\u5ea6tok\/s\u663e\u5b58\u5360\u7528<\/tr>\n\n\n\n\n\n\n
                Q4_K_M<\/td>\n1<\/td>\n2048<\/td>\n22.3<\/td>\n7.2GB<\/td>\n<\/tr>\n
                Q4_K_M<\/td>\n1<\/td>\n8192<\/td>\n18.7<\/td>\n8.1GB<\/td>\n<\/tr>\n
                Q4_K_M<\/td>\n1<\/td>\n32768<\/td>\n14.2<\/td>\n10.5GB<\/td>\n<\/tr>\n
                Q8_0<\/td>\n1<\/td>\n2048<\/td>\n19.8<\/td>\n13.1GB<\/td>\n<\/tr>\n
                Q8_0<\/td>\n1<\/td>\n8192<\/td>\n16.5<\/td>\n14.8GB<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

                8.2 \u591a\u6a21\u6001\u5904\u7406\u8017\u65f6<\/h4>\n\n\u4efb\u52a1\u7c7b\u578b\u8f93\u5165\u89c4\u683cQ4_K_M \u8017\u65f6Q8_0 \u8017\u65f6<\/tr>\n\n\n\n\n\n
                \u56fe\u50cf\u63cf\u8ff0<\/td>\n1080p JPG<\/td>\n32 \u79d2<\/td>\n28 \u79d2<\/td>\n<\/tr>\n
                OCR+\u7406\u89e3<\/td>\nA4 \u626b\u63cf\u4ef6<\/td>\n45 \u79d2<\/td>\n38 \u79d2<\/td>\n<\/tr>\n
                \u97f3\u9891\u8f6c\u5f55<\/td>\n30 \u79d2 WAV<\/td>\n18 \u79d2<\/td>\n15 \u79d2<\/td>\n<\/tr>\n
                \u89c6\u9891\u5206\u6790<\/td>\n10 \u79d2 MP4<\/td>\n120 \u79d2<\/td>\n105 \u79d2<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

                8.3 \u957f\u4e0a\u4e0b\u6587\u538b\u529b\u6d4b\u8bd5<\/h4>\n

                \u4e0a\u4e0b\u6587\u957f\u5ea6 | \u663e\u5b58\u5360\u7528(Q4) | \u9996\u6b21 token \u5ef6\u8fdf | \u7a33\u5b9a\u6027
                \n———–|————-|—————–|——–
                \n4K | 7.1GB | 0.8s | 100%
                \n16K | 7.8GB | 1.2s | 100%
                \n32K | 8.9GB | 1.8s | 100%
                \n64K | 10.5GB | 2.9s | 100%
                \n128K | 13.2GB | 5.1s | 99%
                \n256K | 18.6GB | 9.8s | 95% (\u5076\u6709\u6ce8\u610f\u529b\u6f02\u79fb)<\/p>\n

                \n

                \u4e5d\u3001\u5b9e\u9645\u5e94\u7528\u573a\u666f<\/h3>\n

                9.1 \u672c\u5730\u4ee3\u7801\u52a9\u624b（\u6700\u5f3a\u573a\u666f）<\/h4>\n

                Gemma 4 12B \u5728\u4ee3\u7801\u751f\u6210\u4efb\u52a1\u4e0a\u8868\u73b0\u4f18\u5f02\u3002\u914d\u5408 Continue.dev \u6216 Cursor \u7684\u672c\u5730\u6a21\u578b\u9009\u9879，\u53ef\u5b9e\u73b0：<\/p>\n

                  \n
                • \u79bb\u7ebf\u4ee3\u7801\u8865\u5168（\u652f\u6301 Python\/JavaScript\/C++ \u7b49\u4e3b\u6d41\u8bed\u8a00）<\/li>\n
                • \u4ee3\u7801\u91cd\u6784\u5efa\u8bae<\/li>\n
                • \u81ea\u7136\u8bed\u8a00\u8f6c\u4ee3\u7801（\u5982\u300c\u5199\u4e00\u4e2a\u5e26\u7f13\u5b58\u7684\u722c\u866b\u300d）<\/li>\n
                • Bug \u8bca\u65ad\u4e0e\u4fee\u590d<\/li>\n<\/ul>\n

                  \u914d\u7f6e\u793a\u4f8b（Continue.dev）：<\/p>\n

                  {<\/span>
                  \n "models"<\/span>:<\/span> [<\/span>{<\/span>
                  \n "title"<\/span>:<\/span> "Gemma 4 12B Local"<\/span>,<\/span>
                  \n "provider"<\/span>:<\/span> "ollama"<\/span>,<\/span>
                  \n "model"<\/span>:<\/span> "gemma4:12b"<\/span>,<\/span>
                  \n "apiBase"<\/span>:<\/span> "http:\/\/localhost:11434"<\/span>
                  \n }<\/span>]<\/span>
                  \n}<\/span><\/p>\n

                  9.2 \u591a\u6a21\u6001\u672c\u5730\u77e5\u8bc6\u5e93<\/h4>\n

                  \u5229\u7528 256K \u4e0a\u4e0b\u6587\u7a97\u53e3，\u53ef\u5c06\u6574\u672c\u6280\u672f\u624b\u518c（\u7ea6 10 \u4e07\u5b57）\u76f4\u63a5\u585e\u8fdb\u4e0a\u4e0b\u6587：<\/p>\n

                    \n
                  • \u4e0a\u4f20 PDF\/\u56fe\u7247\u6587\u6863<\/li>\n
                  • \u76f4\u63a5\u63d0\u95ee\u300c\u7b2c\u4e09\u7ae0\u8bb2\u5230\u7684\u914d\u7f6e\u53c2\u6570\u6709\u54ea\u4e9b？\u300d<\/li>\n
                  • \u65e0\u9700 RAG \u67b6\u6784，\u7b80\u5316\u90e8\u7f72<\/li>\n<\/ul>\n

                    9.3 \u9690\u79c1\u654f\u611f\u573a\u666f<\/h4>\n

                    \u533b\u7597\u3001\u6cd5\u5f8b\u3001\u91d1\u878d\u884c\u4e1a\u7684\u672c\u5730\u6587\u6863\u5206\u6790：<\/p>\n

                      \n
                    • \u75c5\u5386\u56fe\u50cf OCR + \u7ed3\u6784\u5316\u63d0\u53d6<\/li>\n
                    • \u5408\u540c\u6761\u6b3e\u5bf9\u6bd4\u5206\u6790<\/li>\n
                    • \u8d22\u52a1\u62a5\u8868\u6570\u636e\u63d0\u53d6<\/li>\n<\/ul>\n

                      9.4 \u79bb\u7ebf\u521b\u4f5c\u52a9\u624b<\/h4>\n

                      \u65e0\u7f51\u7edc\u73af\u5883\u4e0b\u7684\u5199\u4f5c\u8f85\u52a9：<\/p>\n

                        \n
                      • \u957f\u6587\u7eed\u5199（\u5229\u7528 256K \u4e0a\u4e0b\u6587\u4fdd\u6301\u5168\u6587\u8fde\u8d2f\u6027）<\/li>\n
                      • \u591a\u8bed\u8a00\u7ffb\u8bd1（140+ \u8bed\u8a00\u539f\u751f\u652f\u6301）<\/li>\n
                      • \u521b\u610f\u5934\u8111\u98ce\u66b4<\/li>\n<\/ul>\n
                        \n

                        \u5341\u3001\u5e38\u89c1\u5751\u4e0e\u89e3\u51b3\u65b9\u6848<\/h3>\n

                        \u5751 1：Ollama \u4e0b\u8f7d\u4e2d\u65ad\/\u901f\u5ea6\u6162<\/h4>\n

                        \u73b0\u8c61：ollama pull gemma4:12b \u5361\u5728\u67d0\u4e2a\u8fdb\u5ea6\u4e0d\u52a8\u3002<\/p>\n

                        \u89e3\u51b3：<\/p>\n

                        # \u8bbe\u7f6e\u955c\u50cf\u52a0\u901f（\u56fd\u5185\u7528\u6237）<\/span>
                        \nexport<\/span> OLLAMA_HOST<\/span>=<\/span>"0.0.0.0"<\/span>
                        \n# \u6216\u4f7f\u7528 Hugging Face \u955c\u50cf\u624b\u52a8\u4e0b\u8f7d GGUF \u540e\u5bfc\u5165 Ollama<\/span>
                        \n# \u8be6\u89c1\u7f51\u76d8\u4e2d\u7684\u300cOllama \u6a21\u578b\u5bfc\u5165\u811a\u672c\u300d<\/span><\/p>\n

                        \u5751 2：CUDA out of memory<\/h4>\n

                        \u73b0\u8c61：\u6a21\u578b\u52a0\u8f7d\u65f6\u62a5\u663e\u5b58\u4e0d\u8db3\u3002<\/p>\n

                        \u89e3\u51b3：<\/p>\n

                        # Ollama \u964d\u4f4e GPU \u5c42\u6570<\/span>
                        \nollama run gemma4:12b –num-gpu 24<\/span><\/p>\n

                        # llama.cpp \u9650\u5236\u4e0a\u4e0b\u6587\u957f\u5ea6<\/span>
                        \n.\/llama-cli -m<\/span> model.gguf -c<\/span> 8192<\/span> # \u964d\u4f4e\u4e0a\u4e0b\u6587<\/span><\/p>\n

                        # Transformers \u542f\u7528 CPU \u5378\u8f7d<\/span>
                        \nmodel =<\/span> AutoModelForCausalLM.from_pretrained(<\/span>
                        \n model_id,
                        \n device_map<\/span>=<\/span>"auto"<\/span>,
                        \n max_memory<\/span>=<\/span>{<\/span>0<\/span>: "8GiB"<\/span>, "cpu"<\/span>:<\/span> "32GiB"<\/span>}<\/span> # GPU 8GB + CPU 32GB<\/span>
                        \n)<\/span><\/p>\n

                        \u5751 3：\u591a\u6a21\u6001\u8f93\u5165\u65e0\u54cd\u5e94<\/h4>\n

                        \u73b0\u8c61：\u4e0a\u4f20\u56fe\u7247\u540e\u6a21\u578b\u53ea\u8fd4\u56de\u6587\u672c\u63cf\u8ff0，\u4e0d\u5206\u6790\u56fe\u50cf\u5185\u5bb9\u3002<\/p>\n

                        \u89e3\u51b3：<\/p>\n

                          \n
                        • \u786e\u8ba4\u4f7f\u7528\u7684\u662f Gemma 4 12B（\u800c\u975e E2B\/E4B \u5c0f\u7248\u672c，\u5c0f\u7248\u672c\u65e0\u591a\u6a21\u6001）<\/li>\n
                        • \u68c0\u67e5\u662f\u5426\u6b63\u786e\u52a0\u8f7d\u4e86 mmproj \u89c6\u89c9\u6295\u5f71\u6587\u4ef6（llama.cpp \u9700\u8981\u5355\u72ec\u4e0b\u8f7d）<\/li>\n
                        • \u56fe\u50cf\u683c\u5f0f\u5efa\u8bae JPG\/PNG，\u907f\u514d WebP \u7279\u6b8a\u7f16\u7801<\/li>\n<\/ul>\n

                          \u5751 4：\u957f\u4e0a\u4e0b\u6587\u4e0b\u8f93\u51fa\u8d28\u91cf\u4e0b\u964d<\/h4>\n

                          \u73b0\u8c61：\u8d85\u8fc7 64K \u4e0a\u4e0b\u6587\u540e，\u6a21\u578b\u5f00\u59cb\u300c\u9057\u5fd8\u300d\u524d\u6587\u5185\u5bb9\u3002<\/p>\n

                          \u89e3\u51b3：<\/p>\n

                            \n
                          • \u542f\u7528 Flash Attention 2（attn_implementation="flash_attention_2"）<\/li>\n
                          • \u964d\u4f4e\u6e29\u5ea6\u53c2\u6570\u81f3 0.3-0.5，\u63d0\u9ad8\u786e\u5b9a\u6027<\/li>\n
                          • \u8d85\u8fc7 128K \u65f6，\u5efa\u8bae\u5206\u6bb5\u5904\u7406\u800c\u975e\u5355\u8f6e\u8d85\u957f\u4e0a\u4e0b\u6587<\/li>\n<\/ul>\n

                            \u5751 5：Windows WSL2 \u5185\u5b58\u9650\u5236<\/h4>\n

                            \u73b0\u8c61：WSL2 \u4e2d Ollama \u8fd0\u884c\u4e00\u6bb5\u65f6\u95f4\u540e\u5d29\u6e83\u3002<\/p>\n

                            \u89e3\u51b3：<\/p>\n

                            # \u7f16\u8f91 ~\/.wslconfig<\/span>
                            \n[<\/span>wsl2]<\/span>
                            \nmemory<\/span>=<\/span>16GB
                            \nprocessors<\/span>=<\/span>8<\/span>
                            \nswap<\/span>=<\/span>8GB<\/p>\n

                            \n

                            \u5341\u4e00\u3001\u4e0e\u5176\u4ed6 12B \u7ea7\u6a21\u578b\u5bf9\u6bd4<\/h3>\n\n\u5bf9\u6bd4\u9879Gemma 4 12BQwen3.5 9BLlama 3.1 8BDeepSeek V3.2 10B<\/tr>\n\n\n\n\n\n\n\n\n\n
                            \u4e0a\u4e0b\u6587\u957f\u5ea6<\/td>\n256K<\/td>\n128K<\/td>\n128K<\/td>\n64K<\/td>\n<\/tr>\n
                            \u591a\u6a21\u6001<\/td>\n\u539f\u751f\u652f\u6301<\/td>\n\u6587\u672c+\u56fe\u50cf<\/td>\n\u4ec5\u6587\u672c<\/td>\n\u6587\u672c+\u56fe\u50cf<\/td>\n<\/tr>\n
                            \u97f3\u9891\u652f\u6301<\/td>\n\u2705 \u539f\u751f<\/td>\n\u274c<\/td>\n\u274c<\/td>\n\u274c<\/td>\n<\/tr>\n
                            \u91cf\u5316\u540e\u5927\u5c0f<\/td>\n6.9GB (Q4)<\/td>\n5.8GB (Q4)<\/td>\n4.9GB (Q4)<\/td>\n6.2GB (Q4)<\/td>\n<\/tr>\n
                            \u4ee3\u7801\u80fd\u529b<\/td>\n\u5f3a<\/td>\n\u5f3a<\/td>\n\u4e2d\u7b49<\/td>\n\u6781\u5f3a<\/td>\n<\/tr>\n
                            \u4e2d\u6587\u80fd\u529b<\/td>\n\u826f\u597d<\/td>\n\u6781\u5f3a<\/td>\n\u4e2d\u7b49<\/td>\n\u5f3a<\/td>\n<\/tr>\n
                            \u5de5\u5177\u8c03\u7528<\/td>\n\u539f\u751f<\/td>\n\u539f\u751f<\/td>\n\u6709\u9650<\/td>\n\u539f\u751f<\/td>\n<\/tr>\n
                            Arena \u8bc4\u5206<\/td>\n~1450*<\/td>\n~1420<\/td>\n~1380<\/td>\n~1460<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

                            *\u6ce8：Gemma 4 12B \u65e0\u72ec\u7acb Arena \u8bc4\u5206，\u53c2\u8003\u540c\u7cfb\u5217 31B Dense \u7684 1452 \u5206\u63a8\u7b97\u3002<\/p>\n

                            \u9009\u578b\u5efa\u8bae：<\/p>\n

                              \n
                            • \u8981\u591a\u6a21\u6001+\u97f3\u9891\u2192 Gemma 4 12B（\u552f\u4e00\u539f\u751f\u652f\u6301\u97f3\u9891\u7684 12B \u7ea7\u6a21\u578b）<\/li>\n
                            • \u8981\u4e2d\u6587\u6700\u5f3a\u2192 Qwen3.5 9B<\/li>\n
                            • \u8981\u4ee3\u7801\u6700\u5f3a\u2192 DeepSeek V3.2<\/li>\n
                            • \u8981\u751f\u6001\u6700\u6210\u719f\u2192 Llama 3.1 8B<\/li>\n<\/ul>\n
                              \n

                              \u5341\u4e8c\u3001\u603b\u7ed3<\/h3>\n

                              Gemma 4 12B Unified \u662f 2026 \u5e74\u5f00\u6e90\u6a21\u578b\u9886\u57df\u7684\u4e00\u6b21\u91cd\u8981\u8fed\u4ee3\u3002\u5b83\u5728\u4e09\u4e2a\u7ef4\u5ea6\u4e0a\u5b9e\u73b0\u4e86\u7a81\u7834：<\/p>\n

                            • \u67b6\u6784\u521b\u65b0：\u65e0\u7f16\u7801\u5668\u7edf\u4e00\u591a\u6a21\u6001\u67b6\u6784，\u964d\u4f4e\u4e86\u90e8\u7f72\u590d\u6742\u5ea6\u548c\u63a8\u7406\u5ef6\u8fdf<\/li>\n
                            • \u4e0a\u4e0b\u6587\u957f\u5ea6：256K tokens \u5728 12B \u7ea7\u522b\u4e2d\u5904\u4e8e\u9886\u5148\u5730\u4f4d<\/li>\n
                            • \u786c\u4ef6\u53cb\u597d：Q4 \u91cf\u5316\u540e 6.5GB \u663e\u5b58\u95e8\u69db，\u8ba9\u6d88\u8d39\u7ea7\u663e\u5361\u4e5f\u80fd\u8dd1\u8d77\u591a\u6a21\u6001\u5927\u6a21\u578b<\/li>\n

                              \u5bf9\u4e8e\u672c\u5730 AI \u7231\u597d\u8005\u548c\u5f00\u53d1\u8005\u6765\u8bf4，Gemma 4 12B \u662f\u76ee\u524d\u300c\u529f\u80fd\u6700\u5168\u9762\u300d\u7684\u672c\u5730\u90e8\u7f72\u9009\u62e9\u2014\u2014\u5b83\u4e0d\u53ea\u662f\u6587\u672c\u6a21\u578b，\u800c\u662f\u80fd\u770b\u3001\u80fd\u542c\u3001\u80fd\u7406\u89e3\u957f\u6587\u6863\u7684\u591a\u6a21\u6001\u52a9\u624b\u3002<\/p>\n

                              \u5982\u679c\u4f60\u5df2\u7ecf\u6709\u4e00\u5f20 RTX 3060 \u4ee5\u4e0a\u7684\u663e\u5361，\u6216\u8005\u4e00\u53f0 16GB \u5185\u5b58\u7684 MacBook，\u73b0\u5728\u5c31\u662f\u5c1d\u8bd5\u672c\u5730\u90e8\u7f72\u7684\u6700\u4f73\u65f6\u673a\u3002<\/p>\n

                              \n

                              \u8d44\u6e90\u4e0b\u8f7d：<\/p>\n

                              \u7f51\u76d8\u5185\u5305\u542b\u672c\u6587\u63d0\u5230\u7684\u6240\u6709\u914d\u7f6e\u6587\u4ef6\u3001\u4e00\u952e\u90e8\u7f72\u811a\u672c\u548c\u91cf\u5316\u6a21\u578b\u4e0b\u8f7d\u94fe\u63a5：<\/p>\n

                                \n
                              • \u5de5\u4f5c\u6d41\u6a21\u677f + \u914d\u7f6e\u6587\u4ef6\u5408\u96c6：https:\/\/pan.quark.cn\/s\/a7d5cb0d9fbe<\/li>\n
                              • Ollama \u6a21\u578b\u5305 + \u90e8\u7f72\u5de5\u5177\u7bb1：https:\/\/pan.quark.cn\/s\/a8a75ed5ef5a<\/li>\n<\/ul>\n
                                \n

                                \u672c\u6587\u6240\u6709\u5b9e\u6d4b\u6570\u636e\u57fa\u4e8e RTX 4060 Ti 16GB + i5-13600K + 32GB DDR4 \u73af\u5883，\u4e0d\u540c\u786c\u4ef6\u914d\u7f6e\u4e0b\u6027\u80fd\u4f1a\u6709\u5dee\u5f02\u3002\u5982\u6709\u90e8\u7f72\u95ee\u9898，\u6b22\u8fce\u5728\u8bc4\u8bba\u533a\u7559\u8a00\u4ea4\u6d41\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

                                Google Gemma 4 12B Unified \u672c\u5730\u90e8\u7f72\u5b8c\u5168\u6307\u5357：16GB \u663e\u5b58\u8dd1\u8d77\u591a\u6a21\u6001\u5927\u6a21\u578b\u4f5c\u8005：\u8d5b\u535a\u4ed3\u9f20 | 2026-09-142026 \u5e74 8 \u6708，Google DeepMind \u6b63\u5f0f\u5f00\u6e90\u4e86 Gemma 4 \u7cfb\u5217\u6a21\u578b\u3002\u4f5c\u4e3a Gemma \u5bb6\u65cf\u9996\u6b21\u5f15\u5165\u65e0\u7f16\u7801\u5668\u7edf\u4e00\u591a\u6a21\u6001\u67b6\u6784（Encoder-Free Unified）\u7684\u7248\u672c，Gemma 4 12B \u5728 12B \u53c2\u6570\u89c4\u6a21\u4e0a\u5b9e\u73b0\u4e86\u6587\u672c\u3001<\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[66,52,50],"topic":[],"class_list":["post-105298","post","type-post","status-publish","format-standard","hentry","category-server","tag-ai","tag-aigc","tag-50"],"yoast_head":"\nGoogle Gemma 4 12B Unified \u672c\u5730\u90e8\u7f72\u5b8c\u5168\u6307\u5357\uff1a16GB \u663e\u5b58\u8dd1\u8d77\u591a\u6a21\u6001\u5927\u6a21\u578b - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.wsisp.com\/helps\/105298.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"Google Gemma 4 12B Unified \u672c\u5730\u90e8\u7f72\u5b8c\u5168\u6307\u5357\uff1a16GB \u663e\u5b58\u8dd1\u8d77\u591a\u6a21\u6001\u5927\u6a21\u578b - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"Google Gemma 4 12B Unified \u672c\u5730\u90e8\u7f72\u5b8c\u5168\u6307\u5357：16GB \u663e\u5b58\u8dd1\u8d77\u591a\u6a21\u6001\u5927\u6a21\u578b\u4f5c\u8005：\u8d5b\u535a\u4ed3\u9f20 | 2026-09-142026 \u5e74 8 \u6708，Google DeepMind \u6b63\u5f0f\u5f00\u6e90\u4e86 Gemma 4 \u7cfb\u5217\u6a21\u578b\u3002\u4f5c\u4e3a Gemma \u5bb6\u65cf\u9996\u6b21\u5f15\u5165\u65e0\u7f16\u7801\u5668\u7edf\u4e00\u591a\u6a21\u6001\u67b6\u6784（Encoder-Free Unified）\u7684\u7248\u672c，Gemma 4 12B \u5728 12B \u53c2\u6570\u89c4\u6a21\u4e0a\u5b9e\u73b0\u4e86\u6587\u672c\u3001\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/105298.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-09-15T00:24:46+00:00\" \/>\n<meta name=\"author\" content=\"admin\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"\u4f5c\u8005\" \/>\n\t<meta name=\"twitter:data1\" content=\"admin\" \/>\n\t<meta name=\"twitter:label2\" content=\"\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4\" \/>\n\t<meta name=\"twitter:data2\" content=\"6 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/105298.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/105298.html\",\"name\":\"Google Gemma 4 12B Unified \u672c\u5730\u90e8\u7f72\u5b8c\u5168\u6307\u5357\uff1a16GB \u663e\u5b58\u8dd1\u8d77\u591a\u6a21\u6001\u5927\u6a21\u578b - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-09-15T00:24:46+00:00\",\"dateModified\":\"2026-09-15T00:24:46+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/105298.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/105298.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/105298.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"Google Gemma 4 12B Unified \u672c\u5730\u90e8\u7f72\u5b8c\u5168\u6307\u5357\uff1a16GB \u663e\u5b58\u8dd1\u8d77\u591a\u6a21\u6001\u5927\u6a21\u578b\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\",\"url\":\"https:\/\/www.wsisp.com\/helps\/\",\"name\":\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"description\":\"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}\"},\"query-input\":\"required name=search_term_string\"}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\",\"name\":\"admin\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"contentUrl\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"caption\":\"admin\"},\"sameAs\":[\"http:\/\/wp.wsisp.com\"],\"url\":\"https:\/\/www.wsisp.com\/helps\/author\/admin\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"Google Gemma 4 12B Unified \u672c\u5730\u90e8\u7f72\u5b8c\u5168\u6307\u5357\uff1a16GB \u663e\u5b58\u8dd1\u8d77\u591a\u6a21\u6001\u5927\u6a21\u578b - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.wsisp.com\/helps\/105298.html","og_locale":"zh_CN","og_type":"article","og_title":"Google Gemma 4 12B Unified \u672c\u5730\u90e8\u7f72\u5b8c\u5168\u6307\u5357\uff1a16GB \u663e\u5b58\u8dd1\u8d77\u591a\u6a21\u6001\u5927\u6a21\u578b - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"Google Gemma 4 12B Unified \u672c\u5730\u90e8\u7f72\u5b8c\u5168\u6307\u5357：16GB \u663e\u5b58\u8dd1\u8d77\u591a\u6a21\u6001\u5927\u6a21\u578b\u4f5c\u8005：\u8d5b\u535a\u4ed3\u9f20 | 2026-09-142026 \u5e74 8 \u6708，Google DeepMind \u6b63\u5f0f\u5f00\u6e90\u4e86 Gemma 4 \u7cfb\u5217\u6a21\u578b\u3002\u4f5c\u4e3a Gemma \u5bb6\u65cf\u9996\u6b21\u5f15\u5165\u65e0\u7f16\u7801\u5668\u7edf\u4e00\u591a\u6a21\u6001\u67b6\u6784（Encoder-Free Unified）\u7684\u7248\u672c，Gemma 4 12B \u5728 12B \u53c2\u6570\u89c4\u6a21\u4e0a\u5b9e\u73b0\u4e86\u6587\u672c\u3001","og_url":"https:\/\/www.wsisp.com\/helps\/105298.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-09-15T00:24:46+00:00","author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"6 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/105298.html","url":"https:\/\/www.wsisp.com\/helps\/105298.html","name":"Google Gemma 4 12B Unified \u672c\u5730\u90e8\u7f72\u5b8c\u5168\u6307\u5357\uff1a16GB \u663e\u5b58\u8dd1\u8d77\u591a\u6a21\u6001\u5927\u6a21\u578b - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-09-15T00:24:46+00:00","dateModified":"2026-09-15T00:24:46+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/105298.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/105298.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/105298.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"Google Gemma 4 12B Unified \u672c\u5730\u90e8\u7f72\u5b8c\u5168\u6307\u5357\uff1a16GB \u663e\u5b58\u8dd1\u8d77\u591a\u6a21\u6001\u5927\u6a21\u578b"}]},{"@type":"WebSite","@id":"https:\/\/www.wsisp.com\/helps\/#website","url":"https:\/\/www.wsisp.com\/helps\/","name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","description":"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}"},"query-input":"required name=search_term_string"}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41","name":"admin","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/","url":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","contentUrl":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","caption":"admin"},"sameAs":["http:\/\/wp.wsisp.com"],"url":"https:\/\/www.wsisp.com\/helps\/author\/admin"}]}},"_links":{"self":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/105298","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/comments?post=105298"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/105298\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=105298"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=105298"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=105298"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=105298"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}