[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-how-to-run-gemma-4-locally-unsloth-zh":3,"article-related-how-to-run-gemma-4-locally-unsloth-zh":31,"series-industry-8041b1f8-e409-44dc-b574-210938430234":79},{"id":4,"slug":5,"title":6,"content":7,"summary":8,"source":9,"source_url":10,"author":11,"image_url":12,"cover_image":12,"category":13,"language":14,"translated_content":11,"related_article_id":15,"keywords":16,"key_takeaways":23,"views":27,"created_at":28,"published_at":29,"topic_cluster_id":30},"8041b1f8-e409-44dc-b574-210938430234","how-to-run-gemma-4-locally-unsloth-zh","怎麼在本機跑 Gemma 4","\u003Cp data-speakable=\"summary\">用 Unsloth Studio 或 llama.cpp 在本機下載、啟動並聊天 Gemma 4。\u003C\u002Fp>\u003Cp>這篇給想把 \u003Ca href=\"\u002Ftag\u002Fgoogle\">Google\u003C\u002Fa> Gemma 4 跑在筆電、桌機或\u003Ca href=\"\u002Fnews\u002Fvibe-coding-enterprise-software-change-management-zh\">工作\u003C\u002Fa>站上的開發者。照著做完，你會拿到一套可離線使用的本機流程，包含選型、下載、啟動、聊天設定，以及思考模式和多模態輸入的基本做法。\u003C\u002Fp>\u003Cp>你可以走 \u003Ca href=\"https:\u002F\u002Funsloth.ai\u002Fdocs\" target=\"_blank\" rel=\"noopener noreferrer\">Unsloth 文件\u003C\u002Fa>與 \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Funslothai\u002Funsloth\" target=\"_blank\" rel=\"noopener noreferrer\">Unsloth GitHub\u003C\u002Fa> 的瀏覽器介面，也可以走 \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fggerganov\u002Fllama.cpp\" target=\"_blank\" rel=\"noopener noreferrer\">llama.cpp GitHub\u003C\u002Fa> 的直接推理流程。Gemma 4 採用 Apache-2.0 授權，部分型號支援文字、圖片與音訊，量化後可在較小記憶體的機器上執行。\u003C\u002Fp>\u003Ch2>開始之前\u003C\u002Fh2>\u003Cul>\u003Cli>Google 或 Hugging Face 帳號，用來下載模型。\u003C\u002Fli>\u003Cli>本機作業系統為 macOS、Windows、Linux，或 Windows 上的 WSL。\u003C\u002Fli>\u003Cli>Python 3.10+，用於 Unsloth Studio 流程。\u003C\u002Fli>\u003Cli>CMake 3.22+ 與 C++ 編譯器，用於建置 llama.cpp。\u003C\u002Fli>\u003Cli>Git 2.30+，用於取得原始碼。\u003C\u002Fli>\u003Cli>Hugging Face CLI 或 pip 存取權，用於模型下載。\u003C\u002Fli>\u003Cli>NVIDIA GPU 非必需，但可明顯提升推理速度。\u003C\u002Fli>\u003Cli>記憶體至少 8 GB 可跑 Gemma-4-12B 4-bit，或 5 GB 可跑 E2B 4-bit。\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>Step 1: 選定 Gemma 4 型號\u003C\u002Fh2>\u003Cp>先決定\u003Ca href=\"\u002Fnews\u002Fllm-leaderboard-2026-300-models-ranked-zh\">模型\u003C\u002Fa>大小，才能避免下載後才發現記憶體不夠。Gemma 4 有 E2B、E4B、12B Unified、26B-A4B 與 31B 等版本，差異主要在速度、品質與佔用空間。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1780777065583-koml.png\" alt=\"怎麼在本機跑 Gemma 4\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>如果你的目標是筆電或邊緣裝置，先從 E2B 或 E4B 開始；如果要兼顧多模態與可用性，12B Unified 比較平衡；如果你有較大的記憶體預算，26B-A4B 與 31B 會提供更高品質。\u003C\u002Fp>\u003Cp>你應該能寫下具名的記憶體預算，例如「12B 4-bit 需要約 8 GB」、「31B 4-bit 需要約 20 GB」。\u003C\u002Fp>\u003Ch2>Step 2: 安裝 Unsloth Studio\u003C\u002Fh2>\u003Cp>這一步的目的，是先拿到一個能在瀏覽器裡完成搜尋、下載與聊天的本機介面。Unsloth Studio 支援 GGUF 與 MLX 檔案，也會幫你套用常見的推理參數。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1780777068811-r50c.png\" alt=\"怎麼在本機跑 Gemma 4\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>先用 pip 安裝，再啟動本機服務，最後用瀏覽器打開介面並建立第一次登入密碼。\u003C\u002Fp>\u003Cpre>\u003Ccode>python -m pip install unsloth-studio\u003C\u002Fcode>\u003C\u002Fpre>\u003Cp>你應該看到 Studio UI 出現在 \u003Ccode>http:\u002F\u002F127.0.0.1:8888\u003C\u002Fcode>，而且可以切到 Chat 分頁。\u003C\u002Fp>\u003Ch2>Step 3: 下載 Gemma 4 模型檔\u003C\u002Fh2>\u003Cp>這一步的目的，是把符合你硬體的量化模型抓到本機。若使用 Unsloth Studio，就在模型瀏覽器搜尋 Gemma 4 並下載對應量化；若走手動流程，就從 Hugging Face 選 GGUF 或 MLX 版本。\u003C\u002Fp>\u003Cp>初次上手時，E2B 與 E4B 可優先選 8-bit；12B、26B-A4B 與 31B 則建議先看 Dynamic 4-bit 版本，通常更容易塞進本機記憶體。\u003C\u002Fp>\u003Cp>你應該看到完整的模型檔、分片清單，或下載完成的狀態，並保留足夠記憶體給執行時額外開銷。\u003C\u002Fp>\u003Ch2>Step 4: 啟動 Gemma 4 聊天服務\u003C\u002Fh2>\u003Cp>這一步的目的，是把模型變成可互動的本機聊天服務。Gemma 4 使用標準的 system、user、assistant 角色，並可透過 chat template 參數控制是否啟用思考模式。\u003C\u002Fp>\u003Cp>若你用 llama.cpp，建議直接用 \u003Ccode>llama-server\u003C\u002Fcode> 來啟動，並在需要時關掉思考輸出，讓多輪對話只保留最終答案。\u003C\u002Fp>\u003Cpre>\u003Ccode>llama-server -m model.gguf --chat-template-kwargs '{\"thinking\":false}'\u003C\u002Fcode>\u003C\u002Fpre>\u003Cp>你應該看到伺服器啟動訊息，並能在本機端點送出第一個聊天請求。\u003C\u002Fp>\u003Ch2>Step 5: 驗證多模態與記憶體設定\u003C\u002Fh2>\u003Cp>這一步的目的，是確認模型真的能符合你的使用情境。若你要處理圖片或音訊，請先確認所選型號支援該模態，再\u003Ca href=\"\u002Fnews\u002Fllama-benchy-api-benchmark-zh\">測試\u003C\u002Fa>一張小圖或一段短音檔。\u003C\u002Fp>\u003Cp>同時檢查記憶體餘量與上下文長度，避免在長對話時因為超出限制而中斷。若你只需要文字推理，就先把多模態關閉，換取更穩定的本機執行。\u003C\u002Fp>\u003Cp>你應該看到一則成功回覆，且系統監控工具顯示記憶體沒有持續逼近上限。\u003C\u002Fp>\u003Ch2>常見錯誤\u003C\u002Fh2>\u003Cul>\u003Cli>模型下載失敗：先確認 Hugging Face 登入狀態，再重試下載，必要時改用 CLI 下載分片檔。\u003C\u002Fli>\u003Cli>記憶體不足：改用更小的型號或更低位元量化，例如從 12B 換到 E4B，或從 8-bit 換到 4-bit。\u003C\u002Fli>\u003Cli>聊天輸出出現思考內容：在 llama.cpp 端重新檢查 chat-template 參數，並確認你啟動的是支援關閉思考的流程。\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>接下來可以看什麼\u003C\u002Fh2>\u003Cp>如果你已經能在本機穩定跑起來，下一步可以接著看 Gemma 4 的量化調參、\u003Ca href=\"\u002Ftag\u002F長上下文\">長上下文\u003C\u002Fa>設定，以及把本機模型包成 \u003Ca href=\"\u002Ftag\u002Fopenai\">OpenAI\u003C\u002Fa> 相容 \u003Ca href=\"\u002Ftag\u002Fapi\">API\u003C\u002Fa> 的做法。\u003C\u002Fp>","用 Unsloth Studio 或 llama.cpp 在本機下載、啟動並聊天 Gemma 4。","unsloth.ai","https:\u002F\u002Funsloth.ai\u002Fdocs\u002Fmodels\u002Fgemma-4",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1780777065583-koml.png","industry","zh","b7998c1b-8e10-4f65-aef3-59a428f36541",[17,18,19,20,21,22],"Gemma 4","Unsloth Studio","llama.cpp","Hugging Face","量化模型","本機推理",[24,25,26],"先選對 Gemma 4 型號，再決定量化與記憶體預算。","Unsloth Studio 適合瀏覽器式操作，llama.cpp 適合直接本機推理。","下載完成後要驗證聊天端點、思考模式與記憶體餘量。",2,"2026-06-06T20:17:21.259919+00:00","2026-06-06T20:17:21.24+00:00","fe20f6f6-432b-47bf-a410-a5f516d885ed",{"tags":32,"relatedLang":11,"relatedPosts":42},[33,35,36,38,40],{"name":17,"slug":34},"gemma-4",{"name":21,"slug":21},{"name":20,"slug":37},"hugging-face",{"name":18,"slug":39},"unsloth-studio",{"name":19,"slug":41},"llamacpp",[43,49,55,61,67,73],{"id":44,"slug":45,"title":46,"cover_image":47,"image_url":47,"created_at":48,"category":13},"3d7ff80a-4045-4b66-9e21-b6a8eb3b6f6d","openai-europe-privacy-policy-zh","OpenAI 歐洲隱私政策更新重點","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781052479369-yomr.png","2026-06-10T00:47:31.176745+00:00",{"id":50,"slug":51,"title":52,"cover_image":53,"image_url":53,"created_at":54,"category":13},"69002c63-177a-4723-9e63-d28506f08edd","openai-ads-sensitive-chats-policy-zh","OpenAI把廣告擋在敏感對話外是對的","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781051578409-en02.png","2026-06-10T00:32:23.404084+00:00",{"id":56,"slug":57,"title":58,"cover_image":59,"image_url":59,"created_at":60,"category":13},"ea98a8c9-ebe1-4258-8a2b-b0d82b25deed","ai-bootlegs-streaming-royalties-stick-figure-zh","AI bootlegs 正在抽走串流版稅","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781050681742-3rdh.png","2026-06-10T00:17:31.017287+00:00",{"id":62,"slug":63,"title":64,"cover_image":65,"image_url":65,"created_at":66,"category":13},"20d0b5fc-a363-481d-86b2-e30276a49e92","amd-microsoft-windows-ml-acceleration-zh","AMD 與 Microsoft 把 Windows ML 推進 GPU 與 N…","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781047980407-vd5p.png","2026-06-09T23:32:31.304436+00:00",{"id":68,"slug":69,"title":70,"cover_image":71,"image_url":71,"created_at":72,"category":13},"9a0692ba-a9c5-42eb-823d-8a0e6e6ae3fc","openai-ipo-filing-turns-hype-into-scrutiny-zh","OpenAI IPO 讓神話變審核","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781042614962-bj12.png","2026-06-09T22:03:04.524304+00:00",{"id":74,"slug":75,"title":76,"cover_image":77,"image_url":77,"created_at":78,"category":13},"40d4f012-36b6-4b8f-b470-30242a0b8483","skatteetaten-public-sector-ai-should-be-judged-by-outcomes-zh","Skatteetaten 證明公部門 AI 應該看成果，不是看噱頭","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781038986405-p8cf.png","2026-06-09T21:02:32.1198+00:00",[80,85,90,95,100,105,110,115,120,125],{"id":81,"slug":82,"title":83,"created_at":84},"ee073da7-28b3-4752-a319-5a501459fb87","ai-in-2026-what-actually-matters-now-zh","2026 AI 真正重要的事","2026-03-26T07:09:12.008134+00:00",{"id":86,"slug":87,"title":88,"created_at":89},"83bd1795-8548-44c9-9a7e-de50a0923f71","trump-ai-framework-power-speech-state-preemption-zh","川普 AI 框架瞄準電力、言論與州權","2026-03-26T07:12:18.695466+00:00",{"id":91,"slug":92,"title":93,"created_at":94},"ea6be18b-c903-4e54-97b7-5f7447a612e0","nvidia-gtc-2026-big-ai-announcements-zh","NVIDIA GTC 2026 重點拆解","2026-03-26T07:14:26.62638+00:00",{"id":96,"slug":97,"title":98,"created_at":99},"4bcec76f-4c36-4daa-909f-54cd702f7c93","claude-users-spreading-out-and-getting-better-zh","Claude 用戶更分散，也更會用","2026-03-26T07:22:52.325888+00:00",{"id":101,"slug":102,"title":103,"created_at":104},"bd903b15-2473-4178-9789-b7557816e535","openclaw-raises-hard-question-for-ai-models-zh","OpenClaw 逼問 AI 模型價值","2026-03-26T07:24:54.707486+00:00",{"id":106,"slug":107,"title":108,"created_at":109},"eeac6b9e-ad9d-4831-8eec-8bba3f9bca6a","gap-google-gemini-checkout-fashion-search-zh","Gap 把結帳搬進 Gemini","2026-03-26T07:28:23.937768+00:00",{"id":111,"slug":112,"title":113,"created_at":114},"0740e53f-605d-4d57-8601-c10beb126f3c","google-pushes-gemini-transition-to-march-2026-zh","Google 把 Gemini 轉換延到 2026 年 3…","2026-03-26T07:30:12.825269+00:00",{"id":116,"slug":117,"title":118,"created_at":119},"e660d801-2421-4529-8fa9-86b82b066990","metas-llama-4-benchmark-scandal-gets-worse-zh","Meta Llama 4 分數風波又擴大","2026-03-26T07:34:21.156421+00:00",{"id":121,"slug":122,"title":123,"created_at":124},"183f9e7c-e143-40bb-a6d5-67ba84a3a8bc","accenture-mistral-ai-sovereign-enterprise-deal-zh","Accenture 攜手 Mistral AI 賣主權 AI","2026-03-26T07:38:14.818906+00:00",{"id":126,"slug":127,"title":128,"created_at":129},"191d9b1b-768a-478c-978c-dd7431a38149","mistral-ai-faces-its-hardest-year-yet-zh","Mistral AI 迎來最硬的一年","2026-03-26T07:40:23.716374+00:00"]