[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-gemma-4-lands-on-google-cloud-zh":3,"tags-gemma-4-lands-on-google-cloud-zh":35,"related-lang-gemma-4-lands-on-google-cloud-zh":52,"related-posts-gemma-4-lands-on-google-cloud-zh":56,"series-model-release-b8f87962-35c1-4507-a957-2904710abe69":93},{"id":4,"title":5,"content":6,"summary":7,"source":8,"source_url":9,"author":10,"image_url":11,"keywords":12,"language":23,"translated_content":10,"views":24,"is_premium":25,"created_at":26,"updated_at":26,"cover_image":11,"published_at":27,"rewrite_status":28,"rewrite_error":10,"rewritten_from_id":29,"slug":30,"category":31,"related_article_id":32,"status":33,"google_indexed_at":34,"x_posted_at":10,"tweet_text":10,"title_rewritten_at":10,"title_original":10,"key_takeaways":10,"topic_cluster_id":10,"embedding":10,"is_canonical_seed":25},"b8f87962-35c1-4507-a957-2904710abe69","Gemma 4 登上 Google Cloud","\u003Cp>Google Cloud 這次把 \u003Ca href=\"https:\u002F\u002Fcloud.google.com\u002Fvertex-ai\" target=\"_blank\" rel=\"noopener\">Gemma 4\u003C\u002Fa> 接進來，數字很直接：上下文最高 256K toke\u003Ca href=\"\u002Fnews\u002Funsloth-qwen35-partial-fine-tuning-zh\">ns\u003C\u002Fa>。這代表它不只會看圖、聽音檔，還能讀很長的資料。對開發者來說，這種規格不是拿來炫技，是拿來省時間。\u003C\u002Fp>\u003Cp>更有意思的是，Google 沒有只丟一個模型檔案就收工。它把 \u003Ca href=\"https:\u002F\u002Fcloud.google.com\u002Fvertex-ai\u002Fdocs\u002Fgenerative-ai\u002Fmodel-garden\" target=\"_blank\" rel=\"noopener\">Model Garden\u003C\u002Fa>、\u003Ca href=\"https:\u002F\u002Fcloud.google.com\u002Frun\" target=\"_blank\" rel=\"noopener\">Cloud Run\u003C\u002Fa>、\u003Ca href=\"https:\u002F\u002Fcloud.google.com\u002Fkubernetes-engine\" target=\"_blank\" rel=\"noopener\">Google Kubernetes Engine\u003C\u002Fa>，還有 \u003Ca href=\"https:\u002F\u002Fcloud.google.com\u002Ftpu\" target=\"_blank\" rel=\"noopener\">Google Cloud TPUs\u003C\u002Fa> 一起包進來。講白了，就是你可以選 managed、serverless，或自己控到底。\u003C\u002Fp>\u003Ch2>Google 這次到底丟了什麼\u003C\u002Fh2>\u003Cp>先講重點。Gemma 4 是 Google 目前最完整的開放模型家族之一。它不是單一 checkpoint，而是一整套版本。從較小的型號，到 31B dense model，再到 \u003Ca href=\"\u002Fnews\u002Fwebassembly-2026-faster-web-apps-less-javascript-zh\">26\u003C\u002Fa>B mixture-of-experts，選項很齊。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1775239437839-x6ov.png\" alt=\"Gemma 4 登上 Google Cloud\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>這種設計很務實。因為不是每個團隊都需要大模型硬扛。很多產品只需要低延遲、便宜、可控。你如果拿 31B 去做簡單客服，八成是在燒錢。\u003C\u002Fp>\u003Cp>Gemma 4 的賣點很直白。長上下文、多模態、還支援 140 多種語言。再加上 \u003Ca href=\"https:\u002F\u002Fwww.apache.org\u002Flicenses\u002FLICENSE-2.0\" target=\"_blank\" rel=\"noopener\">Apache 2.0\u003C\u002Fa> 授權，商用門檻比很多閉源模型低很多。這點對台灣團隊很實際，因為法務溝通常常比模型測試還慢。\u003C\u002Fp>\u003Cul>\u003Cli>最高 256K token context\u003C\u002Fli>\u003Cli>原生 vision 與 audio\u003C\u002Fli>\u003Cli>支援 140+ 語言\u003C\u002Fli>\u003Cli>Apache 2.0 授權\u003C\u002Fli>\u003Cli>2B 到 31B，多一個 26B MoE 版本\u003C\u002Fli>\u003C\u002Ful>\u003Cp>我覺得這次最聰明的地方，是 Google 把模型和雲端路徑綁在一起。你不是只拿到權重。你是拿到一條能上線的路。\u003C\u002Fp>\u003Cp>Google 也把 Gemma 4 跟自家研究線連在一起，說它和 Gemini 3 用的是同一脈研究成果。這不代表兩者一樣強，但很明顯，Google 想把開放模型做得更像它的旗艦產品。\u003C\u002Fp>\u003Ch2>為什麼企業團隊會在意\u003C\u002Fh2>\u003Cp>企業最怕的不是模型不夠強。最怕的是資料出不去，合規過不了，採購卡住。Gemma 4 的部署方式，剛好就是在處理這些麻煩事。\u003C\u002Fp>\u003Cp>Google Cloud 把它放進不同環境。你可以跑在公有雲，也可以用資料邊界控制。你甚至可以看 \u003Ca href=\"https:\u002F\u002Fcloud.google.com\u002Fsolutions\u002Fsovereign-cloud\" target=\"_blank\" rel=\"noopener\">Sovereign Cloud\u003C\u002Fa>、\u003Ca href=\"https:\u002F\u002Fcloud.google.com\u002Fblog\u002Fproducts\u002Finfrastructure\u002Fintroducing-s3ns\" target=\"_blank\" rel=\"noopener\">S3NS in France\u003C\u002Fa>，或 \u003Ca href=\"https:\u002F\u002Fcloud.google.com\u002Fdistributed-cloud\" target=\"_blank\" rel=\"noopener\">Google Distributed Cloud\u003C\u002Fa> 這類路線。這對金融、醫療、政府案子都很重要。\u003C\u002Fp>\u003Cp>講白了，很多 AI 專案不是死在技術，而是死在流程。POC 做完很漂亮，到了資安審查就卡住。Gemma 4 至少把這些地雷先拆掉一半。\u003C\u002Fp>\u003Cblockquote>“I think the biggest thing is we’re seeing companies realize that AI is not a science project anymore.” — Thomas Kurian, Google Cloud Next 2024 keynote\u003C\u002Fblockquote>\u003Cp>Thomas Kurian 這句話很準。現在企業要的不是 demo，是能交付的軟體。Google 這次的打法也很清楚，就是把模型當基礎設施賣，不是當玩具展示。\u003C\u002Fp>\u003Cp>還有一個細節很值得看。Google 說 26B MoE 版本很快會在 Model Garden 變成 fully managed 和 serverless。這很像在對開發者說：你可以保留 open model 的自由，但不用自己扛全部 MLOps。\u003C\u002Fp>\u003Ch2>部署選項才是主菜\u003C\u002Fh2>\u003Cp>如果你是工程師，你會發現這次最有價值的不是 benchmark，而是部署選擇。\u003Ca href=\"https:\u002F\u002Fcloud.google.com\u002Fvertex-ai\u002Fdocs\u002Fgenerative-ai\u002Fmodel-garden\u002Fuse-gemma\" target=\"_blank\" rel=\"noopener\">Vertex AI\u003C\u002Fa> 適合想快速上線的人。\u003Ca href=\"https:\u002F\u002Fcloud.google.com\u002Frun\" target=\"_blank\" rel=\"noopener\">Cloud Run\u003C\u002Fa> 適合想省維運的人。\u003Ca href=\"https:\u002F\u002Fcloud.google.com\u002Fkubernetes-engine\u002Fdocs\u002Fconcepts\u002Fabout-gke\" target=\"_blank\" rel=\"noopener\">GKE\u003C\u002Fa> 則適合想自己控資源的人。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1775239441824-nv27.png\" alt=\"Gemma 4 登上 Google Cloud\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>Google 也很明顯在推 agent 工作流。Gemma 4 支援 reasoning、function calling、code gener\u003Ca href=\"\u002Fnews\u002Fchatgpt-ads-format-standardization-data-zh\">at\u003C\u002Fa>ion、structured output，還能搭配 \u003Ca href=\"https:\u002F\u002Fgoogle.github.io\u002Fadk-docs\u002F\" target=\"_blank\" rel=\"noopener\">Agent Development Kit\u003C\u002Fa> 做 AI agents。這不是空話，因為現在很多團隊真的在做工具型 agent，不是在做聊天機器人。\u003C\u002Fp>\u003Cp>如果你想跑 inference，Cloud Run 也能接 \u003Ca href=\"https:\u002F\u002Fcloud.google.com\u002Frun\u002Fdocs\u002Fconfiguring\u002Fadvanced\u002Fgpus\" target=\"_blank\" rel=\"noopener\">NVIDIA RTX PRO 6000 Blackwell GPUs\u003C\u002Fa>，而且有 96GB vGPU memory。這種配置對中型模型很實用，尤其是流量忽高忽低的產品。\u003C\u002Fp>\u003Cul>\u003Cli>Cloud Run 可 scale to zero\u003C\u002Fli>\u003Cli>GKE 可搭配 vLLM\u003C\u002Fli>\u003Cli>TPU 可用於 serving、pretraining、post-training\u003C\u002Fli>\u003Cli>vLLM 與 \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FNeMo\" target=\"_blank\" rel=\"noopener\">NVIDIA NeMo\u003C\u002Fa> 都在建議工具鏈內\u003C\u002Fli>\u003Cli>GKE Agent Sandbox 可隔離執行 LLM 產生的 code\u003C\u002Fli>\u003C\u002Ful>\u003Cp>我覺得 GKE Agent Sandbox 這點很有戲。Google 說它能做到 sub-second cold starts，還能到每秒 300 個 sandboxes。這種數字如果在真實環境站得住，對多步驟 agent 會很有幫助。\u003C\u002Fp>\u003Cp>因為 agent 最怕什麼？不是模型不會想，是工具呼叫亂掉。Sandbox 做得好，至少能把風險隔離開來。\u003C\u002Fp>\u003Ch2>跟其他開放模型比，差在哪\u003C\u002Fh2>\u003Cp>Gemma 4 的對手很多。\u003Ca href=\"https:\u002F\u002Fai.meta.com\u002Fllama\u002F\" target=\"_blank\" rel=\"noopener\">Llama\u003C\u002Fa>、\u003Ca href=\"https:\u002F\u002Fmistral.ai\" target=\"_blank\" rel=\"noopener\">Mistral\u003C\u002Fa>、還有一堆自家微調版本，都在搶同一批開發者。差別不在「誰最會講」，而在「誰比較好上線」。\u003C\u002Fp>\u003Cp>Llama 的生態很大。Mistral 的效率也不差。但 Google 這次把模型、雲端、TPU、Sovereign Cloud 一起打包，這就很像把整套廚房都搬給你。你不用自己找瓦斯、找鍋子、找水電。\u003C\u002Fp>\u003Cp>Gemma 4 的優勢也很實際。256K context 很適合長文件、codebase、法規資料。vision 和 audio 則讓它比純文字模型更好用。對做客服、文件分析、會議摘要、內部知識庫的人來說，這些都不是加分題，是基本盤。\u003C\u002Fp>\u003Cul>\u003Cli>\u003Cstrong>Gemma 4\u003C\u002Fstrong>：256K context、vision、audio、140+ languages、Apache 2.0\u003C\u002Fli>\u003Cli>\u003Cstrong>Llama\u003C\u002Fstrong>：生態大，但部署與合規常要自己處理\u003C\u002Fli>\u003Cli>\u003Cstrong>Mistral\u003C\u002Fstrong>：效率不錯，適合自架，但雲端整合較分散\u003C\u002Fli>\u003Cli>\u003Cstrong>Gemma 4 on Google Cloud\u003C\u002Fstrong>：managed、serverless、TPU、GKE 一次到位\u003C\u002Fli>\u003C\u002Ful>\u003Cp>成本也是重點。Cloud Run 的 scale to zero 很適合有尖峰流量的服務。GKE 可以把 autoscaling 玩得很細。TPU 則適合大規模推理或訓練。這些選項放一起，對財務部門會比較好交代。\u003C\u002Fp>\u003Cp>如果你是已經在 Google Cloud 上的團隊，這次幾乎可以直接試。因為整合成本低。反過來說，如果你不在 Google 生態裡，Gemma 4 也還是值得看，因為它把 open model 的商用路徑講得更完整了。\u003C\u002Fp>\u003Ch2>這件事放回產業脈絡看\u003C\u002Fh2>\u003Cp>現在的 open model 市場，已經不是單純比誰權重開得多。大家開始比的是：誰能讓你少養幾個平台工程師。這才是現實。\u003C\u002Fp>\u003Cp>過去很多團隊下載模型後，還要自己處理 serving、監控、快取、版本控管、權限、資料隔離。這一套做下來，很容易變成半個平台專案。Google 這次就是想把這些東西收進自己的雲服務裡。\u003C\u002Fp>\u003Cp>這也解釋了為什麼 Apache 2.0 這麼重要。因為企業不是只看準不準，還看能不能放心用。授權不清楚的模型，再強也會被法務擋下來。\u003C\u002Fp>\u003Cp>另外一個背景是，多模態 AI 已經變成主流需求。很多內部系統不只吃文字，還要吃圖、吃語音、吃 PDF、吃截圖。Gemma 4 的定位，剛好踩在這個需求上。\u003C\u002Fp>\u003Cp>我自己的判斷是，接下來 6 到 12 個月，會有更多雲端業者把 open model 和 managed serving 綁在一起賣。因為大家都知道，模型本身只是入口，真正能留住客戶的是部署體驗。\u003C\u002Fp>\u003Ch2>接下來該怎麼看\u003C\u002Fh2>\u003Cp>如果你正在做內部助理、文件解析、code tool，或多模態 agent，Gemma 4 很值得排進測試清單。先從最小型號開始，別一上來就衝大模型。很多場景根本不需要那麼重。\u003C\u002Fp>\u003Cp>我會特別注意 26B MoE 在 \u003Ca href=\"https:\u002F\u002Fcloud.google.com\u002Fvertex-ai\u002Fdocs\u002Fgenerative-ai\u002Fmodel-garden\" target=\"_blank\" rel=\"noopener\">Model Garden\u003C\u002Fa> 上線後的價格、延遲、吞吐量。這三個數字會決定它是不是能成為很多 Google Cloud 客戶的預設選項。\u003C\u002Fp>\u003Cp>問題很簡單：你要的是一個很會聊天的模型，還是一條能真的上線的路？如果你要後者，Gemma 4 這次給的東西，確實值得動手試。\u003C\u002Fp>","Gemma 4 進入 Google Cloud，支援 256K context、vision、audio 與 Apache 2.0 授權，還能跑在 Vertex AI、Cloud Run、GKE 與 TPU 上。","cloud.google.com","https:\u002F\u002Fcloud.google.com\u002Fblog\u002Fproducts\u002Fai-machine-learning\u002Fgemma-4-available-on-google-cloud",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1775239437839-x6ov.png",[13,14,15,16,17,18,19,20,21,22],"Gemma 4","Google Cloud","Vertex AI","Cloud Run","GKE","TPU","LLM","open model","Apache 2.0","multimodal AI","zh",0,false,"2026-04-03T18:03:40.435555+00:00","2026-04-03T18:03:40.238+00:00","done","080c8e6a-93f6-417f-b041-51dafb538749","gemma-4-lands-on-google-cloud-zh","model-release","94f75563-cdbc-47f2-83c1-0589da2710e1","published","2026-04-07T07:41:09.058+00:00",[36,38,40,42,44,46,48,50],{"name":13,"slug":37},"gemma-4",{"name":22,"slug":39},"multimodal-ai",{"name":18,"slug":41},"tpu",{"name":21,"slug":43},"apache-20",{"name":19,"slug":45},"llm",{"name":16,"slug":47},"cloud-run",{"name":14,"slug":49},"google-cloud",{"name":20,"slug":51},"open-model",{"id":32,"slug":53,"title":54,"language":55},"gemma-4-lands-on-google-cloud-en","Gemma 4 lands on Google Cloud","en",[57,63,69,75,81,87],{"id":58,"slug":59,"title":60,"cover_image":61,"image_url":61,"created_at":62,"category":31},"5b5fa24f-5259-4e9e-8270-b08b6805f281","minimax-m1-open-hybrid-attention-reasoning-model-zh","MiniMax-M1：開源 1M Token 推理模型","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778797859209-ea1g.png","2026-05-14T22:30:38.636592+00:00",{"id":64,"slug":65,"title":66,"cover_image":67,"image_url":67,"created_at":68,"category":31},"b1da56ac-8019-4c6b-a8dc-22e6e22b1cb5","gemini-omni-video-review-text-rendering-zh","Gemini Omni 影片模型怎麼了","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778779280109-lrrk.png","2026-05-14T17:20:42.608312+00:00",{"id":70,"slug":71,"title":72,"cover_image":73,"image_url":73,"created_at":74,"category":31},"d63e9d93-e613-4bbf-8135-9599fde11d08","why-xiaomi-mimo-v25-pro-changes-coding-agents-zh","為什麼 Xiaomi 的 MiMo-V2.5-Pro 改變的是 Coding …","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778689858139-v38e.png","2026-05-13T16:30:27.893951+00:00",{"id":76,"slug":77,"title":78,"cover_image":79,"image_url":79,"created_at":80,"category":31},"8f0c9185-52f9-46f2-82c6-5baec126ba2e","openai-realtime-audio-models-live-voice-zh","OpenAI 即時音訊模型瞄準語音互動","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778451657895-2iu7.png","2026-05-10T22:20:32.443798+00:00",{"id":82,"slug":83,"title":84,"cover_image":85,"image_url":85,"created_at":86,"category":31},"52106dc2-4eba-4ca0-8318-fa646064de97","anthropic-10-finance-ai-agents-zh","Anthropic推10款金融AI Agent","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778389843399-vclb.png","2026-05-10T05:10:22.778762+00:00",{"id":88,"slug":89,"title":90,"cover_image":91,"image_url":91,"created_at":92,"category":31},"6ee6ed2a-35c6-4be3-ba2c-43847e592179","why-claudes-infinite-context-window-wont-autonomous-zh","為什麼 Claude 的「無限」上下文窗口，仍然不會讓 AI 自主運作","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778350250836-d5d5.png","2026-05-09T18:10:27.004984+00:00",[94,99,104,109,114,119,124,129,134,139],{"id":95,"slug":96,"title":97,"created_at":98},"58b64033-7eb6-49b9-9aab-01cf8ae1b2f2","nvidia-rubin-six-chips-one-ai-supercomputer-zh","NVIDIA Rubin 把六顆晶片塞進 AI 機櫃","2026-03-26T07:18:45.861277+00:00",{"id":100,"slug":101,"title":102,"created_at":103},"0dcc2c61-c2a6-480d-adb8-dd225fc68914","march-2026-ai-model-news-what-mattered-zh","2026 年 3 月 AI 模型新聞重點","2026-03-26T07:32:08.386348+00:00",{"id":105,"slug":106,"title":107,"created_at":108},"214ab08b-5ce5-4b5c-8b72-47619d8675dd","why-small-models-are-winning-on-device-ai-zh","小模型為何吃下裝置端 AI","2026-03-26T07:36:30.488966+00:00",{"id":110,"slug":111,"title":112,"created_at":113},"785624b2-0355-4b82-adc3-de5e45eecd88","midjourney-v8-faster-images-higher-costs-zh","Midjourney V8 變快了，也變貴了","2026-03-26T07:52:03.562971+00:00",{"id":115,"slug":116,"title":117,"created_at":118},"cda76b92-d209-4134-86c1-a60f5bc7b128","xiaomi-mimo-trio-agents-robots-voice-zh","小米 MiMo 三模型瞄準代理、機器人與語音","2026-03-28T03:05:08.779489+00:00",{"id":120,"slug":121,"title":122,"created_at":123},"9e1044b4-946d-47fe-9e2a-c2ee032e1164","xiaomi-mimo-v2-pro-1t-moe-agents-zh","小米 MiMo-V2-Pro 登場：1T MoE 模型","2026-03-28T03:06:19.002353+00:00",{"id":125,"slug":126,"title":127,"created_at":128},"d68e59a2-55eb-4a8f-95d6-edc8fcbff581","cursor-composer-2-started-from-kimi-zh","Cursor Composer 2 其實從 Kimi 起步","2026-03-28T03:11:58.893796+00:00",{"id":130,"slug":131,"title":132,"created_at":133},"c4b6186f-bd84-4598-997e-c6e31d543c0d","cursor-composer-2-agentic-coding-model-zh","Cursor Composer 2 走向代理式寫碼","2026-03-28T03:13:06.422716+00:00",{"id":135,"slug":136,"title":137,"created_at":138},"45812c46-99fc-4b1f-aae1-56f64f5c9024","openai-shuts-down-sora-video-app-api-zh","OpenAI 關閉 Sora App 與 API","2026-03-29T04:47:48.974108+00:00",{"id":140,"slug":141,"title":142,"created_at":143},"e112e76f-ec3b-408f-810e-e93ae21a888a","apple-siri-gemini-distilled-models-zh","Apple Siri 牽手 Gemini 的真相","2026-03-29T04:52:57.886544+00:00"]