[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-kimi-k25-review-stronger-still-not-legend-zh":3,"article-related-kimi-k25-review-stronger-still-not-legend-zh":27,"series-ai-agent-2987d097-563f-46c7-b76f-b558d8ef7c2b":86},{"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":11,"views":24,"created_at":25,"published_at":26,"topic_cluster_id":11},"2987d097-563f-46c7-b76f-b558d8ef7c2b","kimi-k25-review-stronger-still-not-legend-zh","Kimi K2.5 評測：更強，但還不是神作","\u003Cp>Kimi K2.5 這次真的不是小修小補。它主打 1 兆參數級模型，還能看圖片、影片、長文件，甚至接代理工作流。說真的，這組合很猛，但實測後也很現實：長文理解進步明顯，跑任務卻常常拖時間。\u003C\u002Fp>\u003Cp>更有意思的是，它現在不是單純聊天機器人了。\u003Ca href=\"https:\u002F\u002Fkimi.moonshot.cn\u002F\" target=\"_blank\" rel=\"noopener\">Kimi\u003C\u002Fa> 把模型、代理層、寫碼工具包在一起賣，部分程式碼也放在 \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fmoonshot-ai\" target=\"_blank\" rel=\"noopener\">Moonshot AI 的 GitHub\u003C\u002Fa>。功能變多了，但免費與付費的界線也更硬了。\u003C\u002Fp>\u003Ch2>Kimi K2.5 到底強在哪\u003C\u002Fh2>\u003Cp>先講結論。Kimi K2.5 最有感的進步，是它更會吃髒資料。長文、圖片、影片這種混在一起的輸入，它比前代更穩。這對做研究、讀報告、整理會議資料的人很實用。\u003C\u002Fp>\u003Cp>在 40 萬字元級別的小說測試裡，它能抓到更細的角色關係。像是誰跟誰有衝突，哪條支線埋了伏筆，它都比一些只會表面摘要的模型更會看。這種能力，對台灣開發者很重要，因為很多工作不是寫一句話，而是要讀一整包資料。\u003C\u002Fp>\u003Cp>它的多模態能力也有進步。像 PDF 指令、30 秒網球影片，它都能拆解得更細。影片上限大約 100MB，這代表它適合短片分析，不適合拿來處理長片素材。講白了，它很會吃小樣本，但還沒到全能。\u003C\u002Fp>\u003Cul>\u003Cli>長文測試：約 40 萬字元小說\u003C\u002Fli>\u003Cli>影片支援：可處理 Apple MOV\u003C\u002Fli>\u003Cli>單段影片上限：約 100MB\u003C\u002Fli>\u003Cli>常見可用長度：約 3 分鐘\u003C\u002Fli>\u003Cli>角色與劇情整理：比 Qwen3-Max 更細\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>代理工具有料，但還不夠穩\u003C\u002Fh2>\u003Cp>現在 AI 產品都愛講 agent。Kimi 也不例外。它有單代理模式，還有 \u003Ca href=\"https:\u002F\u002Fkimi.moonshot.cn\u002F\" target=\"_blank\" rel=\"noopener\">Agent Swarm\u003C\u002Fa>，可以讓多個代理一起做事。再加上 \u003Ca href=\"https:\u002F\u002Fkimi.moonshot.cn\u002Fcode\" target=\"_blank\" rel=\"noopener\">Kimi Code\u003C\u002Fa>，它能接本機開發環境，讀專案檔案，改程式，跑測試，甚至執行命令。\u003C\u002Fp>\u003Cp>聽起來很像把助理、工程師、測試員塞進同一個產品。問題是，跑起來沒有那麼順。複雜任務有時要等 30 分鐘，途中還可能卡住。更麻煩的是，它碰到矛盾指令時，常常不先問你，而是直接照著其中一條路走下去。\u003C\u002Fp>\u003Cp>這種行為適合 demo，不太適合正式流程。你如果拿它做報表、寫 code、處理客戶資料，半小時後才發現方向錯了，真的會想翻白眼。它像一個很會做事的菜鳥，不像你能完全放手的同事。\u003C\u002Fp>\u003Cblockquote>“If you want to build a ship, don’t drum up people to collect wood and don’t assign them tasks and work, but rather teach them to long for the endless immensity of the sea.” — Antoine de Saint-Exupéry, \u003Cem>The Little Prince\u003C\u002Fem>\u003C\u002Fblockquote>\u003Cp>這句話放在 agent 時代很貼切。大家都想要 AI 自己拆任務、自己跑流程。可是真正難的地方，是它要知道什麼時候該停，什麼時候該問人。\u003C\u002Fp>\u003Cp>如果代理只會一路往前衝，那它不是自動化，是自動出包。Kimi K2.5 目前就在這條線上搖擺。能用，但還沒到可以放心交棒。\u003C\u002Fp>\u003Ch2>跟競品比，差距在哪\u003C\u002Fh2>\u003Cp>把 Kimi K2.5 放到同一桌比較，畫面就很清楚。它在長文理解上，這次比 \u003Ca href=\"https:\u002F\u002Fqwenlm.github.io\u002F\" target=\"_blank\" rel=\"noopener\">Qwen\u003C\u002Fa> 系列更穩。面對複雜文本，它比較不會漏掉關係鏈，也比較敢下結論。\u003C\u002Fp>\u003Cp>跟 \u003Ca href=\"https:\u002F\u002Fai.google.dev\u002Fgemini-api\" target=\"_blank\" rel=\"noopener\">Gemini\u003C\u002Fa> 比，視覺指令理解差距不大。可是到了設計輸出，Kimi 還是很容易掉進「像簡報、不像作品」的坑。版面能看，味道很淡。這種東西做內部報告可以，拿去當品牌主視覺就有點尷尬。\u003C\u002Fp>\u003Cp>另一個對照是 \u003Ca href=\"https:\u002F\u002Fopenai.com\u002F\" target=\"_blank\" rel=\"noopener\">OpenAI\u003C\u002Fa> 和 \u003Ca href=\"https:\u002F\u002Fwww.anthropic.com\u002F\" target=\"_blank\" rel=\"noopener\">Anthropic\u003C\u002Fa>。它們的產品線也在往 agent 和 coding 工具靠。差別是，Kimi 現在更像在追趕整套工作流，而不是只比聊天品質。這會讓它更像工具平台，也更考驗穩定度。\u003C\u002Fp>\u003Cul>\u003Cli>長文理解：這次 Kimi 領先 Qwen3-Max\u003C\u002Fli>\u003Cli>視覺指令：大致接近 Gemini\u003C\u002Fli>\u003Cli>設計輸出：完成度有，審美普通\u003C\u002Fli>\u003Cli>Hero 圖：太像 PPT\u003C\u002Fli>\u003Cli>回答意願：比 Qwen3-Max 更敢答\u003C\u002Fli>\u003C\u002Ful>\u003Cp>價格也很現實。Kimi 現在把訂閱推得更明顯，排隊優先、Kimi Code、Agent Swarm 都有付費門檻。最高方案傳出是每月 199 人民幣。這讓它直接站進同一個商業戰場，跟國際大廠搶重度用戶的錢包。\u003C\u002Fp>\u003Cp>對台灣開發者來說，重點不是它會不會聊天，而是它能不能穩定嵌進工作流。只要它還會卡、會誤解、會亂跑，很多團隊就只會把它放在測試環境，不會直接上線。\u003C\u002Fp>\u003Ch2>這次更新透露了什麼產業訊號\u003C\u002Fh2>\u003Cp>Kimi K2.5 其實很像現在 AI 產品的縮影。模型本身越來越會讀長文、看圖片、看影片。產品層則往 agent、寫碼、訂閱制走。說白了，大家都在想辦法把「會聊」變成「會做事」。\u003C\u002Fp>\u003Cp>這也反映一個更大的分岔。開源模型在社群裡越來越有存在感，整合速度也快。封閉式產品則還是握著最完整的體驗和最容易收費的功能。Kimi 想兩邊都吃，策略不差，但執行很容易卡在體驗不一致。\u003C\u002Fp>\u003Cp>如果你是開發者，我會這樣看：Kimi K2.5 值得試，尤其是長文和多模態場景。可是你要先把它當工具，不要先把它當隊友。等它哪天能在開始動手前，先問出更好的問題，才比較像成熟的 agent。\u003C\u002Fp>\u003Cp>我自己的判斷很直接。接下來 3 到 6 個月，Kimi 能不能把 agent 的問答節奏修好，會比參數數字更重要。因為大家最後買單的，不是模型有多大，而是它會不會少搞砸一次。\u003C\u002Fp>\u003Ch2>結尾：它值得誰先試\u003C\u002Fh2>\u003Cp>如果你在做研究、讀長文件、跑多模態整理，Kimi K2.5 值得先放進測試清單。它在長文理解和視覺輸入上，確實有料。你會感覺它比很多只會講漂亮話的模型更能幹。\u003C\u002Fp>\u003Cp>但如果你的場景很吃穩定性，像是 production code、客服流程、正式報告，我會建議先小範圍試。先看它會不會亂跑，再決定要不要把它接進主流程。你如果問我一句話結論：Kimi K2.5 變強了，但還沒強到可以偷懶。\u003C\u002Fp>","Kimi K2.5 加入視覺、寫碼和多代理工具，長文理解更強，但長跑慢、設計感弱、付費牆明顯，還沒到頂級行列。","zhuanlan.zhihu.com","https:\u002F\u002Fzhuanlan.zhihu.com\u002Fp\u002F2001774908666042307",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Fcover-1774595754485-vd5ahn.png","ai-agent","zh","e64aba21-254b-4f93-aa21-837484bb52ec",[17,18,19,20,21,22,23],"Kimi K2.5","Moonshot AI","LLM","agent","多模態 AI","長文理解","AI 寫碼",8,"2026-03-27T07:15:55.277513+00:00","2026-03-27T13:49:01.738+00:00",{"tags":28,"relatedLang":45,"relatedPosts":49},[29,30,32,34,37,38,40,42],{"name":20,"slug":20},{"name":23,"slug":31},"ai-寫碼",{"name":21,"slug":33},"多模態-ai",{"name":35,"slug":36},"研究整理","-",{"name":22,"slug":22},{"name":18,"slug":39},"moonshot-ai",{"name":19,"slug":41},"llm",{"name":43,"slug":44},"多代理AI","-ai",{"id":15,"slug":46,"title":47,"language":48},"kimi-k25-review-stronger-still-not-legend-en","Kimi K2.5 review: stronger, still not a legend","en",[50,56,62,68,74,80],{"id":51,"slug":52,"title":53,"cover_image":54,"image_url":54,"created_at":55,"category":13},"ef96a410-24bd-4e35-8536-439f21f820e6","claude-code-dynamic-workflow-ai-harness-zh","Claude Code 動態工作流：AI 自寫 Harness","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781035378200-qkm9.png","2026-06-09T20:02:21.942031+00:00",{"id":57,"slug":58,"title":59,"cover_image":60,"image_url":60,"created_at":61,"category":13},"9fb91fbe-64cd-4732-aba7-5b20daacf962","agent-orchestration-enterprise-ai-layer-zh","企業 AI 缺的是編排層","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1780984981291-rodj.png","2026-06-09T06:02:30.929215+00:00",{"id":63,"slug":64,"title":65,"cover_image":66,"image_url":66,"created_at":67,"category":13},"2e389faa-a4ab-4f7a-b6da-c2ba69d5f14b","ai-agents-use-blockchain-trust-layer-zh","AI 代理用區塊鏈當信任層","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1780980509390-6s0i.png","2026-06-09T04:48:01.259033+00:00",{"id":69,"slug":70,"title":71,"cover_image":72,"image_url":72,"created_at":73,"category":13},"1c433948-634b-47e4-a119-dd567203a712","8-rag-patterns-demos-into-prod-zh","8 種 RAG 模式把 Demo 變上線","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1780971552397-h12o.png","2026-06-09T02:18:36.130013+00:00",{"id":75,"slug":76,"title":77,"cover_image":78,"image_url":78,"created_at":79,"category":13},"7d860405-aca6-486b-8de0-1c5193a3b06d","fine-tuning-beats-rag-style-not-facts-zh","當目標是文風不是事實時，微調比 RAG 更有效","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1780924689232-5elu.png","2026-06-08T13:17:25.235242+00:00",{"id":81,"slug":82,"title":83,"cover_image":84,"image_url":84,"created_at":85,"category":13},"3d1e5ef7-8f31-4e57-b286-306825d7f38e","openclaw-small-business-ai-staff-zh","OpenClaw把AI變成夜班員工","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1780904888882-6w0v.png","2026-06-08T07:47:27.229503+00:00",[87,92,97,98,103,108,113,118,123,128],{"id":88,"slug":89,"title":90,"created_at":91},"4ae1e197-1d3d-4233-8733-eafe9cb6438b","claude-now-uses-your-pc-to-finish-tasks-zh","Claude 開始幫你操作電腦","2026-03-26T07:20:48.457387+00:00",{"id":93,"slug":94,"title":95,"created_at":96},"5bede67f-e21c-413d-9ab8-54a3c3d26227","googles-2026-ai-agent-report-decoded-zh","Google 2026 AI Agent 報告解讀","2026-03-26T11:15:22.651956+00:00",{"id":4,"slug":5,"title":6,"created_at":25},{"id":99,"slug":100,"title":101,"created_at":102},"95c9053b-e3f4-4cb5-aace-5c54f4c9e044","claude-code-controls-mac-desktop-zh","Claude Code 也能操控 Mac 了","2026-03-28T03:01:58.58121+00:00",{"id":104,"slug":105,"title":106,"created_at":107},"dc58e153-e3a8-4c06-9b96-1aa64eabbf5f","cloudflare-100x-faster-ai-agent-sandbox-zh","Cloudflare 的 AI 沙箱跑超快","2026-03-28T03:09:44.142236+00:00",{"id":109,"slug":110,"title":111,"created_at":112},"1c8afc56-253f-47a2-979f-1065ff072f2a","openai-backs-isara-agent-swarm-bet-zh","OpenAI 挺 Isara 的 agent swarm …","2026-03-28T03:15:27.513155+00:00",{"id":114,"slug":115,"title":116,"created_at":117},"7379b422-576e-45df-ad5a-d57a0d9dd467","openai-plan-automated-ai-researcher-zh","OpenAI 想做自動化 AI 研究員","2026-03-28T03:17:42.090548+00:00",{"id":119,"slug":120,"title":121,"created_at":122},"48c9889e-86df-450b-a356-e4a4b7c83c5b","harness-engineering-ai-agent-reliability-2026-zh","駕馭工程：從「馬具」到「作業系統」，AI Agent 可靠性的終極密碼","2026-03-31T06:42:53.556721+00:00",{"id":124,"slug":125,"title":126,"created_at":127},"96d8e8c8-1edd-475d-9145-b1e7a1b02b65","mcp-explained-from-prompts-to-production-zh","MCP 怎麼把提示詞變工作流","2026-04-01T09:24:39.321274+00:00",{"id":129,"slug":130,"title":131,"created_at":132},"f2ca7720-b471-4ce5-9336-2a9ac2a876fd","amazon-bedrock-agents-multi-agent-workflows-zh","Amazon Bedrock Agents 進入多代理工作流","2026-04-01T09:30:29.945429+00:00"]