[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-5-shifts-in-llms-from-the-last-six-months-zh":3,"article-related-5-shifts-in-llms-from-the-last-six-months-zh":33,"series-industry-be0785a5-7976-4735-8f46-6abd84dac9af":80},{"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":25,"views":29,"created_at":30,"published_at":31,"topic_cluster_id":32},"be0785a5-7976-4735-8f46-6abd84dac9af","5-shifts-in-llms-from-the-last-six-months-zh","5 個 LLM 的半年轉變","\u003Cp data-speakable=\"summary\">這篇整理六個月內 LLM 的五個關鍵轉變，幫你判斷該用雲端前沿模型、開放模型，還是本地工作流。\u003C\u002Fp>\u003Cp>在短短 6 個月內，LLM 的使用方式明顯改變。讀完這 5 項，你可以更快決定：該把預算放在 coding agent、開放模型，還是本地部署。\u003C\u002Fp>\u003Ctable>\u003Cthead>\u003Ctr>\u003Cth>項目\u003C\u002Fth>\u003Cth>代表規格\u003C\u002Fth>\u003Cth>實際意義\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Ctd>Coding agents\u003C\u002Ftd>\u003Ctd>從「常常要修」變成「大多可直接用」\u003C\u002Ftd>\u003Ctd>適合日常開發、重構、測試\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>最佳模型更替\u003C\u002Ftd>\u003Ctd>數月內多次換位\u003C\u002Ftd>\u003Ctd>不能只看靜態排行榜\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>本地／開放模型\u003C\u002Ftd>\u003Ctd>20.9GB 到 1.5TB\u003C\u002Ftd>\u003Ctd>可在筆電到大型伺服器間選擇\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>個人 AI 助手\u003C\u002Ftd>\u003Ctd>Mac mini 常見\u003C\u002Ftd>\u003Ctd>適合長期、持續性任務\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>Benchmark demos\u003C\u002Ftd>\u003Ctd>像 pelican test 這類測試\u003C\u002Ftd>\u003Ctd>更能看出多模態與工具鏈能力\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003Ch2>1. Coding agents 夠日常使用了\u003C\u002Fh2>\u003Cp>最大的變化，不是某一個模型單獨升級，而是 \u003Ca href=\"\u002Ftag\u002Fagentic-coding\">agentic coding\u003C\u002Fa> 的整體品質跳了一級。像 \u003Ca href=\"https:\u002F\u002Fopenai.com\u002Fcodex\u002F\" rel=\"nofollow\">Codex\u003C\u002Fa> 和 \u003Ca href=\"https:\u002F\u002Fclaude.ai\u002Fcode\" rel=\"nofollow\">Claude Code\u003C\u002Fa> 這類工具，開始能在可驗證的回饋下持續優化，結果就是輸出不再像展示品，而是真的能幫忙做事。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1779167646556-uxgd.png\" alt=\"5 個 LLM 的半年轉變\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>這代表工作流也變了。以前你花大部分時間在修錯，現在可以把完整任務交出去，再回來檢查結果。\u003Ca href=\"\u002Fnews\u002F5-indiana-fever-updates-zh\">重點\u003C\u002Fa>不是完美，而是「大多可用」已經足夠進入日常開發。\u003C\u002Fp>\u003Cul>\u003Cli>之前：常常要手動修補\u003C\u002Fli>\u003Cli>之後：大多能直接接手\u003C\u002Fli>\u003Cli>適合：寫程式、重構、補測試、小型功能\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>2. 「最佳模型」換位變得很頻繁\u003C\u002Fh2>\u003Cp>另一個明顯現象，是領先位置在幾個大廠之間快速輪替。短短幾個月，冠軍從 \u003Ca href=\"\u002Ftag\u002Fclaude\">Claude\u003C\u002Fa> Sonnet 4.5、到 GPT-5.1、到 \u003Ca href=\"\u002Ftag\u002Fgemini\">Gemini\u003C\u002Fa> 3、再到 GPT-5.1 Codex Max，\u003Ca href=\"\u002Fnews\u002Fbree-hall-returns-indiana-fever-player-development-zh\">最後\u003C\u002Fa>又回到 \u003Ca href=\"\u002Ftag\u002Fanthropic\">Anthropic\u003C\u002Fa> 的 Claude Opus 4.5。\u003C\u002Fp>\u003Cp>這種變動說明競爭已經非常接近。對實際使用者來說，最好的做法不是相信固定排名，而是拿自己的任務去測。擅長寫 code 的模型，不一定最會做長流程規劃，也不一定最適合影像或工具調用。\u003C\u002Fp>\u003Cul>\u003Cli>Claude Sonnet 4.5：前段時間的領先者\u003C\u002Fli>\u003Cli>GPT-5.1 與 GPT-5.1 Codex Max：中段強勢競爭者\u003C\u002Fli>\u003Cli>Gemini 3：在特定測試上表現突出\u003C\u002Fli>\u003Cli>Claude Opus 4.5：後段重新奪回優勢\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>3. 開放模型和本地模型進步很大\u003C\u002Fh2>\u003Cp>開放模型這邊也跑得很快。像 \u003Ca href=\"https:\u002F\u002Fai.google.dev\u002Fgemma\" rel=\"nofollow\">Gemma 4\u003C\u002Fa>、\u003Ca href=\"https:\u002F\u002Fglm.ai\u002F\" rel=\"nofollow\">GLM-5.1\u003C\u002Fa>、\u003Ca href=\"https:\u002F\u002Fqwenlm.github.io\u002F\" rel=\"nofollow\">Qwen3.6-35B-A3B\u003C\u002Fa> 這些模型，證明 local 或 self-hosted 已經不只是退而求其次，而是可行選項。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1779167648208-i8c2.png\" alt=\"5 個 LLM 的半年轉變\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>真正改變的是「能力對尺寸」的比值。20.9GB 的模型在筆電上就能跑出超乎預期的結果，而 1.5TB 的大模型則能在足夠硬體下展現很強表現。現在的問題不再是能不能本地跑，而是你要選哪種成本與能力的平衡。\u003C\u002Fp>\u003Cul>\u003Cli>Gemma 4：作者認為最強的美系開放模型之一\u003C\u002Fli>\u003Cli>GLM-5.1：體積大、吃硬體，但能力強\u003C\u002Fli>\u003Cli>Qwen3.6-35B-A3B：相對筆電友善\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>4. 個人 AI 助手開始成形\u003C\u002Fh2>\u003Cp>原本不起眼的 Warelay repo，後來變成 \u003Ca href=\"https:\u002F\u002Fopenclaw.ai\u002F\" rel=\"nofollow\">OpenClaw\u003C\u002Fa>，到 2 月時已經吸引大量關注。更重要的是，這類工具開始有了更通用的稱呼：Claws，也就是以 agentic pattern 為核心的個人 AI 助手。\u003C\u002Fp>\u003Cp>這件事重要在於，它把互動模式從一次性的聊天，推進到持續存在的助手。有人甚至會特地買 Mac mini 來跑這些系統。概念很簡單：準備一台專門的機器，讓助手長期處理任務，不要占用主力電腦。\u003C\u002Fp>\u003Cul>\u003Cli>Warelay：最初的 repo 名稱\u003C\u002Fli>\u003Cli>OpenClaw：後來定名並被廣泛注意\u003C\u002Fli>\u003Cli>Claws：逐漸形成的類別名稱\u003C\u002Fli>\u003Cli>常見配置：一台 Mac mini 當助手主機\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>5. Benchmark demo 變得更怪，也更有參考價值\u003C\u002Fh2>\u003Cp>Simon Willison 的 pelican-riding-a-bicycle \u003Ca href=\"\u002Fnews\u002F5-llm-benchmarks-for-business-buyers-2026-zh\">測試\u003C\u002Fa>，會一直被拿來比較模型，因為它荒謬得剛剛好。它難畫、容易辨識，而且不太可能被各家模型專門優化，所以很適合用來看多模態表現。\u003C\u002Fp>\u003Cp>同一時期也出現很多有趣但有意義的 demo，例如用 \u003Ca href=\"https:\u002F\u002Fpyodide.org\u002F\" rel=\"nofollow\">Pyodide\u003C\u002Fa> 在瀏覽器裡跑 WebAssembly，再用 Python 包住 JavaScript。這些例子看起來像玩具，但其實證明了工具鏈已經成熟到足以支持奇怪的實驗。\u003C\u002Fp>\u003Cpre>\u003Ccode>browser → JavaScript → WebAssembly → Pyodide → Python → micro-javascript\u003C\u002Fcode>\u003C\u002Fpre>\u003Cul>\u003Cli>Pelican test：快速檢查模型品質\u003C\u002Fli>\u003Cli>Micro-javascript：小型但很有說服力的實驗\u003C\u002Fli>\u003Cli>重點：工具夠成熟，才會有這些怪但有用的 demo\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>怎麼挑\u003C\u002Fh2>\u003Cp>如果你最在意寫程式效率，先試 \u003Ca href=\"\u002Ftag\u002Fopenai\">OpenAI\u003C\u002Fa> 和 Anthropic 的 agentic 工具，再用自己的 repo 測。若你重視隱私、成本或離線使用，就看 Google、GLM 和 Qwen 的開放模型。若你在做產品，核心判斷已經不是「能不能做」，而是「哪個模型最適合這個任務、這台機器和這個預算」。\u003C\u002Fp>\u003Cp>對多數人來說，最實際的組合是保留一個前沿模型，再配一個本地模型。前者處理難題，後者處理日常工作，會最穩。\u003C\u002Fp>","5 個轉變說明 LLM 為何在 6 個月內快速改變：更強的 coding agents、開放模型與本地工作流。","simonwillison.net","https:\u002F\u002Fsimonwillison.net\u002F2026\u002FMay\u002F19\u002F5-minute-llms\u002F",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1779167646556-uxgd.png","industry","zh","bc80ab6d-7eba-4d56-9e4a-d58aa61328cb",[17,18,19,20,21,22,23,24],"LLM","coding agents","open models","local workflows","Claude","GPT","Gemini","OpenClaw",[26,27,28],"LLM 在半年內的最大變化，是 coding agents 從展示品變成可日常使用。","模型排行榜變動很快，實際任務測試比固定排名更重要。","開放模型與本地部署的可用性提升，讓雲端與離線工作流都更成熟。",5,"2026-05-19T05:13:34.442673+00:00","2026-05-19T05:13:34.431+00:00","caa87b65-9bbc-46fe-bba8-4f4158dd2d8b",{"tags":34,"relatedLang":39,"relatedPosts":43},[35,37],{"name":17,"slug":36},"llm",{"name":21,"slug":38},"claude",{"id":15,"slug":40,"title":41,"language":42},"5-shifts-in-llms-from-the-last-six-months-en","5 shifts in LLMs from the last six months","en",[44,50,56,62,68,74],{"id":45,"slug":46,"title":47,"cover_image":48,"image_url":48,"created_at":49,"category":13},"45eef4b4-fff9-4bbc-9860-a3820395f5c9","webx-2026-speaker-lineup-conference-brief-zh","WebX 2026 把聲量拆成會議簡報","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1783928000041-ukar.png","2026-07-13T07:32:54.333855+00:00",{"id":51,"slug":52,"title":53,"cover_image":54,"image_url":54,"created_at":55,"category":13},"61a27712-a243-481e-9a47-fa84f552ac36","ai-weekly-2026-w29-zh","AI 週報：2026-07-06 ~ 2026-07-13","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1783916422596-zvn0.png","2026-07-13T04:00:33.233975+00:00",{"id":57,"slug":58,"title":59,"cover_image":60,"image_url":60,"created_at":61,"category":13},"9ca76a1c-f59b-4633-9d7e-45a1ce18495d","ai-act-europe-operating-system-ai-zh","AI Act 應被視為歐洲 AI 的作業系統","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1783902767441-90pc.png","2026-07-13T00:32:21.395542+00:00",{"id":63,"slug":64,"title":65,"cover_image":66,"image_url":66,"created_at":67,"category":13},"39624a31-f79e-444f-9850-cabad1885429","booz-allen-openai-deal-real-ai-advantage-zh","Booz Allen 的 OpenAI 合作是真優勢，不是噱頭","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1783900966921-i3t0.png","2026-07-13T00:02:18.55857+00:00",{"id":69,"slug":70,"title":71,"cover_image":72,"image_url":72,"created_at":73,"category":13},"d1753385-8c03-4dec-b939-e5ca8bae9030","opensearch-vector-search-benchmark-5-parts-zh","OpenSearch 向量搜尋基準的 5 種跑法","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1783850566022-b79s.png","2026-07-12T10:02:22.269045+00:00",{"id":75,"slug":76,"title":77,"cover_image":78,"image_url":78,"created_at":79,"category":13},"6e790897-c9af-402c-a928-f2b0cc02f4e6","vector-databases-work-in-production-zh","4 種能上線的向量資料庫選擇","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1783846963245-35py.png","2026-07-12T09:02:23.058273+00:00",[81,86,91,96,101,106,111,116,121,126],{"id":82,"slug":83,"title":84,"created_at":85},"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":87,"slug":88,"title":89,"created_at":90},"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":92,"slug":93,"title":94,"created_at":95},"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":97,"slug":98,"title":99,"created_at":100},"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":102,"slug":103,"title":104,"created_at":105},"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":107,"slug":108,"title":109,"created_at":110},"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":112,"slug":113,"title":114,"created_at":115},"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":117,"slug":118,"title":119,"created_at":120},"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":122,"slug":123,"title":124,"created_at":125},"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":127,"slug":128,"title":129,"created_at":130},"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"]