[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-open-source-llms-run-locally-2026-zh":3,"article-related-open-source-llms-run-locally-2026-zh":37,"series-industry-d686fcc4-b444-4a8e-9c79-477ec86b4c2d":90},{"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":29,"views":33,"created_at":34,"published_at":35,"topic_cluster_id":36},"d686fcc4-b444-4a8e-9c79-477ec86b4c2d","open-source-llms-run-locally-2026-zh","10 款可本地跑的開源 LLM，2026 這樣選","\u003Cp data-speakable=\"summary\">這篇整理 10 款可本地執行的開源 LLM，幫你依硬體、上下文和任務類型挑出合適\u003Ca href=\"\u002Fnews\u002Fmlx-community-apple-silicon-model-weights-zh\">模型\u003C\u002Fa>。\u003C\u002Fp>\u003Cp>開源模型已不再只是備案。看完這 10 項，你可以快速決定：要選最強推理、最省顯存、最\u003Ca href=\"\u002Ftag\u002F長上下文\">長上下文\u003C\u002Fa>，還是最適合代理工作流的本地模型。\u003C\u002Fp>\u003Ctable>\u003Cthead>\u003Ctr>\u003Cth>項目\u003C\u002Fth>\u003Cth>強項\u003C\u002Fth>\u003Cth>代表規格\u003C\u002Fth>\u003Cth>典型 VRAM\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Ctd>\u003Ca href=\"https:\u002F\u002Fwww.alibabacloud.com\u002F\" target=\"_blank\" rel=\"noopener\">Qwen 3 235B-A22B\u003C\u002Fa>\u003C\u002Ftd>\u003Ctd>推理與寫程式\u003C\u002Ftd>\u003Ctd>LiveCodeBench 89%，SWE-Bench 40.0%\u003C\u002Ftd>\u003Ctd>約 132 GB Q4\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>\u003Ca href=\"https:\u002F\u002Fwww.deepseek.com\u002F\" target=\"_blank\" rel=\"noopener\">DeepSeek V4 Pro\u003C\u002Fa>\u003C\u002Ftd>\u003Ctd>數學與技術工作\u003C\u002Ftd>\u003Ctd>GSM8K 96.0%，LiveCodeBench 93.5%\u003C\u002Ftd>\u003Ctd>約 136 GB Q4\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>\u003Ca href=\"https:\u002F\u002Fmoonshotai.com\u002F\" target=\"_blank\" rel=\"noopener\">Kimi K2.6\u003C\u002Fa>\u003C\u002Ftd>\u003Ctd>長上下文流程\u003C\u002Ftd>\u003Ctd>200 萬 token context window\u003C\u002Ftd>\u003Ctd>完整上下文需 80GB+\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>\u003Ca href=\"https:\u002F\u002Fwww.zhipuai.cn\u002Fen\u002F\" target=\"_blank\" rel=\"noopener\">GLM-5 \u002F GLM-5.1\u003C\u002Fa>\u003C\u002Ftd>\u003Ctd>Agentic AI\u003C\u002Ftd>\u003Ctd>Tau2-Bench 89.7%\u003C\u002Ftd>\u003Ctd>64GB+ VRAM\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>\u003Ca href=\"https:\u002F\u002Fai.meta.com\u002Fllama\u002F\" target=\"_blank\" rel=\"noopener\">Llama 3.3 70B\u003C\u002Fa>\u003C\u002Ftd>\u003Ctd>單卡全能型\u003C\u002Ftd>\u003Ctd>MMLU 82%，HumanEval 86.0%\u003C\u002Ftd>\u003Ctd>約 40 GB Q4\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003Ch2>1. Qwen 3 235B-A22B：整體最強的本地選擇\u003C\u002Fh2>\u003Cp>如果你只想選一個模型來扛推理、寫程式和長文工作，\u003Ca href=\"https:\u002F\u002Fwww.alibabacloud.com\u002F\" target=\"_blank\" rel=\"noopener\">Qwen 3 235B-A22B\u003C\u002Fa> 是最完整的答案。它的 MoE 設計每次只啟用 22B 參數，實際運算壓力比名義規模更可控。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781222590157-593t.png\" alt=\"10 款可本地跑的開源 LLM，2026 這樣選\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>代價是硬體門檻很高。約 132 GB VRAM 的 Q4 需求，代表它更像工作站或伺服器模型，而不是一般筆電能碰的選項。\u003C\u002Fp>\u003Cul>\u003Cli>LiveCodeBench：89%\u003C\u002Fli>\u003Cli>SWE-Bench：40.0%\u003C\u002Fli>\u003Cli>授權：Apache 2.0\u003C\u002Fli>\u003Cli>適合：企業代理、複雜程式任務\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>2. DeepSeek V4 Pro：數學與技術推理更狠\u003C\u002Fh2>\u003Cp>\u003Ca href=\"https:\u002F\u002Fwww.deepseek.com\u002F\" target=\"_blank\" rel=\"noopener\">DeepSeek V4 Pro\u003C\u002Fa> 在數學和技術題上特別亮眼。96.0% 的 GSM8K 和 93.5% 的 LiveCodeBench，讓它成為重視正確率的工作流首選。\u003C\u002Fp>\u003Cp>它同樣很吃資源，Q4 約 136 GB VRAM，加上 671B MoE 架構，明顯是多卡或企業等級硬體才有的配置。\u003C\u002Fp>\u003Cul>\u003Cli>GSM8K：96.0%\u003C\u002Fli>\u003Cli>SWE-Bench：67.8%\u003C\u002Fli>\u003Cli>授權：MIT\u003C\u002Fli>\u003Cli>適合：數學、研究、競賽程式\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>3. Kimi K2.6：長上下文處理最有感\u003C\u002Fh2>\u003Cp>\u003Ca href=\"https:\u002F\u002Fmoonshotai.com\u002F\" target=\"_blank\" rel=\"noopener\">Kimi K2.6\u003C\u002Fa> 的核心價值是 200 萬 \u003Ca href=\"\u002Ftag\u002Ftoken\">token\u003C\u002Fa> 上下文。這類模型不是拼單次回答，而是讓你能讀長文件、掃大型 codebase，或維持很長的對話脈絡。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781222586891-e03j.png\" alt=\"10 款可本地跑的開源 LLM，2026 這樣選\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>它的成績比較偏實用面：LiveCodeBench 85%，SWE-rebench 43.8%，再加上 Apache 2.0 授權，部署彈性不錯。\u003C\u002Fp>\u003Cul>\u003Cli>Context window：200 萬 token\u003C\u002Fli>\u003Cli>LiveCodeBench：85%\u003C\u002Fli>\u003Cli>授權：Apache 2.0\u003C\u002Fli>\u003Cli>適合：文件分析、多輪工作流\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>4. GLM-5 \u002F GLM-5.1：做代理任務更對味\u003C\u002Fh2>\u003Cp>\u003Ca href=\"https:\u002F\u002Fwww.zhipuai.cn\u002Fen\u002F\" target=\"_blank\" rel=\"noopener\">GLM-5 \u002F GLM-5.1\u003C\u002Fa> 比較像為 \u003Ca href=\"\u002Ftag\u002Fagent\">agent\u003C\u002Fa> 設計的模型，重點是規劃、工具呼叫與多步驟執行。Tau2-Bench 89.7% 這類指標，正好說明它擅長完成任務鏈，而不只是聊天。\u003C\u002Fp>\u003Cp>如果你要做自動化助理、流程編排或需要模型自己拆解步驟，這一組值得優先測。它同時有 89% 的 LiveCodeBench，程式能力也不弱。\u003C\u002Fp>\u003Cul>\u003Cli>Tau2-Bench：89.7%\u003C\u002Fli>\u003Cli>Quality Index：49.64\u003C\u002Fli>\u003Cli>LiveCodeBench：89%\u003C\u002Fli>\u003Cli>適合：代理、規劃、多步驟任務\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>5. Llama 3.3 70B：最實際的全能型\u003C\u002Fh2>\u003Cp>\u003Ca href=\"https:\u002F\u002Fai.meta.com\u002Fllama\u002F\" target=\"_blank\" rel=\"noopener\">Llama 3.3 70B\u003C\u002Fa> 是很多本地部署場景裡最平衡的選擇。生態成熟、工具支援多，搭配量化後，仍有機會在高階消費級硬體上跑起來。\u003C\u002Fp>\u003Cp>82% 的 MMLU、86.0% 的 HumanEval，加上約 40 GB VRAM 的 Q4 需求，讓它落在「夠強、又不至於太難養」的甜蜜點。\u003C\u002Fp>\u003Cul>\u003Cli>MMLU：82%\u003C\u002Fli>\u003Cli>HumanEval：86.0%\u003C\u002Fli>\u003Cli>VRAM：約 40 GB Q4\u003C\u002Fli>\u003Cli>適合：通用用途、微調\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>6. Gemma 3 27B：中階硬體最均衡\u003C\u002Fh2>\u003Cp>\u003Ca href=\"https:\u002F\u002Fai.google.dev\u002Fgemma\" target=\"_blank\" rel=\"noopener\">Gemma 3 27B\u003C\u002Fa> 很適合想要品質與成本平衡的人。它還支援 vision，對多模態工作有額外價值。\u003C\u002Fp>\u003Cp>約 16 GB VRAM 的 Q4 需求，對單卡\u003Ca href=\"\u002Fnews\u002Fwebassembly-turns-browser-editing-into-desktop-grade-docs-zh\">桌機\u003C\u002Fa>或 MacBook Pro M4 Max 這類設備都更現實。MMLU 約 78.6%，HumanEval 87.8%，整體非常均衡。\u003C\u002Fp>\u003Cul>\u003Cli>MMLU：約 78.6%\u003C\u002Fli>\u003Cli>HumanEval：87.8%\u003C\u002Fli>\u003Cli>多模態：支援\u003C\u002Fli>\u003Cli>適合：單卡部署、影像任務\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>7. Mistral Small 3.1 24B：16GB 顯存的務實解\u003C\u002Fh2>\u003Cp>\u003Ca href=\"https:\u002F\u002Fmistral.ai\u002F\" target=\"_blank\" rel=\"noopener\">Mistral Small 3.1 24B\u003C\u002Fa> 的定位很清楚：在 16 GB VRAM 內，盡量保留長上下文與穩定\u003Ca href=\"\u002Fnews\u002Fwhich-lora-multilingual-instruction-tuning-zh\">指令\u003C\u002Fa>遵循。它不是最大，但很實用。\u003C\u002Fp>\u003Cp>128K context 搭配約 16 GB Q4，讓它很適合 RAG、客服機器人和文件型應用。若你要的是穩定落地，而不是榜單最前面，這款很值得試。\u003C\u002Fp>\u003Cul>\u003Cli>Context window：128K token\u003C\u002Fli>\u003Cli>VRAM：約 16 GB Q4\u003C\u002Fli>\u003Cli>授權：Apache 2.0\u003C\u002Fli>\u003Cli>適合：RAG、長文件\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>8. Phi-4 14B：小模型裡的推理效率派\u003C\u002Fh2>\u003Cp>\u003Ca href=\"https:\u002F\u002Fwww.microsoft.com\u002Fen-us\u002Fresearch\u002Fproject\u002Fphi-4\u002F\" target=\"_blank\" rel=\"noopener\">Phi-4 14B\u003C\u002Fa> 的優勢是小而聰明。對於在意推理效率，而不是單純追求參數量的人，它很有吸引力。\u003C\u002Fp>\u003Cp>約 8 到 10 GB VRAM 的 Q4 需求，讓它能進入邊緣裝置、較小桌機，甚至商業產品。MIT 授權也讓商用整合更單純。\u003C\u002Fp>\u003Cul>\u003Cli>模型大小：14B\u003C\u002Fli>\u003Cli>VRAM：約 8-10 GB Q4\u003C\u002Fli>\u003Cli>授權：MIT\u003C\u002Fli>\u003Cli>適合：邊緣部署、商業應用\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>9. MiMo-V2.5-Pro：代理式寫程式與雙語工作\u003C\u002Fh2>\u003Cp>\u003Ca href=\"https:\u002F\u002Fmimo.xiaomi.com\u002F\" target=\"_blank\" rel=\"noopener\">MiMo-V2.5-Pro\u003C\u002Fa> 偏向 \u003Ca href=\"\u002Ftag\u002Fagentic-coding\">agentic coding\u003C\u002Fa>，重點是長鏈任務與自動化流程。它不是最容易量化比較的那種模型，但方向很明確。\u003C\u002Fp>\u003Cp>如果你的場景包含中英雙語、程式輔助或任務拆解，這款值得放進測試清單。它更像專長型工具，而不是萬用型聊天模型。\u003C\u002Fp>\u003Cul>\u003Cli>焦點：agentic coding\u003C\u002Fli>\u003Cli>強項：長鏈推理\u003C\u002Fli>\u003Cli>授權：open weight\u003C\u002Fli>\u003Cli>適合：自動化、雙語流程\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>10. MiniMax M2.7：多模態輸入最完整\u003C\u002Fh2>\u003Cp>\u003Ca href=\"https:\u002F\u002Fwww.minimax.io\u002F\" target=\"_blank\" rel=\"noopener\">MiniMax M2.7\u003C\u002Fa> 支援文字、影像與音訊，適合處理不只文字的工作。當你的產品需要跨媒體理解時，這種能力比單一榜單分數更重要。\u003C\u002Fp>\u003Cp>它的 SWE-rebench 為 39.6%，而且建議 64GB+，所以不是輕量選項。若你要做創作工具、豐富助理或高階多模態系統，它比較對路。\u003C\u002Fp>\u003Cul>\u003Cli>多模態：文字、影像、音訊\u003C\u002Fli>\u003Cli>SWE-rebench：39.6%\u003C\u002Fli>\u003Cli>VRAM：建議 64GB+\u003C\u002Fli>\u003Cli>適合：創作與多模態應用\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>怎麼挑：先看硬體，再看任務\u003C\u002Fh2>\u003Cp>如果你要的是整體最強，而且硬體足夠，先看 \u003Ca href=\"\u002Ftag\u002Fqwen\">Qwen\u003C\u002Fa> 3 235B-A22B；如果偏數學與技術推理，DeepSeek V4 Pro 更合適；如果你常處理超長文件，Kimi K2.6 最直接。\u003C\u002Fp>\u003Cp>多數人更實際的起點會是 Llama 3.3 70B、Gemma 3 27B 或 Mistral Small 3.1 24B。要做代理，選 GLM-5.1；要小型商用部署，Phi-4 14B 最省心。\u003C\u002Fp>","10 款可本地部署的開源 LLM，從 8GB 到 136GB VRAM 都有對應選擇，適合比對推理、寫程式、長上下文與代理任務。","fungies.io","https:\u002F\u002Ffungies.io\u002Fbest-open-source-llms-2026\u002F",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781222590157-593t.png","industry","zh","e3059303-042d-4d99-9a2d-85151ce0290c",[17,18,19,20,21,22,23,24,25,26,27,28],"open source LLM","local AI","LLM benchmarks","Qwen","DeepSeek","Kimi","GLM","Llama","Gemma","Mistral","Phi-4","multimodal model",[30,31,32],"硬體先決：Q4 顯存需求從約 8GB 到 136GB 差距很大，先看你能跑多大模型。","任務導向：推理選 Qwen\u002FDeepSeek，長上下文選 Kimi，代理任務選 GLM。","實用優先：多數本地使用者會落在 Llama 3.3 70B、Gemma 3 27B、Mistral Small 3.1 24B 這三類。",0,"2026-06-12T00:02:33.72144+00:00","2026-06-12T00:02:33.711+00:00","afac0538-ace3-4ab3-bddc-34bfd44625fd",{"tags":38,"relatedLang":49,"relatedPosts":53},[39,41,43,45,47],{"name":20,"slug":40},"qwen",{"name":21,"slug":42},"deepseek",{"name":17,"slug":44},"open-source-llm",{"name":18,"slug":46},"local-ai",{"name":19,"slug":48},"llm-benchmarks",{"id":15,"slug":50,"title":51,"language":52},"open-source-llms-run-locally-2026-en","10 open source LLMs that run locally in 2026","en",[54,60,66,72,78,84],{"id":55,"slug":56,"title":57,"cover_image":58,"image_url":58,"created_at":59,"category":13},"4fd7980c-c59e-4551-9b72-5b432b05c1a0","latam-stablecoin-engineering-hub-hire-zh","LATAM 已經是招募穩定幣工程師的最佳地區","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781244180869-eh2k.png","2026-06-12T06:02:22.765433+00:00",{"id":61,"slug":62,"title":63,"cover_image":64,"image_url":64,"created_at":65,"category":13},"6e8886a7-f6f9-41ad-bb65-7d95905839eb","anthropic-policy-50b-computing-infrastructure-en-zh","Anthropic 推 500 億美元 AI 基建政策","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781240576407-x3ar.png","2026-06-12T05:02:26.5615+00:00",{"id":67,"slug":68,"title":69,"cover_image":70,"image_url":70,"created_at":71,"category":13},"24da72ed-87c9-43bd-b49a-fb4b74a82a79","mlops-vs-ml-engineer-self-taught-career-guide-zh","MLOps vs ML工程師自學指南","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781239680780-eggn.png","2026-06-12T04:47:28.333267+00:00",{"id":73,"slug":74,"title":75,"cover_image":76,"image_url":76,"created_at":77,"category":13},"a0d5612f-4a5a-4a44-96f4-bb4451b2ac51","liveramp-turns-chatgpt-ads-into-sales-proof-zh","LiveRamp 讓 ChatGPT 廣告變成銷售證據","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781236999736-b7dm.png","2026-06-12T04:02:51.553318+00:00",{"id":79,"slug":80,"title":81,"cover_image":82,"image_url":82,"created_at":83,"category":13},"77f70fd2-47ad-4889-a293-e3800e2a92b0","midjourney-software-first-not-hardware-theater-zh","Midjourney 應該堅持軟體優先，不該追逐硬體秀場","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781228869365-xnuy.png","2026-06-12T01:47:17.318544+00:00",{"id":85,"slug":86,"title":87,"cover_image":88,"image_url":88,"created_at":89,"category":13},"ba2f3a83-fd2b-44fe-8abb-e3d31d2a682a","anthropic-tcs-claude-enterprise-deployments-zh","Anthropic 與 TCS 擴大 Claude 企業部署","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781226171859-9gdp.png","2026-06-12T01:02:22.717696+00:00",[91,96,101,106,111,116,121,126,131,136],{"id":92,"slug":93,"title":94,"created_at":95},"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":97,"slug":98,"title":99,"created_at":100},"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":102,"slug":103,"title":104,"created_at":105},"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":107,"slug":108,"title":109,"created_at":110},"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":112,"slug":113,"title":114,"created_at":115},"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":117,"slug":118,"title":119,"created_at":120},"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":122,"slug":123,"title":124,"created_at":125},"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":127,"slug":128,"title":129,"created_at":130},"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":132,"slug":133,"title":134,"created_at":135},"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":137,"slug":138,"title":139,"created_at":140},"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"]