[{"data":1,"prerenderedAt":1135},["ShallowReactive",2],{"header-counts":3,"footer-counts":6,"model-llama-4":9},{"tools":4,"reviews":5},65,7,{"tools":4,"reviews":5,"playbooks":7,"news":8},10,8,{"id":10,"title":11,"apiCompatible":12,"benchmarks":14,"body":24,"category":1101,"contextWindow":1102,"description":1103,"extension":1104,"maxOutput":1105,"meta":1106,"navigation":304,"path":1107,"pricing":1108,"published":1109,"relatedTools":1110,"releaseDate":1113,"seo":1114,"slug":1115,"stem":1116,"strengths":1117,"updated":1109,"useCases":1123,"vendor":1128,"vendorEn":1128,"weaknesses":1129,"__hash__":1134},"models\u002Fmodels\u002Fllama-4.md","Llama 4",[13],"openai",[15,18,21],{"name":16,"score":17},"SWE-bench Verified","52.3%",{"name":19,"score":20},"HumanEval","82.1%",{"name":22,"score":23},"MMLU","80.2%",{"type":25,"value":26,"toc":1078},"minimark",[27,31,40,43,46,50,53,85,88,91,187,191,203,206,209,246,249,253,256,312,316,374,377,381,410,414,417,598,601,604,664,668,671,691,694,697,763,774,777,903,909,931,935,938,964,967,981,984,1038,1041,1074],[28,29,30],"h2",{"id":30},"概述",[32,33,34,35,39],"p",{},"Llama 4 是 Meta 于 2025 年 7 月发布的开源旗舰模型系列。最大价值是",[36,37,38],"strong",{},"完全免费可商用","——没有 API 费用，只有 GPU 成本。社区生态最强，工具链最完善。",[32,41,42],{},"Llama 系列在开源大模型中处于 \"事实标准\" 地位——Hugging Face 上绝大多数微调模型、量化版本、RAG 教程都基于 Llama。即便实际效果未必是同档最好，\"会用 Llama\" 是企业 AI 团队的基本素质。",[28,44,45],{"id":45},"核心能力",[47,48,49],"h3",{"id":49},"完全开源",[32,51,52],{},"Llama 4 采用 Meta 自有的开源协议（Llama Community License），允许商用。这意味着：",[54,55,56,63,72,75,78],"ul",{},[57,58,59,60],"li",{},"企业可以自由部署，",[36,61,62],{},"数据不出内网",[57,64,65,66,71],{},"可以对模型进行微调（",[67,68,70],"a",{"href":69},"\u002Fwiki\u002Flora.html","LoRA"," \u002F 全量 SFT 都行）",[57,73,74],{},"可以集成到自己的产品中并商业化",[57,76,77],{},"无 API 调用费用，无 token 计费",[57,79,80,81,84],{},"输出内容",[36,82,83],{},"不受任何厂商内容审查策略","约束",[32,86,87],{},"注意\"完全开源\"有边界：月活 > 7 亿的产品需要额外申请商用授权（基本只针对 Meta 的直接竞品）；普通企业用免费即可。",[47,89,90],{"id":90},"多尺寸版本",[92,93,94,113],"table",{},[95,96,97],"thead",{},[98,99,100,104,107,110],"tr",{},[101,102,103],"th",{},"版本",[101,105,106],{},"参数量",[101,108,109],{},"显存需求",[101,111,112],{},"适用场景",[114,115,116,131,145,159,173],"tbody",{},[98,117,118,122,125,128],{},[119,120,121],"td",{},"Llama 4 8B",[119,123,124],{},"8B",[119,126,127],{},"8-16GB",[119,129,130],{},"个人本地、IoT、浏览器 WASM",[98,132,133,136,139,142],{},[119,134,135],{},"Llama 4 70B",[119,137,138],{},"70B",[119,140,141],{},"48-128GB",[119,143,144],{},"中小企业部署",[98,146,147,150,153,156],{},[119,148,149],{},"Llama 4 405B",[119,151,152],{},"405B",[119,154,155],{},"多卡 H100（>800GB）",[119,157,158],{},"云端高性能",[98,160,161,164,167,170],{},[119,162,163],{},"Llama 4 Maverick",[119,165,166],{},"17B 激活 \u002F 400B 总 MoE",[119,168,169],{},"8×H100",[119,171,172],{},"平衡型",[98,174,175,178,181,184],{},[119,176,177],{},"Llama 4 Scout",[119,179,180],{},"17B 激活 \u002F 109B 总 MoE",[119,182,183],{},"4×H100",[119,185,186],{},"10M 超长上下文（实验性）",[47,188,190],{"id":189},"scout-的-10m-上下文","Scout 的 10M 上下文",[32,192,193,194,197,198,202],{},"Llama 4 Scout 号称支持 10M token 上下文——",[36,195,196],{},"业界最长","，远超 ",[67,199,201],{"href":200},"\u002Fmodels\u002Fgemini-2.5-pro.html","Gemini 2.5 Pro","（1M）。但 Meta 自己也承认这是\"实验性\"，超过 1M 后质量下降明显，长文 benchmark 上不如 Gemini Pro。当前最实用范围在 256K-512K。",[47,204,205],{"id":205},"社区生态",[32,207,208],{},"Llama 的社区生态是所有开源模型中最强的：",[54,210,211,223,226,229,237,240,243],{},[57,212,213,217,218,222],{},[67,214,216],{"href":215},"\u002Fcoding\u002Flocal\u002Follama.html","Ollama"," \u002F ",[67,219,221],{"href":220},"\u002Fcoding\u002Flocal\u002Flm-studio.html","LM Studio"," 一键本地部署",[57,224,225],{},"Hugging Face 上有数万微调版本（医疗、法律、编程、角色扮演各领域）",[57,227,228],{},"vLLM \u002F TGI \u002F SGLang 高性能推理框架",[57,230,231,232,236],{},"LangChain \u002F LlamaIndex \u002F ",[67,233,235],{"href":234},"\u002Fcoding\u002Fcli\u002Faider.html","Aider"," 原生支持",[57,238,239],{},"大量量化版本（GGUF \u002F AWQ \u002F GPTQ \u002F EXL2 \u002F MLX）",[57,241,242],{},"Apple Silicon 上跑 MLX 版本，M2\u002FM3 Max 跑 70B 完全可行",[57,244,245],{},"各种 fine-tune 框架（Unsloth \u002F Axolotl \u002F torchtune）默认支持",[28,247,248],{"id":248},"部署方式",[47,250,252],{"id":251},"本地个人开发者","本地（个人开发者）",[32,254,255],{},"最简单：",[257,258,263],"pre",{"className":259,"code":260,"language":261,"meta":262,"style":262},"language-bash shiki shiki-themes github-light github-dark","# Ollama 一键\nollama pull llama4:8b\nollama run llama4:8b\n\n# 或者通过 LM Studio 图形界面\n","bash","",[264,265,266,275,289,299,306],"code",{"__ignoreMap":262},[267,268,271],"span",{"class":269,"line":270},"line",1,[267,272,274],{"class":273},"sJ8bj","# Ollama 一键\n",[267,276,278,282,286],{"class":269,"line":277},2,[267,279,281],{"class":280},"sScJk","ollama",[267,283,285],{"class":284},"sZZnC"," pull",[267,287,288],{"class":284}," llama4:8b\n",[267,290,292,294,297],{"class":269,"line":291},3,[267,293,281],{"class":280},[267,295,296],{"class":284}," run",[267,298,288],{"class":284},[267,300,302],{"class":269,"line":301},4,[267,303,305],{"emptyLinePlaceholder":304},true,"\n",[267,307,309],{"class":269,"line":308},5,[267,310,311],{"class":273},"# 或者通过 LM Studio 图形界面\n",[47,313,315],{"id":314},"企业自部署vllm","企业自部署（vLLM）",[257,317,319],{"className":259,"code":318,"language":261,"meta":262,"style":262},"vllm serve meta-llama\u002FLlama-4-70B-Instruct \\\n    --tensor-parallel-size 4 \\\n    --max-model-len 131072 \\\n    --enable-prefix-caching \\\n    --quantization awq      # 4-bit 量化省显存\n",[264,320,321,336,346,356,363],{"__ignoreMap":262},[267,322,323,326,329,332],{"class":269,"line":270},[267,324,325],{"class":280},"vllm",[267,327,328],{"class":284}," serve",[267,330,331],{"class":284}," meta-llama\u002FLlama-4-70B-Instruct",[267,333,335],{"class":334},"sj4cs"," \\\n",[267,337,338,341,344],{"class":269,"line":277},[267,339,340],{"class":334},"    --tensor-parallel-size",[267,342,343],{"class":334}," 4",[267,345,335],{"class":334},[267,347,348,351,354],{"class":269,"line":291},[267,349,350],{"class":334},"    --max-model-len",[267,352,353],{"class":334}," 131072",[267,355,335],{"class":334},[267,357,358,361],{"class":269,"line":301},[267,359,360],{"class":334},"    --enable-prefix-caching",[267,362,335],{"class":334},[267,364,365,368,371],{"class":269,"line":308},[267,366,367],{"class":334},"    --quantization",[267,369,370],{"class":284}," awq",[267,372,373],{"class":273},"      # 4-bit 量化省显存\n",[32,375,376],{},"OpenAI 兼容 API 自动暴露在 8000 端口，所有支持 OpenAI 的工具直接用。",[47,378,380],{"id":379},"云上托管不想自己运维","云上托管（不想自己运维）",[54,382,383,389,395,404],{},[57,384,385,388],{},[36,386,387],{},"AWS Bedrock","：Llama 4 70B \u002F 405B 都有，按 token 计费但比 Claude 便宜",[57,390,391,394],{},[36,392,393],{},"Groq","：Llama 系列的极速 inference 服务，70B 跑 500 tok\u002Fs（业界最快）",[57,396,397,217,400,403],{},[36,398,399],{},"Together AI",[36,401,402],{},"Replicate","：开发者友好定价",[57,405,406,409],{},[36,407,408],{},"国内","：硅基流动 \u002F 阿里云百炼也有部分 Llama 部署",[28,411,413],{"id":412},"api-调用示例","API 调用示例",[32,415,416],{},"任何 OpenAI 兼容客户端都能用：",[257,418,422],{"className":419,"code":420,"language":421,"meta":262,"style":262},"language-python shiki shiki-themes github-light github-dark","from openai import OpenAI\n\n# 本地 vLLM \u002F Ollama\nclient = OpenAI(api_key=\"dummy\", base_url=\"http:\u002F\u002Flocalhost:8000\u002Fv1\")\n\n# Groq（云端最快）\nclient = OpenAI(api_key=\"gsk_...\", base_url=\"https:\u002F\u002Fapi.groq.com\u002Fopenai\u002Fv1\")\n\nresp = client.chat.completions.create(\n    model=\"llama-4-70b\",\n    temperature=0.7,\n    messages=[{\"role\": \"user\", \"content\": \"...\"}],\n)\n","python",[264,423,424,440,444,449,483,487,493,519,523,534,547,560,593],{"__ignoreMap":262},[267,425,426,430,434,437],{"class":269,"line":270},[267,427,429],{"class":428},"szBVR","from",[267,431,433],{"class":432},"sVt8B"," openai ",[267,435,436],{"class":428},"import",[267,438,439],{"class":432}," OpenAI\n",[267,441,442],{"class":269,"line":277},[267,443,305],{"emptyLinePlaceholder":304},[267,445,446],{"class":269,"line":291},[267,447,448],{"class":273},"# 本地 vLLM \u002F Ollama\n",[267,450,451,454,457,460,464,466,469,472,475,477,480],{"class":269,"line":301},[267,452,453],{"class":432},"client ",[267,455,456],{"class":428},"=",[267,458,459],{"class":432}," OpenAI(",[267,461,463],{"class":462},"s4XuR","api_key",[267,465,456],{"class":428},[267,467,468],{"class":284},"\"dummy\"",[267,470,471],{"class":432},", ",[267,473,474],{"class":462},"base_url",[267,476,456],{"class":428},[267,478,479],{"class":284},"\"http:\u002F\u002Flocalhost:8000\u002Fv1\"",[267,481,482],{"class":432},")\n",[267,484,485],{"class":269,"line":308},[267,486,305],{"emptyLinePlaceholder":304},[267,488,490],{"class":269,"line":489},6,[267,491,492],{"class":273},"# Groq（云端最快）\n",[267,494,495,497,499,501,503,505,508,510,512,514,517],{"class":269,"line":5},[267,496,453],{"class":432},[267,498,456],{"class":428},[267,500,459],{"class":432},[267,502,463],{"class":462},[267,504,456],{"class":428},[267,506,507],{"class":284},"\"gsk_...\"",[267,509,471],{"class":432},[267,511,474],{"class":462},[267,513,456],{"class":428},[267,515,516],{"class":284},"\"https:\u002F\u002Fapi.groq.com\u002Fopenai\u002Fv1\"",[267,518,482],{"class":432},[267,520,521],{"class":269,"line":8},[267,522,305],{"emptyLinePlaceholder":304},[267,524,526,529,531],{"class":269,"line":525},9,[267,527,528],{"class":432},"resp ",[267,530,456],{"class":428},[267,532,533],{"class":432}," client.chat.completions.create(\n",[267,535,536,539,541,544],{"class":269,"line":7},[267,537,538],{"class":462},"    model",[267,540,456],{"class":428},[267,542,543],{"class":284},"\"llama-4-70b\"",[267,545,546],{"class":432},",\n",[267,548,550,553,555,558],{"class":269,"line":549},11,[267,551,552],{"class":462},"    temperature",[267,554,456],{"class":428},[267,556,557],{"class":334},"0.7",[267,559,546],{"class":432},[267,561,563,566,568,571,574,577,580,582,585,587,590],{"class":269,"line":562},12,[267,564,565],{"class":462},"    messages",[267,567,456],{"class":428},[267,569,570],{"class":432},"[{",[267,572,573],{"class":284},"\"role\"",[267,575,576],{"class":432},": ",[267,578,579],{"class":284},"\"user\"",[267,581,471],{"class":432},[267,583,584],{"class":284},"\"content\"",[267,586,576],{"class":432},[267,588,589],{"class":284},"\"...\"",[267,591,592],{"class":432},"}],\n",[267,594,596],{"class":269,"line":595},13,[267,597,482],{"class":432},[28,599,600],{"id":600},"定价",[32,602,603],{},"模型本身免费。成本仅为 GPU 运行费用：",[92,605,606,615],{},[95,607,608],{},[98,609,610,612],{},[101,611,248],{},[101,613,614],{},"月成本估算（中等负载）",[114,616,617,625,632,640,648,656],{},[98,618,619,622],{},[119,620,621],{},"本地（8B，消费级 GPU）",[119,623,624],{},"¥0（电费）",[98,626,627,630],{},[119,628,629],{},"本地（70B，Mac Studio M3 Max）",[119,631,624],{},[98,633,634,637],{},[119,635,636],{},"云 GPU（70B，A100 80G）",[119,638,639],{},"~¥3,000-5,000\u002F月",[98,641,642,645],{},[119,643,644],{},"云 GPU（405B，8×H100）",[119,646,647],{},"~¥30,000-50,000\u002F月",[98,649,650,653],{},[119,651,652],{},"Groq 70B API",[119,654,655],{},"$0.59\u002FM Input · $0.79\u002FM Output",[98,657,658,661],{},[119,659,660],{},"AWS Bedrock 70B",[119,662,663],{},"$0.72\u002FM Input · $0.72\u002FM Output",[47,665,667],{"id":666},"自部署-roi-计算","自部署 ROI 计算",[32,669,670],{},"什么时候自部署划算？",[54,672,673,676,688],{},[57,674,675],{},"如果月 token 用量 > 100 亿，且能接受 70B 而非旗舰：自部署 70B 比走 Claude API 便宜 20-50 倍",[57,677,678,679,217,683,687],{},"如果月用量 \u003C 1 亿：直接走 ",[67,680,682],{"href":681},"\u002Fmodels\u002Fdeepseek-v3.html","DeepSeek-V3",[67,684,686],{"href":685},"\u002Fmodels\u002Fglm-5.2.html","GLM-5.2"," API 更省事",[57,689,690],{},"如果是数据合规驱动（医疗\u002F金融\u002F政府）：自部署是必选项，不算 ROI",[28,692,693],{"id":693},"微调入口",[32,695,696],{},"Llama 4 是微调最方便的开源模型，主流 fine-tune 工具链：",[92,698,699,712],{},[95,700,701],{},[98,702,703,706,709],{},[101,704,705],{},"工具",[101,707,708],{},"适合人群",[101,710,711],{},"特点",[114,713,714,730,741,752],{},[98,715,716,724,727],{},[119,717,718],{},[67,719,723],{"href":720,"rel":721},"https:\u002F\u002Fgithub.com\u002Funslothai\u002Funsloth",[722],"nofollow","Unsloth",[119,725,726],{},"个人 \u002F 小团队",[119,728,729],{},"单卡 7B QLoRA 几小时",[98,731,732,735,738],{},[119,733,734],{},"Axolotl",[119,736,737],{},"企业",[119,739,740],{},"配置化、多任务",[98,742,743,746,749],{},[119,744,745],{},"torchtune",[119,747,748],{},"研究",[119,750,751],{},"PyTorch 官方，灵活",[98,753,754,757,760],{},[119,755,756],{},"TRL",[119,758,759],{},"RLHF \u002F DPO",[119,761,762],{},"HuggingFace 出品",[32,764,765,766,768,769,773],{},"详见 ",[67,767,70],{"href":69}," 与 ",[67,770,772],{"href":771},"\u002Fwiki\u002Ffine-tuning-vs-rag.html","Fine-tuning vs RAG","——大多数业务问题用 RAG 解决，确实需要\"教模型新行为\"才上微调。",[28,775,776],{"id":776},"与其他模型怎么选",[92,778,779,795],{},[95,780,781],{},[98,782,783,786,788,790,792],{},[101,784,785],{},"维度",[101,787,11],{},[101,789,686],{},[101,791,682],{},[101,793,794],{},"Qwen 3",[114,796,797,814,831,846,860,873,886],{},[98,798,799,802,805,808,811],{},[119,800,801],{},"开源",[119,803,804],{},"✅ 完全",[119,806,807],{},"部分（GLM-4 系列开源）",[119,809,810],{},"✅ MoE",[119,812,813],{},"✅ 全系列",[98,815,816,819,822,825,828],{},[119,817,818],{},"商用",[119,820,821],{},"✅（\u003C 7 亿 MAU）",[119,823,824],{},"需授权",[119,826,827],{},"✅ MIT",[119,829,830],{},"✅",[98,832,833,836,839,842,844],{},[119,834,835],{},"编程",[119,837,838],{},"★★★☆☆",[119,840,841],{},"★★★★☆",[119,843,841],{},[119,845,838],{},[98,847,848,851,853,856,858],{},[119,849,850],{},"中文",[119,852,838],{},[119,854,855],{},"★★★★★",[119,857,841],{},[119,859,841],{},[98,861,862,865,867,869,871],{},[119,863,864],{},"多语言",[119,866,841],{},[119,868,838],{},[119,870,838],{},[119,872,855],{},[98,874,875,878,880,882,884],{},[119,876,877],{},"社区",[119,879,855],{},[119,881,838],{},[119,883,841],{},[119,885,855],{},[98,887,888,891,894,897,900],{},[119,889,890],{},"端侧模型",[119,892,893],{},"8B 一档",[119,895,896],{},"9B 一档",[119,898,899],{},"蒸馏版",[119,901,902],{},"0.5B-14B 全覆盖",[32,904,905,908],{},[36,906,907],{},"建议","：",[54,910,911,914,923,928],{},[57,912,913],{},"需要私有化部署且中文要求不高 → Llama 4（社区生态最强）",[57,915,916,917,919,920],{},"中文场景 → ",[67,918,686],{"href":685}," 或 ",[67,921,794],{"href":922},"\u002Fmodels\u002Fqwen-3.html",[57,924,925,926],{},"极致低成本（API） → ",[67,927,682],{"href":681},[57,929,930],{},"端侧多尺寸 → Qwen 3",[28,932,934],{"id":933},"适合-不适合","适合 \u002F 不适合",[32,936,937],{},"✅ 适合：",[54,939,940,943,946,949,958,961],{},[57,941,942],{},"企业私有化部署（数据不出内网）",[57,944,945],{},"模型微调研究 \u002F 实验",[57,947,948],{},"海外多语言应用",[57,950,951,952,957],{},"角色扮演 \u002F 内容创作（无审查限制，",[67,953,956],{"href":954,"rel":955},"https:\u002F\u002Fhuggingface.co\u002Fmodels?search=llama+uncensored",[722],"Hugging Face 上有大量 uncensored fine-tune","）",[57,959,960],{},"学术研究 \u002F 教学",[57,962,963],{},"AI 产品 PoC（先用 Llama 验证，再迁到 API）",[32,965,966],{},"❌ 不适合：",[54,968,969,972,975,978],{},[57,970,971],{},"主力编程（不如 Claude \u002F GLM）",[57,973,974],{},"纯中文场景（不如国产）",[57,976,977],{},"团队没运维能力（API 模型更省心）",[57,979,980],{},"极致质量场景（开源旗舰仍落后闭源旗舰半代）",[28,982,983],{"id":983},"避坑清单",[54,985,986,992,998,1011,1020,1026,1032],{},[57,987,988,991],{},[36,989,990],{},"8B \u002F 70B \u002F 405B 跨档差距巨大","：用 8B 跑不起来的任务不是 Llama 不行，是参数量不够，先升 70B 再下结论。",[57,993,994,997],{},[36,995,996],{},"量化精度不要太低","：q4 \u002F Q4_K_M 是底线，q2 \u002F q3 质量崩坏严重；fp16 \u002F Q8_0 才是真实质量。",[57,999,1000,1006,1007,1010],{},[36,1001,1002,1005],{},[264,1003,1004],{},"Llama-4-*-Instruct"," vs base","：基础模型（base）不会聊天，必须用 ",[264,1008,1009],{},"-Instruct"," 版本；新手常踩。",[57,1012,1013,1016,1017,1019],{},[36,1014,1015],{},"Scout 10M 上下文别迷信","：超过 1M 后实测召回率断崖，长文场景仍推荐 ",[67,1018,201],{"href":200},"。",[57,1021,1022,1025],{},[36,1023,1024],{},"中文场景慎用","：训练语料中文占比低，中文流畅度不如国产；要在中文上用 Llama，先看有没有中文社区微调版本（如 Llama-4-Chinese）。",[57,1027,1028,1031],{},[36,1029,1030],{},"Llama License 不是 MIT","：商用前看条款（7 亿 MAU 限制、不得用 Llama 输出训练非 Llama 模型等）。",[57,1033,1034,1037],{},[36,1035,1036],{},"Groq 限流严格","：免费档 RPM 很低，生产用要付费升级。",[28,1039,1040],{"id":1040},"延伸阅读",[54,1042,1043,1050,1055,1060,1067],{},[57,1044,1045,1046,217,1048],{},"本地部署：",[67,1047,216],{"href":215},[67,1049,221],{"href":220},[57,1051,1052,1053],{},"微调方法：",[67,1054,70],{"href":69},[57,1056,1057,1058],{},"何时微调：",[67,1059,772],{"href":771},[57,1061,1062,1063,217,1065],{},"同档对比：",[67,1064,682],{"href":681},[67,1066,794],{"href":922},[57,1068,1069,1070],{},"推理优化：",[67,1071,1073],{"href":1072},"\u002Fwiki\u002Ftoken.html","Token",[1075,1076,1077],"style",{},"html pre.shiki code .sJ8bj, html code.shiki .sJ8bj{--shiki-default:#6A737D;--shiki-dark:#6A737D}html pre.shiki code .sScJk, html code.shiki .sScJk{--shiki-default:#6F42C1;--shiki-dark:#B392F0}html pre.shiki code .sZZnC, html code.shiki .sZZnC{--shiki-default:#032F62;--shiki-dark:#9ECBFF}html .default .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}html.dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}html pre.shiki code .sj4cs, html code.shiki .sj4cs{--shiki-default:#005CC5;--shiki-dark:#79B8FF}html pre.shiki code .szBVR, html code.shiki .szBVR{--shiki-default:#D73A49;--shiki-dark:#F97583}html pre.shiki code .sVt8B, html code.shiki .sVt8B{--shiki-default:#24292E;--shiki-dark:#E1E4E8}html pre.shiki code .s4XuR, html code.shiki .s4XuR{--shiki-default:#E36209;--shiki-dark:#FFAB70}",{"title":262,"searchDepth":291,"depth":291,"links":1079},[1080,1081,1087,1092,1093,1096,1097,1098,1099,1100],{"id":30,"depth":277,"text":30},{"id":45,"depth":277,"text":45,"children":1082},[1083,1084,1085,1086],{"id":49,"depth":291,"text":49},{"id":90,"depth":291,"text":90},{"id":189,"depth":291,"text":190},{"id":205,"depth":291,"text":205},{"id":248,"depth":277,"text":248,"children":1088},[1089,1090,1091],{"id":251,"depth":291,"text":252},{"id":314,"depth":291,"text":315},{"id":379,"depth":291,"text":380},{"id":412,"depth":277,"text":413},{"id":600,"depth":277,"text":600,"children":1094},[1095],{"id":666,"depth":291,"text":667},{"id":693,"depth":277,"text":693},{"id":776,"depth":277,"text":776},{"id":933,"depth":277,"text":934},{"id":983,"depth":277,"text":983},{"id":1040,"depth":277,"text":1040},"llm",256000,"Meta Llama 4 开源旗舰大模型，完全免费可商用 + 多模态原生支持，HuggingFace 社区生态与微调工具链最完整，权重开放支持 LoRA \u002F 全参微调与私有化部署，适合学术研究、企业自托管与合规敏感场景。","md",16384,{},"\u002Fmodels\u002Fllama-4","开源免费（自行部署成本仅 GPU）","2026-06-21",[1111,1112],"coding\u002Flocal\u002Follama","coding\u002Flocal\u002Flm-studio","2025-07-23",{"title":11,"description":1103},"llama-4","models\u002Fllama-4",[1118,1119,1120,1121,1122],"完全开源，可商用，无使用限制","Scout 版本 10M 超长上下文（实验性）","社区生态最强，工具链完善","多尺寸可选（8B\u002F70B\u002F405B\u002FMaverick\u002FScout）","支持多语言",[1124,1125,1126,1127],"私有化部署（数据不出企业）","学术研究与模型微调","低成本批量处理","定制化模型开发","Meta",[1130,1131,1132,1133],"编程能力弱于 Claude\u002FGPT\u002FGLM","需要自行部署，运维门槛高","10M 上下文为实验性，质量不稳定","中文能力不如国产模型","mQZsmfQMT0dg8BGW1xKfxA6rDL_Y4JowE3kZfNPZXCA",1782316489332]