[{"data":1,"prerenderedAt":966},["ShallowReactive",2],{"header-counts":3,"footer-counts":6,"model-deepseek-v3":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":27,"category":930,"contextWindow":931,"description":932,"extension":933,"maxOutput":934,"meta":935,"navigation":283,"path":936,"pricing":937,"published":938,"relatedTools":939,"releaseDate":943,"seo":944,"slug":945,"stem":946,"strengths":947,"updated":938,"useCases":953,"vendor":958,"vendorEn":959,"weaknesses":960,"__hash__":965},"models\u002Fmodels\u002Fdeepseek-v3.md","DeepSeek-V3",[13],"openai",[15,18,21,24],{"name":16,"score":17},"SWE-bench Verified","61.2%",{"name":19,"score":20},"HumanEval","88.5%",{"name":22,"score":23},"MMLU","84.1%",{"name":25,"score":26},"CMMLU","89.7%",{"type":28,"value":29,"toc":909},"minimark",[30,34,38,41,45,48,61,65,73,162,165,168,171,197,201,209,212,215,219,222,392,396,399,443,446,450,458,461,563,566,604,607,613,616,619,669,672,675,717,720,817,820,871,874,905],[31,32,33],"h2",{"id":33},"概述",[35,36,37],"p",{},"DeepSeek-V3 是深度求索于 2025 年 1 月发布的 671B 参数 MoE 模型，总参数 671B 但每次推理仅激活 37B。最大优势是极致性价比——API 价格是 Claude Sonnet 4 的 1\u002F20，且完全开源可自行部署。",[31,39,40],{"id":40},"核心能力",[42,43,44],"h3",{"id":44},"极致性价比",[35,46,47],{},"Input ¥1\u002FM token，Output ¥2\u002FM token，缓存命中后 Input 仅 ¥0.1\u002FM。这个价格意味着：",[49,50,51,55,58],"ul",{},[52,53,54],"li",{},"100 万字中文处理成本约 ¥2",[52,56,57],{},"一个中型项目全量代码分析约 ¥5",[52,59,60],{},"批量处理 10 万条数据约 ¥20",[42,62,64],{"id":63},"prompt-cache自动命中","Prompt Cache（自动命中）",[35,66,67,68,72],{},"DeepSeek 的 cache 完全自动——任何重复出现的 prompt 前缀（≥64 token）自动命中，Input 价格 -90%。响应里的 ",[69,70,71],"code",{},"usage.prompt_cache_hit_tokens"," 字段显示命中量：",[74,75,80],"pre",{"className":76,"code":77,"language":78,"meta":79,"style":79},"language-python shiki shiki-themes github-light github-dark","resp = client.chat.completions.create(\n    model=\"deepseek-chat\",\n    messages=[...],\n)\nprint(resp.usage.prompt_cache_hit_tokens)   # 命中 cache 的 input token 数\nprint(resp.usage.prompt_cache_miss_tokens)  # 没命中的\n","python","",[69,81,82,98,114,132,138,151],{"__ignoreMap":79},[83,84,87,91,95],"span",{"class":85,"line":86},"line",1,[83,88,90],{"class":89},"sVt8B","resp ",[83,92,94],{"class":93},"szBVR","=",[83,96,97],{"class":89}," client.chat.completions.create(\n",[83,99,101,105,107,111],{"class":85,"line":100},2,[83,102,104],{"class":103},"s4XuR","    model",[83,106,94],{"class":93},[83,108,110],{"class":109},"sZZnC","\"deepseek-chat\"",[83,112,113],{"class":89},",\n",[83,115,117,120,122,125,129],{"class":85,"line":116},3,[83,118,119],{"class":103},"    messages",[83,121,94],{"class":93},[83,123,124],{"class":89},"[",[83,126,128],{"class":127},"sj4cs","...",[83,130,131],{"class":89},"],\n",[83,133,135],{"class":85,"line":134},4,[83,136,137],{"class":89},")\n",[83,139,141,144,147],{"class":85,"line":140},5,[83,142,143],{"class":127},"print",[83,145,146],{"class":89},"(resp.usage.prompt_cache_hit_tokens)   ",[83,148,150],{"class":149},"sJ8bj","# 命中 cache 的 input token 数\n",[83,152,154,156,159],{"class":85,"line":153},6,[83,155,143],{"class":127},[83,157,158],{"class":89},"(resp.usage.prompt_cache_miss_tokens)  ",[83,160,161],{"class":149},"# 没命中的\n",[35,163,164],{},"实测：固定 system prompt + 工具定义共 5000 token，cache 命中后单次调用 input 成本从 ¥0.005 降到 ¥0.0005——批量场景一晚省好几张百元钞。",[42,166,167],{"id":167},"开源",[35,169,170],{},"模型权重完全开源（MIT 协议），可以在自己的 GPU 上部署：",[49,172,173,176,179,190],{},[52,174,175],{},"满血版 FP8：8×H100 (~640GB 显存)",[52,177,178],{},"量化版 INT4：2×H100 或 4×A100",[52,180,181,182,189],{},"通过 ",[183,184,188],"a",{"href":185,"rel":186},"https:\u002F\u002Fgithub.com\u002Fvllm-project\u002Fvllm",[187],"nofollow","vLLM"," \u002F SGLang 高性能推理",[52,191,181,192,196],{},[183,193,195],{"href":194},"\u002Fcoding\u002Flocal\u002Follama.html","Ollama"," 体验（推荐 70B 蒸馏版，本地能跑）",[42,198,200],{"id":199},"moe-架构特性","MoE 架构特性",[35,202,203,204,208],{},"671B 总参数 \u002F 37B 激活——推理时只激活 37B，速度接近 37B 模型，质量接近 671B 模型。但显存要求按总参数算（必须把 671B 都加载进显存），所以",[205,206,207],"strong",{},"自己部署门槛极高","。绝大多数人通过 API 用即可。",[42,210,211],{"id":211},"编程",[35,213,214],{},"SWE-bench Verified 61.2%，接近第一梯队。在 Cursor、Aider 等工具中通过 OpenAI 兼容 API 接入，体验接近 Claude Sonnet 4 的 80% 水平。",[31,216,218],{"id":217},"api-调用示例","API 调用示例",[35,220,221],{},"DeepSeek 提供 OpenAI 兼容 API：",[74,223,225],{"className":76,"code":224,"language":78,"meta":79,"style":79},"from openai import OpenAI\nclient = OpenAI(\n    api_key=\"sk-...\",\n    base_url=\"https:\u002F\u002Fapi.deepseek.com\u002Fv1\",\n)\n\nresp = client.chat.completions.create(\n    model=\"deepseek-chat\",        # V3 别名（推理用 deepseek-reasoner = R1）\n    temperature=0.0,\n    messages=[\n        {\"role\": \"system\", \"content\": \"你是 Python 高级工程师。\"},\n        {\"role\": \"user\", \"content\": \"Review this code...\"},\n    ],\n)\n",[69,226,227,241,251,263,275,279,285,293,307,320,329,358,381,387],{"__ignoreMap":79},[83,228,229,232,235,238],{"class":85,"line":86},[83,230,231],{"class":93},"from",[83,233,234],{"class":89}," openai ",[83,236,237],{"class":93},"import",[83,239,240],{"class":89}," OpenAI\n",[83,242,243,246,248],{"class":85,"line":100},[83,244,245],{"class":89},"client ",[83,247,94],{"class":93},[83,249,250],{"class":89}," OpenAI(\n",[83,252,253,256,258,261],{"class":85,"line":116},[83,254,255],{"class":103},"    api_key",[83,257,94],{"class":93},[83,259,260],{"class":109},"\"sk-...\"",[83,262,113],{"class":89},[83,264,265,268,270,273],{"class":85,"line":134},[83,266,267],{"class":103},"    base_url",[83,269,94],{"class":93},[83,271,272],{"class":109},"\"https:\u002F\u002Fapi.deepseek.com\u002Fv1\"",[83,274,113],{"class":89},[83,276,277],{"class":85,"line":140},[83,278,137],{"class":89},[83,280,281],{"class":85,"line":153},[83,282,284],{"emptyLinePlaceholder":283},true,"\n",[83,286,287,289,291],{"class":85,"line":5},[83,288,90],{"class":89},[83,290,94],{"class":93},[83,292,97],{"class":89},[83,294,295,297,299,301,304],{"class":85,"line":8},[83,296,104],{"class":103},[83,298,94],{"class":93},[83,300,110],{"class":109},[83,302,303],{"class":89},",        ",[83,305,306],{"class":149},"# V3 别名（推理用 deepseek-reasoner = R1）\n",[83,308,310,313,315,318],{"class":85,"line":309},9,[83,311,312],{"class":103},"    temperature",[83,314,94],{"class":93},[83,316,317],{"class":127},"0.0",[83,319,113],{"class":89},[83,321,322,324,326],{"class":85,"line":7},[83,323,119],{"class":103},[83,325,94],{"class":93},[83,327,328],{"class":89},"[\n",[83,330,332,335,338,341,344,347,350,352,355],{"class":85,"line":331},11,[83,333,334],{"class":89},"        {",[83,336,337],{"class":109},"\"role\"",[83,339,340],{"class":89},": ",[83,342,343],{"class":109},"\"system\"",[83,345,346],{"class":89},", ",[83,348,349],{"class":109},"\"content\"",[83,351,340],{"class":89},[83,353,354],{"class":109},"\"你是 Python 高级工程师。\"",[83,356,357],{"class":89},"},\n",[83,359,361,363,365,367,370,372,374,376,379],{"class":85,"line":360},12,[83,362,334],{"class":89},[83,364,337],{"class":109},[83,366,340],{"class":89},[83,368,369],{"class":109},"\"user\"",[83,371,346],{"class":89},[83,373,349],{"class":109},[83,375,340],{"class":89},[83,377,378],{"class":109},"\"Review this code...\"",[83,380,357],{"class":89},[83,382,384],{"class":85,"line":383},13,[83,385,386],{"class":89},"    ],\n",[83,388,390],{"class":85,"line":389},14,[83,391,137],{"class":89},[42,393,395],{"id":394},"与-aider-配合用","与 Aider 配合用",[35,397,398],{},"Aider 是用 DeepSeek-V3 最方便的 CLI 之一：",[74,400,404],{"className":401,"code":402,"language":403,"meta":79,"style":79},"language-bash shiki shiki-themes github-light github-dark","export OPENAI_API_KEY=sk-...\nexport OPENAI_API_BASE=https:\u002F\u002Fapi.deepseek.com\u002Fv1\naider --model deepseek-chat\n","bash",[69,405,406,419,431],{"__ignoreMap":79},[83,407,408,411,414,416],{"class":85,"line":86},[83,409,410],{"class":93},"export",[83,412,413],{"class":89}," OPENAI_API_KEY",[83,415,94],{"class":93},[83,417,418],{"class":89},"sk-...\n",[83,420,421,423,426,428],{"class":85,"line":100},[83,422,410],{"class":93},[83,424,425],{"class":89}," OPENAI_API_BASE",[83,427,94],{"class":93},[83,429,430],{"class":89},"https:\u002F\u002Fapi.deepseek.com\u002Fv1\n",[83,432,433,437,440],{"class":85,"line":116},[83,434,436],{"class":435},"sScJk","aider",[83,438,439],{"class":127}," --model",[83,441,442],{"class":109}," deepseek-chat\n",[35,444,445],{},"成本：一晚做完一个中型 feature 通常 ¥1-3，是用 Claude 的 1\u002F20。",[42,447,449],{"id":448},"在-cursor-中接入","在 Cursor 中接入",[74,451,456],{"className":452,"code":454,"language":455},[453],"language-text","Cursor → Settings → Models → Add Model\n  Provider: OpenAI\n  Base URL: https:\u002F\u002Fapi.deepseek.com\u002Fv1\n  Model: deepseek-chat\n  API Key: sk-...\n","text",[69,457,454],{"__ignoreMap":79},[31,459,460],{"id":460},"关键参数",[462,463,464,480],"table",{},[465,466,467],"thead",{},[468,469,470,474,477],"tr",{},[471,472,473],"th",{},"参数",[471,475,476],{},"推荐",[471,478,479],{},"说明",[481,482,483,497,509,522,535,548],"tbody",{},[468,484,485,491,494],{},[486,487,488],"td",{},[69,489,490],{},"temperature",[486,492,493],{},"0.0-0.3",[486,495,496],{},"编程 \u002F 工具调用",[468,498,499,503,506],{},[486,500,501],{},[69,502,490],{},[486,504,505],{},"1.0",[486,507,508],{},"通用对话",[468,510,511,516,519],{},[486,512,513],{},[69,514,515],{},"top_p",[486,517,518],{},"0.95",[486,520,521],{},"DeepSeek 默认",[468,523,524,529,532],{},[486,525,526],{},[69,527,528],{},"max_tokens",[486,530,531],{},"显式设",[486,533,534],{},"默认 4K，长输出务必调高（上限 8K）",[468,536,537,542,545],{},[486,538,539],{},[69,540,541],{},"frequency_penalty",[486,543,544],{},"0",[486,546,547],{},"一般不动",[468,549,550,555,560],{},[486,551,552],{},[69,553,554],{},"response_format",[486,556,557],{},[69,558,559],{},"{\"type\": \"json_object\"}",[486,561,562],{},"JSON 模式",[31,564,565],{"id":565},"定价",[462,567,568,578],{},[465,569,570],{},[468,571,572,575],{},[471,573,574],{},"项目",[471,576,577],{},"价格",[481,579,580,588,596],{},[468,581,582,585],{},[486,583,584],{},"Input",[486,586,587],{},"¥1 \u002F 百万 token",[468,589,590,593],{},[486,591,592],{},"Input（缓存命中）",[486,594,595],{},"¥0.1 \u002F 百万 token",[468,597,598,601],{},[486,599,600],{},"Output",[486,602,603],{},"¥2 \u002F 百万 token",[35,605,606],{},"这个价格是 GLM-5.2 的一半，是 Claude Sonnet 4 的 1\u002F20。",[35,608,609,612],{},[205,610,611],{},"夜间折扣","：北京时间 00:30-08:30，所有价格再 -50%。批量数据处理可以定时跑在夜间。",[31,614,615],{"id":615},"自行部署",[35,617,618],{},"如果数据敏感不能上云：",[462,620,621,634],{},[465,622,623],{},[468,624,625,628,631],{},[471,626,627],{},"配置",[471,629,630],{},"性能",[471,632,633],{},"成本",[481,635,636,647,658],{},[468,637,638,641,644],{},[486,639,640],{},"8×H100 FP8 满血",[486,642,643],{},"~50 tok\u002Fs 单并发",[486,645,646],{},"~¥30-50 万\u002F月（云租赁）",[468,648,649,652,655],{},[486,650,651],{},"4×H100 INT4 量化",[486,653,654],{},"~30 tok\u002Fs 单并发",[486,656,657],{},"~¥15-25 万\u002F月",[468,659,660,663,666],{},[486,661,662],{},"7B\u002F13B 蒸馏版（Ollama）",[486,664,665],{},"笔记本可跑",[486,667,668],{},"几乎免费",[35,670,671],{},"蒸馏版是 Meta 把 V3 的输出蒸馏到 Llama \u002F Qwen 上的小模型，能力差距明显但本地能跑。",[31,673,674],{"id":674},"适用场景",[49,676,677,683,689,705,711],{},[52,678,679,682],{},[205,680,681],{},"批量处理","：价格极低，适合大规模文本分类、摘要、翻译",[52,684,685,688],{},[205,686,687],{},"私有化部署","：开源协议允许商用，企业可在自有 GPU 上部署",[52,690,691,694,695,699,700,704],{},[205,692,693],{},"编程辅助","：通过 API 接入 ",[183,696,698],{"href":697},"\u002Fcoding\u002Fide\u002Fcursor.html","Cursor"," \u002F ",[183,701,703],{"href":702},"\u002Fcoding\u002Fcli\u002Faider.html","Aider","，低成本替代 Claude",[52,706,707,710],{},[205,708,709],{},"研究实验","：开源权重可用于学术研究和模型微调",[52,712,713,716],{},[205,714,715],{},"后端模型","：Coze \u002F Dify \u002F 自建 Agent 平台的低成本后端",[31,718,719],{"id":719},"与同档对比",[462,721,722,737],{},[465,723,724],{},[468,725,726,729,731,734],{},[471,727,728],{},"维度",[471,730,11],{},[471,732,733],{},"GLM-5.2",[471,735,736],{},"Qwen 3",[481,738,739,752,765,777,791,804],{},[468,740,741,743,746,749],{},[486,742,577],{},[486,744,745],{},"¥1\u002F¥2",[486,747,748],{},"¥2\u002F¥6",[486,750,751],{},"¥0.8\u002F¥2",[468,753,754,757,759,762],{},[486,755,756],{},"SWE-bench",[486,758,17],{},[486,760,761],{},"65.3%",[486,763,764],{},"58.4%",[468,766,767,770,773,775],{},[486,768,769],{},"上下文",[486,771,772],{},"128K",[486,774,772],{},[486,776,772],{},[468,778,779,782,785,788],{},[486,780,781],{},"输出窗口",[486,783,784],{},"8K（短）",[486,786,787],{},"32K",[486,789,790],{},"16K",[468,792,793,795,798,801],{},[486,794,167],{},[486,796,797],{},"✅ 完全",[486,799,800],{},"部分",[486,802,803],{},"✅ 全系列",[468,805,806,809,812,815],{},[486,807,808],{},"缓存折扣",[486,810,811],{},"✅ 自动 -90%",[486,813,814],{},"✅",[486,816,814],{},[31,818,819],{"id":819},"避坑清单",[49,821,822,828,845,851,857],{},[52,823,824,827],{},[205,825,826],{},"8K 输出窗口最短","：长文件 \u002F 长报告生成会被截断。需要 32K+ 输出选 GLM-5.2 或 Qwen。",[52,829,830,839,840,844],{},[205,831,832,835,836],{},[69,833,834],{},"deepseek-chat"," vs ",[69,837,838],{},"deepseek-reasoner","：前者是 V3 通用，后者是 ",[183,841,843],{"href":842},"\u002Fmodels\u002Fdeepseek-r1.html","R1 推理模型","。别选错。",[52,846,847,850],{},[205,848,849],{},"MoE 部署门槛","：想自部署\"满血版\"必须有 8×H100，否则别想，老老实实用 API。",[52,852,853,856],{},[205,854,855],{},"dev 服务器并发","：免费\u002F低 tier RPM 限制较紧，生产前务必充值升级。",[52,858,859,862,863,866,867,870],{},[205,860,861],{},"思维链不要直接喂回去","：deepseek-chat 不输出思维链，但 reasoner 会。Multi-turn 时要把 ",[69,864,865],{},"reasoning_content"," 字段剥掉，只把 ",[69,868,869],{},"content"," 作为 assistant 历史，否则下一轮上下文翻倍。",[31,872,873],{"id":873},"延伸阅读",[49,875,876,882,891,898],{},[52,877,878,879],{},"推理兄弟：",[183,880,881],{"href":842},"DeepSeek-R1",[52,883,884,885,699,888],{},"同档国产：",[183,886,733],{"href":887},"\u002Fmodels\u002Fglm-5.2.html",[183,889,736],{"href":890},"\u002Fmodels\u002Fqwen-3.html",[52,892,893,894],{},"工具集成：",[183,895,897],{"href":896},"\u002Fwiki\u002Ffunction-calling.html","Function Calling",[52,899,900,901],{},"成本控制：",[183,902,904],{"href":903},"\u002Fwiki\u002Ftoken.html","Token",[906,907,908],"style",{},"html pre.shiki code .sVt8B, html code.shiki .sVt8B{--shiki-default:#24292E;--shiki-dark:#E1E4E8}html pre.shiki code .szBVR, html code.shiki .szBVR{--shiki-default:#D73A49;--shiki-dark:#F97583}html pre.shiki code .s4XuR, html code.shiki .s4XuR{--shiki-default:#E36209;--shiki-dark:#FFAB70}html pre.shiki code .sZZnC, html code.shiki .sZZnC{--shiki-default:#032F62;--shiki-dark:#9ECBFF}html pre.shiki code .sj4cs, html code.shiki .sj4cs{--shiki-default:#005CC5;--shiki-dark:#79B8FF}html pre.shiki code .sJ8bj, html code.shiki .sJ8bj{--shiki-default:#6A737D;--shiki-dark:#6A737D}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 .sScJk, html code.shiki .sScJk{--shiki-default:#6F42C1;--shiki-dark:#B392F0}",{"title":79,"searchDepth":116,"depth":116,"links":910},[911,912,919,923,924,925,926,927,928,929],{"id":33,"depth":100,"text":33},{"id":40,"depth":100,"text":40,"children":913},[914,915,916,917,918],{"id":44,"depth":116,"text":44},{"id":63,"depth":116,"text":64},{"id":167,"depth":116,"text":167},{"id":199,"depth":116,"text":200},{"id":211,"depth":116,"text":211},{"id":217,"depth":100,"text":218,"children":920},[921,922],{"id":394,"depth":116,"text":395},{"id":448,"depth":116,"text":449},{"id":460,"depth":100,"text":460},{"id":565,"depth":100,"text":565},{"id":615,"depth":100,"text":615},{"id":674,"depth":100,"text":674},{"id":719,"depth":100,"text":719},{"id":819,"depth":100,"text":819},{"id":873,"depth":100,"text":873},"llm",128000,"深度求索 DeepSeek-V3 开源 MoE 大模型，671B 总参数 \u002F 37B 激活，编程与数学逼近 GPT-4o 第一梯队，API 输入 ¥1\u002FM 全网最低，权重 MIT 协议开放，国内直连无延迟，支持私有化部署。","md",8192,{},"\u002Fmodels\u002Fdeepseek-v3","Input ¥1\u002FM (缓存 ¥0.1\u002FM) · Output ¥2\u002FM","2026-06-21",[940,941,942],"coding\u002Fapi\u002Fopenrouter","coding\u002Flocal\u002Follama","coding\u002Fcli\u002Faider","2025-01-20",{"title":11,"description":932},"deepseek-v3","models\u002Fdeepseek-v3",[948,949,950,951,952],"开源 671B MoE 模型，可自行部署","API 价格全网最低，比 GLM-5.2 还便宜","编程能力接近第一梯队","支持思维链（Chain-of-Thought）推理","国内直连，响应快",[954,955,956,957],"低成本高吞吐 API 调用","私有化部署（开源版本）","编程辅助（通过 API 接入 Cursor 等）","中文 NLP 任务","深度求索","DeepSeek",[961,962,963,964],"8K 输出窗口偏短，长文件生成受限","多步 Agent 场景下稳定性不如 Claude","非编程场景的推理能力略逊 GPT-5","MoE 架构推理部署需要较大显存","bCGleH0DdzspozZcCiOexWK4qL1hRuqerBA0rk8MxcE",1782316489330]