[{"data":1,"prerenderedAt":550},["ShallowReactive",2],{"header-counts":3,"footer-counts":6,"model-kimi-k2-thinking":9},{"tools":4,"reviews":5},70,12,{"tools":4,"reviews":5,"playbooks":7,"news":8},15,13,{"id":10,"title":11,"apiCompatible":12,"benchmarks":15,"body":34,"category":513,"contextWindow":514,"description":515,"extension":516,"maxOutput":517,"meta":518,"navigation":519,"path":520,"pricing":521,"published":522,"relatedTools":523,"releaseDate":527,"seo":528,"slug":529,"stem":530,"strengths":531,"updated":522,"useCases":537,"vendor":542,"vendorEn":543,"weaknesses":544,"__hash__":549},"models\u002Fmodels\u002Fkimi-k2-thinking.md","Kimi K2 Thinking",[13,14],"openai","anthropic",[16,19,22,25,28,31],{"name":17,"score":18},"SWE-bench Verified","71.3%",{"name":20,"score":21},"BrowseComp","60.2%",{"name":23,"score":24},"AIME25 (w\u002F Python)","99.1%",{"name":26,"score":27},"HLE (Heavy Mode)","51.0%",{"name":29,"score":30},"GPQA","84.5%",{"name":32,"score":33},"LiveCodeBench V6","83.1%",{"type":35,"value":36,"toc":500},"minimark",[37,41,51,54,58,71,75,78,81,149,152,174,177,215,330,333,336,421,434,437,467,470,496],[38,39,40],"h2",{"id":40},"概述",[42,43,44,45,50],"p",{},"Kimi K2 Thinking 是月之暗面（Moonshot AI）于 2025 年 11 月 6 日开源的「思维模型」，定位为「thinking agent」——把分步推理（Chain-of-Thought）与动态工具调用交织起来处理复杂长程任务。它是站内 ",[46,47,49],"a",{"href":48},"\u002Fmodels\u002Fkimi-k2.html","Kimi K2"," 的推理强化版，发布时在推理\u002F编码\u002Fagentic 基准上被视为开源 SOTA，部分指标超过 GPT-5、Claude Sonnet 4.5（Thinking）和 Grok-4。",[38,52,53],{"id":53},"核心能力",[55,56,57],"h3",{"id":57},"超长工具调用链",[42,59,60,61,65,66,70],{},"最突出的能力是可维持 ",[62,63,64],"strong",{},"200–300 次连续工具调用","而保持目标连贯，远超此前模型通常 30–50 步就开始退化的水平。这对长链路 agent（深度搜索、多步自动化）是质变。关于工具调用机制见 ",[46,67,69],{"href":68},"\u002Fwiki\u002Ffunction-calling.html","Function Calling","。",[55,72,74],{"id":73},"原生-int4-量化","原生 INT4 量化",[42,76,77],{},"后训练阶段采用量化感知训练（QAT），原生支持 INT4：低延迟模式提速 2x、显存占用下降，且几乎无性能损失。这降低了自部署成本。",[55,79,80],{"id":80},"模型规格",[82,83,84,97],"table",{},[85,86,87],"thead",{},[88,89,90,94],"tr",{},[91,92,93],"th",{},"项目",[91,95,96],{},"参数",[98,99,100,109,117,125,133,141],"tbody",{},[88,101,102,106],{},[103,104,105],"td",{},"架构",[103,107,108],{},"Mixture-of-Experts (MoE)",[88,110,111,114],{},[103,112,113],{},"总参 \u002F 激活",[103,115,116],{},"1T \u002F 32B",[88,118,119,122],{},[103,120,121],{},"专家数",[103,123,124],{},"384（每 token 选 8 + 1 共享）",[88,126,127,130],{},[103,128,129],{},"注意力",[103,131,132],{},"MLA（多头潜在注意力）",[88,134,135,138],{},[103,136,137],{},"上下文",[103,139,140],{},"256K",[88,142,143,146],{},[103,144,145],{},"许可",[103,147,148],{},"Modified MIT（可商用）",[55,150,151],{"id":151},"基准亮点",[153,154,155,162,168],"ul",{},[156,157,158,161],"li",{},[62,159,160],{},"Agentic 搜索","：BrowseComp 60.2%，超过 GPT-5 High 的 54.9%",[156,163,164,167],{},[62,165,166],{},"推理","：AIME25（含 Python）99.1%，GPQA 84.5%",[156,169,170,173],{},[62,171,172],{},"编码","：SWE-bench Verified 71.3%，LiveCodeBench V6 83.1%",[38,175,176],{"id":176},"部署示例",[178,179,184],"pre",{"className":180,"code":181,"language":182,"meta":183,"style":183},"language-bash shiki shiki-themes github-light github-dark","pip install vllm\nvllm serve \"moonshotai\u002FKimi-K2-Thinking\"\n","bash","",[185,186,187,203],"code",{"__ignoreMap":183},[188,189,192,196,200],"span",{"class":190,"line":191},"line",1,[188,193,195],{"class":194},"sScJk","pip",[188,197,199],{"class":198},"sZZnC"," install",[188,201,202],{"class":198}," vllm\n",[188,204,206,209,212],{"class":190,"line":205},2,[188,207,208],{"class":194},"vllm",[188,210,211],{"class":198}," serve",[188,213,214],{"class":198}," \"moonshotai\u002FKimi-K2-Thinking\"\n",[178,216,220],{"className":217,"code":218,"language":219,"meta":183,"style":183},"language-python shiki shiki-themes github-light github-dark","# 推荐 temperature 1.0\nresponse = client.chat.completions.create(\n    model=\"moonshotai\u002FKimi-K2-Thinking\",\n    messages=messages,\n    temperature=1.0,\n    max_tokens=4096,\n)\nprint(response.choices[0].message.content)\nprint(response.choices[0].message.reasoning_content)  # 思考过程\n","python",[185,221,222,228,241,256,267,281,294,300,315],{"__ignoreMap":183},[188,223,224],{"class":190,"line":191},[188,225,227],{"class":226},"sJ8bj","# 推荐 temperature 1.0\n",[188,229,230,234,238],{"class":190,"line":205},[188,231,233],{"class":232},"sVt8B","response ",[188,235,237],{"class":236},"szBVR","=",[188,239,240],{"class":232}," client.chat.completions.create(\n",[188,242,244,248,250,253],{"class":190,"line":243},3,[188,245,247],{"class":246},"s4XuR","    model",[188,249,237],{"class":236},[188,251,252],{"class":198},"\"moonshotai\u002FKimi-K2-Thinking\"",[188,254,255],{"class":232},",\n",[188,257,259,262,264],{"class":190,"line":258},4,[188,260,261],{"class":246},"    messages",[188,263,237],{"class":236},[188,265,266],{"class":232},"messages,\n",[188,268,270,273,275,279],{"class":190,"line":269},5,[188,271,272],{"class":246},"    temperature",[188,274,237],{"class":236},[188,276,278],{"class":277},"sj4cs","1.0",[188,280,255],{"class":232},[188,282,284,287,289,292],{"class":190,"line":283},6,[188,285,286],{"class":246},"    max_tokens",[188,288,237],{"class":236},[188,290,291],{"class":277},"4096",[188,293,255],{"class":232},[188,295,297],{"class":190,"line":296},7,[188,298,299],{"class":232},")\n",[188,301,303,306,309,312],{"class":190,"line":302},8,[188,304,305],{"class":277},"print",[188,307,308],{"class":232},"(response.choices[",[188,310,311],{"class":277},"0",[188,313,314],{"class":232},"].message.content)\n",[188,316,318,320,322,324,327],{"class":190,"line":317},9,[188,319,305],{"class":277},[188,321,308],{"class":232},[188,323,311],{"class":277},[188,325,326],{"class":232},"].message.reasoning_content)  ",[188,328,329],{"class":226},"# 思考过程\n",[42,331,332],{},"推荐推理引擎：vLLM、SGLang、KTransformers。",[38,334,335],{"id":335},"与同类模型怎么选",[82,337,338,353],{},[85,339,340],{},[88,341,342,345,347,350],{},[91,343,344],{},"维度",[91,346,11],{},[91,348,349],{},"MiniMax M2",[91,351,352],{},"Qwen3-Coder",[98,354,355,369,382,394,407],{},[88,356,357,360,363,366],{},[103,358,359],{},"定位",[103,361,362],{},"开源 agent 推理",[103,364,365],{},"开源 agent\u002F编程性价比",[103,367,368],{},"开源 agentic 编码",[88,370,371,373,376,379],{},[103,372,96],{},[103,374,375],{},"1T\u002F32B",[103,377,378],{},"230B\u002F10B",[103,380,381],{},"480B\u002F35B",[88,383,384,386,388,391],{},[103,385,137],{},[103,387,140],{},[103,389,390],{},"大",[103,392,393],{},"256K（可扩 1M）",[88,395,396,398,401,404],{},[103,397,145],{},[103,399,400],{},"Modified MIT",[103,402,403],{},"MIT",[103,405,406],{},"Apache 2.0",[88,408,409,412,415,418],{},[103,410,411],{},"自部署门槛",[103,413,414],{},"极高",[103,416,417],{},"中",[103,419,420],{},"高",[42,422,423,426,427,430,431,70],{},[62,424,425],{},"建议","：要最强开源 agent 推理且有集群资源选 K2 Thinking；资源有限、追性价比看 ",[46,428,349],{"href":429},"\u002Fmodels\u002Fminimax-m2.html","；纯编码选 ",[46,432,352],{"href":433},"\u002Fmodels\u002Fqwen3-coder.html",[38,435,436],{"id":436},"避坑清单",[153,438,439,449,455,461],{},[156,440,441,444,445,70],{},[62,442,443],{},"1T 参数不是消费级硬件能跑的","：自部署需多卡集群，个人优先用官方 API 或 ",[46,446,448],{"href":447},"\u002Fcoding\u002Fapi\u002Fopenrouter.html","OpenRouter",[156,450,451,454],{},[62,452,453],{},"用 INT4 原生权重","：QAT 训练的 INT4 几乎无损，别自己粗暴量化。",[156,456,457,460],{},[62,458,459],{},"Heavy Mode 是并行策略","：HLE 51% 是 8 轨迹并行聚合的结果，单轨调用别按此预期。",[156,462,463,466],{},[62,464,465],{},"temperature 保持 1.0","：官方推荐值。",[38,468,469],{"id":469},"延伸阅读",[153,471,472,484,489],{},[156,473,474,475,477,478,477,480],{},"对比同类：",[46,476,349],{"href":429}," \u002F ",[46,479,352],{"href":433},[46,481,483],{"href":482},"\u002Fmodels\u002Fdeepseek-r1.html","DeepSeek-R1",[156,485,486,487],{},"同系列：",[46,488,49],{"href":48},[156,490,491,492],{},"本地部署：",[46,493,495],{"href":494},"\u002Fcoding\u002Flocal\u002Follama.html","Ollama",[497,498,499],"style",{},"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 .sJ8bj, html code.shiki .sJ8bj{--shiki-default:#6A737D;--shiki-dark:#6A737D}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 .sj4cs, html code.shiki .sj4cs{--shiki-default:#005CC5;--shiki-dark:#79B8FF}",{"title":183,"searchDepth":243,"depth":243,"links":501},[502,503,509,510,511,512],{"id":40,"depth":205,"text":40},{"id":53,"depth":205,"text":53,"children":504},[505,506,507,508],{"id":57,"depth":243,"text":57},{"id":73,"depth":243,"text":74},{"id":80,"depth":243,"text":80},{"id":151,"depth":243,"text":151},{"id":176,"depth":205,"text":176},{"id":335,"depth":205,"text":335},{"id":436,"depth":205,"text":436},{"id":469,"depth":205,"text":469},"reasoning",256000,"月之暗面 2025 年 11 月开源的 agent 推理模型，1T\u002F32B MoE，可连续 200-300 次工具调用，原生 INT4 提速 2x，BrowseComp 超过 GPT-5 High。","md",32000,{},true,"\u002Fmodels\u002Fkimi-k2-thinking","开源可自部署 · 官方 API 按量计费","2026-06-28",[524,525,526],"coding\u002Flocal\u002Follama","coding\u002Fapi\u002Fopenrouter","coding\u002Fcli\u002Fcline","2025-11-06",{"title":11,"description":515},"kimi-k2-thinking","models\u002Fkimi-k2-thinking",[532,533,534,535,536],"开源 agent 推理 SOTA，1T 总参 \u002F 32B 激活 MoE","可维持 200–300 次连续工具调用不退化","原生 INT4 量化（QAT），低延迟模式提速 2x","256K 上下文，Modified MIT 许可可商用","BrowseComp 60.2% 超过 GPT-5 High",[538,539,540,541],"开源 agentic 推理与工具编排","深度搜索 \u002F 多步检索 agent","需可商用开源权重的私有部署","长链路自动化任务","月之暗面","Moonshot AI",[545,546,547,548],"1T 参数自部署门槛极高，需多卡集群","纯编程 SWE-bench 71.3% 略低于闭源旗舰","max output 相对保守","中文生态外文档相对少","qC8eJ-SSZPP3SWmWu_zjplXKKAR9Q50xX5aDtpAh3bA",1782663746802]