初创公司尽调
尽调报告 Artificial Intelligence / Application Software growth 2026-06-19

Moonshot AI

变现速度快的中国 AI 挑战者,产品迭代可信,但估值逻辑仍偏紧,公开披露也偏少。

Moonshot AI 已从亮眼产品故事跨入真实商业相关性,但当前估值已经计入持续超高速增长、更清晰监管和更顺畅 IPO 路径,公开证据还无法证明这些都能兑现。

封面要素

最新轮次 01
2000 USD M [CO023]
报道估值 02
20000 USD M [CO023]
ARR(2026 年 4 月) 03
200 USD M [CV004]
2025 年 Series C 后现金 04
1400 USD M [CO021]
成立时间 05
2023 [CO006]
主要投资方基础 06
Alibaba-led with Tencent, Meituan, and others in later rounds [CO020, CO024, CO026]

公司概况

Moonshot AI 是一家以北京为中心的中国 AI 公司,由杨植麟、周鑫宇、吴雨欣于 2023 年创立。公开产品栈已覆盖 Kimi 消费者助手、Kimi Code、开发者 API 平台,以及快速推进的开放权重 / 开源模型路线图,包括 K1.5、K2、K2.5、K2.6 和 K2.7 Code。商业模式把会员和其他 Kimi 付费访问,与按量计费的 API 结合起来;2026 年报道显示 ARR 已超过 $200 million,融资估值据称达到 $20 billion。投资逻辑有真实产品速度和商业化支撑,但治理、单位经济、客户集中度和上市准备细节的公开披露仍限制判断。

官网
www.moonshot.cn
创始人
Yang Zhilin, Zhou Xinyu, Wu Yuxin
创立地点
Beijing, China
总部
Beijing, China
产品
Kimi 是一套多入口 AI 产品,覆盖聊天、深度研究、代码、网站、幻灯片、表格和 Agent 工作流。Moonshot 也销售兼容 OpenAI 的开发者平台,提供按 token 计价、工具调用和长上下文前沿模型,同时继续发布 Mooncake 等新模型家族和基础设施。
客户
面向消费者、专业人士、开发者,以及早期团队或企业用户,服务知识工作、编码、研究和 Agentic 生产力工作流。
商业模式
消费端与开发者端混合变现:前端依靠会员和 Kimi 付费访问,平台侧依靠按量 API 计费、工具调用费和批处理定价。
阶段
growth
融资情况
2026 年 5 月据报以 $20B 估值融资约 $2B,此前在 2025 年末完成 $500M Series C;公开报道也指向充足现金储备和活跃的香港上市准备。
[CO006, CO010, CO011, CO023, CV004, CU033]

执行摘要

主要优势

  • 消费、编程、开发者和研究工作流里的产品推进速度异常快,模型迭代已推进到 K2.7 Code。
  • Moonshot 已有可见商业化界面,包括公开 API 定价、付费 Kimi 入口,以及据报道超过 $200M 的 ARR。
  • 公司多次拿到大额融资,降低了训练和商业化近期缺钱的风险。
  • Kimi 的长上下文和智能体定位,让 Moonshot 不只是通用聊天,而有差异化工作流故事。
  • GitHub 和 Hugging Face 上的开放权重与外部开发者分发,扩大了技术触达和生态相关性。

主要风险

  • 据报道 $20B 估值相对于有限公开 ARR 证据倍数很高,安全边际很薄。
  • 毛利率、客户集中度、股权条款、债务和现金跑道等公开披露仍然薄弱。
  • 中国 AI 监管、隐私审查和跨境实体结构问题,带来实质法律与合规风险。
  • DeepSeek、Doubao、Qwen、Z.ai、MiniMax 和 Baidu 的竞争,可能迅速压低价格并压缩差异化。
  • 运营叙事仍高度围绕创始人,杨植麟构成重要关键人依赖。
  • 香港上市若延后或弱于预期,可能刺破稀缺性溢价并重估估值。

未决问题

  • 经审计 ARR 桥接、收入确认政策和客户留存细节仍未披露。
  • 公开资料未披露股权条款、清算优先权或投资人权利包袱。
  • 毛利率、算力承诺、月度烧钱速度或现金跑道没有扎实公开信息。
  • 具名企业客户证明和集中度数据,相比消费端与开发者信号仍偏薄。
  • 北京与新加坡之间的具体治理、董事会和运营实体图谱仍不完整。

目录

Chapter 01

01公司概览

1.1 身份、创立、总部与商业模式

Moonshot AI 的官方网站呈现出一种双重姿态:既想像研究实验室,也想像产品公司。Moonshot 主页用面向 AGI 的语言讲“把能量转化为智能”,关联的 Kimi 界面则强调代码、深度研究、网站、幻灯片和电子表格等实际用户工作流。这种组合很关键,因为它是主来源能给出的最清晰商业模式概括:Moonshot 不是靠单一用途应用变现,而是通过消费者助手层和开发者平台,把基础模型能力商业化。平台定价材料记录了基于 token 的 API 计费,也进一步支持这一判断;Kimi Code 则明确把自己描述为与会员绑定的编码入口。 创立和地点记录方向清楚,但口径并不完全统一。TechCrunch 报道称,杨植麟在 2023 年与周鑫宇、吴雨欣共同创立 Moonshot AI,多家 2026 年新闻媒体也把公司描述为总部在北京。同时,Kimi 现行服务条款把新加坡的 Moonshot AI PTE. LTD. 列为服务提供方。因此,最有证据支撑的结论是:Moonshot 是一家以北京为中心运营、服务侧至少有一个新加坡法律实体的公司,而不是一个总部披露干净单一的主体。后续章节应保留这一细微差别,不要压平成单一司法辖区叙事。[CO001, CO002, CO003, CO004, CO005, CO006]

Moonshot AI 快照 KPI 表
指标数值 / 状态日期置信度缺口
公司名称 / 品牌Moonshot AI 运营 Kimi,作为面向公众的助手品牌2026-06-19
成立时间TechCrunch 称为 2023 年2023-01-01未审阅到带有确切成立日期的主要注册文件
总部证据最充分的运营基地是北京;官方条款也标明一家新加坡服务实体2026-06-19官方网站没有发布统一的总部页面
阶段后期私营基础模型公司,至 2026 年已完成多轮大额融资2026-05-07各报道对确切轮次分类不一致
最新估值TechCrunch 和 Forbes 报道为 $20B2026-05-07未审阅到公司确认的投后估值文件
ARR报道称 2026 年 4 月超过 $200M2026-05-07没有经审计收入报表或 cohort 细节
现金Series C 信函后,据报现金超过 RMB 10B2026-01-01需要确认资金存量和烧钱曲线
客户 / 员工数2026-06-19审阅来源中未找到公开客户数量或员工人数

空值反映公开指标缺乏支持,而不是数值为零。

[CO001, CO002, CO006, CO010, CO023, CO028]
FO002: 公司快照逻辑

Moonshot AI 把研究身份、Kimi 产品入口、按用量计费的开发者平台、深厚融资和少数集中依赖连在一起。

该流程是概念性的,不是交易流;它概括公司产品、资本和风险层之间的关系。

[CO003, CO004, CO011, CO018, CO023, CO039]

1.2 领导层可见度、治理与阶段

公开领导层披露高度集中在创始人杨植麟身上。第三方报道反复把他置于创始人和公众面孔的位置,最有实质内容的个人资料也聚焦他的学术和研究履历,包括曾在 Meta AI 和 Google Brain 工作。SCMP 还提供了一个有用的战略框架:引用杨植麟希望结合 OpenAI 式技术理想主义和 ByteDance 式商业纪律的愿景。公开记录没有提供的内容几乎同样重要:主公司界面没有经核验的董事会名单,没有投资者关系页面,也没有清楚披露的财务负责人或更广泛高管梯队。仅靠公开材料评估继任深度、委员会结构或财务控制时,这些缺口会显著增加治理尽调难度。 阶段判断比治理判断更容易。到 2026 年,Moonshot AI 显然已不是种子期实验,也不是前产品实验室。公司已完成多轮后期融资,持续推广反复出现的产品入口,媒体也已围绕估值、ARR 和 IPO 时间表来报道。最稳妥的标签是:处于商业化模式的后期私有基础模型公司。这个判断足以支撑下游分析,但还不足以无保留地统一具体融资轮次分类,因为公开记录在不同时间点同时出现 Series B、追加融资和 Series C 等表述。[CO014, CO015, CO016, CO017, CO018, CO022]

领导层和创始人表
人物 / 职能角色证据创始人-市场匹配或覆盖缺口 / 依赖
Yang Zhilin创始人和公众代表TechCrunch 和 SCMP 报道中点名强研究履历和战略信号,适合前沿模型公司公开记录中关键人物集中度高
Zhou Xinyu联合创始人TechCrunch 创始报道中点名创始团队延续性信号当前公开运营能见度有限
Wu Yuxin联合创始人TechCrunch 创始报道中点名创始团队延续性信号当前公开运营能见度有限
董事会 / 财务领导层审阅到的官方材料未公开呈现未找到董事会名单或 CFO 级页面治理缺口,不等于证明缺席公开治理分析的重大尽调阻碍

这些行混合了已确认个人和明确治理覆盖缺口,因为公开领导层披露很薄。

[CO014, CO015, CO016, CO017, CO018]

1.3 融资历史、估值路径与封面指标

融资故事的核心是快速抬升。TechCrunch 报道,2024 年初公司估值为 $2.5 billion,并完成超过 $1 billion 的 Series B 融资。到 2026 年初,公开报道已进入更大尺度:Caixin 报道 $500 million Series C 和超过人民币 100 亿元现金;TechCrunch 与 Forbes 均报道称,公司以 $20 billion 估值融资约 $2 billion。TechCrunch 还称,Moonshot 在此前六个月融资 $3.9 billion,最新融资中有 Long-Z Investments、Tsinghua Capital、China Mobile 和 CPE Yuanfeng,更广泛投资方包括 Alibaba、Tencent、HongShan、ZhenFund、IDG 和 5Y。这足以判断 Moonshot 已站在中国私有 AI 融资的第一梯队附近。 指标图景仍不完整。2026 年 4 月报道的 ARR 超过 $200 million 很有意义,公司的现金余额和 OpenRouter 使用排名也指向真实商业牵引。但经审阅材料没有给出经审计收入、客户数量、员工规模,也没有调和后的累计融资金额。这些缺口很重要,因为它们限制了对运营效率、稀释和持续性的清晰解读。本章的正确做法是陈述有支撑的数字,把无支撑数字保留为明确缺口,避免把投资者热情转成虚假的精确性。[CO019, CO020, CO021, CO022, CO023, CO024]

利益相关方或投资者地图
利益相关方角色经济或战略重要性公开信号尽调要求
Alibaba多轮 Moonshot 重大融资报道中的投资者具备战略云和生态意义,也提供大市值财务背书TechCrunch 和 Forbes 将其列为支持方确认支票规模、商业绑定和治理权利
HongShan早期主要投资者2024 年机构信心的重要证明TechCrunch 和 SCMP 提到 HongShan 参与厘清后续跟投和稀释影响
Meituan / Long-Z Investments2026 年领投方组合最新轮融资显示国内重量级科技赞助TechCrunch 和 Forbes 在 2026 年融资中引用 Long-Z确认领投是否附带商业分发权
China Mobile2026 年投资参与方可能具备基础设施或企业分发意义TechCrunch 和 Forbes 列出 China Mobile 参与评估关系是财务性质还是战略性质
Tencent / 5Y / ZhenFund / IDG更广泛投资者组合增加财务深度,也释放生态支持信号TechCrunch 和 Forbes 在多轮融资中引用这些名称索要股权结构和董事会权利映射
Tsinghua Capital / CPE Yuanfeng2026 年轮次参与方增加国内机构和国资相关融资意义TechCrunch 2026 年融资报道中点名厘清所有权集中度和治理条款

经济重要性根据报道中的轮次角色推断,因为持股比例和优先权栈未公开。

[CO020, CO024, CO026, CO027]
FO003: 快照 KPI

当前公开 KPI 显示估值和 ARR 动能,但多项尽调关键指标仍未解决。

缺口项表示指标缺乏支持,并不代表数值为零。

[CO021, CO023, CO028, CO029, CO044]

1.4 里程碑、不利事件与后续章节应继承的内容

对一家私有 AI 公司来说,Moonshot AI 的公开里程碑曲线异常清晰。TechCrunch 将基础模型发布放在 2023 年 3 月,将首个 Kimi 聊天机器人发布放在 2023 年 10 月。CNBC 随后记录了后续节奏:2025 年 7 月 Kimi K2、2025 年 11 月 K2 Thinking、2026 年 1 月 K2.5。Moonshot 自有主页又把 WorldVQA、Agent Swarm 和 K2.6 加入 2026 年研究时间线;平台博客则显示,商业化栈自 2024 年起已通过上下文缓存和企业 API 工作持续拓宽。合在一起,这些里程碑支撑了后续章节的一个关键分析点:Moonshot 不只是发布模型,也在把模型包装成更完整的产品和开发者生态。 记录中的负面一侧也足够真实,应随本章进入后续分析。CNBC 报道了 Anthropic 的指控:Moonshot 参与了针对 Claude 的大规模蒸馏行动;Anthropic 自身声明则给出关于数百万次交互和定向能力抽取的详细指称。另据 OECD AI Incidents Monitor 记录,2026 年发生了一起涉及 Kimi 简历泄露的隐私事件。单看任一事件,都不能直接确定 Moonshot 内部控制的法律或技术事实;但两者合在一起,已经构成围绕来源、隐私和治理的非轻微风险。后续章节因此应把本概览作为身份、规模方向和里程碑节奏的事实底座,同时继续把治理深度、客户广度、员工规模和准确融资历史视为未解决尽调项。[CO027, CO030, CO032, CO033, CO034, CO035]

里程碑表
日期事件类型金额 / 状态参与方含义
2023-03-01Moonshot 发布 100B 参数模型产品已报道发布Moonshot AI审阅到的最早技术产品里程碑
2023-10-01Kimi 聊天机器人发布产品声称支持 200k 中文字符上下文Moonshot AI建立早期长上下文消费端定位
2024-02-21Series B 报道估值 $2.5B融资报道超过 $1BAlibaba、HongShan、Meituan、Xiaohongshu 等让公司进入中国 AI 顶级融资讨论
2024-07-01Context Caching 公开测试版产品平台里程碑Kimi Open Platform 平台显示开发者平台建设已超出基础聊天
2024-08-07企业 API 正式发布产品平台里程碑Kimi Open Platform 平台标志企业导向商业化层成形
2025-07-14Kimi K2 发布产品开源模型发布Moonshot AI让 Moonshot 进入全球开放模型竞争
2025-11-06Kimi K2 Thinking 发布产品四个月内第二次重大 K2 更新Moonshot AI释放智能体和推理能力快速迭代信号
2026-01-01Series C 和不急于 IPO 的立场被报道融资报道为 $500M 和 >RMB 10B 现金Moonshot AI 领导层增加跑道,同时优先芯片、K3 和商业化
2026-01-28K2.5 在中国 AI 发布潮中亮相产品声称视频生成和智能体升级Moonshot AI显示一线模型节奏仍在延续
2026-02-03WorldVQA 被列为最新研究产品研究发布Moonshot AI增加多模态评测重点
2026-02-09Agent Swarm 被列为最新研究产品研究发布Moonshot AI增加多智能体编排叙事
2026-02-24Anthropic 蒸馏指控公开反向指控已发布Anthropic、Moonshot AI引入来源和合规风险
2026-04-20Kimi K2.6 被列为最新研究产品研究 / 产品里程碑Moonshot AI公司进入当前旗舰代际
2026-04-21OECD AIM 记录 Kimi 简历泄露事件反向隐私事件记录Kimi 用户 / OECD AIM凸显生产环境隐私控制风险
2026-05-07最新巨额融资报道估值 $20B融资报道为 $2BLong-Z、China Mobile、Tsinghua Capital、CPE Yuanfeng 等将公司重新定位为中国资金最充足的 AI 实验室之一

早期产品发布时间采用来源报道中审阅到的首个公开日期,因为未审阅到一手发布日文件。

[CO019, CO021, CO023, CO032, CO033, CO034]
FO001: 公司里程碑时间线

Moonshot AI 的公开时间线显示,公司从长上下文创业公司转向融资充足的多产品模型平台,2026 年也开始出现重要反向事件。

部分里程碑日期采用首次查阅到的发布日期,而不是公司内部正式发布日时间戳。

[CO019, CO023, CO032, CO033, CO034, CO035]
Chapter 02

02市场分析

2.1 市场边界、纳入支出与现状替代品

分析 Moonshot,应把它放在面向中国的 AI 助手和基础模型应用层,而不是把它当作中国全部 AI 支出的代理。证据显示,Kimi 活跃在公众用户或企业团队购买模型支撑型知识工作的场景:聊天、研究、文档工作流、电子表格辅助、编码和 Agent 任务。Kimi 的 API 文档又通过文件、批处理、工具调用、JSON 模式和网页搜索,把这个边界延伸到开发者工作流。IDC 的中国市场概览也支持同一边界:它把模型和 Agent 平台,与底层芯片、计算、存储和网络栈分开。因此,纳入支出是与模型使用绑定的应用和平台支出;排除支出包括原始半导体、无差异云 IaaS,以及整个中国软件市场。 这个边界重要,因为真实替代品不只是有名有姓的聊天机器人竞品。对中国知识工作者来说,现状仍可能是搜索、文档、电子表格、办公软件或手工分析。对开发者来说,替代品可能是 DeepSeek、Qwen 等兼容 OpenAI 的本土同行,也可能是基于迁移友好 API 自建的内部栈。对企业买家来说,替代品可能是治理更强或母公司分发更强的本地化平台。若把 Moonshot 视为覆盖所有 AI 预算,就会掩盖 Kimi 今天可证明销售内容的窄度,也会掩盖工作流层面的替代强度。[CM001, CM002, CM003, CM004, CM005, CM006]

市场定义表
细分 / 类别纳入支出排除支出买方 / 付款方与 Kimi 的关系
消费者 AI 助手工作流聊天、搜索、摘要、文档分析、幻灯片、电子表格辅助、深度研究不使用模型的通用办公套件、纯人工咨询,以及非 AI 生产力支出直接付费的个人知识工作者或小团队Kimi 网页端和应用的核心触面
开发者 / API 基础模型支出token 计量、工具调用、网页搜索增强、基于文件的问答、批量推理、智能体工作流原始 GPU 采购、通用云托管,以及无关开发者工具开发者、产品团队、平台预算Kimi API 的核心变现轨道
企业本地化知识工作智能体受治理的本地化助手、研究智能体和编码 copilot 推出所有与模型支撑工作流无关的企业软件预算CTO、CIO、平台、创新或业务单元预算通向更大合同的可能路径
中国公共生成式 AI 合规支出与模型部署绑定的日志、标识、治理和本地化成本不含 AI 部署的独立网络安全或法律合规支出平台所有者和受监管运营方重要部署约束和采购标准
被排除的基础设施层None半导体、云 IaaS、存储、网络和通用算力容量基础设施和采购负责人在 Moonshot 直接变现的产品范围之外

边界行把 Kimi 可观察的助手和 API 触面,与更广泛的 AI 基础设施或通用软件预算分开。

[CM001, CM002, CM003, CM005, CM006, CM033]

2.2 受证据约束的规模测算视角与 Kimi 当前位置

公开来源没有给出 Moonshot 所服务的“中国助手 + 基础模型”精确类别的、非付费墙且口径干净的人民币市场规模,因此更可辩护的做法是保留多个规模测算视角。第一是被监测需求。AICPB 2026 年 4 月中国 AI 排名显示,头部产品拥有很大的公众注意力池,并将 Kimi 排在中国 AI 网站第五,访问量 43.69 million;在中国 AI 应用中排第八,MAU 25.33 million。这些数字有意义,但也说明 Kimi 目前不是消费者端领导者:在被监测的公众需求代理指标上,DeepSeek 和 Doubao 要大得多。第二个视角是企业采用。IDC 称,到 2027 年,80% 的中国 C1000 企业将优先考虑 AI 主权,这会扩大国内模型供应商的可服务池,即使精确类别收入仍未披露。第三个视角是变现轨道。Kimi、DeepSeek、Qwen、ERNIE、OpenAI、Anthropic 和 Google 的公开定价页显示,标价阶梯很宽,从低于人民币级别的本土价格地板,到高端美国 API 定价都有。 合并来看,这些视角支撑一个大且增长中的市场,但变现深度仍不确定。IDC 的中国 AI 编码展望明确提到增长很快、月度定价较低,且头部供应商 2025 年收入仍低于人民币 1 亿元。这正是 Moonshot 的核心分析张力:Kimi 已有足够分发来成为重要玩家,但公开证据显示,类别使用量扩张快于标准化利润池。Kimi 当前位置更像一个有真实规模的可信挑战者,而不是一个已证明市场捕获能力的类别所有者。[CM010, CM011, CM012, CM013, CM014, CM015]

TAM / SAM / SOM 或规模测算视角表
发布方年份地域数值CAGR / 采用方法置信度局限
AICPB2026中国所列中国前 10 大 AI 网站合计每月 1.089B 次访问n/a观察到的中国头部 AI 产品网站流量池注意力不等于收入,且排除线下或仅企业端使用
AICPB2026中国所列中国头部 AI 应用合计 1.199B MAUn/a观察到的中国头部 AI 产品应用 MAU 池App MAU 和网站访问量是不同单位,不应合并成一个 TAM
IDC FutureScape 摘录2027中国 C1000 企业80% 将任务关键型 AI 用途中的 AI 主权列为优先事项预测至 2027 年企业采用 / 采购视角衡量企业行为,不是助手收入
IDC FutureScape 摘录2027 / 2030中国智能设备和 AI 智能体渗透率目标为 70% / 90%预测至 2030 年宏观政策和设备渗透视角国家目标宽于 Kimi 当前变现范围
IDC 编码市场摘录2025中国 AI 编码细分市场头部厂商实现 100% 增长,但总收入仍低于 RMB 100m100% 增长细分市场变现视角编码只是市场的一部分
Kimi API Platform 平台2026中国 / 全球开发者K2.6 每 1M tokens 输入 ¥6.50、输出 ¥27n/a一方标价变现轨道标价不等于实际净收入或企业合同组合
OpenAI、Anthropic、Google、Alibaba 与 DeepSeek2026全球或可从中国访问的开发者已发布同业 token 价格从人民币本地低价到美国 API 高价不等n/a可比定价区间视角跨厂商定价不等于份额或使用组合

由于公开来源没有拆出 Kimi 精确类别的非付费墙人民币 TAM,本章保留需求、企业采用和定价视角,而不是编造一个标题式数字。

[CM010, CM018, CM022, CM023, CM024, CM025]
Kimi 当前市场位置表
指标Kimi 数值对比对象对比对象数值Kimi 份额 / 比例含义
中国 AI 网站排名(2026 年 4 月)第 5;43.69M 次访问DeepSeek第 1;486.50M 次访问DeepSeek 访问量的 9.0%Kimi 有可见度,但不是流量领头羊
中国 AI App 排名(2026 年 4 月)第 8;25.33M MAUDoubao第 1;336.04M MAUDoubao MAU 的 7.5%消费端装机基础差距明显
中国 AI 网站池份额前 10 总访问量 1.089B 中的 43.69M中国 AI 前 10 样本池1.089B 次访问4.01%监测到的网页关注度里,Kimi 有分量但不占主导
中国 AI App 池份额上榜 MAU 1.199B 中的 25.33M上榜中国 AI App 样本池1.199B MAU2.11%Kimi 的 App 存在感落后于中国头部消费端 App
全球聊天机器人网站排名(2026 年 4 月)第 9;43.69M 次访问ChatGPT第 1;5.69B 次访问ChatGPT 访问量的 0.8%外部基准差距仍大
相对 Doubao 网站的位置43.69M 次访问Doubao 网站162.89M 次访问26.8%Kimi 必须靠工作流深度取胜,而不是拼原始触达

本表只把 AICPB 监测的网站访问量和 App MAU 当作当前位置的代理指标;它们不等同于收入、留存或企业支出。

[CM011, CM012, CM013, CM014, CM015, CM016]
FM001: 市场规模观察框架

从中国 AI 普及的大背景,到 Kimi 目前可合理变现的较窄公开需求池和工作流池,采用层层收窄的观察口径。

该金字塔有意混合普及、流量和工作流层,因为公开来源没有披露 Kimi 精确品类的独立人民币 TAM。

[CM010, CM011, CM012, CM032, CM033, CM037]
FM002: 中国可访问模型的公开价格带

按每 1M token 人民币计价的低 / 中 / 高挂牌价格带;这里把它当作市场收入约束,而不是市场规模估计。

Baidu ERNIE 数值由每 1K token 价格换算为每 1M token 等价值;图中比较的是公开挂牌区间,不是实际净合同价。

[CM018, CM019, CM022, CM027]

2.3 买家、用户、付款方与采用路径

Kimi 的买方图谱随产品入口而变化。消费端,用户和付款方可以是同一个人:学生、研究人员,或探索搜索、文档、幻灯片、电子表格辅助的一般知识工作者。开发者端,用户可能是工程师或 Agent 构建者;随着 API 使用增长,付款方会从个人卡转向团队或产品预算。企业端,最终用户仍可能是分析师、运营人员或开发者,但实际预算所有者通常落在 CTO、CIO、平台或创新职能,因为采购开始包含治理、吞吐量和推广要求,而不只是模型质量。 采用路径也相应是多步的。公开定价和工具表明,低摩擦入口包括自助使用、编码试用或 API 实验。OpenAI 迁移文档降低初始切换成本,因为团队可以保留熟悉的 SDK 模式。但从试用进入持续支出仍需要工作流验证;从团队使用进入有治理的部署,还要通过合规、日志和本地化的第二道关口。因此,Kimi 宽产品面有助于需求生成,却不能消除后期摩擦。公开证据支持一个真实的买家—用户—付款方阶梯,但也说明收入质量取决于 Moonshot 能否从实验推进到嵌入团队流程的重复性工作。[CM018, CM019, CM020, CM021, CM034, CM035]

细分市场 / 买方地图
细分市场买方用户付款方工作流预算归属采用触发点
个人知识工作者本人本人本人搜索、摘要、文档、幻灯片、电子表格个人或可自由支配支出日常任务立刻提效
学生和研究人员本人或机构本人本人或实验室 / 学校长上下文阅读、笔记综合、研究支持个人、实验室或教育预算需要处理大量中文文本
开发者和智能体构建者开发负责人或产品负责人开发者信用卡、团队或产品预算API 调用、编码、工具调用、文件问答、网页搜索工程或 AI 产品预算API 基准能跑通,token 成本可接受
企业知识工作团队创新负责人、CTO 办公室或业务负责人分析师、运营人员、开发者部门或平台预算本地化助手、编码辅助、内部研究智能体CIO / CTO / 平台 / 业务单元负责人需要受治理的本地部署和中文适配
受监管或主权敏感买方安全、合规和平台负责人内部员工或获批承包商中央 IT / 合规预算可审计的模型使用、标注、日志记录和本地托管选择安全、合规或公共部门预算海外服务缺位或本地化要求

同一底层模型一旦分别作为自助助手、API 或受治理的企业工具使用,面对的买方和付款方会很不一样。

[CM002, CM003, CM021, CM034, CM035]
FM003: 预算负责人和扩张摩擦图

Kimi 在中国主要需求入口上的预算归属和扩张摩擦各不相同。

[CM021, CM034, CM035, CM036, CM037]
FM004: 采用漏斗或价值链图

中国 Kimi 用户或买家从发现到受治理部署的实际路径。

该流程抽象出从自助试用到受治理部署的观察购买路径;部分消费用户可能永远不会进入团队或治理阶段。

[CM020, CM021, CM034, CM036, CM037]

2.4 增长驱动、价格压缩与监管

三个驱动最突出。第一,OpenAI 缺席中国大陆,加上 IDC 强调 AI 主权,为国内替代品创造结构性窗口。第二,中国供应商在更快发布、多模态和生态整合上推进,持续刺激用户试用本土产品。第三,透明的公开价格阶梯让试验成本足够低,能拓宽漏斗顶部。但同一组事实也带来核心约束。DeepSeek 定价显示,本土价格地板可能远低于 Kimi 标价;Qwen、ERNIE 和其他平台也提供各自兼容 OpenAI 或多模态的替代方案。在客户锁定尚未深入前,这会压缩付费意愿。IDC 的编码摘录也表达了同一点:采用增长可能快于已实现收入。 监管又加了一层市场结构。中国 2023 年生成式 AI 办法要求提供者为公共服务管理内容、隐私和安全责任。2025 年标识规则进一步要求对生成内容和元数据添加显式和隐式标识。相对不支持中国大陆服务的外国服务商,这些规则可以帮助国内供应商,但不是免费护城河。它们抬高合规成本,要求运营控制,并让企业级治理更重要。最后,Caixin 关于芯片限制的报道显示,计算供给仍是现实约束。因此,Moonshot 所在市场确有需求顺风,但也有集中的价格压力、政策成本和资本强度,限制使用量转化为持久利润率的速度。[CM007, CM008, CM009, CM022, CM023, CM024]

增长驱动与约束表
驱动 / 约束方向时点含义尽调问题
OpenAI 不支持中国大陆利好国内厂商当前为 Kimi 等本土替代品打开结构性空间Kimi 需求里,有多少来自海外服务不可用,又有多少来自产品偏好?
AI 主权和区域伙伴偏好正向2026-2027扩大国内平台可触达的企业客户池Moonshot 公开拿下了哪些受监管客户?
公开 token 价格低、API 透明利好试用,不利利润率当前放大试验空间,但挤压变现折扣和赠额之后,Moonshot 实际混合价格是多少?
兼容 OpenAI 的迁移路径利好采用当前降低开发者迁入 Kimi 的切换成本试用流量有多少转成持续生产流量?
2025 年中国 AI 标识规则多空交织2025 年起可能利好具备本地合规能力的厂商,但也增加治理开销标识和元数据合规的实际成本有多高?
高端芯片和算力约束负向当前相比供应更充足的对手,可能限制模型迭代速度和利润率Moonshot 已锁定哪些算力和托管合作?
ByteDance、Alibaba、Tencent、Baidu 的母平台生态不利独立厂商当前Kimi 面对的是原生分发更强的厂商没有超级 App 母平台,Kimi 能否持续获客?
公开排名低于 DeepSeek 和 Doubao负向 / 现实校验当前Kimi 可信,但还不是当前品类第一Kimi 最强的场景在哪里:编码、研究,还是企业本地化?

本地可用、开放 API、低成本,这些市场事实一边创造增长,一边也加剧价格压力和分发风险。

[CM007, CM009, CM010, CM021, CM022, CM028]
Chapter 03

03竞争对手

3.1 格局结构:直接同行、平台巨头与外部参照

Kimi 不是在一个整齐的单供应商市场竞争。最强的国内同行分成三类且相互重叠。第一类是 DeepSeek、Z.ai/GLM、Qwen、MiniMax 和 Baidu ERNIE 等直接前沿模型对手,争夺 API、编码和 Agent 工作负载。第二类是 ByteDance、Alibaba、Baidu 和 Tencent 等母平台巨头,可以把模型与既有消费者、商业、搜索、支付或生产力分发结合起来。第三类是外部参照——OpenAI、Anthropic 和 Google——即便它们不是中国大陆默认合法选择,买家仍会拿它们来对标质量、包装和价格。 这种结构重要,因为 Kimi 的强弱取决于比较轴。相对高价美国参照,Kimi 看起来成本有效且本地化。相对 DeepSeek,它可能显得昂贵。相对 Doubao,它在大众市场触达上可能更窄。相对 Baidu 或 Alibaba,它在公开企业控制和分发上可能更薄。因此,严肃的竞争章节必须同时覆盖直接模型同行,以及那些能比独立实验室更激进地补贴或分发 AI 的相邻生态。[CP001, CP005, CP008, CP011, CP013, CP015]

竞争对手画像表
竞争对手类别规模 / 背书目标客群差异化关键限制
Kimi / Moonshot国内直接挑战者Moonshot 2024 年据报以 $2.5B 估值融资 >$1B知识工作者、开发者、企业团队长上下文知识工作、Kimi Code、从 OpenAI 风格技术栈迁移容易公开生态和企业治理证据链落后于大型平台对手
DeepSeek低成本前沿模型和开放平台对手中国出身、能见度高,覆盖聊天、App、API 和开源模型家族开发者、自托管方、成本敏感买方官方定价很低、1M 上下文、兼容 OpenAI / Anthropic母平台分发弱于 ByteDance、Alibaba 或 Baidu
Z.ai / GLM 阵营编码智能体和长周期任务对手Zhipu 支持的技术栈,覆盖 GLM 模型、智能体和移动自动化开发者、智能体构建者、企业1M 上下文 GLM-5.2、长周期智能体定位、免费层和 Claude Code 兼容本次抓取材料里,公开定价清晰度弱于 DeepSeek 或 Kimi
ByteDance Doubao / Seedance消费端助手和多模态平台对手背靠 ByteDance 商业和内容场景大众用户、创作者、开发者文本、代码、视频、图像和语音产品线广,App 分发强公开定价细节少于 Kimi 或 DeepSeek
Baidu ERNIE / Qianfan企业智能体和搜索联动对手背靠 Baidu 搜索和云平台企业构建者、开发者、知识工作流明确的智能体平台、可观测性、审计 / 合规姿态、搜索和百科工具本次材料里,消费端 App 差异化不如 Doubao 或 Qwen 明显
Alibaba Qwen / Model Studio基础模型和商业平台对手背靠 Alibaba 云和商业生态开发者、企业构建者、Alibaba 生态用户兼容 OpenAI 的 Qwen API、第三方模型、低端到旗舰的价格梯度公开抓取材料呈现的直接消费端助手细节,少于 Qwen App 营销材料可能提供的内容
MiniMax国内多模态挑战者中国独立 AI 创业公司,拥有创作者和消费端产品开发者、创作者、消费者M3 编码 + 1M 上下文 + Hailuo + Talkie + token 方案本次材料里,直接官方定价不如 Kimi 或 DeepSeek 透明
OpenAI / ChatGPT外部参照的高端基准全球流量龙头和高端 API 基准全球开发者、企业、消费者全球规模、高端定价、设定基准的角色不是中国大陆默认合法公开选项,也不是本地低价地板

各行综合官方产品页、定价文档和排名来源;比较的是公开定位,而不是经审计的财务深度。

[CP001, CP005, CP008, CP011, CP013, CP015]
FP001: 竞争定位图

X 轴代表企业 / API 就绪度和治理可见度。Y 轴代表消费者触达和分发力。分数是有证据支撑的序数判断,不是市场份额测量。

序数分数综合公开产品广度、定价、分发证据和已抓取来源中的治理可见度。目的在于展示相对定位,不是精确市场份额。

[CP019, CP020, CP023, CP024, CP031, CP032]

3.2 主要竞争者的产品与能力宽度

公开产品页显示,多数严肃 Kimi 竞争者如今卖的远不止单一文本模型。DeepSeek 结合聊天、应用、API 和开源研究资产。Z.ai 强调长周期 Agent、移动自动化、多模态和编码发布。Doubao 的母公司栈横跨文本、代码、视频、图像、语音和实时语音。Baidu Qianfan 进一步延伸到企业 Agent 开发、RAG、搜索、百科工具和可观测性。Qwen 通过 Model Studio 同时运行自有模型和第三方访问。MiniMax 把代码与 Hailuo 视频、音频和 Talkie 结合起来。放在这个竞争场中,Kimi 自身宽度可信:助手工作流、深度研究、电子表格、幻灯片、编码和 API 工具,都有公开证据。 结果是,能力宽度本身不足以成为 Kimi 的护城河,而是入场券。买家完全可能用多个竞品栈解决类似工作。不同对手之间变化的是编码强度、消费者分发、多模态创作、治理和兼容性的平衡。因此,公开证据薄弱的单元格必须保持未知,比较也应聚焦最影响决策的能力,而不是模糊的“AI 领导力”宣称。[CP001, CP002, CP005, CP006, CP008, CP009]

功能 / 能力矩阵
公司文本 / 聊天 API编码 / 智能体视频或丰富多模态迁移友好兼容性企业治理证据消费端 App / 分发证据缺口
Kimi公开强调度强部分具备 / 多模态,但创作者视频广度弱于 Doubao 或 MiniMax借 OpenAI 迁移指南体现较强本次抓取材料里部分具备有分量,但公开触达不到第一梯队需要更多公开企业管理细节
DeepSeek借智能体工具支持体现较强本次抓取材料里有限借 OpenAI + Anthropic 格式体现较强本次抓取材料里未知需要更多公开企业合规细节
Z.ai / GLM 阵营强,且覆盖长周期任务借多模态和移动自动化发布体现为是借 SDK 和 Claude Code 兼容体现较强本次抓取材料里部分具备定价和治理细节仍偏薄
Doubao / Seedance是 / 代码模型视频、图像和语音都很强本次抓取材料里未知本次抓取材料里未知国内很强直接公开定价细节稀疏
Baidu Qianfan / ERNIE是 / 智能体和工具工作流借平台 API 和第三方模型接入部分具备强且明确借更广的 Baidu 生态达到中等需要单独补充 Wenxiaoyan 消费端细节,才能完整比较 App
Qwen / Model Studio是 / 文本、图像、视频借兼容 OpenAI 的 API 体现较强本次抓取材料里部分具备借 Alibaba 场景达到中等到较强需要专门的 Qwen App 页面做消费端比较
MiniMax借 M3 和工具集成体现较强借 Hailuo、音频、Talkie 体现较强借 Anthropic 风格 token 方案文档体现较强本次抓取材料里部分具备直接官方定价页在这里仍偏薄
OpenAIn/a公开包装里很强全球很强中国大陆可用性不在默认公开支持列表内

缺乏支持的单元格标为部分具备或未知,而不是猜测。本表只记录抓取来源集能直接证明的内容。

[CP001, CP002, CP005, CP006, CP008, CP009]
FP002: 功能广度 / 能力图

按本章最重要的购买标准比较相对强弱。未知表示抓取来源中尚未证实,不代表现实中不存在。

强度标签是有序判断,并有证据支撑;它们概括已抓取的来源集合,不声称覆盖全部能力。

[CP021, CP022, CP030, CP031, CP032, CP035]

3.3 定价、包装与分发力量

对 Kimi 最清晰的公开经济威胁不是 OpenAI 的高端定价,而是本土价格压缩。DeepSeek 官方定价显著低于 Kimi。Qwen 和 Baidu 也发布了宽模型阶梯,OpenAI、Anthropic 和 Google 则建立高端外部参照上限。Kimi 仍比许多美国参照更便宜,但如果本土竞争者更便宜,或生态所有者能补贴采用,仅靠这一点不足以保证份额。公开包装证据也不同。Anthropic 和 Baidu 明确突出合规或审计能力。MiniMax 和 Z.ai 突出编码工具集成。Alibaba 和 ByteDance 则把模型与更广生态和用户漏斗绑定。 因此,分发力量是这个市场最大的非对称性之一。CNBC 关于 Agentic commerce 的报道显示,Alibaba 把 Qwen 直接接入 Taobao、Fliggy 和 Alipay,ByteDance 则通过商业和移动设备路径推广 Doubao。这些是 Kimi 在已抓取证据集中并不明显具备的分发和留存护城河。Kimi 最强的补偿杠杆,是降低既有 OpenAI 风格开发者的切换摩擦,以及贴近知识工作的产品套件;但这条护城河比拥有超级应用或商业网络更窄。[CP003, CP004, CP007, CP014, CP016, CP019]

定价 / 包装比较
竞争对手公开价格 / 方案单位包含能力含义未知项
Kimi K2.6¥6.50 输入 / ¥27 输出每 1M tokens多模态旗舰,262,144 上下文相比美国高端 API 有竞争力,但不是本地价格地板定制企业折扣未知
Kimi K2.7 Code HighSpeed¥13 输入 / ¥54 输出每 1M tokens更快的编码档位说明 Kimi 会对性能 / 速度档位变现企业 SLA 条款未知
DeepSeek V4 Pro¥3 输入 / ¥6 输出每 1M tokens1M 上下文加智能体工具兼容对 Kimi 最强的可见本地价格地板威胁合同包装未知
Baidu ERNIE 5.0 / X1.1¥6-10 输入;¥24-40 输出(ERNIE 5.0)以及 ¥1 / ¥4(X1.1)每 1M tokens 等价企业智能体平台和可观测性Baidu 覆盖高端和低价推理档位,企业定位明确这里未覆盖消费端 App 包装
Qwen3.5 旗舰区间$0.1-$1.2 输入和 $2.4-$6 输出地板每 1M tokens兼容 OpenAI 的多模态家族,最高 1M 上下文Alibaba 同时覆盖低端和旗舰价格带Qwen App 的确切包装需要单独证据
MiniMax Token Plan基于方案的包装;本次抓取材料没有清楚露出确切页面价格订阅 / 点数Anthropic 式访问、CLI、MCP、编码工具集成即便直接价格透明度偏弱,打包宽度仍然有力直接官方价格阶梯仍需要更易抓取的页面
OpenAI GPT-5.5 $5 输入 / $30 输出每 1M tokens高端海外参照旗舰 API质量和企业付费意愿的高标杆中国公开可用性受限
Anthropic ClaudeOpus 4.8 $5 / $25;Pro 每月 $17;Max 每月 $100 起API + 席位方案合规 API 与企业部署叙事即便价格偏高,打包护城河仍然明确详细企业合同条款未公开

本表只比较可支撑的公开标价或清晰披露的方案结构;缺少官方证据的部分仍按未知处理。

[CP004, CP007, CP014, CP016, CP018, CP028]

3.4 切换成本、锁定与护城河耐久性

在模型层,切换成本仍低于许多历史软件类别。Kimi 发布 OpenAI 迁移指南;DeepSeek 支持 OpenAI 和 Anthropic 格式;Alibaba 明确提供兼容 OpenAI 的 API;Zhipu 记录 OpenAI-SDK 兼容性;MiniMax token 方案文档暴露了 Anthropic 风格集成。这种组合让买家更容易做价格和基准比较。因此,只靠模型质量取胜的供应商可能很快被替换。 更耐久的护城河在上一层。通过消费者或商业生态分发、明确治理工具、搜索与知识资产、更深工作流嵌入,都会提高真实切换成本。Baidu 的企业控制和搜索资产、Alibaba 的商业图谱、ByteDance 的应用触达、Tencent 的超级应用定位,都在这一更高层运作。Kimi 的护城河更温和,但依然真实:它有可识别品牌、长上下文定位、公开知识工作入口,以及面向已使用 OpenAI 风格工具开发者的迁移便利。问题在于,在本土价格压力和平台支持竞争者进一步缩小差距前,这组能力能否足够快地复利。[CP003, CP006, CP010, CP013, CP015, CP018]

护城河耐久度 / 竞争风险登记表
护城河主张或风险威胁严重性缓释措施 / 尽调问题
OpenAI 式迁移便利,帮助 Kimi 快速拿下开发者多个同业也主打兼容,竞争对手的切换成本同样被压低测试初始迁移之后,Kimi 能否留住用量
知识工作助手界面(文档、幻灯片、表格、研究)Alibaba、Baidu 或 ByteDance 的更大生态套件可以吸收类似任务中高核查用户是否在 Kimi 专属工作流里形成习惯
长上下文与编码定位DeepSeek、Z.ai 和 MiniMax 也在销售长周期或高上下文编码 agent用实际任务成功率和留存做基准,而不只看模型主张
独立实验室聚焦缺少平台巨头的超级应用、商业或搜索分发评估付费获客效率和合作伙伴撬动能力
公开企业证明弱于 Baidu 或 Anthropic 的打包能力受治理约束的买方可能偏好明确的审计 / 合规工具寻找公开标杆客户、管理功能和合规披露
本地价格竞争力相对美国参照DeepSeek 和部分 Qwen 层级仍设定更低的本地价格地板跟踪毛利率韧性和折扣纪律
战略支持方提供资本和入口资本本身不等于渠道所有权或留存区分融资实力与可复用分发优势
中国 AI 网站前五位置不等于品类领导地位;监测用量中 Doubao 和 DeepSeek 更大跟踪未来排名更新中,Kimi 份额是增长还是停滞

本登记表抽出最可能影响耐久度的竞争压力,而不是复述泛泛的市场竞争。

[CP003, CP021, CP023, CP024, CP031, CP032]
FP003: 护城河 / 就绪度 KPI

用一组紧凑指标看 Kimi 当前相对同行的竞争姿态。

[CP019, CP020, CP027, CP029, CP032, CP036]

3.5 竞争风险、替代压力与可能进入者

Kimi 的负面情景很直接。DeepSeek 官方价格地板压制本土变现。Doubao 和 Qwen 受益于强得多的母公司分发。Baidu 展示企业控制和更丰富的搜索绑定平台。Z.ai 持续发布长周期和编码中心产品,争夺 Kimi 需要的同一批开发者心智。MiniMax 通过把多模态创作与编码、消费产品结合,进一步拓宽国内赛场。中国之外,OpenAI、Anthropic 和 Google 仍围绕质量、企业包装和规模设定高端预期。换句话说,Kimi 在价格上被下方挤压,在生态触达上被侧面挤压,在全球基准声誉上被上方挤压。 这不意味着 Kimi 弱。公开流量排名仍显示它是有意义的挑战者,不是事后注脚;Moonshot 融资历史也说明,按创业公司标准它并不缺资本。但公开证据确实表明,Kimi 的护城河仍是有条件的,而非已成定局。缺少披露的胜率、流失率或合同深度数据时,更稳妥的判断是:Kimi 是一个可信、高速度的挑战者,身处全球竞争最激烈的 AI 市场之一。[CP019, CP020, CP025, CP026, CP032, CP036]

Chapter 04

04财务

4.1 收入模式与定价入口

Moonshot 现在有了可见的变现栈,而不是单一不透明的聊天机器人入口。Kimi 官方定价文件显示,公司对实时聊天的输入和输出 token 均计费,按模型家族设置不同标价,对批处理作业打折,并在网页搜索工具使用上叠加明确的单次调用费。这意味着 Moonshot 同时对使用量、模型层级和工作流类型变现。公开文档不能证明已实现价格或利润率捕获,但确立了从模型使用到可计费收入的具体路径。K2.6 和 K2.7 Code 都面向长上下文、Agentic 工作负载定价;Moonshot V1 仍是更便宜的旧模型家族;批处理和工具定价则把变现延伸到可容忍更高延迟或检索更重的工作流。 这些标价入口重要,最清晰的商业证明来自 ARR 加速与企业预付需求的组合。CnTechPost 称,K2.5 发布一个月后,ARR 在 2026 年 3 月初突破 $100 million;TechCrunch 称 ARR 在 4 月超过 $200 million,由订阅和 API 使用驱动。同一篇 CnTechPost 报道称,部分企业客户提出数千万美元预付承诺,以锁定优先算力。这没有揭示经常性收入占比、折扣或合同期限,但说明 Moonshot 不只是收割消费者注意力——它也在把稀缺推理容量变现。因此,本章把 API token、Kimi 付费访问、企业优先容量、批处理和工具调用视为公开证据最能支撑的五个收入入口;已实现 ASP 和续约质量则保留为尽调项,而不是编造事实。[CI008, CI010, CI014, CI018, CI019, CI020]

收入流表
收入流机制单位当前数值 / 状态收入质量尽调问题
实时 API 推理K2.6、K2.7 Code 和 V1 系列的聊天补全按 token 计费1M tokens官方已发布标价;2026 年 API 需求据报强劲中:已明确货币化,但相对算力成本的实际价差未知按模型家族提供净收入、基础设施成本和毛利率
Kimi 付费用量 / 订阅在旗舰模型上叠加消费者或专业消费者付费访问订阅用户或方案TechCrunch 将付费订阅列为 ARR 增长的一部分;未披露方案收入中:原则上可复购,实际不透明按方案披露订阅用户数、ARPU 和流失率
企业优先容量承诺为算力优先级预付承诺或提供保障合同 / 预付余额CnTechPost 称部分客户承诺数千万美元低至中:现金信号有吸引力,但可能集中或一次性提供预付余额滚动表和头部客户集中度
批量推理对延迟不敏感的任务,价格低于标准实时费率1M tokens官方定价为标准费率的 60%中:支撑成本敏感型用量扩张提供批量 token 占比和按任务类别划分的利润率
工具调用收入网页搜索调用单独收费,并对返回内容按 token 计费每次调用加 tokens官方定价为每次搜索调用 ¥0.03中:属于附加收入,但在 agent 工作流里结构性高频披露工具挂载率和毛利贡献

收入流只限于官方文档或 2026 年最新报道直接证明的界面;实际收入组合未公开披露。

[CI010, CI014, CI018, CI019, CI020, CI021]
定价 / 货币化表
产品公开定价证据可能计价基础折扣 / 未知项来源
Kimi K2.6¥1.10 缓存输入 / ¥6.50 非缓存输入 / ¥27.00 输出,每 1M tokens高端多模态推理按用量以 token 计费未披露实际企业折扣或经销商分成官方 K2.6 定价页
Kimi K2.7 Code¥1.30 缓存输入 / ¥6.50 非缓存输入 / ¥27.00 输出,每 1M tokens;HighSpeed 费率翻倍面向编码的 token 计费,并设更高速的溢价层级没有合同最低额或挂载率证据官方 K2.7 Code 定价页
Moonshot V1按上下文窗口设 ¥2/10、¥5/20 和 ¥10/30 输入 / 输出阶梯旧模型家族按上下文层级定价未披露相对 K2.x 家族的蚕食情况官方 V1 定价页
批量 API标准标价的 60%;K2.6 输出每 1M tokens ¥16.20成本敏感的异步工作负载未披露 SLA、完成窗口经济性或组合官方批量定价页
网页搜索工具每次调用 ¥0.03,另加通过聊天补全计费的搜索结果 tokenstoken 消耗之上叠加的附加工具调用费挂载率和搜索结果 token 膨胀未知官方工具定价页

这些是官方标价;不应视作实际收入、混合 ASP 或毛利率。

[CI018, CI019, CI020, CI021, CI022, CI023]
FI001: 收入模型桥

Moonshot 借助分层的 token 与工具栈,把模型使用量变成收入,拉动 ARR 增长。

这张桥图展示结构,不拆分收入构成。公开来源未披露各节点贡献了多少 ARR。

[CI014, CI018, CI019, CI022, CI023, CI027]

4.2 单位经济代理指标与推理成本强度

Moonshot 仍未披露毛利率、CAC、回本周期或贡献利润率,因此唯一可辩护的方法是从公开代理指标入手。官方价格表显示,K2.6 输出 token 价格是未缓存输入 token 的四倍多,K2.7 Code 保持类似经济性,批处理则用延迟换取输出费率降低 40%。这些比例说明 Moonshot 正有意识地把客户引向成本更低、公司能控制基础设施利用率的运行模式。同时,定价中心确认模型按纯 token 吞吐量计费,这意味着毛利高度取决于 Moonshot 自身算力成本与公开价目表之间的价差。 公开需求和基准证据进一步说明,推理容量是关键瓶颈。CnTechPost 称,K2.5 发布后 token-per-minute 配额很快收紧;CoreWeave 和 VentureBeat 则显示,当第三方以不同于官方端点的方式优化 K2.6 时,基础设施性能和单位成本差异可以拉开多大。VentureBeat 还把 K2.6 描述为万亿参数 MoE 模型,每个 token 激活 32 billion 参数,这有助于解释容量管理为何如此重要。结论不是 Moonshot 经济性差,而是其经济性明显受基础设施驱动。没有托管端与转售端成本桥接时,公开记录只能支持方向性结论:Moonshot 有真实变现动能,但其贡献利润率很可能比经典 80% 以上毛利率 SaaS 公司更敏感于推理组合、供应商条款和容量规划。[CI009, CI019, CI020, CI022, CI025, CI026]

单位经济表
指标数值 / 空值置信度重要性尽调问题
ARR(2026 年 3 月上旬)$100M+K2.5 发布后最早的新近货币化证明提供经审计的 ARR 方法和月度桥接表
ARR(2026 年 4 月)$200M+显示货币化极快,但历史很短提供月末 ARR、账单额和流失明细
K2.6 输出 / 非缓存输入价格比4.15x输出偏重的用量会实质改变利润率画像披露提示词、缓存和补全 tokens 的实际组合
批量相对标准 K2.6 输出的折扣低 40%显示延迟与货币化之间的明确取舍披露批量用量相对实时用量的占比
优先算力预付款部分企业客户贡献数千万美元表明需求在完整交付前已经货币化,但可能集中按客户群组提供预付余额
毛利率null核心承销指标仍未披露按模型和交付模式提供毛利率
CAC / 回本期null判断增长是高效还是靠补贴驱动,必须看这一项按渠道提供销售效率、CAC 和回本期
客户集中度null企业预付款可能夸大多元化程度提供前 10 大客户占比和扩张率

所有空值都是有意标出的披露缺口,不代表隐含为零。衍生比率只使用官方标价。

[CI008, CI010, CI014, CI025, CI026, CI036]
FI002: 单位经济模型桥

公开经济数据指向需求强劲,但算力稀缺和模型规模让成本敏感度上升。

这张桥图使用公开定价和需求代理指标;内部成本台账或 cohort 利润率均未公开。

[CI009, CI010, CI019, CI022, CI025, CI026]
FI003: 财务估算区间

公开信息给出了 ARR 加速和估值抬升区间,但没有披露 burn 或 runway。

ARR 区间覆盖 2026 年 3 月和 4 月报道;burn 和 runway 仍不可得。

[CI006, CI008, CI012, CI014, CI017]

4.3 资本充足性、融资历史与融资依赖

Moonshot 的资本故事是公开记录中最强的正面财务信号。TechCrunch 和 SCMP 显示,公司从 2024 年前约 $300 million 的基础,推进到 2024 年超过 $1 billion 的融资;TechNode 又补充了 2024 年 8 月后续一轮,投后估值超过 $3.3 billion。Caixin 随后报道 2025 年末 Series C 融资 $500 million,账上现金超过人民币 100 亿元;接着 TechCrunch 在 2026 年 5 月报道,公司以 $20 billion 估值融资 $2 billion,并在六个月内融资 $3.9 billion。这条曲线强烈暗示,投资者愿意以异常快的节奏,为产品野心和容量扩张同时买单。 更难的问题是,这个资本基础是足够,而不只是大。Caixin 称,2025 年融资将用于 AI 芯片、K3 开发和商业化;一家前沿模型公司在仍需先于收入投入时,正会这样表述。CnTechPost 和 WSJ 也显示,Moonshot 仍在重做公司结构,并为潜在香港上市做准备,说明融资策略仍活跃,并未定型。没有公开来源披露月度烧钱、长期云承诺、债务或优先股条款,本章不能把报道现金转成干净的 runway 数字。因此,最有支撑的答案是有条件的:Moonshot 看起来不缺资本,但仍看起来依赖融资;这种依赖很可能同样落在芯片和基础设施上,而不只是销售和市场。[CI001, CI002, CI003, CI004, CI005, CI006]

资本充足性表
项目公开证据当前数值 / 状态重要性尽调问题
在手现金Caixin 引述的创始人信>10B 元 / 2025 年底约 $1.4B若准确,绝对流动性缓冲很大提供经审计现金、受限现金和未动用授信
月度消耗无公开披露null要把现金换算成现金跑道,必须有该指标提供月度净消耗和前瞻预算
现金跑道月数无法用公开来源计算null没有消耗数据,新融资不等于偿付能力提供基准 / 下行现金跑道假设
计划资金用途Caixin 创始人信AI 芯片、K3 开发、商业化证明现金正在投向算力密集型路线图提供 capex 与 opex 拆分和供应商承诺
下一轮或流动性触发点CnTechPost / WSJ香港 IPO 准备和架构清理仍在推进表明融资策略仍然活跃,尚未结束提供目标时间线、上市前置条件和备选方案
债务 / 项目融资义务无公开披露null表外义务可能大幅缩短现金跑道提供债务明细、云资源预购承诺和优先权结构

历史融资时间线只用于解释当前资本充足性;缺失的消耗和债务指标仍是开放尽调项。

[CI005, CI006, CI007, CI012, CI016, CI017]
FI004: 资本强度 / 现金流图

Moonshot 最新融资看起来投向算力、模型开发和商业化,而不是支撑近期自我造血。

融资用途基于报道和创始人信摘要;现金流量表未公开。

[CI006, CI007, CI012, CI016, CI038]

4.4 财务结论与承销阻断点

Moonshot 已有足够证据支撑一个真实财务叙事:可见的按使用量定价、清晰的 ARR 加速、企业为算力预付的意愿,以及一张仅看报道现金和融资就强于多数私有 AI 实验室的资产负债表。它也带有全球前沿模型同行身上同样可见的资本强度特征。OpenAI 和 Anthropic 都在以明确绑定算力扩张和基础设施控制的尺度融资,Moonshot 自身芯片采购计划也符合这一模式。换句话说,Moonshot 看起来不像没有商业模式的投机聊天机器人,更像一个稀缺的基础设施 + 模型资产,其商业轨迹正开始公开浮现。 阻断点同样清楚。公开来源仍未披露收入确认政策、毛利率、客户集中度、CAC、回本周期、债务或 runway;IPO 准备报道部分在付费墙后;即便最好的外部基准,也只能显示 C3.ai 这类公开 AI 软件公司必须披露什么,而不是 Moonshot 自身每个 token 赚多少。因此,正确结论应谨慎而非看多:Moonshot 看起来商业真实且融资充足,但公开记录仍过薄,无法以申报文件级信心承销收入质量、贡献利润率或资本充足性。下表所有无支撑指标因此都作为尽调缺口呈现,而不是编造数字。[CI031, CI032, CI033, CI034, CI035, CI036]

公开财务缺口表
缺失指标对承销判断的影响具体尽调路径
经审计的按月、按收入流确认收入无法把 ARR 头条与 GAAP 或 IFRS 收入对上要求提供经审计月度收入桥接表,并拆分 API / 订阅 / 服务
按模型和交付模式划分的毛利率无法判断 token 定价能否转化成软件式经济性要求提供托管、合作伙伴托管和批量用量的利润率桥接
客户集中度和续约数据预付款头条可能掩盖单一账户依赖要求提供头部客户集中度、续约分群和用量扩张曲线
CAC、回本期和销售效率指标无法判断增长是买来的还是可持续的要求提供漏斗转化、销售人效和回本期
月度消耗、现金跑道和债务明细无法把现金和新融资转化成偿付能力判断要求提供资金计划表、供应商预购和债务契约
企业预付款收入确认政策预先承诺未必等于经常性收入质量要求提供合同模板和递延收入会计处理

每一行都是真实的开源缺口;没有支撑的数值刻意保留为尽调问题,而不是用猜测回填。

[CI010, CI036, CI037, CI039, CI040]

4.5 附录图表

Chapter 05

05产品与技术

5.1 产品入口与用户工作流

Kimi 的公开产品图谱已经足够宽,应被视为一套产品,而不是单一助手。公司和产品主页强调代码、深度研究、网站、表格和幻灯片;平台主页又增加开发者层,提供网页搜索、记忆、Excel 分析、代码执行和内容抓取等官方工具。Kimi Code 位于这张图谱之内,是明确聚焦编码的入口,被营销为连接 Kimi 会员的终端与 IDE 工作流。开发者平台也没有躲在专有 SDK 后面:快速入门材料称 API 兼容 OpenAI,这会降低已围绕 Chat Completions 风格客户端构建团队的采用摩擦。 这会影响产品分析方式。Moonshot 卖的不是更好的原始模型权重,而是把这些权重包装进一组工作流产品,对应不同买家任务:深度研究服务知识工作,代码服务开发者生产力,网站或幻灯片服务创意产出,API 加工具层服务应用构建者。批处理 API、官方工具文档和迁移指南的存在进一步说明,公司预期用户在 Kimi 之上构建多步 Agentic 工作流,而不是只发单轮文本提示。[CE001, CE002, CE003, CE004, CE005, CE006]

产品模块 / 资产矩阵
模块 / 界面主要用户交付内容状态 / 成熟度差异化尽调缺口
Kimi 助手界面知识工作者深度研究、网站、幻灯片、表格和通用助手工作流已上线消费者界面工作流打包宽于基础聊天型同业公开 ROI 和留存数据未披露
Kimi Code开发者和编码团队使用 K2.7 Code 的终端和 IDE 编码助手已上线产品页编码专属打包,加上 agent 化模型调优公开定价方案细节和席位控制较少
Kimi API 开放平台应用开发者以 OpenAI 兼容 API 访问当前 Moonshot 模型已上线平台降低现有 OpenAI 式技术栈的切换成本缺少公开 SLA、可用性和企业支持承诺
官方工具层agent 工作流构建者网页搜索、记忆、代码执行、抓取、Excel、QuickJS 和实用工具已上线文档界面显示文本生成之外的内建工作流野心工具定价和配额行为尚未集中在一个清晰页面
K2.6 / K2.7 Code 旗舰线高阶用户和开发工作流256K 上下文多模态推理和编码模型当前旗舰强调长上下文和长周期任务独立基准验证有限
Kimi-VL多模态构建者面向 OCR、长视频、长文档和 agent 任务的视觉语言模型开源模型家族配备 MoonViT 编码器的高效 MoE VLM企业打包和支持模式仍不清楚
Mooncake 服务栈Moonshot 内部基础设施 / 高级部署方以 KV-cache 为中心的解耦服务架构生产和开源信号聚焦吞吐量的长上下文基础设施公开可见的运营控制不能替代面向客户的服务保证

行里混合终端用户界面、开发者工具、模型家族和后端系统,因为 Moonshot 现在把四者作为一套商业栈交付。

[CE001, CE002, CE003, CE005, CE008, CE017]
工作流 / 用例表
用户任务当前工作流Moonshot 方案可衡量或声称的收益限制
在终端或 IDE 里交付编码任务使用通用聊天模型或编辑器插件,并手工管理上下文基于 K2.7 Code 的 Kimi Code更快的编码专用模型和原生终端界面方案细节、席位和企业控制披露不足
用新模型后端构建 OpenAI 式应用将现有 SDK 重构到专有接口OpenAI 兼容的 Kimi API 快速入门降低现有应用技术栈的迁移摩擦兼容性仍有模型特定注意事项,例如 reasoning_content 处理
在图像或视频上运行多模态推理拼接 OCR、文件存储和独立 VLM 调用具备视觉能力的 K2.6 / K2.7 Code 和 Kimi-VL,加文件上传流程文本、图像和视频理解跑在单一工作流里仍不支持直接远程图片 URL,且有请求体大小限制
编排长研究或内容生成任务手工串联大量提示词或自定义脚本Agent Swarm、官方工具和适合批量的文档并行化任务拆解,工作流原语更丰富许多能力仍被描述为预览或研究界面
高效服务长上下文请求让预填充和解码留在一个紧耦合集群内Mooncake 解耦式预填充 / 解码服务长上下文负载下,吞吐更高,KV-cache 复用更好后端复杂度和生态依赖抬高运维负担
评估多模态事实性使用长尾覆盖较弱的通用 VLM 基准WorldVQA 基准更容易看清视觉世界知识里的幻觉高发点它仍是公司自建基准,不是外部审计

收益单元格汇总的是已披露主张和基准框架;不应解读为经客户验证的 ROI 数字。

[CE004, CE018, CE020, CE032, CE034, CE037]
FE001: 产品架构图

Moonshot 公开产品架构把面向用户的入口搭在 API 与工具层之上,再往下是模型家族层,最底层是 Mooncake 服务基座。

这套栈综合了官方文档和技术报告,不是厂商发布的单一架构图。

[CE001, CE003, CE008, CE014, CE025, CE037]
FE002: 客户工作流 / 运营流程

文档中的工作流从用户任务开始,经由 Kimi 入口或 API 路由;需要时调用工具和多模态处理,底层依赖长上下文服务基础设施。

该流程是概念图,把多个已记录入口压缩成一个通用工作流。

[CE001, CE004, CE017, CE020, CE037, CE050]

5.2 模型家族、长上下文历史与能力包装

Moonshot 的模型家族已有足够内部历史,能看出一条架构主线。早期公开差异点是长上下文,Kimi 在 2023 年已声称具备超大对话窗口。Kimi k1.5 随后标志着偏重推理的强化学习里程碑;K2 则把家族扩展为 1T 参数 MoE 发布,激活参数 32B、上下文 128K,并具备强 Agentic 基准表现。到当前文档集,K2.5 和 K2.6 把活跃产品线推到 256K 上下文,并强化 Agentic 编码、多模态和长周期行为;K2.7 Code 则进一步专门面向编码,采用 always-thinking 交互模型,并提供快得多的 HighSpeed 变体。 Kimi-VL 再次拓宽组合,而不只是延伸文本模型。其公开材料描述了一个 16B 级 MoE VLM,配有 MoonViT 编码器、128K 上下文,并在 OCR、长视频、长文档和 OS-agent 任务上表现有竞争力。WorldVQA 增加了一个面向抗幻觉视觉世界知识的内部基准资产;这很重要,因为它显示 Moonshot 在构建模型的同时也在搭建评测基础设施。合在一起,这个组合不像单一前沿模型押注,更像一个为编码、多模态生产力和 Agent 工作流优化的扩张模型栈。[CE008, CE009, CE010, CE011, CE012, CE013]

路线图 / 发布 / 开发阶段表
日期 / 阶段功能或里程碑状态影响来源
2024-07-01Context Caching 公测已发布显示 Moonshot 很早就投入长上下文成本控制平台博客
2024-08-07企业 API 正式推出已发布商业化从消费者助手延伸出去平台博客
2025-07-11Kimi K2 博客发布已发布开放式智能体模型家族公开亮相平台博客
2025-08-22K2 HighSpeed 发布已发布延迟优化成为明确产品表面平台博客
2025-09-05K2 模型更新已发布显示 K2 发布后的快速迭代平台博客
2025-11-06K2 Thinking 发布已发布给 K2 系列加入智能体能力和推理升级平台博客 / CNBC
2026-01-28K2.5 公开披露已发布以智能体能力和视频主张释放下一代信号CNBC
2026-02-03WorldVQA研究发布Moonshot 给技术栈补上基准和质量叙事Moonshot 首页
2026-02-09Agent Swarm研究发布Moonshot 推进横向多智能体编排Moonshot 首页
2026-04-20K2.6当前旗舰版本当前产品线转向更强的多模态和编码重点Moonshot 首页
2026-06-17 至 2026-06-18K2.7 Code 和更新后的文档页面当前文档状态截至运行日期,编码专项系列已在文档中列为当前版本模型页面 / 快速入门 / 站点地图

本表追踪已审阅公开资料中的主要开发者平台和模型里程碑,不覆盖每个补丁或内部实验。

[CE012, CE021, CE035, CE036]
FE004: 产品成熟度 / 能力图

当前公开材料显示,Moonshot 在代码、上下文、多模态和 Agent 工具上覆盖面很宽;企业保障和外部审计质量的成熟度较弱。

矩阵单元格是定性判断,仅基于已审阅公开语料。

[CE023, CE025, CE029, CE030, CE037, CE048]

5.3 开发者平台、架构与服务栈

Moonshot 公开架构中最重要的技术点是,产品栈不只是模型包装器,而是工作流和系统栈。API 层,Moonshot 暴露兼容 OpenAI 的端点、工具使用、视觉输入、批处理和模型选择指南。应用层,公司推广 Kimi Code、研究、网站和其他更高层入口。底层公开技术材料显示,其服务和基础设施姿态认真。Mooncake 在论文和代码库中都被描述为以 KVCache 为中心的解耦服务架构:分离 prefill 与 decode,在 CPU、DRAM 和 SSD 资源之间池化缓存,并提升长上下文工作负载下的请求吞吐量。 这个服务层重要,因为 Kimi 的产品主张高度依赖长上下文、多步 Agent 和多模态工作流,而这些都吃基础设施。Mooncake 论文和 USENIX 演示报告了很大的吞吐增益和生产规模;开源 checkpoint-engine 代码库则称,Kimi-K2 权重可在约 20 秒内跨数千块 GPU 更新。因此,架构看起来技术上可信,但也依赖很重:Moonshot 正在构建并贡献一个更广生态,涉及 GitHub 托管代码库、vLLM 风格部署路径、Hugging Face 模型卡,以及第三方硬件或服务集成。这同时带来杠杆和供应商兼容风险。[CE003, CE004, CE005, CE006, CE017, CE018]

技术 / 运营架构表
层级 / 组件作用依赖公开证据质量关键风险
Kimi 应用层把基础模型转成研究、网站、表格、幻灯片等消费者工作流Moonshot UI、当前模型阵容、官方工具官方网站证据质量高功能铺得太快,可能跑在支持和控制披露前面
API 兼容层让开发者用 OpenAI 风格客户端调用 KimiChat-completions 语义和 SDK 兼容性快速入门文档证据质量高假设 OpenAI 默认语义的集成,可能被模型特定行为打乱
推理和工具使用层支持 reasoning_content、多步工具调用和保留思考过程K2.6 / K2.7 Code 推理行为官方指南证据质量高客户端实现必须正确保留特殊字段
多模态摄取层处理 base64 或上传的图像、视频输入Moonshot 文件存储和请求体限制官方视觉指南证据质量高尺寸和格式约束会让生产流水线更复杂
模型家族层提供 K1.5、K2、K2.5、K2.6、K2.7 Code、Kimi-VL 和 Moonshot V1 变体模型训练和打包节奏官方文档和报告证据质量高发布节奏快,会带来弃用和迁移负担
服务基础设施层执行解耦式长上下文服务和缓存池化Mooncake、checkpoint-engine、GPU 集群、RDMA 网络开放技术报告证据质量中高运维复杂度和硬件依赖都很重
外部生态层连接 vLLM、Hugging Face、GitHub、NIM 及相关工具开源和供应商生态兼容性模型卡和代码仓库证据质量中高平台质量取决于能否持续维护多条外部路径

架构根据官方文档和技术报告重建;Moonshot 没有发布一张覆盖整个 Kimi 栈的统一系统图。

[CE004, CE014, CE017, CE025, CE030, CE037]
FE003: 关键依赖图

Moonshot 产品栈依赖自有政策和模型,也依赖外部论文、代码库、硬件生态和服务框架;这些因素共同塑造可部署性和风险。

边表示依赖和风险传导的定性关系,不做定量测算。

[CE005, CE037, CE040, CE041, CE042, CE047]

5.4 信任、质量、合规与路线图风险

Moonshot 的信任包有利有弊。正面看,公司发布了真实政策页面,而不是黑箱:服务条款禁止逆向工程、抓取、滥用自动化、规避安全过滤,以及用服务训练竞争模型;隐私政策描述了加密、备份、监控和事件响应。模型文档还围绕 reasoning_content、多模态限制和兼容性行为给出异常具体的集成指南,这对开发者是质量信号。发布历史也足够活跃,显示真实产品迭代:从 2024 年的上下文缓存和企业 API,到 2025—2026 年的 K2、K2 Thinking、K2.5、Agent Swarm、WorldVQA 和 K2.6/K2.7 Code。 缺口仍然显著。经审阅的公开记录没有披露 SOC 2、ISO 27001、正式 SLA,或采购重的客户常要求的其他企业级保障材料。更重要的是,风险面并不抽象:OECD AI Incidents Monitor 记录了一起 2026 年涉及 Kimi 的隐私事件,Anthropic 也公开指控 Moonshot 对 Claude 进行了工业规模蒸馏。这些事件没有抹掉产品栈的技术成熟度,但改变了尽调负担。买家或投资者可以把技术承销为有野心且日益全栈;在获得更多私有证据前,不应把它承销为低摩擦、认证丰富的企业基础设施。[CE021, CE034, CE035, CE036, CE043, CE044]

信任 / 质量 / 合规表
控制或风险领域状态范围证据缺口 / 影响
使用政策限制已披露条款禁止逆向工程、抓取、训练竞争模型和规避安全机制服务条款政策存在,但执行透明度未公开
训练数据使用和退出机制已披露隐私政策称,用户内容可用于改进或训练服务;在适用法律下,用户可通过支持渠道退出隐私政策受监管买家可能要求更强的默认排除和合同措辞
安全 / 事件响应措辞已披露加密、备份、监控和事件响应措辞已公开隐私政策政策之外,没有配套的审计认证或正常运行时间材料
企业认证 / SLA公开资料未披露已审阅材料中没有 SOC 2、ISO 27001、HIPAA 或公开 SLA官方文档和政策页面采购流程重的客户很可能要求私下提供材料
隐私事件暴露已记录反向信号OECD AI Incidents Monitor 记录了 2026 年 Kimi 简历泄露事件OECD AIM指向不可忽视的生产数据隔离风险
模型来源 / IP 风险已记录反向信号Anthropic 公开指控涉及 Moonshot 的工业规模蒸馏Anthropic 声明和 CNBC 报道引出关于来源、控制和出口叙事的尽调问题
基准 / 质量姿态部分披露Moonshot 发布大量基准导向的模型卡和 WorldVQA官方文档和研究文章大多数证据仍由公司生成,而非第三方审计

本表同时纳入正向控制和反向信号,因为两者都会影响可部署性和承销信心。

[CE032, CE043, CE044, CE045, CE046, CE047]
Chapter 06

06客户

6.1 Kimi 现在覆盖消费者、专业人士、开发者和早期团队工作流

Moonshot 的客户故事比简单消费者聊天机器人更宽,但公开记录仍指向消费者主导的安装基础,而不是已披露的企业基础。Kimi 自有界面显示消费聊天、Agentic 研究、网站和幻灯片生成、深度研究工作流、API 访问和 Kimi Code。帮助中心进一步区分 Membership 与 Kimi Business,说明即便公司没有发布传统企业案例页面,Moonshot 已经按买家、用户和付款方来思考。App Store 描述也异常明确地列出程序员、研究人员、学生、互联网从业者、法律专业人士和一般 AI 好奇用户。其重要性在于,它提示 Kimi 当前客户图谱是角色和工作流驱动的:终端用户从移动端和网页进入,专业构建者通过 API 和开源工具进入,团队则是位于这些入口之上、披露仍很薄的一层。仍缺失的是具名买家名单,能够证明谁已经从有趣的工作流采用跨入持久企业支出。[CU001, CU002, CU003, CU004, CU005, CU006]

客户分层表
客群买方 / 用户 / 付款方主要用例有证据支撑的规模收入 / 战略价值缺口
消费者应用用户移动端或网页端的买方 = 用户 = 付款方聊天、深度研究、文件理解、幻灯片和日常效率辅助App Store 19 万条评分;可见的应用内购买阶梯;中国 AI 应用前 5 排名代理指标最大的公开顶部漏斗,也是最清晰的变现表面未公开免费用户、付费会员和留存月活之间的拆分
专业知识工作者买方可能是个人;用户是同一人;付款方是自助或小团队预算研究、法律 / 金融文件阅读、写作、演示生成和电子表格工作官方表面明确面向研究人员、法律专业人士和互联网从业者可把高频工作流转化为付费计划或 API 使用未公开席位数、ACV 或职能级渗透率
开发者和构建者买方是工程或产品负责人;用户是开发者;付款方是产品或工程预算API 使用、Kimi Code、开源模型部署和智能体应用构建平台称有数百万专业开发者;GitHub / Hugging Face 显示活跃外部拉力价值更高的路径,因为 Kimi 会嵌入下游产品和工作流未公开付费 API 客户或企业开发者合同拆分
团队 / 工作区用户买方可能是经理、运营负责人或 IT;用户是跨职能贡献者;付款方是团队或部门预算共享工作区、协同智能体、文档到技能复用和业务管理帮助中心和 K2.6 材料展示工作区、Claw Groups 和商业计划脚手架最可能成为通向企业支出的先落地再扩张桥梁未公开席位计划、商业计划客户数或具名团队部署
国际用户买方 / 用户 / 付款方随渠道变化英文使用、跨境网页访问、开源采用和中国以外应用安装SEMrush 显示美国和印度流量占比;App Store 列表支持英文可对冲纯国内流量集中未公开区域收入或留存

分层按产品表面和工作流最有力,而不是按已披露 ARR 或客户 logo 数。买方、用户、付款方的区分, 来自官方表面、定价和开发者分发渠道的推断。

[CU001, CU002, CU003, CU004, CU005, CU006]
界面 / 变现路径表
界面主要用户公开可见内容变现路径证据质量
Kimi web 与 app消费者和专业人士免费使用、应用内付费购买、优先访问历史、广泛工作流功能会员、充值和应用商店购买产品存在证据强;转化深度证据中等
Kimi API开发者和产品团队按 token 计价,另有付费网页搜索工具调用按用量计费机制证据强;客户数量披露弱
Kimi Code / 开源开发者和工程团队公开仓库、模型权重和部署文档靠生态拉动和 API 追加销售间接变现
Kimi Business / 工作区团队和企业帮助中心分类,以及工作区 / Claw-Group 提及可能是席位或合同模式,保留来源未公开枚举低至中

本表把可见的变现机制和底层客户质量拆开。Moonshot 披露路径比披露转化、ACV 或集中度更清楚。

[CU005, CU006, CU007, CU008, CU009, CU010]
FU001: 客户旅程图

Kimi 的公开旅程从消费者和专业工作流发现开始,随后分流到付费访问、API 使用和面向团队的功能。

这是方向性工作流图,不是实测转化漏斗。

[CU001, CU002, CU004, CU005, CU006, CU022]

6.2 采用代理指标显示有意义规模,但主要集中在消费者和平台指标

Moonshot 已有足够独立流量和应用证据,说明 Kimi 远不止新奇产品阶段。SEMrush 报告称,2026 年 5 月 kimi.com 访问量为 34.15 million,中国是最大受众,美国和印度也有可见占比。SCMP 另报道称,Kimi 在中国最受欢迎 AI 应用中排名第五,截至 4 月,Moonshot 与 Zhipu 合计月活用户接近 35 million。App Store 界面提供了另一类证明:Kimi 在 190,000 个评分中获得 4.9 / 5,并展示活跃的应用内购买阶梯。这些指标都不等同于已披露付费席位、活跃账户或收入质量 cohort,但合起来证明了真实用户广度。更强的解读不是“Moonshot 已证明企业耐久性”,而是公司已经证明其在网页、移动和排名界面拥有广泛触达,并已开始把其中一部分触达转成付费访问。这种广度重要,因为它降低了证明兴趣的负担,同时留下付费深度和留存这些更难问题。[CU007, CU008, CU010, CU011, CU012, CU013]

客户增长 / 采用轨迹表
指标数值日期来源置信度影响缺失分母
网站访问量34.15M 次访问;较 4 月 +20.02%May 2026SEMrush2026 年 Kimi 的可见网页触达很大访问量不等于登录活跃用户或付费用户
地理分布中国 33.78%;美国 8.12%;印度 6.12%May 2026SEMrush触达仍由中国主导,但已不是纯国内流量占比不能揭示各地区付费使用
应用商店社会证明19 万条评分,4.9 / 52026 年 6 月 12 日快照Apple App Store强消费者复用和满意度信号评分不是收入或队列留存数据
中国应用排名代理指标中国 AI 应用第 5;文章引用其与 Qingyan 合计 MAU 近 35M2026 年 4 月背景SCMP 引用 AICPBKimi 是有规模的中国消费者 AI 表面,不是小众工具合计 MAU 没有隔离出 Kimi 的精确数字
付费消费者阶梯5.2 元 / 4 天至 399 元 / 年的优先使用计划2024 年文章;作为变现证据仍有参考价值SCMPMoonshot 很早就测试了消费者付费意愿没有公开购买率或转化数据
开发者采用代理指标Kimi K2 仓库 10.9k GitHub 星标;Kimi Code 2.6k2026 年 6 月快照GitHub开源社区拉力真实且仍在星标不能揭示付费 API 客户或生产部署

本表混合消费者、网页和开发者采用代理指标,因为 Moonshot 没有发布单一规范客户数指标。每一行都列出分母或披露缺口, 避免过度解读代理指标。

[CU007, CU008, CU011, CU012, CU014, CU015]
留存 / 重复使用 / 满意度表
指标值 / null客群置信度尽调要求
应用商店评分19 万条评分,4.9 / 5消费者应用用户要求 Moonshot 提供按队列划分的 DAU / WAU 和付费会员留存
直接流量占比直接流量 74.57%网页用户要求提供回访用户占比和认证活跃用户趋势
已发布 NRR / GRR企业 / 企业级要求提供 NRR、GRR、logo 留存和续约队列表
已发布流失率企业 / 企业级要求按产品线和区域提供总流失率
公开宕机历史影响3 月 21 日公开提及的过载崩溃所有用户要求提供事件频率、SLA 姿态和复盘纪律
隐私 / 信任事件对留存的影响敏感用户或企业用户要求披露投诉量、信任修复进展和企业安全审查结果

Moonshot 的公开记录在使用量代理指标上更丰富,在 SaaS 级耐久性上更薄。保留来源集中没有找到公开留存披露的地方,null 为有意保留。

[CU013, CU017, CU020, CU035, CU036, CU037]
FU002: 采用 / 部署漏斗

Moonshot 在认知度、使用量和部分变现上有公开证据,但企业留存或集中度证据不足。

前四个阶段有证据支撑。最后阶段刻意保持定性,因为没有找到公开企业 cohort 披露。

[CU007, CU008, CU011, CU014, CU015, CU017]

6.3 公开证明在开发者和可见用户上更密集,在具名企业账户上更稀薄

今天最具体的非消费者采用证明来自 Moonshot 的开发者界面,而不是具名企业 logo。GitHub 和 Hugging Face 显示,Kimi K2 和 Kimi Code 发布节奏快、当前模型家族多,并有有意义的社区活动。CNBC 关于 Kimi K2 的报道补充称,Moonshot 在开源模型的同时尝试以低于西方竞争对手的价格获客,这是典型的构建者获取动作。K2.6 官方页面也称模型可通过网站、应用、API 和 Kimi Code 使用,强化了多入口分发策略。相比之下,保留的公开来源集没有披露具名付费企业客户、合同规模,也没有可清楚标记为生产而非试点的参考部署。这不意味着企业需求不存在;它意味着证据质量不对称。Moonshot 有足够开发者和高强度用户证明,显示真实外部拉力;但公开企业细节不足,不能声称已在大客户中形成多元化生产成熟度。[CU020, CU023, CU025, CU026, CU027, CU028]

具名客户证明表
客户 / 证明表面客群部署 / 用例生产或试点结果限制
Kimi App Store 用户消费者终端用户用移动端聊天、跑智能体工作流、处理办公任务和研究生产级消费者产品19 万条评分和可见付费 SKU 阶梯显示真实规模的终端用户使用未披露用户名和客户 logo;评分不能证明队列留存
优先使用订阅者付费消费者用户在聊天机器人之上叠加更快响应的充值计划生产级变现功能用户规模大到足以让 Moonshot 推出付费优先访问没有公开订阅者数量或转化率
GitHub / Hugging Face 上的开源构建者开发者和 AI 构建者围绕 Kimi K2 和 Kimi Code 微调、部署并集成进下游产品生产级开发者分发大量社区参与显示真实外部构建者拉力开发者活动不等于已披露企业收入或具名客户
企业 / 工作区用户团队和企业潜在客户协同智能体、工作区、成员资格和商业计划工作流产品表面存在,但公开生产证明不完整官方材料显示 Moonshot 已搭出面向团队的工作流和管理概念留存来源未给出付费企业客户名称或参考部署

覆盖故意保持部分,因为 Moonshot 偏消费者且仍是私营公司。最强的具名证明来自平台级用户和开发者证据; 具名企业客户证据则明显缺席于留存公开来源集。

[CU007, CU008, CU017, CU019, CU025, CU026]
FU003: 客户证明矩阵

公开客户证明在消费者和开发者侧最强;具名企业生产使用和续约披露最弱。

矩阵单元格是对证据质量的分析综合,不是 Moonshot 发布的数字评分卡。

[CU003, CU006, CU017, CU025, CU026, CU027]

6.4 变现路径可见,但耐久性和集中度仍披露不足

Moonshot 确有可见的扩张逻辑。用户可以从免费使用开始,购买更快优先访问,为 API 消耗付费,并可能通过工作区、Kimi Code 或商业计划进入更结构化的业务工作流。问题在于,公开披露停在投资者最关心的部分之前。保留来源没有披露 NRR、GRR、流失、续约时间表、活跃商业账户,或消费者订阅、API 收入与任何渠道收入之间的拆分。同样重要的是,信任和合规栈不够干净,不能忽略:SCMP 报道过过度收集数据的认定,OECD 记录了简历泄露,Moonshot 隐私政策也允许根据司法辖区处理用户内容以改进服务和训练模型。这些问题没有抹掉采用故事,但意味着在本应提升变现质量的客户细分中,扩张可能被采购、隐私或信任异议拖慢。因此,当前正确姿态是有条件的:Kimi 明显具备规模和变现路径,但公开证据仍不足以证明持久、低集中度的企业收入。[CU009, CU035, CU036, CU037, CU038, CU039]

扩张与集中风险表
扩张驱动因素集中风险影响尽调路径
免费到付费消费者阶梯免费用户转为付费会员的比例未知消费者覆盖面未必能转化为持久的收入质量要求披露月活付费用户、转化曲线和退款率
API 与 Kimi Code 采用业余开发者与生产客户的组合未知开发者热度可能夸大变现深度要求披露活跃 API 客户、支出分布和前 10 大账户集中度
Business / 工作区功能没有公开点名的企业账户或公开席位表先落地再扩张的故事在概念上成立,但公开证据尚不足以承销要求披露具名客户访谈、ACV 区间和部署成熟度
国际流量中国以外收入占比未知已有一定多元化,但地域变现仍可能偏中国要求披露区域收入和留存拆分
信任敏感型工作流隐私和数据治理问题可能拖慢企业采购可见采用向持久商业合同的转化会被压低要求提供安全尽调包、投诉趋势和企业退出控制

扩张逻辑看得见,但集中度和收入质量披露缺位。这正是客户章节无法从「规模化采用」推进到「耐久性充分验证」的核心原因。

[CU005, CU006, CU009, CU010, CU032, CU035]

6.5 附录图表

Chapter 07

07风险

7.1 监管、法律和隐私暴露是最高严重度的风险栈

Moonshot 面对的监管环境,已经不是泛泛牵连到生成式 AI,而是直接点名生成式 AI。中国《生成式人工智能服务管理暂行办法》直接适用于面向公众的生成式 AI 服务,把合规拉回网络安全、数据安全和个人信息法律框架,并要求训练数据合法、数据收集必要、服务稳定,符合条件的系统还要履行备案义务。2025 年以来,合规门槛没有降低:White & Case 提到强制标识规则和新的安全导向国家标准,中国官方政策页面也显示 AI 标准和司法指引还在扩容。在这个背景下,Moonshot 的公开信任记录比普通消费应用更关键。SCMP 报道 Kimi 被认定存在过度收集数据,OECD 事件追踪器记录了一起简历泄露,并称该事件引发法律行动。Moonshot 自己的条款和隐私材料也留下真实尽调工作:用户与一家新加坡实体按新加坡法律订约,隐私政策仍允许较宽泛的内容处理、关联方共享,以及因法律或运营需要而保留数据。结论是,这套法律栈里,合规、数据处理和信任已经无法干净切开。[CR001, CR002, CR003, CR004, CR006, CR007]

监管 / 法律风险登记表
规则 / 问题司法辖区状态可能性严重性缓释措施剩余敞口尽调路径
中国《生成式人工智能服务管理暂行办法》不合规中国现行监管基线中高关键Moonshot 发布条款和隐私政策,并在备案制度内运营在产品级备案、数据和标识证据得到验证前仍高要求披露公司的 CAC 备案细节、法律备忘录和逐模块合规图
过度或无关数据收集中国 / 跨境企业使用公开批评已经出现Moonshot 有隐私政策和用户权利流程高,因为 SCMP 报道了与监管方有关的过度数据发现要求披露监管往来、整改步骤和逐产品数据最小化控制
隐私泄露 / 数据隔离失效消费者和企业用户2026 年事件已记录关键Moonshot 称已制定泄露响应计划并提供删除权在根因和防复发控制得到举证前仍高要求提供事件报告、整改时间表和会话隔离的独立验证
标识、标准和司法规则收紧中国规则已生效且仍在演进政策页面较早显示了监管走向中高,因为合规范围扩张可能快于初创公司的控制栈成熟要求提供法律监测流程、负责人,以及标识和新标准的落地证据
合同和司法辖区不透明全球用户现行结构性问题中高条款和隐私政策已发布,仲裁地点明确中高,因为北京与新加坡实体的观感会复杂化尽调和执行假设要求提供集团结构图、各产品签约实体和数据控制者职责

各行按剩余下行排序,而不是按新颖性排序。最大担忧不是某一部孤立法律,而是活跃的中国监管与已经可见的信任事件叠加。

[CR001, CR003, CR004, CR006, CR007, CR008]
FR001: 风险热力图

Moonshot 当前最严重的风险不是普通产品竞争,而是隐私 / 合规、Agent 安全和控制栈不透明。

这张热力图基于保留的公开证据做分析综合,不是公司发布的风险评分卡。

[CR013, CR014, CR026, CR030, CR031, CR035]

7.2 运营可靠性、数据处理和智能体安全把下行风险推到普通聊天机器人之外

Moonshot 的运营风险不只是模型会幻觉,而是产品触面已经足够宽,一旦控制失效,可能同时落到多条工作流上。Kimi 现在覆盖深度研究、网站搭建、文档、幻灯片、电子表格、API 工具调用、Kimi Code,以及常驻智能体工作流。SCMP 已记录过一次可见的过载宕机;隐私政策显示,其数据足迹也远宽于窄口径聊天应用:用户内容可包括文件和生成输出,日志可包括设备和会话标识符,设置允许时还可能收集剪贴板数据。IAPS 把风险再往前推一步,认为 Kimi Claw 和 OpenClaw 生态把常驻监控、提示注入、恶意技能,甚至远程代码执行暴露引入了一个中国托管的智能体栈。Hacker News 对 Harmonic 数据的概述又补上企业端视角:在正式治理到位很久之前,员工未经批准采用 Kimi,可能已经把它带进企业环境。几项风险叠加后,运营风险已横跨正常运行时间、隐私、滥用防控,以及智能体自主性带来的安全后果。[CR005, CR017, CR018, CR019, CR020, CR024]

运营 / 质量 / 安全风险登记表
失效模式可能性严重性缓释成熟度剩余敞口未解缺口
高需求界面过载或可用性事件中高低至中未找到公开 SLA 或详细可靠性报告
用户内容或会话隔离失败关键低至中未找到公开复盘或防复发证据
借 Claw / OpenClaw 的常驻 agent 误用或隐蔽外传关键独立负面研究证据强,但官方产品侧控制细节薄
agent 工作流中的提示注入或恶意技能攻陷保留来源未显示公开安全架构和商店审核控制
员工未经批准把敏感数据上传到 Kimi低至中公开来源集显示风险证据,但未显示企业安全部署控制或遥测

这些行强调会迅速打击客户转化和留存的运营信任失效。产品的 agent 化宽度让「质量」和「安全」无法分开。

[CR005, CR014, CR018, CR019, CR020, CR026]
FR002: 风险传导图

Moonshot 的隐私和可靠性失误可能传导到客户信任、监管负担、变现和估值。

传导边表示保留证据中可见的因果路径,不是加权概率。

[CR004, CR005, CR013, CR014, CR018, CR026]

7.3 竞争、价格压力和生态依赖可能迅速传导到利润率和客户风险

Moonshot 所处市场目前奖励快速发布、低定价和生态整合,即便采用势头强,这也是一种高风险运营姿态。CNBC 报道称,中国公司正把用户增长放在基准测试头条之前,Moonshot 也从 K2 到 K2.5 高速迭代,同时定价低于美国主要对手。Kimi 自己的 API 模型又把问题放大:成本与输入、输出和工具使用挂钩,因此重智能体工作流即便客户侧价格看起来有吸引力,服务成本也可能很高。开发者分发是优势,也是一层依赖:GitHub 和 Hugging Face 页面显示,开源发布、公共工具和外部构建者拉力都在快速推进,这会增加维护负担,也提高滥用、故障或不兼容集成传播到 Moonshot 直接支持边界之外的风险。公开记录也没有披露云服务商或基础设施集中度清单,因此投资者能看到规模证据,却还看不到底层成本基础的韧性,也看不到关键交易对手的集中度。[CR022, CR028, CR029, CR035, CR036, CR037]

合作伙伴 / 依赖风险登记表
依赖交易对手 / 生态角色集中度失效场景严重性缓释措施剩余敞口
应用商店计费和分发Apple / 应用商店通道消费者获客和计费退款摩擦、商店政策变化或账户争议损害付费消费者转化中高App 和 web 两条路径都存在
开源开发者生态GitHub / Hugging Face / 外部开发者模型分发和社区采用中高滥用、不兼容 fork 或生态断裂会抬高支持和治理负担开源扩大触达,也降低 go-to-market 成本中高
AI 工具界面搜索、记忆、代码运行器、fetch 和电子表格工具让 Kimi 更适合工作流一个薄弱工具或连接器就可能把滥用和可靠性风险扩散到整个栈平台界面明确,产品侧可以限制
监管批准 / 备案制度CAC 及相关主管部门必需合规和持续注册备案、标识或数据标准收紧,快于 Moonshot 控制落地关键Moonshot 可调整产品披露和合规流程
未公开的基础设施栈未知云和算力交易对手核心服务交付和成本基础unknown集中供应商或算力瓶颈会削弱韧性和利润率弹性未找到公开依赖图

最大依赖问题是隐藏集中度:投资人能看到公开分发界面,却看不到背后的完整运营交易对手名单或集中度上限。

[CR006, CR008, CR022, CR028, CR029, CR037]
FR003: 依赖图

Moonshot 的依赖横跨监管、应用商店、开发者生态和内部工具层;公开证据还无法完全看清这些依赖。

基础设施节点刻意保持泛化,因为保留的公开记录没有披露具体集中度安排或故障切换交易对手。

[CR006, CR008, CR022, CR028, CR029, CR039]

7.4 执行范围本身已经成为风险,投资论证需要可衡量的止损标准

Moonshot 的机会足够大,执行发散已经成了核心承销问题。公司不是在交付一个窄口径答案引擎,而是在同时运营消费者会员、应用商店计费、API 平台、开源模型发布、Kimi Code、Kimi Claw、智能体群,以及法律或金融文档处理这类风险越来越高的工作流。文档和仓库页面很活跃,这是正面信号,但也说明管理层必须同步多少控制平面。公开证据还没有闭合几条重要尽调链:保留材料中没有公开 SLA 或认证包,抓取到的产品页面看不到备案号,没有详细事故复盘,也没有公开的基础设施交易对手地图。这些缺口不能证明公司薄弱,但会实质性抬高剩余执行风险。因此,正确的投资者姿态应当明确,而不是凭印象判断:现在就要求控制证据、事故纪律和依赖关系映射;如果拿不出来,应视为投资论证破裂信号,而不是常规创业公司缺口。[CR023, CR025, CR027, CR028, CR043, CR044]

人员 / 执行风险登记表
角色 / 职能依赖或缺口可能性严重性缓释措施尽调路径
法务 / 合规归口需要统一负责人统筹中国 AI 规则、隐私、消费者条款和跨境产品公开政策已存在,可作为流程设计锚点要求提供具名合规负责人、法律备忘录和董事会汇报节奏
信任 / 安全运营聊天、API 和 agent 需要统一的事件、复盘和漏洞管理纪律中高关键政策在高层次描述了泄露响应要求提供事件 runbook、漏洞赏金姿态和过去一年事件日志
产品 / 工程优先级界面快速扩张可能跑在质量保证和支持能力前面活跃仓库和帮助中心更新显示已有人员投入要求提供发布治理流程、QA 闸口和缺陷 backlog 指标
客户运营 / 支持消费者和开发者渠道扩张快于支持时,退款、计费和信任问题会叠加中高中高应用分发商和联系邮箱提供了基础渠道要求提供支持 SLA、投诉解决指标和退款 / 拒付趋势
高管聚焦和依赖图谱未找到公开云交易对手图或逐产品合规矩阵即便未公开,也可通过尽调材料补齐要求提供架构图、供应商集中度上限和故障切换设计

Moonshot 的执行风险更多来自需要同步运转的控制面太多,而不是缺少野心:法务、产品、支持、安全和伙伴管理都要一起跑。

[CR023, CR027, CR028, CR043, CR045, CR046]
缓释与终止标准表
风险可监测触发因素阈值 / 事件行动含义
隐私 / 数据治理失效新的重大泄露、监管行动或反复出现的数据收集批评下一个刷新周期内再次发生经核实的用户数据泄露,或监管发现仍未解决升级为红旗尽调;没有独立整改证据前,不承销企业耐久性
可靠性退化公开宕机或发布时过载核心用户界面反复出现多小时宕机,或缺少复盘纪律在运营控制改善前,把利润率和客户扩张假设视为高估
控制栈不透明缺少公开或私下合规材料Moonshot 在尽调中无法提供备案细节、DPA / 安全包或事件证据对受监管或跨境客户承销,从继续研究转为回避
竞争性低价在成本纪律不可见的情况下持续低价发布主要新版本继续压低价格,同时客户质量和留存数据仍未披露假设收入质量更低,并要求更严格的入场纪律
依赖集中没有供应商图或云韧性证据管理层无法识别主要基础设施交易对手和故障切换计划将运营韧性视为未验证,并折减扩张假设

每个触发因素都绑定具体的公开材料或尽调材料,投资团队因此可以把风险变化当作可观察事项,而不是靠直觉判断。

[CR005, CR008, CR013, CR014, CR026, CR037]

7.5 图表资料

Chapter 08

08估值

8.1 融资背景和当前私募估值

目前支撑最充分的估值锚,是 TechCrunch 2026 年 5 月报道称 Moonshot 以 $20 billion 估值融资约 $2 billion。该文还称,公司在此前六个月融资 $3.9 billion,并在 4 月把 ARR 推到 $200 million 以上,这足以解释为什么投资者把 Moonshot 视为中国开源权重前沿实验室中商业化最强的一家。CnTechPost 补充了更早的 3 月数据点:K2.5 发布一个月后,ARR 已超过 $100 million;Caixin 则称,Moonshot 在 $500 million Series C 后,于 2025 年底持有超过 10 billion yuan 现金。合在一起看,这段融资历史不像一串救急轮次,更像一家公司反复把技术动能转化为一级市场资本。 但这个估值锚的结构仍然脆弱。The Wall Street Journal 报道称,Moonshot 在更高审查压力下筹备香港上市时,曾讨论过约 $18 billion 的融资;CnTechPost 称,公司正在拆除红筹结构,因为保住 VIE 豁免的可能性看起来很低。这些报道重要,因为在 $20 billion 估值上,Moonshot 已不再像一家有信息缺口的高速增长私营 SaaS 公司被定价,而是像一个战略资本市场资产被定价。因此,这个估值或许说得通,但前提是商业化和 IPO 执行的下一章能按时到来。[CV001, CV002, CV003, CV004, CV006, CV007]

建议摘要表
维度数值支持证据决策含义
建议继续研究Moonshot 确有变现和资本可得性,但价格支撑仍依赖不透明变量没有数据室前,不按当前估值投入
信心2026 年新来源印证了规模、融资和 IPO 准备继续尽调,而不是停止覆盖
风险评级监管重组、披露缺口和稀缺性溢价下行仍然重要任何交易都应按高波动、受条款敏感处理
估值立场偏高当前估值已计入乐观增长和上市假设要么要求更好披露,要么要求更好入场价格

建议对价格敏感,并不意味着 Moonshot 缺少产品或市场质量。

[CV001, CV004, CV029, CV032, CV033, CV034]
FV001: 推荐逻辑

按当前公开证据,Moonshot 像一家可投公司,但还不像一只可投证券。

这张图综合推荐逻辑;它不是带校准权重的因果模型。

[CV001, CV004, CV006, CV031, CV032, CV034]

8.2 投资论证与反论证

正面逻辑很直接。Moonshot 现在有官方定价,有订阅和 API 变现的报道,有愿意预付算力费用的企业客户,也有足够现金继续训练和商业化,不需要马上紧急融资。这样的组合在中国前沿模型公司中很少见。更重要的是,公司还处在收入曲线较早阶段,如果接下来一年保持类似执行,规模图景可能发生实质变化。相比许多仍躲在产品演示背后的基础模型实验室,Moonshot 至少已经开始公开投资者想看的商业化触面。 反论证同样有力。即便接受 4 月 ARR 数据,$20 billion 标价也意味着一个已经折现多年超高速增长的倍数,并假设上市执行、利润率纪律和客户留存都会兑现,但这些还没有公开证据。监管重组不是表面修饰,而是不利因素。股权结构和优先权栈缺失。公开估值支撑仍部分依赖中国 AI IPO 中可见的稀缺性溢价,而不是一套完整披露的基本面账本。因此,正确结论不是「回避」——业务看起来太真实,不能这么处理——而是在估值偏紧的前提下「继续研究」。[CV025, CV026, CV029, CV030, CV031, CV032]

正方 / 反方论点表
论点支持因素何种变化会改变判断
Moonshot 是中国变现最快的开放权重前沿实验室ARR 从 3 月的 $100M+ 增至 4 月的 $200M+;官方定价和预付需求均已公开独立审计收入或留存低于预期会削弱这一判断
该资产具备公开上市可选性已融资、结构清理和香港准备工作都指向仍在推进的 IPO 路径监管延误或放弃上市计划会实质削弱稀缺性溢价
当前价格已折现大部分利好基于公开 ARR 证据的 $20B 估值意味着当前倍数很高更强的审计后变现基础或更低入场价格会降低这一担忧
安全性定价风险不同于公司质量风险股权结构、清算优先权和投资者权利仍未披露完整融资文件和清晰分配瀑布可迅速改善投资论点

正方论点真实存在,但仍对证据敏感;反方论点关乎安全性定价和披露,而不是产品无关紧要。

[CV004, CV025, CV029, CV030, CV031, CV044]
FV004: 投资 KPI

Moonshot 在市场和验证上得分高,但证据质量和估值舒适度得分低。

评分是基于章节证据得出的 IC 式启发判断,不是机械模型输出。

[CV004, CV026, CV030, CV031, CV034]

8.3 可比估值框架

Moonshot 落在一组尴尬的可比公司之间,这正是当前价格难以承销的原因。OpenAI 和 Anthropic 说明,当算力、开发者需求和分发同时复合时,私营前沿模型龙头仍可拿到极高融资倍数。但这些公司对经常性收入、产品宽度和基础设施战略的披露,也远多于今天的 Moonshot。公开市场一侧,C3.ai 给出了低规模、弱利润率 AI 软件的严苛下限;Palantir 则说明,一旦执行已经被证明,市场仍愿意为 AI 驱动增长和现金生成支付极高溢价。 中国 AI IPO 让图景更复杂。Yicai 和 TechNode 显示,Zhipu 和 MiniMax 上市时,市场愿意按远高于美国中盘软件可比公司可解释水平的价格资本化前沿模型稀缺性。这有助于解释为什么 Moonshot 当前估值并不明显荒谬。但这不等于这个估值本身有吸引力。最站得住脚的解释是:Moonshot 现在更接近一份公开市场稀缺性期权,而不是带折价的私募轮次。上行仍有可能,但只有当商业化和上市准备的收敛速度持续快过怀疑情绪时才成立。[CV010, CV011, CV012, CV013, CV014, CV015]

可比估值表
可比对象关键指标倍数 / 估值相关性局限
Moonshot AI(私有,2026 年 5 月)2026 年 4 月 ARR >$200M据报道估值 $20B;按公开最低锚点计算,隐含 ARR 倍数约 100x最接近本次决策的直接锚点公开 ARR 是地板信号,不是经审计收入
Anthropic(私有,2026 年 5 月)运行率收入 $47B投后估值 $965B;约 20.5x 运行率收入展示规模化前沿模型赢家能拿到的定价规模、全球化和披露程度都远高于 Moonshot
OpenAI(私有,2026 年 3 月)收入 $2B/月(年化约 $24B)投后估值 $852B;约 35.5x 年化收入展示极强分发优势下的前沿稀缺溢价分发、算力控制和地缘政治背景不同
Zhipu AI(公开首秀,2026 年 1 月)2025 年上半年收入 CNY190.9M;亏损严重首秀市值 $7.4B显示香港投资者即便面对亏损,也会给中国 AI 稀缺性资本化收入期不是全年,业务组合也不同
MiniMax(公开首秀,2026 年 1 月)IPO 募资 $619M;首秀市值 >$11.5B;9M25 毛利率 69.4%公开市场接受一家亏损中国 AI 发行人的极高估值有用的中国前沿模型先例首日市值波动大,不代表稳态估值
C3.ai(公开,2026 年 6 月)FY2026 收入 $250.3M市值 $1.50B;3.9x EV/Sales经济性偏弱的小规模 AI 软件公司估值下沿不是前沿模型平台,增速也低得多
Palantir(公开,2026 年 6 月)TTM 收入 $5.22B市值 $307.98B;57.46x EV/Sales已证明盈利能力和现金生成的 AI 驱动增长公司公开市场上沿比 Moonshot 成熟且盈利能力强得多

这是一组样本可比公司,不覆盖所有 AI 发行人。用途是给 Moonshot 估值划出区间,而不是精确定价。

[CV001, CV004, CV011, CV012, CV014, CV015]
FV002: 估值敏感性

Moonshot 的估值敏感性主要由 ARR 规模,以及投资者愿意为前沿稀缺性支付多少溢价驱动。

评分是序数值(1-5 影响),不是概率估计。

[CV005, CV007, CV031, CV043]

8.4 牛市 / 基准 / 熊市情景和建议

在现有证据水平下,建议是继续研究,估值立场是偏紧。核心原因在于,只有当 Moonshot 继续极快复合收入,同时成功把监管清理转化为可信的香港上市选项时,$20 billion 估值才守得住。牛市情景下,ARR 扩至约 $500 million,投资者继续给 55-70x 稀缺性溢价,对应 $27.5-$35 billion 结果。基准情景下,ARR 达到约 $350 million,溢价回落到 50-60x,留下 $17.5-$21 billion 区间,使今天价格更接近公允而非便宜。熊市情景下,ARR 在 $250 million 附近停滞,随着审查、稀释或需求转弱削弱热情,倍数压缩到 30-40x,对应 $7.5-$10 billion,并带来实质下行。 这些情景刻意避免对 IRR 做虚假精确化,因为股权结构和清算优先权栈未知。它们只回答一个更窄的问题:当前估值是否已经计入了大部分好消息?答案是肯定的。Moonshot 有足够证据留在可投资名单上,但没有足够证据支持在缺少额外尽调的情况下穿透当前价格买入。更积极的立场需要两件事之一:经审计的商业化数据显著强于 4 月 ARR 所暗示的水平,或入场价格显著更好。[CV005, CV032, CV033, CV034, CV035, CV036]

牛市 / 基准 / 熊市情景表
情景核心假设估值 / 回报逻辑概率信号关键风险
乐观ARR 向 ~$500M 复合增长,IPO 路径保持打开,55-70x 稀缺溢价守住隐含估值 $27.5B-$35.0B;存在上行空间,但取决于持续高增长需要 4 月之后变现持续推进,披露也更清晰监管或利润率一旦滑坡,该情景会很快被压缩
基准ARR 达到 ~$350M,溢价回落到 50-60x隐含估值 $17.5B-$21.0B;当前估值大致合理最可能的前提是 Moonshot 仍是品类龙头,但不是全球级爆发异类新资金安全边际很薄
悲观ARR 卡在 ~$250M 附近,倍数压缩到 30-40x隐含估值 $7.5B-$10.0B;较当前水平有实质下行IPO 延后、付费采用放慢或利润率质量偏弱都会触发该情景下轮降估值、治理折价、估值重置

情景测算仅作示意,且有意取整;股权结构条款、稀释和经审计 ARR 均未披露。

[CV035, CV036, CV037]
FV003: 估值 / 回报区间

在现有证据下,情景区间说明 Moonshot 更适合放进观察名单,而不是按一个干净价格直接买入。

区间采用 ARR 和倍数的情景假设,不是 DCF 或只看可比公司的模型。

[CV035, CV036, CV037]

8.5 最后尽调要求和投资论证破裂触发器

缺失证据异常集中在估值关键区域。投资者需要经审计的月度收入桥、股权结构和优先权栈、算力采购承诺,以及香港上市准备的真实时间表。没有这些,公司仍可能是高质量资产,但无法判断新投资者与底层业务之间隔着多少下行保护、稀释或现金强度风险。因此,尽调负担不在于弄清 Moonshot 有没有需求——公开证据已经显示大概率有——而在于弄清这些需求能否在当前价格下转化为一只耐久、可投资的证券。 投资论证破裂触发器也相应具体。上市流程失败或实质延迟,会刺破稀缺性溢价。低于当前估值的融资会确认估值透支。ARR 停在支撑 2026 年热情所需水平之下,会推翻「变现最快实验室」叙事。披露包若显示毛利率薄弱或激进优先权悬顶,会把一家有前景的公司变成一只糟糕证券。这些不是遥远边缘情形;Moonshot 是从观察名单进入组合,还是从观察名单转为放弃,应由这四个问题决定。[CV038, CV039, CV040, CV041, CV042, CV043]

投资逻辑失效与止损触发表
触发项阈值对投资逻辑的传导行动含义
IPO 执行失败结构清理后,上市明显延后或放弃稀缺溢价和退出可选性坍塌从「继续研究」转为「回避」,直到融资条款重设
降估值或平估值融资下一轮新股融资低于当前私人估值,或没有明显高于当前估值证实估值跑在基本面前面重切情景表,并假设下行伴随大量稀释
增长停滞证据仍停留在 2026 年 4 月 ARR 地板附近,而不是继续复合增长乐观和基准情景失去支撑要么要求更低入场价,要么放弃机会
利润率披露偏弱Data room 显示毛利率差,或存在大额算力预购成长故事变成低质量证券除非价格大幅重置,否则按回避处理
优先权悬顶高级投资者权利或反稀释堆叠严重压在新进入者身上上行可能属于早期轮次,而不是新资本要求更强治理,否则放弃
监管收紧香港或内地政策变化限制上市或外资持有人结构退出时点和投资者池同时收缩重估为持有期更长、倍数更低的私人资产

这些触发项用于监测,不是预测;它们是最可能打破估值逻辑的具体事件。

[CV038, CV039, CV040, CV042, CV043, CV045]
最终尽调清单
主题缺失证据重要性负责人 / 尽调路径
经审计变现桥月度确认收入、ARR、递延收入和预收余额判断 4 月 ARR 是否足够持久,能否支撑当前估值财务团队 / 审计师材料
股权结构表和优先股堆叠完整股权结构表、清算瀑布、反稀释、ROFR 和投资者权利条款判断真实证券质量和下行保护公司法律顾问和 CFO
毛利率桥自托管与合作伙伴路由经济性、算力承诺和支持成本分摊揭示增长更像软件,还是严重依赖基础设施转手FP&A 加基础设施运营
IPO 准备材料结构清理状态、上市时间表、承销商状态和监管反馈当前估值部分押注公开市场可选性总法律顾问、外部律师和投行
客户集中度 / 留存头部客户占比、预付依赖和留存队列检验变现故事是否分散且持久收入运营和销售负责人
竞争切换风险相对 DeepSeek、Zhipu、MiniMax 和第三方 K2.6 托管方的赢单 / 输单分析显示当前需求属于 Moonshot,还是属于暂时性能窗口产品、GTM 和大客户团队

这些问题按它们改变定价观点的速度排序,而不是按获取难度排序。

[CV041, CV042, CV043, CV044, CV045]

8.6 图表资料

免责声明

本报告是基于公开证据的尽调快照,不构成投资建议。重要的财务、法律、技术和合同事实仍未公开;任何投资决策前,都应向管理层和一手文件直接核验。

证据索引

结论
编号陈述可信度来源
CO001 Moonshot AI is the company name used on the official corporate homepage. SO001
CO002 Kimi is the public-facing assistant brand presented alongside Moonshot AI on the official web surfaces. SO001, SO002
CO003 The official Moonshot AI site frames the company mission as seeking the optimal conversion from energy to intelligence and pursuing AGI research. SO001
CO004 Moonshot AI currently operates both consumer Kimi surfaces and a developer-facing Kimi API open platform. SO001, SO002, SO003
CO005 The current Kimi product surface emphasizes code, deep research, websites, sheets, and slides rather than chat alone. SO001, SO002
CO006 TechCrunch reported that Moonshot AI was founded in 2023 by Yang Zhilin, Zhou Xinyu, and Wu Yuxin. SO007
CO007 TechCrunch reported that the startup name was inspired by Pink Floyd's The Dark Side of the Moon. SO007
CO008 TechCrunch in May 2026 described Moonshot AI as a Beijing-based AI lab. SO008
CO009 The current Kimi Terms of Service identify Moonshot AI PTE. LTD. in Singapore as the provider of the services. SO005
CO010 The safest headquarters formulation is that Moonshot AI appears to operate from Beijing while also using a Singapore service entity, and the official sites do not publish a single consolidated headquarters page. SO005, SO008
CO011 Moonshot AI's public monetization stack combines usage-based API billing with paid plans or memberships tied to Kimi surfaces. SO003, SO004, SO005
CO012 Kimi Code is marketed as a coding-focused perk of Kimi membership and drops into terminal and IDE workflows. SO004
CO013 The Kimi API bills both input and output tokens and keeps file extraction interfaces temporarily free, which is consistent with a developer-usage revenue model. SO003
CO014 Yang Zhilin is the founder most prominently associated with Moonshot AI across current profile and funding coverage. SO007, SO010
CO015 SCMP reported that Yang described Moonshot AI as aiming to combine OpenAI's technology idealism with ByteDance's business philosophy. SO010
CO016 TechCrunch reported that Yang previously worked at Meta AI and Google Brain. SO007, SO008
CO017 The reviewed official surfaces do not provide a public board roster, investor-relations page, or named finance leader. SO001, SO002, SO005
CO018 The visible public record is founder-centric enough that key-person dependence on Yang Zhilin remains a material diligence issue. SO010, SO017
CO019 TechCrunch reported in February 2024 that Moonshot AI had raised more than $1 billion in a Series B round at a reported $2.5 billion valuation. SO007
CO020 TechCrunch reported that Alibaba and HongShan co-led that 2024 round, with Meituan and Xiaohongshu also participating. SO007
CO021 Caixin reported that Moonshot AI completed an oversubscribed $500 million Series C round and held more than RMB 10 billion in cash at year-end 2025. SO011
CO022 Caixin reported that Yang said Moonshot AI was not in a rush to go public and would use new funds to buy AI chips, accelerate K3 development, and pursue commercialization and revenue growth. SO011
CO023 TechCrunch reported in May 2026 that Moonshot AI raised about $2 billion at a $20 billion valuation. SO008
CO024 TechCrunch named Long-Z Investments, Tsinghua Capital, China Mobile, and CPE Yuanfeng as participants in the 2026 round. SO008
CO025 TechCrunch reported that Moonshot AI raised $3.9 billion over the prior six months, after being valued at $4.3 billion at the end of 2025 and over $10 billion following an earlier 2026 raise. SO008
CO026 Forbes independently reported the same $2 billion financing and $20 billion valuation and added that Alibaba, Tencent, and 5Y Capital had already joined earlier 2026 financings. SO009
CO027 Public round labels and capital totals are inconsistent enough that lifetime capital raised should be treated as directionally very large rather than precisely settled from public sources alone. SO007, SO008, SO009, SO011
CO028 TechCrunch reported that Moonshot AI's annual recurring revenue topped $200 million in April 2026, citing a Huafeng Capital post. SO008
CO029 The reviewed public corpus does not provide audited revenue, customer count, or headcount disclosures for Moonshot AI. SO001, SO002, SO008, SO009
CO030 Forbes reported that Cursor was using Kimi K2.5 as a customer reference by May 2026. SO009
CO031 TechCrunch reported that Kimi K2.6 had become the second-most used LLM on OpenRouter by May 2026. SO008
CO032 TechCrunch reported that Moonshot AI launched a 100 billion-parameter model in March 2023. SO007
CO033 TechCrunch reported that Moonshot launched the Kimi chatbot in October 2023 with a claim of supporting 200,000 Chinese characters in one conversation. SO007
CO034 CNBC reported that Moonshot released Kimi K2 in July 2025 as a low-cost open-source model. SO012
CO035 CNBC reported that Moonshot released Kimi K2 Thinking in November 2025 as its second major AI update in four months. SO013
CO036 CNBC reported that Moonshot revealed Kimi K2.5 in January 2026 and said the model claimed video-generation and agentic capabilities. SO014
CO037 The official Moonshot AI homepage lists WorldVQA on 2026-02-03, Agent Swarm on 2026-02-09, and Kimi K2.6 on 2026-04-20 as its latest research milestones. SO001
CO038 The Kimi open-platform blog records context-caching and enterprise-API milestones from 2024 onward, showing a steady expansion from long-context API primitives into a broader product platform. SO025
CO039 CNBC reported in February 2026 that Anthropic accused Moonshot AI of participating in a large-scale model-distillation campaign using fraudulent accounts. SO015
CO040 Anthropic's own statement claimed Moonshot generated more than 3.4 million exchanges with Claude while targeting agentic reasoning, tool use, coding, data analysis, and computer vision. SO016
CO041 The OECD AI Incidents Monitor recorded a 2026 Kimi incident in which one user's resume data was reportedly disclosed to another user. SO017
CO042 Kimi's privacy policy says prompts, images, videos, files, and other user content may be processed to provide and improve the services, including model training and optimization. SO006
CO043 The Kimi Terms of Service describe subscriptions, recurring billing, and paid features, reinforcing that Moonshot monetizes more than pure research output. SO005
CO044 Later chapters should treat customer count, headcount, board composition, and exact lifetime capital raised as open diligence items rather than settled facts. SO005, SO008, SO009
CM001 Moonshot positions Kimi K2.6 as a natively multimodal model with coding and agent performance rather than as a raw infrastructure or chip offering. SM001, SM006
CM002 Kimi’s consumer interface exposes website, document, slides, spreadsheet, deep research, Kimi Code, Kimi Claw, and agent-cluster workflows, indicating an application-layer knowledge-work product surface. SM001, SM002
CM003 Kimi’s API documentation includes files, batch, tool calls, JSON mode, partial mode, and web search, extending the product beyond simple chat into developer and agent workflows. SM005
CM004 Kimi’s current public model set emphasizes 256K-context multimodal and coding models, showing that Moonshot competes in long-context assistants and developer APIs rather than generic consumer search alone. SM006
CM005 Because Moonshot’s public surfaces sell assistant, research, and API workflows, the relevant market excludes raw semiconductors, general cloud IaaS, and the entire China software economy. SM001, SM002, SM022
CM006 IDC’s China AI market-glance separates model, application-development, and agent-development platforms from chips and infrastructure, supporting an upper market boundary centered on model and agent software rather than compute hardware. SM022
CM007 OpenAI’s supported-countries policy says service is unsupported wherever a location is absent from the list, and mainland China is absent while Taiwan is listed. SM012
CM008 China’s 2023 CAC generative-AI measures apply to services offered to the public inside China and require providers to manage prohibited content, discrimination, intellectual-property risk, privacy, and service security. SM013
CM009 China’s 2025 AI-labeling regime adds explicit and implicit marking requirements for AI-generated text, image, audio, video, and virtual-scene content, with effect from 2025-09-01. SM014, SM015
CM010 IDC forecasts that by 2027, 80% of China C1000 enterprises will prioritize AI sovereignty through nonpublic hosting, open technologies, and regional partners for mission-critical uses. SM022
CM011 AICPB ranks Kimi fifth in China AI websites for April 2026 with 43.69 million monthly visits. SM010
CM012 AICPB ranks Kimi eighth in China AI apps for April 2026 with 25.33 million monthly active users. SM010
CM013 Kimi accounts for about 4.01% of the top-10 China AI website-visit pool reported by AICPB for April 2026. SM010
CM014 Kimi accounts for about 2.11% of the top-listed China AI app MAU pool reported by AICPB for April 2026. SM010
CM015 On AICPB’s April 2026 China AI rankings, Kimi’s website traffic is only about 26.8% of Doubao’s and Kimi’s app MAU is only about 7.5% of Doubao’s. SM010
CM016 On the same AICPB ranking, Kimi’s website traffic is about 9.0% of DeepSeek’s and Kimi’s app MAU is about 18.2% of DeepSeek’s. SM010
CM017 AICPB’s April 2026 global chatbot website ranking places Kimi ninth at 43.69 million visits, behind ChatGPT, Claude, DeepSeek, Doubao, and several other leaders. SM011
CM018 Kimi K2.6 is priced at ¥1.10 cached input, ¥6.50 uncached input, and ¥27.00 output per 1 million tokens with a 262,144-token context window. SM003
CM019 Kimi K2.7 Code keeps the same ¥6.50 uncached-input and ¥27.00 output list price as K2.6, while the highspeed tier raises the rates to ¥13.00 input and ¥54.00 output per 1 million tokens. SM004
CM020 Kimi’s web-search tool costs ¥0.03 per tool call, and successful search results also add billable search tokens to the next chat-completions call. SM008
CM021 Moonshot explicitly documents OpenAI-SDK compatibility and says many applications can migrate by replacing the base URL and API key, keeping model-layer switching costs low. SM007
CM022 DeepSeek’s API docs advertise OpenAI and Anthropic compatibility with up to 1 million tokens of context and much lower list prices than Kimi’s local coding and multimodal models. SM024, SM025
CM023 Alibaba’s Model Studio provides official Qwen APIs, OpenAI-compatible APIs, and multimodal text, image, and audio/video support across China and overseas deployment regions. SM026
CM024 Alibaba’s published model list shows flagship Qwen3.5 models with 262,144 to 1,000,000 tokens of context and minimum input prices from $0.1 to $1.2 per 1 million tokens, reinforcing the breadth of low-cost alternatives around Kimi. SM027
CM025 OpenAI lists GPT-5.5 at $5.00 input and $30.00 output per 1 million tokens, a premium external benchmark well above local Chinese RMB-denominated list prices. SM019
CM026 Anthropic lists Opus 4.8 at $5 input and $25 output per MTok and pairs paid consumer/enterprise plans with compliance and admin features from $17 Pro to $100 Max. SM020
CM027 Google Cloud lists Gemini 3.1 Pro Preview at $2 input and $12 output per 1 million tokens and Gemini 3.5 Flash at $1.5 input and $9 output globally. SM021
CM028 IDC says China AI coding vendors used low pricing around RMB 20 per month and still kept total 2025 revenue below RMB 100 million despite 100% growth, indicating fast adoption but thin monetization. SM022
CM029 CNBC reports Chinese AI firms are competing through faster model rollouts, open-source distribution, lower prices, and ecosystem integration rather than only headline benchmark wins. SM017
CM030 Caixin says Chinese AI startups still face material compute and funding constraints because U.S. restrictions on advanced Nvidia chips limit computational resources relative to U.S. peers. SM018
CM031 TechCrunch reported Moonshot raised over $1 billion at a reported $2.5 billion valuation in 2024, highlighting both capital intensity and strategic dependence on large Chinese backers. SM016
CM032 IDC’s China AI+ strategy slide targets 70% penetration of smart devices and AI agents by 2027 and 90% by 2030, supporting a broad long-run domestic adoption backdrop. SM022
CM033 Moonshot’s realistic SAM is narrower than China’s total AI spend because its public products most clearly map to consumer assistants, developer APIs, research/search workflows, and enterprise knowledge-work agents. SM001, SM002, SM005, SM022
CM034 Moonshot’s public surfaces imply distinct buyer-user-payer combinations between self-serve knowledge workers, API developers, and enterprise teams purchasing localized productivity or agent workflows. SM001, SM002, SM007
CM035 IDC’s China AI market-glance maps demand into software development, operations, finance, sales and marketing, HR, customer service, and industry verticals, indicating multiple budget owners beyond a single chatbot line item. SM022
CM036 The combined CAC service rules and labeling rules make local compliance, traceability, and content governance a structural cost of operating in China’s public generative-AI market. SM013, SM014, SM015
CM037 Kimi’s public ranking and product evidence support a meaningful domestic position, but not category leadership, so Moonshot still needs to win on workflow depth, coding performance, and localization rather than raw consumer reach. SM010, SM011, SM001, SM006
CM038 OpenAI’s absence from mainland China and IDC’s note that multinational vendors banning Chinese users accelerate local capability development both support the structural demand tailwind for domestic substitutes such as Kimi. SM012, SM022
CP001 Moonshot’s public surfaces combine Kimi assistant workflows, deep research, website/document/slides/spreadsheet tools, Kimi Code, Agent Swarm, and API access. SP001, SP002
CP002 Kimi’s current public model lineup centers on K2.6 multimodal and K2.7 Code models with 256K context and explicit coding and agent upgrades. SP005
CP003 Moonshot documents OpenAI-SDK migration by swapping the base URL and API key, which keeps model-layer switching costs lower than a proprietary rewrite would. SP006
CP004 Kimi K2.6 is listed at ¥6.50 uncached input and ¥27 output per 1M tokens, while K2.7 Code HighSpeed raises pricing to ¥13 input and ¥54 output. SP003, SP004
CP005 DeepSeek’s public home page spans chat, app, open platform, status page, and an extensive open-source research catalog. SP019
CP006 DeepSeek API docs support both OpenAI and Anthropic formats and say mainstream agent and coding tools can use DeepSeek as a backend model. SP020
CP007 DeepSeek V4 Pro is priced at 3 yuan uncached input and 6 yuan output per 1M tokens with a 1M context window, undercutting Kimi’s flagship list prices. SP021
CP008 Zhipu’s documentation presents a one-stop platform covering text, vision, image, video, audio, OCR, knowledge bases, agents, model deployment, and OpenAI-SDK compatibility. SP022
CP009 Z.ai release notes say GLM-5.2 supports 1M lossless context and GLM-5.1 is designed for long-horizon tasks that can run for up to 8 hours in a single run. SP023
CP010 Z.ai also advertises GLM-4.7-Flash as a free-tier model and highlights Claude Code compatibility in the GLM-4.5 series release notes. SP023
CP011 Volcengine’s Doubao page shows a wide ByteDance lineup spanning code, lite/pro/mini text models, Seedance video, Seedream image, and multiple speech and realtime voice products. SP025
CP012 Doubao therefore competes as a multimodal consumer-plus-cloud stack, not only as a single domestic chatbot. SP025, SP026
CP013 Baidu Qianfan positions itself as an enterprise agent platform with multi-agent orchestration, RAG, observability, logging, and audit-compliance features. SP027, SP028
CP014 Baidu publicly lists ERNIE 5.0, ERNIE X1.1 Preview, ERNIE 4.5 Turbo, and third-party DeepSeek services on the same platform, with explicit RMB token pricing for ERNIE models. SP027
CP015 Alibaba Model Studio integrates the full Qwen series and mainstream third-party LLMs via official and OpenAI-compatible APIs. SP015
CP016 Alibaba’s published model list gives Qwen3.5-Plus text, image, and video input, up to 1M context on flagship tiers, and minimum input prices from $0.1 to $1.2 per 1M tokens. SP016
CP017 MiniMax’s public home page spans M3 coding/agentic models, Hailuo video, audio, Talkie, Code, and a Token Plan developer surface with 1M context marketing. SP017
CP018 MiniMax’s token-plan docs expose Anthropic-base-url compatibility, MiniMax CLI, MCP/web-search tooling, and integrations for Claude Code, Cursor, Codex, OpenCode, and other coding tools. SP029
CP019 AICPB’s April 2026 China AI rankings put Kimi at 43.69M website visits and 25.33M app MAU versus DeepSeek at 486.50M website visits and 138.98M app MAU and Doubao at 162.89M website visits and 336.04M app MAU. SP007
CP020 AICPB’s global chatbot ranking places Kimi ninth at 43.69M visits while ChatGPT leads at 5.69B and Claude ranks third at 839.88M. SP008
CP021 CNBC says Chinese AI firms are competing through faster releases, open-source and low-cost strategies, and ecosystem integration rather than only benchmark wins. SP009
CP022 CNBC reports Moonshot revealed K2.5 with video-generation and agentic claims during the 2026 release race, showing Kimi is still a credible product competitor even if it is not the usage leader. SP009
CP023 CNBC’s agentic-commerce reporting shows Alibaba connecting Qwen to Taobao, Fliggy, and Alipay while ByteDance upgraded Doubao to handle tasks through Douyin-linked commerce flows. SP010
CP024 The same CNBC piece says super-app ecosystems give Alibaba, Tencent, and ByteDance integrated data, payments, logistics, and consumer familiarity that independent labs lack. SP010
CP025 TechCrunch reported Moonshot raised over $1 billion at a reported $2.5 billion valuation in 2024 with strategic Chinese backers including Alibaba and other large internet platforms. SP011
CP026 IDC’s China excerpt lists Moonshot, Doubao, Qwen, ERNIE, DeepSeek, MiniMax, and Zhipu among major foundational-model players and says open-source models can run at one-third to one-half Claude-series cost in coding. SP018
CP027 IDC also says China AI coding vendors used pricing around RMB 20 per month while total 2025 revenue stayed below RMB 100 million despite 100% growth, highlighting commoditization risk. SP018
CP028 OpenAI prices GPT-5.5 at $5 input and $30 output per 1M tokens, establishing a premium outside-reference price point. SP012
CP029 Anthropic pairs premium API rates with Compliance API, enterprise desktop deployment, and paid Pro and Max plans at $17 and from $100 per month. SP013
CP030 Google Cloud lists Gemini 3.1 Pro at $2 input and $12 output per 1M tokens and Gemini 3.5 Flash at $1.5 input and $9 output globally. SP014
CP031 Kimi’s public differentiation is strongest in long-context knowledge-work assistant surfaces and easy migration for OpenAI-style developers rather than in public evidence of enterprise compliance packaging. SP001, SP002, SP006
CP032 Kimi’s distribution is materially weaker than Doubao and DeepSeek on public China website visits and app MAU, which limits its default consumer reach. SP007
CP033 Multi-homing is easier than in many software markets because Kimi, DeepSeek, Alibaba/Qwen, Zhipu, and MiniMax all emphasize migration-friendly or widely compatible APIs and coding tools. SP006, SP020, SP022, SP029
CP034 Switching costs rise when vendors bundle model access with search, commerce, app ecosystems, governance tooling, or embedded agent platforms rather than only an API endpoint. SP010, SP027, SP028
CP035 Baidu Qianfan and Anthropic both surface audit or compliance features more explicitly in the fetched public pages than Moonshot does in the Kimi pages reviewed here. SP013, SP027, SP028, SP001, SP002
CP036 DeepSeek is the clearest local price-floor threat because its official docs show far lower list pricing than Kimi alongside long context and agent-tool compatibility. SP020, SP021, SP003, SP004
CP037 Qwen and Doubao are the clearest parent-platform threats because official and CNBC sources show broad model breadth plus stronger distribution through Alibaba and ByteDance ecosystems. SP010, SP015, SP016, SP025
CP038 Z.ai is a likely entrant threat in coding and agents because its release velocity emphasizes long-horizon agents, mobile automation, free tiers, and Claude Code compatibility. SP023
CP039 MiniMax is the closest China-native multimodal challenger on creator-plus-code breadth, even though direct public pricing detail is thinner in this fetched set than for Kimi or DeepSeek. SP017, SP029
CP040 OpenAI remains the outside reference for global scale because AICPB shows ChatGPT traffic vastly exceeding any China-origin chatbot in the same monitored ranking. SP008, SP012
CP041 Kimi is still competitive enough to sit close to Qianwen on monitored China AI website traffic, so it should be treated as a serious challenger rather than a fringe player. SP007
CP042 Kimi’s competitive risk comes from a stack of pressures at once: cheaper APIs below it, ecosystem giants beside it, and premium global references above it. SP007, SP009, SP010, SP012, SP013, SP014, SP021
CI001 Before the 2024 mega-round, TechCrunch reported Moonshot previously raised $200 million at a $300 million valuation from HongShan and ZhenFund. SI001
CI002 TechCrunch reported in February 2024 that Moonshot had raised over $1 billion in a Series B round at an implied $2.5 billion valuation. SI001
CI003 SCMP reported Alibaba disclosed a total investment of approximately $0.8 billion for an approximately 36% Moonshot stake, implying a valuation near $2.2 billion. SI002
CI004 TechNode reported Moonshot last raised about $300 million in August 2024 at a post-money valuation of over $3.3 billion. SI003
CI005 Caixin reported Moonshot completed a significantly oversubscribed $500 million Series C round in late 2025. SI007
CI006 Moonshot held more than 10 billion yuan, or about $1.4 billion, in cash after the late-2025 Series C according to the founder letter cited by Caixin. SI007
CI007 Caixin said Moonshot planned to use the new funds to buy AI chips, accelerate K3 development, and focus on commercialization and revenue growth. SI007
CI008 CnTechPost reported Moonshot ARR surpassed $100 million in early March 2026, roughly one month after Kimi K2.5 launched. SI004
CI009 CnTechPost reported Moonshot's API tokens-per-minute quota tightened quickly after the K2.5 launch. SI004
CI010 CnTechPost reported some enterprise clients made spending commitments and prepaid guarantees in the tens of millions of dollars to secure priority computing resources. SI004
CI011 CnTechPost said Moonshot's latest two funding rounds collectively exceeded $1.2 billion before the May 2026 mega-round. SI004
CI012 TechCrunch reported Moonshot raised about $2 billion at a $20 billion valuation in May 2026, led by Meituan's Long-Z Investments with Tsinghua Capital, China Mobile, and CPE Yuanfeng participating. SI006, SI005
CI013 TechCrunch reported Moonshot raised $3.9 billion over the prior six months. SI006
CI014 TechCrunch reported Moonshot ARR topped $200 million in April 2026, driven by paid subscriptions and API usage. SI006
CI015 TechCrunch reported Moonshot's valuation moved from $4.3 billion at the end of 2025 to $10 billion in early 2026 before the May 2026 $20 billion round. SI006
CI016 CnTechPost reported Moonshot is dismantling its red-chip structure because the company is unlikely to secure a waiver to keep a VIE model for a Hong Kong listing. SI005
CI017 The Wall Street Journal reported Moonshot was considering a corporate-structure change for a Hong Kong IPO and a contemporaneous private round that would value the company at around $18 billion. SI008
CI018 Moonshot's pricing hub says chat-completion billing charges both input and output tokens, while document extraction interfaces are temporarily free. SI009
CI019 The official Kimi K2.6 price sheet lists 1M-token prices of ¥1.10 for cached input, ¥6.50 for uncached input, and ¥27.00 for output, with a 262,144-token context window. SI010
CI020 The official Kimi K2.7 Code sheet lists 1M-token prices of ¥1.30 cached input, ¥6.50 uncached input, and ¥27.00 output; the HighSpeed version doubles those rates. SI011
CI021 The Moonshot V1 pricing page lists 1M-token prices ranging from ¥2.00 input and ¥10.00 output on the 8k model to ¥10.00 input and ¥30.00 output on the 128k model. SI012
CI022 The official batch-pricing page says Batch API jobs are priced at 60% of standard rates; K2.6 batch pricing is ¥0.66 cached input, ¥3.90 uncached input, and ¥16.20 output per 1M tokens. SI013
CI023 The official tools page prices web search at ¥0.03 per invocation, and the associated search-result tokens are billed through the next chat-completion call. SI014
CI024 Moonshot's official site presents AGI research and open-source-community work, while Kimi's product homepage markets K2.6 around multimodal, coding, and agent performance. SI015, SI016
CI025 Official K2.6 list pricing implies output tokens are roughly 4.15 times as expensive as uncached input tokens. SI010
CI026 Official batch pricing implies a 40% discount to real-time K2.6 output pricing, lowering the output rate from ¥27.00 to ¥16.20 per 1M tokens. SI010, SI013
CI027 Moonshot's public monetization surface is usage-based rather than seat-based, spanning real-time API tokens, batch tokens, model-tier choice, and paid tool calls. SI009, SI010, SI013, SI014
CI028 CoreWeave said independent benchmarking delivered Kimi K2.6 at 205 output tokens per second and about $0.7 per million tokens blended, showing how aggressively third-party inference providers are competing on speed and cost. SI024
CI029 VentureBeat reported a benchmark in which Cerebras completed a standard Kimi K2.6 request in 5.6 seconds versus 163.7 seconds on the official Kimi endpoint, highlighting performance dispersion across infrastructure providers. SI025
CI030 VentureBeat described Kimi K2.6 as a one-trillion-parameter MoE model with 32 billion activated parameters per token and a 256,000-token context window. SI025
CI031 OpenAI raised $40 billion at a $300 billion post-money valuation in March 2025 and said the capital would scale compute infrastructure. SI020
CI032 OpenAI said its 2026 round brought in $122 billion at an $852 billion valuation, alongside an explicit compute-flywheel thesis and revenue of $2 billion per month. SI021
CI033 Anthropic said its Series G valued the company at $380 billion on $14 billion run-rate revenue, then Series H valued it at $965 billion on $47 billion run-rate revenue, underscoring how capital hungry frontier-model competition has become. SI022, SI023
CI034 C3.ai's SEC-filed 2026 earnings release reported $250.3 million of revenue, 31% GAAP gross margin, and $575.4 million of cash, showing what public disclosure looks like for an AI software vendor with real but still challenged economics. SI017, SI018
CI035 Stock Analysis shows C3.ai trading at roughly 3.9x EV/Sales with negative free cash flow, illustrating how public markets punish low-margin, cash-consuming AI software even after revenue scale is reached. SI019, SI017
CI036 Moonshot has not publicly disclosed audited revenue, gross margin, CAC, payback, customer concentration, or revenue-recognition policy.
CI037 Moonshot has not publicly disclosed monthly burn, runway months, debt facilities, or project-finance obligations.
CI038 The best-supported capital-adequacy conclusion is that Moonshot has unusually strong access to capital and a large reported cash balance, but open sources still do not permit a filing-grade runway calculation. SI006, SI007
CI039 Moonshot's public revenue quality looks strongest on monetization breadth and demand, but weakest on realized pricing, subscription retention, and margin transparency. SI004, SI006, SI009, SI010, SI014
CI040 Moonshot's public disclosures are sufficient to support a growth-and-capacity narrative but insufficient to underwrite revenue quality, contribution margin, or capital sufficiency without a data room. SI006, SI007, SI017, SI019
CE001 Moonshot AI presents Kimi as a multimodal productivity surface spanning code, deep research, websites, sheets, and slides rather than a chat-only assistant. SE001, SE002
CE002 Kimi Code is positioned as a terminal-and-IDE coding surface linked to Kimi membership. SE027
CE003 The Kimi open platform is marketed as a developer surface for building with Kimi APIs and official tools. SE003
CE004 Moonshot documents Kimi API as fully compatible with the OpenAI API format. SE004
CE005 The platform homepage exposes official tools for web search, memory, Excel analysis, code execution, QuickJS, date handling, URL fetch, conversion, and base64 operations. SE003
CE006 The llms.txt index shows workflow guides for batch jobs, web search, official tools, Kimi CLI, OpenClaw, and OpenAI migration. SE013
CE007 Moonshot bills chat-completion usage on both input and output tokens and keeps file-related extraction interfaces temporarily free. SE008
CE008 The current platform models page lists K2.7 Code, K2.7 Code HighSpeed, K2.6, K2.5, Moonshot V1 8k/32k/128k, and vision-preview models. SE005
CE009 The models page says both K2.7 Code variants provide 256K context windows. SE005
CE010 The models page says K2.6 upgraded K2.5 in agentic coding, long-context reasoning, long-cycle execution, and front-end design. SE005
CE011 The models page says K2.5 delivered open-source state-of-the-art performance across agent, code, vision, and general-intelligence tasks with 256K context. SE005
CE012 The models page says the older K2-series preview models were retired on 2026-05-25 and kimi-latest was retired on 2026-01-28. SE005
CE013 Moonshot's K2.6 pricing page describes the model as text-image-video capable, available in thinking and non-thinking modes, and compatible with ToolCalls, JSON Mode, Partial Mode, web search, and automatic context caching. SE006
CE014 Moonshot's thinking-model guide says K2.7 Code is always-on thinking with preserved thinking always enabled, while K2.6 allows reasoning to be disabled or retained via thinking.keep. SE011
CE015 The K2.7 Code quickstart says the HighSpeed variant is the same model running roughly 5-6x faster, around 180 tokens per second typically and up to 260 tokens per second in short-context coding. SE010
CE016 The K2.7 Code product docs describe it as a multimodal coding model supporting text, image, and video input for agent tasks. SE007, SE010
CE017 Moonshot's vision guide says K2.6 and the K2.7 Code variants can understand both images and videos. SE012
CE018 The vision guide supports base64 uploads or Moonshot file IDs via ms://, supports up to 100 MB request bodies, and does not support direct URL image inputs. SE012
CE019 The vision guide documents multi-turn dialogue, streaming output, tool calling, JSON Mode, and Partial Mode for the vision-capable models. SE012
CE020 The K2.6 quickstart demonstrates multimodal tool loops for video clip analysis, showing that Moonshot is productizing agent workflows rather than simple prompt-response completion. SE009
CE021 Moonshot's platform blog shows that context caching entered public beta on 2024-07-01 and enterprise API formally launched on 2024-08-07. SE014
CE022 TechCrunch reported in early 2024 that Moonshot's original differentiation was unusually long-context language models and that Kimi claimed 200,000 Chinese characters in one conversation. SO007
CE023 The Kimi k1.5 technical report describes an RL-trained multimodal model that matched OpenAI o1 on several reasoning benchmarks. SE016
CE024 Moonshot's Kimi-k1.5 repository identifies the model family as the company's reinforcement-learning scaling milestone. SE015
CE025 Kimi K2 is described in Moonshot's model card and technical report as a 1T-parameter MoE model with 32B activated parameters trained on 15.5T tokens using MuonClip. SE018, SE019
CE026 The K2 model card lists a 128K context window and MLA attention. SE018
CE027 Moonshot distinguishes between K2 Base and K2 Instruct, with the instruct checkpoint positioned for drop-in chat and agentic experiences without long thinking. SE017, SE018
CE028 CNBC reported that the original K2 API pricing was 15 cents per million input tokens and $2.50 per million output tokens. SE033
CE029 K2.6 supersedes the retired K2 previews by adding 256K context and more explicit multimodal, thinking, and agent-task positioning in the official docs. SE005, SE006
CE030 Kimi-VL is documented as an MoE vision-language model with roughly 16B total parameters, about 3B activated, 128K context, and a MoonViT visual encoder. SE020, SE021, SE022
CE031 The updated Kimi-VL 2506 release says it cut thinking length by about 20 percent while raising image resolution support to 1792x1792 or 3.2 million pixels. SE020
CE032 WorldVQA is documented as a 3,500-example benchmark across 9 categories designed to measure factual visual world knowledge and long-tail hallucination resistance. SE038
CE033 Moonshot's public surfaces show that Kimi now spans code generation, websites, slides, spreadsheets, document work, and deep research. SE001, SE003
CE034 The Agent Swarm research post says K2.5 Agent Swarm can deploy up to 100 sub-agents, execute more than 1,500 tool calls, and deliver better results 4.5x faster than sequential execution. SE037
CE035 The Moonshot homepage dates Agent Swarm to 2026-02-09 and K2.6 to 2026-04-20 in the current research timeline. SE001
CE036 The open-platform blog shows a release cadence of K2 in July 2025, K2 HighSpeed in August 2025, K2 model updates in September 2025, K2 Thinking in November 2025, and long-thinking API support in July 2025. SE014
CE037 Mooncake is described in the paper as a KVCache-centric disaggregated architecture that separates prefill and decoding clusters. SE024
CE038 The Mooncake paper reports up to 525% simulated throughput improvement and 75% more requests under real workloads versus the baseline method. SE024
CE039 USENIX FAST reported Mooncake operating across thousands of nodes, processing over 100 billion tokens daily, and enabling 115% and 107% more requests on NVIDIA A800 and H800 clusters respectively. SE025
CE040 The Mooncake repository says the serving stack powered Kimi K2 deployment on 128 H200 GPUs with 224k tokens per second prefill throughput and 288k tokens per second decode throughput. SE023
CE041 Checkpoint-engine says updating Kimi-K2 weights across thousands of GPUs takes about 20 seconds. SE026
CE042 The Mooncake repository says its high-performance P2P store has already been applied in K1.5 and K2 production training. SE023
CE043 Moonshot's Terms of Service prohibit reverse engineering, automated scraping, high-frequency abusive usage, safety-filter evasion, and training competing models from the service. SE028
CE044 Moonshot's privacy policy says user prompts, media, and files may be used to operate, improve, and train the services, with opt-out available by contacting support in accordance with applicable law. SE029
CE045 The privacy policy describes encryption, security checks, backups, and a cybersecurity incident response process. SE029
CE046 The OECD AI Incidents Monitor records a Kimi incident in which one user's resume data was leaked to another user, highlighting a production privacy and data-isolation risk. SE030
CE047 Anthropic and CNBC publicly framed Moonshot as part of a large-scale distillation campaign, creating external provenance, export-control, and IP-compliance risk even before any adjudication. SE031, SE032
CE048 The reviewed public corpus does not disclose SOC 2, ISO 27001, HIPAA, audited uptime, or formal enterprise SLA commitments for Kimi surfaces or the API. SE028, SE029
CE049 Because K2.7 Code requires preserved reasoning content across turns and cannot disable thinking, naïve OpenAI-compatible clients still face integration friction even though the API surface is nominally compatible. SE004, SE011
CE050 Moonshot's docs show batch APIs, official tools, web search, and agent-support guides, indicating that the developer surface is designed for longer workflows instead of single-shot chat only. SE013
CU001 The Kimi help center publicly organizes the product around Agent Mode, Kimi Claw, Kimi Code, Deep Research, Docs & Sheets, Websites, Membership, and Kimi Business. SU003
CU002 Moonshot says Kimi K2.6 is available through the Kimi website, Kimi App, Kimi API, and Kimi Code. SU005
CU003 The Kimi API Platform markets itself as trusted by millions of professional developers. SU006
CU004 Kimi’s App Store description explicitly targets programmers, researchers, students, internet workers, legal professionals, and broad AI-curious users. SU007
CU005 Moonshot says Kimi K2.6 is free to use and that paid plans are available for users who want more features or workflow enhancement. SU005
CU006 The help center says Membership covers plans, billing, credits, upgrades, and invoices, while Kimi Business covers enterprise benefits, pricing, team management, and workspaces. SU003
CU007 The App Store lists Kimi in-app purchases ranging from low-value tips to annual plans priced as high as 1,948 yuan. SU007
CU008 SCMP reported that Kimi introduced six priority-use top-up plans ranging from 5.2 yuan for four days to 399 yuan for one year. SU010
CU009 Kimi API pricing is usage-based and includes an extra $0.004 web-search charge per invocation. SU004
CU010 CNBC reported that Kimi K2 was free in Moonshot’s app and browser while API prices were $0.15 per million input tokens and $2.50 per million output tokens. SU012
CU011 SEMrush reported 34.15 million visits to kimi.com in May 2026, up 20.02% from April. SU016
CU012 SEMrush reported that Kimi’s largest website audience share was in China, followed by the United States and India. SU016
CU013 SEMrush reported that 74.57% of kimi.com traffic was direct in May 2026, with Google contributing 10.9%. SU016
CU014 SCMP reported that Kimi ranked fifth among China’s 10 most popular AI applications as of April 2026. SU021
CU015 SCMP reported that Kimi and Zhipu’s Qingyan had a combined total of nearly 35 million monthly active users as of April 2026, citing AICPB. SU021
CU016 AICPB says its AI App Rankings are based on April 2026 app monthly active users and updated monthly using a standardized methodology. SU015
CU017 Kimi’s China App Store page showed a 4.9 out of 5 rating from 190,000 ratings as of the June 12, 2026 version listing. SU007
CU018 Kimi’s App Store listing was published in Simplified Chinese, Traditional Chinese, and English, indicating some international packaging beyond mainland-only Chinese. SU007
CU019 SCMP said Kimi started charging for faster responses after user numbers surged, showing visible monetization pressure on a large consumer base. SU010
CU020 SCMP said Kimi’s app and website crashed for hours on March 21 because of overload issues. SU010
CU021 TechCrunch described Moonshot’s early customer wedge as long-context use cases such as legal documents, fiction writing, and deeper financial analysis. SU011
CU022 Moonshot says K2.6 can deliver output across websites, documents, slides, and spreadsheets from a single coordinated run. SU005
CU023 Kimi’s App Store description says Agent Swarm can dispatch up to 100 agents across 1,500 steps and that Kimi Claw supports 24/7 scheduled tasks with long-term memory. SU007
CU024 Moonshot says Claw Groups let multiple agents with different tools, contexts, and models work together inside a shared workspace. SU005
CU025 Moonshot’s GitHub organization shows 38 repositories with fresh updates in June 2026 across Kimi Code, kimi-cli, kimi-agent-sdk, and infrastructure components. SU018
CU026 Moonshot’s GitHub organization page showed the Kimi K2 repository at 10.9k stars and 856 forks on the fetch date. SU018
CU027 Moonshot’s GitHub organization page showed the Kimi Code repository at 2.6k stars and 301 forks on the fetch date. SU018
CU028 The Kimi K2 repository describes K2 as a 1T-parameter MoE model designed for tool use, reasoning, and autonomous problem-solving, and offers builder-oriented deployment guidance. SU019
CU029 Moonshot’s Hugging Face organization page shows multiple Kimi model collections and recent updates, indicating continued external distribution into the developer ecosystem. SU020
CU030 Moonshot’s Hugging Face organization page displays multi-million model activity counters and four-digit community reactions on current Kimi models, although the exact metric labels are platform-specific. SU020
CU031 CNBC quoted Counterpoint’s Wei Sun saying Kimi K2 was globally competitive and open-sourced, but still needed better integration tooling for developers to switch from rivals. SU012
CU032 Moonshot says K2.6 code and weights are publicly available on GitHub and Hugging Face. SU005, SU019, SU020
CU033 In the retained public source set, Moonshot does not disclose a named paying enterprise customer or contract reference for Kimi. SU001, SU003, SU005, SU006, SU007, SU018, SU020
CU034 Moonshot’s fetched official materials show business and workspace scaffolding, but they do not surface a public seat schedule or public enterprise case study. SU003, SU005
CU035 No public NRR, GRR, churn, renewal, or cohort table appeared in the retained Moonshot customer source set. SU001, SU003, SU005, SU006, SU007, SU008, SU009
CU036 Kimi’s App Store ratings and direct-traffic mix are useful repeat-use proxies, but neither metric substitutes for account-level retention disclosure. SU007, SU016
CU037 Moonshot’s privacy policy says user content can be processed to provide and improve the service, including model training and optimization depending on jurisdiction. SU009
CU038 SCMP reported that Chinese authorities found Kimi had accessed data irrelevant to its functions. SU021
CU039 OECD’s AI incident monitor recorded a 2026 event in which Kimi exposed one user’s resume to another, making public trust risk a current rather than hypothetical issue. SU023
CU040 White & Case says China’s Interim AI Measures and September 2025 labeling rules create an active compliance layer for generative AI providers. SU024, SU025
CU041 Kimi’s terms place disputes under Singapore law and SIAC arbitration, adding cross-border contractual complexity for some customers. SU008
CU042 IAPS argues that Kimi Claw’s always-on agents widen data-exposure risk enough to deter some foreign procurement even if model quality is improving. SU022
CU043 Public Moonshot sources do not disclose what share of customer monetization comes from consumer subscriptions, API usage, or any partner channel. SU003, SU004, SU005, SU006, SU007, SU008, SU010
CU044 Public Moonshot sources do not disclose top-customer exposure or whether adoption is primarily direct rather than partner-mediated. SU001, SU003, SU006, SU018
CR001 China’s Interim Measures for Generative AI Services took effect on August 15, 2023 and apply to generative AI services offered to the public in China. SR010, SR012
CR002 The Interim Measures are grounded in China’s Cybersecurity Law, Data Security Law, Personal Information Protection Law, and related statutes. SR010
CR003 The Interim Measures require providers to use lawfully sourced data and base models and to obtain consent or another lawful basis when training data contains personal information. SR010, SR012
CR004 The Interim Measures say providers must not collect unnecessary personal information or illegally retain or disclose user inputs and usage records. SR010
CR005 The Interim Measures require providers to offer safe, stable, and continuous service to users. SR010
CR006 The Interim Measures require providers with public-opinion or social-mobilization capability to conduct security assessments and complete algorithm filing procedures. SR010, SR011, SR012
CR007 The CAC announced that 748 generative AI services had completed filing and 435 applications or functions had completed registration by the end of 2025. SR011
CR008 The CAC said live generative-AI applications should disclose the model name and filing or registration number in a prominent place or on product-detail pages. SR011
CR009 White & Case says China’s September 1, 2025 labeling rules made implicit labels mandatory and explicit labels required where applicable for AI-generated content. SR012, SR032
CR010 White & Case says three national standards covering data annotation, pre-training and fine-tuning data security, and basic service security took effect on November 1, 2025. SR012, SR013, SR033
CR011 China’s State Council said the country planned to formulate more than 50 national and industrial AI standards by 2026. SR013
CR012 China’s Ministry of Justice said courts would refine judicial rules around AI, data rights, and AI-generated content during the 2026-2030 planning period. SR014
CR013 SCMP reported that Chinese cyber authorities found Kimi had accessed data irrelevant to its functions. SR015
CR014 OECD.AI logged a 2026 incident in which Kimi disclosed one user’s private resume to another and noted that legal action was underway. SR016
CR015 Moonshot’s privacy policy identifies Moonshot AI PTE. LTD. as the provider and controller of the website, app, and browser-extension services. SR002
CR016 Moonshot’s terms say users contract with a Singapore company and that disputes are governed by Singapore law and SIAC arbitration. SR001
CR017 Moonshot’s privacy policy says user content includes prompts, audio, images, videos, files, and generated content. SR002
CR018 Moonshot’s privacy policy says user content may be used to provide and improve the service, including training and optimizing models depending on jurisdiction. SR002
CR019 Moonshot’s privacy policy says log and usage data can include device identifiers, conversation IDs, interaction patterns, and clipboard data where permitted by settings. SR002
CR020 Moonshot’s privacy policy says personal information may be shared with service providers, affiliates, and public authorities under stated conditions. SR002
CR021 Moonshot’s privacy policy says transaction information may be retained after account deletion as necessary for legal, financial, and operational obligations. SR002
CR022 Moonshot’s terms say payments are generally non-refundable and that app-store billing, cancellation, and refund policies are controlled by the distributor for app purchases. SR001, SR006
CR023 Moonshot’s terms reserve the right to suspend or terminate access for legal or regulatory compliance, harmful activity, or misuse. SR001
CR024 Moonshot’s terms say users may opt out of allowing their content to be used for model improvement or research by contacting Moonshot. SR001
CR025 Moonshot’s terms prohibit automated crawling, prompt injection, competitive model development, and uploading business data that the user lacks legal rights to use. SR001
CR026 SCMP reported that Kimi’s app and website crashed for hours on March 21 because of overload issues. SR019
CR027 Kimi’s official surfaces now span websites, slides, spreadsheets, deep research, Kimi Claw, and agent-swarm workflows, widening Moonshot’s operational scope. SR004, SR005, SR006
CR028 Moonshot’s GitHub organization shows 38 repositories with active June 2026 updates across code, CLI, agent SDK, help-center, and infrastructure projects. SR023
CR029 The Kimi K2 repository says the model is designed for tool use and autonomous problem-solving, increasing the need for tool-governance controls when used in production. SR024, SR005
CR030 IAPS says Kimi Claw is an always-on browser-tab agent that can observe, collect, shape, and act upon nearly everything a user does digitally. SR017
CR031 IAPS argues that the combination of Chinese legal exposure and OpenClaw ecosystem vulnerabilities could make Kimi Claw a larger national-security risk than TikTok-like single-app platforms. SR017
CR032 IAPS catalogs malicious skills, prompt-injection-mediated data exfiltration, and remote-code-execution vulnerabilities in the OpenClaw ecosystem. SR017
CR033 The Hacker News summarized Harmonic Security data showing that nearly 8% of 14,000 sampled US and UK employees had used China-based GenAI tools including Kimi and that 535 sensitive-data incidents were observed. SR018
CR034 The Hacker News said Kimi and peer Chinese GenAI services are often used without security-team approval, widening data-residency and compliance exposure for enterprises. SR018
CR035 CNBC reported that Chinese AI companies are prioritizing user growth and ecosystem integration over headline benchmark wins. SR021
CR036 CNBC reported that Moonshot released K2.5 only about three months after K2 as Chinese AI competition accelerated against U.S. rivals. SR021
CR037 CNBC reported that Moonshot made Kimi K2 free in app and browser while charging API prices below major U.S. rivals. SR020
CR038 Kimi API pricing is usage-based and charges separately for input, output, and web-search invocations, which means heavy agentic use can compound cost exposure. SR007, SR020
CR039 The Kimi API Platform bundles web search, memory, code execution, URL fetch, and file-analysis tools for professional developers, increasing the governance surface beyond simple chat. SR008
CR040 Kimi’s App Store listing shows 190,000 ratings and in-app purchases, implying that any trust or reliability problem can propagate across a very large consumer surface. SR006
CR041 Kimi’s App Store listing identifies a Beijing provider entity while the privacy policy and terms point users to a Singapore company, creating a corporate-structure diligence question. SR001, SR002, SR006, SR017
CR042 White & Case says AI regulation now overlaps IP, data protection, litigation, financial regulation, and global trade, increasing the number of legal fronts Moonshot must manage. SR012
CR043 No retained official source published a public SLA, certification pack, or detailed trust-control packet for Kimi. SR003, SR004, SR005, SR008, SR009
CR044 The retained public product surfaces did not visibly publish a Kimi-specific filing number or a jurisdiction-by-jurisdiction compliance mapping. SR003, SR004, SR005, SR006, SR010, SR011
CR045 No retained official source disclosed a cloud-provider concentration schedule or backup counterparty map for Kimi’s production stack. SR004, SR005, SR008, SR009
CR046 Moonshot’s public legal and privacy materials remain high-level rather than module-specific on retention, logging, and red-team detail. SR001, SR002, SR004, SR005
CR047 Moonshot’s terms say output is not professional advice and may not be used for high-stakes decisions about identifiable people. SR001
CR048 Moonshot’s privacy policy says it may obtain publicly available information from websites, datasets, and open forums to improve and train models. SR002
CR049 For app-based subscriptions, Moonshot routes cancellation and refund control through the app distributor, which can complicate unified customer remediation. SR001, SR006
CR050 AICPB and SEMrush both show that Kimi remains a scaled consumer surface in 2026, which increases the blast radius of outages or trust failures. SR015, SR026, SR027, SR028
CR051 Moonshot’s help-center repository has 302 commits and bilingual English and Simplified Chinese trees, underscoring the documentation breadth Moonshot must keep current across the ecosystem. SR030
CR052 TechCrunch described Moonshot’s early differentiation as long-context handling for workflows such as legal documents and deeper financial analysis, making quality failures in those domains especially trust-sensitive. SR022
CR053 SCIO said Chinese authorities issued May 2026 implementation guidelines for AI agents that stress safety, controllability, standardization, and application-driven rollout. SR031
CR054 SCIO said Chinese regulators punished three online platforms in April 2026 for violating AI-generated-content labeling rules, showing active enforcement rather than paper-only regulation. SR032
CV001 TechCrunch reported Moonshot raised about $2 billion at a $20 billion valuation in May 2026. SV001
CV002 TechCrunch reported Moonshot raised $3.9 billion over the prior six months. SV001
CV003 CnTechPost reported Moonshot ARR surpassed $100 million in early March 2026. SV002
CV004 TechCrunch reported Moonshot ARR topped $200 million in April 2026, driven by subscriptions and API usage. SV001
CV005 A $20 billion valuation against the reported April ARR level implies a rough 100x ARR multiple if the $200 million figure is annualized rather than substantially exceeded. SV001
CV006 Caixin reported Moonshot held more than 10 billion yuan of cash after its late-2025 $500 million Series C, which reduces near-term financing pressure even if valuation support remains debatable. SV004
CV007 CnTechPost reported Moonshot is dismantling its red-chip structure because a waiver to preserve the VIE setup now looks unlikely. SV003
CV008 The Wall Street Journal reported a private-funding discussion around an approximately $18 billion valuation while Moonshot pursued Hong Kong listing preparation under heightened scrutiny. SV005
CV009 Moonshot moved from reported valuation anchors of about $2.2-$2.5 billion in early 2024 to over $3.3 billion in August 2024, then to $18-$20 billion by 2026. SV018, SV019, SV020, SV001, SV005
CV010 OpenAI announced $40 billion of new funding at a $300 billion post-money valuation in March 2025. SV006
CV011 OpenAI announced $122 billion in committed capital at an $852 billion post-money valuation in March 2026. SV007
CV012 OpenAI said it was generating $2 billion of revenue per month in 2026, implying roughly $24 billion of annualized revenue at the time of the $852 billion funding round. SV007
CV013 Anthropic's Series G valued the company at $380 billion on $14 billion run-rate revenue. SV008
CV014 Anthropic's Series H valued the company at $965 billion on $47 billion run-rate revenue. SV009
CV015 Yicai reported Zhipu AI closed its Hong Kong debut with a market capitalization of HK$57.5 billion, or about $7.4 billion. SV010
CV016 Yicai reported Zhipu AI raised over HK$4.3 billion in its IPO and earmarked 70% of proceeds for general-purpose AI model R&D. SV010, SV029
CV017 Yicai reported Zhipu AI generated CNY190.9 million of first-half 2025 revenue and a CNY2.4 billion net loss, showing that public AI valuations in China can remain rich even with steep losses. SV010
CV018 Reuters reported via Yahoo Finance that MiniMax raised HK$4.82 billion, or about $619 million, in its Hong Kong IPO at HK$165 per share. SV011
CV019 TechNode reported MiniMax briefly exceeded an $11.5 billion market capitalization on its trading debut. SV012
CV020 TechNode reported MiniMax remained in a high-investment phase with a $512 million net loss in the first three quarters of 2025 and a 69.4% gross margin. SV012
CV021 C3.ai reported FY2026 revenue of $250.3 million in its results release and SEC exhibit. SV013, SV014
CV022 Stock Analysis showed C3.ai at roughly $1.50 billion market cap, $975 million enterprise value, and 3.9x EV/Sales in June 2026. SV015
CV023 Stock Analysis showed Palantir at roughly $307.98 billion market cap, $300.17 billion enterprise value, and 57.46x EV/Sales in June 2026. SV016
CV024 Stock Analysis showed Palantir at $5.22 billion trailing-twelve-month revenue and $1.63 billion Q1 2026 revenue. SV017
CV025 Official K2.6 list pricing is ¥6.50 per 1M uncached input tokens and ¥27.00 per 1M output tokens, with batch jobs priced at 60% of standard and web search at ¥0.03 per call. SV021, SV022, SV023
CV026 Moonshot's product and research homepages position K2.6 as a flagship multimodal, coding, and agent model rather than a single-purpose chatbot. SV024, SV030
CV027 CoreWeave said independent testing priced K2.6 at about $0.7 per million blended tokens at 205 tokens per second, showing how quickly third-party infrastructure can compress or repackage inference economics. SV026
CV028 VentureBeat reported a benchmark where the official Kimi endpoint took 163.7 seconds to complete a standard coding request while Cerebras completed it in 5.6 seconds. SV025
CV029 Moonshot's current private mark appears to depend more on scarcity premium, China AI capital-market enthusiasm, and expected future scale than on disclosed current fundamentals. SV001, SV010, SV012, SV015, SV016
CV030 The strongest thesis is that Moonshot is the fastest-monetizing Chinese open-weight frontier lab, with fresh capital, official pricing, and enough product traction to sustain a public-listing option. SV001, SV002, SV024, SV030
CV031 The strongest anti-thesis is that the 2026 valuation already discounts aggressive revenue scaling while margin quality, preference stack, and IPO execution remain opaque. SV001, SV003, SV005, SV015
CV032 A supportable recommendation at the current evidence set is research-more rather than buy, because the business looks real but the current price relies on too many undisclosed variables. SV001, SV004, SV014, SV015, SV016
CV033 Confidence in that recommendation should be medium rather than low because multiple fresh sources corroborate scale and financing, but key valuation mechanics remain unverified. SV001, SV004, SV010, SV011, SV014
CV034 The right valuation stance is stretched: not impossible in the current China AI market, but demanding enough that future upside depends on continued hypergrowth and successful listing execution. SV001, SV010, SV012, SV015, SV016
CV035 A reasonable bull case is $27.5-$35.0 billion if Moonshot can compound ARR toward roughly $500 million and preserve a 55-70x scarcity premium into a listing window. SV001, SV010, SV012, SV016
CV036 A reasonable base case is $17.5-$21.0 billion if Moonshot can sustain ARR around $350 million and hold a 50-60x market premium, leaving today's mark roughly fair but not attractive. SV001, SV010, SV015, SV016
CV037 A reasonable bear case is $7.5-$10.0 billion if ARR stalls near $250 million and the multiple compresses toward 30-40x amid regulatory delay or weaker monetization. SV001, SV003, SV005, SV015
CV038 The main thesis-break trigger is a failed or materially delayed Hong Kong listing process tied to structure cleanup or regulatory scrutiny. SV003, SV005
CV039 A second thesis-break trigger is revenue or ARR stalling below the level required to make the current valuation fair, especially if public evidence remains capped near the April 2026 ARR figure. SV001, SV002
CV040 A third thesis-break trigger is a future financing below the current implied private mark, which would reveal that scarcity premium outran fundamentals. SV001, SV005
CV041 The most important diligence ask is an audited monthly revenue bridge across subscriptions, API, and enterprise prepaids.
CV042 A second critical diligence ask is the cap table, preference stack, and rights package for the 2025 and 2026 rounds.
CV043 A third critical diligence ask is a gross-margin bridge by model, hosted-versus-partner routing, and compute-procurement commitments.
CV044 A milestone that would move the call more positive is evidence that ARR has moved well beyond the April 2026 level while the company preserves IPO eligibility and margin discipline. SV001, SV003, SV005
CV045 Moonshot is not yet fully exit-ready on public evidence because audited financials, preference terms, and the IPO timetable remain unavailable. SV003, SV004, SV005
来源
编号出版方标题引文
SO001 Moonshot AI Moonshot AI homepage
SO002 Moonshot AI Kimi homepage
SO003 Moonshot AI Kimi pricing overview
SO004 Moonshot AI Kimi Code product page
SO005 Moonshot AI Kimi Terms of Service
SO006 Moonshot AI Kimi Privacy Policy
SO007 TechCrunch Moonshot AI reportedly raises over $1B at a $2.5B valuation
SO008 TechCrunch China’s Moonshot AI raises $2B at $20B valuation
SO009 Forbes Kimi is closing a $2 billion funding round at a $20 billion valuation
SO010 South China Morning Post Meet Yang Zhilin, the Moonshot AI founder
SO011 Caixin Global Moonshot AI rules out quick IPO after raising $500 million
SO012 CNBC Moonshot releases Kimi K2
SO013 CNBC Moonshot releases Kimi K2 Thinking
SO014 CNBC Chinese tech companies accelerate AI model rollouts
SO015 CNBC Anthropic flags distillation campaigns by Chinese AI firms
SO016 Anthropic Detecting and preventing distillation attacks
SO017 OECD AI Incidents Monitor Kimi large language model leaked user resume data
SO018 MoonshotAI GitHub - Kimi k1.5
SO019 arXiv Kimi k1.5: Scaling Reinforcement Learning with LLMs
SO020 MoonshotAI GitHub - Kimi-K2
SO021 Hugging Face moonshotai/Kimi-K2-Instruct model card
SO022 arXiv Kimi K2 technical report
SO023 MoonshotAI GitHub - Kimi-VL
SO024 arXiv Kimi-VL Technical Report
SO025 Moonshot AI Kimi open platform blog overview
SM001 Moonshot AI Moonshot AI
SM002 Kimi Kimi AI 官网 - K2.6 上线
SM003 Kimi API 开放平台 多模态模型 Kimi K2.6 定价 - Kimi API 开放平台
SM004 Kimi API 开放平台 编程模型 Kimi K2.7 Code 定价 - Kimi API 开放平台
SM005 Kimi API 开放平台 Kimi API 开放平台
SM006 Kimi API 开放平台 模型列表 - Kimi API 开放平台
SM007 Kimi API 开放平台 从 OpenAI 迁移到 Kimi API - Kimi API 开放平台
SM008 Kimi API 开放平台 联网搜索定价 - Kimi API 开放平台
SM009 AICPB AICPB – The Global Standard for AI Rankings | AI Apps, AI Websites & AI Models
SM010 AICPB China AI Rankings by Users — Apr 2026 Edition | AICPB
SM011 AICPB AI ChatBot Rankings by Users — Apr 2026 Edition | AICPB
SM012 OpenAI Help Center OpenAI API - Supported Countries and Territories | OpenAI Help Center
SM013 中国网信网 生成式人工智能服务管理暂行办法_中央网络安全和信息化委员会办公室
SM014 中国网信网 四部门联合发布《人工智能生成合成内容标识办法》_中央网络安全和信息化委员会办公室
SM015 中国网信网 关于印发《人工智能生成合成内容标识办法》的通知_中央网络安全和信息化委员会办公室
SM016 TechCrunch China's Moonshot AI zooms to $2.5B valuation, raising $1B for an LLM focused on long context | TechCrunch
SM017 CNBC One year after DeepSeek, Chinese AI firms from Alibaba to Moonshot race to release new models
SM018 Caixin Global Cover Story: Chinese AI Startups Make Gains in Challenge to U.S.-based OpenAI
SM019 OpenAI Pricing | OpenAI API
SM020 Anthropic Plans & Pricing | Claude by Anthropic
SM021 Google Cloud Agent Platform Pricing | Google Cloud
SM022 IDC China AI: Redefining Global Competition for the Agentic Future (IDC FutureScape 2026 excerpt)
SM023 DeepSeek DeepSeek | 深度求索
SM024 DeepSeek API Docs 首次调用 API | DeepSeek API Docs
SM025 DeepSeek API Docs 模型 & 价格 | DeepSeek API Docs
SM026 Alibaba Cloud OpenAI-Compatible Qwen & Multimodal Model Access - Model Studio - Alibaba Cloud
SM027 Alibaba Cloud Supported Models and Capabilities Overview - Model Studio - Alibaba Cloud
SM028 MiniMax MiniMax
SP001 Moonshot AI Moonshot AI
SP002 Kimi Kimi AI 官网 - K2.6 上线
SP003 Kimi API 开放平台 多模态模型 Kimi K2.6 定价 - Kimi API 开放平台
SP004 Kimi API 开放平台 编程模型 Kimi K2.7 Code 定价 - Kimi API 开放平台
SP005 Kimi API 开放平台 模型列表 - Kimi API 开放平台
SP006 Kimi API 开放平台 从 OpenAI 迁移到 Kimi API - Kimi API 开放平台
SP007 AICPB China AI Rankings by Users — Apr 2026 Edition | AICPB
SP008 AICPB AI ChatBot Rankings by Users — Apr 2026 Edition | AICPB
SP009 CNBC One year after DeepSeek, Chinese AI firms from Alibaba to Moonshot race to release new models
SP010 CNBC China's tech giants enter 'agentic commerce' race
SP011 TechCrunch China's Moonshot AI zooms to $2.5B valuation, raising $1B for an LLM focused on long context | TechCrunch
SP012 OpenAI Pricing | OpenAI API
SP013 Anthropic Plans & Pricing | Claude by Anthropic
SP014 Google Cloud Agent Platform Pricing | Google Cloud
SP015 Alibaba Cloud OpenAI-Compatible Qwen & Multimodal Model Access - Model Studio - Alibaba Cloud
SP016 Alibaba Cloud Supported Models and Capabilities Overview - Model Studio - Alibaba Cloud
SP017 MiniMax MiniMax
SP018 IDC China AI: Redefining Global Competition for the Agentic Future (IDC FutureScape 2026 excerpt)
SP019 DeepSeek DeepSeek | 深度求索
SP020 DeepSeek API Docs 首次调用 API | DeepSeek API Docs
SP021 DeepSeek API Docs 模型 & 价格 | DeepSeek API Docs
SP022 智谱AI开放文档 平台介绍 - 智谱AI开放文档
SP023 Z.ai New Released - Overview - Z.AI DEVELOPER DOCUMENT
SP024 Z.ai Z.ai - Advanced AI Chatbot & Agent powered by GLM-5.2
SP025 火山引擎 火山引擎-你的AI云
SP026 Doubao Doubao App
SP027 百度智能云 千帆大模型平台-企业级一站式大模型开发及应用开发平台-百度智能云
SP028 百度智能云 百度千帆·大模型服务及Agent开发平台 -百度智能云
SP029 MiniMax API Docs Quick Start - MiniMax API Docs
SI001 TechCrunch China's Moonshot AI zooms to $2.5B valuation, raising $1B for an LLM focused on long context Moonshot AI has raised over $1 billion in a Series B round ... value Moonshot AI at $2.5 billion.
SI002 South China Morning Post Alibaba emerges as major backer of high-flying Chinese start-up Moonshot AI Alibaba has invested a total of approximately US$0.8 billion ... for an approximately 36 per cent equity interest.
SI003 TechNode China’s Moonshot AI reportedly raising several hundred million dollars in new funding round Moonshot AI last raised about $300 million in August 2024 at a post-money valuation of over $3.3 billion.
SI004 CnTechPost Kimi maker Moonshot's annual recurring revenue tops $100 million after K2.5 launch In early March ... Moonshot's ARR surpassed $100 million.
SI005 CnTechPost Kimi creator Moonshot revamps corporate structure to prepare for Hong Kong IPO Moonshot is raising about $2 billion in its ongoing funding round, bringing its valuation to more than $20 billion.
SI006 TechCrunch China's Moonshot AI raises $2B at $20B valuation as demand for open source AI skyrockets Moonshot’s annual recurring revenue topped $200 million in April, driven by rapid growth in paid subscriptions and API usage.
SI007 Caixin Global Moonshot AI Rules Out Quick IPO After Raising $500 Million Moonshot AI raised $500 million ... now holding over $1.4 billion in cash.
SI008 The Wall Street Journal China’s Moonshot AI Seeks Listing in Hong Kong Under Heightened Scrutiny The company is raising a new round of private funding that would value it at around $18 billion.
SI009 Kimi API Open Platform 模型推理价格说明 Chat Completion 接口收费:我们对 Input 和 Output 均实行按量计费。
SI010 Kimi API Open Platform 多模态模型 Kimi K2.6 定价 ["kimi-k2.6", "1M tokens", "¥1.10", "¥6.50", "¥27.00", "262,144 tokens"]
SI011 Kimi API Open Platform 编程模型 Kimi K2.7 Code 定价 ["kimi-k2.7-code", "1M tokens", "¥1.30", "¥6.50", "¥27.00", "262,144 tokens"]
SI012 Kimi API Open Platform 生成模型 Moonshot V1 定价 ["moonshot-v1-8k", "1M tokens", "¥2.00", "¥10.00", "8,192 tokens"]
SI013 Kimi API Open Platform 批量推理定价 Batch API 即批量推理 API,批量推理 API 费用为标准模型价格的 60%。
SI014 Kimi API Open Platform 联网搜索定价 ["联网搜索", "1 次", "¥0.03", ...]
SI015 Moonshot AI Moonshot AI Our research team works toward AGI while sharing the latest research with the global open-source community.
SI016 Kimi Kimi AI with K2.6 | Better Coding, Smarter Agents K2.6 is a natively multimodal model, powerful coding capabilities, and Agent performance.
SI017 U.S. Securities and Exchange Commission C3 AI Announces Fiscal Fourth Quarter and Full Fiscal Year 2026 Results (EX-99.1) Total Revenue was $250.3 million ... Gross margin 31% ... cash, cash equivalents, and marketable securities was $575.4 million.
SI018 C3 AI C3 AI Announces Fiscal Fourth Quarter and Full Fiscal Year 2026 Results The sales performance over recent quarters has been entirely unacceptable, to the point of surreal.
SI019 Stock Analysis C3.ai (AI) Statistics & Valuation C3.ai has a market cap ... $1.50 billion. The enterprise value is $975.44 million. EV / Sales 3.90.
SI020 OpenAI New funding to build towards AGI Today we’re announcing new funding—$40 billion at a $300 billion post-money valuation.
SI021 OpenAI OpenAI raises $122 billion to accelerate the next phase of AI Today, we closed our latest funding round with $122 billion in committed capital at a post money valuation of $852 billion.
SI022 Anthropic Anthropic raises $30 billion in Series G funding at $380 billion post-money valuation Today, our run-rate revenue is $14 billion.
SI023 Anthropic Anthropic raises $65B in Series H funding at $965B post-money valuation Our run-rate revenue crossed $47 billion earlier this month.
SI024 CoreWeave CoreWeave Leads Kimi K2.6 Inference Benchmarks CoreWeave ... delivering 205 token/sec at $0.7 per million tokens blended price.
SI025 VentureBeat Cerebras says its chips run a trillion-parameter AI model nearly 7 times faster than GPU clouds For a standard agentic coding request ... 163.7 seconds on the official Kimi endpoint.
SE001 Moonshot AI Moonshot AI homepage
SE002 Moonshot AI Kimi homepage
SE003 Moonshot AI Kimi open platform homepage
SE004 Moonshot AI Start using Kimi API
SE005 Moonshot AI Kimi API models page
SE006 Moonshot AI Kimi K2.6 pricing and model page
SE007 Moonshot AI Kimi K2.7 Code pricing and model page
SE008 Moonshot AI Kimi pricing overview
SE009 Moonshot AI Kimi K2.6 quickstart
SE010 Moonshot AI Kimi K2.7 Code quickstart
SE011 Moonshot AI Using thinking models
SE012 Moonshot AI Using Kimi vision models
SE013 Moonshot AI Kimi API llms.txt index
SE014 Moonshot AI Kimi open platform blog overview
SE015 MoonshotAI GitHub - Kimi k1.5
SE016 arXiv Kimi k1.5: Scaling Reinforcement Learning with LLMs
SE017 MoonshotAI GitHub - Kimi-K2
SE018 Hugging Face moonshotai/Kimi-K2-Instruct model card
SE019 arXiv Kimi K2 technical report
SE020 MoonshotAI GitHub - Kimi-VL
SE021 Hugging Face moonshotai/Kimi-VL-A3B-Instruct model card
SE022 arXiv Kimi-VL Technical Report
SE023 kvcache-ai GitHub - Mooncake
SE024 arXiv Mooncake: A KVCache-centric Disaggregated Architecture for LLM Serving
SE025 USENIX FAST Mooncake best-paper presentation
SE026 MoonshotAI GitHub - checkpoint-engine
SE027 Moonshot AI Kimi Code product page
SE028 Moonshot AI Kimi Terms of Service
SE029 Moonshot AI Kimi Privacy Policy
SE030 OECD AI Incidents Monitor Kimi large language model leaked user resume data
SE031 Anthropic Detecting and preventing distillation attacks
SE032 CNBC Anthropic flags distillation campaigns by Chinese AI firms
SE033 CNBC Moonshot releases Kimi K2
SE034 CNBC Moonshot releases Kimi K2 Thinking
SE035 CNBC Chinese tech companies accelerate AI model rollouts
SE036 NVIDIA Kimi K2.6 model card on NVIDIA NIM
SE037 Moonshot AI Kimi Agent Swarm blog post
SE038 Moonshot AI WorldVQA research post
SU001 Kimi Kimi AI with K2.6 | Better Coding, Smarter Agents
SU002 Moonshot AI Moonshot AI
SU003 Kimi Help Center Home | Kimi Help Center
SU004 Kimi Help Center API pricing - Kimi Help Center
SU005 Kimi Kimi K2.6 | Leading Open-Source Model in Coding & Agent
SU006 Kimi API Platform Kimi API Platform
SU007 Apple App Store Kimi App - App Store 4.9 满分 5 分;19万 个评分。
SU008 Kimi Terms of Service
SU009 Kimi Privacy Policy
SU010 South China Morning Post Moonshot AI’s Kimi Chatbot offers paid service in bid to profit from mass users Moonshot AI is offering six tiers of “top-up” plans, ranging from 5.2 yuan for four days to 399 yuan for a year of “priority use”.
SU011 TechCrunch China’s Moonshot AI zooms to $2.5B valuation, raising $1B for an LLM focused on long context
SU012 CNBC Alibaba-backed Moonshot releases new Kimi AI model that beats ChatGPT, Claude in coding — and it costs less
SU013 CNBC One year after DeepSeek, Chinese AI firms from Alibaba to Moonshot race to release new models
SU014 AICPB AICPB – The Global Standard for AI Rankings
SU015 AICPB AI App Rankings by App MAU — Issue 21 (Apr 2026 Edition)
SU016 SEMrush kimi.com Website Traffic, Ranking, Analytics [May 2026]
SU017 Sensor Tower Kimi - Apple App Store - China - Category Rankings, Keyword Rankings, Sales Rankings, Research, Performance, and Growth Metrics
SU018 GitHub Moonshot AI
SU019 GitHub GitHub - MoonshotAI/Kimi-K2
SU020 Hugging Face moonshotai (Moonshot AI)
SU021 South China Morning Post China accuses “AI tigers” Zhipu, Moonshot of collecting excessive data Moonshot’s Kimi had accessed data irrelevant to its functions.
SU022 Institute for AI Policy and Strategy Kimi Claw: Risks from Chinese-Hosted “Always On” AI Agents
SU023 OECD.AI Kimi AI Model Leaks User Resume Data, Causing Privacy Breach in China The Kimi large language model mistakenly disclosed a user's private resume to another user during a routine task.
SU024 Cyberspace Administration of China 生成式人工智能服务管理暂行办法
SU025 White & Case AI Watch: Global regulatory tracker - China
SR001 Kimi Terms of Service
SR002 Kimi Privacy Policy User Content includes prompts, audio, images, videos, files, and any content you input or generate while using our products and services.
SR003 Kimi Kimi AI with K2.6 | Better Coding, Smarter Agents
SR004 Kimi Help Center Home | Kimi Help Center
SR005 Kimi Kimi K2.6 | Leading Open-Source Model in Coding & Agent
SR006 Apple App Store Kimi App - App Store
SR007 Kimi Help Center API pricing - Kimi Help Center
SR008 Kimi API Platform Kimi API Platform
SR009 Moonshot AI Moonshot AI
SR010 Cyberspace Administration of China 生成式人工智能服务管理暂行办法
SR011 Cyberspace Administration of China 国家互联网信息办公室关于发布2025年生成式人工智能服务已备案信息的公告
SR012 White & Case AI Watch: Global regulatory tracker - China
SR013 State Council of the People’s Republic of China China to formulate over 50 standards for AI sector by 2026
SR014 Ministry of Justice of the People’s Republic of China China to refine AI-related legal framework
SR015 South China Morning Post China accuses “AI tigers” Zhipu, Moonshot of collecting excessive data Moonshot’s Kimi had accessed data irrelevant to its functions.
SR016 OECD.AI Kimi AI Model Leaks User Resume Data, Causing Privacy Breach in China
SR017 Institute for AI Policy and Strategy Kimi Claw: Risks from Chinese-Hosted “Always On” AI Agents
SR018 The Hacker News Overcoming Risks from Chinese GenAI Tool Usage
SR019 South China Morning Post Moonshot AI’s Kimi Chatbot offers paid service in bid to profit from mass users
SR020 CNBC Alibaba-backed Moonshot releases new Kimi AI model that beats ChatGPT, Claude in coding — and it costs less
SR021 CNBC One year after DeepSeek, Chinese AI firms from Alibaba to Moonshot race to release new models
SR022 TechCrunch China’s Moonshot AI zooms to $2.5B valuation, raising $1B for an LLM focused on long context
SR023 GitHub Moonshot AI
SR024 GitHub GitHub - MoonshotAI/Kimi-K2
SR025 Hugging Face moonshotai (Moonshot AI)
SR026 AICPB AICPB – The Global Standard for AI Rankings
SR027 AICPB AI App Rankings by App MAU — Issue 21 (Apr 2026 Edition)
SR028 SEMrush kimi.com Website Traffic, Ranking, Analytics [May 2026]
SR029 Sensor Tower Kimi - Apple App Store - China - Category Rankings, Keyword Rankings, Sales Rankings, Research, Performance, and Growth Metrics
SR030 GitHub GitHub - MoonshotAI/kimi-help-center
SR031 State Council Information Office China unveils guidelines to regulate, boost innovative development of AI agents
SR032 State Council Information Office Chinese internet platforms punished for AI-generated content labeling violations
SR033 Digital China Summit 人工智能高质量发展的制度基石与行动指引
SV001 TechCrunch China's Moonshot AI raises $2B at $20B valuation as demand for open source AI skyrockets Moonshot’s annual recurring revenue topped $200 million in April.
SV002 CnTechPost Kimi maker Moonshot's annual recurring revenue tops $100 million after K2.5 launch Moonshot's ARR surpassed $100 million.
SV003 CnTechPost Kimi creator Moonshot revamps corporate structure to prepare for Hong Kong IPO The AI startup plans to dismantle its red-chip structure to meet regulatory requirements.
SV004 Caixin Global Moonshot AI Rules Out Quick IPO After Raising $500 Million Moonshot AI now holds more than 10 billion yuan ($1.4 billion) in cash.
SV005 The Wall Street Journal China’s Moonshot AI Seeks Listing in Hong Kong Under Heightened Scrutiny The company is raising a new round of private funding that would value it at around $18 billion.
SV006 OpenAI New funding to build towards AGI Today we’re announcing new funding—$40 billion at a $300 billion post-money valuation.
SV007 OpenAI OpenAI raises $122 billion to accelerate the next phase of AI Today, we closed our latest funding round with $122 billion in committed capital at a post money valuation of $852 billion.
SV008 Anthropic Anthropic raises $30 billion in Series G funding at $380 billion post-money valuation Today, our run-rate revenue is $14 billion.
SV009 Anthropic Anthropic raises $65B in Series H funding at $965B post-money valuation Our run-rate revenue crossed $47 billion earlier this month.
SV010 Yicai Global Zhipu AI Soars in Hong Kong Stock Market Debut as Chinese Startup Becomes World's First LLM Firm to Go Public Zhipu AI ... bringing its market capitalization to HKD57.5 billion (USD7.4 billion).
SV011 Reuters / Yahoo Finance China's AI startup MiniMax Group raises $619 million in Hong Kong IPO MiniMax Group said ... it raised HK$4.82 billion ($618.60 million) in its Hong Kong initial public offering.
SV012 TechNode MiHoYo-backed AI firm MiniMax jumps on Hong Kong debut MiniMax ... briefly pushing the company’s market capitalisation above HK$90 billion ($11.5 billion).
SV013 C3 AI C3 AI Announces Fiscal Fourth Quarter and Full Fiscal Year 2026 Results Total Revenue was $250.3 million.
SV014 U.S. Securities and Exchange Commission C3 AI Announces Fiscal Fourth Quarter and Full Fiscal Year 2026 Results (EX-99.1) Total Revenue was $250.3 million ... cash ... $575.4 million.
SV015 Stock Analysis C3.ai (AI) Statistics & Valuation C3.ai has a market cap ... $1.50 billion. The enterprise value is $975.44 million. EV / Sales 3.90.
SV016 Stock Analysis Palantir Technologies (PLTR) Statistics & Valuation Palantir has a market cap ... $307.98 billion. EV / Sales 57.46.
SV017 Stock Analysis Palantir Technologies (PLTR) Revenue 2018-2026 Palantir had revenue of $1.63B in the quarter ending March 31, 2026 ... TTM $5.22B.
SV018 South China Morning Post Alibaba emerges as major backer of high-flying Chinese start-up Moonshot AI That values Beijing-based Moonshot AI at approximately US$2.2 billion.
SV019 TechCrunch China's Moonshot AI zooms to $2.5B valuation, raising $1B for an LLM focused on long context Moonshot AI has raised over $1 billion ... value Moonshot AI at $2.5 billion.
SV020 TechNode China’s Moonshot AI reportedly raising several hundred million dollars in new funding round Moonshot AI last raised about $300 million in August 2024 at a post-money valuation of over $3.3 billion.
SV021 Kimi API Open Platform 多模态模型 Kimi K2.6 定价 ["kimi-k2.6", "1M tokens", "¥1.10", "¥6.50", "¥27.00", "262,144 tokens"]
SV022 Kimi API Open Platform 批量推理定价 Batch API ... priced at 60% of standard model price.
SV023 Kimi API Open Platform 联网搜索定价 联网搜索 ... ¥0.03
SV024 Moonshot AI Moonshot AI Our research team works toward AGI while sharing the latest research with the global open-source community.
SV025 VentureBeat Cerebras says its chips run a trillion-parameter AI model nearly 7 times faster than GPU clouds 163.7 seconds on the official Kimi endpoint.
SV026 CoreWeave CoreWeave Leads Kimi K2.6 Inference Benchmarks 205 token/sec at $0.7 per million tokens blended price.
SV027 Pandaily Kimi Nears $600 Million Funding Round, IDG Reportedly to Join Kimi Nears $600 Million Funding Round, IDG Reportedly to Join
SV028 Pandaily Kimi Operator Moonshot AI Valued at $20B+ After $2B Funding Round Kimi Operator Moonshot AI Valued at $20B+ After $2B Funding Round
SV029 Pandaily Zhipu AI Launches Hong Kong IPO With HK$3 Billion in Cornerstone Commitments, Poised to Be 2026’s Largest Opening IPO Zhipu AI Launches Hong Kong IPO With HK$3 Billion in Cornerstone Commitments
SV030 Kimi Kimi AI with K2.6 | Better Coding, Smarter Agents K2.6 is a natively multimodal model, powerful coding capabilities, and Agent performance.