初创公司尽调
尽调报告 AI foundation models (large language and multimodal models) Series B+ / pre-IPO 2026-07-21

StepStar

StepStar (StepFun / 阶跃星辰) 尽调报告

StepStar 技术底子扎实、资金充足,是中国基础模型 “Big Six” 成员;多模态和 AI 终端打法有辨识度,但财务不披露、估值标记争议大且快速抬升,结论仍是继续研究,不是买入。

封面要素

成立时间 01
2023 year [CO001]
最新披露轮次 02
Series B+ (>RMB5B / ~$717M) [CO019, CV003]
据报道 IPO / IPO 前估值 03
~$10-12B (contested) USD [CV008, CV012]
累计融资 04
RMB5B+ (~$717M B+ alone) [CO019]
公开收入披露 05
Not audited; media-estimated ~RMB500M (2025) [CO039]
当前建议 06
research-more / track [CV045]

公司概况

StepStar 以 StepFun(阶跃星辰)为品牌,注册主体为 Shanghai Jieyue Xingchen Intelligent Technology Co., Ltd.,是一家中国 AI 基础模型公司。公司由前 Microsoft 全球副总裁 Jiang Daxin 于 2023 年 4 月 6 日创立,技术联合创始人为 Zhu Yibo 和 Jiao Binxing。公司开发大语言与多模态模型,包括万亿参数 Step-2 MoE 模型和 321B 参数 Step-3 多模态模型,同时推进 StepAI 智能体平台和 STEPX「AI + 终端」设备战略。市场通常把它归入中国基础模型创业公司的「大六家 / AI 六小虎」。公司在 2024 年 12 月以约 $1 billion 估值完成 Series B,并在 2026 年 1 月完成超过 RMB5 billion(~$717M)的 Series B+,同时准备 2026 年赴港 IPO。详细财务数据——收入、利润率、烧钱速度、客户数量——尚未公开披露。

官网
www.stepfun.com
成立时间
2023-04-06
创始人
Jiang Daxin, Zhu Yibo, Jiao Binxing
创立地点
Shanghai, China
总部
Shanghai, China (with Beijing and Hangzhou presence)
产品
Step-2(万亿参数 MoE LLM)、Step-3 / Step-3.5 Flash 多模态模型、Step-1V 视觉模型、StepAI 智能体平台、兼容 OpenAI 的 API、在 Hugging Face/GitHub 发布的开放权重,以及 STEPX 品牌 / Step AOS 智能体手机(STEPX Neo)。
客户
企业客户、通过 API 接入的开发者、OEM 设备与汽车伙伴(OPPO、Honor、ZTE、Geely),以及通过 STEPX 设备和 StepAI 应用触达的消费者。
商业模式
模型 / API 访问、面向 OEM 与汽车伙伴的端侧模型授权,以及正在形成的消费者「AI + 终端」硬件 / 智能体打法;变现经济性尚未公开披露。
阶段
Series B+ / pre-IPO
融资情况
2024 年 12 月按约 $1B 估值完成 Series B 后,2026 年 1 月完成超过 RMB5 billion(~$717M)的 Series B+;正在准备 2026 年赴港 IPO。
[CO001, CO005, CO008, CO009, CO019, CO029, CO039]

执行摘要

主要优势

  • 创始人与研究背景顶尖:前 Microsoft VP Jiang Daxin、IEEE Fellow 领衔;上海国资、Tencent、Qiming 等国资与战略投资人同场入局。
  • 多模态和低成本 MoE 模型族差异化明显:Step-2 万亿参数、Step-3 321B;“AI + 终端”设备策略已触达数千万台 OEM 设备。
  • 2026 年融资通道强:创纪录的 RMB5B+ Series B+,以及已推进到后期的 Hong Kong IPO 路径;Zhipu 和 MiniMax 已先行验证这条路。

主要风险

  • 收入、毛利率、烧钱速度和客户均未公开披露,快速抬升的估值无法像正常软件投资那样由基本面承接。
  • DeepSeek、Qwen、Doubao 和同属 “Big Six” 的创业公司竞争激烈、价格压力大,拥挤市场会挤压变现。
  • 地缘政治与供应链暴露(US chip export controls)叠加中国 AI 监管收紧,推高成本、合规和执行风险。

未决问题

  • 经审计收入、毛利率、算力承诺、烧钱速度和现金跑道均未公开披露。
  • 各轮估值历史在公开来源之间自相矛盾(大致 $4B 到 $12B 标记),股权结构表和优先条款不可得。
  • API 和 STEPX 设备业务的具名企业客户、付费用户数、留存和单位经济性均未公开披露。

目录

Chapter 01

01公司概况

1.1 身份、布局与产品逻辑

本报告用 StepStar 作为 StepFun 的英文尽调标签;StepFun 是 Shanghai Jieyue Xingchen Intelligent Technology Co., Ltd. 的商号。公开登记和公司资料均指向 2023 年 4 月 6 日成立、注册地址在上海徐汇;官方入口则把读者导向 StepFun 官网、开发者平台、Studio 和 Step AI 助手。产品身份不是单一聊天机器人:公司展示的是模型平台、消费者助手、模型文档,以及暗示 API 商业化路径的定价 / 限速页面。官方愿景接近口号:放大每个人的可能,让每个人都强十倍;模型页面重点展示 Step 3.7 Flash、Step 3.5 Flash、Step 2 和 Step 1。已抓取证据支持上海和北京运营信号,但本章未能验证用户要求核查的杭州办公室;应把它列为地点尽调缺口,而不是既成事实。[CO001, CO002, CO003, CO004, CO005, CO006]

快照 KPI 表
指标数值 / 状态日期置信度缺口 / 注意事项
法定名称 / 品牌Shanghai Jieyue Xingchen 智能科技有限公司 / StepFun当前StepStar 是英文尽调别名,不是官方商号
成立April 6, 20232023-04-06除公开登记 / 公司资料来源外,未审阅完整设立文件包
总部上海徐汇登记地址当前未披露确切办公租赁面积
其他地点信号有北京子公司或办公室信号;杭州办公室未核实当前杭州体现为资本 / 投资人敞口,不是已确认办公室
产品模式消费助手 + API 平台 + Step 模型家族2026-07-21未披露企业收入拆分
最新已完成融资轮B+ 轮超过 RMB5B2026-01-26确切持股和优先权条款未公开
早期独角兽标记Series B;Crunchbase 估值 $1B2024-1236Kr 支持轮次规模 / 投资人信息,但不支持全部估值细节
IPO 路径据报道准备香港 IPO,募资约 $500M,估值目标 $10B-$12B2026未审阅招股书或交易所申报
累计融资信号China AI Atlas 列示已披露 / 已宣布融资 >$3.2B2026-07-20数据库式画像,不是经审计股权结构表
员工数估计公开估计约 400-500 人;社保登记下限 214 人2025-2026无法获得准确薪资人数、承包商和地点拆分
收入信号媒体估计 2025 年接近 RMB500M,2026 年预计约 RMB1.2B2026-04-15本文未获得审计或管理层确认
终端规模据报道设备安装 42M+,日服务近 20M2025-end安装、活跃设备和用户的定义未经审计
反向信号公开文件称标的公司处于大额亏损,并承担市场 / 政策 / 经营风险2026-05-27未审阅完整财务报表

快照混合了登记事实、官方产品页面、新闻报道融资、分析师画像估计和一份上市公司文件;无支持指标按缺口列示,而不是当作经审计事实。

[CO001, CO002, CO004, CO005, CO007, CO016]
FO002: 公司快照逻辑

概览投资逻辑把创始人履历、上海与国资支持,同模型 / API 平台和终端分发策略串起来;治理和地域布局仍留有证据缺口。

[CO003, CO004, CO006, CO008, CO009, CO010]
FO003: 快照 KPI

KPI 视角突出融资和分发主张很强,但经审计的经营指标和 IPO 风险仍未解开。

融资、估值、员工数、收入和使用量指标均为公开报道或估计数字,不是经审计的管理层数据;币种单位按报道保留,避免制造虚假精度。

[CO001, CO008, CO014, CO016, CO019, CO022]

1.2 创始人、领导层与治理披露面

公司最值得承销的资产,仍是围绕 Jiang Daxin 的技术领导层。Baidu 和 China AI Atlas 支撑了关键人物叙事:Jiang 在 Microsoft/MSRA 体系约 16 年,做到首席科学家或全球副总裁,并在 2024 年成为 IEEE Fellow。公开创始人及准创始人资料也显示,Zhu Yibo 是 CTO / 系统负责人,具备大规模系统经验;Jiao Binxing 来自 Microsoft Bing,是搜索与数据负责人。2026 年,治理叙事扩展:Yin Qi 据报道并在登记信息中出任董事长 / 法定代表人,Jiang 仍任董事和经理。这有利于运营厚度,但公开材料没有披露董事会投票权、优先股条款、观察员权利、创始人归属安排,或经核对的股东名册。因此,尽调判断是:创始人—市场匹配强,但治理透明度不足。[CO008, CO009, CO010, CO011, CO012, CO013]

领导层和创始人表
人员 / 群体公开角色或状态背景 / 证据创始人-市场匹配或职能覆盖关键人 / 尽调备注
Jiang Daxin创始人 / 联合创始人,CEO,董事兼经理前 Microsoft/MSRA/STCA 负责人;Microsoft 全球副总裁 / 首席科学家;IEEE Fellow 2024深度 NLP、搜索、Bing/Cortana/Azure 和基础模型可信度关键人依赖高;创始人股权和投票权未公开
Zhu Yibo联合创始人 / CTO 或系统负责人Baidu 公司资料和 Jademond 将他与 Microsoft、ByteDance、Google 及大规模系统联系起来覆盖模型基础设施和工程执行具体职衔历史和当前汇报线需管理层确认
Jiao Binxing联合创始人 / 数据或搜索系统负责人Baidu 公司资料和 Jademond 将他与 Microsoft Bing 核心搜索联系起来覆盖数据挖掘、搜索索引和 NLP 质量系统当前股权和运营角色在官网不可见
Yin Qi2026 年登记 / 新闻来源中的董事长及法定代表人AI 创业者 / 运营者;据报道 2026 年 1 月出任董事长围绕创始人主导实验室,增加治理和商业化厚度厘清董事长权限、董事会票数,以及任命是否改变控制权
Zhang XiangyuBaidu 资料中的首席科学家公司资料材料列名为首席科学家在创始人 CEO 之外增加研究领导力需要官方履历、激励和当前雇佣确认
更广泛董事会 / 监事Aiqicha 列示董事和监事Aiqicha 列名 Jiang、Li Jing、Zhang Xiangyu、He Miao、Liu Shanquan、Sun Qian 及监事表明治理层比公开创始人叙事更大未披露股东权利、优先权条款或董事会观察员名单

列举不完整,因为公开来源对创始人和董事的命名并不均匀,也未披露投票控制权或投资人董事会权利。

[CO008, CO009, CO010, CO011, CO012, CO013]

1.3 融资、投资人和 IPO 准备度

StepFun 的融资记录,是本章最清晰的外部验证。36Kr 和 1ai 称,2024 年 12 月 Series B 融资规模为数亿美元,参与方包括上海国资、Tencent、FiveYuan/Wuyuan Capital 和 Qiming Venture Partners;Crunchbase 另将 StepStar 列入 2024 年 12 月估值 $1 billion 的独角兽。2026 年 1 月 B+ 轮佐证更强:Eastmoney 和 AIBase 报道融资超过 RMB5 billion,由上海国资和产业资本领投,Tencent、Qiming、FiveYuan 跟投。到 2026 年春天,叙事从私募融资转向上市准备:Yicai、The Standard、Reuters 镜像、Tencent News 等来源提到红筹拆除、可能的 pre-IPO 融资,以及募资约 US$500 million、估值 US$10 billion 至 US$12 billion 的香港 IPO 目标。它带来强资本市场选项,也带来监管结构风险和估值—披露张力。[CO017, CO018, CO019, CO020, CO021, CO022]

利益相关方或投资人地图
利益相关方角色控制权或经济重要性重要性证据尽调问题
Shanghai State-owned Capital Investment / 上海国资基金Series B 和 B+ 锚定资本与上海 AI 和终端政策一致的地方政府战略资本36Kr、1ai、Crunchbase、Eastmoney 将国资与这些融资联系起来确认领投实体身份、董事会权利、政策约束和清算优先权
Tencent重复战略投资人 / 老股东贯穿 Series B 和 B+ 的大平台与战略资本信号36Kr、1ai、Eastmoney 和 AIBase 提及 Tencent 参投或跟投支持厘清商业集成、云 / 算力依赖和持股
Qiming Venture Partners重复创投投资人顶级 VC 信号,且很可能是早期私募市场尽调背书方36Kr、1ai、Eastmoney 和 AIBase 在投资人名单中提及 Qiming确认基金实体、轮次进入价格和治理权利
FiveYuan / Wuyuan Capital重复创投投资人融资叙事中的持续股东36Kr/1ai 和 Eastmoney 提及 FiveYuan/Wuyuan 参投确认不同译名是否指同一投资实体
China Life Equity、Pudong Venture Capital、Xuhui Capital、Wuxi Liangxi Fund、Xiamen ITG 等机构B+ 国资和产业资本阵营拓宽资本基础,并强化上海 / 长期资本逻辑Eastmoney 在 RMB5B 以上 B+ 轮中列名这些机构要求披露各投资人分配,以及任何产业政策约束
Huaqin Technology 及其他终端 / 供应链投资人产业投资人和设备生态验证者可能关系到 AI + 终端分发和硬件合作Eastmoney 和 IPO 报道指出 Huaqin 及其他供应链资本区分纯财务投资与实际产品 / 渠道承诺
Lotus Holding / 杭州相关投资人敞口寻求少数股权敞口的上市公司投资人增加文件级反向来源,包括大额亏损风险表述CNINFO 文件披露拟投资和风险警示审阅投资合同、估值条款,以及杭州资本是否改变业务足迹
香港 IPO 投资人 / 基石买家潜在公开市场利益相关方若提交招股书,可能重置估值和披露义务Yicai、The Standard、Reuters 镜像和 IPO 报道讨论上市计划确认申报日期、HKEX 意见、基石定价和募资用途

利益相关方重要性根据公开融资和文件报道推断;未获得股权结构表、董事席位名单、优先股堆叠或老股出售时间表。

[CO017, CO018, CO019, CO020, CO021, CO022]

1.4 里程碑、规模信号与需延续核查的缺口

公开时间线解释了 StepFun 为什么很快进入「大六家」讨论:2023 年成立,早期模型推进很快,2024 年 7 月 WAIC 上完成 Step-2 定位里程碑,2024 年 12 月获独角兽融资,2026 年 1 月完成 B+ 融资并升级董事长,2026 年发布产品 / 模型,同时排出 2026 年 IPO 时间表。规模证明有意义,但不均衡。Eastmoney 报道超过 42 million 的设备模型安装基数、近 20 million 日服务量、手机品牌渗透、Geely 座舱部署,以及与芯片和云厂商的生态合作。Tencent News 也报道了 2025 年和 2026 年收入估计,但这些是未经审计的媒体估算,不是管理层财务数据。最重要的反向来源是 Lotus/CNINFO 文件,它称 StepFun 处于大额亏损状态,并提示市场、政策和运营风险。因此,后续章节应复用身份和融资事实,但独立检验产品—市场匹配、客户质量、经审计收入、毛利率、算力烧钱和 IPO 准备度。第二个需延续核查的问题是证据来源:若干规模数字来自媒体文章或数据库资料,而非发行人认证披露;后续章节应把它们当作待验证线索,而不是最终承销输入。[CO029, CO030, CO031, CO032, CO033, CO034]

里程碑表
日期事件类型金额 / 估值 / 状态参与方含义
2023-04-06公司在上海成立成立Shanghai Jieyue Xingchen / StepFun 设立Jiang Daxin、Zhu Yibo、Jiao Binxing 等创始团队形成标准身份和创始人-市场匹配起点
2023-05-17登记 / 资料来源出现北京子公司或办公室信号治理Beijing Jieyue Xingchen 实体 / 海淀信号StepFun 集团支持北京足迹,但不能验证杭州办公室
2023据资料来源,数月内完成首个 100B 参数模型训练产品早期 Step-1 / 100B 级模型里程碑StepFun 技术团队显示成立后很快跑出模型迭代
2024-07WAIC / Step-2 定位进入公开模型叙事产品Step-2 在公开资料中被描述为万亿参数 MoEStepFun把公司从创业公司画像推到前沿模型竞争者
2024-12Series B 完成,并报道独角兽估值融资数亿美元;Crunchbase 称估值 $1BShanghai State-owned Capital Investment、Tencent、FiveYuan、Qiming 等投资方确立首个标准融资估值标记
2026-01-26B+ 融资完成融资超过 RMB5B;过去 12 个月中国大模型最高融资轮上海国资基金、China Life Equity、Pudong、Xuhui、Wuxi、Xiamen ITG、Huaqin、Tencent、Qiming、FiveYuan为 AI + 终端战略增加长期资本和产业资本支持
2026-01-26Yin Qi 董事长任命随 B+ 融资被报道治理董事长任命 / 法定代表人信号Yin Qi、StepFun 董事会提升商业化 / 治理层,但提出控制权尽调问题
2026-02-02Step 3.5 Flash 模型发布出现在模型目录和官方 / GitHub 材料中产品开源旗舰推理模型;目录中总参数 196-197BStepFun、开发者生态增加融资故事之外可检查的技术实物
2026-04-13Reuters 报道为 IPO 路径拆除红筹 / 离岸架构反向CSRC 审查趋严背景下的 IPO 结构调整StepFun、监管机构、IPO 顾问可能帮助适配境内监管,但会延迟或复杂化时间表
2026-04-15Tencent News 报道 IPO 时间、收入估计及公司无回应反向可能于 2026 年中前后申报;2025/2026 收入估计;无回应Tencent News / StepFun提高披露质量和时间表不确定性
2026-05-27Lotus Holding 文件披露投资及大额亏损风险警示反向投资最高 RMB300M;标的公司大额亏损Lotus Holding、StepFun增加文件级下行证据和风险表述
2026-mid香港 IPO 和 Pre-IPO 融资报道延续融资近 US$2.5B Pre-IPO 轮;$10B-$12B 估值目标;~$500M IPO 募资Yicai、The Standard、Startup Wired、市场来源让 StepFun 从单纯私募融资故事,变成公开市场准备度尽调案例

无法获得准确内部交割日期时,行项目采用公开公告或来源发布时间;未披露内部里程碑和未提交 IPO 文件不纳入。

[CO001, CO005, CO017, CO018, CO019, CO020]
FO001: 公司里程碑时间线

StepFun 从 2023 年 4 月成立,到成为独角兽、拿下创纪录 B+ 轮、发布开源模型产物、推进 IPO 重组,并在大约三年内进入申报级风险披露。

[CO001, CO008, CO012, CO017, CO018, CO019]

1.5 图表

Chapter 02

02市场分析

2.1 市场边界:StepStar 卖的是模型能力,不是整个 AI 经济

StepStar 正确的市场边界,是以中国为中心的基础模型和生成式 AI 技术栈;不是整个 AI 经济,也不只是消费者聊天。StepStar 自家平台展示了面向生产的智能体型 Step 3.7 Flash 模型、无代码模型体验中心、API 使用,以及面向消费电子、内容创作、智能汽车、本地生活、金融、制造、游戏和政府的垂直方案。由此纳入的市场横跨四个变现面:MaaS/API 模型消耗、企业私有或托管部署、消费者设备和应用集成,以及政府或国企数字化服务。应排除的支出包括没有 AI 工作负载的通用云容量、传统分析、非 AI SaaS 和未变现的开源实验。替代品集合异常宽,因为买方可以选择 DeepSeek、Qwen、Doubao、GLM、全球 API 平台、开放权重本地部署,或超大规模云厂商的模型路由器,而不是 StepStar 端点。这个边界很关键:市场可以很大,但 StepStar 可变现的池子只限于中国买方愿意为能力、延迟、合规、集成或分发付费,而不是使用免费开放权重的那一块。[CM001, CM002, CM003, CM004, CM005, CM035]

市场定义表
细分 / 类别纳入支出排除支出买方 / 付款方对 StepStar 的相关性
基础模型 MaaS / API按 token 计费的文本、多模态、推理、向量嵌入、工具使用和智能体 API 消耗未付费开源实验和非 AI API 使用开发者、创业公司、AI 平台团队可直接对标 StepFun 平台 / API 以及 DeepSeek/Qwen 价格基准
企业私有或托管部署私有模型部署、RAG、工作流智能体、数据连接器、治理和支持不含基础模型工作负载的通用 IT 咨询CIO、CTO、业务单元或 AI 平台预算如果 StepStar 证明可靠性和合规性,这是最可信的大型 B2B 变现路径
消费级 AI 应用由模型驱动的聊天、搜索、创作、教育、陪伴、娱乐和应用订阅没有模型收入分成的广告支持流量消费应用所有者和订阅用户如果 StepStar 赢得应用分发或嵌入消费产品,则相关
设备和汽车 AI端侧助手、多模态搜索、座舱、手机、PC 和智能硬件集成没有 AI 模型许可的商品设备销售OEM 产品团队和战略伙伴StepFun 平台列出消费电子和智能汽车解决方案
政府和国企 AI+公共服务助手、城市运营、法律咨询、工业 AI 和国有部门采购没有 LLM 层的通用电子政务软件政府部门、SOE、国资基金政策可以创造需求,但会提高合规和采购复杂度
现状替代品DeepSeek、Qwen、GLM、Doubao、Kimi、全球 API、超大规模云厂商路由、本地开放权重仅非 AI 搜索或传统分析开发者、企业、平台所有者给通用模型访问制造价格压力,并降低切换成本

边界把可变现的模型 / 应用工作负载,与广义 AI 基础设施和未变现开放权重使用区分开;类别彼此重叠,因此不应加总。

[CM001, CM002, CM003, CM004, CM005, CM017]
FM001: 市场规模测算视角

StepStar 可触达池子从全球 AI 宏观市场,收窄到中国基础模型商业化,再落到 StepStar 能覆盖的工作负载。

Grand View 的 China 2025 窄口径数值由公开 2030 收入和 2025-2030 CAGR 反推;广义 AI 数字不能与 GenAI 数字相加。

[CM006, CM007, CM008, CM009, CM010, CM011]

2.2 规模估算方向利好,但数字彼此不兼容

市场规模故事很有吸引力,但数字混乱。口径较窄的 Grand View / Horizon 国家页面称,中国生成式 AI 到 2030 年应达到 $17.6 billion,CAGR 为 39.1%;MarketsandMarkets 则把中国 2025 年生成式 AI 市场定为 $7.0 billion,并预测 2030 年接近 $98.8 billion,CAGR 为 45.8%。Tianxia Gongchang 的中国大模型口径接近这个 2025 年数字,约 RMB49.5-51.0 billion;若纳入 AI 赋能软件,则扩至 RMB100-130 billion。Axis Intelligence 采用更宽的中国 AI 收入口径,把 2025 年放在约 $28-31 billion,并引用更宽的 IDC 式活动代理指标,约 $62 billion。全球口径同样不一致:Gartner 预计 2026 年 AI 支出为 $2.596 trillion,主要由基础设施主导;而 Grand View、Precedence 和 MarketsandMarkets 的 LLM 专项预测集中在个位数十亿到数百亿美元区间。尽调结论不是给这些数字取平均。对 StepStar 而言,TAM 是中国和全球基础模型需求;SAM 是中国 MaaS、私有部署、AI 应用以及设备 / 政府方案支出;SOM 取决于定价权、分发和工作负载份额,而公开来源没有拆出这些变量。[CM006, CM007, CM008, CM009, CM010, CM011]

TAM / SAM / SOM 或规模测算视角表
发布方 / 视角年份地域 / 范围数值CAGR方法置信度局限
Grand View / Horizon 中国生成式 AI2030 预测;2025-2030 CAGR中国生成式 AI2030 收入 $17.609B39.1%国家数据手册预测;软件 / 服务细分页面披露 2030 和 CAGR,但没有完整公开逐年表
MarketsandMarkets 中国生成式 AI2025-2030中国生成式 AI2025 $7.036B;2030 $98.756B45.8%自上而下 / 自下而上的市场测算与交叉验证2030 年显著高于 Grand View;宽泛分层口径可能抬高边界
Tianxia Gongchang 中国大模型市场2025中国 LLM / 模型服务 + 部署 + 付费应用RMB49.5-51.0B (~$6.9-7.1B);广义口径 RMB100-130B (~$13.9-18.1B)公开摘录未说明行业研究报告;按市场层级拆解发布方方法论透明度较低;RMB 按约 7.2/USD 折算
Axis Intelligence 保守中国 AI 收入2025;2032 路径中国 AI 收入2025 年约 $28-31B;2032 年约 $200B32.5%汇总公开市场与政策数据的跨来源综合口径宽于生成式 AI,且部分来自二手来源综合
Axis / IDC 式广义中国 AI 活动代理指标2025包括基础设施 / 活动在内的中国 AI~$62B未说明Axis 引用的广义追踪代理指标可能包括算力 / 基础设施,并非直接软件收入
Gartner 全球 AI 支出2026全球 AI 支出总额 $2.596T;基础设施 $1.432T;软件 $453B;模型 $32.6B总支出同比 47%按细分领域预测的全球支出多为厂商 / 超大规模云厂商支出,远宽于 StepStar 收入池
Grand View 全球 LLM 市场2024-2030全球 LLM 收入2024 年 $5.617B;2030 年 $35.434B36.9%全球行业市场预测仅覆盖 LLM,但为全球口径;非中国专项
Precedence 全球 LLM 市场2025-2035全球 LLM 收入2025 年 $7.77B;2026 年 $10.57B;2035 年 $149.89B34.44%全球市场预测预测期较长,边界也不同于 Grand View
MarketsandMarkets 全球 LLM 市场2030 年预测全球 LLM 收入2030 年 $36.1B33.2%市场报告摘要搜索页摘录而非完整报告;发布时间较早

数值沿用发布方单位;RMB 数值按近似美元折算以便比较。本表刻意保留彼此不兼容的边界,而不是取平均值。

[CM006, CM007, CM008, CM009, CM010, CM011]
FM002: 市场估计区间

同一单位的区间视图显示,市场边界从 GenAI 扩到广义 AI 活动后,中国 2025 AI / GenAI 收入估计如何拉宽。

所有点位均为 USD billions;RMB 换算使用 ~7.2 RMB/USD,后续刷新应替换为来源原生财务数据。

[CM006, CM007, CM008, CM009, CM011]

2.3 买方分散在企业平台、开发者、消费者、政府和设备伙伴之间

StepStar 面向的不是单一买方画像。企业需要私有数据锚定、工作流智能体、合规和集成时,预算可能来自 CIO、CTO、数字化转型、业务部门或 AI 平台团队。开发者和创业公司通过 API 平台购买或试用,每 token 低价、长上下文、SDK 兼容性和模型路由器位置决定份额。消费者使用通常通过应用、手机 OEM、内容产品或订阅间接付费,而不是直接收到基础模型账单。政府和国企买方在中国是独立付款群体,因为国家 AI+ 政策明确鼓励公共服务和产业落地。设备和汽车伙伴也是一条路径,因为 StepStar 宣传消费电子和智能汽车方案;在这条路径上,用户可能从不知道底层模型来自 StepStar。因此,采用路径从能力证明和免费 token,走向开发者使用、企业试点、集成、治理审查和规模化生产。关键门槛是预算所有者是否清晰、模型切换成本多高、数据控制要求多强,以及 StepStar 是否嵌入了能把使用转成付费经常性需求的渠道。[CM017, CM018, CM019, CM020, CM021, CM022]

细分市场 / 买方图谱
细分市场买方用户付款方 / 预算所有者工作流采用触发因素
企业 AI 平台CIO、CTO、AI 平台负责人开发者、分析师、运营团队技术、数字化转型或业务单元预算私有数据锚定、智能体、文档工作流、自动化ROI 清晰、安全控制到位,并能接入现有数据
开发者 / API 用户开发者、创始人、应用 AI 工程师构建者或应用用户信用卡、创业公司工程预算、平台额度模型调用、长上下文工作流、编码、工具使用低价格、长上下文、SDK 兼容性、模型路由器可见度
消费者应用产品经理或应用运营方终端消费者订阅、广告、应用 P&L、OEM 收入分成聊天、内容生成、搜索、娱乐、教育黏性用例与分发,而不只是基准测试领先
政府与国企买方政府局办、国企数字化办公室公务员、市民、工业运营人员财政、国企或政府引导基金支持的项目预算政务服务助手、城市运营、法律问答、工业 AI政策要求叠加安全、本地化、可控部署
设备与汽车合作伙伴OEM AI 产品或车舱团队手机、PC、汽车或智能设备用户OEM 产品预算或战略合作嵌入式助手、多模态搜索、座舱智能能提升硬件差异化的端侧或云边体验
超大规模云厂商 / 生态平台云平台总经理或市场负责人企业客户和开发者云 / 市场平台商业预算模型市场、路由、托管智能体、评测供应商广度以及对中文 / 本地化模型的需求

各细分市场的买方、用户和付款方并不相同;只要合同保住模型收入分成,StepStar 即便不拥有终端客户,也能拿到使用量。

[CM017, CM018, CM019, CM020, CM021, CM022]
FM003: 买方 / 细分矩阵

StepStar 买方在预算归属、采用触发点和约束上各不相同;用户常常不是付款方。

准备度为定性判断,因为公开来源未披露 StepStar 管线或细分收入拆分。

[CM017, CM018, CM020, CM021, CM022, CM035]
FM004: 采用漏斗与价值链图

采用从模型验证和免费 / 开发者使用开始;收入放大前,还要过集成、治理和单位经济性关卡。

该流程是商业化逻辑图,不是实测转化漏斗;要量化流失,需要 StepStar 私有队列数据。

[CM013, CM017, CM018, CM021, CM022, CM023]

2.4 Token 增长和政策支持真实存在,但定价、算力和信任限制价值捕获

最强的市场驱动,是中国模型使用似乎已经从演示跨入高容量部署:Digital in Asia 报道,到 2026 年 3 月,日 AI token 达到 140 trillion;DigitalApplied 报道,2026 年 Q2 中国供应商拿下 OpenRouter 流量的 45% 以上。政府政策也是顺风:国务院 AI+ 意见推动跨行业采用,Axis 则引用了私人投资总额低估的国家和引导基金资本。全球企业需求也在从试点走向生产,Deloitte 报道员工 AI 访问增长 50%,McKinsey 则警告 AI 可能吃掉多达三分之一的变革预算。同一组证据也界定了约束。DeepSeek 官方定价显示 token 价格很低;AWS 和 Azure 让模型路由和多供应商选择变成常态;Qwen 和 DeepSeek 的开放权重让能力更可替代;中国特有监管还增加安全评估、实名义务、标识和内容控制风险。因此,市场判断是:使用量和政策利好,收入质量喜忧参半,对无差异模型 API 不利。StepStar 需要分发、垂直集成或智能体工作流控制,才能把中国 AI 增长转成可防守收入。[CM026, CM027, CM028, CM029, CM030, CM031]

增长驱动因素与约束表
驱动因素 / 约束方向时点含义尽调问题
据报道,到 2026 年 3 月中国 AI 日 token 使用量达 140T正向当前显示使用量已越过试点阶段并具备规模核验国家数据局原始指标以及 StepStar 的 token 份额
2026 年 Q2 中国供应商占 OpenRouter 流量超过 45%正向当前开发者需求可能迅速转向中国模型要求 StepStar 提供 API 流量、留存和路由器排名数据
国务院 AI+ 政策与公共部门采用议程正向2025-2026为公共部门和国企的垂直应用创造需求梳理 StepStar 真正能进入的政策驱动采购项目
企业 AI 访问与生产环境使用正在上升正向近期扩大智能体和模型支撑工作流的买方基础区分实验性使用与付费生产合同
政府引导基金与国资混合当前至长期资本和算力支持能加速供给,但也会扭曲市场信号识别哪些补贴或基金直接触达 StepStar
芯片 / 出口管制与国产加速器约束负向当前算力稀缺会放慢前沿训练,或迫使架构取舍审计算力供给、单位经济性和国产芯片兼容性
Qwen 与 DeepSeek 带来的开放权重浪潮负向当前抬高基准测试底线,削弱通用 API 能力的护城河测试 StepStar 相比开放权重替代品的差异化
API 价格战与超大规模云厂商路由负向当前使用量增长可能只能以被压缩的毛利率变现对照 DeepSeek / Qwen / AWS / Google 价格带测算模型毛利率
监管、内容控制与信任要求负向当前公共部署和企业采购需要承担合规开销审查备案、审批、标识、数据安全和审计控制

这些驱动因素和约束都有证据支撑,但可量化程度不同;其中几项取决于 StepStar 非公开流量、定价和算力数据。

[CM026, CM027, CM028, CM029, CM030, CM031]

2.5 图表

Chapter 03

03竞争格局

3.1 中国前沿模型格局:StepStar 可信,但不是量级领先者

StepStar 所在的是全球最拥挤的基础模型市场之一。中国竞品集合不只是拿了风投的钱的「六小虎」;还包括 DeepSeek 的低成本开放推理栈、Alibaba/Qwen 的云与开放权重机器、ByteDance/Doubao 的消费者分发、Baidu/ERNIE 和 Tencent/Hunyuan 的存量企业渠道,以及 Baichuan 等垂直化同行。证据支持 StepStar 是严肃玩家:Step3 是大型多模态 MoE 模型,STEPX Neo 则给公司一条终端级产品逻辑。反向解读是,独立 2026 年 Q2 供应商地图把 StepFun 排在最大 API 量级领导者之后,因此公司当前公开优势不是原始 API 份额。差异化必须来自多模态效率和设备 / 智能体集成,而不是默认开发者后端。尽调需要这一点,因为买方比较按工作负载展开,而不是按公司名展开:文档智能体、编码智能体、设备助手、云托管企业模型和自托管开放权重,都能满足重叠任务。因此,StepStar 需要证据证明,Step3 加终端这条路能改变使用频率或服务成本,而且便宜 API 难以复制。[CP001, CP003, CP007, CP008, CP009, CP036]

中国 LLM 供应商对比表
供应商所有者 / 实体类别规模 / 融资信号产品范围与切入点定价 / 开放态度与 StepStar 对比的局限
StepStar / StepFunShanghai Jieyue Xingchen / StepFun多模态基础模型创业公司独角兽;据报道 2026 年 B+ 轮融资超过 5B RMB,并有 IPO 估值目标报道Step3 多模态 MoE,加上 STEPX Neo 智能体手机 / 终端策略公开模型检查点和兼容 API;公开企业包装细节较薄终端差异化逻辑可信,但据报道 API 份额小于领先者
DeepSeekDeepSeek低成本开放推理实验室私营公司;全球冲击力比披露融资更显眼R1 推理模型,以及带 1M 上下文定价页的低成本 API开源模型并公开 API 价格压低 StepStar 定价,并让内部自建更容易
Moonshot / KimiMoonshot AI长上下文与智能体模型创业公司公开公司资料显示其获得大量私募融资Kimi K 系列、百万 token 上下文,聚焦编码和知识工作基于 token 的 Kimi API 定价,以及 Kimi K2 模型发布在智能体知识工作上与 StepStar 高度重合
Zhipu / GLM / Z.ai 体系Zhipu AI企业 / 开放模型实验室公开资料称其完成大额私募轮,并有 IPO 讨论GLM 智能体推理 / 编码模型,以及 BigModel / Z.ai 平台MIT 许可 GLM 模型和 API 平台对自托管买方更具宽松许可优势
MiniMaxMiniMax Group多模态与智能体模型创业公司公开资料和市场报道将其列为中国主要 AI 创业公司之一M3、M1、视频、语音、音乐、编码智能体,以及 1M 上下文主张开放权重 M1 模型卡,加上 API / 聊天机器人入口在长上下文和多模态智能体上直接竞争
BaichuanBaichuan Intelligence垂直化模型 / 应用公司公开资料称其为 Wang Xiaochuan 创办的中国 AI 创业公司Baixiaoyi 医疗 / 家庭健康助手和模型 APIAPI 文档可用;可见切入点是医疗健康作为广义前沿威胁较弱,但在健康垂直领域更清晰
ByteDance / DoubaoByteDance消费级与多模态既有巨头背靠 ByteDance 分发,而非风险融资信号Doubao 聊天机器人 / 产品族,以及创意 / 多模态用例在已审阅来源中,定价不是主要公开切入点消费端分发规模可能碾压创业公司声量
Alibaba / QwenAlibaba Cloud开放权重与云既有巨头上市公司云生态,而非创业融资Qwen3 开放模型、Model Studio、多模态模型、企业部署大量开放权重,加上云端定价 / 部署设定开放性和企业分发基准
Baidu / ERNIEBaidu AI Cloud云既有巨头模型平台上市公司既有巨头Qianfan / Wenxin 一站式企业大模型平台云平台包装;此处未完整提取定价借既有企业采购通道竞争
Tencent / HunyuanTencent Cloud云 / 内容既有巨头模型平台上市公司既有巨头Hunyuan MoE、256K 上下文、搜索 / 内容生态集成Tencent Cloud 平台包装借生态捆绑和内容访问竞争

列举范围覆盖章节简报明确要求的中国供应商以及 StepStar 本身;规模和融资单元格使用公开信号,而非经审计的私营资本数据。

[CP001, CP003, CP004, CP005, CP007, CP008]
FP001: 竞争定位图

StepStar 靠多模态终端策略做出差异化,但在分发和 API 用量证据上落后于头部厂商。

轴向分数是基于公开证据的序位估计,不是经审计的市场份额或基准值。

[CP003, CP008, CP012, CP024, CP026, CP027]

3.2 能力与开放性:Step3 有效率优势,但开放生态已很拥挤

Step3 架构是竞争论点的正面核心:总参数 321B、每 token 激活 38B、多模态推理、公开模型检查点,以及兼容 OpenAI/Anthropic 的访问方式。对需要在受限加速器上跑视觉—语言推理的买方来说,这是可信的技术切入点。问题在于,开放已经不稀缺。DeepSeek 开源 R1 和蒸馏模型,Qwen3 在广泛模型家族中开放稠密与 MoE 权重,Zhipu 以宽松许可发布 GLM 模型,Moonshot 和 MiniMax 发布大型 MoE 模型卡,Meta 则通过 Llama 在国际市场维持开放权重压力。因此,StepStar 的策略更像选择性开放叠加终端分发,而不是独一无二的开源护城河。买方可以跨这些 API 和模型检查点多宿主,降低切换成本,也削弱单独模型访问的溢价。比较方式也会改变基准测试解读。强模型卡是进入候选名单的必要条件;但如果买方能用开放模型检查点、云托管 Qwen 或 GLM 部署,或 DeepSeek 式低成本路由复现可接受表现,它还不足以赢单。StepStar 必须证明,多模态效率在生产中到底在哪里可量化。[CP001, CP002, CP010, CP011, CP015, CP017]

功能 / 能力矩阵
公司多模态 / 智能体实力开放权重或自托管信号API / 定价可见度分发 / 市场进入能力信任 / 合规姿态竞争含义
StepStar高:Step3 多模态 MoE 加 STEPX 终端逻辑中:模型检查点和模型卡可用中低:提到兼容 API,但审阅过的公开页面未完全显示定价中:有潜在设备 / 合作伙伴路径,但 API 份额较小未知-中:需要披露企业控制能力如果终端工作流转化为使用量,就能形成差异化
DeepSeek中高:聚焦推理和编码高:R1 和蒸馏模型已开源高:公开按 token 定价和兼容端点引发冲击后,在开发者中热度高受监管企业控制能力未知-中主要价格威胁和内部自建威胁
Qwen / Alibaba高:覆盖语言、图像、翻译、安全和智能体的广泛模型族高:稠密模型和 MoE 权重公开高:云端 Model Studio 和定价路径借 Alibaba Cloud 达到很高水平高:云页面上有合规和企业主张设定开放性与企业基准
Kimi / Moonshot高:长上下文、多模态、编码、知识工作中高:Kimi K2 模型检查点 / 模型卡高:Kimi API 定价页中高:Kimi 消费品牌强未知-中直接的长上下文智能体竞争者
GLM / Zhipu高:智能体推理 / 编码,加上视觉姊妹模型引用高:MIT 开源许可中高:Z.ai 和 Zhipu API 平台引用在中国企业市场为中高对自托管买方为中高强宽松许可替代方案
MiniMax高:编码 / 智能体、视频、语音、音乐、长上下文中高:M1 开放权重模型卡中:API / 聊天机器人入口;此处定价较不清晰消费者 / 媒体声量中高未知-中直接的多模态智能体竞争者
Baidu / Tencent中高:ERNIE 与 Hunyuan 企业模型低中:既有巨头平台更封闭中:云平台包装借既有巨头生态达到很高水平国内企业采购能力高难撼动的分发替代品
OpenAI / Anthropic / Gemini / Llama前沿基准覆盖面很高混合:Llama 开放;其他以 API 优先高:公开商业 / API 定价和模型页面全球生态触达很高企业控制能力基准高抬高全球买方预期

序数单元格是基于公开模型卡、官方页面和独立对比来源做出的分析师判断;没有支撑的企业控制能力单元格标为未知,而非推断。

[CP001, CP002, CP010, CP011, CP013, CP015]
FP002: 功能广度 / 能力图

StepStar 的功能广度强在多模态效率和终端策略;同业则在开放度、价格透明度或分发上领先。

单元格是基于公开证据的序位评估,并保留企业控制能力的未知缺口。

[CP001, CP010, CP013, CP017, CP020, CP025]

3.3 定价、融资和全球标杆抬高门槛

融资故事很强,但不够。公开来源显示,StepStar 到 2024 年底已成为独角兽,2026 年据报完成超过 RMB 5 billion 的 B+ 轮,并与估值目标更高的潜在香港 IPO 绑定。但 Moonshot、Zhipu、MiniMax、Baichuan 等私营公司的融资数据参差不齐,单靠资本化无法干净排序。采购可比性在定价和包装上更清楚。DeepSeek、Kimi、Gemini、OpenAI 和 Baichuan 发布 API 或商业定价入口;Alibaba 把 Qwen 包进云部署和合规工具;OpenAI、Anthropic 和 Google 则提供国际企业级控制和模型广度标杆。StepStar 有公开技术模型故事,但已审阅证据在标价、企业管理、支持承诺和使用控制上更薄。落到实际采购,StepStar 是高潜力但不够清晰的选项。技术团队可能喜欢 Step3;CIO 或平台负责人仍会追问价格可预测性、安全姿态、支持、使用分析和备用模型。全球领导者和若干中国存量厂商已经把这些买方控制项摆了出来。[CP004, CP005, CP006, CP014, CP016, CP018]

定价 / 包装对比
供应商已审阅公开包装价格可见度包含能力对 StepStar 的含义
StepStarStep3 GitHub / Hugging Face 与 STEPX 发布报道API 兼容性可见;已留存来源中未找到公开定价多模态 MoE 模型、模型检查点、终端设备验证要在采购中竞争,需要更清晰的企业 / 定价包装
DeepSeekAPI 文档和开放模型仓库高:已发布按百万 token 计价1M 上下文、OpenAI / Anthropic 兼容端点、开放推理模型直接锚定低成本基准
Kimi / MoonshotKimi 首页、API 定价、Kimi K2 仓库高:token 计费页面和模型列表可见K3 / K2 系列、长上下文、代码和多模态模型争夺长文档和智能体工作负载
Qwen / AlibabaQwen 网站、GitHub、Hugging Face、Alibaba Cloud Model Studio中高:云定价路径可见开放权重、多模态模型、企业部署、合规声明生态广度很难被击败
Baichuan官方产品页和 API 文档中:API 文档可见,未提取当前精确价格医疗产品加聊天补全端点垂直方向清晰,但平台压力不那么广
OpenAIGPT-5 发布和 Business 定价高:Business 定价和控制项可见前沿模型、工作空间、分析、连接器、SSO、预算全球企业级打包标杆
Anthropic模型概览和 Claude Opus 页面中高:模型文档可见;保留的来源集中未把定价作为核心Claude 模型家族,高级推理 / 编码定位技术买家的全球能力标杆
Google GeminiGemini 模型和定价文档高:按百万 token 定价可见Gemini 3 预览版、Gemini 2.5 Pro / Flash、事实锚定选项公开 API 经济性和广度的标杆

该对比强调公开采购时能否看清;实际谈判的企业价格可能不同,仍是尽调缺口。

[CP001, CP002, CP010, CP014, CP026, CP030]
融资 / 估值对比
公司公开融资 / 估值信号来源置信度战略解读尽调保留项
StepStar据报道 B+ 轮融资超过 5B RMB;IPO 估值目标报道约 $12B;截至 2024 年底已是独角兽资本充足,足以支撑训练和设备战略继续推进私募融资条款和 IPO 状态需要一手确认
Moonshot AI公开资料概述了规模可观的私募融资和 Kimi 产品动能中低资金充足的长上下文同业各轮估值需用一手股权结构表或监管文件刷新
Zhipu / Z.ai 体系公开资料概述了大额私募轮和 IPO 讨论中低有融资厚度的企业级 / 开放模型同业IPO 时点和估值仍是波动的私募市场事实
MiniMax公开资料概述了 MiniMax 融资和产品扩张中低资本充足的多模态智能体同业本报告未审计私募轮条款和现金跑道
Baichuan公开资料和官方产品入口显示其垂直 AI 战略仍在推进中低覆盖面可能更窄,但更垂直化需要更新融资和客户牵引证据
DeepSeek相比颠覆性低成本模型冲击,融资不是核心竞争威胁来自经济性和开放性,不是报道估值需要核查当前所有权和资本能力
Alibaba / ByteDance / Baidu / Tencent背后是上市公司或大平台,不适合按创业公司估值口径横比身份可信度高,独立模型 P&L 可见度低在位者资产负债表和分发能力压过创业公司渠道独立模型单元经济未披露
OpenAI / Anthropic / Google / Meta全球前沿标杆,企业与生态可见度更强产品证据高,私募估值参差定义全球能力、定价、控制和开放权重预期并非都能直接替代中国市场采购标的

该表刻意把报道的私募融资信号与产品证据分开;多数中国创业公司估值需要一手融资文件才能形成投资判断。

[CP004, CP005, CP006, CP016, CP018, CP021]
FP003: 护城河 / 就绪 KPI

竞争逻辑可信,但仍需证明终端分发和模型效率能转化为持久使用。

KPI 数值把来源披露值与分析师标签合在一起;API 份额标签只是方向性判断,不是经审计的精确份额。

[CP001, CP004, CP008, CP009, CP036, CP046]

3.4 护城河耐久性:DeepSeek 冲击和较小 API 份额构成反向情景

反向竞争情景很直接。DeepSeek 重置了买方对价格和开放性的预期;Qwen 和 GLM 让自托管和云部署更可信;MiniMax 和 Kimi 争夺长上下文与智能体工作负载;Doubao、Baidu、Tencent、OpenAI、Anthropic 和 Google 带来的分发能力,StepStar 很难快速复制。StepStar 的答案是一条整合「AI + 终端」路径;如果 STEPX 设备和伙伴成为经常性用户界面,它可能形成工作流级锁定。这个结果尚未得到公开证明。因此,尽调负担不是 StepStar 是否有技术可信的模型——它有——而是 StepStar 能否赢下任务级部署;在这些场景中,买方也可以因成本把工作路由给 DeepSeek,因开放性选 Qwen 或 GLM,因长上下文智能体选 Kimi 或 MiniMax,或因企业控制选国际前沿供应商。由此得出的尽调姿态是:把 StepStar 视为差异化挑战者,而不是已经证明的品类赢家。下一步最重要的证据,是重复部署案例能证明终端访问、模型效率和多模态推理共同产生持久优势,胜过更便宜或分发更强的替代品。[CP012, CP024, CP027, CP028, CP035, CP037]

护城河耐久性 / 竞争风险台账
护城河主张威胁路径严重程度证据缓释措施 / 尽调要求
成本高效的多模态 MoEDeepSeek、Qwen、GLM、MiniMax 和 Kimi 也发布高效 MoE 或开放模型Step3、DeepSeek R1、Qwen3、GLM-4.5、MiniMax-M1 和 Kimi K2 均提到 MoE 或开放模型路径跑任务级基准,验证 Step3 在质量调整后成本上胜出
AI + 终端分发设备发布未必能转化为经常性工作流或开发者采用中高STEPX Neo 发布是正面信号,但 API 份额证据仍弱于头部核验激活设备、留存和合作伙伴渠道经济性
开放模型检查点和兼容 API多模型并用和自托管降低切换成本竞争对手发布兼容 API、价格或开放权重在真实客户部署中测量切换成本
中国市场准入Alibaba、ByteDance、Baidu 和 Tencent 具备更强在位渠道Qwen、Doubao、ERNIE / Qianfan 和 Hunyuan 都借助大型生态找出 StepStar 拥有排他合作伙伴或终端铺位的场景
资金跑道竞争对手和在位者也资金充足或有资产负债表支撑公开来源显示 StepStar 完成融资,但同业资本化信息不完整获取股权结构表、现金、烧钱速度和算力承诺
基准可信度外部排行榜尚未证明 StepStar 是广义前沿领导者中高Artificial Analysis 和 LMArena 给买方提供外部对照针对具名中国和全球同业跑独立评测

严重程度是基于公开证据的分析师判断;缓释项是尽调问题,不是已确认的管理层计划。

[CP012, CP034, CP035, CP036, CP037, CP038]

3.5 图表

Chapter 04

04财务情况

4.1 融资与估值轨迹

StepStar 的融资记录,是其财务画像中最可见的一块;但即便在这里,公开证据也更偏标题级,而非尽调级。多个来源称,2024 年 12 月 Series B 有上海国资、Tencent、Qiming Venture Partners 和 5Y Capital 参与;后续来源称,2026 年 1 月 B+ 轮超过 RMB 5 billion,约 US$700 million 以上,并新增国资、产业资本和存量财务投资人。估值路径没那么干净:2024 年报道把公司放在约 US$1 billion 档位;2026 年 1 月中文报道估计投后区间约 RMB 20-30 billion;后续面向 IPO 的英文报道则抛出更高美元目标。后面的 IPO 估值应视为市场报道,不是已完成融资条款。最强结论不是精确估值曲线,而是 StepStar 已成为少数能在同业面对更挑剔融资市场时,仍吸引超大规模战略资本的中国基础模型创业公司。因此,下表记录融资事实,并把已确认轮次金额与报道或推断的估值锚分开。[CI001, CI002, CI003, CI004, CI005, CI006]

资本充足性表
融资 / 指标公开证据支持的数值或状态它说明什么它不能说明什么
2024 年 12 月 B 轮据报道数亿美元;来源称估值约 US$1B政府与战略投资者支持的早期证明新股 / 老股精确拆分、优先权和现金余额
2026 年 1 月 B+ 轮据报道超过 RMB 5B / 大约 US$700M+大额注资,也是同期中国模型赛道最大融资轮之一现金跑道、月度烧钱,以及募资是否足以覆盖路线图
累计融资公开来源称计入 B+ 后累计融资超过 RMB 5B相比典型创业公司,确认资本化程度可观精确累计新股融资额和剩余现金
B+ 后估值中文报道估计约 RMB 20-30B;数据网站和 IPO 文章口径不一估值较 2024 年显著上台阶没有已提交股权结构表或条款清单来核对报道估值
投资者组合报道出现国资基金、China Life PE、PDVC、Xuhui、Wuxi Liangxi、Xiamen ITG、Huaqin、Tencent、Qiming 和 5Y投资团带来资本、政府、硬件和风险投资背书治理权、跟投承诺和战略限制
募资用途据报道用于基础模型研发、算力 / 基础设施、人才和 AI 终端铺开资本绑定高成本技术和商业化目标预算分配、里程碑覆盖和成本超支风险

轮次金额只是公开报道锚点,不是经审计现金余额。后续 IPO 估值目标不纳入本表,因为它们不是已完成融资条款。

[CI001, CI002, CI003, CI004, CI005, CI006]
FI004: 资本密集度 / 现金流图

公开资本叙事从大额私募轮次,走向重算力执行和可能的公开募资;隐藏烧钱速度是核心变量。

[CI005, CI006, CI021, CI022, CI032, CI033]

4.2 收入模式与变现披露

公开记录不足以支撑常规收入模型拆解。StepStar 官方入口展示消费者聊天、开放平台、模型卡和 Step Plan 促销页面,包括限时免费 token 优惠;但没有提供完整收入表、实际成交价格、客户集中度、确认收入、年经常性收入(ARR)、留存或毛利率。独立中文报道主张公司有终端优先的商业化路径,并引用手机安装、日服务量、汽车部署和 API 增长;这些证据能说明商业化方向,但仍不等于经审计收入。最可防守的收入模型,是终端与平台混合:平台上的 API 使用或套餐订阅,与手机和汽车伙伴的授权或集成经济性,以及未来企业或智能体工作流套餐。每条收入流的承销信息仍不足,因为计量单位、价格兑现、收入分成、支持义务和收入确认政策都未公开。因此,本章把公开变现视为方向上得到证明,但财务上无法量化。[CI011, CI012, CI013, CI014, CI015, CI016]

收入流表
收入流机制计量单位当前公开价值 / 状态质量尽调要求
开放平台 API 使用开发者基于 StepFun 模型和智能体构建应用token / API 调用 / 套餐平台和 Step Plan 入口存在;完整实际价格未披露可能形成经常性使用收入,但财务可见度有限索取按模型拆分的 API 收入、活跃付费开发者、token 用量、折扣和毛利率
Step Plan / 订阅入口官网把用户导向 Step Plan 和购买流程套餐订阅或 token 包官方材料宣传限时免费 token 体验;付费转化未公开显示打包意图,不代表已确认收入索取套餐档位、付费席位、流失率和促销转付费转化
手机 OEM 模型部署模型嵌入 OPPO、Honor、ZTE 或其他设备工作流授权、按设备收费、收入分成或服务费报道称装机设备 42M+ 且有日常服务使用,但未披露合同经济性分发信号强,收入证据弱索取 OEM 合同、收入分成公式、最低保底和支持义务
汽车座舱 / Agent OS语音和智能体模型接入 Geely 车载系统按车授权、项目费或服务费报道提到 Galaxy M9 / Agent OS 部署和百万辆车目标;定价未知可能是高价值垂直场景,但大概率为定制项目索取汽车合同金额、收入确认政策、质保 / 支持成本和续约条款
企业智能体工作流套件面向企业或合作伙伴销售生产级智能体模型能力席位、用量或定制合同未以收入项目公开拆分平台成熟后可能带来上行空间索取企业销售管线、已签约 ARR、试点转化、CAC 和实施成本

公开来源显示产品和分发入口,但不显示实际收入、收入分成或已确认收入;各行只是受公开证据约束的收入流假设。

[CI011, CI012, CI013, CI014, CI015, CI016]
定价 / 变现表
入口公开价格或合同口径标价与实际状态折扣 / 未知项来源含义
StepFun 首页链接到聊天、开放平台、Studio 和 Step Plan 购买入口标价式打包存在,但未公开看到详细价格表促销、免费 token 赠送、付费转化、企业折扣未知官方入口支持变现意图判断,但不能证明收入
StepFun 开放平台推广模型 / API 访问和应用开发API 变现具备可能性;实际使用收入未披露token 定价、量级返利、免费额度和合作伙伴条款不清楚建模收入前需要平台流水账
Step Plan官方 Step Plan 页面称提供限时免费 token 体验 / token 赠送能看到促销;付费采纳和实际 ARPU 未公开免费 token 补贴可能压低近期收入质量付费转化和补贴成本是重要尽调问题
OEM / 设备合作报道的装机量和终端集成无标价;大概率是谈判合同授权单位、收入分成和最低承诺未披露分发可以很大,但单设备收入仍未知
IPO 市场叙事英文报道讨论可能的香港 IPO 和估值目标融资叙事,不是定价证据没有招股书级收入或利润率披露不应作为变现证明

官方定价证据不完整。该表区分公开打包和实际经济性,后者仍属私人信息。

[CI011, CI012, CI013, CI014, CI015, CI032]
FI001: 收入模型桥

公开证据支撑从模型和终端分发走向潜在收入的路径,但还撑不起每个转化环节的经济性。

[CI011, CI012, CI013, CI014, CI015, CI016]

4.3 成本结构、单位经济性与烧钱速度

StepStar 更不像轻 SaaS 公司,而像资本密集型基础模型和 AI 设备基础设施公司。公司自有和第三方材料强调大型多模态模型、Step 3.7 Flash、Step-3 效率、AI 智能体、端侧部署、车载语音模型和模型迭代。中文报道称,B+ 轮资金将用于基础模型研发和 AI 终端推出;长篇文章也明确把算力基础设施和顶尖 AI 人才列为投资重点。这些事实使 GPU / 算力、高级研究人才、模型服务基础设施、数据运营和集成工程,成为最可能的烧钱类别。但没有来源披露月度烧钱、云承诺、GPU 资本开支、单用户推理成本、毛利率或伙伴收入分成。反向市场背景很重要,因为行业来源描述了更广泛的转向:从不受约束地烧钱,转向变现、效率和公开市场审视。因此,单位经济性只能做方向性估计:如果效率主张成立,StepStar 或许拥有更低边际推理成本,但公开记录不足以把它转译为毛利率或现金跑道。[CI020, CI021, CI022, CI023, CI024, CI025]

单元经济表
指标 / 驱动因素公开数值或状态置信度重要性尽调要求
已确认收入 / ARR所审阅公开来源未披露估值和收入质量的核心分子索取按产品、客户分层和季度拆分的收入桥
毛利率未披露决定 AI 终端模式扣除推理和支持成本后,能否呈现软件式利润率索取按算力、数据、支持、合作伙伴分成和折旧拆分的收入成本
每次 API 调用 / 每设备用户推理成本未披露;效率主张意味着可能有优势关键在于,若单次使用成本高,庞大装机基础仍可能亏损索取 token 组合、模型路由策略、GPU 利用率和每 1K token 成本
训练 / 研发算力无美元金额;B+ 募资用途包括基础模型研发和基础设施优先事项可能是最大可支配烧钱项索取 GPU / 云合同、预留容量、资本开支和训练路线图
人才烧钱无薪酬数字;战略需要顶尖模型、系统和商业化人才高级 AI 团队在收入规模化前就会形成高固定烧钱索取人员计划、薪酬、承包商支出和招聘承诺
销售效率 / CAC 回本未披露需要判断终端合作是否降低 GTM 成本索取合作伙伴获取成本、实施成本、回本周期和续约数据
营运资本 / 债务义务无公开债务或营运资本披露若硬件或车辆集成产生预付款或支持义务,该项重要索取债务、应付款、最低承诺和表外义务

公开证据缺失时,数值刻意留空;置信度指该指标的公开可得性,不评价公司质量。

[CI020, CI021, CI022, CI023, CI024, CI025]
FI002: 单位经济性桥

在公开毛利证据出现前,成本基础很可能主要由模型 R&D、算力、推理和伙伴集成主导。

[CI021, CI022, CI023, CI024, CI025, CI026]
FI003: 财务估计区间

只有融资和估值区间有公开锚点;没有私有数据,经营指标仍基本无法框定。

经营区间使用的是 0 个已披露公开数字,并不代表公司收入或烧钱为 0。融资区间反映四舍五入后的公开报道。

[CI001, CI002, CI007, CI008, CI020, CI032]

4.4 资本充足性与 IPO 融资背景

B+ 轮显著改善 StepStar 的资本充足性,但仍不足以承销现金跑道。按中国 AI 标准,超过 RMB 5 billion 是一笔大型私募融资,可能给公司继续训练模型、补贴推理、招聘和深化终端伙伴关系留下空间。但公开来源没有显示轮前或轮后现金、月度净烧钱、计划算力承诺、债务义务、清算优先权,或融资是否包含老股流动性。2026 年后续报道称 StepStar 正在探索香港 IPO,并可能募资约 US$500 million,这又增加了一个融资依赖信号:公司可能希望趁私募市场仍接受 AI 终端故事时引入公共资本。这些 IPO 报道可作为市场背景,但不等同于已提交的招股书。因此,财务判断是平衡的:StepStar 拥有异常强的资本获取能力和一个可能差异化的商业化渠道,但详细财务仍未公开,决定性尽调阻塞点仍是收入、利润率、烧钱速度、现金跑道、股权结构条款和伙伴经济性。[CI032, CI033, CI034, CI035, CI036, CI037]

公开财务缺口表
缺失的私有指标影响严重程度具体尽调路径
收入 / ARR / 已确认收入无法区分真实变现与分发或使用牵引关键索取按产品、地域、客户和收入确认政策拆分的经审计月度收入
毛利率和收入成本无法判断模型服务和伙伴支持能否随规模盈利关键索取 COGS 桥:GPU 推理、训练摊销、数据、带宽、支持、合作伙伴分成
月度烧钱、现金余额和现金跑道无法评估 B+ 轮后或可能 IPO 前的融资依赖关键索取现金余额、净烧钱、已承诺支出,以及模型训练情景下的现金跑道敏感性
算力合同和 GPU 获取隐性最低承诺可能主导资本需求关键索取云 / GPU 合同、预留容量、利用率和芯片供应约束
股权结构表和优先股堆叠表面估值可能不反映投资者回报经济性重大索取股份类别、清算优先权、老股交易、期权池和战略权利
合作伙伴经济性和客户集中度如果经济性靠补贴或客户集中,设备装机可能只能转化为低收入重大索取 OEM / 汽车合同、收入分成、最低保底、流失率和集中度上限

这些是把公开融资证据转成可承做财务模型所需的最低私有输入。

[CI016, CI017, CI018, CI019, CI020, CI026]

4.5 图表

Chapter 05

05产品与技术

5.1 产品定义与组合宽度

StepFun 不是单产品 AI 创业公司;公开产品定义是一套分层平台。可见组合从消费者 StepFun AI 应用和桌面智能体开始,加上用于实验和创作的 AI Studio,开放带 API 文档和定价的 Open Platform,把面向编码与智能体工具的 Step Plan 订阅打包进去,现在又通过 STEPX Neo 智能体手机概念延伸到终端硬件。按工作流看,公司想掌握从基础模型到智能体执行界面的路径:开发者通过 OpenAI 风格 API 调用模型,创作者在 Studio 里实验,消费者使用 StepFun AI,设备伙伴则可把多模态助手能力嵌入手机、车辆或其他终端。组合宽度是正面承销信号,因为它展示了多条分发路径,而不是单纯研究演示。需要警惕的是,宽度也会分散执行:模型质量、API 可靠性、应用 UX、开放权重支持、订阅经济性和硬件—智能体集成,需要不同运营肌肉。[CE001, CE018, CE019, CE028, CE029, CE030]

产品组合 / 资产矩阵
资产或模块主要用户状态 / 成熟度差异化尽调缺口
StepFun AI 应用和桌面智能体消费者和准专业知识工作者公开 Web、移动端和桌面端入口可配置行为、以 OS 级任务为框架的智能体式工作伙伴留存、活跃用户和任务成功率未公开
AI Studio创作者、开发者和模型评测者公开实验与创作入口在同一环境里整合聊天、搜索、展示、资产库和 PlaygroundStudio 使用量、协作控制和企业治理未公开
Open Platform / Step API 产品开发者和企业集成方有文档的 API,提供定价和 OpenAI 兼容迁移主打稳定、高性能、易集成,并提供模型目录实际正常运行时间、延迟分布和支持 SLA 未公开
Step Plan智能体和编码工具用户订阅产品,含每月 Credit 额度提供专用 API key,可从 Claude Code、Trae、Cursor 等智能体 / 编码工具访问实际转化率和每 Credit 毛利未披露
Step-3 开放模型开发者、研究人员和自托管用户开放权重 / 代码,通过 GitHub 和 Hugging Face 分发321B 多模态 MoE,38B 激活,MFA / AFD 成本效率逻辑独立基准复现和企业部署审计未公开
STEPX Neo / Step AOS / Amoo 终端线智能体手机用户和生态合作伙伴2026 年 7 月新闻报道中的终端概念 / 产品原生智能体手机,整合模型、OS、硬件和生态服务官方规格、出货规模、定价和用户采用未公开

枚举覆盖已审阅来源中出现的公开产品和资产入口,不包括任何私有 SKU 或未发布的企业合同。

[CE001, CE007, CE008, CE018, CE028, CE029]
模型能力对比表
模型 / 系列公开能力定位已披露核心指标交付入口主要限制
Step-2万亿参数级基础 LLM 和推理底座官方站点称其为万亿参数自研底座;二手来源描述为 1T+ 多模态范围官方站点 / 历史发布报道未找到包含完整架构的主要公开技术卡
Step-3旗舰开放多模态推理 MoE321B 总参数、38B 激活、65,536 上下文、48 个专家、MFA 注意力API、GitHub、Hugging Face、ModelScope 等渠道基准主要由发行方发布;自托管仍高度依赖硬件
Step-3.7 Flash面向智能体的当前旗舰多模态 Flash 推理模型198B 总参数、11B 激活、256K 上下文,原生支持图像和视频输入Open Platform、Step Plan、官方博客2026 新模型;独立生产案例未公开
Step-3.5 Flash面向智能体和编码的快速语言推理模型256K 上下文;针对工具调用、规划、数学、编码和研究优化Open Platform、Step Plan 等入口相较 Step-3.7 的多模态输入,定位为纯文本
Step-1V / Step-1.5V较早的视觉 / 多模态理解谱系第三方报告称 2024 年中文视觉模型排名强历史产品报道和官方站点引用基准语境较旧,不如 Step-3 / Flash 线当前
StepAudio / Step Image / Step Router 扩展语音、图像编辑和路由扩展文档在支持模型集中列出 StepAudio 2.5、Step Image Edit 2 和 Step Router V1Open Platform 和 Step Plan特定能力质量指标和客户结果未公开

该表枚举与产品尽调相关的主要公开模型系列;行内指标仅限抓取到的公开来源可见内容。

[CE002, CE003, CE016, CE017, CE019, CE022]
FE001: 产品架构图

StepFun 公开展示的产品架构把模型、API、智能体工具、应用和终端实验叠在一起。

[CE004, CE005, CE007, CE018, CE028, CE030]
FE002: 客户工作流 / 运营流

用户意图可从应用、Studio、API 或终端进入,再路由到模型 / 工具执行。

[CE018, CE019, CE020, CE028, CE031, CE032]

5.2 模型架构、成本效率与基准证据

最强的产品技术证据是 Step-3。公开技术材料描述了一个 321B 参数多模态 MoE,每 token 激活 38B 参数,支持 65,536 上下文,采用 MFA attention、48 个专家且每 token 选择 3 个专家。系统论文和模型博客对工程逻辑写得异常明确:不是单纯堆参数,而是降低解码成本。MFA 降低 KV-cache 和 attention 计算;AFD 把 attention 与 FFN 工作拆到专门推理子系统。StepFun 还披露了预训练规模——超过 20T 文本 tokens 加 4T 图文混合 tokens——并报告 MMMU、MATH-Vision、AIME25、GPQA-Diamond、LiveCodeBench 等任务的基准结果。成本故事可信,因为模型、attention 算术、稀疏激活和部署指引彼此一致。但这些基准仍主要由公司发布,比较器复现设置很关键;已审阅公开记录中没有找到独立复现。因此,Step-3 可以作为严肃技术资产来承销,但还不能被独立认证为持久性能护城河。[CE002, CE003, CE004, CE005, CE006, CE009]

技术 / 运营架构表
层或流程作用依赖风险或尽调含义
MFA 注意力为 Step-3 解码降低 KV-cache 和注意力 FLOPsStepFun 模型架构和实现选择若能在发行方测试外复现,就是强成本效率杠杆
MoE 稀疏激活VLM 总参数达到 321B 时,将激活参数压在 38B专家路由稳定性和训练质量独立评审提示专家失活问题,需要尽调
AFD 服务设计将注意力和 FFN 工作拆到专门推理子系统分布式推理软件和硬件拓扑开源指南称 AFD 支持仍在推进中
训练数据管线提供 20T+ 文本 tokens 和 4T 图文混合 tokensWeb、授权出版商数据和内部解析 / 过滤数据权利和安全过滤证据来自公司描述
开放部署引擎为 Step-3 提供 vLLM 和 SGLang 部署路径GPU 内存、张量并行、nightly 依赖最低部署占用限制了轻量自托管

架构行基于公开 Step-3 技术材料,不应解读为对私有训练或服务基础设施的审计。

[CE002, CE003, CE004, CE006, CE009, CE010]
FE005: 基准与效率信号条形图

公司披露的部分 Step-3 指标释放出较强的基准和服务效率信号,但还不是独立复现。

基准值来自公司发布或评测转述的公开数据;单位混合了分数和服务指标,仅用于可视化比较。

[CE005, CE012, CE013, CE014]

5.3 部署、开发者工作流与智能体界面

StepFun 的使用路径先面向开发者,也先面向智能体。Open Platform 和 API 参考文档记录了 chat-completion 调用、兼容 OpenAI 的迁移、按 token 计费、速率限制和分模型定价。Step Plan 为使用 OpenClaw、Claude Code、Trae、Cursor 等编码和智能体平台的开发者增加了一层订阅。开放模型侧,Step-3 出现在 GitHub 和 Hugging Face,在中国通过 ModelScope 分发;部署指南支持 vLLM 和 SGLang,但也显示大模型自托管仍然吃硬件。这个分叉对尽调很重要。托管 API 和订阅产品降低采用摩擦,开放权重提升开发者心智和自托管选择权。与此同时,开放权重也会商品化部分模型层,最低部署足迹仍大到不适合普通企业随手自托管。只有当 StepFun 能把模型层转成稳定工具、记忆、编排、设备集成和可支持的企业工作流,智能体策略才更可防守。[CE007, CE008, CE009, CE010, CE018, CE020]

工作流 / 用例表
用户任务当前工作流StepFun 方案可衡量收益信号限制
构建 AI 应用开发者集成模型 API 并处理计费Open Platform、OpenAI 兼容迁移和聊天补全 API有具体文档和 token 定价未找到公开正常运行时间或延迟 SLA
运行编码或智能体工作流用户从智能体 / 编码工具调用模型Step Plan,提供专用 API key 和 Credit 额度为重复智能体使用提供订阅包装公开转化、流失或工作负载结构缺失
自托管或检查旗舰模型研究者拉取权重并用 vLLM / SGLang 部署Step-3 GitHub、Hugging Face 和部署指南开放 Apache-2.0 权重和模型卡硬件内存要求和 AFD 缺口仍然重要
创作或实验资产创作者使用 studio / playground 工作流AI Studio,提供聊天、搜索、展示、库和 PlaygroundAPI 文档之外可见产品入口使用规模和协作控制未公开
使用智能体终端消费者向手机智能体描述意图STEPX Neo,搭载 Step AOS 和 Amoo 智能体新闻报道显示模型 / OS / 硬件整合及生态伙伴官方规格、定价、出货和留存未公开

该表把产品说法转成用户工作流;可衡量收益是公开信号,不是已验证的客户结果。

[CE007, CE018, CE020, CE021, CE028, CE029]
FE003: 关键依赖图

StepFun 的护城河取决于模型架构、GPU 服务、分发渠道和伙伴生态。

[CE004, CE009, CE018, CE020, CE030, CE039]

5.4 路线图、信任控制与技术风险判断

路线图推进很快:2024 年 WAIC 展示 Step-2,2024 年推出 Step-1V 和其他多模态线,2025 年与 Geely 开放视频 / 音频模型,2025 年推出 Step-3,2026 年推出 Step-3.7 Flash 和 Step Plan,2026 年 7 月推出 STEPX Neo。这个节奏支持一个判断:StepFun 具有真实的模型开发速度,也有宽广的多模态议程。信任和部署证据更薄。公司发布了隐私、用户协议和平台管理规则,定价 / 限速文档也比通用营销页面更具体。但已审阅公开记录没有识别到官方状态页、uptime SLA、SOC 2 或等同认证、独立模型风险审计,或旗舰 Step-3 基准的第三方复现。反向观点因此不是 StepFun 没有技术,而是大型开放多模态模型栈可以被复制,基准叙事可由发行方选择,生产级企业保证还没有追上产品野心。[CE016, CE017, CE022, CE023, CE024, CE025]

信任 / 质量 / 合规与路线图表
日期或控制项公开状态范围含义剩余缺口
2024 WAIC / Step-2据报道发布万亿参数 Step-2基础模型路线图显示早期规模野心未找到详细官方技术报告
2025 Step-3开放模型、论文、代码仓库和模型卡旗舰多模态推理模型最可检查的技术资产仍需独立复现和生产审计
2026 Step-3.7 Flash官方当前面向真实世界智能体的 Flash 模型智能体编码、企业搜索和多模态输入显示模型节奏和智能体焦点外部采用指标未公开
2026 Step Plan订阅 / Credit 套餐开发者和智能体工具变现把模型调用转成打包的重复使用单位经济和续费数据未公开
2026 STEPX Neo新闻报道的智能体手机,搭载 Step AOS 和 Amoo终端硬件 / OS / 模型整合把产品线扩展到软件 / API 之外官方规格表、价格和出货未公开
隐私政策、用户协议、管理规则公开法律和平台行为文件Open Platform 治理强于没有政策入口未找到 SOC 2、状态页、SLA 或模型风险审计

路线图条目混合官方来源和明确标注的第三方报道;信任控制条目仅是公开文件证据。

[CE016, CE017, CE022, CE023, CE030, CE037]
FE004: 产品成熟度 / 能力图

StepFun 在模型宽度和 API 打包上最强,独立审计的生产保障较弱。

基于公开证据的分析师序数评分;低表示缺失,或在已审阅来源中未获独立验证。

[CE012, CE013, CE014, CE037, CE038, CE041]
FE006: 产品与模型时间线

StepFun 的发布节奏从 Step-2 和早期多模态模型,推进到 Step-3、Flash 模型、Step Plan 和 STEPX Neo。

[CE016, CE018, CE023, CE026, CE030, CE037]

5.5 图表

Chapter 06

06客户情况

6.1 客户分层:伙伴主导分发先于直客证明

StepFun 的公开客户故事异常依赖渠道。最强证据不像经典 SaaS 客户名单,更像通过手机 OEM、汽车座舱、应用商店消费者、API 开发者,以及能让 Amoo 在 STEPX Neo 内执行有用动作的生态伙伴来分发。据报道,通过 OPPO、Honor 和 ZTE 的设备集成提供了最清楚的规模化使用界面;Geely 则是最清晰的具名汽车部署伙伴。消费者 StepFun AI 应用提供评分和评论,但不提供活跃用户或变现数字。开发者 / API 来源展示目录、价格和开源界面,而不是付费账户队列。这个分层很重要,因为 StepFun 可能更接近嵌在伙伴渠道里的模型与智能体基础设施供应商,而不是拥有可独立观察终端客户需求的公司。因此,本章把伙伴、用户和付款方分开处理,而不是把所有触达都折叠成客户。实际尽调应逐个分层追问:StepFun 是否控制账户关系,是否只是提供模型能力,还是依赖另一平台暴露需求。[CU001, CU002, CU003, CU004, CU005, CU014]

客户 / 合作伙伴分群图
分群买方 / 用户 / 付款方用例公开规模信号战略价值主要缺口
手机 OEM 设备伙伴OEM 买方 / 手机用户 / StepFun 或 OEM 变现不清楚OPPO、Honor、ZTE 级设备内的端侧模型和助手能力据报道到 2025 年底 42M+ 台设备最大分发渠道和最强触达证明无伙伴级收入、排他性或活跃使用分母
汽车合作伙伴车企买方 / 司机-乘客用户Geely 智能座舱、语音模型、AgentOS、Galaxy M9 交互City News 称 Geely 部署预计到 2026 年底超过 1M 辆车具名的近生产合作伙伴和高频车内用例无合同经济、续费或单车激活指标
StepFun AI 消费者 App 用户个人 App 用户 / 订阅或免费用户助手、文档理解、任务执行、StepClaw 智能体App Store 4.7,912 个评分;Google Play 3.7,67 条评论直接消费者入口和反馈回路下载量、MAU、付费转化、留存和 ARPU 未披露
STEPX Neo 生态合作伙伴服务合作伙伴 / 手机用户 / 交易平台付款方不清楚智能体跨支付、旅行、网约车、本地服务、生产力、内容执行首批合作伙伴包括 Alipay、Meituan、Didi、Trip.com、CapCut、WPS、Baidu、JD.comAmoo 要在聊天之外有用,生态必不可少集成深度、商业条款和消费者采用未验证
企业 / API 开发者开发者或企业买方 / 应用用户按 token 定价的 API,覆盖编码、RAG、智能体、视觉、长上下文和工具调用工作负载第三方目录跟踪定价和七个模型硬件伙伴之外的可扩展变现路径付费开发者账户和留存未披露
开源开发者社区开发者用户 / 不一定有直接付款方Step-Audio2 和 Step-Audio-R1 代码仓库及模型实验GitHub 组织和公开代码仓库可见生态可信度和招聘 / 开发者飞轮star 数 / 下载量和向付费 API 转化未披露

分群把用户、伙伴和付款方分开,因为公开来源很少披露谁直接向 StepFun 付费。

[CU001, CU002, CU003, CU005, CU010, CU014]
FU001: 客户旅程图

StepFun 的公开客户路径先靠伙伴触达,再进入应用 / 开发者使用,最后才走到可变现留存证明。

旅程阶段是定性证据类别,不是实测转化率。

[CU001, CU002, CU009, CU010, CU014, CU018]

6.2 采用证明:42M+ 设备触达很强,但分母不均衡

最具体的采用指标,是据报道截至 2025 年底通过手机伙伴达到超过 42 million 的设备安装基数。这是有意义的分发信号,也比单纯 logo 页更强,因为它把 StepFun 模型绑定到已出货设备。但分析一旦从伙伴触达转向使用质量,公开记录就更薄。Geely 报道指向智能座舱共研和 2026 年部署预期,但本章没有生产激活率、单车使用量或合同经济性。STEPX Neo 具有战略重要性,因为它可能让 StepFun 成为直接消费者硬件拥有者;但发布报道反复提示价格、规格、出货目标和销售时间缺失。StepFun AI 应用评分显示真实消费者存在,API 定价来源则显示开发者可接入。没有一个来源披露留存、ARPU、付费用户转化,或按渠道拆分的收入贡献。因此,设备数字是漏斗顶部 KPI,而不是完整采用指标;应与日活请求、经常性合同和伙伴发票核对。[CU002, CU005, CU006, CU007, CU009, CU010]

客户增长 / 采用轨迹表
指标或里程碑数值 / 观察日期或新鲜度来源可信度含义缺失分母
手机设备安装基数通过手机品牌合作超过 42M 台设备针对 2025 年底的报道本章最强规模证明活跃用户、模型调用、收入分成、伙伴拆分
主要手机品牌覆盖约占中国主要手机品牌的 60%;具名 OPPO、Honor、ZTE2026 年报道显示分发宽度合同期限和排他性
Geely 车辆部署预计到 2026 年底超过 1M 辆车2026 年预期汽车渠道可能成为大型装机基数实际生产激活和单车使用
STEPX Neo 发布智能体手机随 Step AOS 和 Amoo 亮相2026 年 7 月潜在的直接消费者硬件切入价格、规格、上市日期、出货
StepFun AI iOS 评分4.7,来自 912 个评分2026-07-21 观察消费者 App 可见,且已有足够使用量产生评分下载量、MAU、付费转化
StepFun AI Google Play 评分3.7,来自 67 条评论;2026 年 7 月 9 日更新2026-07-21 观察消费者 App 有 Android 反馈,但可见评论基数更小下载量、留存、地域
API 定价入口Step 3.5 Flash 定价和模型目录可见第三方于 2026 年 6 月下旬验证开发者采用可建立在价格 / 性能上付费账户、使用量、SLA 层级结构

数值混合已披露触达、应用商店信号和第三方 API 目录观察;空缺分母是明确的尽调缺口。

[CU002, CU003, CU004, CU006, CU009, CU012]
具名客户证明表
实体分群部署或用例生产 vs 试点 / 伙伴状态结果信号限制
OPPO手机 OEM端侧 StepFun 模型集成具名手机品牌合作属于据报道 42M+ 设备基数的一部分无公开收入、排他性或模型调用指标
Honor手机 OEM端侧 StepFun 模型集成具名手机品牌合作属于据报道 42M+ 设备基数的一部分无公开收入、排他性或模型调用指标
ZTE手机 OEM / 行业参与者端侧 StepFun 模型集成和 WAIC 智能体手机语境具名手机品牌合作属于据报道 42M+ 设备基数的一部分无公开收入、排他性或模型调用指标
Geely汽车智能座舱、语音模型、AgentOS、Galaxy M9 类人 AI 智能体战略技术生态 / 联合开发伙伴部署预计到 2026 年底超过 1M 辆车实际活跃使用和合同经济未披露
Alipay支付生态伙伴AI 原生支付基础设施和 STEPX / Amoo 任务执行战略生态伙伴让智能体进入支付路径商业条款和 API 深度未披露
Meituan / Didi / Trip.com / CapCut 等伙伴消费者服务生态伙伴Amoo 工作流中的本地服务、网约车、旅行预订、视频编辑首批生态合作伙伴让智能体手机工作流更可信不能证明这些实体向 StepFun 付费
Kingdee企业软件合作伙伴行业智能体服务合作战略合作企业服务渠道信号未披露付费客户数或部署结果
Huaqin制造 / ODM 合作伙伴据报道为首款 AI 智能体手机制造商制造合作伙伴 / 依赖支撑 STEPX Neo 硬件落地不是客户需求信号

该清单只列出审阅来源中有具名公开证据的设备与合作关系;不是非公开客户的完整列表。

[CU002, CU003, CU005, CU006, CU007, CU010]
设备 / 合作关系表
合作伙伴或设备渠道关系类型据报道规模或角色客户证据强度主要限制
OPPO手机 OEM 集成见于 42M+ 设备基数报道分发触达证据强收入或排他条款未公开
Honor手机 OEM 集成见于 42M+ 设备基数报道分发触达证据强收入或活跃使用分母未公开
ZTE手机 OEM / 设备生态见于设备基数报道和 WAIC 智能体手机语境分发触达证据中高合作深度和付费使用未公开
Geely汽车智能座舱突出展示 AgentOS 和语音模型;2026 年预期覆盖 1M+ 辆车具名汽车部署证据强激活量和合同经济性未公开
Huaqin制造 / ODM据报道为 STEPX Neo 制造商硬件落地证据中制造角色不代表客户需求
Alipay / Meituan / Didi / Trip.com / CapCut 等生态伙伴STEPX Neo 服务生态Amoo 任务执行的首批应用集成生态深度证据中不是直接付费客户证据
Kingdee企业软件合作伙伴围绕行业智能体服务的战略合作企业渠道证据中部署结果或付费账户数未公开

清单仅限具名公开关系;私有商业客户和合同条款未披露。

[CU002, CU003, CU005, CU006, CU007, CU010]
FU002: 采用 / 部署漏斗

公开证据从广泛伙伴触达收窄到极有限的直接变现证据。

零值表示公开披露指标为零,不代表实际客户或留存为零。

[CU002, CU006, CU014, CU015, CU026, CU027]
FU004: 留存 / 复购队列可见度

所有可见渠道都缺留存证据。

因为未找到公开留存百分比,表中展示披露可见度,而非实际留存表现。

[CU016, CU023, CU027, CU038]
FU005: 设备触达 KPI 条形图

披露的设备和车辆触达规模远高于已披露的直接用户和留存指标。

零值表示缺少公开披露,不代表实际付费客户或留存为零。

[CU002, CU006, CU014, CU015, CU026, CU027]

6.3 具名客户证明:伙伴有名字,付费客户和留存没有

具名证明表刻意保守。OPPO、Honor、ZTE、Geely、Alipay、Kingdee 和 Huaqin 都是公开报道中的真实具名实体,但角色不同。手机厂商和 Geely 更像分发或部署伙伴;Alipay、Meituan、Didi、Trip.com、CapCut 等应用公司是生态参与者;Kingdee 和 Alipay 报道指向企业服务与支付基础设施合作;Huaqin 是制造依赖。这些信号有价值,但不等同于已披露付费客户数量或可引用企业合同。留存几乎完全私有:已审阅公开材料没有出现净留存率(NRR)、总留存率(GRR)、流失率、续约率、合同长度、队列留存、头部客户集中度或分层 ARPU。这意味着客户质量必须通过伙伴耐久性和分发杠杆承销,而不是传统经常性收入指标。后续客户访谈应区分试点、付费生产部署和联合营销公告,避免把伙伴可见度误认成合同客户质量。[CU003, CU005, CU007, CU010, CU014, CU015]

留存 / 复用 / 满意度表
指标数值或公开状态分部置信度尽调问题
NRR / GRR / 流失未披露全分部索取按渠道拆分的队列留存、总留存和净留存
续约 / 合同期限未披露OEM、汽车、企业 / API索取合同条款、续约日期和终止权
StepFun AI iOS 满意度4.7 分;912 个评分消费者应用索取活跃用户、付费转化和评论队列历史
StepFun AI Android 满意度3.7 分;67 条评论消费者应用索取国家拆分、安装量和月度留存
API 可靠性代理指标LLM Stats 跟踪模型的 7 天成功率 98.3%开发者/API索取内部 SLA、正常运行时间、付费用量和支持工单
头部客户集中度未披露;反向来源称依赖少数终端伙伴OEM / 汽车索取前 5 大合作伙伴收入和排他条款

留存经济性基本空白;应用评分和 API 可靠性只是弱代理指标,不能替代队列数据。

[CU014, CU015, CU016, CU017, CU020, CU027]
FU003: 客户证明矩阵

伙伴分发的证据质量最高,留存经济性的证据最弱。

矩阵单元格是对公开证据的定性判断。

[CU003, CU005, CU010, CU014, CU015, CU020]

6.4 耐久性风险:OEM 锁定和生态权限仍是关键测试

反向证据直接削弱了简单的客户牵引叙事。GSMArena 质疑智能体手机发布背后有多少实质;Gogi 警告买方不要等待一款时间、定价或国际可用性都未确认的产品。中文分析对 StepFun 伙伴模式更尖锐:Toutiao 认为 OEM 关系并非独家,OPPO 或 Geely 可以切换或多源采购模型,而少数终端伙伴周围的收入集中可能变成断崖风险。BigGo 的发布分析增加了另一个运营约束:Amoo 的有用性取决于超级应用是否提供深且持久的接口,以及用户是否愿意交出敏感权限。尽调结论不是客户牵引不存在,而是公开证据支持广泛分发,却不支持留存耐久性。后续尽调应要求合同、独家条款、活跃用户队列、伙伴级收入、续约历史和权限接口深度。这个反向姿态尤其重要,因为如果最重要的应用限制自动化,或把关键工作流留给自家助手,智能体手机的用处就会下降。[CU012, CU013, CU028, CU029, CU030, CU031]

扩张与集中度风险表
驱动因素或风险对客户的影响当前判断降低风险的证据
OEM 装机基数扩张可借已出货手机快速放大触达强,但由伙伴控制已签多年合同、活跃使用数据、模型调用量
Geely 智能座舱渠道可能把模型能力转成大规模车端使用具名证据和 2026 年部署预期量产激活、车内使用、续约经济性
STEPX Neo 直接硬件可能掌握用户记忆和智能体交互层尚未验证;价格、规格、日期未披露零售发布、出货量、复用、伙伴 API 深度
中国超级应用集成让 Amoo 能用于支付、出行、本地生活和内容有潜力,但依赖权限深层接口、Tencent / WeChat 访问、可审计执行指标
开发者/API 价格差低价模型可能吸引开发者目录可见,但无账户队列数据付费开发者数、留存、用量、企业 SLA
非排他合作伙伴关系合作伙伴可多源采购或切换模型反向分析中的实质风险排他性、切换成本证据、伙伴层面总留存

风险评级为定性判断,因为合作伙伴合同和收入集中度未公开。

[CU006, CU010, CU012, CU020, CU021, CU022]

6.5 图表

Chapter 07

07风险

7.1 监管、IPO 与法律结构风险

StepFun 的风险栈从监管开始,因为公司卖的不是普通软件;它运行在全球监管最活跃的 AI 体系之一。中国生成式 AI 规则、当前 CAC 备案公告、深度合成规定和国家 AI 内容标识标准,为任何面向公众的模型、API、智能体或设备界面列出一张实时运营清单。StepFun 自家法律页面展示了基础条款和隐私披露,但不能证明模型备案、标识落地、安全审计或企业控制准备已经完成。IPO 层进一步放大这个问题:Reuters 转发报道称,在北京审查红筹上市的背景下,公司拆除了离岸结构;其他来源则描述股改和香港 IPO 准备。这并不意味着上市受阻,但会让法律执行成为高影响路径依赖,而不是行政细节。[CR001, CR002, CR003, CR004, CR005, CR006]

监管 / 法律风险登记表
风险类别可能性影响证据缓释措施 / 尽调问题
生成式 AI 备案和模型展示缺口监管CAC 规则和 2026 年 7 月备案名单要求备案并展示模型名称或编号核验 StepFun 模型备案、产品页展示和重大变更备案流程
AI 生成内容标识不合规监管 / 法律GB 45438-2025 和标识办法要求显式与隐式标识审查可见标识、元数据 / 水印设计、下载留存和平台传递控制
深度合成治理缺口监管中高深度合成规则要求用户注册、算法审核、伦理审查、内容审核和应急响应索取内部安全管理制度、滥用响应日志和监管沟通记录
AI 智能体自主性治理缺口监管 / 产品2026 年智能体框架将自主感知、记忆、决策、交互和执行列为单独政策类别将 StepFun 智能体功能映射到用户授权、人工控制、工具使用和审计要求
隐私和客户内容处理风险法律 / 数据中高StepFun 隐私政策覆盖用户输入、输出、企业信息、API 密钥、支付、设备和日志审查 PIPL 依据、留存、删除、训练使用政策和企业 DPA 条款
条款只是基础卫生项,不是企业合规证明法律 / 商业StepFun 发布 ToS、责任限制和仲裁条款索取安全认证、DPA、客户审计权、事件流程和特定行业附录
红筹架构调整延缓 IPOIPO / 法律架构中高Reuters 转载报道称,北京审查推动 StepFun 解除境外架构,类似动作可能拖慢上市审查股改、CSRC/HKEX 律师备忘录、税务影响、投资人同意和备案时间表
间接制裁或实体清单风险敞口地缘政治 / 法律低-中美国芯片和中国 AI 政策报道显示,即使 StepFun 未被点名,管制也会快速变化,仍有进入实体清单的风险监测 BIS / 实体清单变化、客户地域、投资人敞口和美国原产技术依赖

清单按公开法律、监管、IPO 和地缘政治来源做严重性排序;可能性和影响是尽调判断,不是公司披露的风险评分。

[CR001, CR002, CR003, CR004, CR005, CR006]
监管义务表
义务来源 / 制度StepFun 风险敞口公开证据状态剩余风险尽调问题
面向公众的生成式 AI 服务备案CAC 暂行办法和备案公告在中国提供的模型、应用和 API 集成功能制度已生效;此处未审阅 StepFun 专属备案文件中高获取模型备案编号和重大变更备案历史
展示已备案模型信息CAC 2026 年 7 月公告产品页和 API / 应用详情CAC 称线上应用应披露模型名称以及备案号或上线编号检查所有 StepFun 网页、应用、API 和设备触点
防止禁止或有害输出生成式 AI 暂行办法文本、图像、语音、视频、API 和智能体输出一般规则适用;公开安全控制细节有限中高审查红队结果、内容过滤日志和升级流程
深度合成服务提供者控制深度合成规定语音、图像、视频和多模态生成用例规则要求注册、审计、伦理审查、内容审核和应急系统中高索取深度合成治理政策和审计证据
显式和隐式 AI 内容标识GB 45438-2025 和 2025 年标识办法平台和设备合作伙伴上的生成内容和可下载文件国家标准在本次报告日期前已生效测试可见标识、元数据、水印和伙伴传递
智能体自主性治理2026 年 AI 智能体实施意见StepAI、工具使用、设备、汽车和智能体工作流框架较新,可能继续演变中高绘制智能体权限、人工控制、日志和安全边界
个人信息保护隐私政策加 PIPL / 数据安全 / 网络安全框架输入、输出、企业信息、API 密钥、支付、设备和日志隐私政策已发布,但企业控制文档未公开审查留存、删除、训练使用排除、DPA、跨境传输和泄露流程

义务来自监管机构、法律分析和 StepFun 法律页面来源;公开证据不能证明 StepFun 具体执行质量。

[CR002, CR003, CR004, CR005, CR006, CR007]
FR004: 风险时间线与监测条

关键风险监测窗口从已生效的旧规延伸到 2026 年的智能体、芯片、备案和 IPO 事件。

[CR002, CR003, CR005, CR006, CR007, CR008]

7.2 算力、地缘政治与供应链风险

最具体的运营依赖是算力。StepFun 的大模型、API、手机和汽车野心,都需要可靠的模型训练、推理和边缘优化。公开来源没有显示 StepFun 的内部 GPU 合同,但行业背景已经足够不利,应把算力视为核心残余风险。美国出口管制分析、2026 年 H200 许可争论,以及海外中国实体受限报道都显示,Nvidia 级硬件获取可能由政策改变,而不只取决于采购能力。国产替代是部分缓释,不是完整解法:以 Huawei 为代表的产能正在中国提升份额,但 CFR 技术分析认为,Huawei 在前沿性能上仍明显落后于 Nvidia。对 StepFun 而言,GPU 风险是双向的:过度依赖美国来源芯片会带来地缘政治敞口,快速迁移到国产芯片又会带来优化、成本、性能和交付风险。[CR027, CR028, CR029, CR030, CR031, CR032]

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决缺口
受控 Nvidia 访问拖慢训练或推理规模部分;H200 访问似乎有条件且依赖政策算力合同和应急容量未公开
迁移到国产 Ascend/Huawei 后表现不及前沿 Nvidia 栈部分;国产芯片份额上升,但性能差距仍有争议中高StepFun 工作负载上的基准性能对齐和每 token 成本未公开
设备和汽车铺开暴露安全、标识和支持缺口手机 / 汽车给出早期分发信号,但运营控制未公开中高伙伴 QA、召回、日志和用户升级流程未公开
API 企业用户提交敏感客户内容中高已有隐私政策和条款中高需要训练使用政策、DPA、删除 SLA 和租户隔离证据
智能体 / 工具使用输出引发未授权操作或审计失败2026 年框架释放政策方向;StepFun 控制措施未公开需要权限边界、人工批准和操作日志
价格战驱动的成本削减压低可靠性或支持投入无直接证据显示 StepFun 削减投入毛利率、支持团队配置和事故历史未公开

运营严重性根据公开监管、芯片、设备和法律证据推断;未找到 StepFun 停机或召回来源。

[CR009, CR010, CR012, CR013, CR027, CR028]
合作伙伴 / 依赖风险登记表
依赖交易对手 / 领域角色集中度失效情景严重性缓释措施剩余敞口
高端 GPU 供应Nvidia / 美国出口管制制度训练和推理容量行业敞口高许可证、配额或海外出货管制削减容量国产与受控进口双轨策略
国产 AI 芯片Huawei 及其他中国加速器中国本土算力替代中高性能或软件栈差距推高成本并拖慢迭代模型-芯片优化和本地生态支持中高
手机 OEM 分发Oppo、Honor、ZTE 及主要手机品牌模型装机基数和消费者触达可见集中度高OEM 战略变化或伙伴经济性削弱采用证据拓宽 OEM 和应用分发渠道中高
工业供应链投资方Huaqin、Longcheer、OmniVision、ZTE资金、零部件和设备生态战略投资方优先优化自身硬件路线图,而不是 StepFun 利润率中高清晰的商业合同和转让定价经济性
监管审批路径CAC、MIIT、NDRC、标准机构上线权利和产品约束备案、标识或智能体规则拖慢发布专职合规团队和监管沟通中高
香港资本市场CSRC、HKEX、公开市场投资者IPO 流动性和估值验证中高红筹重组或估值审查延误上市境内重组、审计财务和保荐人准备度

依赖项把已点名的 StepFun 合作伙伴与市场层面的芯片、监管依赖放在一起;集中度基于公开证据,而不是内部合同数据。

[CR014, CR015, CR016, CR019, CR020, CR021]
FR003: 依赖关系图

StepFun 的依赖横跨监管方、算力供应商、国产芯片、设备 OEM、产业投资人和资本市场。

[CR002, CR014, CR016, CR019, CR020, CR021]

7.3 竞争、变现与资本风险

竞争风险比普通的“市场拥挤”警示更重。StepFun 处在中国“AI 六小虎”叙事里,但公开反向来源认为,这组同业本身正承受 DeepSeek、Alibaba、ByteDance、Qwen 以及开放权重价格 / 性能模型的压力。Forbes 报道 DeepSeek 突然降价,其他来源也把中国模型市场概括为效率、开放分发和长期价格战。这迫使 StepFun 证明,除通用 API 访问之外还能变现。AsiaICT 的反向表述尤其相关,因为它直接点出核心问题:盈利模式未证实、依赖少数硬件厂商、成本重估压力。大额融资和 IPO 野心买来了时间,也抬高了证明门槛。如果 StepFun 不能把设备安装、企业使用和智能体功能转成持久利润率,公开市场投资者可能会认为估值叙事跑在经济性前面。[CR017, CR018, CR019, CR020, CR021, CR033]

FR001: 风险热力图

StepFun 剩余风险最高的部分集中在监管合规、算力获取、价格战变现、伙伴集中和 IPO 执行。

可能性和影响的序数判断来自公开证据和尽调判断,不是公司预测。

[CR003, CR006, CR008, CR018, CR027, CR031]

7.4 人才、执行与伙伴依赖风险

StepFun 的人才叙事很强,但这一叙事本身也是风险因素。Jiang Daxin 的 Microsoft、STCA 和 IEEE Fellow 履历是有价值的信号,但也让公司故事异常依赖一位创始人的科学声望,以及他持续凝聚顶尖模型人才的能力。2026 年 1 月 Yin Qi 出任董事长,是有意义的缓释:公司在团队里加入了商业化与 AI 硬件负责人。它也带来执行挑战:同一批公开来源强调 AI + 硬件、智能汽车、设备和 Qianli 相关经验,因此公司如今要同时协调前沿模型、手机 OEM、汽车伙伴、芯片约束和 IPO 工作。公开报道列出了核心管理团队,却没有披露董事会架构、继任计划、留任激励或运营节奏,投资者因此难以有把握地承销关键人和跨领域执行风险。[CR020, CR021, CR022, CR023, CR024, CR025]

人员 / 执行风险清单
角色 / 职能依赖或缺口可能性严重性缓释措施尽调路径
创始人 / CEO Jiang Daxin公司叙事和技术可信度高度依赖创始人声誉Microsoft、STCA、IEEE 和论文履历强评估决策权、接班安排、留任和梯队厚度
董事长 Yin Qi带来商业化和硬件能力,但也增加双重角色和协调复杂度中高已获任董事长,统筹战略和技术方向审查与 Qianli 的运营节奏、StepFun 董事会角色和冲突控制
核心技术领导层首席科学家和 CTO 已公开具名,但团队留任经济安排未披露公开管理团队不只创始人一人要求提供组织架构图、留任计划、人才流失,以及竞业 / 禁止挖角风险
IPO 执行团队红筹拆除、股改、估值、审计和 HKEX 流程必须合拢中高据报道已推进股改和红筹拆除审查保荐人时间表、审计准备度、法律重组备忘录和税务成本
AI + 设备执行公司必须协调模型、芯片、手机、汽车、智能体和 API 路线图工业投资方和 OEM 合作伙伴可见要求提供项目管理指标、合作伙伴 SLA、发布 QA 和上线后支持证据

人员风险基于公开领导层报道和 IPO 报道;薪酬、接班、董事会监督和留任文件仍属私密。

[CR022, CR023, CR024, CR025, CR026, CR038]
FR002: 风险传导图

监管、算力和价格战冲击会传导到产品节奏、利润率质量、IPO 准备度和估值。

[CR008, CR014, CR015, CR022, CR024, CR027]

7.5 缓释因素、监测指标与否决条件

今天可见风险没有单项致命,但剩余风险堆栈仍高,因为多条风险链相互强化。监管合规影响产品发布节奏;产品发布节奏影响企业和设备变现;变现影响 IPO 准备度;IPO 准备度又影响公司继续为昂贵算力和人才融资的能力。公开缓释因素存在:国资相关和产业投资人、StepFun 法律页面、扩大的领导团队、主要设备伙伴,都降低了这是空转概念故事的概率。它们不能替代私下尽调。投资逻辑破裂触发器应当具体:CAC 备案和模型公示证据缺失或过期、AI 内容标识执行薄弱、没有清晰算力应急方案、DeepSeek 式价格压力下看不到毛利率路径、红筹或股改问题未解决,或者上市前核心技术领导团队离职。[CR040, CR041, CR042, CR043, CR044, CR045]

缓释措施与止损标准表
风险可监控触发项阈值 / 事件行动含义
生成式 AI 合规CAC 备案和展示证据面向公众的产品缺少模型备案号,或展示信息过期备案文件齐备前,暂停产品风险研判
标识和深度合成控制AI 标识、元数据和滥用日志GB 45438-2025 生效后,生成内容仍没有显式 / 隐式标识将受监管发布准备度视为未证实
算力供给芯片获取和国产迁移证明没有可信的 H200 / 国产算力计划,或负载下每 token 成本上升提高烧钱和执行风险折价
价格战下的变现毛利率和付费使用队列API / 设备收入增长,但在 DeepSeek / Qwen 降价下单位经济性恶化要求更低入场价,或等待证据
合作伙伴集中度OEM 和工业合作伙伴经济性多个合作伙伴的设备装机没有转化为付费或留存使用下调客户和渠道投资逻辑
IPO 执行香港申报、股改、审计和估值区间红筹或 CSRC / HKEX 问题让申报延后,错过下一个预期窗口将流动性和估值叙事视为受损
关键人物和领导层创始人 / 董事长 / CTO 留任与治理IPO 准备完成前,核心领导离任或角色冲突升级至投资逻辑破裂复盘

止损标准是尽调阈值,不是公司指引;当私密财务和申报材料可用时,应重新审视这些阈值。

[CR003, CR004, CR006, CR007, CR014, CR015]

7.6 图表

Chapter 08

08估值

8.1 投资建议与价格纪律

估值判断对价格敏感,而不是对公司质量敏感。StepStar 融资动能异常强:公开来源核验了 2024 年 12 月由上海国资、Tencent、Qiming 等投资人支持的 Series B;多家 2026 年来源核验了超过 RMB5 billion 的 Series B+;后续报道还描述了来自产业硬件玩家、国资相关投资人和既有创投股东的 Pre-IPO 或 IPO 路径融资。这一组合让香港上市具备可信度。承销难题在于,头部价格比公开基本面变化更快。本次审阅来源把公司在 Pre-IPO 语境下估到大约 $4 billion 至 $6 billion;部分融资数据集约 $10 billion;IPO 报道为 $12 billion;更激进市场传闻接近 RMB90 billion 或更高。本文审阅的 StepStar 官方页面和公开文件没有提供经审计收入、毛利率、烧钱速度、算力承诺、优先股堆叠或确定版招股书。因此,新投资人不应把最高 IPO 标记视为已证实价值;那只是动能标记,需要数据室确认。[CV001, CV002, CV003, CV004, CV005, CV006]

建议摘要表
决策字段当前观点决策含义
建议继续研究 / 观察继续跟进 HKEX 申报,但在缺少经审计运营指标时,不要为最高头条估值买单。
信心融资和可比公司证据面较宽;StepStar 自身收入、烧钱和条款仍是私密信息或媒体估计。
风险评级IPO 执行、上市后波动、算力成本和私募条款不透明,都可能压缩价值。
估值立场偏高$10B-$12B 在可比窗口内有合理性,但近 RMB90B 乃至更高传闻,需要招股书级别证据。
入场纪律要求已提交招股书或折价买方在支付头条价格前,应要求看到官方收入、利润率、算力、客户和优先权细节。

这是 IC 建议表:它把公司动量与价格质量拆开,并在申报出现前,将私募媒体估计视为未经验证。

[CV001, CV005, CV006, CV007, CV008, CV014]
估值历史与冲突标记表
日期 / 窗口估值或融资标记来源姿态尽调解读
2024 年 12 月 Series B 轮数亿美元 Series B 轮;提示中 ~$1B 的估值锚点未被保留的公开来源证实。SCMP、SiliconAngle、TMTPost 和 Dealroom 证实该轮融资 / 资方,但没有给出一致估值。将其视为早期独角兽锚点,同时保留重大估值披露缺口。
2026 年 1 月 Series B+ 轮融资超过 RMB5B / 约 $700M-$717M。36Kr、KR Asia、Tencent News 和 Aibase 在规模和投资方组合上趋同。资本获取能力信号强;估值仍未统一披露。
2026 年 2–4 月 Pre-IPO 规划Caijing / Sina 报道分批融资,投前估值约 $4B 和 $5B-$6B。媒体称截至发稿管理层尚未公开回应。比 $12B 更可承销,因为它更贴近近期私募融资语境。
2026 年 5 月工业轮报道腾讯转载报道中,据称融资近 $2.5B,投后估值 $5B-$6B。报道称工业投资方包括 Huaqin、Longcheer、OmniVision、ZTE 和 HKIC。支持战略财团溢价,但也显示与 $10B-$12B 目标冲突。
2026 年 6 月 IPO 报道据报道 IPO 估值目标最高 $12B / 超过 RMB80B。Sina、163、StartupWired 和 The AI Chronicle 承载高 IPO 叙事。IPO 要价有合理性;尚不是已完成的公开市场估值。
2026 年 7 月市场传闻来源BestStartup 和 Oryndex 引用约 $10B;Newsglobenow 提到近 RMB90B 量级或更高的二级市场需求。各报道的来源质量和单位换算差异很大。明确保留所有数字;不要把它们压成一个虚假精确的标记。

列举并不完整:它覆盖截至 2026-07-21 保留的公开融资 / IPO 标记,并刻意保留不一致估值,而不是取平均。

[CV001, CV002, CV003, CV004, CV005, CV006]
FV001: 推荐逻辑

融资动能支撑 IPO 准备度,但估值标记冲突、申报文件缺失压住推荐上限。

这是定性的投委会逻辑链,不是数学模型。

[CV001, CV005, CV006, CV014, CV015, CV039]
FV002: 估值时间线与敏感性

公开估值标记从 Series B/B+ 阶段到 IPO 传闻大幅抬升,但单位和信源质量彼此冲突。

USD 等值已四舍五入;本图保留彼此冲突的公开估值标记,并不表示所有数字同样可靠。

[CV002, CV003, CV006, CV008, CV011, CV012]

8.2 估值支撑与倍数上限

支撑溢价估值的最强证据不是披露的软件倍数,而是战略稀缺性。StepStar 背后有国资相关资本、Tencent/Qiming/FiveYuan 式财务投资人,以及后续报道中的 Huaqin、Longcheer、OmniVision、ZTE 等硬件生态投资人。来源还指向终端设备部署、移动 / 汽车合作,以及已有公开开发者入口的模型平台。这些都是真实战略信号,因为中国基础模型赢家需要资本、分发、算力获取和嵌入硬件的路径。限制因素是,基于公开证据无法干净地做收入倍数工作。Caijing 与一篇转载的 Tencent 文章报告 2025 年收入接近 RMB500 million、2026 年预计约 RMB1.2 billion,后续市场报道也重复这些数字;但这些是媒体估计,不是经审计的公司披露。用它们去计算 $10 billion、$12 billion 或接近 RMB90 billion 标记下的市销率,会制造虚假精确。因此,本章采用里程碑 / 情景估值,把收入披露视为尽调缺口,而不是表格分母。[CV016, CV017, CV018, CV019, CV020, CV021]

投资逻辑 / 反向逻辑表
论点方向什么会改变观点
国资相关、Tencent / Qiming / FiveYuan 以及工业资本参与,给 StepStar 带来异常深的融资通道。投资逻辑如果最终招股书显示融资偏优先权堆叠,或主要由内部人支持,投资逻辑会走弱。
硬件和终端生态投资方可以把模型能力转成嵌入式分发。投资逻辑若披露部署经济性、API 收入和客户集中度,投资逻辑会增强。
香港 AI IPO 可比案例显示,投资者能以很高估值资本化稀缺的中国模型实验室上市标的。投资逻辑如果 StepStar 定价前 Z.AI 或 MiniMax 估值波动延续,投资逻辑会走弱。
官方公开渠道没有披露经审计收入、烧钱、利润率或股权结构条款。反向逻辑一份已提交且财务可信的招股书,会补上最大承销缺口。
公开估值标记彼此冲突,从 $4B-$6B、$10B-$12B 到近 RMB90B 或更高传闻不等。反向逻辑如果有约束力的基石价格、最终发行区间和机构簿记质量证实需求,观点会改善。
中国 AI 泡沫批评认为,算力约束和盈利能力弱会让头条估值变脆。反向逻辑如果 StepStar 展示出明显优于同行的收入质量、毛利率和算力效率,风险会下降。

表格刻意保持对称:StepStar 可以具备战略价值,但在错误入场估值下仍可能过贵。

[CV016, CV017, CV018, CV019, CV020, CV021]
乐观 / 基准 / 悲观情景表
情景关键假设估值 / 回报逻辑概率信号
乐观HKEX 申报顺利落地,收入估计得到验证,终端 / 汽车 / API 部署转化为持久收入,可比公司保持强势。$12B+ IPO 定价可以通过;只有官方财务和簿记质量都强,近 RMB90B 区间才撑得住。在申报收入、亏损和投资者锁定结构可见前,概率较低。
基准IPO 进程继续,但最终区间落在近期私募标记与最激进头条之间。约 $8B-$10B 是工作区间;投资者回报取决于能否避开 MiniMax 式上市后压缩。最可能,因为它调和了战略需求和冲突的公开标记。
悲观申报延误、收入不及预期、算力成本压力、公开市场可比公司走弱,或重优先权条款浮现。下轮降价或破发式 IPO 重置到 $4B-$6B 变得合理,与早前 Pre-IPO 报道吻合。概率偏高,因为公开市场可比公司波动大,官方招股书仍缺席。
止损 / 投资逻辑破裂没有已提交招股书、财务披露差,或估值在缺乏证据下定在 $12B 以上。除非大幅低价或保护性条款抵消证据缺口,否则避免投新钱。如果稀缺性溢价在基本面追上前消退,下行会是二元式。

除非行明确提到 RMB,否则区间都是以十亿美元计的情景包络;由于官方收入和利润率未披露,它们不是 DCF 输出。

[CV023, CV024, CV025, CV026, CV027, CV028]
FV003: 估值 / 回报区间

由于官方披露不足以计算常规倍数,情景估值区间统一以 USD billions 表示。

区间是里程碑 / 情景包络,不是 DCF 或收入倍数输出。

[CV006, CV008, CV012, CV013, CV027, CV029]
FV005: 投资 KPI

StepStar 在融资通道和可比窗口上得分高,但披露质量和估值支撑得分低。

1-5 的序数评分综合公开证据;下行风险数值越高,风险越大。

[CV003, CV014, CV023, CV027, CV039, CV040]

8.3 中国 AI 可比公司与 IPO 窗口

最有用的可比对象不是美国软件倍数,而是 2026 年香港 AI 上市参照和中国后期模型实验室融资轮。Zhipu/Z.AI 与 MiniMax 提供公开市场类比,Moonshot 提供私募估值类比。关键信号是,香港投资者愿意给稀缺 AI 上市公司很大的市值:StockAnalysis 显示,Z.AI 在 2026 年 7 月 20 日约 HK$397 billion,4 月 30 日约 HK$386.99 billion;MiniMax 5 月 14 日约 HK$266.59 billion,之后到 7 月 20 日跌至约 HK$60.56 billion。波动性很关键:同一组可比公司既支持 IPO 需求的乐观情景,也支持上市后估值压缩的悲观情景。Moonshot 据报道 $20 billion 至 $30 billion 的私募标记说明,StepStar 的 $10 billion 至 $12 billion IPO 目标在中国前沿实验室语境下并不荒唐;但可比表也表明,市值可能跑赢已披露收入,然后迅速修正。[CV027, CV028, CV029, CV030, CV031, CV032]

可比估值表
可比对象指标 / 估值标记与 StepStar 的相关性局限
Zhipu / Z.AIStockAnalysis 显示 2026-07-20 市值约 HK$397.02B,2026-04-30 为 HK$386.99B;SCMP 和 Straits Times 描述其 IPO / 售股背景。为中国基础模型标的设定了很高的香港公开市场边界。已上市可比公司自身有波动;不能直接证明 StepStar 价值。
MiniMaxStockAnalysis 显示 2026-05-14 市值为 HK$266.59B,但到 2026-07-20 只有 HK$60.56B;CNBC 描述其首日翻倍以及收入 / 亏损背景。同时显示 IPO 胃口和上市后压缩风险。业务组合和时间点不同;市值数月内剧烈变化。
Moonshot AI / KimiTechCrunch 和其他来源报道估值约 $20B;Yahoo 报道 Moonshot 在 Kimi K3 后接近 $30B。私募前沿实验室可比案例,说明 StepStar 的 $10B-$12B 目标并非孤例。私募轮,不是可交易的公开市场标记;来源集合来自媒体报道。
StepStar Pre-IPO 私募标记Caijing / Tencent 报道投前 $4B、随后 $5B-$6B 的分批融资;BestStartup / Oryndex 引用约 $10B。基准情景估值的直接公司锚点。媒体标记冲突;最终发行区间不可见。
StepStar 高 IPO 目标Sina / 163 / AI Chronicle / StartupWired 报道最高 $12B 或超过 RMB80B。与预期香港定价诉求直接相关。IPO 目标不是已执行市场价值,可能调整。
AI 估值过高 / 泡沫批评ChinaBizInsider 和 NationPress 警告,收入 / 盈利能力不足以支撑估值,且存在算力压力。为偏高私募标记提供反向边界。行业层面批评;需要 StepStar 自身财务来量化影响。

列举是直接相关的中国 AI 公开 / 私募估值参考和反向市场校验样本,不是完整的全球 AI 可比公司范围。

[CV027, CV028, CV029, CV030, CV031, CV032]
FV004: 中国 AI 可比公司矩阵

可比公司显示,IPO 稀缺性上行空间与上市后压缩风险之间差距很大。

矩阵条目是筛选后的估值标记,不是标准化倍数。

[CV006, CV008, CV011, CV012, CV013, CV027]

8.4 下行触发器与最终尽调问题

反向情景很直接:中国 AI 估值可能被资本稀缺、国家冠军叙事和 IPO 稀缺性拉动,跑得比基本面追赶更快。反向来源警告,算力成本、token 配给,以及相对收入 / 盈利能力的高估值,可能把当前周期推成类似泡沫的结构。对 StepStar,具体的降轮或 IPO 破发风险包括:重组或备案审核耗时超预期,导致 HKEX 时间表滑坡;招股书显示收入或亏损显著弱于媒体估计;公开市场投资者给私募或基石价格打折;MiniMax 式波动削弱下一单上市的需求;或者优先股条款过重,使头部估值不能很好代表普通股价值。因此,决定性尽调问题必须具体:已提交招股书、经审计收入、毛利率、算力承诺、客户集中度、股权结构表、清算优先权以及锁定期 / 老股出售结构。在这些材料可得之前,建议维持继续研究 / 观察,估值立场偏高。[CV039, CV040, CV041, CV042, CV043, CV044]

投资逻辑破裂与止损触发器表
触发项阈值 / 事件对投资逻辑的传导行动含义
HKEX 申报延误在已报道重组和 6 月时间窗口之后,仍无招股书或正式申报路径。把 IPO 准备就绪叙事转成执行风险。转为等待申报;不要支付 IPO 溢价。
收入证明缺口申报收入显著低于媒体估计,或 API / 终端 / 汽车需求没有分部经济性。打破倍数支撑故事。按 $4B-$6B 或更低重新承销。
算力成本压力毛利率、云容量或 token 配给暴露扩展性差。高增长变贵,可持续倍数下降。要求折价,并尽调算力承诺。
公开可比公司压缩StepStar 定价前,Z.AI 或 MiniMax 市值继续下跌。降低投资者对又一家中国模型实验室 IPO 的胃口。推迟入场,或要求更小估值区间。
优先权悬顶新老投资者持有偏重下行保护的条款、清算优先权或特殊权利。头条估值高估普通股价值。投资前先搭瀑布模型。
缺乏支撑的高目标最终区间在没有经审计财务证据下瞄准 $12B+ 或近 RMB90B。稀缺性溢价压过承销纪律。回避,或只有强下行保护时参与。

触发器设计成可监控项,等招股书、基石簿记或融资文件可用时即可检查。

[CV023, CV026, CV028, CV029, CV033, CV039]
最终尽调问题表
主题缺失证据为什么重要尽调路径
官方申报HKEX 申请证明、招股书、风险因素、募资用途和发行区间。把媒体报道转成有法律责任的披露。跟踪 HKEXnews 和公司公告。
收入和利润率经审计的 2024-2026 年收入、毛利率、算力成本和亏损桥。判断 $8B-$12B 能否绑定销售质量。财务资料室和招股书审查。
客户和部署质量API、终端设备、汽车、企业和消费者产品之间的收入拆分。区分装机量叙事和可变现需求。客户访谈、合同、使用日志和合作伙伴确认。
股权结构表 / 优先权清算优先权、反稀释棘轮、按比例跟投权、锁定期、老股出售条款和基石分配。判断头条标记背后的普通股价值。法律文件审查和瀑布模型。
算力和模型经济性GPU / 云承诺、推理单位成本、利用率和容量权利。算力约束是行业层面的反向风险。对容量合同做技术 / 财务尽调。
IPO 需求质量簿记覆盖、基石集中度、机构组合和上市后锁定安排。靠稀缺性驱动的簿记上市后可能撑不住。承销商尽调和可比交易敏感性分析。

这些问题是把建议从继续研究推向买入或回避所需的最低资料室包。

[CV014, CV015, CV023, CV024, CV026, CV039]

8.5 图表

免责声明

本报告仅基于截至 2026-07-21 已审阅的公开来源,不能替代非公开财务、法律、技术和客户尽调。本报告将 StepStar 作为以 StepFun(阶跃星辰)为品牌的公司来分析。

证据索引

结论
编号陈述可信度来源
CO001 StepFun is the trade name of Shanghai Jieyue Xingchen Intelligent Technology Co., Ltd., founded on April 6, 2023 with headquarters or registered address in Shanghai. SO008, SO009, SO011
CO002 The company's official public website is https://www.stepfun.com and its official developer platform is platform.stepfun.com. SO001, SO003
CO003 Official copy frames the company vision as scaling up possibilities for everyone and making each person ten times more capable. SO001, SO002
CO004 StepFun's official product surface combines a consumer AI assistant, an API platform, Studio, and Step-series model documentation. SO001, SO003, SO004
CO005 Fetched registry and profile sources support Shanghai as headquarters and a Beijing subsidiary or office signal, while no fetched authoritative source confirmed an active Hangzhou office. SO009, SO011, SO027
CO006 The official homepage advertises Step 3.7 Flash, Step 3.5 Flash, Step 2, Step 1, API access, Studio, and a downloadable Step AI assistant. SO001, SO004
CO007 StepFun publishes pricing and rate-limit tables for API models, indicating a commercial developer API model rather than only a research-lab posture. SO006, SO003
CO008 Public profile sources identify Jiang Daxin, Zhu Yibo, and Jiao Binxing as StepFun founders or founding technical leaders. SO008, SO009, SO027
CO009 Jiang Daxin is StepFun's founder or co-founder and CEO, formerly spent 16 years at Microsoft or MSRA/STCA, rose to Chief Scientist or Global Vice President, and became an IEEE Fellow in 2024. SO010, SO026, SO009
CO010 Zhu Yibo is described as CTO or system head with prior Microsoft, ByteDance, and Google experience and responsibility for large-scale systems. SO009, SO027
CO011 Jiao Binxing is described as data head or co-founder with prior responsibility for Microsoft Bing core search systems. SO009, SO027
CO012 Yin Qi became StepFun chairman in January 2026, adding an experienced AI operator to the governance layer above Jiang's founder-CEO role. SO018, SO025, SO012
CO013 Baidu Baike identifies Zhang Xiangyu as chief scientist and Zhu Yibo as CTO/system head in StepFun's core technical team. SO009
CO014 Aiqicha lists Yin Qi as legal representative and chairman, Jiang Daxin as director and manager, and several additional directors and supervisors. SO012
CO015 Public registry aggregators show registered capital, patents, trademarks, software copyrights, and 214 insured employees, but not a complete operating headcount or cap table. SO011, SO012
CO016 The best public headcount range for StepFun is approximate: Jademond reports about 400 to 500 employees from company statements, while Aiqicha shows a lower 214 insured-person registry figure. SO027, SO012
CO017 StepFun completed a December 2024 Series B financing of several hundred million dollars involving Shanghai state-owned capital, Tencent, FiveYuan Capital, and Qiming Venture Partners. SO016, SO017
CO018 Crunchbase News listed StepStar among December 2024 newly minted unicorns, saying the Series B led by Shanghai State-owned Capital Investment valued the one-year-old Shanghai company at $1 billion. SO030
CO019 On January 26, 2026, StepFun completed a B+ financing of more than RMB5 billion, backed by Shanghai state-owned funds, China Life Equity, Pudong Venture Capital, Xuhui Capital, Wuxi Liangxi Fund, Xiamen ITG, Huaqin Technology, Tencent, Qiming, and FiveYuan. SO013, SO018, SO014, SO015
CO020 The January 2026 B+ financing was described as supporting foundation-model upgrades, frontier model research, and deeper AI-plus-terminal deployment in cars and phones. SO013, SO014
CO021 Tencent, Qiming Venture Partners, and FiveYuan/Wuyuan Capital appear as repeated investors across the Series B and B+ narratives. SO013, SO016, SO017
CO022 Eastmoney and AIBase characterize the RMB5 billion-plus B+ round as a record or highest single financing in China's large-model sector over the prior 12 months. SO013, SO018
CO023 Yicai and The Standard reported that StepFun was pursuing or completing an almost US$2.5 billion pre-IPO financing tied to Hong Kong listing preparations. SO019, SO022
CO024 Reuters reported through U.S. News and Economic Times that StepFun was unwinding an offshore incorporation structure to pave the way for a planned Hong Kong IPO amid tighter scrutiny of red-chip structures. SO020, SO021
CO025 Tencent News reported that StepFun had not responded by publication time to questions about whether it was dismantling an offshore structure for an IPO. SO015
CO026 Multiple 2026 reports describe StepFun's Hong Kong IPO plan as targeting roughly US$500 million of proceeds and a valuation band around US$10 billion to US$12 billion. SO014, SO019, SO022, SO023, SO031
CO027 Yicai reported that StepFun had dismantled its red-chip structure and was accelerating Hong Kong IPO preparation after the large pre-IPO financing. SO019, SO022
CO028 Lotus Holding disclosed a planned investment in StepFun and warned that the target company was in a state of large losses as of the announcement date. SO029
CO029 StepFun is repeatedly grouped among China's AI Six Tigers or Big Six foundation-model startups in 36Kr, Eastmoney, Yicai, The Standard, and China AI Atlas sources. SO016, SO013, SO019, SO022, SO025
CO030 Hubpy summarizes StepFun as a Chinese AI startup founded in 2023 that raised about US$718 million in January 2026 and had released 11 foundational models. SO028
CO031 Eastmoney reported that by the end of 2025 StepFun's model install base exceeded 42 million devices and served nearly 20 million daily users. SO013
CO032 Eastmoney reported that StepFun had deep cooperation with 60 percent of leading domestic smartphone brands and with Geely/Chongqing Qianli on an AgentOS smart cockpit for the Geely Galaxy M9. SO013
CO033 Eastmoney reported StepFun ecosystem partnerships with Biren Technology, Shanghai Yidian Smart Computing, and applications such as Guotai Junan intelligent customer service. SO013
CO034 36Kr framed StepFun's distinctive Big Six label as solid technology and said the Series B would fund foundation model R&D, multimodal and complex reasoning, and C-end ecosystem coverage. SO016
CO035 StepFun's official documentation presents Step 3.5 Flash as a flagship reasoning model for complex planning, tool use, software engineering, and deep research tasks. SO005
CO036 The official GitHub repository describes Step 3.5 Flash as StepFun's most capable open-source foundation model for frontier reasoning and agentic capabilities. SO007
CO037 Independent model catalogs describe Step 3.5 Flash as released on February 2, 2026 with roughly 196 to 197 billion total parameters, about 11 billion active parameters, and a 256K-token context. SO033, SO034
CO038 The official model overview lists Step 3.7 Flash as a recommended multimodal reasoning flagship with 256K context and support for agent workflows. SO004, SO001
CO039 Tencent News reported media estimates that StepFun's 2025 revenue was nearly RMB500 million and expected 2026 revenue was about RMB1.2 billion, but those figures were not audited in this chapter. SO015
CO040 Public profiles and official sources associate StepFun with Step-2 trillion-parameter MoE work and Step-3 multimodal model capabilities. SO001, SO027, SO028
CO041 China AI Atlas lists StepFun at about US$3.2 billion in cumulative disclosed or announced funding and a reported US$10 billion IPO target valuation. SO025
CO042 No fetched public source disclosed audited revenue, customer-count definitions, detailed board rights, preference terms, debt terms, or secondary-sale terms for StepFun. SO001, SO003, SO015, SO029
CO043 The fetched evidence supports Hangzhou-related capital exposure through a Lotus/Hangzhou investment notice but does not verify an independent Hangzhou office. SO029, SO001, SO009
CO044 CNINFO's Lotus filing warns that an investment in StepFun could face market, policy, and operating-management risks and could lead to investment losses. SO029
CM001 StepFun platform positions StepStar around production Agent models, model/API experimentation, and vertical AI solutions rather than only a consumer chatbot. SM027
CM002 The included market for StepStar spans foundation-model MaaS/API, enterprise private deployment, consumer application embedding, device/automotive AI, and government AI+ workloads. SM006, SM011, SM027
CM003 Excluded spend should include generic non-AI cloud, legacy analytics, non-AI SaaS, and open-source usage that produces no model or application revenue for StepStar. SM007, SM008, SM015, SM023
CM004 Status-quo substitutes include DeepSeek, Qwen, GLM, Doubao, global APIs, hyperscaler routing, and local open-weight deployment. SM004, SM013, SM014, SM015, SM020, SM021, SM024
CM005 StepStar has listed vertical surfaces including consumer electronics, content creation, smart vehicles, local services, finance, manufacturing, gaming, and government. SM027
CM006 Grand View / Horizon says China generative AI is projected to reach $17.609 billion in 2030 and grow at a 39.1% CAGR from 2025 to 2030. SM001
CM007 MarketsandMarkets values China generative AI at $7.0359 billion in 2025 and projects $98.7557 billion by 2030, a 45.8% CAGR. SM002
CM008 Tianxia Gongchang estimates China large-model market revenue at roughly RMB49.5-51.0 billion in 2025 and a broader AI-enabled software definition near RMB100-130 billion. SM006
CM009 Axis Intelligence places conservative China AI market revenue around $28-31 billion in 2025 and projects a path toward $200 billion by 2032 at a 32.5% CAGR. SM003
CM010 Axis Intelligence reports a broader IDC-style China AI tracker proxy around $62 billion for 2025, likely including infrastructure and AI-enabled activity. SM003
CM011 Public China sizing lenses range from roughly $3.4 billion to $62 billion for 2025 depending on whether the boundary is narrow GenAI revenue, large-model revenue, broad AI revenue, or infrastructure-inclusive activity. SM001, SM002, SM003, SM006
CM012 Gartner forecasts worldwide AI spending of $2.595667 trillion in 2026, including $1.431509 trillion of AI infrastructure, $453.209 billion of AI software, and $32.604 billion of AI models. SM007
CM013 IDC identifies AI Infrastructure Provisioning as the largest AI investment area and says AI-enabled customer service and self-service represented $16.7 billion of spending in 2024. SM008
CM014 Precedence Research calculates the global LLM market at $7.77 billion in 2025, $10.57 billion in 2026, and $149.89 billion by 2035 at a 34.44% CAGR. SM018
CM015 Grand View Research estimates the global LLM market at $5.6174 billion in 2024 and $35.4344 billion by 2030 at a 36.9% CAGR. SM017
CM016 MarketsandMarkets projects the global LLM market to reach $36.1 billion by 2030 at a 33.2% CAGR. SM026
CM017 StepStar-relevant enterprise buyers are likely CIOs, CTOs, platform leaders, and business-unit owners who pay for private data, agents, governance, and integration. SM009, SM010, SM020, SM021, SM027
CM018 Developer/API buyers evaluate providers on per-token pricing, context length, SDK compatibility, model-router visibility, and tool-use support. SM013, SM014, SM019, SM020, SM021, SM023
CM019 Consumer applications and device integrations often make the application owner or OEM the payer while the end user experiences the underlying model indirectly. SM027, SM004, SM005
CM020 Government and SOE demand is a distinct China buyer segment because the State Council AI+ opinion explicitly promotes AI deployment across public services and industries. SM011, SM027
CM021 Azure AI Foundry presents model benchmarking, intelligent model routing, and agent orchestration as enterprise platform capabilities, reinforcing buyer expectations for routing and governance. SM021
CM022 AWS Bedrock lists many model providers including DeepSeek, Qwen, OpenAI, Anthropic, and others, and offers batch inference at a 50% lower price than on-demand for selected models. SM020
CM023 OpenAI Business pricing presents team workspaces at $20 per user per month and enterprise custom pricing with security, analytics, and administrative controls. SM022
CM024 Deloitte reports worker access to AI rose by 50% in 2025 and that companies with at least 40% of projects in production are expected to double within six months. SM009
CM025 McKinsey says AI can consume up to a third of companies' change budgets while also adding technology run costs. SM010
CM026 Digital in Asia reports China daily AI token usage surpassed 140 trillion in March 2026, up from 100 billion at the start of 2024. SM005
CM027 DigitalApplied reports Chinese AI providers served more than 45% of OpenRouter traffic in Q2 2026, up from less than 2% a year earlier. SM004
CM028 China policy and state-backed capital are structural market drivers, with Axis citing a national AI industry fund, a broader venture-capital guidance-fund target, and other state-capital channels. SM003, SM011
CM029 WIPO reports China-based inventors filed the highest number of GenAI patents and that the patent landscape contained about 54,000 GenAI inventions in the decade through 2023. SM025
CM030 Digital in Asia and Tianxia Gongchang both identify chip supply, domestic accelerator substitution, and compute sovereignty as central constraints for China foundation-model vendors. SM005, SM006
CM031 Qwen3 was released as an open-weight family including MoE models, with pretraining on about 36 trillion tokens across 119 languages and dialects. SM015
CM032 DeepSeek-R1 is openly available on Hugging Face with distilled models based on Llama and Qwen, showing how reasoning capability can diffuse into smaller and substitutable models. SM024
CM033 DeepSeek official API pricing lists deepseek-v4-flash at $0.14 per million cache-miss input tokens and $0.28 per million output tokens, with much lower cache-hit pricing. SM013
CM034 Google Gemini and AWS Bedrock pricing show that buyers can choose across free tiers, batch discounts, model tiers, and multiple provider routes, raising pricing transparency and substitution pressure. SM019, SM020
CM035 The CAC Interim Measures require generative-AI service providers to meet content, security, data, and service obligations before and during public deployment. SM012
CM036 The State Council AI+ opinion and CAC generative-AI measures together show China incentivizes adoption while also imposing a controlled deployment regime. SM011, SM012
CM037 StepStar's segment fit is strongest where its model is embedded into agent workflows, vertical solutions, device partners, or government/enterprise deployments rather than used as a commodity API endpoint. SM027, SM013, SM015, SM020, SM021
CM038 Open-source model diffusion and low per-token prices can make usage grow faster than revenue or gross profit for undifferentiated model providers. SM013, SM015, SM020, SM024
CM039 Enterprise adoption requires governance and trust because Deloitte highlights preparedness gaps in infrastructure, data, risk, and talent even as AI access expands. SM009
CM040 The commercialization funnel for StepStar should be evaluated from model proof to developer trial, enterprise pilot, governance clearance, unit economics, and scaled recurring revenue. SM009, SM010, SM013, SM020, SM021, SM027
CM041 The main StepStar diligence gap is not whether China AI demand exists, but whether StepStar can show paid usage share, segment-level revenue, gross margin, and defensible distribution. SM002, SM003, SM004, SM013, SM027
CM042 Public sources do not disclose StepStar revenue, API traffic, customer conversion, or segment mix, so SOM cannot be responsibly quantified beyond a qualitative share-of-SAM framing.
CP001 StepFun’s Step3 is a 321B-total-parameter multimodal MoE model with 38B active parameters per token, positioning StepStar around cost-efficient multimodal reasoning rather than a dense-model-only strategy. SP005, SP006
CP002 Step3 is distributed through public checkpoints and an OpenAI/Anthropic-compatible API path, but public evidence does not show StepStar matching Qwen or DeepSeek in ecosystem breadth. SP005, SP006, SP030, SP013
CP003 StepFun’s STEPX Neo launch shows a deliberate AI-plus-terminal strategy that pairs StepStar models with agentic device workflows. SP007, SP044
CP004 AIbase reported in January 2026 that StepStar completed B+ financing of more than 5 billion RMB. SP008
CP005 The AI Chronicle described a prospective StepFun Hong Kong IPO with a valuation target around $12 billion. SP009
CP006 36Kr and Crunchbase coverage support that StepStar had already reached unicorn status by late 2024 before the 2026 B+ financing report. SP010, SP011
CP007 Digital Applied’s Q2 2026 provider report says Chinese providers collectively crossed 45% of OpenRouter traffic, making China’s model market a mainstream API battleground rather than a local-only market. SP001
CP008 The same Q2 2026 provider report ranks StepFun below larger API-volume leaders such as Alibaba/Qwen, MiniMax, Zhipu, and DeepSeek, making smaller API share an adverse competitive signal. SP001
CP009 Independent landscape mapping identifies GLM/Zhipu, Kimi/Moonshot, DeepSeek, MiniMax, Qwen/Alibaba, Doubao/ByteDance, Baidu, Tencent, and Baichuan as relevant Chinese alternatives for buyers. SP001, SP002, SP003
CP010 DeepSeek publishes low per-million-token API pricing, 1M context on current API pages, and OpenAI/Anthropic-compatible endpoints, creating direct cost and integration pressure. SP012, SP013
CP011 DeepSeek-R1’s open-source release and distilled Qwen/Llama checkpoints make internal build and self-hosting more credible alternatives to StepStar API adoption. SP013, SP014
CP012 DeepSeek’s combination of low-cost API access and open reasoning models is the clearest commoditization threat to StepStar’s standalone model economics. SP012, SP013, SP014
CP013 Moonshot positions Kimi around million-token context, multimodal capability, programming, knowledge work, and deep reasoning. SP015, SP016
CP014 Kimi API pricing pages list token-based billing and current Kimi model families, giving Moonshot clearer public developer packaging than StepStar’s platform pages exposed in the reviewed evidence. SP016, SP005
CP015 Moonshot’s Kimi K2 is a 1T-total-parameter MoE model with 32B active parameters and explicit coding and agentic benchmark comparisons. SP017, SP018
CP016 Moonshot’s public profile and Kimi product surface make it a long-context and agentic-workflow competitor rather than only a consumer chatbot peer. SP015, SP017, SP018
CP017 Zhipu/Z.ai’s GLM-4.5 is a 355B-total-parameter, 32B-active-parameter agent foundation model released under an MIT license. SP019, SP020
CP018 Zhipu’s public profile and open platform make it a stronger enterprise and self-hosting competitor than StepStar in cases where permissive licensing is decisive. SP019, SP020, SP021
CP019 MiniMax’s official site emphasizes coding, agentic models, 1M context, and multimodal language, video, voice, and music coverage. SP022
CP020 MiniMax-M1 is described as a 456B-total-parameter hybrid-attention MoE model with 45.9B active parameters and 1M native context, giving MiniMax a direct long-context efficiency story. SP022, SP023, SP024
CP021 MiniMax’s product and financing profile make it a Chinese six-tiger peer with broader consumer-media and agent surfaces than StepStar’s currently cited model and terminal surfaces. SP022, SP023, SP024
CP022 Baichuan’s current public positioning emphasizes Baixiaoyi medical and family-health workflows, making it a narrower verticalized competitor than Qwen, DeepSeek, GLM, or Kimi. SP025, SP027
CP023 Baichuan still exposes developer API documentation, so it remains a model-platform alternative even if its visible wedge is healthcare-specific. SP026, SP025
CP024 ByteDance’s Doubao is relevant because ByteDance distribution can turn consumer and multimodal AI usage into scale that a standalone model startup cannot easily match. SP028, SP003
CP025 Qwen3 publishes dense and MoE open-weight models and lists 235B-total, 22B-active parameters for Qwen3-235B-A22B. SP029, SP030, SP031
CP026 Alibaba Cloud’s Qwen model studio adds cloud distribution, multimodal products, pricing, compliance claims, and enterprise deployment around the open model family. SP032, SP030
CP027 Baidu Qianfan/Wenxin and Tencent Hunyuan compete through enterprise platform bundling and incumbent ecosystem access rather than startup-style model openness alone. SP033, SP034
CP028 Tencent Hunyuan’s public page says the model uses MoE, supports up to 256K context, and connects to search and Tencent content ecosystems. SP034
CP029 OpenAI, Anthropic, Google, and Meta set global benchmark pressure through frontier model releases, business packaging, published API pricing, and open-weight alternatives. SP035, SP036, SP037, SP039, SP041
CP030 OpenAI’s GPT-5 and Business pricing pages show a broad frontier-plus-enterprise package with analytics, budgets, connectors, SSO, and spend controls. SP035, SP036
CP031 Anthropic’s model overview and Claude Opus page make Claude a benchmark for advanced reasoning and coding even when the buyer is evaluating Chinese alternatives. SP037, SP038
CP032 Google’s Gemini model and pricing pages combine model breadth, published pricing, and Google grounding options, raising the procurement baseline for StepStar. SP039, SP040
CP033 Meta Llama 4 reinforces that open-weight model access remains a global substitute for proprietary API dependence. SP041
CP034 Artificial Analysis and LMArena provide external benchmarking surfaces that buyers can use to compare StepStar’s models against global and Chinese alternatives. SP042, SP043
CP035 The public benchmark evidence reviewed here supports Step3 as credible in multimodal efficiency but does not establish StepStar as the broad frontier leader across all model leaderboards. SP005, SP006, SP034, SP042, SP043
CP036 StepStar’s strongest competitive differentiation is the combination of Step3’s cost-efficient multimodal MoE design and an AI terminal strategy through STEPX Neo. SP005, SP006, SP007
CP037 StepStar’s weakest competitive signal is smaller reported API share versus Chinese leaders, especially when Qwen, MiniMax, Zhipu, and DeepSeek have clearer volume or openness stories. SP001, SP030, SP019, SP012, SP023
CP038 StepStar’s openness strategy is real but partial: checkpoints and compatible APIs exist, while DeepSeek, Qwen, GLM, Kimi, MiniMax, and Llama set a higher public bar for open ecosystems. SP005, SP006, SP013, SP017, SP019, SP023, SP030, SP041
CP039 Multi-homing risk is high because many competitors publish OpenAI-compatible APIs, token pricing, or downloadable checkpoints that reduce technical switching costs. SP005, SP012, SP016, SP017, SP019, SP030, SP040
CP040 StepStar appears well-funded, but public funding and valuation disclosures across Chinese private labs remain uneven enough that relative capitalization cannot be underwritten from public sources alone. SP008, SP009, SP010, SP018, SP021, SP024
CP041 The Chinese “AI six little tigers” framing puts StepStar in a peer group where commercialization pressure intensified after DeepSeek’s disruption. SP044, SP014, SP001
CP042 Status quo and internal build are viable substitutes for some buyers because Qwen, DeepSeek, GLM, MiniMax, Step3, and Llama all expose open or self-hostable model paths. SP005, SP006, SP013, SP019, SP023, SP030, SP041
CP043 Capability niches are already crowded: Qwen is broad and enterprise-backed, DeepSeek is cost-efficient, Kimi is long-context, Doubao is distribution-led, GLM is open/enterprise, and MiniMax is multimodal-agentic. SP002, SP003, SP004
CP044 StepStar’s terminal strategy could create switching cost only if STEPX or partner devices become recurring workflow surfaces rather than one-off launch proof. SP007, SP003, SP039
CP045 The adverse investment implication is not that StepStar lacks technology, but that technology alone is not scarce enough when DeepSeek, Qwen, GLM, MiniMax, Kimi, and international labs all publish credible alternatives. SP001, SP012, SP017, SP019, SP023, SP030, SP035, SP037, SP039
CP046 A buyer evaluating StepStar in 2026 should require task-level win-loss proof against DeepSeek, Qwen, Kimi, GLM, MiniMax, Doubao, OpenAI, Anthropic, Gemini, and Llama before underwriting a durable moat. SP001, SP003, SP004, SP034, SP042, SP043
CI001 StepStar completed a January 2026 Series B+ financing round of more than RMB 5 billion. SI004, SI006, SI007, SI008, SI010
CI002 The January 2026 B+ round was widely described as roughly US$700 million-plus, with Yicai reporting about US$719 million. SI008, SI010
CI003 The B+ round was reported as one of the largest Chinese large-model financings in the prior twelve months. SI004, SI006, SI008
CI004 The B+ investor group included state-owned capital, China Life PE, PDVC, Xuhui Capital, Wuxi Liangxi Fund, Xiamen ITG, Huaqin Technology, Tencent, Qiming Venture Partners, and 5Y Capital across public reports. SI004, SI008, SI010, SI013
CI005 StepStar completed a December 2024 Series B round with participation from Shanghai state capital, Qiming Venture Partners, Tencent, and 5Y Capital according to later reporting. SI008, SI011
CI006 December 2024 coverage reported StepStar raising hundreds of millions of dollars in Series B financing and becoming valued around US$1 billion. SI011, SI012, SI021
CI007 Chinese January 2026 coverage estimated StepStar’s post-B+ valuation at roughly RMB 20-30 billion. SI005, SI009
CI008 Later 2026 IPO-oriented articles floated materially higher US dollar valuation targets, but those are reported IPO expectations rather than completed private-round terms. SI016, SI017, SI018, SI019
CI009 Huaqin Technology publicly acknowledged participating as an industrial investor in StepStar’s January 2026 B+ financing. SI013, SI014
CI010 Huaqin Technology said the specific amount of its StepStar investment was non-public material commercial information unless disclosure thresholds are reached. SI013, SI014
CI011 StepStar’s official homepage links users to chat, open-platform, Studio, and Step Plan surfaces rather than publishing a full financial model. SI001, SI002, SI003
CI012 StepStar’s open platform promotes model and agent application development, supporting an API or usage-based monetization path. SI002, SI003
CI013 The Step Plan page and platform material show limited-time free token promotion, indicating customer acquisition subsidies may exist before paid usage is measured. SI002, SI003
CI014 No reviewed official surface discloses realized API revenue, ARR, customer concentration, retention, or gross margin. SI001, SI002, SI003
CI015 No reviewed source provides a complete list-versus-realized pricing schedule for StepStar’s API, Step Plan, OEM, or automotive contracts. SI001, SI002, SI003, SI013, SI014
CI016 Chinese coverage cites StepStar model deployment across more than 42 million devices and daily service volume, but not contract revenue. SI005, SI009, SI023
CI017 Public coverage describes Geely automotive deployment and million-vehicle ambitions for StepStar-powered systems, but not per-vehicle economics. SI005, SI008, SI009
CI018 Device and vehicle distribution signals are usage or channel proof, not direct evidence of recognized revenue or margin. SI005, SI008, SI009, SI023
CI019 The most plausible public revenue streams are API or plan usage, OEM licensing or revenue share, automotive integration, and future enterprise agent packages. SI001, SI002, SI003, SI005, SI008
CI020 Reviewed public sources do not disclose audited revenue, ARR, recognized revenue, or customer concentration for StepStar. SI001, SI002, SI003, SI020, SI021, SI022
CI021 Public reports say B+ proceeds will be used for foundation-model R&D and AI+terminal strategy rollout. SI004, SI006, SI010, SI008
CI022 Long-form Chinese coverage says StepStar will increase compute infrastructure and attract top AI talent alongside model iteration. SI005, SI009
CI023 StepStar’s technical and distribution posture implies material inference and serving costs as API, phone, and vehicle usage scale. SI002, SI003, SI016, SI017, SI027
CI024 The company’s cost base likely includes foundation-model training, inference serving, senior AI talent, platform operations, and partner integration engineering. SI002, SI003, SI005, SI008, SI027, SI028
CI025 StepStar’s Step-3 efficiency claims and terminal deployment strategy may reduce cost-to-serve if they translate into real lower inference cost. SI005, SI009, SI023
CI026 No reviewed public source discloses StepStar gross margin, cost of revenue, inference cost per token, or model-serving P&L. SI001, SI002, SI003, SI020, SI021, SI022
CI027 Without partner contract terms, StepStar’s installed-device scale cannot be converted into revenue per device or gross profit per user. SI005, SI008, SI009, SI013, SI014
CI028 Gartner and McKinsey describe AI infrastructure and run-cost pressure as major 2026 budget themes, increasing the importance of cost discipline for model vendors. SI027, SI028
CI029 CNBC reports that enterprise AI buyers are shifting toward model routing and efficiency, an adverse signal for premium pricing without clear ROI. SI029, SI030
CI030 Sector sources describe Chinese AI model companies moving from cash-burn narratives toward monetization, operational efficiency, and public-market scrutiny. SI024, SI025, SI026
CI031 Comparable Chinese AI IPO candidates have faced scrutiny for high burn and compute costs, making StepStar’s undisclosed burn a material diligence issue. SI024, SI026
CI032 Several 2026 English sources reported that StepStar was considering a Hong Kong IPO. SI015, SI016, SI017, SI018, SI019
CI033 Edgen and related IPO coverage reported a possible roughly US$500 million Hong Kong IPO raise for StepStar. SI015, SI017
CI034 IPO valuation reports are not a filed prospectus and do not provide audited revenue, margin, or burn disclosures. SI015, SI016, SI017, SI018, SI019
CI035 StepStar’s reported cumulative funding exceeds RMB 5 billion when the January 2026 B+ round is included, but precise primary capital retained is not public. SI004, SI008, SI010, SI021, SI022
CI036 The January 2026 B+ round likely extends operating capacity materially but does not reveal how many months of runway StepStar has. SI001, SI004, SI008, SI010, SI028
CI037 Reviewed public sources do not disclose StepStar cash balance, monthly burn, net burn, or runway. SI001, SI002, SI004, SI008, SI020, SI021, SI022
CI038 Reviewed public sources do not disclose debt obligations, compute minimum commitments, liquidation preferences, or secondary-sale details. SI004, SI008, SI013, SI014, SI020, SI021, SI022
CI039 StepStar is well funded by headline round size but not financially underwritten from public data because revenue, margin, burn, and runway remain absent. SI004, SI005, SI008, SI024, SI025, SI026, SI029, SI030
CI040 The minimum diligence package should include revenue by stream, gross margin bridge, cash balance, burn, runway, compute contracts, partner economics, and cap-table terms. SI013, SI014, SI024, SI026, SI027, SI028, SI029, SI030
CE001 StepFun’s public product surface spans a consumer StepFun AI app, an Open Platform/API, AI Studio, Step Plan subscription, open model weights, and the STEPX Neo agent-phone concept. SE001, SE002, SE010, SE013, SE014, SE020, SE030
CE002 Step-3 is documented as a 321B-total-parameter multimodal MoE model with 38B active parameters per token. SE018, SE019, SE020, SE022
CE003 The Step-3 public model card lists 65,536 maximum context length, 48 experts, 3 selected experts per token, MFA attention, and DeepSeek V3 tokenizer. SE020, SE022
CE004 Step-3’s technical report attributes the cost-efficiency thesis to Multi-Matrix Factorization Attention and Attention-FFN Disaggregation. SE018, SE019, SE020
CE005 Step-3’s paper reports up to 4,039 tokens per second per GPU under a 50ms TPOT SLA at 4K context, compared with 2,324 for DeepSeek-V3 in the same setup. SE019
CE006 The Step-3 blog says pretraining processed more than 20T text tokens and incorporated 4T image-text mixed tokens for multimodal training. SE018
CE007 The Step-3 repository says API access is available through StepFun’s platform and that OpenAI/Anthropic-compatible API modes are provided. SE020, SE011
CE008 The Step-3 repository says model checkpoints are stored in BF16 and block-FP8 format and the code and weights are Apache-2.0 licensed. SE020, SE022
CE009 The open-source Step-3 deployment guide says FP8 deployment requires about 326G memory and has an 8xH20 minimum deployment unit, while BF16 requires about 642G and 16xH20. SE021
CE010 The public Step-3 deployment guide says the AFD implementation from the system report is not yet in the open-source guide and remains work in progress with the open-source community. SE021
CE011 Independent review coverage says Step-3 can run on 8x48GB GPUs using int8 quantization for non-attention parameters, but this claim was not independently replicated in the primary StepFun repository text reviewed here. SE025, SE020, SE021
CE012 The Step-3 company blog’s benchmark table reports Step 3 scores including MMMU 74.2, MATH-Vision 64.8, AIME25 82.9, GPQA-Diamond 73.0 and LiveCodeBench 67.1. SE018
CE013 StepFun’s own benchmark note marks some comparator results as reproduced under the same settings, so parts of the benchmark comparison remain company-issued rather than third-party-replicated. SE018
CE014 AI Indigo’s review argues that teams needing maximum raw performance may still prefer o3 or Gemini 2.5 Pro and that teams needing cloud API simplicity may prefer proprietary cloud models. SE025
CE015 AI Indigo identifies a known Step-3 dead-expert phenomenon under investigation by StepFun, creating a technical-risk diligence item despite the model’s cost-efficiency claims. SE025
CE016 StepFun’s official model docs position Step 3.7 Flash as a 198B total / 11B active sparse-MoE multimodal reasoning model with native image and video input and 256K context. SE005, SE007
CE017 StepFun’s official docs position Step 3.5 Flash as a fast flagship language-reasoning model for complex task decomposition, planning, tool use, coding, math and research with 256K context. SE005, SE006
CE018 Step Plan is a subscription service for calling StepFun flagship models from coding tools and agent platforms using a dedicated API key and monthly Credit allowance. SE010
CE019 Step Plan currently lists support for step-3.7-flash, step-3.5-flash, stepaudio-2.5 models, step-router-v1 and step-image-edit-2. SE010, SE004
CE020 StepFun’s pricing page lists step-3.7-flash at 1.35 yuan per 1M uncached input tokens, 0.27 yuan cached input and 8.1 yuan output, and step-3.5-flash at 0.7, 0.14 and 2.1 yuan respectively. SE009, SE008
CE021 StepFun’s billing introduction says image input for multimodal models is converted into token consumption, making multimodal usage a metered API cost rather than a flat feature. SE008, SE009
CE022 StepFun’s official site says Step 2 is a trillion-parameter self-developed foundation model with deep reasoning and multi-layer instruction-following positioning. SE001
CE023 Wikipedia summarizes that StepFun launched Step-2, a trillion-parameter LLM, at WAIC 2024 alongside Step-1.5V and Step-1X. SE028
CE024 36Kr reported that StepFun had released 11 self-developed foundation models spanning language, image and video understanding, image and video generation, and speech capabilities. SE027
CE025 36Kr reported Step-2 ranked first among domestic base models in a LiveBench list released in November 2024, second only to OpenAI o1 and Claude. SE027
CE026 36Kr reported Step-1V ranked first among Chinese visual large models on an LMSYS Chatbot Arena list released in November 2024. SE027
CE027 StepFun’s platform surfaces model categories across reasoning, real-time speech interaction, vision understanding, speech, image generation/editing and model routing. SE004, SE005, SE016, SE017
CE028 StepFun’s consumer app and download pages frame StepFun AI as a configurable work partner and desktop agent that can discover and proactively complete tasks. SE013, SE015
CE029 AI Studio exposes chat, search, showcase, asset library and Playground surfaces, making it a creation and experimentation surface adjacent to the API platform. SE014, SE002
CE030 Tencent News reported StepFun released STEPX Neo as its first terminal product, with Step AOS and built-in Amoo agent integrating model, software system and terminal hardware. SE030
CE031 Tencent News reported Step AOS supports voice, image and text multimodal inputs, context memory and environment-aware service matching. SE030
CE032 Tencent News listed first ecosystem partners for STEPX Neo including Meituan, WPS, Jianying, Ctrip, Amap, Alipay, Baidu, Didi, JD.com and Weibo. SE030
CE033 StepFun’s platform pitches industry solutions for consumer electronics, content creation, smart vehicles, local services, finance, manufacturing, gaming and government. SE002
CE034 36Kr reported StepFun had become a large-model technology partner of leading mobile-phone manufacturers such as Honor and OPPO and that multimodal API invocation volume rose more than 45 times in the second half of 2024. SE027
CE035 Hubpy describes Step-2 as having 1T+ parameters with text, image, video and audio capabilities and lists Honor, Oppo and ZTE partnerships, but it is a secondary summary rather than primary technical documentation. SE029
CE036 SiliconFlow presents Step3 use cases across multimodal scientific discovery, code debugging, financial analysis and compliance/system audits. SE026
CE037 StepFun’s privacy policy, user agreement and management rules provide legal and platform-behavior controls, but they are not equivalent to a public SOC 2 report, model card safety audit, uptime SLA or enterprise security certification. SE034, SE035, SE036
CE038 The reviewed public record did not identify an official status page, uptime SLA, third-party security certification, model-risk audit or independently replicated Step-3 benchmark report. SE001, SE002, SE018, SE034, SE035, SE036
CE039 Open model distribution through GitHub, Hugging Face and ModelScope improves developer access but also exposes StepFun to open-weight commoditization pressure as rivals can inspect, fine-tune and benchmark against the released stack. SE020, SE022, SE023, SE024, SE025
CE040 StepFun’s moat therefore depends less on a single model release and more on maintaining an integrated loop across flagship model design, lower serving cost, API distribution, agent surfaces, terminal partnerships and rapid model cadence. SE001, SE002, SE018, SE019, SE020, SE027, SE030
CE041 The public product maturity pattern is unusually broad for a young foundation-model company: app, Studio, API, subscription plan, open weights, and device/agent-phone experiments are all visible, but enterprise deployment controls remain comparatively thin. SE001, SE002, SE010, SE013, SE014, SE020, SE030, SE034
CE042 StepFun’s product portfolio is best read as a cost-efficient multimodal model platform plus agentic distribution strategy rather than a single chatbot or a pure model-lab story. SE001, SE002, SE005, SE010, SE018, SE020, SE030
CU001 StepFun's public customer base is best segmented into OEM/device partners, automotive partners, consumer app users, STEPX Neo ecosystem partners, enterprise/API developers, and open-source developers. SU001, SU006, SU008, SU011, SU019, SU020, SU021, SU024
CU002 StepFun reportedly had models integrated into more than 42 million devices by the end of 2025 through phone-brand partnerships. SU001, SU006, SU027
CU003 The named phone brands tied to that device reach include Oppo, Honor, and ZTE. SU001, SU006, SU027, SU029
CU004 City News Service reports that those phone partnerships covered about 60 percent of China's major phone brands. SU001, SU006
CU005 StepFun's automotive partner proof is strongest around Geely, where StepFun voice models and AgentOS are described as featured in Geely cars. SU006, SU011, SU013, SU016
CU006 City News Service reports StepFun voice models and AgentOS in Geely cars are expected to surpass one million vehicles by the end of 2026. SU006
CU007 Geely and StepFun jointly showcased Agent OS and a Galaxy M9 human-like AI agent at WAIC 2025. SU011, SU012, SU013
CU008 StepFun's CES 2026 smart-cockpit coverage ties the company's end-to-end voice model to Geely Galaxy M9 cockpit interaction upgrades. SU017, SU018
CU009 STEPX Neo is positioned as a consumer hardware product built around Step AOS and the personal agent Amoo rather than a conventional app-centric smartphone. SU001, SU002, SU003, SU008
CU010 The first-wave STEPX Neo ecosystem partners named publicly include Alipay, Meituan, Amap, Didi, JD.com, Baidu, Weibo, WPS, Trip.com/Ctrip, and CapCut/Jianying. SU001, SU003, SU008, SU004
CU011 The partner list is China-centric, which limits near-term usefulness in markets without equivalent local integrations. SU002, SU005, SU008
CU012 StepFun had not disclosed STEPX Neo retail price, full specifications, sale date, or shipment figures in the reviewed launch coverage. SU001, SU002, SU003, SU004, SU005
CU013 BigGo Finance says StepFun framed the July event as a first unveiling rather than a formal product launch, with more details to follow after 100 days. SU003, SU008
CU014 The StepFun AI Assistant app has visible App Store consumer proof with a 4.7 rating and 912 ratings on the fetched China App Store page. SU019
CU015 The StepFun Google Play listing showed a 3.7 rating, 67 reviews, and a July 9, 2026 update. SU020
CU016 Neither app-store listing reviewed disclosed monthly active users, downloads, paid subscriber count, ARPU, or retention cohorts. SU019, SU020
CU017 Public consumer evidence for StepFun AI is therefore review-and-rating evidence, not a hard user-base or monetization metric. SU019, SU020
CU018 StepFun's developer surface includes a GitHub organization and public Step-Audio repositories. SU021, SU022, SU023
CU019 Step-Audio2 is presented on GitHub as an end-to-end multimodal model for industry-standard speech-to-speech conversation. SU022
CU020 LLM Stats reported StepFun API input pricing from $0.10 per one million tokens and a 98.3 percent seven-day success rate for the tracked model at fetch time. SU024
CU021 LLM Reference reported seven tracked StepFun models across coding, RAG, agents, long context, vision, and JSON/tool-use workloads, last verified on 2026-06-29. SU025
CU022 AI API Prices listed Step 3.5 Flash at $0.090 input and $0.300 output per one million tokens, verified on 2026-06-27. SU026
CU023 The API and open-source signals support a developer/customer segment, but they do not disclose paying developer accounts, retention, or usage volume. SU021, SU024, SU025, SU026
CU024 The Paper reported a strategic cooperation between StepFun and Kingdee around industry agentic services. SU015
CU025 AIbase reported StepFun and MiniMax working with Alipay around AI-native payment infrastructure. SU014
CU026 Public sources reviewed in this chapter name partners and deployment channels more often than direct paying enterprise customers. SU001, SU006, SU008, SU011, SU014, SU015, SU019, SU020
CU027 No reviewed public source disclosed NRR, GRR, churn, renewal rate, contract length, cohort retention, or top-customer concentration percentage for StepFun. SU001, SU006, SU019, SU020, SU024, SU027
CU028 Toutiao argued StepFun's OEM and automotive cooperation is not exclusive and that OPPO or Geely could switch or add alternative model suppliers. SU027
CU029 Toutiao characterized StepFun revenue concentration around a small number of major terminal partners such as OPPO, Honor, and Geely as a cliff-risk if partners self-build or switch. SU027
CU030 人人都是产品经理 warned that terminal and industry rollouts face long cycles and that control of the user entrance remains in partners' hands. SU030
CU031 GSMArena was skeptical that Step AOS was more than an Android skin and noted zero phone specifications had been revealed at the time of its article. SU004
CU032 Gogi advised readers not to wait for STEPX Neo because pricing, timing, and India availability were not confirmed. SU002
CU033 BigGo Finance said the breadth of Amoo tasks depends on whether super-apps open deep enough interfaces and whether users accept permission delegation. SU003
CU034 The likely adoption path runs from partner-embedded device reach to app-store usage and developer API usage, but only the device-reach metric has a public multi-million scale number. SU001, SU006, SU019, SU020, SU024
CU035 StepFun's customer proof matrix is strongest for named OEM and automotive partners, moderate for consumer app ratings, and weakest for retention and direct paying-customer economics. SU006, SU011, SU019, SU020, SU027, SU030
CU036 The named STEPX Neo service partners are best treated as ecosystem integrations rather than evidence of StepFun customers paying directly for its models. SU001, SU003, SU008, SU004
CU037 Huaqin is reported as the manufacturing partner for StepFun's first AI agent smartphone, adding an OEM manufacturing dependency to the customer story. SU028, SU003
CU038 The reviewed 2026 sources support a thesis of broad distribution through partners but leave paying-customer count, retention, and ARPU largely undisclosed. SU006, SU019, SU020, SU024, SU027, SU030
CR001 China regulates AI through a sectoral stack rather than a single comprehensive AI law. SR001, SR003
CR002 Public-facing generative AI services in China are subject to the Interim Measures for Generative AI Services and CAC filing obligations. SR004, SR005
CR003 The July 2026 CAC announcement reported 988 generative-AI services filed and 598 applications or functions registered as of June 30, 2026. SR005, SR034
CR004 Online generative-AI applications or functions should display the registered model name, filing number, or launch number in a prominent place or product-detail page. SR005, SR034
CR005 China’s deep-synthesis rules require providers to implement security responsibilities, user registration, algorithm review, ethics review, content review, data security, and emergency response systems. SR006, SR001
CR006 GB 45438-2025 is the national standard for AI-generated synthetic content labeling and is effective from September 1, 2025. SR007, SR008, SR033
CR007 The AI-content labeling measures impose both explicit visible labels and implicit machine-readable metadata or watermarking obligations. SR008, SR033
CR008 China’s 2026 AI-agent framework treats autonomous agents as systems capable of perception, memory, decision-making, interaction, and execution. SR002, SR024
CR009 The AI-agent framework creates a regulatory watch item for StepFun because StepFun is positioning models and applications around agents, devices, and tool-use workflows. SR002, SR009, SR032
CR010 StepFun’s Open Platform terms describe large-model API technology for enterprise clients and individual developers. SR009, SR011
CR011 StepFun’s terms include limitation-of-liability and individual arbitration language. SR009, SR011
CR012 StepFun’s April 2026 privacy policy says it collects account, enterprise-authentication, user-input, output, payment, API-key, device, and log information for platform services. SR010
CR013 That privacy-policy footprint creates data-governance risk because API users can submit text, voice, images, video, and other content to the platform. SR010, SR004
CR014 Reuters-republished sources reported that StepFun was unwinding an offshore incorporation structure to pave the way for a Hong Kong IPO. SR012, SR013, SR014
CR015 The same reporting said Beijing’s red-chip scrutiny could delay some listings and make legal restructuring costly enough that some companies might abandon IPO plans. SR012, SR013, SR014
CR016 The Standard reported that StepFun completed a roughly US$2.5 billion round, dismantled its red-chip structure, and was pursuing a Hong Kong IPO that earlier market rumors sized around US$500 million. SR015, SR030
CR017 Public IPO valuation narratives vary materially, with AsiaICT discussing a rumored US$10 billion target and StartupWired discussing a possible nearly US$12 billion value. SR016, SR017
CR018 AsiaICT explicitly framed StepFun’s IPO case as carrying an unproven profit model, dependence on a few hardware manufacturers, and cost re-evaluation pressure. SR016
CR019 City News Service reported that StepFun’s Series B+ exceeded RMB 5 billion and that existing backers included Tencent and Qiming alongside state-owned and industrial investors. SR032, SR030
CR020 StepFun’s device-distribution proof is heavily tied to phone and automotive ecosystems, including over 42 million devices and major phone brands such as Oppo, Honor, and ZTE. SR032, SR015
CR021 Yicai/Shanghai Information Office reported that StepFun’s 2026 funding round attracted supply-chain investors including Huaqin, Longcheer, OmniVision, and ZTE. SR030, SR015
CR022 Jiang Daxin is StepFun’s founder and CEO and was previously a Microsoft Global Vice President and STCA chief scientist. SR027, SR028, SR029
CR023 Jiang Daxin’s public technical reputation is unusually central to the StepFun narrative, including his IEEE Fellow selection for context-aware search and language scaling contributions. SR027, SR028, SR029
CR024 Yin Qi’s January 2026 appointment as chairman broadened StepFun’s management bench and placed him in charge of strategy and technical direction. SR030, SR031, SR032
CR025 Yin Qi also serves as chairman of Qianli Technology and has an AI-plus-hardware background, which supports StepFun’s device strategy but adds coordination and dual-role complexity. SR031
CR026 Public sources name Jiang Daxin, Yin Qi, Zhang Xiangyu, and Zhu Yibo as core management figures, but they do not disclose board committees, succession plans, or incentive retention packages. SR030, SR031, SR032
CR027 The United States continues to shape China’s AI compute access through controls and conditional licensing of advanced AI chips. SR018, SR019, SR020
CR028 IAPS described the January 2026 H200 policy as allowing exports under conditions while limiting H200 exports to China to less than 50 percent of total U.S. sales. SR020, SR019
CR029 CNBC reported in May 2026 that the United States moved to halt Nvidia AI-chip shipments to Chinese firms outside China. SR021
CR030 TechXplore/AP reported that Nvidia’s advanced-chip sales in China stalled while local chipmakers led by Huawei gained share in the domestic market. SR022, SR019
CR031 CFR argued that Huawei remains materially behind Nvidia on frontier AI-chip performance, so domestic substitution does not fully eliminate performance and scaling risk. SR018, SR022
CR032 StepFun’s large-model and device strategy is exposed to compute-supply risk because controlled Nvidia access, domestic-chip transition, and edge-device optimization must all work at once. SR018, SR019, SR020, SR032
CR033 Forbes reported that DeepSeek announced a 75 percent promotional discount on V4-Pro and cut cache-hit prices to one-tenth of prior levels. SR023, SR024
CR034 VaaSBlock framed DeepSeek and Qwen as competing on efficiency, price-performance, and open-weight distribution rather than only closed-frontier capability. SR025, SR023
CR035 Sohu/TMTPost carried an adverse view that many Chinese AI unicorns raise substantial funding while struggling to generate sustainable revenue. SR026, SR016
CR036 The “AI Six Tigers” label increases StepFun’s competitive risk because it places the company in a crowded peer set that also includes firms facing DeepSeek, Alibaba, and ByteDance pressure. SR015, SR026, SR025
CR037 DeepSeek-led price compression and open-weight alternatives make standalone API monetization harder for StepFun unless device distribution, enterprise workflow depth, or agentic integration carries differentiated value. SR016, SR023, SR024, SR025
CR038 StepFun’s high fundraising cadence and IPO preparation reduce near-term capital risk but increase public-market execution pressure to show revenue quality, margin path, and governance maturity. SR015, SR016, SR017, SR030, SR032
CR039 The entity-list risk is indirect rather than named: public sources reviewed here do not identify StepFun on a U.S. entity list, but 2026 chip and geopolitical reporting shows policy can change quickly for Chinese AI firms. SR020, SR021, SR024
CR040 No active StepFun enforcement action or litigation event was identified in the reviewed public sources for this chapter. SR009, SR010, SR012, SR013, SR034
CR041 StepFun’s visible legal pages are baseline hygiene rather than proof of full enterprise compliance, model-risk governance, or regulator-facing audit readiness. SR009, SR010, SR011, SR004, SR005
CR042 The regulatory risk is high-residual because StepFun must manage generative-AI filing, model display, deep-synthesis controls, AI-content labeling, privacy obligations, and emerging agent governance simultaneously. SR004, SR005, SR006, SR007, SR008, SR010, SR033
CR043 The IPO-execution risk is high-impact because red-chip restructuring, share reform, valuation expectations, and Hong Kong listing timing all have to converge before public-market access is secured. SR012, SR013, SR014, SR015, SR017
CR044 The most material partner-dependency risk is not a single supplier; it is the stacked dependence on regulators, chip suppliers, domestic hardware ecosystems, phone OEMs, industrial investors, and Hong Kong capital markets. SR015, SR018, SR019, SR021, SR030, SR032
CR045 The strongest visible mitigation is that StepFun has state-linked and industrial backers, a broadened leadership bench, published legal terms, and distribution through major device partners. SR009, SR010, SR019, SR030, SR031, SR032
CR046 The strongest adverse reading is that those same mitigations may become dependencies if regulators, hardware partners, or capital markets demand slower growth and clearer compliance. SR012, SR016, SR018, SR023, SR026
CR047 A thesis break would occur if StepFun cannot show compliant filings and labeling, stable compute access, differentiated monetization, or IPO-ready governance before the next financing or filing window. SR005, SR007, SR012, SR016, SR020, SR023
CR048 Risk monitoring should focus on CAC filing/display updates, labeling enforcement, H200 or overseas-chip license changes, domestic-chip migration proof, DeepSeek/Qwen price moves, and Hong Kong IPO filings. SR005, SR007, SR019, SR021, SR023, SR015
CR049 StepFun’s residual risk profile is high because regulatory, compute, monetization, partner, people, and IPO risks reinforce one another rather than remaining isolated. SR012, SR016, SR018, SR023, SR026, SR030
CR050 The main private diligence needs are compliance filings, security and labeling implementation evidence, compute contracts, partner economics, burn/revenue cohort data, board materials, and IPO restructuring documents. SR009, SR010, SR012, SR016, SR018, SR030
CV001 Public sources verify that StepFun raised several hundred million dollars in a December 2024 Series B backed by Shanghai state capital, Tencent, Qiming, FiveYuan, and related investors. SV001, SV002, SV003, SV034
CV002 The retained public Series B sources do not consistently disclose a precise December 2024 valuation, so the prompt-level approximately $1 billion anchor should be treated as uncorroborated in this chapter. SV001, SV002, SV003, SV034
CV003 StepFun completed a January 2026 Series B+ financing of more than RMB5 billion, roughly $700 million to $717 million depending on the source. SV004, SV005, SV006, SV014
CV004 The January 2026 B+ round was described as one of the largest recent Chinese foundation-model financings. SV004, SV005, SV006
CV005 The B+ investor set included state-linked capital, insurance capital, local government funds, industrial investors, and existing backers such as Tencent, Qiming, and FiveYuan. SV004, SV005, SV006
CV006 Caijing/Sina reported StepStar Pre-IPO financing tranches at roughly $4 billion pre-money and $5 billion to $6 billion pre-money. SV007, SV006
CV007 Tencent News republished reporting that a later financing could put StepStar at a $5 billion to $6 billion post-money valuation. SV008
CV008 Sina Finance reported that major investors proposed a StepStar IPO valuation as high as $12 billion, while warning that the final valuation may adjust. SV009
CV009 NetEase republished reporting that StepStar had secretly submitted an HKEX IPO application with a proposed valuation up to $12 billion. SV010
CV010 The AI Chronicle framed StepFun’s reported Hong Kong IPO valuation as nearing $12 billion and explicitly noted criticism that such figures may be inflated by national-champion sentiment. SV011
CV011 Newsglobenow reported unusually aggressive StepFun secondary-market and pre-IPO valuation chatter, including near-RMB90 billion-style or unit-ambiguous headline figures. SV013
CV012 BestStartup.Asia reported StepFun had raised $2.5 billion at a $10 billion valuation while preparing for a Hong Kong IPO. SV012
CV013 Oryndex describes StepFun as having a $10 billion valuation and a rapid funding trajectory ahead of a planned 2026 IPO. SV015
CV014 The public StepStar valuation record is internally inconsistent across $4 billion to $6 billion, $10 billion, $12 billion, and near-RMB90 billion-style marks. SV006, SV007, SV008, SV009, SV010, SV012, SV013, SV015
CV015 The valuation stance should be stretched because the highest marks are IPO targets or market chatter rather than completed, prospectus-backed public valuations. SV009, SV010, SV011, SV013, SV021, SV025
CV016 Strategic investors in reported StepStar rounds include industrial hardware and device-chain players, supporting a distribution and ecosystem premium. SV008, SV012, SV015
CV017 State-linked capital participation supports the view that StepStar is treated as a strategic Chinese AI infrastructure asset. SV001, SV004, SV005, SV006
CV018 Public sources describe StepStar as pursuing terminal-device, smartphone, automotive, and enterprise/industry scenarios rather than only a consumer chatbot. SV006, SV008, SV015, SV016
CV019 StepFun maintains public developer surfaces through its official platform, GitHub organization, and Hugging Face profile. SV016, SV017, SV018
CV020 The reviewed official StepFun surfaces do not provide audited revenue, gross margin, burn, compute commitments, or cap-table preference terms. SV016, SV017, SV018
CV021 Caijing/Sina and Newsglobenow report media-sourced revenue estimates of roughly RMB500 million for 2025 and RMB1.2 billion expected for 2026. SV007, SV013
CV022 Because the revenue figures are media estimates rather than audited company disclosure, conventional revenue multiples at the reported valuations are not computable with diligence-grade confidence. SV007, SV013, SV016
CV023 The appropriate public method for StepStar is milestone and scenario valuation rather than a DCF or precise revenue multiple. SV007, SV013, SV016, SV033
CV024 A bull case requires a clean HKEX filing, validated revenue estimates, credible terminal or API monetization, and a durable Z.AI/MiniMax-style public market window. SV007, SV008, SV020, SV024, SV028
CV025 A base case reconciles StepStar’s strategic demand with conflicting valuation marks by centering the range around roughly $8 billion to $10 billion. SV006, SV007, SV012, SV015
CV026 A bear case resets toward roughly $4 billion to $6 billion if IPO proof, financial disclosure, or comp support disappoints. SV006, SV007, SV028, SV033, SV035
CV027 StockAnalysis shows Z.AI with a market cap of about HK$397.02 billion on July 20, 2026 and about HK$386.99 billion on April 30, 2026. SV024
CV028 Zhipu/Z.AI’s IPO and subsequent market capitalization create a high public-market ceiling for Chinese foundation-model comparables. SV020, SV021, SV022, SV023, SV024
CV029 StockAnalysis shows MiniMax at about HK$266.59 billion on May 14, 2026 but about HK$60.56 billion by July 20, 2026. SV028
CV030 MiniMax’s Hong Kong IPO evidence supports both the possibility of strong debut demand and the risk of post-listing compression. SV025, SV026, SV027, SV028, SV029
CV031 CNBC reported MiniMax revenue of $53.4 million in the nine months ended September 30, 2025 and an ongoing loss, illustrating that large AI market caps can coexist with early financial profiles. SV029
CV032 TechCrunch reported Moonshot AI raised about $2 billion at a $20 billion valuation in May 2026. SV030
CV033 Yahoo Finance reported Moonshot neared a $30 billion valuation after Kimi K3, extending the Chinese frontier-lab private valuation boundary. SV032, SV031
CV034 Moonshot’s $20 billion to $30 billion reported range makes a $10 billion to $12 billion StepStar IPO target plausible in category context but not automatically attractive. SV030, SV031, SV032, SV009
CV035 Newsglobenow’s comp summary reports Zhipu and MiniMax market values above HK$400 billion and HK$200 billion respectively, supporting the idea that AI listing scarcity influenced private-market StepStar demand. SV019
CV036 StepStar’s $10 billion to $12 billion target sits below some Z.AI observed market-cap marks and below Moonshot’s highest private reports, but above the lower StepStar Pre-IPO marks. SV007, SV009, SV012, SV024, SV030, SV032
CV037 The comparable set is useful for boundary-setting but not for direct multiple comping because StepStar lacks official revenue and margin disclosure. SV016, SV024, SV028, SV030
CV038 Hong Kong AI listing momentum directly affects StepStar because several sources frame it as a likely next large-model company to pursue HKEX. SV006, SV007, SV008, SV009, SV010, SV019
CV039 ChinaBizInsider warns that compute cost and token-rationing pressures can make current AI valuation projections fragile. SV033
CV040 NationPress reports investor concern that Chinese AI firms appear overvalued relative to current revenue and profitability fundamentals. SV035
CV041 The main thesis-break triggers are filing delay, revenue proof gap, compute-cost pressure, comp compression, preference overhang, and unsupported high IPO pricing. SV006, SV007, SV009, SV028, SV033, SV035
CV042 Cap-table and preference terms remain a material valuation gap because headline private marks do not disclose common-equity economics. SV007, SV008, SV009, SV010
CV043 Final diligence should require a prospectus, audited revenue and margin bridge, customer/deployment economics, cap-table terms, compute contracts, and IPO demand quality. SV016, SV021, SV025, SV029, SV033
CV044 The Dec 2024 to 2026 financing arc is directionally steep, but the exact starting valuation is a diligence gap rather than a hard public fact. SV001, SV002, SV003, SV034, SV004, SV005
CV045 The appropriate recommendation is research-more / track with medium confidence, high risk, and a stretched valuation stance. SV006, SV007, SV009, SV024, SV028, SV033, SV035
来源
编号出版方标题引文
SO001 StepFun 阶跃星辰 Scale-up possibilities for everyone / 智能阶跃,十倍每个人的可能.
SO002 StepFun 阶跃星辰 company page Scale-up possibilities for everyone.
SO003 StepFun 阶跃星辰开放平台 Step API 稳定 · 高性能 · 易集成.
SO004 StepFun 模型能力总览 - StepFun 开放平台文档中心 Step 3.7 Flash is listed as a recommended multimodal reasoning flagship with 256K context.
SO005 StepFun Step 3.5 Flash - StepFun 开放平台文档中心 step-3.5-flash 是阶跃星辰的旗舰语言推理模型.
SO006 StepFun 定价与限速 - StepFun 开放平台文档中心 定价明细.
SO007 GitHub Step-3.5-Flash/README.md at main · stepfun-ai/Step-3.5-Flash Step 3.5 Flash is our most capable open-source foundation model.
SO008 Wikipedia StepFun Founded April 6, 2023; founders Jiang Daxin, Zhu Yibo, Jiao Binxing; headquarters Shanghai.
SO009 Baidu Baike Shanghai Jieyue Xingchen Intelligent Technology Co., Ltd. StepFun was established on April 6, 2023, with registered address on the 30th floor, No. 701 Yunjin Road, Xuhui District, Shanghai.
SO010 Baidu Baike Jiang Daxin He joined Microsoft Research Asia in 2007 and later was promoted to Microsoft Global Vice President; he was elected an IEEE Fellow in 2024.
SO011 Qichacha 上海阶跃星辰智能科技股份有限公司 上海阶跃星辰智能科技股份有限公司 存续.
SO012 Aiqicha 上海阶跃星辰智能科技股份有限公司 - 阶跃星辰 - 爱企查 法定代表人为印奇,参保人数为214人.
SO013 Eastmoney / International Finance News 上海大模型企业阶跃星辰完成超50亿元B+轮融资_天天基金网 阶跃星辰(StepFun)完成超50亿元B+轮融资,刷新过去12个月中国大模型赛道单笔最高融资纪录.
SO014 AIBase 估值或超 200 亿!中国 AI 独角兽阶跃星辰传赴港 IPO:前微软大拿坐镇,腾讯领投 B+ 轮 预计集资约5亿美元.
SO015 Tencent News / 21st Century Business Herald AI独角兽阶跃星辰,加速赴港IPO?_腾讯新闻 公司未回应; 拆除架构可能导致部分上市计划推迟.
SO016 36Kr / Intelligent Emergence 36氪_让一部分人先看到未来 Chinese large model unicorn StepStar has recently completed its Series B financing, with a total financing amount of several hundred million US dollars.
SO017 1ai.net Big model unicorn Step Star has closed Series B round totaling "hundreds of millions of dollars," sources say Core investors including Shanghai State-owned Capital Investment Company Limited and strategic and financial investors including Tencent Investment, Wuyuan Capital, Qiming Venture Capital.
SO018 AIBase Breaking Industry Records! Step Star Achieves Over 5 Billion Yuan in Funding, Yindi Officially Appointed as Chairman StepZen officially announced that it has completed a B+ round financing of over 5 billion RMB.
SO019 Yicai Global StepFun to Raise Nearly USD2.5 Billion as Chinese AI Startup Advances Hong Kong IPO StepFun, one of China's six artificial intelligence tigers, is set to complete a new funding round worth almost USD2.5 billion.
SO020 U.S. News / Reuters Chinese AI Startup StepFun to Unwind Offshore Structure to Pave Way for IPO, Sources Say StepFun is unwinding its offshore incorporation structure to pave the way for a planned Hong Kong initial public offering, three sources said.
SO021 Economic Times / Reuters Chinese AI startup StepFun to unwind offshore structure to pave way for IPO - The Economic Times Chinese AI agent StepFun is unwinding its offshore incorporation structure to pave the way for a planned Hong Kong IPO.
SO022 The Standard Stepfun, China's AI Six Tigers, finishes new US$2.5b funding round for HK IPO Stepfun, one of China's AI Six Tigers, has reportedly completed a new US$2.5 billion funding round.
SO023 Startup Wired StepFun Plans Huge Hong Kong IPO Amid AI Boom Market experts believe the company could reach a value of nearly $12 billion.
SO024 BestStartup.Asia StepFun China AI Funding 2026: $2.5 Billion, $10 Billion Valuation and a Hong Kong IPO StepFun China AI funding 2026 has rewritten the record books.
SO025 Tech Buzz China / China AI Atlas StepFun (阶跃星辰) - China AI Atlas Yin Qi became chairman January 2026; valuation $10B reported IPO target valuation.
SO026 Tech Buzz China / China AI Atlas JIANG Daxin (姜大昕) — China AI Atlas Co-founder & CEO, StepFun; IEEE Fellow (2024); Ex-MSRA 16 years, rose to Chief Scientist.
SO027 Jademond StepFun (Step Models): History, IPO & Key Facts Employees (2025, per company statements) approximately 400-500 people.
SO028 Hubpy.io Stepfun (阶跃星辰) Guide 2026: The $718M AI Unicorn With a Trillion Parameters Stepfun raised $718M in January 2026 and offers 1T+ parameter multimodal AI models.
SO029 CNINFO / Lotus Holding 莲花控股股份有限公司 关于对外投资的公告 截至本公告披露日,标的公司处于大额亏损状态.
SO030 Crunchbase News Crunchbase Unicorn Board Tops $1T In Funding Raised Foundation model company StepStar raised a Series B led by Shanghai State-owned Capital Investment; valued at $1 billion.
SO031 The AI Chronicle StepFun IPO: $12B Valuation and China’s AI Sovereignty A startup seeking a public listing with a valuation nearing $12 billion.
SO032 NXplace StepFun: The "Ex-Microsoft" AI Lab That Chose Independence Over Partnership StepFun—Shanghai Jieyue Xingchen Intelligent Technology Co., Ltd—is rewriting the script.
SO033 DataLearnerAI Step 3.5 Flash: Specs, Benchmarks & Model Details Step 3.5 Flash is a chat model from StepFunAI, released on 2026-02-02.
SO034 Airank Step-3.5-Flash by StepFun: Complete Performance Review & Benchmarks (2026) Step-3.5-Flash, released by StepFun on February 2, 2026, is a Mixture-of-Experts large language model.
SM001 Grand View Research China Generative AI Market Size & Outlook, 2030
SM002 MarketsandMarkets China Generative AI Market Size, Share, Trends, Growth Analysis Report, 2030
SM003 Axis Intelligence China AI Statistics 2026: Market Size, Investment & Global Competitive Position
SM004 DigitalApplied Chinese AI Models Q2 2026: 10-Provider Landscape Report
SM005 Digital in Asia What is China's AI Strategy in 2026? A Comprehensive Analysis of Models, Chips, and State Policy
SM006 Tianxia Gongchang Research China AI Large Language Models and Applications: 2026 In-Depth Industry Market Size and Competitive Landscape Research Report
SM007 Gartner Gartner Forecasts Worldwide AI Spending to Grow 47% in 2026
SM008 IDC IDC's Global Outlook on AI and Generative AI Spending - Use Case Insights
SM009 Deloitte The State of AI in the Enterprise - 2026 AI report
SM010 McKinsey Recalibrating technology budgets for the AI era
SM011 State Council of the PRC 国务院关于深入实施“人工智能+”行动的意见
SM012 Cyberspace Administration of China 生成式人工智能服务管理暂行办法
SM013 DeepSeek Models & Pricing | DeepSeek API Docs
SM014 Alibaba Cloud 选择模型
SM015 Qwen Qwen3: Think Deeper, Act Faster
SM016 DeepSeek Introducing DeepSeek-V3
SM017 Grand View Research Large Language Models Market Size | Industry Report, 2030
SM018 Precedence Research Large Language Model Market Size to Surpass USD 149.89 Billion by 2035
SM019 Google AI for Developers Gemini Developer API pricing
SM020 AWS Amazon Bedrock Pricing
SM021 Microsoft Azure Microsoft Foundry - Pricing
SM022 OpenAI Business Pricing
SM023 GitHub QwenLM/Qwen3 repository
SM024 Hugging Face deepseek-ai/DeepSeek-R1
SM025 WIPO China-Based Inventors Filing Most GenAI Patents, WIPO Data Shows
SM026 MarketsandMarkets Large Language Model (LLM) Market - Global Forecast to 2030
SM027 StepFun 阶跃星辰开放平台
SP001 Digital Applied Chinese AI Models Q2 2026: 10-Provider Landscape Report Chinese AI providers now serve over 45% of all OpenRouter traffic, and StepFun appears below the largest Chinese API-volume leaders.
SP002 Wenhao Free Blog Mapping the Chinese AI Landscape: DeepSeek, GLM, Kimi, MiniMax, and Qwen Explained The article maps company, consumer product, model API, developer tools, and multimodal capabilities across Chinese AI players.
SP003 China AI Tour Chinese LLMs 2026 — Qwen, DeepSeek, Doubao, Kimi, Pangu, SenseNova Compared The guide describes Qwen as a general-purpose leader, DeepSeek as cost-efficient reasoning, Doubao as multimodal/creative, and Kimi as long-context.
SP004 NextFuture Chinese LLMs 2026: Qwen vs DeepSeek vs Kimi vs GLM Compared The practitioner comparison frames Qwen, DeepSeek, Kimi, MiniMax, and GLM as production-ready Chinese frontier options.
SP005 StepFun GitHub - stepfun-ai/Step3 Step3 is a multimodal reasoning model built on a MoE architecture with 321B total parameters and 38B active.
SP006 StepFun stepfun-ai/step3 · Hugging Face The model card says Step3 is accessible by API and the checkpoints are available for inference with Hugging Face Transformers.
SP007 City News Service / Shanghai Daily StepFun Launches World’s First Mass-Market Agentic Smartphone StepFun launched the STEPX Neo, described as a mass-market agentic smartphone powered by a built-in AI agent.
SP008 AIbase Breaking Industry Records! Step Star Achieves Over 5 Billion Yuan in Funding, Yindi Officially Appointed as Chairman AIbase reported StepStar completed B+ financing of over 5 billion RMB.
SP009 The AI Chronicle StepFun IPO: $12B Valuation and China’s AI Sovereignty The article describes a prospective Hong Kong IPO and a valuation target around $12 billion.
SP010 36Kr Europe Big model unicorn Step Star has closed Series B round totaling hundreds of millions of dollars, sources say 36Kr reported StepStar’s Series B financing involved state-owned, strategic, and financial investors.
SP011 Crunchbase News Crunchbase Unicorn Board Tops $1T In Funding Raised Crunchbase’s unicorn board coverage listed StepFun among newly minted AI unicorns in December 2024.
SP012 DeepSeek Models & Pricing | DeepSeek API Docs DeepSeek publishes per-million-token prices and OpenAI/Anthropic-compatible base URLs for its API.
SP013 DeepSeek GitHub - deepseek-ai/DeepSeek-R1 DeepSeek states that it open-sourced DeepSeek-R1, DeepSeek-R1-Zero, and distilled models based on Qwen and Llama.
SP014 Wikipedia DeepSeek The page summarizes DeepSeek as a Chinese AI company whose models drew global attention for low-cost performance.
SP015 Moonshot AI Moonshot AI Moonshot’s homepage describes Kimi as built for long-horizon programming, knowledge work, and deep reasoning with million-token context.
SP016 Kimi API 模型推理价格说明 - Kimi API 开放平台 The Kimi API page says Chat Completion input and output are billed by token and lists Kimi K3, Kimi K2.7 Code, and Kimi K2.6.
SP017 Moonshot AI GitHub - MoonshotAI/Kimi-K2 Kimi K2 is described as a 1T-parameter MoE model with 32B activated parameters and agentic optimization.
SP018 Wikipedia Moonshot AI The page summarizes Moonshot AI, Kimi, and public funding history.
SP019 Z.ai GitHub - zai-org/GLM-4.5 GLM-4.5 has 355B total parameters and 32B active parameters, and the series is released under an MIT open-source license.
SP020 Z.ai zai-org/GLM-4.5 · Hugging Face The model card presents GLM-4.5 as an agent foundation model for reasoning, coding, and intelligent agents.
SP021 Wikipedia Z.ai The page summarizes Zhipu AI, also branded Z.ai, and its public financing and commercialization history.
SP022 MiniMax MiniMax MiniMax describes M3 as a coding/agentic frontier model with a 1M context and highlights language, video, voice, and music models.
SP023 MiniMax MiniMaxAI/MiniMax-M1-80k · Hugging Face MiniMax-M1 is an open-weight hybrid-attention MoE reasoning model with 456B total parameters and 45.9B activated per token.
SP024 Wikipedia MiniMax Group The page summarizes MiniMax Group and its financing, products, and company background.
SP025 Baichuan Intelligence 百川大模型-百川智能 Baichuan’s page emphasizes Baixiaoyi as an AI family doctor and clinical-assistance product.
SP026 Baichuan Intelligence 百川大模型-汇聚世界知识 创作妙笔生花-百川智能 Baichuan’s API documentation provides a chat completions endpoint and authorization requirements.
SP027 Wikipedia Baichuan Intelligence The page summarizes Baichuan Intelligence as a Chinese AI company founded by Wang Xiaochuan.
SP028 Wikipedia Doubao The page summarizes Doubao as a ByteDance AI chatbot and product family.
SP029 Qwen Qwen Qwen’s site lists active releases across chat, image, translation, safety, and API surfaces.
SP030 Alibaba Cloud GitHub - QwenLM/Qwen3 Qwen3 makes dense and MoE model weights available, including 235B-A22B, and emphasizes reasoning, agent, and multilingual capabilities.
SP031 Alibaba Cloud Qwen/Qwen3-235B-A22B · Hugging Face The Qwen3-235B-A22B model card lists 235B total parameters, 22B activated parameters, and MoE architecture.
SP032 Alibaba Cloud 千问大模型_AI大模型_一站式大模型推理和部署服务-阿里云 Alibaba Cloud’s Qwen page emphasizes model studio, multimodal models, pricing, compliance, and enterprise deployment.
SP033 Baidu AI Cloud 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 Baidu Cloud presents Qianfan/Wenxin as an enterprise one-stop large-model development and application platform.
SP034 Tencent Cloud 腾讯混元大模型_大语言模型_自然语言大模型- 腾讯云 Tencent Cloud says Hunyuan uses a MoE structure, supports up to 256K context, and is deployed across text, math, code, and search scenarios.
SP035 OpenAI Introducing GPT-5 OpenAI introduces GPT-5 as a frontier model release.
SP036 OpenAI Business Pricing OpenAI Business pricing highlights usage analytics, budgeting, connectors, SSO, and spend controls.
SP037 Anthropic Models overview Anthropic’s model overview lists available Claude models and capability positioning.
SP038 Anthropic Claude Opus Anthropic markets Claude Opus as a frontier model for advanced reasoning and coding work.
SP039 Google AI for Developers Models | Gemini API | Google AI for Developers Google’s Gemini model page lists Gemini 3 preview, Gemini 2.5 Pro, Gemini 2.5 Flash, and creative models.
SP040 Google AI for Developers Gemini Developer API pricing | Gemini API | Google AI for Developers Google’s Gemini pricing page publishes per-million-token paid-tier pricing for Gemini models.
SP041 Meta Unmatched Performance and Efficiency | Llama 4 Meta positions Llama 4 around performance and efficiency for open model use.
SP042 Artificial Analysis Comparison of AI Models across Intelligence, Performance, and Price Artificial Analysis compares AI models across intelligence, performance, output speed, latency, context window, and price.
SP043 LMArena Arena Leaderboard | Compare & Benchmark the Best Frontier AI Models LMArena presents a leaderboard for comparing and benchmarking frontier AI models.
SP044 Wikipedia StepFun The page summarizes StepFun as a Shanghai-based AI company and describes its place among Chinese foundation-model startups.
SI001 StepFun 阶跃星辰 Step Plan 从 Coding 到 Agent,皆可构建。
SI002 StepFun Open Platform 阶跃星辰开放平台 Step Plan 限时免费体验,海量 Token 放送中。
SI003 StepFun Open Platform 阶跃星辰开放平台 - Step Plan Step 3.7 Flash 的升级绝非仅做视觉能力优化。
SI004 Sina Finance “大模型六小虎”之一阶跃星辰B+融资超50亿,多地国资参投 完成超50亿元人民币B+轮融资。
SI005 Sina Finance / TMTPost 阶跃星辰凭什么拿最多的钱 市场正式进入“去泡沫”的结构性调整期。
SI006 Tencent News 刷新纪录!阶跃星辰完成超50亿元人民币B+轮融资_腾讯新闻 刷新过去12个月中国大模型赛道单笔融资纪录。
SI007 Eastmoney Fund 上海大模型企业阶跃星辰完成超50亿元B+轮融资_天天基金网 完成超50亿元B+轮融资。
SI008 KrASIA China’s investors double down on AI frontrunners as StepFun raises RMB 5 billion StepFun has raised more than RMB 5 billion (USD 700 million) in a Series B+ funding round.
SI009 KrASIA As AI consolidates, what makes StepFun worth a RMB 5 billion raise? The raise underscores consolidation taking shape across China’s AI sector.
SI010 Yicai Global Chinese AI Firm Stepfun Raises USD719 Mln for Model Development, AI Agent Rollout Chinese AI Firm Stepfun Raises USD719 Mln for Model Development.
SI011 SiliconANGLE Chinese AI model maker Stepfun raises hundreds of millions in Series B funding Chinese AI model maker Stepfun raises hundreds of millions in Series B funding.
SI012 TMTPost Chinese AI Unicorn Stepfun Secures $100 Million in New Funding Round Chinese AI Unicorn Stepfun Secures $100 Million in New Funding Round.
SI013 Eastmoney Guba / Huaqin Technology 华勤技术:公司作为产业投资人参与了阶跃星辰于2026年1月所完成的B+轮融资 华勤技术作为产业投资人参与了阶跃星辰于2026年1月所完成的B+轮融资。
SI014 10jqka iNews / Huaqin Technology 华勤技术:华勤技术作为产业投资人参与了阶跃星辰于2026年1月所完成的B+轮融资 关于本次投资的具体金额,属于公司非公开的重大商业信息。
SI015 Edgen Tencent-Backed StepFun Eyes $500M Hong Kong IPO Tencent-Backed StepFun Eyes $500M Hong Kong IPO.
SI016 The AI Chronicle StepFun IPO: $12B Valuation and China’s AI Sovereignty StepFun IPO: $12B Valuation and China’s AI Sovereignty.
SI017 StartupWired StepFun Plans Huge Hong Kong IPO Amid AI Boom StepFun Plans Huge Hong Kong IPO Amid AI Boom.
SI018 NewsGlobeNow StepFun Eyes Hong Kong IPO After Reported $2.5B Raise StepFun Eyes Hong Kong IPO After Reported $2.5B Raise.
SI019 XIX AI Chinese AI Unicorn Jieyu Star Eyes Hong Kong IPO with Potential $20B Valuation Chinese AI Unicorn Jieyu Star Eyes Hong Kong IPO with Potential $20B Valuation.
SI020 Tracxn StepFun StepFun
SI021 Parsers.vc StepFun – Funding, Valuation, Investors, News StepFun Funding, Valuation, Investors, News.
SI022 Oryndex StepFun Funding & Company Data StepFun Funding & Company Data.
SI023 Hubpy Stepfun (阶跃星辰) Guide 2026: The $718M AI Unicorn With a Trillion Parameters The $718M AI Unicorn With a Trillion Parameters.
SI024 BigGo Finance China's AI Model Race Abandons Cash-Burn Narrative: Three Ledgers Will Determine Winners China's AI Model Race Abandons Cash-Burn Narrative.
SI025 AsiaICT China’s AI Industry: A Unified Pivot Towards Monetization? China’s AI Industry: A Unified Pivot Towards Monetization?
SI026 Business Standard Two more Chinese AI players prepare for IPOs, but the burn rate is high Two more Chinese AI players prepare for IPOs, but the burn rate is high.
SI027 Gartner Gartner Forecasts Worldwide AI Spending to Grow 47% in 2026 Gartner Forecasts Worldwide AI Spending to Grow 47% in 2026.
SI028 McKinsey Recalibrating technology budgets for the AI era AI is gobbling up to a third of companies’ change budgets while adding to run costs.
SI029 CNBC Model routing is a fix for AI overspending. That’s a problem for OpenAI and Anthropic Companies are shifting from running everything on the most powerful AI model to matching each task to the right one.
SI030 CNBC OpenAI and Anthropic face new AI reality as users shift from tokenmaxxing to efficiency OpenAI and Anthropic face new AI reality as users shift from tokenmaxxing to efficiency.
SE001 StepFun 阶跃星辰 Scale-up possibilities for everyone; Step API stable, high-performance, easy to integrate; Step 3.7 Flash is for real Agent workflows.
SE002 StepFun Open Platform 阶跃星辰开放平台 I build with models, I create my agent; language large model Step 3.7 Flash; industry solutions include consumer electronics, content creation, smart vehicles, finance, manufacturing and government.
SE003 StepFun Open Platform StepFun Open Platform A Powerful Platform for Building AI Apps; Step API · Stable · High-Performance · Easy-to-Integrate.
SE004 StepFun Open Platform Docs 模型能力总览 - StepFun 开放平台文档中心 The model catalog groups public models by capability and lists Step 3.7 Flash, Step 3.5 Flash, speech, vision, image editing and routing entries.
SE005 StepFun Open Platform Docs 推理模型总览 - StepFun 开放平台文档中心 Step 3.7 Flash is a flagship multimodal reasoning model with native image and video input, 198B total / 11B active sparse MoE architecture, and 256K context; Step 3.5 Flash is a flagship language reasoning model.
SE006 StepFun Open Platform Docs Step 3.5 Flash - StepFun 开放平台文档中心 Step 3.5 Flash is optimized for agent and code tasks, preserving flagship reasoning and tool-calling capability while improving token efficiency and speed.
SE007 StepFun Step 3.7 Flash — A high-efficiency Flash model for Real-World Step 3.7 Flash is presented as a high-efficiency Flash model for real-world agents, with sections on agentic coding, enterprise search and agents that can see.
SE008 StepFun Open Platform Docs 计费介绍 - StepFun 开放平台文档中心 The platform charges by total model input and output token usage; image input for multimodal models is also converted into token consumption.
SE009 StepFun Open Platform Docs 定价与限速 - StepFun 开放平台文档中心 Pricing lists step-3.7-flash at 1M tokens with input 1.35 yuan, cached input 0.27 yuan and output 8.1 yuan; step-3.5-flash at 0.7 yuan, 0.14 yuan and 2.1 yuan.
SE010 StepFun Open Platform Docs Step Plan 概述 - StepFun 开放平台文档中心 Step Plan is a subscription service for calling flagship models from coding tools and agent platforms such as OpenClaw, Claude Code, Trae and Cursor using a dedicated API key and monthly Credit allowance.
SE011 StepFun Open Platform Docs 从 OpenAI 迁移至阶跃星辰 - StepFun 开放平台文档中心 StepFun says its models support personal and enterprise calls and can be used with OpenAI-compatible invocation patterns after creating an API key.
SE012 StepFun Open Platform Docs Chat Completions API - StepFun 开放平台文档中心 The chat completions API reference documents request and response fields for StepFun model calls.
SE013 StepFun 阶跃AI The StepFun chat surface shows New conversation, StepClaw, Research, API Platform and login-gated history.
SE014 StepFun AI Studio | StepFun AI Studio exposes new chat, search, showcase library, asset library and Playground surfaces.
SE015 StepFun 阶跃AI download The download page describes StepFun AI as an agent on the user operating system that discovers and proactively completes tasks, with MacOS and Windows clients.
SE016 StepFun Open Platform Docs Step Image Edit 2 - StepFun 开放平台文档中心 Step Image Edit 2 is described as the latest lightweight iterative image-editing model.
SE017 StepFun Open Platform Docs StepAudio 2.5 TTS - StepFun 开放平台文档中心 StepAudio 2.5 TTS is documented as a Contextual TTS model in the StepFun model catalog.
SE018 StepFun Step3: Cost-Effective Multimodal Intelligence Step3 is a 321B-parameter multimodal reasoning model with 38B active parameters; during pretraining it processed over 20T text tokens and 4T image-text mixed tokens.
SE019 arXiv Step-3 is Large yet Affordable: Model-system Co-design for Cost-effective Decoding The paper introduces Step-3, a 321B-parameter VLM with Multi-Matrix Factorization Attention and Attention-FFN Disaggregation, achieving up to 4,039 tokens per second per GPU under a 50ms TPOT SLA.
SE020 GitHub GitHub - stepfun-ai/Step3 Step3 is a Mixture-of-Experts model with 321B total parameters, 38B active, 65,536 max context, OpenAI/Anthropic-compatible API access and Apache-2.0 code and weights.
SE021 GitHub Step3/docs/deploy_guidance.md at main · stepfun-ai/Step3 The deployment guide says FP8 requires about 326G memory and the smallest deployment unit is 8xH20; BF16 requires about 642G and 16xH20; AFD open-source support is still in progress.
SE022 Hugging Face stepfun-ai/step3 · Hugging Face The Hugging Face model card mirrors the Step3 321B / 38B active configuration and provides Transformers inference guidance for the open checkpoints.
SE023 GitHub stepfun-ai The StepFun GitHub organization lists public repositories including Step3 and agent/research infrastructure projects with visible stars and languages.
SE024 ModelScope step3 ModelScope exposes a StepFun Step3 model page, indicating an additional China model-hub distribution surface.
SE025 AI Indigo StepFun Step-3: Cost-Effective Multimodal Intelligence at 321B Parameters The review says teams needing maximum performance may prefer o3 or Gemini 2.5 Pro, cloud API simplicity may favor proprietary cloud models, and Step-3 has a known dead-expert phenomenon under investigation.
SE026 SiliconFlow step3 - Model Info, Parameters, Benchmarks - SiliconFlow SiliconFlow positions Step3 for multimodal scientific discovery, code analysis, financial insights, and multimodal system/compliance audits.
SE027 36Kr 36氪_让一部分人先看到未来 36Kr reported StepFun had released 11 self-developed foundation models across language, image understanding, video understanding, image generation, video generation and speech, and that Step-2 ranked first among domestic base models in a LiveBench list.
SE028 Wikipedia StepFun Wikipedia summarizes that StepFun launched Step-2, a trillion-parameter LLM, at WAIC 2024, later open-sourced Step-Video-T2V and Step-Audio with Geely, and released Step 3 in July 2025.
SE029 Hubpy Stepfun (阶跃星辰) Guide 2026: The $718M AI Unicorn With a Trillion Parameters Hubpy describes Step-2 as a 1T+ parameter model with multimodal capabilities across text, image, video and audio and says StepFun offers API access for developers.
SE030 Tencent News / Pacific Tech 阶跃星辰首款智能体终端手机STEPX Neo发布_腾讯新闻 Tencent News reported StepFun released STEPX Neo, described as a large-model-native agent smartphone with a rear interactive screen, Step AOS and built-in Amoo agent.
SE031 1ai.net Big model unicorn Step Star has closed Series B round totaling hundreds of millions of dollars, sources say 1ai reported StepFun had launched Leap Ask and Photo Ask, a multimodal visual-search function based on its visual understanding model.
SE032 Baidu Baike Jiang Daxin Baidu Baike identifies Jiang Daxin as founder of Shanghai Jieyue Xingchen Intelligent Technology Co., Ltd.
SE033 AIbase 估值或超 200 亿!中国 AI 独角兽阶跃星辰传赴港 IPO AIbase described StepFun as developing foundation models and pushing an AI + terminals strategy to embed large-model capability deeply into hardware ecosystems.
SE034 StepFun Open Platform Docs StepFun 开放平台隐私政策 - StepFun 开放平台文档中心 The privacy policy latest effective date is May 25, 2026 and describes handling of user personal information for the StepFun Open Platform.
SE035 StepFun Open Platform Docs StepFun 开放平台用户协议 - StepFun 开放平台文档中心 The user agreement latest effective date is May 25, 2026 and governs use of the StepFun Open Platform.
SE036 StepFun Open Platform Docs 开放平台管理规则及公约 - StepFun 开放平台文档中心 The management rules and convention cite Chinese internet-information-service and generative-AI regulatory requirements for platform behavior.
SU001 City News Service StepFun Launches World's First Mass-Market Agentic Smartphone StepFun announced ecosystem partnerships with Trip.com, Alipay, Didi, Meituan, WPS, and ByteDance's Jianying but did not reveal the STEPX Neo's retail price.
SU002 Gogi STEPX Neo: China's StepFun Unveils the World's First Agentic AI Phone, Runs a Custom Step AOS Instead of Android Skins If you are looking for a phone now, do not wait on the STEPX Neo. StepFun has not confirmed pricing, launch timing, or India availability.
SU003 BigGo Finance StepFun Unveils World's First AI Agent Phone STEPX Neo, Ecosystem Partnerships Seen as Key to Breakthrough Whether it can persuade more super-apps to open their ecosystems, convince users to hand over critical permissions, and establish a sustainable payment model beyond hardware will be the core tests facing this experiment.
SU004 GSMArena An unlikely Chinese company claims they have the first "AI Agentic Phone" That's all we know so far about this hype building "announcement" that was mostly buzzwords with little substance.
SU005 News24 World’s first agentic phone launched, can use AI without internet, name is…, made by… The StepX Neo AI phone is among the first smartphones launched by the China-based giant for the Chinese market. However, no official updates on a global launch have been confirmed.
SU006 City News Service StepFun Secures Record 5-Billion-Yuan Funding, Appoints New Chairman By the end of 2025, StepFun's models would already be used on over 42 million devices through partnerships with 60 percent of China's major phone brands, including Oppo, Honor, and ZTE.
SU007 The Next Web China is rebuilding the smartphone around AI agents. ZTE’s NaviX sold out in hours. China is rebuilding the smartphone around AI agents. ZTE’s NaviX sold out in hours.
SU008 36Kr Global World's First AI Agent Smartphone Launch: Seamlessly Integrated with Alipay, Meituan, Didi & Baidu Ecosystems The first batch of ecosystem partners includes Alipay, Meituan, Amap, Trip.com, CapCut, JD.com, Didi, Baidu, Weibo, WPS, and more.
SU009 Oton Technology China’s StepFun Debuts First Agentic Phone, Specs Still Unknown China’s StepFun Debuts First Agentic Phone, Specs Still Unknown
SU010 Mega Mobile Content StepX Neo: The First Phone Built on AI Agents, Not Apps StepX Neo: The First Phone Built on AI Agents, Not Apps
SU011 Business Wire Geely Auto Group Teams Up with StepFun for a Joint Showcase at the 2025 World Artificial Intelligence Conference Geely showcased a suite of new products... And the Galaxy M9—highlighted as the world's first vehicle equipped with a Human-like AI agent.
SU012 Bitauto Geely launched its latest achievement codeveloped with StepFun at WAIC 2025 Geely launched its latest achievement codeveloped with StepFun at WAIC 2025.
SU013 CNEVPost Geely unveils industry's first AI agent-powered car cockpit Geely unveils industry's first AI agent-powered car cockpit.
SU014 AIbase MiniMax、阶跃星辰联手支付宝:AI 原生支付基建迎来“大模型国家队” MiniMax、阶跃星辰联手支付宝:AI 原生支付基建迎来“大模型国家队”
SU015 The Paper 阶跃星辰与金蝶达成战略合作,布局行业智能体服务 阶跃星辰与金蝶达成战略合作,布局行业智能体服务
SU016 Sina Finance 吉利加速整合“AI+智驾”:印奇“双线”任职,阶跃星辰超50亿元融资落定 吉利加速整合“AI+智驾”:印奇“双线”任职,阶跃星辰超50亿元融资落定
SU017 Tencent News “活人感”智能座舱原来如此丝滑!阶跃星辰端到端语音模型海外“出圈” 阶跃星辰端到端语音模型海外“出圈”
SU018 Tencent News CES 2026:阶跃星辰端到端语音模型亮相 助力吉利银河M9智能座舱交互升级 CES 2026:阶跃星辰端到端语音模型亮相 助力吉利银河M9智能座舱交互升级
SU019 Apple App Store StepFun - StepFun AI Assistant App - App Store 4.7 out of 5; 912 Ratings.
SU020 Google Play StepFun - Apps on Google Play 3.7; 67 reviews.
SU021 GitHub stepfun-ai stepfun-ai
SU022 GitHub GitHub - stepfun-ai/Step-Audio2 Step-Audio 2 is an end-to-end multi-modal large language model designed for industry-standard speech-to-speech conversation.
SU023 GitHub GitHub - stepfun-ai/Step-Audio-R1 GitHub - stepfun-ai/Step-Audio-R1
SU024 LLM Stats StepFun: API Pricing, Performance & Model Catalog StepFun hosts 1 active AI models, with input pricing from $0.10 per 1M tokens, with median throughput of 177 characters/sec, and P95 time to first token of 1.07s, with 98.3% success rate over 7 days.
SU025 LLM Reference StepFun — AI Model API StepFun offers 7 tracked models... last verified 2026-06-29.
SU026 AI API Prices StepFun API Pricing (2026) — Cost per Token for Every Model Cheapest StepFun model: Step 3.5 Flash at $0.090 in / $0.300 out per 1M tokens.
SU027 Toutiao 装机4200万台、营收仅5亿,阶跃星辰百亿估值是泡沫吗? 合作不具备排他性:吉利可以同时接入豆包,OPPO也能转投其他服务商。
SU028 Sohu 阶跃星辰将推出首款AI智能体手机,代工企业为华勤技术 阶跃星辰将推出首款AI智能体手机,代工企业为华勤技术
SU029 All-Weather TMT WAIC智能体手机潮涌调研:荣耀、阶跃、中兴各有什么筹码? WAIC智能体手机潮涌调研:荣耀、阶跃、中兴各有什么筹码?
SU030 人人都是产品经理 阶跃星辰深度拆解:产品、技术、客户与它真正的护城河 入口在合作伙伴手里,话语权是一场持久战。
SR001 Deep Lex China AI Regulation — Deep Lex China operates the most extensive binding sectoral AI regulatory regime globally, with no single comprehensive AI law to date.
SR002 NYU Shanghai Research Institute for Technology and Society China Issues First National Policy Framework Dedicated to AI Agents China’s CAC, NDRC, and MIIT jointly released the Implementation Opinions on the Standardized Application and Innovative Development of Intelligent Agents.
SR003 China Crunch China’s AI Regulation 2026: Building a Global Framework for Responsible Algorithms Beijing is moving from sector-specific guidelines toward a unified system that regulates algorithmic design, data use, and ethical deployment.
SR004 Cyberspace Administration of China 生成式人工智能服务管理暂行办法 利用生成式人工智能技术向中华人民共和国境内公众提供生成文本、图片、音频、视频等内容的服务,适用本办法。
SR005 Cyberspace Administration of China 关于发布生成式人工智能服务已备案信息的公告(2026年5月至6月) 截至6月30日,累计有988款生成式人工智能服务完成备案,598款生成式人工智能应用或功能完成登记。
SR006 Cyberspace Administration of China 互联网信息服务深度合成管理规定 深度合成服务提供者应当落实信息安全主体责任,建立健全用户注册、算法机制机理审核、科技伦理审查等管理制度。
SR007 State Administration for Market Regulation 国家标准|GB 45438-2025 中文标准名称:网络安全技术 人工智能生成合成内容标识方法。实施日期 2025-09-01。
SR008 Covington Inside Privacy China Releases New Labeling Requirements for AI-Generated Content The Labeling Rules impose explicit and implicit labeling obligations on internet information service providers.
SR009 StepFun Terms of Service - StepFun Documentation StepFun provides artificial intelligence large-model API technology.
SR010 StepFun StepFun开放平台隐私政策 为了向您提供智能对话及内容生成服务,我们会收集您主动输入的信息。
SR011 StepFun Terms of Services YOUR PARTICULAR ATTENTION IS DRAWN TO THE LIMITATION OF LIABILITY CONTAINED IN SECTION 8.
SR012 Reuters via Yahoo Finance Chinese AI startup StepFun to unwind offshore structure to pave way for IPO, sources say StepFun is unwinding its offshore incorporation structure to pave the way for a planned Hong Kong initial public offering.
SR013 Reuters via U.S. News Chinese AI Startup StepFun to Unwind Offshore Structure to Pave Way for IPO, Sources Say Experts have said the move could delay some listings as red-chip companies scramble to change their domicile back to China.
SR014 The Economic Times Chinese AI startup StepFun to unwind offshore structure to pave way for IPO Some might even have to abandon their IPO plans as changing the legal structure of the company could be cost-prohibitive.
SR015 The Standard Stepfun, China's AI Six Tigers, finishes new US$2.5b funding round for HK IPO Stepfun, one of China's AI Six Tigers, has reportedly completed a new US$2.5 billion funding round.
SR016 AsiaICT Is StepFun's $10 Billion Valuation a Cure for AI Monetization or a New Source of Anxiety? Behind the glossy capital narrative lie an unproven profit model, a channel structure highly dependent on a few hardware manufacturers, and pressure from cost re-evaluation.
SR017 StartupWired StepFun Plans Huge Hong Kong IPO Amid AI Boom Market experts believe the company could reach a value of nearly $12 billion.
SR018 Council on Foreign Relations China’s AI Chip Deficit: Why Huawei Can’t Catch Nvidia and U.S. Export Controls Should Remain Huawei is not a rising competitor to Nvidia but has a large and growing performance deficit relative to Nvidia.
SR019 The Diplomat Nvidia’s H200 Chips Re-enter China – But Beijing Isn’t Giving up on Huawei Even with controlled access to H200 chips, China will continue to incentivize the growth of domestic chipmakers.
SR020 Institute for AI Policy and Strategy New BIS Licensing Policy for H200s: Tough Guidelines, Weak Enforcement The policy limits H200 exports to China to less than 50% of total U.S. sales.
SR021 CNBC U.S. takes step to halt Nvidia AI chip shipments to Chinese firms outside China U.S. takes step to halt Nvidia AI chip shipments to Chinese firms outside China.
SR022 TechXplore / Associated Press Nvidia's AI chip sales in China stall, as local chipmakers like Huawei take the lead Chinese companies like Huawei overtake global industry leaders like Nvidia in their home market.
SR023 Forbes China’s DeepSeek V4 And Qwen Reshape The Open-Source AI Race DeepSeek announced a 75% promotional discount on V4-Pro and cut input cache hit prices to one-tenth.
SR024 Big Hat Group China AI Weekly: DeepSeek's $7.4B Raise, World's First Agentic AI Law, and the Permanent Price War June 2026 marks a structural inflection point for China’s AI ecosystem.
SR025 VaaSBlock Chinese AI 2026: DeepSeek, Qwen, ByteDance | VaaSBlock DeepSeek and Qwen had chosen a different board: the efficiency frontier and open-weight distribution.
SR026 Sohu / TMTPost Only DeepSeek, Alibaba, and ByteDance Will Survive AI Competition in China as "Six Tigers" Fall Many unicorns of the AI sector were likened to unipigs—companies that raise substantial funding but struggle to generate sustainable revenue.
SR027 Baidu Baike Jiang Daxin In 2023, he founded Shanghai Step Star Intelligence Technology Co., Ltd., launching the Step Series Multimodal Large Models.
SR028 Alibaba Cloud Startup 阶跃星辰创始人、CEO 姜大昕博士入选 2025 IEEE Fellow IEEE 给姜大昕博士的入选理由是:对上下文感知搜索和语言 Scaling 方法做出的贡献。
SR029 Tencent Cloud Developer Community AI人物传:阶跃星辰创始人、CEO姜大昕 姜大昕是阶跃星辰的创始人兼CEO,曾任微软全球副总裁和微软亚洲互联网工程研究院(STCA)的首席科学家。
SR030 Shanghai Information Office / Yicai Economic News | StepFun to raise nearly USD2.5 billion as Chinese AI startup advances Hong Kong IPO StepFun, one of China's six artificial intelligence tigers, is set to complete a new funding round worth almost USD2.5 billion.
SR031 Gasgoo Personnel Changes | Yin Qi Appointed Chairman of StepFun Yin Qi was appointed chairman, responsible for setting the overall strategy and technical direction.
SR032 City News Service / Shanghai Daily StepFun Secures Record 5-Billion-Yuan Funding, Appoints New Chairman By the end of 2025, StepFun's models would already be used on over 42 million devices through partnerships with 60 percent of China's major phone brands.
SR033 Regulations.ai Measures for the Identification of AI-Generated (Synthetic) Content The Measures set a mandatory national baseline requiring that AI-generated or AI-synthesized content be clearly identified.
SR034 Digital Policy Alert Cyberspace Administration's domestic generative AI services filing list Policy Area: Authorisation, registration and licensing; Policy Instrument: Business registration requirement.
SV001 South China Morning Post Shanghai firm helps AI start-up Stepfun raise 'hundreds of millions of dollars' A Shanghai-backed investment vehicle helped Stepfun raise hundreds of millions of dollars in its latest funding round.
SV002 SiliconANGLE Chinese AI model maker Stepfun raises hundreds of millions in Series B funding Stepfun raised hundreds of millions of dollars in Series B funding.
SV003 TMTPost Chinese AI Unicorn Stepfun Secures $100 Million in New Funding Round The round attracted state-owned capital, strategic backers, and financial investors including Shanghai State-owned Capital Investment and Tencent.
SV004 KR Asia China’s investors double down on AI frontrunners as StepFun raises RMB 5 billion StepFun has raised more than RMB 5 billion (USD 700 million) in a Series B+ funding round.
SV005 36Kr 阶跃星辰拿到50亿新年最大融资,资本看中了什么? 阶跃星辰完成了过去一年中国基础大模型领域金额最大的单轮融资,超50亿元人民币的B+轮。
SV006 Tencent News AI独角兽阶跃星辰,加速赴港IPO? 2026年1月,阶跃星辰完成超50亿元B+轮融资,刷新中国大模型赛道此前近一年单笔融资纪录。
SV007 Sina Finance 独家|阶跃星辰计划年内港股上市,2025年收入约5亿元 阶跃星辰正在进行新一轮Pre-IPO融资,第一拨投前估值约40亿美元,第二拨投前估值50亿-60亿美元。
SV008 Tencent News 大模型独角兽阶跃星辰将完成近25亿美元融资,冲刺港股IPO 本轮融资完成后,阶跃星辰投后估值已达50亿美元~60亿美元。
SV009 Sina Finance 阶跃星辰将 IPO!估值或超 800 亿 主要投资方提出的估值最高可达120亿美元,但最终估值仍可能调整。
SV010 NetEase AI“六小虎”之一阶跃星辰,据传已秘密递表港交所,估值达120亿美元 阶跃星辰传已秘密向港交所递交IPO申请,主要投资方提出的估值最高可达120亿美元。
SV011 The AI Chronicle StepFun IPO: $12B Valuation and China’s AI Sovereignty Critics argue that such figures are inflated by national champion sentiment and state-backed investment vehicles.
SV012 BestStartup.Asia StepFun China AI Funding 2026: $2.5 Billion, $10 Billion Valuation and a Hong Kong IPO StepFun China AI funding 2026 raised $2.5B at $10B valuation.
SV013 Newsglobenow StepFun Valuation Hits at Least $90 Billion Before IPO Push StepFun's revenue rose from 30 million yuan in 2024 to 500 million yuan in 2025, with 2026 revenue expected at 1.2 billion yuan.
SV014 Aibase Valuation May Exceed 20 Billion! Chinese AI Unicorn Jieyu Star Reports to Go Public in Hong Kong The B+ round raised more than RMB 5 billion and provided confidence for the IPO.
SV015 Oryndex StepFun Funding & Company Data The rapid succession of large funding rounds, a $10 billion valuation, and partnerships with major smartphone brands indicate aggressive expansion.
SV016 StepFun 阶跃星辰开放平台 阶跃星辰开放平台
SV017 GitHub stepfun-ai stepfun-ai
SV018 Hugging Face stepfun-ai (StepFun) stepfun-ai (StepFun)
SV019 Newsglobenow StepFun Eyes Hong Kong IPO After Reported $2.5B Raise Zhipu AI and MiniMax have already listed in Hong Kong, with reported market values above HK$400 billion and HK$200 billion respectively.
SV020 South China Morning Post China’s Zhipu AI launches US$560 million share sale amid heated IPO tech race The company’s post-listing market valuation is estimated at HK$51.16 billion.
SV021 HKEXnews Zhipu AI Global Offering Prospectus Prospective investors should carefully consider all of the information set out in this prospectus.
SV022 The Straits Times China’s OpenAI rival Zhipu rises after $715 million IPO Zhipu’s market capitalisation of US$6.6 billion based on the issue price values the company lower than several chipmakers.
SV023 Yicai Global Zhipu AI Soars in Hong Kong Stock Market Debut as Chinese Startup Becomes World's First LLM Firm to Go Public Some 70 percent of the net proceeds from the IPO will be invested in research and development of general-purpose artificial intelligence models.
SV024 StockAnalysis Z.AI Co., Ltd. (HKG:2513) Market Cap & Net Worth Z.AI Co., Ltd. has a market cap or net worth of 397.02 billion as of July 20, 2026.
SV025 HKEXnews MiniMax Group Inc. Global Offering Prospectus MiniMax Group Inc. GLOBAL OFFERING.
SV026 MiniMax MiniMax Investor Relations MiniMax is a global AI foundation model company.
SV027 M&A Insights MiniMax completes HK$4.8 billion IPO on Hong Kong Stock Exchange, shares surge 42% on debut MiniMax completed HK$4.8 billion IPO on Hong Kong Stock Exchange, shares surge 42% on debut.
SV028 StockAnalysis MiniMax Group (HKG:0100) Market Cap & Net Worth MiniMax Group has a market cap or net worth of 60.56 billion as of July 20, 2026.
SV029 CNBC MiniMax doubles in Hong Kong debut, marking yet another Chinese AI listing MiniMax served over 200 million cumulative users and reported revenue of $53.4 million in the nine months ended Sept. 30, 2025, though it still posted a loss.
SV030 TechCrunch China's Moonshot AI raises $2B at $20B valuation as demand for open source AI skyrockets Moonshot AI has raised about $2 billion at a valuation of $20 billion.
SV031 Entrepreneur Loop Moonshot AI Funding Reaches $20B as China's Open-Weight AI Bet Pays Off Moonshot AI funding has ballooned from a $4.3 billion valuation to a jaw-dropping $20 billion.
SV032 Yahoo Finance Moonshot Nears $30 Billion Valuation After Kimi K3 Release Moonshot's valuation had already increased from $4.3 billion in December to $20 billion within five months.
SV033 ChinaBizInsider China AI Compute Crunch: Bubble Risk Grows in 2026 If pricing doesn't bend sharply downward before the capital runs out, revenue will never reach the projections embedded in current valuations.
SV034 Dealroom.co StepFun Secures Series B Funding in Millions StepFun completed a Series B financing round, raising hundreds of millions of dollars.
SV035 NationPress China leads US in AI apps but firms face overvaluation risk Chinese AI firms appear increasingly overvalued relative to their current revenue and profitability fundamentals.