初创公司尽调
尽调报告 Multimodal foundation models / embedded-device AI (China) Pre-IPO (Hong Kong listing filed) 2026-07-01

StepFun

真实的设备与汽车分发楔子,但公开披露暂时撑不起其定价

StepFun 同时握有真实设备和汽车分发、2026 年创纪录融资冲刺,但其 $10-12 billion IPO 目标仍无法由当前公开披露支撑;披露既薄、未经审计,又多来自第三方。

封面要素

IPO 目标估值 01
10-12 USD billion (HK listing filed Jun 2026) [CO039]
IPO 前轮 02
2.5 USD billion (near-close, May 2026) [CO034]
B+ 轮 03
718-733 USD million (closed Jan 2026) [CO026]
报道口径 2025 年营收 04
73.3 USD million (~RMB 500M, third-party-reported) [CO042]
设备部署 05
40M+ smartphones; 1M+ target vehicles (Geely) [CO040]
成立时间 06
Apr 2023 [CO001]
员工规模 07
500+ staff, ~80% technical (2026) [CO013]

公司概况

StepFun(阶跃星辰),正式主体为上海阶跃星辰智能科技有限公司,2023 年 4 月由前 Microsoft 副总裁 Jiang Daxin 在上海创立,共同创始人包括 Zhu Yibo 和 Jiao Binxing,通常被归入中国 AI「六小虎」。公司打造多模态「Step」模型族,覆盖文本、视觉、视频和音频;分发渠道包括按量计费的开发者 API 平台、GitHub / Hugging Face 开放权重、自己的跃问 (Yuewen) 消费者聊天应用,以及与 OPPO、Honor、ZTE、Geely 合作的手机和车端嵌入式部署。2026 年创纪录的资本冲刺——超过 RMB5 billion 的 B+ 轮、接近 $2.5 billion 的 IPO 前轮,以及完成红筹架构拆除——最终指向 2026 年 6 月提交香港 IPO,目标估值 $10-12 billion;但其收入基础仍薄,且大多未经审计。

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, Xuhui District, China
产品
按量计费的 API 平台(platform.stepfun.com / platform.stepfun.ai),服务「Step」多模态模型族,以旗舰 Step 3.7 Flash 智能体式视觉语言模型为核心;同时开源 Step-Audio、Step-Video 与 Step1X 图像 / 音频 / 视频工具,通过 GitHub / Hugging Face 分发,并通过 Yuewen(跃问)聊天应用和嵌入式手机 / 车载助手触达消费者。
客户
手机 OEM(OPPO、Honor、ZTE)和汽车合作伙伴(Geely,经由 Qianli Technology)将 StepFun 模型嵌入端侧;同时服务使用按量 API 平台的开发者和企业。
商业模式
聊天、视觉、语音和智能体模型按 token / 调用计费,并向设备与汽车合作伙伴授权嵌入式多模态功能;不是按席位收取固定软件订阅费。
阶段
Pre-IPO (Hong Kong listing filed June 2026, targeting $10-12B valuation)
融资情况
2026 年 1 月完成超过 RMB5 billion(约 $718-733M)的 B+ 轮;截至 2026 年 5 月,接近完成近 $2.5 billion 的 IPO 前轮;2026 年 6 月 9 日提交香港 IPO,目标估值 $10-12 billion,拟融资约 $500 million。
[CO001, CO013, CO026, CO034, CO039, CO042]

执行摘要

主要优势

  • 真实、具名的设备级量产证明:OPPO、Honor、ZTE 和 Geely(经 Qianli)都在量产硬件中交付由 StepFun 驱动的功能,不是试点。
  • 2026 年融资冲刺创纪录——Series B+ 超过 RMB5 billion、近 $2.5 billion 的 pre-IPO 轮,加上 HKIC 和产业投资方背书——显示资金来源很深,也在变多元。
  • 独立佐证的成本效率架构(稀疏 MoE 路由、MFA/AFD 注意力设计)是真正技术差异点,不只是基准测试包装。
  • 创始人与市场匹配度强:CEO Jiang Daxin(前 Microsoft VP)和领导层来自 ByteDance、Microsoft Bing 以及 ResNet 研究谱系。

主要风险

  • $10-12 billion IPO 目标对据报道 2025 年收入 $73.3 million 隐含 57-68x forward revenue multiple,但公司没有可用的审计财报或已递交招股书。
  • CSRC 推动的红筹离岸架构拆除,以及 2026 年密集监管日程,带来上市 timing 和执行风险。
  • 董事长 Yin Qi 同时担任 Qianli Technology 董事长;Qianli 是 StepFun 唯一已验证的汽车分发伙伴,并且过去四个已披露期间有三个核心净利润为负,治理和渠道风险集中在这段关联关系上。
  • 没有来源披露付费客户数、留存 / NRR、合同条款或股权结构;StepFun 自有 Yuewen 消费 app 虽有 40M+ 设备安装,在独立 MAU 排名中仍落后对手。

未决问题

  • 没有公开的经审计 2025 财报、现金头寸、烧钱速度或毛利率。
  • 完整股权结构和董事会名单仍未披露,包括 HKIC 和国资持股比例。
  • 最终 IPO 定价、融资规模和基石锁定条款尚未确定。
  • 设备与 OEM 伙伴收入分成条款(Geely、OPPO、Honor、ZTE)未披露,因此无法核实 40M+ 设备覆盖背后的 StepFun 实际经济性。

目录

Chapter 01

01公司概览

1.1 身份、创立和产品 / 技术架构

StepFun 是上海阶跃星辰智能科技有限公司的商业名称,公司注册在上海徐汇区。Baidu Baike 的工商登记条目显示,公司成立于 2023 年 4 月 6 日,法定代表人为 Jiang Daxin,初始注册资本 RMB 20 million;Wikipedia、Reuters 等英文报道均指向「2023 年 4 月」,但没有给出更精确的街道地址。Reuters 和 Wikipedia 都将曾参与 Bing 和 Microsoft 365 的前 Microsoft 全球 / 高级副总裁 Jiang Daxin 列为创始人,并列出共同创始人 Zhu Yibo 和 Jiao Binxing。公司的技术身份围绕「Step」多模态基础模型族(语言、视觉、视频、音频)展开;据 MIT Technology Review,StepFun 仍在明确追求通用人工智能,而不少中国同行已经降低了这一目标优先级。StepFun 官方平台目前主推「Step 3.7 Flash」这一面向生产的智能体式 Flash 模型;消费者聊天产品「跃问」(Yuewen) 和配套应用冒泡鸭,则让它在开发者 API 之外也有直达用户的分发。媒体反复将 StepFun 与 Zhipu AI、MiniMax、Moonshot AI、01.AI、Baichuan Intelligence 一起列入中国「六小虎」,并引用 36Kr 采访中的 StepFun 高管说法,称其策略更接近「中国的 xAI 加 Tesla」,而不是纯聊天应用;这反映出公司更强调把模型嵌入物理终端,而非只卖云端订阅。[CO001, CO002, CO003, CO004, CO005, CO006]

KPI 快照表
指标数值/状态日期置信度缺口
成立2023 年 4 月(工商登记为 2023 年 4 月 6 日)
总部上海市徐汇区Baidu Baike 的具体街道地址未获英文报道独立佐证。
公司定位中国「AI 六小虎」之一;聚焦 AGI 的多模态 + 物理 AI 策略2026-07-01
团队规模500+ 名员工,约 80% 为技术人员(2026 年访谈);更早公开数字为 150+2026-07-01没有单一权威的截至某日员工数披露。
B+ 轮(已完成)超过 RMB 5 billion(约 $718-733 million)2026-01-26
IPO 前轮次(接近完成)接近 $2.5 billion;投前估值据报为 $4-6 billion2026-05-08最终规模、定价和交割日期尚未由经审计文件确认。
香港 IPO 申请已提交;目标估值 $10-12 billion,募资约 $500 million2026-06-09实际上市前条款可能变化。
据报 2025 年收入约 CNY 500 million(约 $73.3 million)2025-12-31来自二级市场报道,并非经审计文件。
据报 2026 年收入预测约 CNY 1.2 billion(约 $175.9 million)2026-12-31前瞻性市场估算,不是公司指引。
设备部署规模40M+ 部智能手机;通过 Geely 触达 1M+ 目标车辆;约 20M 日活用户2026-06-09部署数字来自公司/合作伙伴,未经过独立审计。
公开董事会名单已查阅公开资料没有披露完整现任董事会或股权结构表。

融资、收入和部署各行应视为来自二级市场报道的公开参考点,而不是经审计披露。空值标记已查阅公开记录无法清晰验证的事实。

[CO001, CO010, CO020, CO026, CO034, CO039]
FO002: 公司快照逻辑

StepFun 将 AGI 导向的多模态模型组合连接到物理世界终端和快速推进的资本 / IPO 路径;同时,监管重组和披露不足带来下行风险。

[CO006, CO010, CO040, CO049, CO031, CO043]

1.2 领导层、治理和关键人依赖

StepFun 公开可见的领导层比同阶段不少同行更完整,但治理缺口仍然存在。KrASIA 的公司画像列出四位技术负责人:创始人兼 CEO Jiang Daxin;CTO Zhu Yibo,曾负责 ByteDance AI 基础设施和大集群运营;数据负责人 Jiao Binxing,曾任 Microsoft Bing 核心搜索团队负责人;首席科学家 Zhang Xiangyu,是 ResNet 架构核心作者之一。2026 年 1 月 26 日,StepFun 宣布任命 Yin Qi 为董事长,负责整体战略和技术方向;Yin Qi 是 Megvii(Face++)联合创始人、前 CEO,也是 Geely 关联智能驾驶公司 Qianli Technology 的现任董事长。Gasgoo 报道称,Yin Qi 的双重角色明确要把 StepFun 的模型研发接入真实硬件部署数据,形成「Physical AI」反馈回路,而不是只担任顾问型董事长。The Standard 引述对 Jiang Daxin 的采访称,StepFun 团队已超过 500 人,其中约 80% 从事算法和技术岗位;这比 Chinese Wikipedia 较早时期所述约 150 人团队明显扩大。公开记录缺少的是完整现任董事会名单、委员会结构以及股权 / 所有权拆分;考虑到 StepFun 的资本基础高度关联国资,而 Jiang 和 Yin 的战略影响力又很集中,这些缺口很关键。[CO013, CO014, CO015, CO016, CO017, CO018]

领导层与创始人表
人名角色背景创始人-市场契合或职能覆盖关键人物依赖
Jiang Daxin创始人兼 CEO曾任 Microsoft 全球/高级副总裁,负责过 Bing 和 Microsoft 365定下面向 AGI 的多模态路线图,也是公司最主要的公开发言人
Zhu Yibo联合创始人兼 CTO曾负责 ByteDance AI 基础设施和大集群掌管万亿参数模型训练背后的系统和算力策略
Jiao Binxing联合创始人兼数据负责人曾任 Microsoft Bing 核心搜索团队负责人负责支撑模型训练的数据挖掘和 NLP 流水线
Zhang Xiangyu首席科学家ResNet 架构核心作者撑起技术可信度和研究路线图
Yin Qi董事长(2026 年 1 月起)Megvii(Face++)联合创始人、前 CEO;Geely 关联 Qianli Technology 董事长把基础模型研发接到汽车和硬件商业化上

本表覆盖公开具名且角色明确的领导层人物。StepFun 在已查阅材料中没有公布完整董事会或高管名单,因此本表刻意只列部分名单。

[CO013, CO014, CO015, CO016, CO017, CO018]

1.3 融资历史、投资方和估值轨迹

StepFun 的融资历史显示,公司起步不快,但随后加速放大。KrASIA 和 BigGo Finance 都称,天使 / 首轮融资凭借创始人网络把估值推到独角兽水平,投资方包括 Sequoia China(HongShan)、Qiming Venture Partners 和 IDG Capital;2024 年 A 轮引入美元基金 5Y Capital、Shunwei Capital 以及 Legend Capital;2024 年中启动、约 2024 年 12 月关闭的 B 轮,则加入 Shanghai State-owned Capital Investment 及关联基金、Hong Kong Investment Corporation(HKIC)和 Tencent。2026 年 1 月 26 日,TechNode、KrASIA、Yicai 和 C114 均报道 StepFun 完成超过 RMB 5 billion(约 $718-733 million)的 B+ 轮融资,被称为此前 12 个月中国大模型领域最大单笔融资;新进产业和国资关联投资方包括 Shanghai SDIC/S&T Capital InnoFund、China Life Private Equity、Pudong Venture Capital、Xuhui Capital、Wuxi Liangxi Fund、Xiamen ITG Group、Huaqin Technology,Tencent、Qiming、5Y Capital 也继续跟投。到 2026 年 4 月,StepFun 已改制为股份有限公司,并拆除 Cayman Islands「红筹」结构;到 2026 年 5 月,KrASIA 和 CnTechPost 报道称公司接近完成近 $2.5 billion 的 IPO 前轮,投资方包括 HKIC 迄今唯一披露的生成式 AI 投资,以及供应链投资方 Huaqin、Longcheer、OmniVision 和 ZTE。Reuters 引用 Caijing 报道称,该 IPO 前轮目标估值最高 $6 billion,为香港上市前接近 $10 billion 的锚定估值做铺垫;FAQ.com.tw 报道 StepFun 于 2026 年 6 月 9 日提交香港 IPO,目标估值 $10-12 billion,拟融资约 $500 million。[CO023, CO024, CO025, CO026, CO027, CO028]

关键利益方或投资人图谱
关键利益方角色控制权或经济重要性尽调问题
Sequoia China(HongShan)天使/早期投资人支持 StepFun 首轮融资;该轮凭创始人网络达到独角兽估值确认当前持股,以及是否有按比例跟投权或优先权。
Qiming Venture Partners早期且多轮投资人从天使轮一直参与到 2026 年 B+ 轮要求披露各轮累计持股比例。
IDG Capital天使投资人通过创始人网络进入的早期支持方确认当前持股,以及是否有观察员/董事会权利。
Tencent Holdings战略性多轮投资人参与 B 轮和 B+ 轮说明股权是否绑定产品、云或分发合作。
Shanghai State-owned Capital Investment / Shanghai Guosheng Group(国资投资方)国资背景领投方在 B 轮、B+ 轮和 IPO 前轮次中锚定国资支持评估其对 IPO 地点和架构决策的影响。
Hong Kong Investment Corporation(HKIC,投资方)IPO 前投资人按 KrASIA 披露,这是其迄今唯一公开的生成式 AI 投资确认持股规模及与香港上市绑定的任何条件。
Huaqin Technology产业/战略投资人(B+ 轮和 IPO 前轮次)全球头部智能手机 ODM;显示其看重硬件-软件集成说明与投资绑定的商业供货协议。
ZTE Corporation产业/战略投资人(IPO 前轮次)手机合作伙伴,已量产出货 StepFun AI 手机功能确认共同开发条款及任何排他性。

公开股权结构表不完整,因此本图谱列出已披露的战略和财务参与方,而不是最终股权登记簿。

[CO023, CO024, CO025, CO026, CO027, CO035]

1.4 商业化里程碑——模型、合作与规模

StepFun 的产品节奏从一开始就很快:Chinese Wikipedia 和 KrASIA 都把约 100-billion 参数的 Step-1 发布时间放在 2023 年 8 月,也就是公司成立数月后;随后是 2023 年 11 月的多模态 Step-1V,以及 2024 年 3 月万亿参数 Step-2 Mixture-of-Experts 预览。在 2024 年 7 月 WAIC 上,StepFun 正式发布 Step-2、Step-1.5V 多模态模型和 Step-1X 图像生成器。2025 年 2 月,公司开源与 Geely 共同开发的 Step-Video-T2V 和 Step-Audio;到 WAIC 2025,这段汽车合作已经成熟为双方联合展示 Galaxy M9。BusinessWire 称,Galaxy M9 是全球首款搭载「类人 AI agent」的车型,由 StepFun 端到端语音模型和双方共建的 Agent OS 智能座舱系统驱动。StepFun 于 2025 年 7 月 31 日开源旗舰 Step 3(总参数 321 billion、激活参数 38 billion,Apache 2.0);技术论文描述了 Multi-Matrix Factorization Attention 和 Attention-FFN Disaggregation,旨在相对 DeepSeek-V3 和 Qwen3 MoE 235B 降低解码成本。2026 年 2 月,公司继续开源 Step 3.5 Flash(196B / 11B active,256k context)。Reuters 报道,Step 3.5 Flash 此后已跻身 OpenClaw agent 平台使用量前三。部署侧,FAQ.com.tw 和 BigGo Finance 均称 StepFun 模型已为超过 40 million 部智能手机(OPPO、Honor、ZTE)提供 AI 功能,并瞄准超过 1 million 辆 Geely 汽车,服务估计 20 million 日活用户。[CO039, CO040, CO041, CO044, CO045, CO046]

里程碑表
日期事件类型金额/估值/状态参与方含义
2023-04-06StepFun 由 Jiang Daxin、Zhu Yibo 和 Jiao Binxing 在上海创立创立注册资本 RMB 20 millionJiang Daxin、Zhu Yibo 与 Jiao Binxing为后续章节确立可复用的创立锚点。
2023-08-01StepFun 首个大模型 Step-1 发布产品约 100B 参数语言模型StepFun显示创立后数月内技术执行很快。
2023-11-01Step-1V 多模态模型发布产品约 100B 参数多模态模型StepFun释放早期「多模态优先」产品策略信号。
2024-03-01Step-2 预览版(万亿参数 MoE)亮相;A 轮融资融资A 轮引入 5Y Capital、Shunwei Capital、Legend CapitalStepFun、5Y Capital、Shunwei Capital 与 Legend Capital确认独角兽估值首秀后投资人继续支持。
2024-07-04WAIC 2024——Step-2 正式版、Step-1.5V 和 Step-1X 发布产品万亿参数 MoE LLM,加多模态和图像生成模型StepFun巩固 StepFun 在国内同业中的多模态和规模可信度。
2024-12-01B 轮完成融资据 BigGo Finance,为数亿美元Shanghai SSCI、Hong Kong Investment Corporation 与 Tencent在 2025-2026 年扩张前,加深国资和战略投资人背书。
2025-02-01StepFun 与 Geely 开源 Step-Video-T2V 和 Step-Audio产品30B 参数文生视频模型加语音交互模型StepFun; Geely标志着与 Geely 的正式汽车 AI 合作启动。
2025-07-26WAIC 2025——搭载 StepFun 语音智能体和 Agent OS 的 Geely Galaxy M9 展出合作被称为全球首款搭载「类人 AI 智能体」的车辆StepFun; Geely; Qianli Technology展示实验室演示之外的真实物理 AI 部署。
2025-07-31Step 3 发布并开源产品321B 总参数 / 38B 激活参数,Apache 2.0 许可证StepFun确立 StepFun 旗舰多模态推理模型和成本效率叙事。
2026-01-26B+ 轮完成;Yin Qi 出任董事长融资超过 RMB 5 billion(约 $718-733 million);该领域 12 个月融资纪录StepFun、Tencent、Qiming 与 Yin Qi一次公告同时带来面向市场的领导层和创纪录国资/产业资本。
2026-02-01Step 3.5 Flash 开源产品196B 总参数 / 11B 激活参数,Apache 2.0,256k 上下文StepFun把 StepFun 的开源模型节奏延续到 2026 年。
2026-04-02股份制改造完成;红筹架构拆除监管由有限责任公司变更为股份有限公司StepFun在 CSRC 收紧红筹上市规则下,这是香港上市的前置步骤。
2026-04-13Reuters 报道 StepFun 正拆除离岸注册架构反向CSRC 审查红筹上市之际,StepFun 推进拆除 Cayman Islands 架构StepFun; CSRC给本章后续留下实时监管时点风险。
2026-05-08IPO 前轮次接近 $2.5 billion 交割融资投前估值据报为 $4-6 billion;HKIC 加入投资StepFun、HKIC、Huaqin、Longcheer、OmniVision 与 ZTE显示 IPO 申请前数周资本市场需求很强。
2026-06-09StepFun 提交香港 IPO 申请融资目标估值 $10-12 billion,募资约 $500 millionStepFun把多年私募融资推向近期公开市场检验。

本表是第 1 章唯一基准时间线;除非更新证据取代它,否则后续应复用。

[CO002, CO023, CO025, CO026, CO029, CO030]
FO001: 公司里程碑时间线

StepFun 的公开轨迹从 2023 年 4 月创立开始,经过 2024-2025 年模型发布和 Geely 合作,进入 2026 年创纪录融资冲刺、离岸架构拆除,以及 2026 年 6 月香港 IPO 申报。

[CO026, CO029, CO030, CO034, CO039, CO050]

1.5 负面信号——监管重组、披露缺口与 IPO 执行风险

StepFun 第一章记录中最清晰的负面线索不是声誉,而是监管。Reuters 于 2026 年 4 月 13 日报道,中国证券监管机构要求部分红筹公司在赴境外上市前迁回境内后,StepFun 正在拆除离岸「红筹」注册架构;国资重度支持的 StepFun 也认为境内架构更合适。BigGo Finance 的分析,以及 Reuters 自身关于近期获批香港上市队列中约五分之一涉及离岸控股的报道,都把这件事放在行业级合规潮里理解:架构调整会拖慢上市,在一些情况下还会因法律结构变更成本太高而让公司放弃 IPO。时间风险之上,还有估值支撑问题:BigGo Finance 汇编数据将 StepFun 2025 年营收放在约 CNY 500 million(约 $73.3 million),而香港 IPO 目标估值为 $10-12 billion;本章把这条差距明确列为尽调警讯,而不是既定结论,因为营收数据来自二级市场报道,不是经审计申报文件。最后,限制领导层可见度的同一种不透明也延伸到所有权:已审阅公开来源没有披露完整董事会名单或股权结构;对于一家刚改制为股份有限公司、数月内就寻求公开市场资本的公司,这一点尤其值得注意。[CO022, CO030, CO031, CO032, CO033, CO042]

FO003: 快照 KPI

公开证据支持的指标显示,StepFun 是一家快速扩张、资本充足的多模态 AI 公司,正走向近期香港 IPO;收入与估值之间的缺口是最突出的风险标记。

[CO020, CO026, CO034, CO042, CO039, CO031]

1.6 展示材料

Chapter 02

02市场分析

2.1 市场边界、纳入支出和现状替代方案

StepFun 应被视为一家在中国基础模型应用层和嵌入式设备 AI 层竞争的公司,而不是中国整个 AI 或半导体经济的替身。IDC 的中国 AI agent market-glance 正是沿这条线切分:它把模型开发、应用开发和 agent 开发平台,与底层算力、存储、网络和半导体基础设施堆栈区分开。在这条边界内,StepFun 自己的 platform.stepfun.com 文档显示,公司通过聊天、视觉、语音和智能体式「step-3.x」模型按 token 计费,而不是售卖固定软件许可证;Step-Audio 语音模型已经作为嵌入式座舱软件装进 Geely 量产 Galaxy M9,也把汽车智能座舱授权而非仅消费者聊天明确纳入可服务支出。第四个容易被忽视的部分是合规:截至 2026 年 4 月 30 日,中国已有 868 个国家级 CAC 备案生成式 AI 服务和 530 个地方备案应用,StepFun 的每一个公开界面都要通过同样的备案、安全评估和内容标识门槛;合规因此是持续经营成本,不是法律脚注。排除在边界之外的是更大的基础设施层——Forrester 估计中国 2026 年 AI 基础设施支出超过 $70 billion——StepFun 消耗这层支出作为成本投入,而不是直接变现。StepFun 各界面的现状替代方案,从没有嵌入生成式 AI 的原生手机或车机软件,到 OEM 合作伙伴可改选的更大 hyperscaler 分发替代品(Doubao、Qwen)不等。[CM001, CM002, CM003, CM004, CM007]

市场定义表
细分/类别纳入支出排除支出买方/付费方与 StepFun 的关联
消费者 AI 助手应用聊天/语音智能体应用使用、应用内增购通用消息应用、人工客服人力个人消费者StepFun 的 Yuewen(跃问)聊天应用和冒泡鸭陪伴应用
开发者/API 基础模型支出聊天、视觉、语音及智能体式 Step-3.x 模型的 token 计费;OpenClaw 智能体运行时消耗原始 GPU 采购、通用云 IaaS、无关 DevOps 工具开发者、智能体构建者、产品/工程预算platform.stepfun.com API 通道;Step 3.5 Flash 在 OpenRouter/OpenClaw 排名前 3
智能手机内嵌 AI 授权捆绑进手机软件的端侧/边云多模态功能与 AI 无关的手机硬件 BOM、电信运营商支出OEM(OPPO、Honor、ZTE)工程/软件预算已披露 40M+ 部智能手机部署规模
汽车座舱内嵌 AI智能座舱语音/智能体 OS 授权、按车计的软件收入整车硬件 BOM、EV 电池/传动系统支出汽车 OEM(Geely)研发和产品预算Geely Galaxy M9 中由 Step-Audio 驱动的「Agent OS」;100 万辆目标
中国合规/智能体治理支出与已部署生成式 AI 和智能体服务绑定的备案、标识和安全评估成本与 AI 无关的独立网络安全支出、通用法律支出平台所有者和受监管运营方StepFun 的每个触点都必须通过与 868 个全国备案服务相同的备案和标识关口
排除的基础设施层NoneGPU/半导体、云 IaaS、电信网络资本开支基础设施和芯片采购方不在 StepFun 直接变现范围内;算力是投入成本和约束,不是收入

这些行把 StepFun 可观察的消费者、开发者/API、内嵌设备和合规触点,与它不直接变现、规模大得多的 AI 基础设施层拆开;边界参考 IDC 的中国 AI 智能体市场概览, 以及 StepFun 官方产品/定价页面。

[CM001, CM002, CM003, CM004, CM007]

2.2 没有干净单一数字的市场,只能用证据约束几组规模镜头

公开市场没有一个数字能干净框定「StepFun 的市场」,所以本章保留几组有来源支撑的镜头,而不是人为拼出一个 TAM。最窄的应用镜头来自 IMARC 的中国生成式 AI(AIGC)市场估计:2025 年规模约 $5.16 billion,到 2034 年增至 $19.56 billion,CAGR 为 15.96%;这很可能低估 StepFun 的真实可服务池,因为它排除了嵌入式设备授权。更大但相邻的镜头是基础设施:Forrester 预计中国 AI 基础设施支出将在 2026 年超过 $70 billion,Gartner 则把全球 AI 支出放在 $2.59 trillion(同比 +47%);这些数字位于 StepFun 实际变现层的上游。使用量镜头来自中国公有云模型即服务消费,2025 年达到 1,944 trillion tokens(同比 16x),仅 ByteDance Volcano Engine 就占 49.5% 份额,说明使用量扩张远快于任何已经稳定的营收池。监管规模镜头显示,截至 2026 年 4 月 30 日,全国备案生成式 AI 服务有 868 个,仅两个月增加 72 个;全球份额镜头(Nikkei,经 TrendForce)则显示,中国开源模型在全球使用份额一年内从约 1% 跳到 15%。有用的是,IDC 自己的中国基础模型跟踪(CHC5530125)也把 StepFun 的「Step Reasoner mini」与 Qwen3-Max、DeepSeek、Kimi K2、GLM-4.6 并列为其跟踪的前沿模型之一;IDC 还单独设定国家目标:到 2027 年智能设备 / agent 渗透率达到 70%,到 2030 年达到 90%,并预计中国 top-1,000 企业中 80% 到 2027 年会优先考虑 AI 主权。这些镜头不能相加成一个总数;合在一起,它们描绘的是一个增长快、使用量充沛、但商业化仍未定型的市场。[CM008, CM009, CM010, CM011, CM012, CM013]

TAM / SAM / SOM 或规模测算视角表
发布方年份地域数值CAGR / 采用率方法置信度局限
IMARC Group2025中国USD 5.16B 生成式 AI(AIGC)应用市场 -> 2034 年 USD 19.56B15.96% CAGR 2026-2034联合市场规模测算报告应用口径较窄;排除内嵌设备授权,完整方法未公开
Gartner2026全球USD 2.59T 总 AI 支出+47% YoY;GenAI 模型 +110%分析师支出预测全球口径,非中国或 StepFun 品类专属
Forrester2026中国> USD 70B AI 基础设施支出属于中国科技支出增长 7% 的一部分分析师支出预测仅基础设施层;排除 StepFun 实际变现的模型/应用层收入
IDC(经 Volcano Engine / China Daily)2025中国公有云 MaaS 调用 1,944 trillion tokens16x YoYIDC 汇总的厂商上报使用量使用量/token,不是收入;主要覆盖超大云平台(Doubao、Qwen),而不是 StepFun 这类独立厂商
Cyberspace Administration of China(CAC,监管机构)2026中国868 个备案服务 + 530 个备案应用2026 年 3-4 月全国 +72 / 地方 +49监管备案数量统计服务数,不是收入或使用量;StepFun 单项备案状态未公开逐项列出
中商产业研究院(经 seccw.com)2026中国RMB 182.8B 智能座舱解决方案市场高于 RMB 154.5B(2025 年)和 RMB 129B(2024 年)行业研究预测座舱软件和硬件打包统计;StepFun 的单车收入分成未披露
IDC2026中国147M 部 AI 手机出货,占 278M 总量的 53%首年 AI 手机占比过半厂商出货预测口径把具备 AI 能力的手机算得较宽;StepFun 是几个嵌入式模型供应商之一(OPPO、Honor、ZTE),不是独家供应商

各口径使用互不兼容的单位(USD、RMB、tokens、注册量、整机出货量),不能加总成一个 TAM;每个口径只覆盖 StepFun 可触达机会中不同且部分重叠的切片。

[CM008, CM009, CM010, CM011, CM012, CM014]
FM001: 市场规模测算视角

从中国广义「AI+」渗透目标,到更窄且不重叠的代理指标,逐层定位 StepFun 可触达的市场切片。

这个金字塔有意混合政策目标、基础设施支出、使用 token 和设备数,因为没有单一公开来源披露 StepFun 精确的「嵌入式设备 + Agent API」类别的独立 RMB 或 USD TAM。

[CM014, CM010, CM011, CM017, CM006]

2.3 设备嵌入式 AI:StepFun 真正追逐的是手机、PC 和座舱界面

StepFun 自己的定位很明确:它不是只争云端聊天流量,而是把模型嵌进智能手机、汽车和工业硬件。公开报道将 StepFun 旗舰模型与超过 40 million 部智能手机的 AI 功能、以及通过 Geely 合作瞄准超过 1 million 辆汽车的部署联系起来;只有底层设备品类本身也在增长,这些数字才有意义。手机侧,IDC(另有一家中国媒体交叉印证)预测 2026 年中国出货量将下降 2.2% 至约 278 million 台,但下一代 AI 手机出货量会升至 147 million 台,占总量 53%,也是 AI 手机首次跨过市场一半,模式是「端云协同」。PC 侧,IDC 预计 2026 年中国市场约 42.22 million 台、大体持平,但 GenAI PC 同比增长 146.5%,到 2029 年估计达到品类的 36.5%。汽车座舱侧,中商产业研究院行业预测显示,中国智能座舱解决方案市场 2026 年达到 RMB 182.8 billion,高于 2025 年的 RMB 154.5 billion;语音交互渗透率一年内几乎翻倍,从 26% 升至 48%。Geely 的 Galaxy M9 在 WAIC 2025 亮相,双方都称其为行业首款搭载端到端语音 LLM 驱动「Agent OS」座舱的量产车,底层采用 StepFun 的 Step-Audio 模型。这些设备品类都在扩张,但 StepFun 只是多个争取 OEM 定点的嵌入式模型供应商之一;目前只有 Geely 这一段关系被公开确认达到规模,独立分析也把这种集中度列为潜在天花板,限制 StepFun 真实可服务市场的独立性。[CM017, CM018, CM019, CM020, CM021, CM022]

FM004: 采用漏斗或价值链地图

StepFun OEM(智能手机 / 汽车)渠道从模型展示到规模化、合规设备部署的实际路径。

该流程抽象了从展示到规模化设备部署的已观察路径;并非每一段 StepFun-OEM 关系都一定按这个精确顺序经历所有阶段。

[CM021, CM017, CM003, CM004, CM022]

2.4 四个界面的买方、用户、付款方与采用路径

StepFun 的买方地图不是一条漏斗,而至少是四条并行路径。个人消费者为其消费者聊天和语音应用自付费,或使用免费层。开发者和 agent 构建者直接在 platform.stepfun.com 上按使用量计费;StepFun 2026 年 3 月推出的 Step Plan 订阅从标准版 RMB 49 到 RMB 699 / 月分四档,并长期给予开发者社区折扣,明显是为 OpenClaw-agent 和 AI coding 工作流设计。独立的使用量证据也支撑这一段牵引力:OpenRouter 真实调用排行榜显示,Step 3.5 Flash(free)在 2026 年 3 月 9-15 日当周以 1.34 trillion tokens 位居全球第二,环比增长 79%。智能手机和汽车 OEM 是第三类、结构完全不同的买方:采用触发点是旗舰产品周期或联合展示,例如 Galaxy M9 在 WAIC 2025 亮相,而不是自助试用;因此这一渠道依赖少数高价值关系,而不是广泛转化。第四类买方是企业或受监管部署方,上线前必须通过 CAC 备案和内容标识要求;IDC 自己的中国 AI agent market-glance 显示,这类买方分布在多种职能(金融、HR、客服、供应链)和行业(FSI、制造、汽车、零售)中,没有单一企业垂直领域主导机会。同一开发者细分也很拥挤:IDC 追踪的中国 AI coding IDE 竞争者至少十余家,包括 Alibaba Qoder、ByteDance Trae、Baidu Comate、Tencent Codebuddy 等;即便 IDC 预计到 2027 年中国 35% 专业开发者会采用 AI-native coding 平台,2025 年头部厂商 coding 收入在三位数使用量增长下仍低于 RMB 100 million,StepFun 自身份额还未计算在内。[CM024, CM025, CM026, CM027, CM028, CM038]

细分市场 / 买方地图
细分市场买方用户付费方工作流预算归属采用触发因素
消费级聊天 / 语音应用(Yuewen / 冒泡鸭)个人消费者同上同上(或广告 / 免费层)日常聊天、语音和陪伴应用任务本人应用商店发现、社交声量
开发者 / 智能体 API(platform.stepfun.com)开发者或智能体构建者同上,或下游应用用户工程 / 产品预算,或个人信用卡智能体工作流、编程、OpenClaw 集成工程 / AI 产品预算低价自助试用;OpenRouter / OpenClaw 排名可见度
智能手机 OEM 嵌入式 AIOPPO / Honor / ZTE 硬件与软件团队手机终端消费者OEM 研发 / 软件授权预算打包进手机软件的端侧多模态功能OEM 产品 / 软件预算旗舰机周期、AI 手机功能跟进压力
汽车 OEM 嵌入式座舱 AIGeely Galaxy 工程 / 产品团队车辆驾驶员 / 乘客OEM 研发 / 单车软件预算语音和智能体座舱交互汽车研发 / 产品预算新车型上市(如 Galaxy M9)、WAIC 展示
企业 / 监管场景部署企业 IT 和合规团队企业员工 / 客户平台 / 合规预算上线前完成网信办备案、标识和安全评估中央 IT / 合规 / 法务预算监管备案截止期或审计要求

这些行描述了 StepFun 自有产品 / 定价页、Geely 合作报道和 IDC 中国 AI 智能体市场速览所能证明的四条不同买方阶梯;StepFun 在这些阶梯上的收入结构未公开拆分。

[CM024, CM025, CM026, CM027, CM028]
FM003: 买方 / 分群地图

StepFun 四个主要中国需求面的预算归属、采用触发和摩擦差异。

[CM024, CM026, CM027, CM022, CM033]

2.5 增长驱动、价格竞争、算力约束和监管

三个结构性驱动有利于 StepFun。中国 2026 年 5 月 CAC / NDRC / MIIT《实施意见》是国内首个面向 AI agent 的国家级专门政策框架,正式定义 AI agent,并点名科研、制造、交通、金融、医疗、政务等 19 个优先部署场景。IDC 的「AI+」目标——2027 年智能设备 / agent 渗透率 70%、2030 年 90%,以及 2027 年中国 top-1,000 企业中 80% 优先考虑 AI 主权——为国产、可嵌入模型创造持久需求。中国设备换新周期(AI 手机、GenAI PC、智能座舱)也在扩大嵌入式 AI 的物理触点。但同样的力量也构成约束。中国 2023 年生成式 AI 办法和 2025 年 AI 内容标识规则(2025 年 9 月 1 日生效)对每一个已部署界面都施加持续的内容安全、数据和显性 / 隐性标识义务。算力获取在 2026 年 5 月 31 日进一步收紧:U.S. Commerce Department 堵上漏洞,禁止中国总部企业通过海外子公司无许可证购买先进 Nvidia / AMD 芯片;出口管制从交付地延伸到所有权。价格压缩同样迅速:2026 年上半年,中国实验室至少六次下调 API 价格,DeepSeek、Xiaomi MiMo 和 Kimi 将降价永久化,使每 token 价格比西方前沿 API 低 5-34x;StepFun 自己公布的 step-3.5-flash 和 step-3.7-flash 费率也处于同一竞争区间,而不是溢价区间。信任也承压:中国主管部门已就 OpenClaw agent 平台上的植入木马假冒技能包发布高级别警示,而 Step 3.5 Flash 正是在这个生态的使用量榜单靠前。最后,分发并不对称:ByteDance Doubao 单独就覆盖超过 500 million 部手机终端和超过 7 million 辆汽车,约比 StepFun 已披露足迹高一个数量级;全球车企也越来越多与本地伙伴(Huawei HarmonyOS、XPeng、Momenta)共研面向中国的座舱 AI,而不是押注单一供应商。与此同时,中国基础模型层 2025 年融资额约减半至约 22 笔、RMB 9.4 billion,独立玩家与这些更大平台竞争时可用资本更紧。[CM029, CM030, CM031, CM032, CM033, CM034]

增长驱动与约束表
驱动 / 约束方向时点影响尽调问题
国家「AI+」智能体政策框架(网信办 / 发改委 / 工信部,2026 年 5 月)驱动已生效;优先场景覆盖 2027-2030 年为智能体部署背书并拉齐标准,打开 19 个优先行业相比超大云厂的智能体平台,StepFun 的嵌入式模型现实中能覆盖 19 个优先场景里的多少个?
AI 生成内容标识规则(2025 年 9 月生效)约束已实施每个已部署触点都要承担显式 / 隐式标识工程负担StepFun 每个触点的标识合规成本和审计状态如何?
网信办生成式 AI 服务备案(868 项服务,2026 年 4 月)约束兼驱动持续推进,约每两个月新增 72 项备案备案已经是市场准入门票,也说明可触达服务数量在快速增长StepFun 的 Step 平台是否在备案服务名单中,对应哪些备案编号?
AI 手机 / AI PC / 智能座舱设备周期(IDC、中商产业研究院)驱动2026 年换机周期扩大 StepFun 这类模型的设备端嵌入触点StepFun 实际拿下多少 AI 手机和智能座舱设计定点,相比 Doubao 或 Qwen 的份额如何?
芯片出口管制收紧(美国封堵漏洞,2026 年 5 月)约束立即生效提高 StepFun 下一代模型训练算力获取的不确定性和成本StepFun 的训练算力供应链对已被封堵的海外子公司漏洞有多大敞口?
中国 LLM API 价格战(DeepSeek、MiMo、Kimi,2026 年上半年)约束持续且在加速压缩可实现 API 利润率,迫使 StepFun 跟进低价StepFun 能否把 step-3.x 价格维持在接近 DeepSeek / Kimi 的水平,同时保持盈利?
OpenClaw 智能体平台安全事件(木马技能包)约束2026 年初以来持续Step 3.5 Flash 排名靠前的平台若出现安全疑虑,企业采用智能体可能放慢StepFun 对自有智能体平台分发采用哪些安全和验证控制?

有几行同时是驱动也是约束(如备案量、设备周期);方向列反映对 StepFun 可触达市场的近期净影响,不是永久分类。

[CM029, CM030, CM031, CM032, CM033, CM034]
FM002: 市场估算区间

StepFun 自有 API 模型与中国领先前沿模型对手的公开价格带:低位(缓存命中)/ 中位(标准输入)/ 高位(输出),单位为 CNY / 1M tokens。

StepFun 价格来自 platform.stepfun.com,以 CNY 标价。DeepSeek、Qwen 和 Kimi 价格按 apidog 发布的 USD/MTok 数字,以约 7.15 CNY/USD 汇率换算;低位 = 缓存命中价,中位 = 标准(未命中缓存)输入价(Kimi K2.6 的挂牌输入价为分层 $0.16-$2.00,因此中位采用公开层级下限),高位 = 输出价。这些区间是公开标价,不是实际净合同价。

[CM033, CM034]

2.6 展示材料

Chapter 03

03竞争格局

3.1 格局:同业创业公司、大厂嵌入式模型和相邻入局者

StepFun 的真实竞争集比第一章用来描述创始批次的「六小虎」框架更宽。到 2026 年 Q2,中国基础模型领域已经收敛到约十家拥有可观公开使用份额的供应商——Xiaomi、Alibaba(Qwen)、Zhipu AI(国际上改名为 Z.ai)、DeepSeek、Moonshot AI、MiniMax、StepFun、ByteDance、Baidu 和 Tencent——其后还有更长尾的垂直或小众玩家(Baichuan Intelligence、Yi、Xunfei、KwaiKAT)。在这组公司里,StepFun 与 Zhipu AI、Moonshot AI、MiniMax、Baichuan Intelligence 同属风险资本支持的纯基础模型创业公司;这不同于 Alibaba Qwen、Baidu ERNIE、ByteDance Doubao / Seed 等嵌入大厂的项目,后者依托既有云、搜索或消费者应用分发栈;也不同于 DeepSeek,后者在 2026 年中首次外部融资前一直由 High-Flyer 资助、独立于外部投资人。 还有两类相邻竞争者和具名同业同样重要。第一,Xiaomi MiMo 模型族并非 StepFun 传统上被拿来比较的同业,但它单独就在 2026 年 Q2 占中国供应商 OpenRouter 周度 token 量约 21%,超过任何一家纯创业公司;这说明设备巨头仅靠智能手机和开发者工具捆绑,就能在分发上压过所有融资充足的 AI 实验室,而这正是 StepFun 试图用作自身楔子的渠道。第二,对于部分受监管中国买方而言,内部自建或依赖受阻 / 受出口管制的海外前沿 API 仍是现状替代方案;但数据本地化规则和芯片出口管制,会把大多数国内需求推向本章画像中的十家中国供应商。[CP001, CP002, CP003, CP004, CP041]

竞争对手画像表
竞争对手类别规模 / 融资信号目标细分市场差异化局限
StepFun独立多模态基础模型创业公司接近 $2.5B 的 IPO 前轮;已递交香港 IPO,目标估值 $10-12B(背景见第 1 章)消费级聊天、开发者 API、智能手机 / 汽车 OEM 嵌入嵌入式多模态设备模型,加上一条明确且仍在推进的 AGI 路线图相比已融资或上市同行,消费应用触达偏弱,收入基础也基本未经审计
Zhipu AI(Z.ai,竞品)独立基础模型创业公司,香港上市首家上市 LLM 公司;FY2025 收入约 CNY724M(同比 +131.9%);市值 > HK$400B企业 / API 加消费级 chat.z.aiMIT 许可的 GLM-5(744B MoE)基于 Huawei Ascend 训练;深度接入 ByteDance / Alibaba / Tencent收入高速增长,但 FY2025 净亏损约 CNY4.72B
Moonshot AI (Kimi)独立基础模型创业公司2026 年 5 月以 $20B 估值融资 $2B,较 2025 年底 $4.3B 大幅抬升消费级 Kimi 应用加开放权重开发者 API长上下文、100 智能体集群架构;Kimi K2.5 支撑 Cursor Composer 2Kimi 消费应用 MAU(约 25M)仍比 Doubao 低一个数量级
MiniMax独立基础模型创业公司,香港上市2026 年 1 月香港 IPO 融资约 $619M,估值约 $6.5B消费应用(如 Xingye)加智能体开发者 API自进化智能体 M 系列(M2-M2.7);OpenRouter 供应商份额第 4刚上市,仍要面对 2026 年 AI 公司上市常见的收入质量和 IPO 首日涨幅审视
Baichuan Intelligence独立基础模型创业公司,转向垂直场景2024 年 A 轮融资约 $691M,估值 $2.7B;2026 年 4 月 B 轮估值仍接近该下限医疗 AI(医院、诊所)加企业 LLM垂直医疗 LLM(Baichuan-M2)在部分开放模型基准上优于 OpenAI已解散金融 / 教育团队;平台野心比 StepFun 更窄
DeepSeek投资人独立的前沿实验室(High-Flyer 孵化)首轮外部融资:2026 年 6 月以 > $50B 估值融资约 $7.4B,创始人控制结构开放权重 API 加消费级聊天应用2026 年已公开降价幅度市场最陡;V4 家族采用 MIT 许可投资人结构让多数外部投资者没有投票权,限制常规治理可见度
Alibaba Qwen大科技公司内嵌模型家族由 Alibaba Group 资产负债表供血;Qwen3.6-Plus 接入 Wukong / Qwen App企业云 API(Model Studio)加消费级 Qwen App发布节奏最宽;OpenRouter 周份额约 13.9%,仅次于 Xiaomi分发依托 Alibaba 既有云和合规栈,而不是独立护城河
Baidu ERNIE大科技公司内嵌模型家族由 Baidu 资产负债表供血;ERNIE 5.1 训练成本约为同业预训练成本的 6%搜索内嵌消费级助手加企业 Qianfan API通过 Baidu Search 获得最深的单一消费分发触点(约 200M MAU)ERNIE 5.1 未开源,不同于 ERNIE 4.5 或多数具名同行
ByteDance Doubao (Seed)大科技公司内嵌模型家族由 ByteDance 资产负债表供血;没有单独披露融资轮中国最大的消费级 AI 应用加企业 Volcano Engine APIMAU 约 336-345M,月会话约 54.8 次 / 用户;Seed 2.0 Pro 定价约为 GPT-5.2 的十分之一独立披露极少;市场主要靠应用使用量判断表现,而不是模型层面申报

规模 / 融资数字混合了已披露私募轮金额、IPO 募资和母公司内部供血,不能直接横向比较;更窄的同口径比较见定价和护城河表。StepFun 的规模数字在本章新来源下重述第 1 章背景。

[CP001, CP002, CP005, CP006, CP007, CP009]
FP001: 竞争定位图

按独立基准强度和消费者 / 设备分发能力,对九家画像供应商做序数映射。

坐标轴是基于 BenchLM 2026 年 7 月榜单分数和独立 MAU / 使用份额报道合成的有证据支撑的序数判断,不是单一标准化指数;StepFun 的位置反映真实的 OpenRouter 免费模型前三排名,但被其缺席全国前五榜单和前 50 消费者应用排名所抵消。

[CP023, CP025, CP029, CP030, CP032, CP034]

3.2 竞争者画像:规模、融资和战略方向

Zhipu AI 在国际上改名为 Z.ai,并通过 2026 年 1 月香港 IPO 成为首家上市的大语言模型公司;其 FY2025 年报显示,收入约 CNY 724 million(约 $105 million,同比增长 131.9%),净亏损约 CNY 4.72 billion,到 2026 年中市值高于 HK$400 billion。其旗舰 GLM-5 是 744-billion 参数 MoE 模型,采用 MIT license,并端到端在 Huawei Ascend 芯片上训练,把 Zhipu 定位成面向国资关联和企业买方的抗出口管制选择。Moonshot AI 是 Kimi 模型族的开发者,2026 年 5 月由 Meituan Long-Z Investments 领投,以 $20 billion 估值融资约 $2 billion;这一估值高于 2025 年底的 $4.3 billion。Kimi K2.6 / K2.5 模型在长上下文和 agent-swarm 协同上竞争,并嵌入西方开发者工具 Cursor Composer 2。MiniMax 也在 2026 年 1 月窗口期赴港上市,以约 $6.5 billion 估值融资约 $619 million,GIC、Baillie Gifford 和 Abu Dhabi Investment Authority 为基石投资者,并推出自进化智能体 M 系列(M2-M2.7)。 Baichuan Intelligence 走了另一条路:它从更广的平台战略转向垂直医疗 AI,解散金融和教育团队,聚焦 Baichuan-M2 等模型;Baichuan-M2 是开源医疗 LLM,在 HealthBench benchmark 上得分高于 OpenAI gpt-oss-120b,同时仍获得 Alibaba、Tencent、Xiaomi 和多只国资背景基金支持。DeepSeek 是 High-Flyer 对冲基金孵化公司,在 2026 年 6 月之前没有外部投资人;该轮融资超过 CNY 50 billion(约 $7.4 billion),估值高于 CNY 330 billion,结构上多数外部投资人没有投票权,而 National AI Industry Investment Fund 直接投资并拥有完整投票权。Alibaba Qwen、Baidu ERNIE 和 ByteDance Doubao / Seed 构成大厂嵌入式梯队:Qwen3.6-Plus 接入 Alibaba 的 Wukong 企业平台和消费者 Qwen App;ERNIE 5.1(不同于 ERNIE 4.5,未开源)声称以约 6% 的同业训练成本排到 Arena Search 全球第 4;Doubao / Seed 2.0 则是 ByteDance 以消费者为先的旗舰,未披露独立融资轮。[CP005, CP006, CP007, CP009, CP010, CP013]

3.3 比较维度:能力、定价、分发和信任姿态

原始能力上,独立的 2026 年 7 月 benchmark 跟踪把 Qwen3.7 Max 和 GLM-5.2 放在中国模型排行榜首位(各 90 分),DeepSeek V4 Pro Max 以 87 分紧随其后,Kimi K2.6 为 81 分;StepFun 的模型虽然凭 Step 3.5 Flash 在 OpenRouter 免费模型榜 #3 的排名成为真实的「惊喜选手」,但没有出现在该榜单全国前五中国模型中。定价上,StepFun 自己的 step-3.7-flash(每 million tokens CNY 0.27 cache-hit / 1.35 standard input / 8.1 output)和 step-3.5-flash(0.14 / 0.7 / 2.1)处在与 DeepSeek、Qwen、Kimi 各档类似的 CNY 计价宽区间里;但 2026 年 API 价格战几乎完全由其他实验室驱动:DeepSeek、Xiaomi MiMo 和 Moonshot Kimi 都宣布 2026 年降价永久化,DeepSeek V4-Pro 现在 output 价格为 $0.87(约比 GPT-5.5 低 34x),Kimi K2.6 则以 $0.07 保住市场最低 cache-hit 底价。 分发能力比价格分化更尖锐。ByteDance Doubao 是中国使用量最高的消费者 AI 应用,MAU 达 336-345 million;Alibaba 的 Qianwen 和 Quark 应用合计约 380 million;Baidu 的 ERNIE 驱动助手(Wenxiaoyan)月活约 200 million、注册用户超过 300 million;即便 DeepSeek 自有消费者应用也以约 139 million MAU 位列全国 #4。相比之下,Kimi 消费者应用以约 25 million MAU 排第 8,StepFun 的 Yuewen 根本没有进入中国 AI 应用 MAU 前 50。信任和硬件姿态上,GLM-5(Huawei Ascend)和 Baidu ERNIE(Baidu Kunlun)在国产硬件独立性上领先,这对受出口管制敏感的买方尤其相关;StepFun 则在 NVIDIA Hopper 上训练,这与行业大多数玩家相同,但不是差异化。由于这些 API 大多兼容 OpenAI,且数家提供宽松许可的开放权重,企业和开发者买方可以多供应商并用或切换供应商,技术摩擦相对较低;因此,包括 StepFun 在内,任何单一供应商都难以仅靠定价或模型质量榨出太多锁定效应。[CP008, CP011, CP012, CP015, CP021, CP022]

功能 / 能力矩阵
能力StepFunZhipu GLMMoonshot KimiMiniMaxBaichuanDeepSeekAlibaba QwenBaidu ERNIEByteDance Doubao
开放权重 / 自托管可用性部分强(MIT)强(开放权重)强(开放 M2)强(MIT)强(宽松许可)有限(5.1 闭源)有限(多数闭源)
消费应用分发规模弱(MAU 未进前 50)有限中(约 25M MAU)有限至中弱(临床 / 垂直)强(约 139M MAU)强(约 220M + 162M MAU)强(约 200M MAU)最强(约 336-345M MAU)
设备 / OEM 嵌入(智能手机、汽车)中(40M+ 部手机,Geely 目标)有限有限有限有限有限有限(Alibaba 生态设备)有限中(ByteDance / Douyin 生态设备)
企业 / API 定位广度强(企业优先)中(智能体工作流)窄(医疗垂直)强(自托管友好)最强(Alibaba Cloud 栈)强(Baidu Qianfan)中(Volcano Engine)
编程 / 智能体基准实力国内前五未上榜强(77.8% SWE-Verified)强(Cursor Composer 底座)强(80.2% SWE-Bench)非编程重点强(IMO / IOI 金牌;约 49% 编程 token 来自 MiMo)搜索原生,不以编程为先非主要基准重点
定价激进度(2026 年价格战)中档中档(输出 $3.20)缓存命中底价最低智能体折扣激进不参与 API 价格战整体最便宜(输出约 $0.87)中档(输出 $3.90)非 API 价格战主要参与者超低价、贴近消费端的档位
国产硬件独立性NVIDIA HopperHuawei Ascend混合混合混合当前 NVIDIA,计划转 AscendNVIDIA + 混合Baidu Kunlun混合
资本获取 / 国家资本背书披露中(私有,IPO 前)强(香港上市)强($20B 私募估值)强(香港上市)中(国资基金,私有)强(国资基金持股,私募估值 > $50B)强(Alibaba 资产负债表)强(Baidu 资产负债表)强(ByteDance 资产负债表)

评级是基于保留的官方产品 / 定价页和独立市场份额 / 基准报道综合得出的有证据支撑的定性判断,不来自单一厂商中立评分源;证据不足的单元格标为「有限」或「非主要重点」,而不是猜测。

[CP007, CP009, CP011, CP015, CP021, CP023]
定价 / 打包比较
供应商 / 模型缓存命中输入(每 1M tokens)标准输入(每 1M tokens)输出(每 1M tokens)上下文备注
StepFun step-3.7-flash 模型CNY 0.27CNY 1.35CNY 8.10原生StepFun 开放平台原生 CNY 标价
StepFun step-3.5-flash 模型CNY 0.14CNY 0.70CNY 2.10256K开放权重 flash 档
DeepSeek V4-Pro(2026 促销,现已成为标准价)$0.003625$0.435$0.871MMIT 许可;DeepSeek 文档显示,促销 75% 折扣已永久化
DeepSeek V4-Flash$0.0028$0.14$0.281MDeepSeek 最便宜档位
Alibaba Qwen3 Max未公布$0.78$3.90262K定位为质量 / 成本平衡,不是最低价档
Moonshot Kimi K2.6$0.07分层 $0.16-$2.00$1.72(K2.5 参考)262K2026 年主要中国前沿 API 中跟踪到的缓存命中底价最低
Zhipu GLM-5(直连 API)未公布$0.80$2.56(Turbo $4.00)200KMIT 许可权重
ByteDance Doubao Seed 2.0 Lite未公布$0.10-$0.25~$2.00未披露超低价、贴近消费端的档位;ByteDance 很少披露独立模型层面定价细节

StepFun 价格来自其自有平台的原生 CNY;其他行按公开发布或独立 2026 定价比较跟踪为 USD,因此没有把单元格换算为同一种货币——跨行比较只宜看方向,不能当作精确同口径数字。

[CP008, CP012, CP021, CP022, CP027, CP033]
FP002: 功能广度 / 能力地图

对九家画像供应商的开放权重广度、分发、设备嵌入、基准强度和资本获取做能力对照。

[CP003, CP007, CP009, CP021, CP024, CP025]

3.4 护城河耐久性、商品化风险和竞争者负面证据

StepFun 看起来最耐久的主张不是 benchmark 领先,也不是消费者触达,而是嵌入式多模态设备分发(手机预装和 Geely 汽车关系,见第二章)叠加对 AGI 导向模型范围的明确、持续追求;若干同行已经降低了这一追求的优先级。这个楔子是真实的,但有条件:Xiaomi MiMo 系列已经通过同一设备渠道在使用量上超过 StepFun;Geely 自己 2026 年 4 月「Full-Domain AI 2.0」材料也描述了多供应商、越来越内生化的座舱和 ADAS 战略,而不是排他依赖任何单一外部模型供应商。对于集中在一个 OEM 的关系来说,这是实质性替代风险。 负面证据不止来自某一个竞争者。第一,资本获取:Zhipu 和 MiniMax 已经上市,Moonshot 以 $20 billion 估值融资,DeepSeek 以超过 $50 billion 估值融资(后者还获得 National AI Industry Investment Fund 直接且有投票权的投资),这些资本基础都比 StepFun 报道中的近 $2.5 billion IPO 前轮更深、更透明。第二,消费者品牌:StepFun 的 Yuewen 应用在披露 MAU 上落后本章所有具名竞争者;与 Doubao(336-345 million)或即便 DeepSeek 自有消费者应用(约 139 million)相比,差距至少一个数量级。第三,benchmark 不透明:关于中国 LLM 排行榜的独立评论指出,测试集泄漏和刷榜风险(有争议的 C-Eval 结果、未披露方法论的自报分数)会降低外界对任何单一供应商自报排名的信心,StepFun 也包括在内,除非结果得到独立复现。第四,商品化:2026 年 API 价格战由 DeepSeek、Xiaomi 和 Moonshot 发起并维持,不是 StepFun;这意味着 StepFun 必须在资金更厚的对手设定的价格底部里保住利润,而不是掌控这个底部。上述证据并未抹掉 StepFun 的设备楔子,但意味着这个楔子——而非 benchmark 排名、应用规模或资产负债表深度——才是本章找到的唯一耐久相对优势。[CP037, CP040, CP041, CP042, CP043, CP045]

护城河耐久度 / 竞争风险台账
护城河主张竞争威胁严重性证据缓解措施 / 尽调问题
嵌入式多模态设备分发(智能手机、Geely)更大的设备既有玩家(Xiaomi MiMo、ByteDance / Doubao 生态)可以用远大得多的规模捆绑自有模型"仅 Xiaomi MiMo 就占中国供应商 OpenRouter 周份额约 21%;Doubao 的 MAU 远大于 Yuewen"要求 StepFun 提供其 OEM 交易的设备级搭载率和续约数据,并与公开设备出货声明对照
AGI 导向的技术雄心与差异化资金更充足的同行(Zhipu、Moonshot、MiniMax、DeepSeek)在公开基准分数上追平或超过 StepFunBenchLM 2026 年 7 月榜单把 Qwen3.7 Max、GLM-5.2、DeepSeek V4 Pro Max 和 Kimi K2.6 排在 StepFun 已列模型之上要求提供独立复现的基准结果,不要只依赖自报分数
开源 / 自托管开发者好感DeepSeek、Zhipu、Moonshot 都发布了覆盖更广、许可更宽松的开放权重旗舰模型2026 年跟踪中,DeepSeek V4(MIT)和 GLM-5(MIT)的排名都高于 StepFun 的开放发布版本厘清 StepFun 自身开源发布节奏和许可条款,并与这些同行对照
消费者品牌与应用层分发按 MAU 计,Doubao、Qianwen/Quark、DeepSeek、Tencent Yuanbao 均进入中国 Top 10 应用;Yuewen 未进入 Top 50AICPB 2026 年 4 月中国 AI 应用 MAU 排名索取 Yuewen 的实际 MAU/DAU 和留存数据,确认或反驳公开的「未进 Top 50」信号
价格竞争力2026 年 API 价格战(DeepSeek、Xiaomi、Kimi 宣布降价永久化)压缩了包括 StepFun 在内所有中腰部供应商的利润空间Apidog 2026 年 5 月五家实验室定价对比;DeepSeek 自身定价文档确认 StepFun 的实际成交价(而非标价)以及相对价格战底部的单 token 利润
资本获取与国家资金支持Zhipu 和 MiniMax 已上市;Moonshot、DeepSeek 分别以 $20B+、$50B+ 估值融资,后者有国家基金直接参与Zhipu/MiniMax IPO 报道;TrendForce 对 DeepSeek National AI Industry Investment Fund 持股的报道将 StepFun 披露的投资者名单和锁定条款,与这些已上市或有国家资金支持的同行对比
汽车 / OEM 渠道集中度少数设备和汽车合作伙伴撑起了大部分设备叙事;若某笔交易流失或重谈,这部分规模会被大幅削弱Geely 2026 年「Full-Domain AI 2.0」材料强调多供应商,并不断强化自研座舱 AI 的目标,而非独家依赖一家外部模型供应商索取 StepFun OEM 合作的合同排他条款和续约日期

严重度衡量的是 StepFun 竞争耐久性和定价权风险,而不是中国 AI 总需求风险;该风险台账使用上方画像和对比表引用的同一批来源,并纳入证据清单中的反向视角来源。

[CP035, CP037, CP038, CP040, CP041, CP042]
FP003: 护城河 / 就绪度 KPI

一张精简评分卡,概括当前帮助或削弱 StepFun 防御力的竞争特征。

数值是综合本章引用来源形成的定性判断,不是标准化评分指数,也不是管理层提供的竞争自评。

[CP037, CP041, CP042, CP043, CP045, CP046]

3.5 展示材料

Chapter 04

04财务

4.1 营收模式与定价:官方价目表可见,实际经济性仍不可见

StepFun 官方开放平台公布了罕见细的价目表,覆盖文本、视觉和音频模型:step-3.7-flash、step-3.5-flash、step-1o-turbo-vision 按每 1M tokens 定价,实时语音、文本转语音、联网搜索和文件存储也分别计费。标价透明度是真实的,但旁边还有一条明确战略表述:2025 年 7 月,一位 StepFun 副总裁称,公司有意避开高度定制化项目交付,因为这会输给开源替代;也避开纯 token 计价 API,因为大平台可以把模型访问与云销售捆绑后商品化。相反,StepFun 把商业化放在产品化 to-B 和 to-C 收入上,依靠手机、汽车、IoT、具身机器人等设备合作伙伴,以及专门金融合资公司、与 Guotai Haitong 合作等垂直伙伴关系。独立报道还称,企业服务据称覆盖多家 Fortune 500 客户,AgentAPI 调用量显著增长;但两项说法都没有公司披露数据独立佐证,面向国产芯片厂商的芯片适配授权渠道也增加了一个仍未量化的收入面。结果是,StepFun 的定价机制比多数私有 AI 同行更可见,但公开信息仍无法把任何一个收入面桥接到已实现的、分板块的营收。[CI001, CI002, CI003, CI004, CI005, CI006]

收入来源表
收入来源机制计价单位当前价值 / 状态证据质量尽调事项
直接 API 使用(文本、视觉、音频)step-3.x 模型家族按 token 计价、随用随付RMB / 1M tokens官方已公布多模态推理、标准推理、视觉和音频模型的标价标价置信度高,实际收益率置信度低提供模型组合、缓存命中率以及任何企业折扣表。
企业 / 垂直行业交易(金融、内容、零售)借「财跃星辰」金融合资公司,以及 Guotai Haitong 等合作伙伴签订定制合同定制合同金融、内容、零售领域均有深度合作协议报道;未公开合同条款存在性置信度中,经济性置信度低提供 MSA 样本、最低承诺量和合同级定价。
设备 / OEM AI 授权(手机、汽车、IoT)与设备厂商(ZTE/Nubia、Geely 及其他国产手机品牌)达成授权或收入分成安排按设备计费或收入分成报道显示,手机和汽车 OEM 已采用 StepFun,覆盖超过半数领先国产手机品牌;未公开单台费用存在性置信度中,单位经济性置信度低提供单台授权费或收入分成条款,以及销量承诺。
芯片适配 / 生态授权借助 Model-Chip Ecosystem Innovation Alliance,与国产芯片厂商联合适配并授权定制 / 技术授权据报道,Step 3 已授权并适配多家国产芯片厂商存在性置信度中,变现细节置信度低厘清芯片适配工作是否单独收费,还是纯技术合作。
平台增值服务官方平台上的互联网搜索、文件存储及其他附加服务按调用 / 按 GB-day官方价目表列示:搜索调用 RMB0.04/次,文件存储 RMB0.5/GB/day标价置信度高,采用结构置信度低提供增值服务附加率及其收入占比。
面向消费者的应用变现可能存在订阅或应用内变现,并打包进设备体验不明确本章所审阅来源未单独量化要求披露区别于 OEM 授权的任何直接消费者变现。

标价来自官方文档;企业、设备、芯片适配各行概括的是报道中的交易类型,并非已确认的合同经济性。公司概览 讨论设备 / 用户部署规模;本表讨论变现机制,而非用户数量。

[CI001, CI002, CI003, CI004, CI008, CI009]
定价 / 变现表
模型 / SKU价格 / 单位标价 / 实际成交价折扣 / 未知项来源
step-3.7-flash(多模态推理)每 1M tokens:RMB1.35 / RMB0.27 / RMB8.1(输入未命中缓存 / 输入命中缓存 / 输出)公开标价未披露企业折扣或缓存命中率StepFun 开放平台文档
step-3.5-flash(标准推理)每 1M tokens:RMB0.7 / RMB0.14 / RMB2.1公开标价同上StepFun 开放平台文档
step-1o-turbo-vision 模型每 1M tokens:RMB2.5 / RMB0.5 / RMB8(图片默认约 169 tokens)公开标价细节模式 token 数取决于图片尺寸,未单独定价StepFun 开放平台文档
stepaudio-2.5-realtime 模型每 1M tokens:RMB10 / RMB2 / RMB70公开标价未公开批量折扣表StepFun 开放平台文档
step-tts-2 / stepaudio-2.5-tts 模型每 10,000 字符:RMB5.8公开标价未公开企业批量价StepFun 开放平台文档
第三方市场挂牌价(OpenRouter / PricePerToken)StepFun 2026 年 1–5 月发布版本:每 1M tokens 输入 $0.09-$0.20、输出 $0.30-$1.15独立聚合器标价相较 2026 年 1 月版本,2026 年 5 月旗舰版本价格上调而非下调;不同于通常的模型代际价格压缩PricePerToken;OpenRouter
增值服务互联网搜索 RMB0.04/call;文件存储 RMB0.5/GB/day公开标价未披露附加服务附加率及其收入占比StepFun 开放平台文档

除非另有说明,价格均为 StepFun 自有文档中的 RMB 标价;第三方聚合器给出的 USD 数字跟踪的是另一套英文市场价目表,不应与 RMB 各行相加。

[CI001, CI002, CI003, CI004, CI005, CI006]
FI001: 收入模式桥接

StepFun 通过直接 API 用量、企业合同、设备 / OEM 授权和芯片适配合作变现,但从标价走到 2025 年据报约 RMB500M 收入,中间仍隔着未披露的折扣、分成和成本层。

该桥接基于官方定价页和第三方对合作结构的报道定性搭建;StepFun 没有发布分部到收入的勾稽表。

[CI001, CI002, CI003, CI008, CI009, CI010]

4.2 GTM 动作与设备合作:分发真实,但销售效率代理指标缺失

StepFun 的 go-to-market 动作看起来由设备牵引,而不是只由开发者牵引。报道把 ZTE Nubia Z80 Ultra 称为首款在操作系统内原生嵌入 StepFun OpenClaw AI agent 的智能手机,而不是以独立应用形式存在;Tencent Cloud 也与 StepFun 签署战略合作协议,建设「AI Cockpit Agent」,把 Tencent 音乐、视频、地图、支付和出行服务连接到车内使用场景,深化了从 Tencent B 轮投资开始的关系。报道还称,StepFun 近期融资轮集中吸纳消费电子供应链投资方——Huaqin、Longcheer、OmniVision 和 ZTE——而非纯财务投资人,并称截至 2025 年中,StepFun 的智能手机 Agent 合作覆盖中国头部国产手机品牌超过一半。这种集中度有两面:它给 StepFun 带来真实、具名的分发伙伴,也把近期营收增长绑定在少数制造商的出货周期和谈判条款上。本文审阅来源没有披露任何设备或企业交易的获客成本、回本周期、净收入留存或合同级定价,因此本章无法验证这些分发到底高效,还是只是可见。[CI008, CI009, CI010, CI011, CI012, CI013]

FI002: 单位经济桥接

公开证据在标价和合作采用层最强,在 B2B2C 收入结构层较弱;一旦进入 2025 年指引缺口、利润率和资本问题,证据最弱。

这座桥刻意停在公开证据止步处;下游节点仍是待尽调问题,而非已确认指标。

[CI001, CI005, CI009, CI010, CI015, CI016]

4.3 报道口径营收、融资和资本获取:真实信号、落空的指引目标和缺失的审计基线

公司概览已经覆盖 StepFun 逐轮融资历史;本章的问题是,近期报道对当下资本获取和收入质量意味着什么。Caijing 报道 StepFun 2025 年营收约 RMB500 million(约 $73.3 million),2026 年预测接近 RMB1.2 billion(约 $175.9 million)。这一基数明显低于 StepFun 自身领导层在 2025 年 7 月提出的约 RMB1 billion 全年 2025 年营收目标;彼时据称 H1 已签约数亿 RMB 收入,意味着公司没有达到自己的年中指引。资本获取侧,2026 年 2 月一篇报道描述了两批次 Pre-IPO 轮(投前约 $4 billion 和 $5-6 billion),并计划到 2026 年中提交香港上市申请,基石定价接近 $10 billion;这与公司概览已确立的 $10-12 billion 区间和 2026 年 6 月 9 日申请日期一致。另有被引用内部文件指向可能的 2027 年中股份 lock-up。独立审阅的注册资本文件显示,StepFun 在 2026 年两次提高境内注册资本,并于 2026 年 5 月注册香港实体,两者都符合积极准备上市的状态。对比已上市同行,Zhipu AI 披露经审计 2025 年营收 RMB724 million、毛利率 41%(仍净亏损 RMB4.72 billion),MiniMax 披露 2025 年营收约 RMB543 million;二者都比 StepFun 仍未经审计的数字更大、披露更多。[CI015, CI016, CI017, CI018, CI019, CI020]

单位经济模型表
指标数值 / 空缺置信度重要性尽调事项
报道的 2025 年收入约 RMB500 million(约 $73.3 million),第三方报道为约 $10 billion 基石估值目标提供收入基准提供经审计的 2025 年财务报表或已提交招股书数据。
报道的 2026 年收入目标 / 预测约 RMB1.2 billion(约 $175.9 million)检验收入能否在一年内合理翻倍以上提供 2026 年中期管理账或 2026 年 H1 实际数据。
2025 年中全年收入指引2025 年 7 月提出约 RMB1 billion 全年目标约 RMB500 million 的实际收入显示原指引未达成,属于执行风险信号调和两组数字及其底层收入定义。
毛利率StepFun 未公开披露决定设备 / 企业收入结构能否在算力成本下支撑盈利提供分部毛利率,并与 Zhipu 41%、MiniMax 25.4% 对比。
账上现金 / 烧钱速度 / 跑道未公开披露资本充足性测算的核心输入,不能只靠收入推断提供最新资产负债表和月度现金流量表。
设备与企业交易的 CAC / 回本周期 / NRR未公开披露判断设备合作到底是盈利分发,还是补贴获客提供最大设备 / 企业合作伙伴的合同级经济性。
设备合作伙伴收入分成行业惯例据称为 token / 收入量的 30%-60%,非公司专属数据若 OEM 合作伙伴分走大比例收入,StepFun 自身留存会被明显压缩提供 StepFun 与具名合作伙伴的收入分成条款。

标注为第三方报道的数值只是方向性信号,并非经审计业绩的独立确认。

[CI015, CI016, CI020, CI021, CI022, CI034]
资本充足性表
指标公开数值 / 状态置信度重要性尽调事项
账上现金未披露缺少该项就无法测算跑道提供最新现金和短期投资余额。
月度烧钱速度未披露;公开信息只有收入估计决定 IPO 或下一轮融资的紧迫性提供月度现金消耗以及研发 / 算力成本拆分。
跑道未披露缺少跑道数据,资本依赖程度就无法量化提供基准和下行情形下的跑道测算。
注册资本备案(境内注资代理指标)2026 年 4 月从 RMB23.94 million 增至 RMB56.27 million,至 2026 年 6 月约为 RMB60.52 million确认上市前有新增资本在境内登记,是部分资本获取信号将注册资本变化与报道的 Series B+ 和 Pre-IPO 现金金额调和。
Pre-IPO 融资结构据报道分为两批:一批约 $4.0 billion 投前估值、募资 RMB2-3 billion,约 2026 年 3 月关闭;另一批 $5.0-6.0 billion 投前估值,约 2026 年 4 月关闭显示资本获取分阶段完成,而非一次关闭,有助于判断跑道时点提供最终签署条款和实际收到的总募集资金。
计划中的香港 IPO 募资据公司概览,计划按 $10-12 billion 估值募资约 $500 million;已于 2026 年 6 月 9 日递表若上市按期完成,跑道可能显著延长提供已提交招股书、募资用途章节和更新后的时间表。
债务 / 项目融资 / GPU 抵押义务StepFun 具体情况未披露;行业惯例显示,AI 算力部署中 GPU 资产支持融资很常见表外或资产支持义务可能改写真正杠杆水平提供任何 GPU 融资、云承诺或预留容量合同的细节。

本表将历史 Series B/B+ 融资时间线视为公司概览 已覆盖的背景,重点放在前瞻性资本获取信号;注册资本变化在本章中由独立来源支持。

[CI017, CI018, CI019, CI023, CI024, CI025]
FI003: 财务估计区间

据报数字横跨亿美元级收入基数、约为该基数两倍的 2025 年中指引目标、数十亿美元级 Pre-IPO 与 IPO 估值台阶,以及规模相近的同行收入——几乎全部来自第三方报道,而非审计数据。

除非另有说明,所有数值均为百万美元;本章来源集依赖公开报道,而非独立审阅的 StepFun 财务报表。

[CI015, CI016, CI017, CI018, CI020]

4.4 成本结构与资本强度:国产芯片联盟并未消除算力账单

StepFun 公开成本叙事围绕算力,而不是面向客户的基础设施。2025 年 7 月,StepFun 与约十家国产芯片和 AI 基础设施厂商共同发起 Model-Chip Ecosystem Innovation Alliance,成员包括 Huawei Ascend、MetaX、Biren、Enflame、Iluvatar CoreX、Moore Threads 和 Cambricon,明确目标是降低对海外 GPU 的依赖;StepFun 称其 Step 3 模型在国产芯片上的推理吞吐最高约为 DeepSeek-R1 的 3x,在 Nvidia Hopper 硬件上相对 DeepSeek-R1 也有 70%+ 吞吐提升。但到 2026 年 6 月,Baidu Baike 汇总的行业评论称,该联盟进展仍集中在硬件适配和基础集成测试,距离大规模部署或 Nvidia 平台已具备的广泛开发者生态仍进展有限。关键在于,无论芯片来自哪里,前沿推理都很贵:独立分析师数据把全球 hyperscaler 单季 AI capex 放在 $130 billion 以上,预计 2026 年云服务商合计 capex 超过 $520 billion;单张 H100 GPU 价格为 $25,000-$40,000,B200 为 $30,000-$70,000;以 GPU 为抵押的资产支持融资,也越来越常见地用于在不稀释股权的情况下资助 100-10,000 GPU 部署。对照 36Kr 估计的中国大模型市场——2024 年不足 RMB30 billion,到 2026 年增长至 RMB70 billion+——StepFun 的算力野心与机会规模相称;但公开记录没有量化 StepFun 自身算力支出、芯片组合或由此带来的单位成本节省。[CI026, CI027, CI028, CI029, CI030, CI031]

FI004: 资本强度 / 现金流图谱

StepFun 的收入和合作可见度尚可,但各条商业线的成本、利润率和资本充足性仍不透明;即便国产芯片替代 Nvidia GPU,算力强度仍高。

矩阵标签是基于已审阅来源总结的证据强度序数判断,不是 StepFun 内部指标。

[CI001, CI011, CI013, CI017, CI018, CI023]

4.5 财务结论与尽调阻塞项:真实商业牵引、没有审计基线,以及未解释的指引落空

StepFun 的正向财务案例是可见的:官方定价覆盖五个以上变现界面,具名设备和企业伙伴真实存在,2026 年注册资本和香港实体文件确认公司在积极且资金充足地准备上市。负面案例同样具体。本文审阅来源没有披露 StepFun 的现金余额、月度 burn、runway、毛利率或按产品线的单位经济性,因此资本充足性和利润质量无法独立判断。AsiaICT 的分析给出最尖锐的担忧版本:约 $10 billion 基石估值目标,对应的是低数亿营收基数,意味着收入倍数达到数十倍;且截至 2026 年 5 月 20 日,尽管市场广泛报道 2026 年 6 月上市计划,StepFun 尚未正式递交 HKEX Application Proof,这在公司概览已覆盖的 CSRC 红筹回拆之外,又增加执行时点风险。2025 年中约 RMB1 billion 营收指引与报道口径约 RMB500 million 的 2025 年结果之间的差距,在所有审阅来源中都没有得到解释;负面报道引用的 30%-60% 设备伙伴收入分成惯例,也没有针对 StepFun 具体量化。在经审计报表或正式招股书补齐这些缺口之前,本章结论保持谨慎:StepFun 看起来是一家真实、快速放大的商业公司,但公开证据仍支持继续研究,而不是干净的财务背书。[CI034, CI035, CI036, CI037, CI038, CI039]

公开财务缺口表
缺失的私有指标影响精确尽调路径
经审计财务报表 / HKEX 招股书缺少这些文件,收入、毛利率和亏损数字仍只是第三方估计在 HKEXnews 可获取后索取已提交的 A1/PHIP 招股书,或索取管理层分享的经审计 2025 年年报。
现金、烧钱速度和跑道不能只靠收入报道支撑资本充足性判断索取月度现金流量表,以及基准 / 下行增长情形下的 12–24 个月跑道模型。
分部毛利率与单位经济模型无法判断设备 / 企业收入结构是盈利还是补贴索取分部 P&L,以及头部设备和企业账户的单合作伙伴贡献毛利。
设备 / OEM 收入分成条款行业惯例的 30%-60% 抽成率可能显著改变 StepFun 实际收入索取具名合作伙伴合同条款,或至少披露汇总后的混合抽成率。
算力成本与芯片联盟经济性国产芯片联盟降低了部分 GPU 依赖,但其成本 / 性能收益未单独量化索取 StepFun 按芯片厂商拆分的算力支出,以及相对 Nvidia 推理产生的单位成本节省。
2025 年收入指引与实际结果调和2025 年中约 RMB1 billion 目标与报道约 RMB500 million 实际结果之间尚未调和索取管理层对差距的解释,以及更新后的 2026 年指引节奏。

截至 2026-07-01,上述缺口阻碍 StepFun 在公开信息口径下形成清晰投资论证。

[CI016, CI017, CI028, CI034, CI036, CI037]

4.6 展示材料

Chapter 05

05产品与技术

5.1 产品线:从 Step 3.7 Flash 到开源音频、视频和图像工具

StepFun 的商业入口是一套统一 API 平台(中国区 platform.stepfun.com、全球区 platform.stepfun.ai),按一组模型计量收费,而不是押注单一旗舰产品。当前旗舰 Step 3.7 Flash 是一个 198B 参数的稀疏 MoE 视觉语言模型(196B 语言骨干加 1.8B 视觉编码器),每个 token 大约激活 11B 参数,支持 256K 上下文窗口,输入价格为每百万 token RMB1.35/RMB0.27(缓存未命中/缓存命中),输出为 RMB8.1。上一代 Step 3.5 Flash(196B/11B 激活参数、262K 上下文)仍在提供服务,并在 2026 年 2 月发布后不久排到 OpenRouter 免费模型榜第 3 位。Flash 层级之下是 Step-2,StepFun 在 WAIC 2024 推出的原始万亿参数 MoE 模型,以及 Step-1V/Step-1.5V 多模态产品线;StepFun 称这条线支撑了超过 40 million 个嵌入式智能手机 AI 功能。另一条创作工具线覆盖文生图和图像编辑(step-2x-large、step-image-edit-2,按生成图像收费),还有庞大的音频家族:Step-Audio 2 是开源端到端语音对话 LLM,支持检索增强生成和工具调用;实时、TTS、语音克隆和 ASR 产品层级各自按不同计量单位收费,也可通过 pip 和 GitHub 安装。StepFun 还开源了 30B 参数文生视频模型 Step-Video-T2V,并维护模型之外面向开发者的工具,包括 STEP-GUI GUI-agent 框架(仓库 gelab-zero)和 Step-Realtime-CLI/Console 语音智能体工具。消费端没有单独承载这些模型的品牌;模型藏在阶跃AI(跃问 / Yuewen)聊天应用和嵌入式设备助手背后,因此 StepFun 的消费分发本质上完全是推理后端关系,而不是独立产品。[CE001, CE002, CE004, CE005, CE006, CE007]

产品模块 / 资产矩阵
模块 / 产品用户 / 买家状态 / 成熟度差异化尽调缺口
Step 3.7 Flash(198B/11B 激活参数 MoE VLM)智能体 / 代码开发者GA(2026 年 5 月 29 日)以 Flash 级成本提供原生视觉 + 智能体式工具编排主要基准大多由厂商自述,仍待独立审计
Step 3.5 Flash(196B/11B 激活参数 MoE)开发者、OpenRouter 用户GA,仍在服务(2026 年 2 月)发布时 OpenRouter 免费模型排名第 3相对 Step 3.7 Flash 的退役 / 支持时间线未披露
Step-2(1T 参数 MoE LLM)企业 / 消费者聊天后端2024 年 7 月 GA,未确认继任版本公告中国首个万亿参数 MoE LLM尚不清楚当前产品背后是否仍在生产中主动使用
Step-1V / Step-1.5V(多模态)消费者应用(Yuewen)、OEM 设备GA支撑 40M+ 嵌入式智能手机 AI 功能Yuewen 当前背后的模型版本未披露
step-2x-large / step-image-edit-2(图像生成 / 编辑)开发者、消费者创意使用GA;免费层于 2026 年 6 月 12 日结束按图片计量的生成和编辑未找到独立图像质量基准
Step-Audio 2 / Step-Audio R1.5 家族(语音 LLM)构建语音智能体的开发者GA,已开源端到端 ASR + 语音对话 + RAG + 工具调用语音克隆滥用防护尚无独立文档
Step-Video-T2V / TI2V(文本 / 图像转视频)开发者、内容创作者2025 年 2 月开源,未确认 2026 年更新30B 参数 Video-VAE + DiT 文本转视频是否跟上 2026 年竞品视频模型尚不清楚(超出本章范围)
STEP-GUI(gelab-zero)智能体框架开发者开源,仓库活跃GUI 智能体研究框架,2.2k GitHub 星标生产就绪度相对研究演示不清楚
Step-Realtime-CLI / Console构建语音 UI 的开发者开源,提交活跃(2026 年 6 月)实时语音 API 的参考工具采用指标未披露

状态 / 成熟度和星标数反映访问日期(2026-07-01);图像 / 视频模型参数量以官方技术报告披露为准。

[CE001, CE002, CE004, CE005, CE006, CE007]

5.2 架构:稀疏 MoE、注意力解耦和国产芯片切口

Step 3.7 Flash 的成本效率靠细粒度 MoE 路由撑住:围绕该版本的独立技术分析称,45 个语言层中有 42 个是路由 MoE 层,每层有 288 个专家,学习型路由器每个 token 只选前 8 个——也就是任一 token 只动用约 5.6% 的权重。同一家族中另一个更大的模型 Step-3(321B 总参数 / 38B 激活参数)进一步押注硬件感知效率:其协同设计论文引入 Multi-Matrix Factorization Attention(MFA)来压缩 KV-cache,并用 Attention-FFN Disaggregation(AFD)把注意力和 FFN 计算拆到专门子系统;论文声称在 Hopper 硬件上解码吞吐达到 4,039 tokens/second/GPU,同一测试设置下 DeepSeek-V3 为 2,324——这是 StepFun 自行运行、且对 StepFun 有利的比较,不是独立基准。StepFun 也试图直接降低算力供应风险:2025 年 7 月,它与约 10 家国产芯片和平台厂商共同发起“模芯生态创新联盟”(包括 Huawei Ascend、MetaX、Biren、Enflame、Iluvatar CoreX、Moore Threads、Cambricon 等),并把 Step 3 的 MFA 架构适配到包括 MetaX GPU 在内的国产加速器上。但截至 2026 年 6 月,行业评论仍把联盟进展描述为集中在硬件适配和基础集成测试,而不是大规模生产部署。产品效率一侧,Step 3.7 Flash 的 Advisor Mode 把智能体轨迹大部分交给廉价执行器,只在规划或失败恢复拐点升级到更大模型;这是 StepFun 自己落地的一种升级模式,StepFun 将其归因于 Anthropic。其据称“涌现”的跨工具组合行为(在视觉搜索循环中独立调用 Python 裁剪 / 缩放工具)也只是厂商观察,StepFun 自身尚未交给独立复现。[CE002, CE013, CE014, CE015, CE044, CE042]

技术 / 运营架构表
层级 / 组件作用依赖风险
视觉语言骨干(Step 3.7 Flash:196B LM + 1.8B ViT)原生多模态感知与推理专有训练数据与算力基准测试数据大多来自厂商表述
稀疏 MoE 路由(288 个专家取 top-8,198B 中约 11B 激活)在大标称规模下压低推理成本稳定的专家路由训练工具组合涌现行为是厂商观察,尚未被独立复现
MFA + AFD 架构(独立 Step-3,321B/38B 激活)降低 KV-cache 规模与解码成本需要 Hopper 级 GPU 部署才能达到披露吞吐StepFun 自己运行并表述头对头对比,场景也偏有利
国产芯片协同设计(Model-Chip Ecosystem Innovation Alliance)降低对受出口管制 Nvidia GPU 的依赖依赖约 10 家国产芯片厂商的硬件成熟度2026 年 6 月评论:仍处于硬件适配阶段,尚未全规模部署
Agent 运行框架兼容层(Claude Code、KiloCode、Hermes Agent、OpenClaw、MCP)降低接入第三方 agent 框架的成本StepFun 不控制外部运行框架生态OpenClaw 平台另被提示存在夹带木马的假冒技能包
开源权重分发(GitHub、Hugging Face、ModelScope;约 32 个公开仓库)以 Apache 2.0 作为开发者信任与采用渠道社区维护入口为 opensource@stepfun.com迭代很快——部分发布(如 Audio R1.5)已在数月内替代前一版本
API 计量与平台基础设施(platform.stepfun.com/.ai,V0-V5 档)承载商业交付和基于公平性的限流账户累计充值决定吞吐门槛(TPM/RPM)未披露可用性 SLA;只有低声誉第三方监控在跟踪状态

风险列反映独立来源或 StepFun 自身披露中识别出的缺口和注意事项,不是正式风险评级。

[CE002, CE013, CE014, CE015, CE044, CE017]
FE001: StepFun 产品架构栈

从消费者 / 开发者触点,经模型家族下沉到服务基础设施和 API 平台的分层栈。

[CE001, CE002, CE016, CE029, CE030]
FE003: StepFun 关键依赖图

支撑 StepFun 模型训练、服务、分发和监管风险敞口的依赖。

[CE015, CE024, CE025, CE027]

5.3 客户工作流、部署足迹和发布路线图

StepFun 最具体的效率卖点是智能体编程经济性:Step 3.7 Flash 的 Advisor Mode 据称能以约 $0.19 跑完一个智能体任务,而更大的 顾问模型约需 $1.76;StepFun 自己也把这个说法框定为内部脚手架基准,而不是独立审计比较。编程之外,模型线覆盖企业文档 / 截图解析(原生多模态感知)、网页 / 视觉搜索增强问答、基于 Step-Audio 2 的端到端对话式语音智能体,以及基于 STEP-GUI 框架的 GUI 驱动桌面自动化。纸面上部署确实灵活:Step 3.7 Flash 支持云端、数据中心和本地使用,已集成 vLLM、SGLang、Hugging Face Transformers、llama.cpp、NVIDIA NeMo 生态和 NVIDIA NIM,并在发布时登陆 StepFun 自有平台、OpenRouter 和 NVIDIA NIM(DeepInfra、Fireworks AI 和 Modal 宣布随后接入)。标准 API 访问采用五级限速阶梯(V0 到 V5),门槛看账户累计充值,而不是固定开发者套餐;另有较新的订阅制 “Step Plan”。真正本地部署仍有硬件底线:即便已发布的最小量化版本,也至少需要 128GB 统一内存,普通消费级笔记本因此出局,本地运行被限制在 DGX Spark、Mac Studio M4 Max 或 AMD Ryzen AI Max+ 395 这类工作站级硬件上。StepFun 在 2025-2026 年的发布节奏——2025 年 7 月 Step-Audio 2 和 Step-3 技术论文、同年 8-9 月开源 Step-Audio 2 mini 变体和 vLLM 后端、2026 年 2 月 Step 3.5 Flash、2026 年 5 月 Step 3.7 Flash——显示它大约每 1 到 3 个月更新一次主要模型家族;Step-Audio 2 也以标准 pip 可安装 PyPI 包的形式与 GitHub 源码一并分发。[CE003, CE016, CE029, CE030, CE033, CE036]

工作流 / 使用场景表
用户任务当前工作流StepFun 方案可衡量收益限制
自主交付代码变更 / 测试人工开发者或昂贵前沿模型智能体循环Step 3.7 Flash + Advisor Mode 升级~$0.19/任务,对比更大规模顾问模型的 ~$1.76/任务(厂商说法,约 9.3x)成本倍数是 StepFun 自有评测框架估算,未经独立审计
解析密集财报、UI 截图和文档人工抽取或 GPT-4o 级视觉 APIStep 3.7 Flash 原生多模态解析单次摄取混合视觉 / 文本输入SimpleVQA/V* 分数由厂商自述,待独立复现
回答需要新鲜、多来源证据的问题级联搜索 + 摘要流水线Step 3.7 Flash 的网页与视觉搜索增强79.2 SimpleVQA(Search)、75.82% BrowseComp(厂商)对比模型来自不同评测框架,不能严格同口径比较
带工具使用的自然双向语音对话级联 ASR + LLM + TTS 流水线Step-Audio 2 端到端音频 LLM,带 RAG + 工具调用厂商宣称在多个开放 ASR / 副语言学基准上达到 SOTA对比完全由 StepFun 撰写,并非独立评估
中国境内通用助手 / 搜索Doubao、ERNIE、DeepSeek 消费者应用跃问(Yuewen)应用 + 嵌入式智能手机助手40M+ 嵌入式设备(据公司概览)按 MAU 计,Yuewen 本身未进入中国前 50 大应用(见竞争对手章节)
在自有硬件上私有运行 agent仅云端可用的前沿 API面向 llama.cpp/vLLM/SGLang 的开源权重 GGUF 量化版一项独立 DGX Spark 测试中工具调用成功率 100%需要 >=128GB 统一内存;本地吞吐约 27 tok/s,对比云端约 400 tok/s
借助屏幕理解操作桌面软件手写 UI 脚本或 RPA 工具STEP-GUI(gelab-zero)框架已开源,GitHub 2.2k 星标定位为研究级 GUI agent 方案;生产可靠性尚未由独立方验证

除非标注为独立来源,收益数据均来自厂商表述;限制项标出独立佐证缺失或只覆盖部分证据的地方。

[CE003, CE009, CE016, CE022, CE035, CE036]
路线图 / 发布 / 开发阶段表
日期 / 阶段功能 / 里程碑状态含义来源
2024 年 7 月(WAIC 2024)Step-2 正式发布(万亿参数 MoE)、Step-1.5V、Step-1X正式可用确立中国首个万亿参数 MoE,并铺开多模态和图像生成产品线Entireweb 公司摘要
2025 年 2 月Step-Video-T2V(30B)+ Step-Video-TI2V 开源开源并正式可用确立多模态开源发布节奏arXiv 2502.10248 与 GitHub Step-Video-T2V
2025 年 7 月Step-Audio 2 技术报告;Step-3(321B/38B)协同设计论文;Model-Chip Ecosystem Innovation Alliance 成立技术论文发布 / 联盟成立确立面向硬件的高效解码架构和国产芯片抱负arXiv 2507.16632、arXiv 2507.19427 与 Baidu Baike
2025 年 8–9 月Step-Audio 2 mini / mini Base / mini Think 开源;vLLM 后端发布正式可用扩大音频开源家族,并补上推理引擎支持GitHub Step-Audio2
2026 年 2 月 2 日Step 3.5 Flash 发布(196B/11B 激活,262K context)正式可用发布后不久按使用量成为 OpenRouter 免费模型 #3Hugging Face Step-3.5-Flash; ComputerTech 评测
2026 年 5 月 29 日Step 3.7 Flash 发布(198B,原生视觉,256K context,Apache 2.0)正式可用定位为 Step 3.5 Flash 的生产级 agent 式视觉后继版本StepFun 官方博客;Hugging Face Step-3.7-Flash
2026 年 6 月 12 日step-2x-large 图像模型结束限时免费档(转为 RMB0.1/image)定价调整从获客免费额度转向计量变现StepFun 平台定价文档
持续中(据 2026 年 6 月评论)Model-Chip Ecosystem Innovation Alliance 国产芯片适配硬件适配 / 测试阶段算力多元化缓释仍未到生产规模Baidu Baike

日期取每个里程碑已找到的最早公开披露;来源为上方 sources[] 列表中的出版方简称。

[CE006, CE007, CE011, CE009, CE014, CE015]
FE002: Step 3.7 Flash 的智能体任务流程

智能体任务如何经过推理层级选择、多模态感知、工具编排,以及可选的 Advisor Mode 升级。

[CE003, CE042, CE018]

5.4 基准可信度和竞争定位

StepFun 的发布材料罕见地明说,部分头部对比来自厂商口径而非独立复现:其 Step 3.7 Flash 基准图注称,某竞争对手的 Terminal-Bench 2.1 分数来自 StepFun 自己的内部测试,其他厂商数字则是各自报告的另一个版本(Terminal-Bench 2.0)结果;它还主打更高但被标记为污染的 76.5% SWE-Bench Verified 分数,尽管按独立分析,更抗污染的 SWE-Bench Pro 分数(56.3%,在 35 个追踪模型中排第 15)才是更干净的信号。基于 NVIDIA DGX Spark 的独立上手测试确实给出一个正面结果:在多步骤评估套件中工具调用成功率达到 100%,与模型 67.1 的 ClawEval-1.1 第一名成绩一致(但不能证明该成绩)。不过,更广泛、中立的聚合器 Crafiq 给 Step 3.7 Flash 的总体智能排名只有 #108、编程排名 #89——这能修正 StepFun 自有材料强调的狭窄智能体基准叙事。StepFun 还报告了强劲的搜索增强分数(79.2 SimpleVQA-with-tool、75.82% BrowseComp、92.82% DeepSearchQA)和企业职业覆盖(44 个职业的 45.8% GDPval、Tau2-bench Telecom 各层级 98%+ 通过率),并把自家 Step-SWE-Bench 的跨评测性能方差从 43-73%(Step 3.5 Flash)收窄到 64.5-71.5%——这些仍是自测。相比之下,Step 3.5 Flash 在 2026 年 2 月亮相时获得的独立评测覆盖,确实在 Hacker News 和本地推理社区激起讨论;StepFun 的 GitHub 组织也提供佐证,约有 1,700 名关注者和 32 个公开仓库,包括 2,100+ star 的 Step-3.5-Flash 仓库,以及增长很快的 Step-3.7-Flash 仓库。[CE017, CE018, CE019, CE020, CE021, CE022]

FE004: 产品成熟度 / 能力图谱

StepFun 主要模型线的成熟度和独立验证覆盖。

[CE021, CE027, CE028, CE032, CE040]

5.5 信任、安全、安保和合规

StepFun 的开放权重版本(Step 3.7 Flash、Step 3.5 Flash、Step-Audio 2、Step-Video-T2V、Step1X-Edit)一贯采用宽松的 Apache 2.0 许可证分发;相较完全闭源的竞争对手,这是真实的信任和透明度加分项。但与之相对,本章审阅的公开记录仍有多项信任缺口未被说明。中国《人工智能生成合成内容标识办法》自 2025 年 3 月发布起生效,要求对 AI 生成的音频、图像和视频内容做显式与技术性(水印 / 元数据)标识——这项义务正好覆盖 Step-Audio 的语音克隆 / TTS 输出,以及 step-2x-large/step-image-edit-2 图像模型;但已审阅来源没有描述 StepFun 专属的标识或水印实现。另一个问题是,中国主管机构(CNCERT/CC、工信部和国家计算机病毒应急处理中心)已警示 OpenClaw 智能体平台上存在夹带木马的仿冒技能包——StepFun 自己的发布材料也把同一生态列为 Step 3.5 和 Step 3.7 Flash 支持且排名靠前的集成——但没有披露 StepFun 专属的技能验证或沙箱控制。platform.stepfun.com API 服务未披露独立 SOC 2、ISO 27001 或同类安全认证;唯一可得的可靠性信号是一个声誉较低的众包可用性监测器(截至 2026 年 7 月 1 日运行正常,30 天内有 26 条问题报告),而不是 StepFun 发布的 SLA 或事故复盘。最后,Step-Audio 2 声称相对其他开源和商业音频模型达到最先进基准表现,但整个比较由 StepFun 团队撰写,并非独立评估者。[CE024, CE025, CE026, CE027, CE028, CE040]

信任 / 质量 / 合规表
控制项 / 认证 / 指标状态范围缺口
Apache 2.0 开源权重许可已由 GitHub/Hugging Face 发布页确认Step 3.7/3.5 Flash、Step-Audio 2、Step-Video-T2V、Step1X-Edit 的模型权重与推理代码未找到独立的许可证合规审计
中国《人工智能生成合成内容标识办法》(2025 年生效)全行业有约束力的监管义务适用于 Step-Audio TTS / 语音克隆,以及 Step-1X/step-2x-large 图像生成输出StepFun 自身的标识 / 水印实现没有独立文档佐证
OpenClaw agent 平台安全警示(木马技能包)仍在监管关注中(CNCERT/CC、MIIT、国家计算机病毒应急处理中心)Step 3.5/3.7 Flash 使用量排名靠前的共享生态未披露 StepFun 专属的技能验证或沙箱控制
SWE-Bench Pro vs. Verified 的披露方式StepFun 与独立跟踪方都并列发布两项分数编码 agent 基准测试声称多个头部指标(Terminal-Bench 2.1、Advisor-Mode 成本)仍由厂商表述,尚未进入公开独立榜单
平台可用性 / 事故披露StepFun 未发布正式 SLA 或状态页platform.stepfun.com API 交付唯一独立证据是低声誉众包监控;没有经验证的历史可用率 % 或事故复盘
独立安全认证(SOC 2 / ISO 27001)已审阅公开来源中未找到开发者账户与 API 使用数据处理未找到认证、审计报告或信任中心页面

状态反映截至 2026 年 7 月 1 日访问日的公开披露;公开来源未出现某项控制,并不必然意味着内部没有该控制。

[CE024, CE025, CE027, CE028, CE026, CE019]

5.6 证据要点

Chapter 06

06客户情况

6.1 消费应用和开发者生态:触达广,但独立测量弱

StepFun 最直接的客户关系位于分发栈两端:免费使用的 Yuewen(跃问)消费应用,以及围绕 Hugging Face 和 GitHub 搭建的开放权重开发者生态。两者单独看,都没有显示出强独立测量牵引力。Yuewen 的 Google Play 页面截至 2026 年 5 月更新时只有 58 条评论、评分 4.2,多名具名评论者抱怨应用拦住 Google、邮箱或社交登录,只接受手机号——这种摩擦足够严重,以至于有评论者称不提供手机号就完全无法聊天或复制文本。另一个独立信号是,AICPB 2026 年 4 月中国 AI 应用月活榜列出二十多款应用,从 ByteDance 的 Doubao 336 million MAU 一直到最小上榜应用低于 1 million,Yuewen 没有出现在榜单任何位置。相比之下,开发者侧更健康:stepfun-ai 的 Hugging Face 组织下,个别开放权重模型变体下载量达到几十万;stepfun-ai GitHub 组织维护活跃仓库,包括一个 vLLM fork,且提交活动延续到 2026 年 6 月。Step 3.7 Flash 也是 OpenRouter 截至 2026 年 6 月 4 日的十天窗口内列出的五个主要新模型之一,与 Anthropic 等前沿发布并列,说明该模型在第三方市场上仍具商业时效性。但这些开发者信号都无法区分付费 API 账户和免费试用,本章审阅的来源也没有披露开发者到付费的转化率。[CU001, CU002, CU020, CU003, CU004, CU033]

客户分层
客群买方 / 用户 / 付款方主要用例规模或战略价值关键证据缺口
消费者(Yuewen 应用)个人终端用户(免费;付费模式未披露)聊天、文档解析、StepClaw 任务 agent、创意生成Google Play 58 条评论,评分 4.2;未进入 AICPB 中国 AI 应用榜前列未披露 MAU/DAU、收入或订阅用户数
开发者 / 开源个人和组织开发者(免费 API 档 + 付费档)微调、自托管开源权重模型、agent 运行框架集成(OpenClaw、Claude Code、KiloCode)每个模型变体在 Hugging Face 下载量达数十万级;GitHub 组织活跃未披露付费 API 账户数或开发者转化率
OEM 智能手机(OPPO、Honor、ZTE)设备品牌授权 / 集成模型;终端消费者在手机 UI 内免费使用端侧多模态功能(屏幕捕捉与总结、记忆助手、养老语音指令)覆盖 60%+ 国内头部手机品牌;累计设备安装基数 42M+未披露单个 OEM 收入、版税率或续约期限
汽车(Geely / Qianli Technology)车企集团授权座舱软件;购车用户在车内免费使用AgentOS 智能座舱语音 agent,编排导航 / 娱乐 / 车控Galaxy M9 三个月售出约 40,000 辆;2026 年目标 1M+ 辆未披露按车计费的版税或与 Qianli/Geely 的合同期限
企业垂直(媒体、金融 / 新闻)企业伙伴(Huanxi Media、Shanghai United Media Group)共同出资设立合资公司或授权AI 原生电影 / 电视 / 游戏内容;金融新闻垂直搜索一份已披露合资公司框架协议(Huanxi Media,2025 年 11 月);一个已具名媒体伙伴(SUMG)未披露 Huanxi Media 合资公司的收入、时间表或完成状态
机器人 / 具身 AI生态伙伴(Zhiyuan Robotics)集成多模态模型具身智能感知与交互StepFun 2025 年 2 月 Step Up 生态日具名除合作公告外,没有产品、出货或收入细节

买方 / 用户 / 付款方划分与规模数据来自本章审阅的公司及第三方披露;所有行均未独立披露细分收入或账户数。

[CU001, CU003, CU004, CU021, CU026, CU009]
FU001: StepFun 客户旅程图:细分客群、采用触点与扩张循环

展示 StepFun 如何通过五类客户触点分发:从免费开放权重 / API 访问,到设备嵌入和垂直行业生产部署。

旅程阶段是根据公开公告定性重建;StepFun 未披露阶段之间的转化漏斗。

[CU021, CU024, CU025, CU026, CU007, CU028]

6.2 设备、汽车和企业客户证明:具名部署真实,单账户细节薄

StepFun 最强的客户证据来自设备。中国独立科技媒体直接把具名、已出货功能与 StepFun 绑定:《科创板日报》报道称 OPPO Find X8 Ultra 旗舰机的“一键记忆捕捉”功能运行在 StepFun 多模态技术上;Honor 因 StepFun 支撑 Magic8 系列 YOYO 助手记忆功能而授予其“最佳协作奖”;ZTE 在 2025 年 7 月与 StepFun 联合推出养老语音助手,其 Nubia Z80 Ultra 手机也成为首台原生嵌入 OpenClaw 智能体的设备,而 StepFun 发布材料将 OpenClaw 列为受支持集成。新浪科技对 StepFun 2025 年末 GELab-Zero 开源发布的报道给出了触达数字:覆盖中国国产手机品牌超过 60%。汽车侧,Geely Auto Group 和 Qianli Technology 在 WAIC 2025 与 StepFun 联合展示 AgentOS 智能座舱语音智能体,首款车型 Geely Galaxy M9 前三个月销量约 40,000 台,StepFun 目标是 2026 年覆盖超过 1 million 辆车。Tencent Cloud 也另行宣布与 StepFun 建立战略合作,基于其模型打造 AI 座舱智能体;Huanxi Media Group 签署了一份在香港备案的框架协议——由 HKEXnews 监管文件和 Wilson Sonsini 律所共同确认——计划成立合资公司,把 StepFun 模型用于电影、电视和游戏内容。这些关系共同缺少的是单账户商业细节:任何一个 OEM、汽车或企业伙伴都没有披露单台授权费、合同期限或收入分成,因此这些证据证明的是功能级生产使用,而不是已验证的商业规模。[CU024, CU025, CU027, CU012, CU026, CU007]

已具名客户证据表
客户 / 伙伴客群部署 / 用例量产 vs. 试点结果 / 证据局限
OPPOOEM 智能手机Find X8 Ultra“一键记忆截取”“一键全局搜索”“一键屏幕问答”量产(旗舰功能已出货)具名、功能级媒体确认(科创板日报,经 NetEase)未披露 OPPO 关系的出货量、收入或续约期限
HonorOEM 智能手机Magic8 系列“YOYO”助手多模态记忆功能量产(旗舰功能已出货)Honor 2025 年 10 月开发者大会授予的“最佳合作奖”奖项是 Honor 授予的认可,不是独立审计的使用指标
ZTEOEM 智能手机养老“老年 AI 助手”(语音驱动医院挂号、社保缴费);Nubia Z80 Ultra 原生 OpenClaw agent量产(分别为 2025 年 7 月出货与 Nubia Z80 Ultra)Sina Technology 联合发布报道;独立 Nubia Z80 Ultra 评测确认集成 OpenClaw独立评测提示嵌入式 agent 仍有未解决的安全担忧
Geely Auto Group / Qianli Technology(汽车渠道)汽车座舱Geely Galaxy M9 及其他 Galaxy 车型中的 AgentOS 端到端语音座舱量产(量产车)WAIC 2025 联合展示;M9 三个月售出约 40,000 辆交付主体 Qianli Technology 披露,对 Geely 自身的收入集中度为 30-42%
Tencent / Tencent Cloud汽车座舱 / 云基础设施联合开发 AI 座舱 agent,集成 Tencent 生态服务(地图、支付、内容)已宣布合作;产品尚未在市场上获得独立评测AAStocks 与 Edgen 报道战略合作和融资关联截至本章,未独立确认已出货产品、用户数或收入
Huanxi Media Group企业垂直(媒体 / 娱乐)面向 AI 原生电影、电视、互动娱乐和游戏内容的合资公司已签框架协议;合资公司设立和商业条款尚未敲定HKEXnews 监管文件与 Wilson Sonsini 法律确认未披露合资公司交割日期、收入分成条款或已交付产品
开发者 / 开源社区开发生态自助微调和自托管开源权重 Step 系列模型生产(在线、可下载模型具备真实使用遥测)Hugging Face 下载量与活跃 GitHub 提交历史下载量无法区分生产部署与实验

列举范围限于本章研究中至少找到一项独立或官方公开确认的关系;StepFun 未发布完整客户 / 伙伴名单,因此可能仍有未公布关系(见证据缺口)。

[CU024, CU025, CU012, CU027, CU007, CU014]
采用轨迹指标
指标披露数值截至来源置信度缺失分母
智能手机累计安装基数42 million+ 台设备2026 年 1 月Xinmin Evening News / Sohu未按 OEM 或功能拆分
已合作的国内头部手机品牌占比60%+(OPPO、Honor、ZTE 等)2025 年 12 月Sina Technology(GELab-Zero 发布)未列出全部合作品牌名单
估算设备日活用户~20 million2026 年 1 月Xinmin Evening News未获独立分析来源佐证
Terminal Agent API 调用量增长连续 3 个季度约 170% QoQ截至 2025 年 Q4Xinmin Evening News(StepFun 来源)未披露绝对调用量基数
Geely Galaxy M9(搭载 AgentOS)销量~40,000 辆开售前 3 个月(截至约 2025 年 11 月)Xinmin Evening News;ECARX 新闻稿(24h 内 40,000 订单)销量不是 StepFun 专属指标;车辆由 Geely 销售,不是 StepFun
2026 年汽车部署目标1,000,000+ 辆车2026 自然年目标Xinmin Evening News(StepFun 来源)前瞻性公司目标,不是已达成数据
Hugging Face 模型下载量(单一变体)数十万级(如一个 Step-3.7-Flash 变体约 309k)截至 2026 年 6 月下旬Hugging Face 组织页面未披露所有模型变体和时间段的合计数据

所有设备 / 用户 / 车辆数据均来自第三方报道或媒体报道中的公司来源,未经独立审计;应视为方向性的采用信号,而不是经验证的 KPI。

[CU021, CU026, CU030, CU031, CU008, CU003]
FU002: 从采用到部署的漏斗

近似呈现从广泛开发者 / API 访问到具名生产客户部署的量级衰减。

数值混合了真正可比的计数(下载量、仓库)和粗略代理计数(品牌、部署),用来展示数量级衰减;这不是 StepFun 披露的转化漏斗。

[CU003, CU004, CU026, CU024, CU025, CU009]
FU003: 客户证明矩阵:证据质量 vs. 生产成熟度

按证据独立可验证程度和走向生产部署的进展,绘制每个客户 / 伙伴关系。

定性评级(高 / 中 / 低)反映本章对公开证据的判断,不是 StepFun 披露的评分卡。

[CU024, CU025, CU012, CU013, CU007, CU014]

6.3 留存和耐久性:任何客户分部都没有披露指标

本章审阅的来源没有披露 StepFun 任一客户分部——设备、汽车、开发者或企业——的净收入留存、总收入留存、续约率或流失率。最接近的代理指标是 StepFun 自身口径称,终端 Agent API 调用量截至 2025 年末已连续三个季度环比增长近 170%;但这个数字没有披露绝对基数,没有独立分析供应商佐证,衡量的是调用量而非客户留存。消费端, Yuewen 应用稀薄且评价混杂的 Google Play 评论基础(58 条评论、4.2 分,反复出现登录摩擦投诉)是唯一公开可见的满意度信号,样本太小,无法外推留存。合同层面的细节同样缺失:没有来源说明任何具名 OEM、汽车或企业关系的合同期限、续约日期或最低用量承诺,因此无法判断 StepFun 的设备和车辆合作是耐久的多年承诺,还是可能短期到期或重新谈判的协议。本章把留存、付费客户数和合同条款可见度列为明确且重大的证据缺口,而不是估算占位数字。[CU030, CU034, CU035, CU036, CU001, CU002]

留存、重复使用与满意度信号
指标披露值客群置信度尽调追问
净收入留存(NRR)所有细分市场n/a(未披露)要求提供设备、汽车和企业账户的队列 NRR/GRR
客户流失 / 续约率所有细分市场n/a(未披露)要求提供最大 OEM 与汽车合同的续约率数据
消费者应用商店评分4.2 / 5(58 条评论)Yuewen 消费者应用(Google Play 英语市场上架页)要求提供 Yuewen 中国应用商店评分数量和 iOS 评分,以补全图景
消费者应用登录摩擦投诉多位具名评论者称只能用手机号登录,阻碍继续使用Yuewen 消费者应用评估登录摩擦是否压低可衡量留存 / 注册
Terminal Agent API 调用量增长(准留存代理)连续 3 个季度约 170% QoQ设备内嵌 Agent API低(公司来源,未给基数)要求提供增长百分比背后的绝对调用量基数和单设备使用率
OEM / 汽车合同续约或最低量条款OPPO、Honor、ZTE 与 Geely/Qianlin/a(未披露)要求提供头部设备和汽车伙伴的合同级续约及最低承诺条款

标为 null 的单元格表示未找到公司或第三方披露;本章将未披露的留存指标视为明确证据缺口,而不是估算值。

[CU001, CU002, CU030, CU035, CU036]

6.4 扩张和集中度风险:宽设备足迹压在窄伙伴集上

StepFun 的扩张故事看上去很宽——手机、汽车、云、媒体、机器人——但底层分发依赖少数关系。三家手机 OEM 集团(OPPO、Honor、ZTE)承载了整个 42 million 台设备、60% 品牌份额的头部叙事;Qianli Technology 作为 StepFun 的汽车座舱载体,自身财务也承压:其 IPO 招股书披露,2022 年至 2025 年中前五大客户贡献收入 41.5%-48.1%,仅 Geely Group 一家就贡献 30.8%-39.7%;Qianli 在最近四个披露期间中有三个期间净亏损。新浪财经 2026 年 6 月报道还补充称,Qianli 控股集团把 2026 年关联交易上限同比提高逾 80%,并批准以约 2 倍溢价收购一家关联且资不抵债的雷达感知公司,却没有设置业绩补偿条款——对于依赖 Qianli 作为 StepFun 主要汽车分发渠道的投资者,这些治理信号值得担忧。另一份独立商业资料指出,StepFun 董事长也控制 Qianli,意味着两家公司处在同一个 Geely 锚定集团内,而不是保持独立交易对手关系。Tencent 进一步强化了这种模式:它同时是 StepFun 多轮股东,又是 AI 座舱智能体的具名商业伙伴,这一结构让外界更难确认汽车合作是否按独立交易定价。企业垂直多元化(Huanxi Media、Shanghai United Media Group、Zhiyuan Robotics)仍停留在框架协议或早期合作阶段,没有完成合资公司或披露收入,因此还不足以对冲设备 / 汽车集中度。[CU037, CU038, CU015, CU016, CU017, CU018]

扩张与集中度风险
扩张驱动集中度 / 依赖风险潜在影响尽调路径
60%+ 国内手机品牌覆盖(OPPO、Honor、ZTE)带来宽广的设备内嵌分发面覆盖面依赖三家 OEM 集团持续支持并续约,而不是更分散的客户基础失去一个主要 OEM 关系,或重谈协议,都可能砍掉已披露设备安装基数的很大一块索取各 OEM 的合同条款、续约日期和收入分成
汽车 AgentOS 计划在 2026 年覆盖 100 万+ 辆目标车辆落地载体是 Qianli Technology;该公司自身 30.8%-39.7% 的收入仅来自 Geely Group,且 2023-2025 年持续录得净亏损Qianli/Geely 的财务或战略压力,可能扰乱 StepFun 依赖其证明汽车赛道能力的座舱渠道索取 StepFun 与 Qianli 相关的直接敞口 / 应收账款,以及 Qianli 表现不及预期时的合同保护
Tencent Cloud 合作把汽车触达延伸进 Tencent 的消费生态Tencent 同时是多轮股权投资方和商业合作伙伴,谈判公平商业条款时可能存在利益冲突关联方关系可能意味着汽车合作条款不能代表独立市场定价询问 Tencent Cloud 商业协议是否独立于 Tencent 的董事席位 / 股权位置谈判或批准
借 Huanxi Media 合资公司和 Shanghai United Media Group 合作拓展企业垂直场景目前只有一个企业合资项目(媒体 / 娱乐)走到已签框架协议阶段;合资实体、资金和交付时间表尚未最终敲定即便已有企业垂直收入,短期内相对设备 / 汽车规模也很可能不重要跟踪合资公司设立、出资和首个商业交付里程碑
开发者生态(Starlight Program、开放权重发布)扩大长尾应用基础没有披露指标区分已承诺 / 付费开发者和免费试用者;项目成效只做定性描述如果不能转化为已披露的付费使用,开发者生态可能比设备 / 汽车更弱,难以拉动扩张索取 Starlight Program 参与者数量、获资助项目数量,以及由此产生的商业部署

风险框架结合了 Qianli Technology 自身 IPO 招股书披露(一个独立且邻近 StepFun 的数据点)和 StepFun 自己的合作伙伴公告;StepFun 本身不发布客户集中度指标。

[CU015, CU016, CU017, CU018, CU022, CU023]

6.5 分发渠道依赖和尽调结论

按证据可被独立验证的强弱给 StepFun 客户渠道排序,层次很清楚。监管文件和第三方平台遥测排在最上方:HKEXnews 文件和 Wilson Sonsini 的法律确认独立佐证了 Huanxi Media 合资事项,Hugging Face/GitHub 下载和提交数据独立佐证了开发者生态活动。设备和汽车证据位于中间:具名功能媒体报道(OPPO、Honor、ZTE)和联合展示或新闻稿材料(Geely/Qianli、Tencent)虽由独立渠道报道,但最终仍回到公司或合作伙伴对某个功能或事件的披露,而不是审计后的使用或收入数据。最弱的是消费应用:Yuewen 缺席 AICPB 独立 MAU 榜单,Google Play 评论基础又薄且投诉密集,两者都指向同一个结论——StepFun 尚未把设备级触达转化为可独立测量的自有消费用户。总体看,本章在设备和汽车功能层面找到了真实、可验证的客户证明,但整个组合没有披露付费客户数、留存指标或合同期限;Qianli/Geely/Tencent 的所有权和投资结构也把汽车渠道集中在一个小型、互相交织的集团里,而不是分散客户基础中。[CU040, CU041, CU042, CU020, CU001]

分销渠道依赖与证据质量
渠道主要证据类型证据独立性剩余缺口
OEM 智能手机(设备内嵌)具名功能媒体报道(科创板日报、Sina Technology)加上一项供应商授予奖项(Honor)独立媒体报道,但最终仍来自公司 / PR 对具体功能的披露没有按 OEM 拆分的独立出货量或收入数据
汽车座舱(Geely/Qianli/AgentOS)WAIC 联合展示(合作伙伴验证)、ECARX/Nasdaq 新闻稿(合作伙伴验证)、Qianli IPO 招股书财务披露(接近申报文件)混合:合作伙伴发布的 PR,加上一份真正独立的财务披露(Qianli 招股书)StepFun 自身来自汽车渠道的收入分成未披露
云 / 超大规模云厂商(Tencent Cloud)与 StepFun 融资轮绑定的战略合作财经新闻报道独立财经媒体,但 Tencent 同时也是 StepFun 投资方尚无已交付产品或独立用户数据
企业垂直场景(Huanxi Media、媒体 / 金融)香港监管文件(HKEXnews)和律所交易确认(Wilson Sonsini)独立性高(监管文件、法律顾问)只有框架协议;合资公司尚未完成,也无收入数据
开发者 / 开源Hugging Face 和 GitHub 平台遥测(下载量、提交活动)独立性高(第三方平台数据)遥测无法区分生产使用与评估 / 测试
消费者应用(Yuewen)Google Play 评论(客户撰写)和 AICPB 独立 MAU 排名(缺席)两者独立性都高,但都显示牵引力弱或未被测量没有 StepFun 披露的 MAU/DAU/收入可与独立信号对照

本表概括每个分销渠道的客户证据可被独立验证的程度;监管文件和第三方平台遥测的权重高于合作伙伴发布的新闻稿和公司自报。

[CU024, CU025, CU026, CU007, CU014, CU015]

6.6 证据要点

Chapter 07

07风险

7.1 监管和上市风险:红筹回拆叠加密集的 2026 合规日程

StepFun 最迫近的结构性风险不是技术,而是监管。Reuters 2026 年 4 月 13 日报道称,StepFun 正在拆除其离岸 Cayman Islands “红筹”注册架构,为计划中的香港 IPO 铺路;此前中国证券监管机构因股东透明度和合规风险,要求部分离岸注册公司拆除此类架构。两名消息人士告诉 Reuters,这一决定反映 StepFun “受到国资重度支持”,包括上海市级和区级政府投资平台;到 2026 年 4 月,公司已完成股份制改造,Qianli Technology 董事长 Yin Qi 正式登记为 StepFun 董事长。自 2023 年 3 月起适用的《试行办法》规定,通过红筹架构间接境外上市,需在提交境外上市申请后三个工作日内向中国证监会(CSRC)备案,并面临 20 个工作日审核窗口;法律评论指出,拆除既有红筹架构在法律上可能复杂且成本高昂,一些公司因此推迟甚至放弃 IPO 计划。叠加在上市架构风险之上的,是密集的合规日程:中国修订后的《网络安全法》已于 2026 年 1 月 1 日生效,分级处罚显著加重(关键信息基础设施违规最高可罚 RMB 10 million 或更多),并明确加入 AI 治理表述;AI 生成内容标识办法和强制性 GB 45438-2025 标准也已自 2025 年 9 月 1 日起要求显式 / 隐式双重标识。机会一侧,HKEX 2026 年 3 月咨询文件提议下调同股不同权上市门槛;如果在 StepFun 申报窗口关闭前获采纳,可能放松(而不是收紧)其通往创始人友好型上市结构的路径。[CR001, CR002, CR003, CR004, CR005, CR006]

监管 / 法律风险登记表
规则 / 许可 / 案件司法辖区状态可能性严重性缓释措施剩余敞口尽调路径
CSRC 对红筹离岸架构回拆的要求中国 / 香港生效中 —— StepFun 据报道已于 2026 年 4 月改制为股份有限公司在递交 HKEX 文件前已完成境内股份制改造IPO 时间线和估值仍暴露在 CSRC 对红筹必要性的进一步审查之下在假设上市日期确定前,确认 CSRC 已结束对 StepFun 架构的审查。
美国 BIS 先进计算出口管制指引,堵上海外子公司漏洞(2026-05-31)美国(域外适用)已生效;追溯敞口至 2023-11StepFun 的国产芯片联盟只能部分对冲,并非已披露的正式合规缓释措施任何 StepFun 关联实体若通过非中国子公司采购 Nvidia/AMD 芯片,都面临算力供应和交易对手筛查风险索取 StepFun 的芯片采购实体结构和出口管制合规筛查。
修订后《网络安全法》分级处罚制度(2026-01-01 生效)中国(域外触达)已生效一般合规计划;未发现 StepFun 专属披露罚款最高可达人民币 1000 万元(造成重大危害时可能更高),并可能暂停服务或吊销许可索取 StepFun 的 CSL 合规状态和关键信息基础设施运营者分类。
美国 Entity List / 五角大楼「1260H」涉军公司名单扩容(2026-06)美国生效中;截至 2026-07-01 未确认 StepFun 被列名未披露;邻近风险来自同行(Alibaba、Baidu、DeepSeek、CXMT)已被加入或正被考虑加入若国家支持的 AI 独角兽越来越成为目标,StepFun 面临声誉和资本市场邻近风险持续监控 Entity List / 1260H 后续更新,看 StepFun 是否被纳入。
CAC 生成式 AI 内容标识办法 + GB 45438-2025 标准(2025-09-01 生效),以及 2026 年 4 月 CAC 关于拟人化交互式 AI 的草案中国已生效,另有一项草案待落地(目标约 2026-07)显式 / 隐式标识在全行业技术上已较成熟,但 StepFun 自身落地和审计状态未经验证工程和审计负担会反复出现;不合规可能导致内容下架或服务暂停索取 StepFun 的 CAC 备案和内容标识审计证据。
HKEX 加权投票权(Chapter 8A)上市框架咨询(2026 年 3 月)香港已提出,尚未采纳N/A —— 提议会放松而非收紧要求WVR 门槛的监管不确定性,可能影响上市时创始人控制权结构跟踪 HKEX 最终规则采纳时间线与 StepFun 申报窗口的相对关系。

各行按严重性从高到低排序;可能性 / 严重性是基于所引监管、法律和新闻来源的序数判断,而非已披露的内部风险模型。覆盖范围不完整:这里只纳入本章保留的 2026 年最重要监管 / 法律事项,不是逐司法辖区的完整清单。

[CR001, CR002, CR003, CR004, CR006, CR007]

7.2 算力和供应链风险:出口管制收紧,国产芯片栈仍落后

StepFun 的第二个结构性风险在上游,也就是支撑其持续训练前沿竞争力模型所需的算力。2026 年 5 月 31 日,美国商务部工业与安全局(BIS)堵上了一个漏洞:此前中国总部企业可通过新加坡、马来西亚和台湾的海外子公司无出口许可证购买先进 Nvidia/AMD 芯片;新规依托自 2023 年 11 月起技术上已生效的“最终母公司”测试,也让此类交易产生追溯合规风险。Tech Times 报道称,Trump 政府官员 2026 年 6 月内部对该漏洞是否源自 2025 年停止执行 Biden 时代 AI Diffusion Rule 后的意外政策空档意见不一。同月,BIS 又把数十家与中国军工复合体相关、从事先进 AI 和高性能芯片开发的中国实体加入 Entity List;同时 CNBC 报道,美国暂缓正式将 DeepSeek、CXMT 及 100 多家其他被标记公司列入黑名单,以避免加剧与北京的紧张关系——五角大楼 2026 年 6 月更新的“1260H 清单”也加入了 Alibaba、Baidu 和 BYD。截至本章运行日期,没有已审阅来源显示 StepFun 直接出现在任一清单上,但两份名单快速扩张,给任何有国资背景的中国 AI 独角兽都释放了邻近风险信号。StepFun 自己的对冲,是 2025 年 7 月与约 10 家国产芯片厂商共同发起的“模芯生态创新联盟”,成员包括 Huawei Ascend、MetaX、Biren、Enflame、Iluvatar CoreX 和 Moore Threads;Huawei 最新 Ascend 950PR 的推理吞吐约为 Nvidia 出口获批 H20 的 2.8 倍,但功耗高约 200W。独立分析认为,国产栈在单芯片效率和 CUDA 生态成熟度上仍落后于 Nvidia Blackwell 级芯片,只能用集群规模聚合来补偿,功耗接近 4 倍。两类风险之下还有前沿 AI 本身的资本强度:2026 年,一个 100 到 10,000 GPU 集群需要 $5 million 到超过 $500 million 前期投入;计入网络、电力、冷却和人员后,硬件本身只占五年总拥有成本约 35%——这是一道 StepFun 必须在远大于自身的竞争对手面前跨过的资本门槛。[CR016, CR017, CR018, CR019, CR020, CR021]

7.3 竞争和定价风险:行业整合、价格战持续、分发差距扩大

StepFun 所处的中国基础模型赛道,KrASIA 和 JPMorgan 都描述为正在快速整合。KrASIA 将 StepFun 与 01.AI、Baichuan AI、MiniMax、Moonshot AI、Zhipu 一并归为中国“AI 虎”,并指出到 2026 年初 Zhipu 和 MiniMax 已在香港上市,其余公司(包括 StepFun)仍为私有且承受融资压力。JPMorgan 2026 年 2 月行业报告认为,可行的中国模型开发商已从 200 多家缩减到少于 10 家,未来赢家取决于部署规模和变现效率,而不是模型发布节奏;该行对 Zhipu 和 MiniMax 发起公开分析师覆盖并给出“Overweight”评级,却没有覆盖 StepFun,这与 StepFun 相对稀薄的公开披露相符。收入差距很实质:Zhipu 2025 年收入达到 RMB 724.3 million(2024 年为 RMB 312.4 million),明显高于 StepFun 先前报道的约 RMB 500 million;ByteDance 一家公司计划 2026 年在 AI 采购上投入超过 RMB 160 billion,Tencent 2025 年资本开支达到 RMB 79.2 billion——这些基础设施规模比 StepFun 已披露规模高一个数量级。行业整合还叠加主动价格战:中国实验室在 2026 年反复下调 API 价格,单 token 收入在算力成本仍高度资本密集时被压缩,StepFun 的单位经济模型同时被需求侧和供给侧挤压;本章审阅的来源没有证据表明 StepFun 相较规模更大、资本更充足的对手已有差异化定价权。[CR028, CR029, CR030, CR031, CR032]

7.4 客户集中和渠道风险:设备足迹压在财务脆弱的汽车伙伴上

StepFun 最清晰的设备相邻证明点——与 OPPO、Honor、ZTE 的智能手机嵌入,以及与 Geely 的汽车座舱 AI——也构成最窄的渠道依赖。智能手机侧,三家 OEM 集团据称承载了 StepFun 全部约 4200 万台设备分发叙事;OPPO 自己的公开合规声明只承诺其 60 多个国家的一般法律和数据保护合规,没有披露任何 StepFun 专属服务级条款。汽车侧,渠道完全通过 Qianli Technology 运行;这家 Geely 关联的汽车 AI 载体也由 StepFun 董事长 Yin Qi 担任董事长。Qianli 2025 年年报显示,对 Geely 关联公司的销售占总收入 29.26%(2022-2025 年区间为 30.8%-39.7%);其股东于 2025 年 12 月 31 日以超过 99% 赞成票批准 2025-2026 年与 Geely 实体的关联交易上限。2025 年 6 月 cninfo 文件显示,Qianli、Geely、Megvii 关联的 Yin Qi 实体和重庆国资基金共同搭建一家智能驾驶合资公司,承诺资本 RMB 1.5 billion,这进一步加深而非分散了关系。关键是,Qianli 自己的 2026 年二季度招股书更新显示,尽管 2025 年收入增长 42.1%,其扣除非经常性损益后的核心净利润已连续三年为负,汽车业务毛利率只有 2.8%,母公司累计未弥补亏损接近 RMB 948 million——也就是说,承载 StepFun 汽车证明的渠道本身正处在真实财务压力下。StepFun 对智能手机集中度的并行对冲,是“模芯生态创新联盟”和具国资关联的资本基础(上海市 / 区基金和 Hong Kong Investment Corporation),但这些都无法降低汽车收入对 Qianli 的单点依赖。[CR033, CR034, CR035, CR036, CR037, CR038]

合作伙伴 / 依赖风险登记表
依赖交易对手角色集中度失效情景严重性缓释措施剩余敞口
汽车座舱渠道Geely Auto Group / Qianli Technology(汽车渠道)唯一已确认具备规模的汽车定点项目(Galaxy M9 搭载 AgentOS)单一 OEM / 单一集成商集中;Qianli 自身 29%-40% 的收入来自 Geely 关联交易失去 Geely/Qianli 关系或重谈协议,将拿掉 StepFun 唯一已验证的汽车收入线除关系本身外未披露其他缓释措施索取单车授权分成条款和合同续约日期。
Qianli Technology 自身财务脆弱性Qianli Technology(与 Geely 关联;亦由 StepFun 董事长 Yin Qi 担任董事长)StepFun 座舱 AI 的汽车落地载体尽管收入增长 42.1%,Qianli 2025 年核心(剔除非经常性)净利润仍连续第三年为负Qianli 财务承压,可能扰乱 StepFun 所依赖的汽车渠道;这是其最清晰的设备邻近验证未披露;未发现 StepFun 获得直接合同保护索取 StepFun 对 Qianli 的应收账款 / 敞口,以及任何最低销量保证。
智能手机 OEM 集团OPPO、Honor、ZTE (Nubia)面向消费者 AI 功能的设备内嵌合作伙伴据报道,三家 OEM 集团承载全部约 4200 万台设备、60% 品牌份额的头条指标失去一个主要 OEM 关系,会拿掉已披露设备触达的很大一块OPPO 发布一般合规 / 数据保护声明;未发现 StepFun 专属 SLA索取各 OEM 的合同条款和续约日期。
国产芯片供应商联盟Huawei Ascend、MetaX、Biren、Enflame、Iluvatar CoreX 与 Moore Threads(Model-Chip Ecosystem Innovation Alliance)算力多元化 / 出口管制对冲联盟覆盖约十家供应商;Huawei 自身 Ascend 950PR 在能效和软件生态成熟度上仍落后 Nvidia Blackwell 级芯片国产芯片如果性能不达标或供应受限,可能拖慢 StepFun 相对不受出口限制竞争对手的训练 / 推理服务与多家供应商真实技术协同,降低单一供应商集中度索取 StepFun 实际训练 / 推理芯片组合和利用率。
国家关联资本上海市 / 区级投资平台、Hong Kong Investment Corporation (HKIC)IPO 基石和战略资本Reuters 将离岸回拆决定直接归因于 StepFun「得到国有资本大力支持」继续依赖国家关联资本,可能约束独立治理,或招致地缘政治 / Entity List 邻近审查未披露正式缓释措施索取完整 cap table,披露国家关联所有权比例。

Qianli Technology 的集中度数据来自其自身披露的年报比例,并由中文二级报道转述,不是 StepFun 专属分部披露;各行按严重性从高到低排序。

[CR033, CR034, CR035, CR036, CR037, CR038]
FR003: 依赖关系图

StepFun 运营受制于一组高度集中的汽车、设备、芯片和资本依赖,所有依赖最终都汇入 HKEX / CSRC 上市闸门。

[CR033, CR041, CR023, CR043, CR001]

7.5 治理和披露风险:交叉任职叠加无审计基线

StepFun 的治理画像有两个相互叠加的缺口。第一,董事长 Yin Qi 同时担任 Qianli Technology 董事长——后者是 StepFun 主要且唯一已证明的汽车分发交易对手,又与 Geely 关联——这意味着两家公司最重要的商业关系是在交叉任职董事之间谈判,而不是按独立交易完成。本章审阅的来源中,没有发现覆盖 StepFun-Qianli 交易的董事会会议纪要、关联交易审批流程或独立董事证明。第二,截至本章运行日期,无法检索到 StepFun 经审计的 2025 年财务报表或已提交的 HKEX 招股书,因此其报道的约 RMB 500 million 2025 年收入和目标 $10-12 billion 估值缺少独立验证基线;在隐含收入倍数如此高的情况下,这个缺口很重要。StepFun 的投资者基础进一步放大透明度问题:Reuters 报道称其资本中包括上海市级和区级投资平台等国资关联资金;2026 年 pre-IPO 轮还另有 Hong Kong Investment Corporation 参与,这是一只由香港政府全资拥有的基金,其披露文件称同时承担财务回报和提升香港竞争力的双重使命。双重使命本身不一定负面,但意味着 StepFun 的所有权和治理决策——包括 IPO 地点选择——部分受公共政策目标塑造,而纯私营资本公司不会背负这一点。StepFun 从天使轮到 pre-IPO 轮的完整股权结构,仍只能通过媒体报道部分拼出,而不是来自一手文件。[CR041, CR042, CR043, CR044]

人员 / 执行风险登记表
角色 / 职能依赖或缺口可能性严重性缓释措施尽调路径
审计财务基准在目标 $10-12B 估值前,公开渠道尚无经审计的 2025 年财务报表或已提交招股书香港上市流程最终会要求经审计的招股书披露招股书或审计财务一旦可得,即索取。
董事长角色(Yin Qi)同时担任 Qianli Technology 董事长,而 Qianli 是 StepFun 主要汽车渠道交易对手——这是董事互锁结构,不是公平交易合作未披露;治理结构本身就是风险索取 StepFun-Qianli 交易的董事会纪要和关联交易审批流程。
创始人 CEO 关键人依赖(Jiang Daxin)公开技术和战略身份集中在一位创始人身上具名技术领导梯队(CTO Zhu Yibo、数据负责人 Jiao Binxing)提供一定冗余索取管理层接班和关键人连续性披露。
Cap table / 所有权透明度天使轮、Series A/B/B+ 和 pre-IPO 轮的完整所有权拆分未完全披露;国家关联和产业投资方(Huaqin、Longcheer、OmniVision、ZTE、HKIC)持股未披露轮次参与方已有部分媒体报道披露索取完整 cap table 和投票权结构。
国家投资方对 IPO 地点 / 结构的影响Reuters 将境内重组决定部分归因于 StepFun「得到国有资本大力支持」,意味着国家关联投资者会影响战略决策未披露索取国家关联投资者持有的治理权利。

严重性 / 可能性是本章基于所引证据作出的序数判断;各行按严重性从高到低排序。

[CR041, CR042, CR043, CR044, CR004, CR014]

7.6 技术和产品风险:基准不透明、安全姿态未验证、标识负担加重

StepFun 的技术风险不太在原始模型质量,而在于无法独立验证的部分。针对 2.8 million 条 LMArena 模型比较记录的行业级分析发现,选择性提交通过挑选样本把排行榜分数最多抬高了 100 分,主要实验室只提交表现最好的私有测试变体,而不是已上线生产模型;这种模式并非 StepFun 特有,但本章审阅的来源没有找到任何经 StepFun 授权的独立复现来验证其旗舰智能体基准主张,因此同样的不透明风险在 StepFun 身上仍未解除。另一个缺口是,没有找到 StepFun 平台、Step-Audio 或 Step-Video 的独立安全认证、漏洞赏金计划或公开事故响应披露;也没有找到点名 StepFun 的公开诉讼、数据泄露披露或集体诉讼记录——这种缺席应解读为证据缺口,而不是确认记录清白。合规上,中国显式 / 隐式双轨 AI 内容标识制度已自 2025 年 9 月起强制执行,Tencent、Douyin、RedNote、Weibo、DeepSeek 等主要平台都已发布合规声明,但没有找到 StepFun 针对 Yuewen 消费应用或 API 输出的同等公开声明。合规面还在扩张:国家网信办 2026 年 4 月发布征求意见稿,拟对模拟人类交互的 AI 服务增加强制伦理审查、应急程序和内容治理义务,预计约在 2026 年 7 月生效;这会在 StepFun 消费端已经必须满足的标识制度之外,再增加持续工程和审计负担。[CR045, CR046, CR047, CR048, CR049, CR050]

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
没有独立安全认证或公开事故披露记录未披露;没有负面报道不能证明记录清白无法排除嵌入消费设备和车辆的 Step-Audio/Step-Video 管线存在未披露漏洞索取 SOC2/ISO27001 等同认证或事故响应披露。
基准不透明 / 中国与全球实验室的榜单博弈独立榜单审计(如 LMArena)正在出现,但未发现针对 StepFun 的独立基准审计投资者无法验证 StepFun 的 agentic-benchmark 主张是否来自生产模型,而不是优化过的展示版本索取对 StepFun 已交付模型(非展示模型)的独立复现实测基准。
StepFun 自有触点的 AI 内容标识落地和审计状态未经验证全行业已有成熟的双轨标识制度(Tencent、Douyin、DeepSeek 已发布合规声明);未发现 StepFun 等同声明如果 StepFun 的 Yuewen 应用或 API 输出被发现未标识,将带来持续合规和声誉风险索取 StepFun 的 CAC 备案 / 标识审计证据。
拟人化 / 交互式 AI 服务的 CAC 草案规则继续扩围(2026 年 4 月)仍处草案阶段,预计约 2026 年 7 月生效Yuewen 等消费者聊天 / 语音触点将增加伦理审查和应急程序义务跟踪最终规则文本和 StepFun 的合规准备窗口。
未发现 StepFun 的公开诉讼、数据泄露或法律案件记录N/A(没有负面记录,不等于确认记录清白)限制了将 StepFun 运营风险历史与已披露事故同行对标的能力将其视为证据缺口,而非健康证明;每次刷新时重新检查。

严重性 / 可能性是本章作者基于所引证据作出的序数判断,不是 StepFun 已披露的风险评分框架;各行按严重性从高到低排序。

[CR045, CR046, CR047, CR048, CR049, CR050]

7.7 综合:缓释成熟度、可监控触发器和投资逻辑失效点

合起来看,StepFun 最严重的风险集中在三处:CSRC 红筹回拆和密集 2026 合规日程带来的上市执行风险;美国出口管制执行收紧,而国产芯片替代在效率和软件成熟度上仍落后 Nvidia 带来的算力供应风险;以及 StepFun 董事长与唯一已证明汽车渠道交易对手 Qianli Technology 之间交叉任职所放大的治理 / 渠道风险,而 Qianli 自身又承受真实财务压力。单看其中任何一个风险都不致命——红筹回拆在其他公司已成功完成,国产芯片一代代改进,交叉任职在中国国资关联科技行业也常见——但这些风险的缓释成熟度普遍很低:本章风险清单中几乎没有任何事项披露了 StepFun 的正式缓释措施,除了底层关系或重组本身。最清晰的投资逻辑失效触发器因此都能从公司外部观察:CSRC 备案闭环,或 HKEX 招股书在 2026 年底后撤回 / 延迟;StepFun、Qianli 或具名芯片供应商被列入 Entity List 或五角大楼 1260H 清单;出现点名 StepFun 的 CAC 执法行动;或出现披露 StepFun 与 Qianli 之间非独立交易条款的关联交易文件。每一项都是公开且可监控事件,而不是私下事件,因此尽调可以在两次报告刷新之间追踪 StepFun 的风险轨迹,而不只依赖公司披露。[CR001, CR009, CR016, CR028, CR030, CR033]

缓释措施与否决标准表
风险可监控触发项阈值 / 事件行动含义
红筹 / CSRC 上市风险CSRC 备案状态或 HKEX 招股书发布CSRC 审查未完成,或 HKEX 申请撤回 / 延至 2026-12-31 之后推迟进入 / 退出时点;重估与 IPO 绑定的估值假设。
芯片出口管制收紧BIS Entity List 或五角大楼 1260H 名单更新StepFun、Qianli 或具名算力供应商被直接列名重新评估算力供应连续性,并对制裁风险适用估值折价。
汽车渠道集中Qianli Technology 季度财务披露核心净利润继续为负,或 Geely 收入占比下降并释放渠道压力信号下调汽车渠道收入假设;寻找 Geely/Qianli 之外的多元化证据。
内容标识 / Cybersecurity Law 执法CAC 对生成式 AI 同行的公开执法行动或罚款任何有文件记录的 StepFun 专属暂停、罚款或服务限制视为打破投资论点的触发项,需要立即重新承保。
价格战压缩利润率中国实验室公开 API 价目表变化StepFun 在 2026 年已被压缩的价格水平下进一步降价,且没有用量抵消重建单位经济模型和每 token 收入轨迹。
基准可信度独立、供应商中立的基准发布StepFun 在经审计榜单上的排名被实质性下调重新评估产品差异化论点。
治理 / 董事互锁董事会披露或关联交易文件任何 StepFun-Qianli 关联交易披露为非公平条款升级治理尽调;考虑把董事会独立性作为投资先决条件。

触发项是本章作者选取的可监控公开披露事件,不是 StepFun 自身发布的阈值。

[CR001, CR016, CR033, CR036, CR009, CR012]
FR001: 风险热力图

StepFun 最严重的风险集中在监管 / 上市执行、芯片出口管制敞口,以及汽车渠道 / 治理集中度。

各单元格都是序数判断,依据是保留的公开监管、法律和新闻证据,而不是 StepFun 披露的内部风险评分模型。

[CR001, CR002, CR009, CR012, CR016, CR019]
FR002: 风险传导图

监管、算力、竞争和治理风险共同传导至收入质量、融资通道和 IPO 估值。

[CR001, CR009, CR012, CR016, CR028, CR030]

7.8 证据要点

Chapter 08

08估值

8.1 建议:真实设备切口定价过高,公开披露还无法承销

StepFun 进入本次尽调时,确实有一条差异化分发故事:完成股改、拆除红筹架构,并完成近 $2.5 billion 的 pre-IPO 轮,Hong Kong Investment Corporation 与智能手机 ODM 和零部件投资方 Huaqin、Longcheer、OmniVision、ZTE 一同入局。这组投资人不是普通财务资本;它正是 StepFun 要把模型持续嵌入设备所需的供应链。创始人 Jiang Daxin 曾任 Microsoft 副总裁,自 2023 年 4 月以来搭建了约 500 人、80% 技术人员的团队;公司也获得外部认可,被视为 DeepSeek 之外值得关注的中国 AI 实验室之一。 但这些都没有回答本章必须回答的定价问题。StepFun 香港 IPO 目标估值 $10-12 billion,对应其报道的 2025 年收入 $73.3 million 和市场预测的 2026 年收入 $175.9 million——隐含倍数约为过去收入 137-164x、远期收入 57-68x。独立评论明确把 $12 billion 估值与“national champion”情绪和国资支持资本相连,而不是与已披露基本面相连;本章审阅的来源也没有提供审计财报、现金头寸或已提交招股书。面对真实但未验证的投资逻辑、偏高但非极端的倍数,以及薄弱证据基础,合适结论是继续研究,而不是买入:跟踪申报和上市后前两个季度;没有审计数字时,不要在区间低端以上投入资本。[CV001, CV002, CV003, CV005, CV036, CV038]

建议摘要表
决策字段当前观点决策含义
建议继续研究跟踪 StepFun 已提交的香港 IPO;在没有经审计招股书前,不要按当前 $10-12B 目标估值投入新资本。
置信度可比公司组和倍数数据较强;StepFun 专属披露(现金、利润率、cap table)偏薄。
风险评级IPO 时点 / 执行风险叠加指数驱动的同行波动,并压在第 7 章标出的监管和渠道风险之上。
估值立场偏高隐含 57-68x 远期收入倍数只有在 2026 年收入预测兑现、定价靠近 $10B 而非 $12B 时才站得住。
持有 / 退出姿态穿透招股书和上市后前两个季度继续跟踪Zhipu 和 MiniMax 上市后前六个月均出现 40-70%+ 波动;上市时价格发现尚未完成。
价格纪律在缺少审计财务前,不在 $10-12B 区间低端以上投入资本国家支持的基石需求可以支撑发行价,但不能验证底层倍数。

单元格反映本章对下文可比公司组和披露证据的综合判断,不是公司提供的预测。

[CV001, CV036, CV037, CV038, CV039, CV045]
FV001: 建议逻辑

建议的关键,是 StepFun 真实的设备分发切口,能否抵消偏高入场价和偏薄公开披露。

[CV001, CV002, CV036, CV038, CV045]

8.2 价格背景:相对公开基准偏高,相对中国 AI 虎同伴处于中游

这里最有用的视角不是单个倍数,而是把一组真实可比公司叠起来看。Zhipu AI 是第一家在香港上市的中国大模型公司(2026 年 1 月 8 日),披露 FY2025 收入 RMB724 million、净亏损 RMB4.72 billion;按 2026 年中约 HK$495.8 billion 的市值计算,静态市销率约 576x——即便相对独立分析认为全球顶级 SaaS 极端区间为 20-40x 的上沿,这也是离群值。MiniMax 晚一天上市,披露 FY2025 收入 $79 million、毛利率 25.4%,当前市值约 $16.8 billion,较 52 周高点跌去超过三分之二,隐含约 212x 收入。Moonshot AI 私募估值在五个月内从 $4.3 billion 升至 $20 billion,背后是 $200 million 年化经常性收入,接近 100x ARR;DeepSeek 首轮外部融资把估值定在 $50 billion 以上,却完全没有披露收入指标;Baichuan Intelligence 正转向更窄的医疗 AI,估值约 $2.7-2.8 billion,处在这组可比公司的低端。 放在这一背景下,StepFun 隐含 57-68x 远期倍数相对 Palantir 约 ~53.5x 的公开市场基准偏高,但并不是自己可比组中最贵的定价——Zhipu 和 MiniMax 按收入倍数都高得多,其中一部分由稀薄自由流通盘、Hang Seng Tech Index 纳入资金和基石锁定机制托住,而不是由审计利润率或现金生成证据托住。因此,StepFun 自身定价只有在 2026 年收入预测兑现、且交易按 $10 billion 而非 $12 billion 附近定价时,才看起来可辩护。[CV010, CV011, CV012, CV013, CV020, CV021]

投资论点 / 反论点表
论点方向改变观点的条件
StepFun 的 4000 万+ 设备覆盖和 Geely 汽车项目,给了它一个中国模型同行少有能匹敌的分销楔子。正方经验证的单设备经济性或单元级收入分成条款,会把它从分销故事强化为变现论点。
2026 年创纪录的融资冲刺(Series B+、约 $2.5B pre-IPO 轮、HKIC 支持)显示资本获取深、来源也在多元化。正方IPO 失败或大幅下调定价,都会削弱这一信号。
隐含的 57-68x 预期收入倍数高于 Palantir 约 53.5x 的公开市场基准,但在同组可比公司里并非最极端(Zhipu 约 576x,MiniMax 约 212x)。支持论点若 IPO 定价靠近 $12B 而非 $10B,倍数会更难自圆其说。
公开来源没有披露 StepFun 经审计的现金、烧钱速度、利润率或股权结构。反向论点若递交招股书并附经审计财务数据,投资论证的可承做性会明显提高。
独立评论者把约 $12B 目标更多归因于国家冠军叙事和国资背书,而不是已披露基本面。反向论点若收入基数获得独立验证,或盈利路径可信,才能反驳这一批评。
这组可比公司中,2026 年已披露上市的中国 AI 虎公司(Zhipu、MiniMax)股价都较上市后高点回撤 40-70%+。反向论点若上市后能走出一段稳定期,任何 StepFun 入场价面对的这项风险都会下降。

这些论点刻意对价格敏感:它们讨论当前 $10-12B 目标估值已经假设了什么,而不是泛泛评价公司质量。

[CV002, CV005, CV012, CV013, CV024, CV036]
可比估值表
可比对象指标倍数 / 估值 / 状态参考意义局限
StepFun(标的)目标 IPO 估值相对 2025/2026 年报告收入约 137-164x 往绩,约 57-68x 预期标的公司;给出本尽调必须论证的倍数。估值只是目标,不是已定价结果;收入数据来自第三方报道,未经审计。
Zhipu AI(Z.ai,HKEX 02513,竞品)市值相对 FY2025 收入约 576x 静态 P/S(据 36Kr)最接近的已披露公开可比:首家上市的中国纯 LLM 公司。倍数被指数纳入资金和稀薄自由流通盘推高,不一定是可复制基准。
MiniMax (HKEX 0100)市值相对 FY2025 收入约 212x(HKD130.79B 市值相对 $79M 收入)第二家在 HKEX 上市的中国 AI 虎公司;显示该群体呈现的上市后波动模式。账面 $1.87B 净亏损主要是非现金优先股重估,并非现金烧钱信号。
Moonshot AI (Kimi)私募估值相对年化经常性收入约 100x ARR($20B 估值 / $200M ARR)按 ARR 倍数看,最接近的中国私募可比公司。私募标记,不是交易所交易价格;ARR 数据来自公司 / 顾问报道。
DeepSeek私募估值;无披露收入指标估值高于 $50B,无公开收入数据显示 2026 年中国 AI 私募估值的天花板。无法计算收入倍数;创始人控制的治理结构与 StepFun 差异很大。
Baichuan Intelligence私募估值~$2.7-2.8B显示已披露中国 AI 虎公司估值区间的低端。战略不同(转向垂直医疗 AI),直接可比性有限。
Palantir(PLTR,公开市场基准)市值相对 TTM 收入约 ~53.5x P/SAI 溢价软件定价的公开市场天花板参考。盈利、美国上市、政府业务敞口高;不是直接行业同业。

这些倍数混合了公开市场市值比、私募 ARR 倍数和一个无法测算案例(DeepSeek);只能作方向性背景,不是逐项可比筛选。

[CV010, CV011, CV012, CV020, CV021, CV022]
FV002: 估值敏感性

StepFun 的 IPO 目标是否站得住,只在很窄的收入倍数假设区间内才显得稳妥。

收入门槛只是用本章来源引用的取整倍数搭出的 EV / 收入桥,不是折现现金流输出。

[CV036, CV037, CV034]

8.3 情景逻辑:指数资金驱动的上行真实存在,但执行失误也会快速重估

乐观情景很直接:如果 StepFun 的设备和汽车足迹(40M+ 智能手机、1M+ 目标 Geely 车辆)能转化为可变现的 API 和授权收入,并达到或超过 $175.9 million 的 2026 预测;且 IPO 靠近区间上端定价,并获得与 Zhipu 股价相似的指数纳入和基石锁定支撑,StepFun 有可能在随后一年复合到 $12-20 billion 区间,接近 Moonshot 当前估值。这个结果要求执行零失误:不能重演第四章已标出的 2025 年中收入指引落空,设备伙伴继续配合,CSRC 驱动的上市审核也不能转向不利。 基准和悲观情景更容易由公开证据支撑。Zhipu 和 MiniMax 上市后六个月内都较高点波动超过 40-70%,因此即便 StepFun 以接近 $10-11 billion 顺利上市,也不能保证价格稳定——2026 年中国 AI 实验室估值既取决于底层经济性,也同样取决于指数资金、基石锁定和稀缺性叙事。悲观情景还多一条下行路径:红筹和 CSRC 审核导致上市延迟或缩水、2025 年指引落空重演,或设备伙伴要求惩罚性收入分成;任一情况都可能把估值重置回 2026 年早些时候报道已披露的 $4-6 billion 投前估值区间。[CV026, CV027, CV036, CV040, CV041, CV042]

乐观 / 基准 / 悲观情景表
情景假设估值 / 回报逻辑关键风险概率信号
乐观2026 年收入达到或超过 $175.9M 预测,设备 / 汽车分发转化为可变现的 API 和授权收入,IPO 定价接近 $12B,并获得类似 Zhipu 的指数纳入支撑。以 $12B 对应 $175.9M 收入,倍数约 68x;若 2027 年收入再翻倍且倍数稳定,估值会接近 $20-24B,接近 Moonshot 当前标记。需要 IPO 执行零失误、设备伙伴持续配合,且 2025 年中业绩指引落空不能重演。
基准IPO 定价约 $10-11B,2026 年收入接近但不高于预测,股票呈现 Zhipu / MiniMax 式波动,而不是稳定重估。隐含价值维持在 $10-11B 附近,上市后大幅波动;若没有经审计利润率披露,难以形成持续重估。解禁后自由流通盘扩大,指数资金流和基石锁定支撑会衰退。中高
悲观持续中的红筹 / CSRC 审查导致上市延迟或缩小规模,2025 年中式业绩指引落空重演,或设备伙伴分成压缩 StepFun 自身留存。估值重置到 2026 年早些时候披露的 $4-6B 投前区间,较 $10-12B 目标折价 50-60%。监管时点风险、收入分成压缩,以及 2026 年 API 价格战带来的利润率天花板会相互叠加。

这些情景采用收入倍数框架,而非折现现金流模型,因为公开证据对 StepFun 自身的利润率和现金流转化披露太少。

[CV026, CV036, CV042, CV043, CV045, CV046]
FV003: 估值 / 回报区间

可行估值区间很宽,因为倍数怎么选、IPO 何时发生,与收入增长同样重要。

区间本身是供投委会讨论的情景化收入倍数输出,不是管理层指引,也不是已定价结果。

[CV026, CV036, CV045, CV046]

8.4 退出准备度和最终尽调:哪些要求会改变本结论

StepFun 已补齐上市的结构性前置条件——股改、红筹拆除,以及已经纳入国资和产业资本的基石阵容;这类资本曾支撑 Zhipu 上市后的韧性。上述条件必要,却不足以支撑积极估值结论。证据缺口很具体,也能补齐:经审计招股书披露现金、 烧钱速度、毛利率和分部收入;股权结构表说明本轮后 HKIC、上海国资和产业投资人实际持股;最终 IPO 定价和募资规模; 以及 StepFun 自身设备和 OEM 收入分成条款,而不是本尽调其他部分引用的行业 30-60% 常态。 每补上一个缺口,就能关掉一个会打穿投资论点的触发器。若没有新增披露,上市估值定在 $12 billion 或更高,一旦剔除指数资金流影响, 前瞻倍数会超过 Zhipu 已被拉伸、又受指数推高的基准。若上市后波动达到 Zhipu 和 MiniMax 已经历的幅度,就会确认流动性驱动的重估模式同样适用于 StepFun;也就是说,任何上市入场价都只能先按临时标记处理。Zhipu、Kimi、MiniMax 到 2026 年中已经出现又一轮 API 降价; 无论 IPO 如何定价,这都给 StepFun 自己按 token 计价的业务压上毛利天花板。[CV002, CV038, CV039, CV044, CV047, CV048]

论点失效与退出触发条件表
触发条件证据 / 阈值对投资论点的影响行动含义
未新增披露却以 $12B 或更高定价 IPO从申报到上市期间没有新增经审计财务数据隐含预期倍数被推高到 68x 以上;剥离资金流影响后,甚至超过 Zhipu 被指数纳入推高的基准不追加资金;等待上市后披露。
上市后两个季度内波动超过 50%上市六个月内出现这种幅度的股价变动,类似 Zhipu 和 MiniMax证实可比公司组由流动性驱动的重定价模式同样适用于 StepFun把任何上市后入场价都视为临时价格,而非稳定标记。
设备伙伴收入分成被证明过高StepFun 专属收入分成披露达到或高于行业惯例 60% 上沿压缩 40M+ 设备覆盖带来的实际留存,削弱分发论点将估值立场从偏高下调为昂贵。
算力配给 / API 价格战加剧Zhipu、Kimi 或 MiniMax 进一步下调 token 价格或出现排队表明整个可比组都有利润率天花板,包括 StepFun 自身按 token 定价的业务上调悲观情景概率并重新论证。
监管或 CSRC 红筹审查延迟或阻断上市上市时间表滑到 2026 年之后移除该估值依赖的近期流动性催化剂在重新申报前,将立场转向继续研究 / 回避。

这些触发条件专门服务于估值判断,不是复述第 7 章已经覆盖的运营风险清单。

[CV006, CV014, CV017, CV019, CV041, CV042]
最终尽调问题表
主题缺失证据重要性负责人或尽调路径
经审计财务 / 招股书现金、烧钱速度、毛利率和分部收入若不核验背后的收入基数,无法论证 57-68x 预期倍数HKEXnews 发布后的申报文件;管理层尽调资料室。
股权结构 / 所有权HKIC、上海国资和产业投资者持股比例决定稀释幅度,也判断 $10-12B 目标反映真实经济权益还是一轮融资的表面规模向公司或法律顾问索取股权结构表和股东名册。
IPO 定价结果最终发行价、募资规模和基石配售直接解答入场倍数落在约 57x 还是 68x+跟踪 HKEXnews 上市公告和定价补充文件。
设备 / OEM 收入分成条款与 Geely、OPPO、Honor 和 ZTE 的单设备或收入分成比例决定 40M+ 设备覆盖中有多少转化为 StepFun 自身收入,而不是伙伴收入通过管理层或伙伴 IR 索取合作合同条款。
上市后稳定计划绿鞋机制、锁定期安排和指数纳入时间表Zhipu / MiniMax 先例显示,这些机制驱动上市初期的大部分价格波动招股书提交后审阅稳定价格章节。

每个问题都可能直接改变上述建议、信心或可接受入场价,因此被列入。

[CV002, CV038, CV039, CV044, CV048]
FV004: 投资 KPI

StepFun 在分发和资本通道上得分高,但披露质量和估值支撑偏弱。

分数为 0-10 的序数判断,基于本章证据综合而来,供投委会讨论;不是供应商提供的评分卡。

[CV002, CV036, CV038, CV045, CV047]

8.5 图表

免责声明

本报告仅供信息参考,基于截至 2026-07-01 的公开来源,不构成投资建议。StepFun 仍是私营、pre-IPO 公司;文中引用的财务和经营指标大多来自第三方报告,且未经审计。作出任何投资决定前,应独立核验这些指标。

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. SO009, SO011
CO002 StepFun was founded in April 2023 by Jiang Daxin, Zhu Yibo, and Jiao Binxing. SO009, SO023
CO003 Baidu Baike's corporate registry lists StepFun's founding date as April 6, 2023, registered capital of RMB 20 million, and Jiang Daxin as legal representative. SO011
CO004 StepFun's registered headquarters address is Floor 30, No. 701 Yunjin Road, Xuhui District, Shanghai. SO011
CO005 English-language coverage, including Wikipedia, generally describes StepFun's headquarters only as Shanghai, China, without the more specific street address found in Chinese-language corporate registries. SO009
CO006 StepFun focuses on multimodal foundation models spanning text, image, video, and audio under its "Step" model family. SO009, SO004
CO007 StepFun's official platform highlights "Step 3.7 Flash" as a production-oriented Flash model for agentic use with native multimodal understanding. SO002
CO008 StepFun operates the consumer chat product "跃问" (Yuewen) at chat.stepfun.com, offering chat, knowledge-base Q&A, and image-generation tools. SO003
CO009 Chinese Wikipedia describes StepFun's companion/role-play consumer app "冒泡鸭," which targets open-world and character-interaction scenarios. SO010
CO010 StepFun is widely described in press and encyclopedia coverage as one of China's "Six AI Tigers," alongside Zhipu AI, MiniMax, Moonshot AI, 01.AI, and Baichuan Intelligence. SO026, SO021
CO011 MIT Technology Review reported that StepFun released 11 foundational AI models in 2024 spanning language, vision, video, audio, and multimodal systems while pursuing AGI, a goal many domestic peers have deprioritized. SO026, SO009
CO012 A StepFun executive told 36Kr the company's strategic ambition is closer to being "China's xAI plus Tesla" than China's OpenAI or Anthropic, reflecting its physical-world terminal focus. SO013
CO013 StepFun's CEO Jiang Daxin is a former Microsoft global/senior vice president who previously worked on products including Bing and Microsoft 365. SO023, SO013
CO014 StepFun's CTO Zhu Yibo previously led AI infrastructure and large-scale cluster operations at ByteDance before co-founding StepFun. SO013
CO015 StepFun's chief scientist Zhang Xiangyu is a core author of the ResNet architecture. SO013
CO016 Chinese Wikipedia identifies Jiao Binxing as StepFun's data lead, a former head of Microsoft Bing's core search team focused on data mining and NLP algorithms. SO010
CO017 On January 26, 2026, StepFun announced the appointment of Yin Qi as chairman, responsible for overall strategy and technical direction. SO018, SO013
CO018 Yin Qi co-founded and previously served as CEO of Megvii (Face++), one of China's earlier "AI Dragons," and concurrently chairs Qianli Technology, a Geely-linked smart-driving company. SO013, SO018
CO019 Yin Qi's dual role at StepFun and Qianli Technology is intended to link StepFun's foundation-model R&D to real-world hardware deployment data, forming a "Physical AI" feedback loop. SO018
CO020 The Standard reported, citing an interview with founder Jiang Daxin, that StepFun's team exceeds 500 members, with nearly 80% in algorithm and technical roles. SO021
CO021 Chinese Wikipedia's description of StepFun's earlier R&D team put headcount above 150 professionals, indicating substantial headcount growth by 2026. SO010
CO022 Public materials reviewed for this chapter do not disclose a complete current board roster or a cap-table/ownership breakdown for StepFun.
CO023 StepFun's first (angel-stage) funding round reached unicorn valuation shortly after founding, backed by Sequoia China (HongShan), Qiming Venture Partners, and IDG Capital. SO013, SO024
CO024 StepFun's 2024 Series A round added US-dollar investors 5Y Capital and Shunwei Capital alongside strategic investor Legend Capital. SO013, SO024
CO025 StepFun's Series B round launched in mid-2024 and closed around December 2024, backed by Shanghai State-owned Capital Investment and affiliated funds, the Hong Kong Investment Corporation, Tencent, and returning shareholders. SO013, SO024
CO026 On January 26, 2026, StepFun closed a Series B+ round exceeding RMB 5 billion (about $718-733 million), described by TechNode and KrASIA as the largest single financing in China's large-model sector over the prior 12 months. SO012, SO013, SO016
CO027 New investors in StepFun's Series B+ round included Shanghai SDIC/S&T Capital InnoFund, China Life Private Equity Investment, Pudong Venture Capital, Xuhui Capital, Wuxi Liangxi Fund, Xiamen ITG Group, and Huaqin Technology, alongside follow-on participation from Tencent, Qiming Venture Partners, and 5Y Capital. SO012, SO014, SO020
CO028 Xiamen ITG Group's participation in StepFun's Series B+ round was reportedly its first investment in a large-model developer, and Huaqin Technology's involvement signaled hardware-software integration interest from a top smartphone ODM. SO014
CO029 StepFun completed a shareholding reform converting from a limited liability company to a joint stock company around early April 2026, a step regarded as a prerequisite for a Hong Kong listing. SO015, SO024
CO030 As part of the same restructuring, StepFun dismantled its offshore "red-chip" structure, which had been incorporated via the Cayman Islands. SO015, SO023
CO031 Reuters reported that StepFun's unwind of its offshore structure followed new CSRC guidance instructing some red-chip companies to re-domicile before overseas listings, and that StepFun judged an onshore structure more appropriate given its heavy state-capital backing. SO023, SO024
CO032 Reuters reported that Chinese law firm Hankun found roughly one-fifth of 131 Hong Kong listings China approved in the prior year involved offshore holdings, mostly using the red-chip structure. SO023
CO033 BigGo Finance reported that red-chip restructuring can delay IPO timelines and, in some cases, lead companies to abandon listing plans given the cost of changing legal structure. SO024
CO034 StepFun was nearing a pre-IPO funding round of almost $2.5 billion as of May 2026, according to KrASIA's translation of an IPO Zaozhidao report. SO015, SO017, SO025
CO035 Industrial investors in StepFun's pre-IPO round included Huaqin Technology, Longcheer Technology, OmniVision Technologies, and ZTE Corporation, spanning smartphone ODM, camera-sensor, and handset-brand supply-chain roles. SO015, SO021
CO036 The Hong Kong Investment Corporation joined StepFun's pre-IPO round as its only disclosed generative-AI investment to date. SO015, SO021
CO037 BigGo Finance's compiled funding table put StepFun's Series B+ (January 2026) round at just over CNY 5 billion (about $733.1 million), and described StepFun's ongoing pre-IPO round as targeting a pre-money valuation of $4-6 billion across two tranches. SO024
CO038 Reuters reported that Chinese financial outlet Caijing said StepFun planned to raise CNY 2-3 billion in its pre-IPO round at a valuation of up to $6 billion, then file for a Hong Kong IPO by end of June 2026 at a $10 billion anchor valuation. SO023, SO024
CO039 On June 9, 2026, StepFun filed for a Hong Kong IPO targeting a $10-12 billion valuation and an approximate $500 million raise. SO022
CO040 FAQ.com.tw reported that StepFun's flagship models power AI features in more than 40 million smartphones and that its Geely partnership targets deployment in more than one million vehicles. SO022
CO041 BigGo Finance reported that market sources placed StepFun's device deployment above 42 million units from manufacturers including OPPO and Honor, serving close to 20 million daily users. SO024
CO042 BigGo Finance's compiled financial data put StepFun's 2025 revenue at roughly CNY 500 million (about $73.3 million), with a 2026 projection of about CNY 1.2 billion (about $175.9 million). SO024
CO043 The gap between StepFun's roughly $73 million of reported 2025 revenue and its targeted $10-12 billion IPO valuation is a valuation-support risk this chapter flags rather than resolves, since the revenue figure comes from secondary market reporting rather than an audited filing. SO024, SO022
CO044 The Standard reported that fellow "Six AI Tigers" members Zhipu AI and MiniMax had already listed in Hong Kong by January 2026, with market capitalizations of roughly $49.3 billion and $33.9 billion respectively. SO021
CO045 StepFun launched its first large model, the roughly 100-billion-parameter Step-1, in August 2023, within months of founding. SO010, SO013
CO046 StepFun released the multimodal Step-1V model in November 2023, followed by a Step-2 trillion-parameter Mixture-of-Experts preview in March 2024. SO010
CO047 At WAIC 2024 in July, StepFun officially launched Step-2 (a trillion-parameter MoE LLM), the Step-1.5V multimodal model, and the Step-1X image-generation model. SO009, SO010
CO048 In February 2025, StepFun open-sourced Step-Video-T2V (a 30-billion-parameter text-to-video model) and Step-Audio jointly with Geely. SO009, SO010
CO049 At WAIC 2025, Geely Auto Group and StepFun jointly showcased the Galaxy M9, described as the world's first vehicle with a "human-like AI agent," powered by StepFun's end-to-end voice model, alongside the jointly built Agent OS smart-cockpit system. SO019
CO050 StepFun open-sourced Step 3, a 321-billion-parameter multimodal reasoning model with 38 billion active parameters, on July 31, 2025, under the Apache 2.0 license. SO006, SO007, SO008, SO010
CO051 Step 3's technical paper describes a Multi-Matrix Factorization Attention (MFA) mechanism and Attention-FFN Disaggregation (AFD) system co-design intended to cut decoding costs versus DeepSeek-V3 and Qwen3 MoE 235B. SO008, SO006
CO052 In February 2026, StepFun released Step 3.5 Flash, an open-source Apache-2.0 Mixture-of-Experts model with 196 billion total and 11 billion active parameters supporting a 256k-token context window and tool use. SO009
CO053 Reuters reported that since its February 2026 launch, Step 3.5 Flash has ranked among the three most-used models on the OpenClaw AI-agent platform, alongside MiniMax M2.5 and Kimi K2.5. SO023
CO054 At WAIC 2025, StepFun helped announce the "Model-Chip Ecosystem Innovation Alliance" with Chinese AI-chip makers Huawei, Biren Technology, Moore Threads, and Enflame, and joined a separate Shanghai General Chamber of Commerce AI Committee alongside SenseTime, MiniMax, MetaX, and Iluvatar CoreX. SO009
CO055 It was reported in February 2026 that StepFun was seeking an IPO on the Hong Kong Stock Exchange. SO009
CO056 Beyond Baidu Baike's Chinese-language corporate registry, reviewed sources do not provide an independently corroborated street address for StepFun, leaving an identity-verification gap for this chapter. SO009, SO011
CO057 The clearest adverse milestone in StepFun's chronology is the 2026 red-chip/offshore-structure unwind undertaken to satisfy CSRC guidance ahead of a Hong Kong listing. SO023, SO024
CM001 StepFun's official platform.stepfun.com documentation prices its foundation models -- including step-3.7-flash and step-3.5-flash -- on a per-token API metering basis rather than flat software licensing, indicating the company's primary directly monetized surface is developer and agent API consumption. SM020
CM002 IDC's China AI agent market-glance (4Q25) separates model-development, application-development, and agent-development platforms from the underlying compute, storage, network, and semiconductor infrastructure layer, providing an independent boundary between application-layer AI spend and infrastructure spend. SM007
CM003 StepFun's Step-Audio end-to-end voice model and its jointly built 'Agent OS' smart-cockpit software were unveiled by Geely at WAIC 2025 as the industry's first mass-produced-vehicle deployment of an end-to-end voice large language model, placing automotive embedded-cockpit licensing inside StepFun's addressable market rather than vehicle-hardware spend. SM011
CM004 As of April 30, 2026, mainland China had 868 generative-AI services registered nationally with the Cyberspace Administration of China and 530 AI applications or functions registered locally, with 72 new national registrations and 49 new local registrations added in March-April 2026 alone. SM003, SM022
CM005 China's 2023 Measures for the Administration of Generative Artificial Intelligence Services (Article 12) require providers offering generative-AI services to the Chinese public to label generated images and video per the Provisions on the Administration of Deep Synthesis Internet Information Services, establishing a compliance boundary around any public-facing generative-AI service, including StepFun's. SM015
CM006 OpenRouter's and OpenClaw's real API-call-based leaderboards -- independent of any StepFun self-reporting -- rank StepFun's Step 3.5 Flash among the most-used models globally, corroborating company claims of rapid developer adoption with independently observed usage data rather than marketing claims alone. SM009, SM010
CM007 Forrester estimates China's AI-infrastructure (data-center and cloud-platform buildout) spending will exceed US$70 billion in 2026, a distinct and much larger spend pool than the application-layer generative-AI market that StepFun monetizes directly. SM024
CM008 IMARC estimates China's generative AI (AIGC) market reached USD 5.16 billion in 2025 and forecasts it will grow to USD 19.56 billion by 2034 at a 15.96% CAGR over 2026-2034. SM001
CM009 Gartner forecasts worldwide AI spending will total US$2.59 trillion in 2026, a 47% year-over-year increase, and has revised short-term GenAI model-consumption growth up to 110% for the year. SM017
CM010 Forrester forecasts China's AI infrastructure spending will exceed US$70 billion in 2026, driven by Alibaba, ByteDance, and MIIT's industrial-digitalization mandate, even as broader enterprise IT spending growth slows amid weak domestic demand. SM024
CM011 China's public-cloud model-as-a-service token consumption reached 1,944 trillion tokens in 2025, a 16-fold year-over-year increase, with ByteDance's Volcano Engine holding a 49.5% share of that public-cloud MaaS call volume. SM012
CM012 As of April 30, 2026, China had 868 nationally CAC-registered generative-AI services and 530 locally registered generative-AI applications, a regulatory-scale proxy for how many distinct generative-AI offerings are now operating in the market StepFun competes in. SM003
CM013 Nikkei-sourced reporting cited by TrendForce found Chinese-developed generative AI models held about 15% global usage share in November 2025, up from roughly 1% a year earlier, driven largely by DeepSeek's and Qwen's open-source releases. SM023
CM014 IDC's China AI FutureScape excerpt sets national targets of 70% penetration of smart devices and AI agents by 2027 and 90% by 2030 under China's 'AI+' strategy, and separately projects 80% of China's top-1,000 enterprises will prioritize AI sovereignty (nonpublic hosting, open technologies, regional partners) by 2027. SM007
CM015 IDC's China foundation-model landscape (CHC5530125, 中国基础大模型综合评估) independently lists StepFun's 'Step Reasoner mini' alongside Alibaba Qwen3-Max, DeepSeek R1/V3.2-Exp, Moonshot Kimi K2, Zhipu GLM-4.6, and ByteDance Doubao 1.6 as one of the frontier Chinese foundation models it tracks as of late 2025. SM007
CM016 Published sizing lenses are large but non-overlapping proxies for StepFun's addressable market: a roughly $5-20 billion narrow China AIGC-applications estimate (IMARC), a >$70 billion China AI-infrastructure-spend lens that sits upstream of StepFun's monetized layer (Forrester), and a >1,944-trillion-token public MaaS consumption lens (IDC-sourced Volcano Engine data) that is dominated by hyperscaler platforms rather than independent vendors like StepFun. SM001, SM024, SM012
CM017 Public reporting credits StepFun's flagship models with powering AI features in more than 40 million smartphones and targeting deployment in more than one million vehicles through its Geely partnership, describing StepFun's strategy as embedding models inside devices, vehicles, and industrial hardware rather than delivering AI purely through cloud subscriptions. SM027
CM018 IDC forecasts China smartphone shipments will fall 2.2% to about 278 million units in 2026, while next-generation AI-phone shipments rise to 147 million units -- 53% of the total and the first time AI phones surpass half the market -- under an 'edge-cloud synergy' (端云结合) model blending on-device inference with cloud AI. SM002, SM025
CM019 IDC forecasts China's total PC market will ship about 42.22 million units in 2026 (-0.8% year-over-year) while GenAI PC shipments grow 146.5% year-over-year, reaching an estimated 36.5% of the total PC market by 2029 at a 58.7% 2025-2029 CAGR. SM014
CM020 中商产业研究院 forecasts China's smart-cockpit solutions market will reach RMB 182.8 billion in 2026, up from RMB 154.5 billion in 2025 and RMB 129 billion in 2024, with in-vehicle voice-interaction penetration rising from 26% to 48% within a year. SM013
CM021 Geely's Galaxy M9 hybrid SUV, unveiled at the 2025 World AI Conference, is described by both Geely and StepFun as the industry's first mass-produced vehicle to ship an end-to-end voice-large-language-model-powered 'Agent OS' smart cockpit, built on StepFun's Step-Audio model. SM011
CM022 Independent analysis of StepFun's strategy argues that its Geely tie-up gives fast, large-scale deployment but also 'a looming constraint': rival automakers may be reluctant to hand a vehicle's AI 'brain and eyes' to a supplier so closely associated with one carmaker, which can cap StepFun's addressable market as an ostensibly independent model vendor. SM026
CM023 China's foundation-model layer saw only about 22 disclosed financing deals worth roughly RMB 9.4 billion in 2025, about half the prior year's deal value, indicating capital available to independent Chinese model vendors like StepFun was contracting even as device-embedded deployment expanded. SM026
CM024 StepFun's buyer map splits into at least four groups: individual consumers using its consumer chat/voice apps (self-pay), developers and agent builders metering platform.stepfun.com API tokens (engineering/product budget), smartphone and automotive OEMs licensing embedded models (hardware R&D or BOM budget), and enterprises subject to CAC registration and content-labeling compliance (platform/compliance budget). SM020, SM008, SM027, SM003
CM025 StepFun's Step Plan monthly subscription, launched March 23, 2026, targets OpenClaw-agent and AI-coding developers specifically, pricing four tiers from a standard RMB 49/month (developer-community discount RMB 25/month) up to RMB 699/month (discounted RMB 349/month), distinct from its per-vehicle or per-device OEM licensing. SM021
CM026 OpenRouter's real API-call leaderboard placed Step 3.5 Flash (free tier) second among all models globally in the week of March 9-15, 2026, with weekly usage of 1.34 trillion tokens, up 79% week-on-week, evidencing rapid usage-verified adoption specifically in the developer/agent-builder segment. SM009
CM027 For smartphone and automotive OEM buyers, the adoption trigger is typically a flagship product cycle or joint showcase -- such as Geely's WAIC 2025 Galaxy M9 unveiling -- rather than a self-serve trial, meaning StepFun's device channel depends on a small number of high-value OEM relationships rather than broad self-serve funnel conversion. SM011, SM026
CM028 IDC's China AI agent market-glance segments enterprise buyers by function (finance, HR, customer service, sales & marketing, supply chain, legal, analytics, operations) and by industry (FSI, government, manufacturing, transportation, energy, healthcare, retail, automobile, robotics, devices, internet), indicating StepFun's enterprise/API buyer segment is fragmented across many smaller budget pools rather than concentrated in one large procurement category. SM007
CM029 China's May 2026 CAC/NDRC/MIIT 'Implementation Opinions on the Standardized Application and Innovative Development of Intelligent Agents' is the country's first dedicated national policy framework for AI agents, defining agents as autonomous perception-memory-decision-execution systems and naming 19 priority deployment scenarios spanning scientific research, manufacturing, transportation, finance, healthcare, education, and government. SM004
CM030 China's 2025 Measures for Labeling AI-Generated and Synthetic Content, effective September 1, 2025, require explicit (user-visible) and implicit (metadata/watermark) labeling of AI-generated text, images, audio, video, and virtual scenes, adding a recurring compliance obligation on top of the 2023 generative-AI service measures. SM016
CM031 Chinese authorities (CNCERT/CC, MIIT, and the National Computer Virus Emergency Response Center) have issued high-level warnings about vulnerabilities and counterfeit skill packages embedded with Trojan viruses on the OpenClaw agent platform -- the same agent ecosystem where StepFun's Step 3.5 Flash ranks near the top of usage charts -- creating a live security and trust constraint on agent-platform growth. SM019
CM032 The U.S. Commerce Department closed a loophole on May 31, 2026 that had allowed overseas subsidiaries of China-headquartered firms to buy advanced Nvidia and AMD AI chips without a license, extending compute-access restrictions to entities based on headquarters location rather than delivery location. SM005
CM033 Chinese frontier labs cut LLM API prices at least six times in the first half of 2026, with three of the cuts (DeepSeek, Xiaomi MiMo, Moonshot Kimi) declared permanent, pushing per-token prices for some Chinese models 5-34x below comparable Western frontier APIs and compressing the margin available to smaller vendors like StepFun that must price competitively against DeepSeek, Qwen, and Kimi. SM006
CM034 StepFun's own official pricing lists step-3.7-flash at RMB 1.35 (input) / RMB 8.1 (output) per million tokens and step-3.5-flash at RMB 0.70 (input) / RMB 2.10 (output) per million tokens, positioning it within the same competitive price band as DeepSeek V4-Pro, Qwen3 Max, and Kimi K2.6 rather than at a premium. SM020
CM035 China's 'AI+' penetration targets (70% by 2027, 90% by 2030) and its enterprise AI-sovereignty preference (80% of top-1,000 enterprises by 2027) function as structural demand drivers for domestic, embeddable models such as StepFun's, while simultaneously inviting larger, better-resourced entrants (ByteDance, Alibaba, Baidu, Huawei) into the same embedded/device layer. SM007
CM036 ByteDance's Doubao models alone reach over 500 million phone terminals across 9 of the world's top 10 smartphone makers and are installed in more than 7 million smart vehicles across 50-plus brands and 145 models, a distribution scale roughly an order of magnitude larger than StepFun's disclosed 40-million-phone and one-million-vehicle target, indicating StepFun competes for embedded-AI share against much larger, better-distributed platforms. SM012, SM027
CM037 Global automakers including BMW, Mercedes-Benz, and Volkswagen are increasingly co-developing China-specific cockpit AI with local partners (Huawei HarmonyOS, XPeng, Momenta) rather than importing a single foreign or independent AI vendor, intensifying competition for the same automotive embedded-AI budget that StepFun targets through Geely. SM018
CM038 IDC's China AI coding market analysis shows top vendors' 2025 revenue for AI coding tools remained below RMB 100 million domestically despite over 100% year-over-year growth, illustrating that fast usage growth in China's AI economy does not yet imply a large realized revenue pool. SM007
CM039 IDC forecasts 35% of professional developers in China will adopt AI-native ('vibe coding') development platforms for production apps by 2027 as enterprise-grade capabilities mature, expanding the addressable developer segment for API-based model vendors including StepFun. SM007
CM040 IDC's China AI coding-market data shows a fragmented competitive set of at least a dozen named China AI-coding-IDE vendors (Alibaba Qoder, ByteDance Trae, Baidu Comate, Tencent Codebuddy, AWS Kiro, Sensetime Raccoon, Zhipu CodeGeex, and others), indicating StepFun's addressable developer/agent market is contested by well-resourced incumbents even before StepFun's own share is counted. SM007
CP001 By Q2 2026, China's foundation-model competitive set has consolidated to roughly ten providers with meaningful public usage share: Xiaomi, Alibaba (Qwen), Zhipu AI (Z.ai), DeepSeek, Moonshot AI, MiniMax, StepFun, ByteDance, Baidu, and Tencent. SP024
CP002 StepFun sits alongside Zhipu AI, Moonshot AI, MiniMax, and Baichuan Intelligence as one of the venture-funded, pure-play Chinese foundation-model startups, distinct from big-tech-embedded programs such as Qwen, ERNIE, and Doubao/Seed. SP002, SP005, SP009, SP012
CP003 Xiaomi's MiMo model family, not one of StepFun's traditionally named peers, alone commands roughly 21% of weekly Chinese-provider OpenRouter token volume in Q2 2026, more than any single pure-play foundation-model startup. SP024
CP004 Internal build or reliance on overseas frontier APIs remains a status-quo substitute for some Chinese enterprise buyers, but export controls and data-localization rules push most domestic demand toward the ten-provider Chinese set that includes StepFun. SP024
CP005 Zhipu AI, rebranded internationally as Z.ai, lists GLM-5.2 as its current flagship chatbot/agent model on its consumer-facing chat.z.ai product. SP002
CP006 Zhipu AI's FY2025 annual report (its first as a Hong Kong-listed company) showed revenue of roughly CNY 724 million (about $105 million), up 131.9% year over year, against a net loss of about CNY 4.72 billion (about $686 million), with market capitalization trading above HK$400 billion (about $51 billion). SP004
CP007 GLM-5 is a 744-billion-parameter Mixture-of-Experts model (44B active) released under an MIT license and trained end-to-end on Huawei Ascend silicon, positioning Zhipu as an export-control-resilient option for Chinese state-linked and enterprise buyers. SP024
CP008 GLM-5 direct API pricing is published at $0.80 input / $2.56 output per million tokens (GLM-5 Turbo at $1.20/$4.00), and independent tracking places GLM-5 output pricing at about $3.20 per million tokens in a May 2026 price-war comparison. SP024, SP028
CP009 Moonshot AI's flagship 2026 releases are branded Kimi K2.6/K2.5 open-weight large language models, offered through both a consumer chat app (kimi.com) and a developer API. SP005, SP006
CP010 Moonshot AI raised about $2 billion in a round valuing it at $20 billion in May 2026, up from a $4.3 billion valuation at the end of 2025, led by Meituan's Long-Z Investments with Tsinghua Capital, China Mobile, and CPE Yuanfeng participating. SP007, SP008
CP011 Kimi's consumer app ranked 8th among China's AI apps by MAU in April 2026 at about 25.3 million monthly active users, an order of magnitude below ByteDance's Doubao. SP025
CP012 Kimi K2.6 holds the lowest cache-hit input price floor among major Chinese frontier APIs tracked in May 2026, at $0.07 per million cached tokens. SP028
CP013 MiniMax describes its mission as building general intelligence 'with everyone' toward AGI, with a multimodal model line spanning language, video, audio, and music plus consumer products including MiniMax Code, MiniMax Hub, and MiniMax Audio. SP009
CP014 MiniMax's January 2026 Hong Kong IPO raised about HK$4.8 billion ($619 million), priced at the top of its range, valuing the company at about $6.5 billion, with cornerstone investors including GIC, Baillie Gifford, and the Abu Dhabi Investment Authority alongside existing backer Alibaba. SP010, SP011
CP015 MiniMax's M-series (M2, M2.1, M2.5, M2.7) shipped four numbered releases in under six months in early 2026, with M2.7 reaching #4 on OpenRouter at 1.34 trillion weekly tokens and roughly 50x cheaper than Claude Opus on comparable agentic workloads. SP024
CP016 Baichuan Intelligence pivoted its commercial strategy toward vertical healthcare AI, disbanding its financial and education teams to focus on models such as Baichuan-M2, an open-source medical LLM designed for private clinical deployment that Baichuan says can run on a single RTX 4090 GPU after quantization. SP012, SP013
CP017 On the HealthBench medical benchmark, Baichuan-M2 scored 60.1, ahead of OpenAI's gpt-oss-120b at 57.6, according to independent reporting on Baichuan's release. SP013
CP018 Baichuan Intelligence's 2024 Series A raised about $691 million (RMB 5 billion) at a $2.7 billion valuation, backed by Alibaba, Tencent, and Xiaomi alongside state-backed funds including the Beijing AI Industry Investment Fund, Shanghai AI Industry Investment Fund, and Shenzhen Capital Group; its April 2026 Series B, led by 37 Interactive Entertainment, kept the company at or above that valuation floor. SP014
CP019 DeepSeek, spun out of quantitative hedge fund High-Flyer, operated without any external investors from its 2023 founding until its first external funding round in June 2026, self-funding development from High-Flyer's balance sheet. SP017
CP020 DeepSeek's June 2026 funding round raised over CNY 50 billion (about $7.4 billion) at a valuation above CNY 330 billion (over $50 billion), structured so most outside investors hold no voting rights and must invest through a founder-controlled limited partnership with a five-year lock-up, while the National AI Industry Investment Fund invested directly with full voting rights. SP017
CP021 DeepSeek-V4-Pro and V4-Flash, released April 24, 2026 under an MIT license, support a 1-million-token context window; DeepSeek's official pricing lists V4-Pro at $0.435 input (standard) / $0.87 output and V4-Flash at $0.14 input / $0.28 output per million tokens, with cache-hit input as low as $0.0028-$0.003625. SP016, SP015
CP022 DeepSeek V4-Pro's output pricing is roughly 34x cheaper than GPT-5.5 per independent 2026 tracking, and DeepSeek's price cuts in the first half of 2026 (up to 97.5% off list) are described as permanent rather than promotional. SP028
CP023 DeepSeek V3.2, released December 2025, is an open-weight 685-billion-parameter MoE model with DeepSeek Sparse Attention and gold-medal performance on the 2025 IMO and IOI competitions, and holds roughly 5.6% weekly OpenRouter share as of Q2 2026. SP024
CP024 Alibaba's Qwen3.6-Plus, launched April 2, 2026, is positioned as an agentic-AI flagship integrated into Alibaba's Wukong enterprise automation platform and its consumer-facing Qwen App. SP020
CP025 Qwen models are the most widely used open-weight family tracked globally, and Alibaba's combined Qwen line held roughly 13.9% weekly OpenRouter share (2.77 trillion weekly tokens) in Q2 2026, the second-highest of any single provider after Xiaomi. SP024
CP026 Alibaba Cloud markets Model Studio as its managed platform for deploying Qwen and other foundation models to enterprise customers, positioning Qwen's enterprise distribution inside Alibaba's existing cloud sales and compliance infrastructure rather than as a standalone API business. SP019
CP027 Qwen3 Max output pricing is tracked at $3.90 per million tokens in a May 2026 five-lab comparison, positioned as a price-quality balance rather than the cheapest tier, while Qwen3.5 Flash undercuts it at $0.065 input / $0.26 output per million tokens. SP028, SP024
CP028 Baidu released ERNIE 5.1 in May 2026, a Mixture-of-Experts model that compresses total and active parameters versus ERNIE 5.0 while using about 6% of the pre-training cost of comparable models, and it is not released with open weights (unlike the earlier ERNIE 4.5). SP022
CP029 ERNIE 5.1 scored 1,223 on the Arena Search leaderboard on May 9, 2026, claiming 4th place globally and 1st among Chinese models, per Baidu's own release announcement. SP022
CP030 Baidu's ERNIE-powered AI assistant (originally Wenxin Yiyan, rebranded Wenxiaoyan) reached about 200 million monthly active users and over 300 million registered users, integrated into Baidu Search, Maps, and Health, with users able to switch between ERNIE and DeepSeek inside the same assistant. SP030
CP031 Baidu ERNIE 5.0 is a 2.4-trillion-parameter omnimodal flagship trained on Baidu's own Kunlun silicon and integrated with Baidu's search engine as its primary discovery surface. SP024
CP032 ByteDance's Doubao app is China's most-used consumer AI app, with independent trackers reporting 336-345 million monthly active users in Q1/Q2 2026 and average engagement of about 54.8 sessions per user per month, versus StepFun's Yuewen app, which does not appear in the top-50 China AI app MAU rankings. SP025
CP033 ByteDance's underlying Seed 2.0 model family (branded Doubao) reports 155 million weekly active users, with its Pro variant said to match GPT-5.2 performance at roughly one-tenth the cost, plus Lite and Mini tiers priced as low as $0.10-$0.25 input per million tokens. SP024
CP034 StepFun's Step 3.5 Flash (196B MoE, 11B active, 262K context, released February 2, 2026) ranks #3 on the OpenRouter free-model leaderboard at 1.38 trillion weekly tokens and trains on NVIDIA Hopper rather than Huawei Ascend hardware, reaching up to 350 tokens per second at inference. SP024
CP035 Independent 2026 benchmark tracking places Qwen3.7 Max and GLM-5.2 at the top of the Chinese-model leaderboard (score 90 each), with DeepSeek V4 Pro Max at 87 and Kimi K2.6 at 81; StepFun's models are not listed among the top five nationally ranked Chinese models on this leaderboard. SP026
CP036 StepFun's own published API pricing for step-3.7-flash (CNY 0.27 cache-hit / 1.35 standard input / 8.1 output per million tokens) and step-3.5-flash (CNY 0.14/0.7/2.1) sits inside the same broad price band as DeepSeek, Qwen, and Kimi's Chinese-priced tiers, but StepFun has not disclosed an open-weight release cadence comparable to DeepSeek's or Zhipu's MIT-licensed flagships. SP001, SP016
CP037 StepFun does not appear among China's top-10 AI apps by MAU as of April 2026, while ByteDance's Doubao (#1), Alibaba's Qianwen and Quark (#2-3), and DeepSeek's own app (#4) all exceed 130 million monthly active users, indicating StepFun's Yuewen consumer app lags every other named peer in this chapter on pure consumer reach. SP025
CP038 Chinese labs cut frontier API list prices six times in the first half of 2026, with three cuts (DeepSeek, Xiaomi MiMo, Moonshot Kimi) declared permanent rather than promotional, compressing per-token economics across the entire Chinese frontier-model tier that StepFun competes in. SP028
CP039 Chinese flagship-tier API pricing undercuts comparable US frontier pricing (e.g. GPT-5.4 at $2.50/$15.00, Claude Opus 4.6 at $5.00/$25.00 per million tokens) by roughly 2.5-5x on input and 4-8x on output at comparable context lengths, a pricing floor every Chinese provider, including StepFun, competes under. SP024
CP040 Geely's April 2026 'Full-Domain AI 2.0' cockpit and ADAS platform, unveiled at Auto China 2026, shows Chinese automakers deepening in-house and multi-vendor AI integration across cockpit, navigation, and driver-assist systems -- the same OEM channel StepFun relies on for its own automotive distribution. SP027
CP041 Xiaomi's MiMo model family alone captured about 21.1% of weekly Chinese-provider OpenRouter token volume in Q2 2026 (three times OpenAI's global share), showing a smartphone/EV hardware incumbent can out-distribute every pure-play foundation-model startup, including StepFun, through device and developer-tooling bundling alone. SP024
CP042 State-linked investment vehicles participate directly in Chinese foundation-model financing: Baichuan's funding included the Beijing AI Industry Investment Fund, Shanghai AI Industry Investment Fund, and Shenzhen Capital Group, and DeepSeek's 2026 round included a direct, voting investment from the National AI Industry Investment Fund. SP014, SP017
CP043 By mid-2026, Zhipu AI, MiniMax, Moonshot AI, and DeepSeek have each completed a public listing, IPO, or external funding round at valuations of roughly $6.5-20 billion or more, giving each materially deeper disclosed capital access than StepFun's reported near-$2.5 billion pre-IPO round and $10-12 billion Hong Kong IPO target (background: chapter 1). SP004, SP010, SP007, SP017
CP044 Because DeepSeek, Qwen, GLM, and Kimi all publish OpenAI-compatible or similarly documented APIs and several ship permissively licensed open weights, Chinese enterprise and developer buyers can multi-home or switch providers with comparatively low technical friction, limiting how much lock-in any single Chinese model vendor, including StepFun, can rely on. SP016, SP024
CP045 Independent commentary flags that Chinese LLM benchmark leaderboards have faced test-set leak and opacity concerns (e.g. disputed C-Eval results), and that developers may optimize for publicized metrics rather than disclosing training data or methodology, which limits confidence in any single Chinese provider's self-reported ranking, including StepFun's. SP026
CP046 StepFun's clearest remaining differentiation versus better-funded or better-distributed peers is its embedded multimodal device strategy (smartphone and automotive OEM deployment) and its explicit, still-active pursuit of AGI-oriented model scope, rather than consumer app scale, benchmark leadership, or disclosed capital access. SP024, SP027
CP047 DeepSeek's April 24, 2026 V4 release under an MIT license was followed by a rapid 75% promotional discount within 48 hours, a pattern independent coverage frames as reshaping the entire Chinese open-source pricing race that StepFun's own open-source models must be judged against. SP029
CP048 ByteDance's Doubao product surface (doubao.com) functions as an integrated consumer chat and search assistant rather than a narrow developer API, reinforcing ByteDance's distribution-first strategy versus StepFun's more API/device-centric go-to-market. SP023
CI001 StepFun's official open platform prices the step-3.7-flash multimodal reasoning API at RMB1.35 per 1M input tokens (cache miss), RMB0.27 per 1M input tokens (cache hit), and RMB8.1 per 1M output tokens. SI001
CI002 StepFun's step-3.5-flash reasoning model is priced at RMB0.7/RMB0.14/RMB2.1 per 1M tokens for input-cache-miss, input-cache-hit, and output respectively. SI001
CI003 StepFun's step-1o-turbo-vision model is priced at RMB2.5/RMB0.5/RMB8 per 1M tokens, with image inputs counted at a default of 169 tokens per image unless detail mode is enabled. SI001
CI004 StepFun publishes separate usage-based pricing for speech and audio models, including stepaudio-2.5-realtime at RMB10/RMB2/RMB70 per 1M tokens and step-tts-2 text-to-speech at RMB5.8 per 10,000 characters, plus value-added services such as internet search at RMB0.04 per call and file storage at RMB0.5 per GB per day. SI001
CI005 Independent pricing aggregator PricePerToken shows StepFun's cheapest tracked model at $0.09 per 1M input tokens for a January 2026 release, while the May 2026 flagship release is priced higher at $0.20 per 1M input and $1.15 per 1M output tokens. SI010
CI006 StepFun's models are listed on the independent OpenRouter developer marketplace alongside other frontier model providers, giving it a distribution channel beyond its own platform. SI009
CI007 In July 2025, StepFun vice president Li Jing said the company deliberately avoids two business models it considers unhealthy: highly customized project delivery, which is vulnerable to open-source substitution and linear delivery costs, and pure token-priced APIs, which large platforms can commoditize by bundling model access with cloud sales. SI018
CI008 StepFun instead frames its commercialization around productized to-B and to-C revenue with device partners across phones, cars, IoT, and embodied robotics, alongside vertical partnerships such as a dedicated finance-sector joint venture ('财跃星辰') and cooperation with Guotai Haitong Securities. SI018
CI009 As of July 2025, StepFun's smartphone Agent partnerships reportedly covered more than half of China's leading domestic phone brands. SI022, SI018
CI010 Independent reporting describes StepFun's commercialization as a 'B2B2C dual-drive' model in which enterprise services reportedly cover multiple Fortune 500 clients while AgentAPI call volumes have grown significantly, though neither figure is independently corroborated with company-disclosed data. SI002
CI011 ZTE's Nubia Z80 Ultra smartphone was reported as the first phone to natively embed StepFun's OpenClaw AI agent directly into the operating system rather than as a separate app, with the device starting at RMB4,999 (about $730). SI020
CI012 Tencent Cloud and StepFun signed a strategic cooperation agreement to build an 'AI Cockpit Agent' linking Tencent's music, video, maps, payment, and mobility services for in-vehicle scenarios, deepening a relationship that began with Tencent's participation in StepFun's Series B round. SI019, SI021
CI013 Reporting describes StepFun's Series B+ and Pre-IPO fundraising rounds as concentrated with consumer-electronics supply-chain investors -- Huaqin, Longcheer, OmniVision, and ZTE -- rather than purely financial investors. SI005, SI011
CI014 The concentration of hardware-supply-chain investors alongside a stated 'AI + devices' strategy implies StepFun's near-term revenue growth is closely tied to the shipment cycles and commercial terms of a small number of manufacturing partners. SI002, SI005
CI015 Caijing reported StepFun's 2025 revenue at roughly RMB500 million (about $73.3 million), with 2026 revenue projected at roughly RMB1.2 billion (about $175.9 million). SI003, SI002
CI016 In July 2025 StepFun's leadership publicly targeted roughly RMB1 billion of full-year 2025 revenue, citing several hundred million RMB of contracted revenue already booked in H1 2025; the roughly RMB500 million figure later reported for full-year 2025 implies the company fell meaningfully short of its own mid-year guidance. SI018, SI003
CI017 A February 2026 Caijing report described StepFun's Pre-IPO financing as two tranches: a first tranche priced at roughly a $4 billion pre-money valuation targeting an RMB2-3 billion raise, led by a telecom operator and expected to close by March 6, 2026, and a second tranche priced at a $5-6 billion pre-money valuation expected to close in mid-to-late April 2026. SI003
CI018 The same February 2026 report said StepFun planned to file for a Hong Kong listing by June 30, 2026 with expected cornerstone pricing near $10 billion and a year-end 2026 listing target, consistent with the $10-12 billion valuation range and June 9, 2026 filing date already covered in Company Overview. SI003
CI019 Internal documents cited by Sina Finance indicate StepFun could face a share lock-up expiry around mid-2027, a detail not previously covered in earlier reporting on the listing timeline. SI004
CI020 Comparative reporting places StepFun's revenue scale below already-listed peers: Zhipu AI reported RMB724 million of audited 2025 revenue (up 131.9% year-on-year) with a 41% gross margin, and MiniMax reported roughly RMB543 million of 2025 revenue, both disclosed through Hong Kong-listed annual reporting, versus StepFun's still-unaudited roughly RMB500 million figure. SI026, SI004
CI021 Zhipu AI's 2025 annual report disclosed a net loss of roughly RMB4.72 billion (about $686.5 million) alongside its revenue growth, showing that even the most comparable newly listed peer remains deeply loss-making despite far more revenue scale and audited disclosure than StepFun currently provides. SI026
CI022 Sina Finance reported MiniMax's 2025 adjusted net loss at roughly RMB1.75 billion and Zhipu's at roughly RMB3.18 billion (non-IFRS), with respective gross margins of 25.4% and 41%, providing an approximate loss-and-margin band against which StepFun's still-undisclosed figures could eventually be measured. SI004
CI023 Business-registration filings reviewed via Aiqicha and cited by Tencent News show Shanghai Jieyue Xingchen Intelligent Technology Co., Ltd. (StepFun) increased its registered capital from RMB23.9436 million to RMB56.2681 million in April 2026, coinciding with its conversion from a limited liability company to a joint-stock company. SI006, SI007
CI024 A further registered-capital increase to roughly RMB60.5199 million was approved around June 12, 2026, following an earlier step to RMB57.268 million in May 2026, indicating StepFun filed multiple onshore capital injections in quick succession while preparing for listing. SI004
CI025 A Hong Kong company-registry filing reviewed via QCC shows an entity named '上海階躍星辰智能科技股份有限公司' was registered in Hong Kong as a non-Hong Kong company on May 18, 2026, under business registration number 80414928, an administrative step consistent with preparing a Hong Kong listing. SI008
CI026 StepFun co-founded the 'Model-Chip Ecosystem Innovation Alliance' (模芯生态创新联盟) in July 2025 with roughly ten domestic chip and AI-infrastructure vendors, including Huawei Ascend, MetaX, Biren, Enflame, Iluvatar CoreX, Moore Threads, and Cambricon, aiming to reduce dependence on foreign GPUs. SI022, SI024
CI027 StepFun says its Step 3 model achieves inference throughput on domestic chips of up to roughly 3x that of DeepSeek-R1, and a throughput improvement of more than 70% over DeepSeek-R1 when run on Nvidia Hopper-architecture chips. SI018, SI022
CI028 As of June 2026, industry commentary summarized by Baidu Baike said the chip alliance's progress remained concentrated in hardware adaptation and basic integration testing, with limited progress on large-scale deployment or building the kind of broad developer ecosystem that competing, mostly Nvidia-based, platforms already have. SI024
CI029 Independent analyst data put NVIDIA's global data-center GPU annualized revenue run-rate at roughly $228 billion and top hyperscaler AI capex above $130 billion in a single quarter in early 2026, underscoring how capital-intensive frontier model training and inference remain industry-wide even for companies pursuing domestic-chip alternatives. SI015
CI030 TrendForce data reported that combined capex from eight major global cloud providers is projected to exceed $520 billion in 2026, a 24% year-on-year increase, as GPU procurement and custom AI ASIC development both scale. SI016
CI031 Analyst pricing trackers put a single Nvidia H100 GPU at $25,000-$40,000 and an early-production B200 at $30,000-$70,000 in 2026, with volume orders for the newest chips facing 12-18 month lead times. SI017
CI032 Deployments of 100 to 10,000 GPUs typically require $5 million to over $500 million in upfront capital, and companies increasingly use asset-backed financing, with the GPUs themselves as collateral, to accelerate procurement without immediately diluting equity. SI023
CI033 36Kr Research Institute estimates China's overall large-model market reached roughly RMB29.4 billion in 2024 and will exceed RMB70 billion by 2026, with the multimodal segment alone estimated at about RMB15.6 billion. SI025
CI034 No source reviewed in this chapter discloses StepFun's cash on hand, monthly cash burn, or runway; the only public financial signals are third-party-reported revenue estimates, leaving liquidity adequacy unverifiable from public information alone.
CI035 AsiaICT's analysis frames a valuation-to-revenue mismatch explicitly, noting that a roughly $10 billion targeted cornerstone valuation against a reported low-hundreds-of-millions-of-dollars revenue base implies a revenue multiple in the tens of times, an unusually rich multiple for a company that has not yet demonstrated a path to breakeven. SI002
CI036 Public sources reviewed here do not disclose StepFun's gross margin, cost of revenue, or unit economics by product line, making it impossible to independently verify whether its device- and enterprise-weighted revenue mix carries better or worse margins than the 25%-41% range disclosed by newly listed peers Zhipu and MiniMax.
CI037 No audited financial statement or filed HKEX prospectus for StepFun was available in the source set reviewed for this chapter as of the runDate, so all revenue, margin, and loss figures cited here remain third-party-reported rather than independently verified.
CI038 AsiaICT reported that StepFun had not filed a formal HKEX Application Proof (A1) as of May 20, 2026, despite widely reported plans to submit by June 2026, indicating some execution timing risk even before considering the CSRC red-chip unwind covered in Company Overview. SI002
CI039 An analyst quoted in Sina Finance argued that, absent a public listing, a private company's headline valuation has limited connection to realized revenue, and that model quality, API call volume, and financing continuity matter more to outside observers than the RMB500 million revenue figure itself. SI004
CI040 Using the 30%-60% device/OEM AI partnership revenue-sharing norm cited in adverse reporting, a meaningful share of StepFun's device-linked revenue may be shared with hardware partners before reaching StepFun's own income statement, a factor not separately quantified in any source reviewed. SI002
CI041 No source reviewed discloses a customer-acquisition cost, payback period, net revenue retention, or contract-level pricing for StepFun's enterprise or device deals, so unit economics for the core commercial motion remain a diligence gap rather than a quantified metric.
CI042 StepFun's Step 3 flagship model was reported in July 2025 as already licensed to and adapted for multiple domestic chip vendors ahead of the chip alliance's formal launch, suggesting some commercial value already flows from chip-adaptation work independent of end-customer API revenue. SI018
CE001 StepFun's official platform homepage and product blog headline Step 3.7 Flash as a production-oriented, high-efficiency agentic Flash model with native multimodal understanding, web/visual search enhancement, reliable tool orchestration, and agent-ecosystem compatibility with frameworks such as Claude Code, KiloCode, Hermes Agent, and OpenClaw. SE001, SE004
CE002 Step 3.7 Flash is a 198B-parameter sparse Mixture-of-Experts vision-language model combining a 196B-parameter language backbone with a 1.8B-parameter vision encoder, activating roughly 11B parameters per token via top-8-of-288 expert routing, with a 256,000-token context window and three selectable reasoning tiers (low, medium, high). SE003, SE013
CE003 Step 3.7 Flash delivers up to 400 tokens/second cloud throughput per StepFun, and independent measurement by Artificial Analysis recorded 415.9 tokens/second output throughput, while an independent DGX Spark deployment of the IQ4_XS local quantization measured only about 27 tokens/second. SE013
CE004 StepFun's official pricing documentation lists step-3.7-flash at RMB1.35 (cache-miss) / RMB0.27 (cache-hit) input and RMB8.1 output per million tokens, and step-3.5-flash at RMB0.70 / RMB0.14 input and RMB2.1 output per million tokens. SE002
CE005 Step 3.5 Flash is a 196B-parameter MoE model with roughly 11B active parameters and a 262,000-token context window, released February 2, 2026, and ranked #3 on OpenRouter's free-model leaderboard at 1.38 trillion weekly tokens shortly after release. SE022, SE021
CE006 Step-2 is StepFun's trillion-parameter Mixture-of-Experts LLM, officially launched at WAIC 2024 in July 2024 with a 16,000-token context window, and remains listed among the reasoning-model tiers in StepFun's current pricing documentation. SE017, SE002
CE007 Step-1V and Step-1.5V are StepFun's multimodal large models, with Step-1.5V launched alongside Step-2's official release at WAIC 2024, underpinning the more-than-40-million embedded smartphone AI features StepFun has publicly cited. SE017
CE008 StepFun's text-to-image and image-editing lineup (step-2x-large and step-image-edit-2) is billed per generated image at RMB0.1 and RMB0.02 respectively, with step-2x-large's limited-time free tier ending June 12, 2026. SE002
CE009 Step-Audio 2 is an end-to-end, open-source multimodal large language model for audio understanding and speech conversation that integrates a latent audio encoder, reasoning-centric reinforcement learning, and retrieval-augmented generation with external tool-calling (e.g., web search, audio search for timbre switching). SE010, SE006
CE010 The Step-Audio product family on StepFun's platform spans realtime conversational (stepaudio-2.5-realtime, stepaudio-2.5-chat), text-to-speech (stepaudio-2.5-tts, step-tts-2, step-tts-mini), voice-cloning, and automatic speech recognition (stepaudio-2.5-asr, step-asr, and streaming variants) tiers, each billed on distinct per-character, per-clone, or per-hour metering. SE002
CE011 Step-Video-T2V is a 30-billion-parameter text-to-video foundation model using a deep-compression Video-VAE (16x16 spatial / 8x temporal compression) and a Diffusion Transformer with 3D full attention trained via Flow Matching, open-sourced in February 2025 alongside an image-to-video variant, Step-Video-TI2V. SE011, SE007
CE012 StepFun's GitHub organization also maintains STEP-GUI (repository gelab-zero), described as a GUI-agent solution built by the StepFun-GELab team, and Step-Realtime-CLI/Console tooling for real-time voice and agent development, both with active 2026 commit history. SE009, SE005, SE026
CE013 Step 3.7 Flash's sparse MoE design routes each token through 8 of 288 experts per layer across 42 routed MoE layers of 45 total language layers, so roughly 5.6% of weights compute on any given token, per independent technical analysis of the release. SE013
CE014 Step-3, a separate 321-billion-parameter (38-billion-active) multimodal reasoning model, introduces Multi-Matrix Factorization Attention (MFA) and Attention-FFN Disaggregation (AFD), reducing KV-cache size and decoupling attention/FFN compute into specialized subsystems, achieving a claimed 4,039 tokens/second/GPU decoding throughput on Hopper GPUs versus 2,324 for DeepSeek-V3 in the same head-to-head test setup. SE012, SE023
CE015 StepFun co-founded the Model-Chip Ecosystem Innovation Alliance in July 2025 with roughly ten domestic chip and platform vendors (including Huawei Ascend, MetaX, Biren, Enflame, Iluvatar CoreX, Moore Threads, and Cambricon), adapting Step 3's Multi-Matrix Factorization Attention architecture to domestic accelerators including MetaX GPUs to reduce reliance on foreign chips. SE024
CE016 Step 3.7 Flash supports flexible cloud, data-center, and local deployment, running on NVIDIA DGX Station, AMD Ryzen AI Max+ 395 systems, and Mac Studio/MacBook Pro devices with at least 128GB unified memory, and integrates with vLLM, SGLang, Hugging Face Transformers, llama.cpp, the NVIDIA NeMo ecosystem (AutoModel, Megatron Core, Megatron Bridge), and NVIDIA NIM. SE004
CE017 StepFun's own Step 3.7 Flash benchmark disclosure mixes internally-reproduced ('vendor-stated') and officially-reported third-party figures within the same comparison chart, e.g. DeepSeek V4 Flash's Terminal-Bench 2.1 score comes from StepFun's own internal testing while Kimi K2.6, GPT 5.5, and Claude Opus 4.7 figures are those vendors' own reported Terminal-Bench 2.0 results, a different benchmark version. SE004
CE018 Step 3.7 Flash narrowed cross-harness performance variance from a 43-73% spread on Step 3.5 Flash to 64.5-71.5% on StepFun's in-house Step-SWE-Bench across coding harnesses including Hermes Agent, OpenClaw, Claude Code, KiloCode, OpenCode, and RooCode. SE013, SE004
CE019 On SWE-Bench Pro (a 1,865-task, four-language, contamination-mitigated benchmark), Step 3.7 Flash's vendor-stated 56.3% placed 15th of 35 tracked models and 2nd among non-Anthropic models at release, while its higher 76.5% SWE-Bench Verified score sits on a 500-task, Python-only benchmark that frontier labs including OpenAI have flagged for training-data contamination. SE013
CE020 Independent hands-on testing on an NVIDIA DGX Spark reported a 100% tool-call success rate for Step 3.7 Flash across a multi-step tool-calling evaluation suite, which the reviewer frames as consistent with -- but not a substitute for -- the model's first-place 67.1 ClawEval-1.1 score. SE016
CE021 Independent benchmark aggregator Crafiq ranks Step 3.7 Flash #108 overall on intelligence and #89 on coding among tracked models, a materially more modest position than the 'first place' / top-tier framing StepFun's own release material emphasizes for narrower agentic benchmarks like ClawEval-1.1. SE014
CE022 StepFun's Advisor-Mode cost claim -- roughly $0.19 per agentic task for Step 3.7 Flash versus about $1.76 for a larger advisor model, a roughly 9.3x difference -- is StepFun's own internal scaffold-based estimate rather than an independently audited comparison, per independent analyst review. SE013
CE023 Independent review coverage described Step 3.5 Flash as trading blows with GPT-5.2-class and Claude Opus 4.5-class closed models on selected benchmarks while running at a fraction of the compute, generating substantial Hacker News and local-inference community attention around Mac Studio deployment shortly after its February 2026 release. SE021
CE024 China's national AI-Generated Synthetic Content Labeling Measures (issued March 14, 2025 by the Cyberspace Administration of China, MIIT, Ministry of Public Security, and National Radio and Television Administration) require visible and technical (watermark/metadata) labeling of AI-generated text, image, audio, and video content on Chinese platforms -- an obligation that applies to StepFun's Step-Audio voice/TTS outputs and Step-1X/step-2x-large image generation, but StepFun's own product-level labeling or watermarking implementation is not independently documented in the public sources reviewed for this chapter. SE019
CE025 Chinese authorities (CNCERT/CC, MIIT, and the National Computer Virus Emergency Response Center) issued high-level 2026 warnings about vulnerabilities and Trojan-laced counterfeit skill packages on the OpenClaw agent platform -- the same agent ecosystem in which StepFun's own release material lists both Step 3.5 Flash and Step 3.7 Flash as compatible and highly-ranked -- and no StepFun-specific skill-verification or sandboxing control is disclosed in the public sources reviewed. SE018, SE004
CE026 A third-party crowd-sourced uptime monitor recorded StepFun operating normally as of July 1, 2026, with only 26 user-submitted problem reports over the trailing 30 days and no listed historical outage events, though this reflects automated/crowd monitoring rather than a StepFun-published SLA or incident postmortem. SE017
CE027 Step 3.7 Flash, Step 3.5 Flash, Step-Audio 2, Step-Video-T2V, and Step1X-Edit are all distributed under the Apache 2.0 license on GitHub and Hugging Face, permitting unrestricted commercial and enterprise use of the open-weight releases. SE003, SE006, SE007, SE008
CE028 Public sources reviewed for this chapter do not disclose a StepFun-specific SOC 2, ISO 27001, or other independent security/data-protection certification for the platform.stepfun.com API service, even though the platform's tiered rate-limiting (V0-V5) is gated on cumulative account recharge rather than any disclosed compliance tier.
CE029 Step 3.7 Flash is available at release on StepFun's own Open Platform (platform.stepfun.ai globally, platform.stepfun.com in China), OpenRouter, and NVIDIA NIM, with DeepInfra, Fireworks AI, and Modal availability announced as forthcoming. SE004
CE030 StepFun's standard pay-as-you-go rate-limiting tiers scale from V0 (RMB0 cumulative recharge; 5 concurrent requests, 10 RPM, 5,000,000 TPM) up to V5 (RMB200,000+ cumulative recharge; 100,000,000 TPM), explicitly separate from the subscription-based 'Step Plan' monthly-credit tier. SE002
CE031 StepFun's GitHub organization (contact opensource@stepfun.com) maintains at least 32 public repositories and had approximately 1,700 followers as of the access date, with flagship pinned repositories Step-3.5-Flash (2,100+ stars, 85 forks) and Step-3.7-Flash (257 stars, 22 forks shortly after its May 2026 release). SE005
CE032 Step1X-Edit, StepFun's open-source image-editing model targeting parity with GPT-4o- and Gemini 2 Flash-class closed editors, has accumulated 2,232 GitHub stars and 101 forks, while the STEP-GUI GUI-agent repository (gelab-zero) has 2,209 stars and 196 forks. SE008, SE009
CE033 StepFun's publicly-dated release cadence across 2025-2026 runs from the Step-Audio 2 technical report and Step-3 co-design paper (July 2025), through open-sourced Step-Audio 2 mini variants and a vLLM backend (August-September 2025), Step 3.5 Flash (February 2, 2026), and Step 3.7 Flash (May 29, 2026) -- roughly one major model-family release or update every one to three months. SE010, SE012, SE006, SE022, SE004
CE034 StepFun's own release material describes an 'emergent ability' in which Step 3.7 Flash combined visual and non-visual tools (independently calling Python crop/zoom/bounding-box utilities) inside multi-step visual-search workflows without being explicitly trained to do so -- a vendor-observed behavior that independent analysis characterizes as a promising but not independently reproduced signal. SE013
CE035 StepFun's consumer-facing distribution runs through the 阶跃AI (跃问 / Yuewen) chat app at chat.stepfun.com and embedded smartphone/automotive assistants, positioning the Step-series foundation models as the reasoning backend behind that consumer surface rather than as a separately branded consumer product. SE001, SE025
CE036 Minimum local-hardware requirements for Step 3.7 Flash exclude consumer laptops: even the smallest published quantization (Q3_K_M, 94GB) plus a 4GB vision-projector and OS overhead requires at least 128GB of unified/system memory, restricting genuinely local deployment to workstation-class hardware such as NVIDIA DGX Spark, Mac Studio M4 Max, or AMD Ryzen AI Max+ 395 systems. SE016
CE037 On search-augmented benchmarks, StepFun reports Step 3.7 Flash scoring 79.2 on SimpleVQA (with tool/search), 75.82% on BrowseComp, and 92.82% F1 on DeepSearchQA -- the latter StepFun frames as comparable to Kimi K2.6, a considerably larger 1-trillion-parameter/32-billion-active model. SE004
CE038 On enterprise/occupation-coverage testing, StepFun reports Step 3.7 Flash reaching 45.8% on GDPval across 44 occupations and passing at over 98% across reasoning-difficulty tiers on Tau2-bench Telecom. SE004
CE039 On Toolathlon (multi-tool coordination) and ClawEval-1.1 (real-world autonomous task execution), StepFun reports Step 3.7 Flash scoring 49.5% and 67.1% respectively, describing the ClawEval-1.1 score as first place among tracked peers as of the benchmark's May 2026 snapshot. SE004
CE040 Step-Audio 2's technical report claims state-of-the-art performance versus other open-source and commercial solutions on automatic speech recognition and paralinguistic (emotion/accent/style) understanding benchmarks, a comparison authored entirely by the StepFun team rather than an independent evaluator. SE010
CE041 PyPI hosts an installable step-audio2 package alongside the GitHub source release, giving Python developers a standard package-manager path (pip install step-audio2) to the model's inference code separate from manually cloning the repository. SE020
CE042 Step 3.7 Flash's Advisor Mode routes the bulk of an agent trajectory through the low-cost Step 3.7 Flash executor and escalates only to a larger advisor model at planning or failure-recovery inflection points, an escalation pattern StepFun's release material attributes to an advisor strategy described by Anthropic. SE004
CE043 Step-Video-T2V generates videos up to 204 frames using bilingual (English/Chinese) text encoders and a video-based Direct Preference Optimization (Video-DPO) step intended to reduce visual artifacts, evaluated on StepFun's own Step-Video-T2V-Eval benchmark rather than a third-party video-generation leaderboard. SE011
CE044 As of June 2026, industry commentary summarized by Baidu Baike characterized the Model-Chip Ecosystem Innovation Alliance's progress as still concentrated in hardware adaptation and basic integration testing rather than large-scale production deployment or a broad developer migration. SE024
CE045 Step 3.7 Flash is also listed on OpenRouter's model marketplace, though the page's primary content renders client-side and did not yield extractable pricing/benchmark text during this review, illustrating an access gap in independently verifying third-party marketplace claims. SE015
CU001 StepFun's Yuewen (跃问) Android app on Google Play has a 4.2-star rating from 58 reviews and offers chat, document understanding, and a 'StepClaw' task-execution agent as of its May 2026 update. SU001
CU002 Multiple named Yuewen Android users publicly complain that the app only accepts phone-number login and blocks Google, email, or social sign-in, with one reviewer reporting they can no longer chat or copy text without providing a phone number. SU001
CU003 StepFun's open-weight models are hosted under the stepfun-ai Hugging Face organization, where individual model variants such as Step-3.7-Flash-NVFP4 show cumulative download counts in the hundreds of thousands. SU002
CU004 The stepfun-ai GitHub organization actively maintains multiple public repositories (including Step-3.7-Flash, Step-Realtime-CLI, Step-Realtime-Console, and a vLLM fork) with ongoing commit activity and star/fork counts as of June 2026. SU003
CU005 StepFun's roughly $2.5 billion pre-IPO financing round is led by consumer-electronics supply-chain investors Huaqin Technology, Longcheer Technology, OmniVision Group, and ZTE, alongside the Hong Kong Investment Corporation as the sole state fund participant. SU004
CU006 A January 2026 StepFun financing announcement ties the more-than-RMB5-billion B+ round explicitly to accelerating its 'AI+terminal' (AI+终端) strategy across smartphones and vehicles. SU005
CU007 Geely Auto Group publicly showcased its strategic technology partnership with StepFun at WAIC 2025, jointly exhibiting the Geely Galaxy M9 with an industry-first 'Human-like AI agent' cockpit feature. SU006, SU011
CU008 Orders for the Geely Galaxy M9 -- the flagship vehicle carrying the StepFun-linked AgentOS cockpit -- surpassed 40,000 within 24 hours of pre-sales opening in August 2025, per ECARX's press release on the cockpit platform powering the launch. SU007
CU009 Huanxi Media Group, a Hong Kong-listed film and TV company, entered a strategic framework agreement with Shanghai Stepfun Intelligent Technology Co., Ltd. on 11 November 2025 to form a joint venture applying large-model AI to film, TV, interactive entertainment, and game development. SU009, SU008
CU010 Law firm Wilson Sonsini Goodrich & Rosati confirms it advised Huanxi Media on the StepFun strategic-cooperation agreement, corroborating that the counterparty is 'Shanghai Stepfun Intelligent Technology Co., Ltd.' SU008, SU027
CU011 The Huanxi Media-StepFun joint venture's disclosed scope covers data cooperation, joint model/agent R&D, and co-production of AI-native film, interactive entertainment, and game content, with capital contribution terms still to be negotiated as of the November 2025 filing. SU027
CU012 ZTE's Nubia Z80 Ultra became the first smartphone to natively embed the OpenClaw AI agent -- a framework StepFun's own release material lists as a supported, highly-ranked integration -- directly into the operating system rather than as a separate app. SU010
CU013 Independent coverage of the Nubia Z80 Ultra's native OpenClaw integration flags unresolved security questions around embedding a system-level AI agent with broad device permissions, even as Nubia opened a public beta test of the feature. SU010
CU014 Geely Auto Group, Qianli Technology, and StepFun jointly launched a preview version of the 'Agent OS' AI-native intelligent cockpit at WAIC 2025, built on StepFun's multimodal and end-to-end voice foundation models. SU011, SU006
CU015 Qianli Technology's (千里科技, formerly Lifan) IPO prospectus discloses that its top five customers accounted for 41.5%-48.1% of total revenue between 2022 and mid-2025, with Geely Group alone contributing 30.8%-39.7% of revenue in the same period. SU013
CU016 Qianli Technology's prospectus also shows the company posted a net loss in 2023-2025 (net margin falling from +2.0% in 2022 to as low as -4.7% in 2024), meaning StepFun's primary automotive distribution partner is itself commercially fragile and Geely-dependent. SU013
CU017 In June 2026, Sina Finance reported that Qianli Technology's intelligent-driving subsidiary agreed to acquire a related-party radar-sensing company at roughly double its net-asset value without any performance-compensation clause, while total related-party transaction limits for 2026 were raised over 80% year-on-year to RMB21.885 billion. SU012
CU018 An independent business profile of StepFun chairman Yin Qi describes StepFun and Qianli Technology (the AgentOS automotive partner) as sibling ventures both anchored to Geely, raising the question of how independent StepFun's automotive commercial terms are from a single controlling group. SU025
CU019 The Wire China's independent profile of StepFun describes the company's late-2025 wave of image- and video-generation tools as consumer-facing releases distributed through its own apps rather than through third-party enterprise integrations. SU014
CU020 AICPB's April 2026 China AI app ranking by monthly active users lists 27+ distinct apps with disclosed MAU figures (from ByteDance's Doubao at 336.04 million down to under 1 million for the smallest listed app), and StepFun's Yuewen does not appear among the ranked apps. SU015
CU021 A January 2026 Sohu report states StepFun's AI models are integrated into more than 42 million smartphones (spanning OPPO, Honor, and ZTE) and that its AgentOS smart cockpit, co-built with Qianli Technology, is already deployed in Geely Galaxy M9 production vehicles. SU016
CU022 Tencent Cloud and StepFun announced a strategic partnership to jointly develop AI cockpit agents that integrate StepFun's foundation models with Tencent's content, payments, maps, and entertainment ecosystem for in-vehicle use. SU021, SU017
CU023 Tencent has been a StepFun shareholder since its Series B round and continued participating in subsequent financing including the roughly $2.5 billion pre-IPO round, making Tencent simultaneously an investor and a named commercial automotive-AI partner. SU021
CU024 科创板日报 reported that OPPO's flagship Find X8 Ultra's network-first 'one-tap memory capture' (一键闪记) feature -- which recognizes on-screen content, summarizes it, and files it into the phone's memory app -- is powered by StepFun's multimodal technology, alongside two earlier OPPO AI features StepFun also supported. SU018
CU025 Honor awarded StepFun a 'Best Collaboration Award' at its October 2025 global developer conference for powering the YOYO agent's multimodal memory feature in the Honor Magic8 flagship series, which lets the assistant recall fragmented daily information such as schedules, IDs, and addresses. SU019
CU026 Sina Technology reported that StepFun open-sourced its GELab-Zero GUI-agent suite in late November 2025 and disclosed partnerships covering more than 60% of China's domestic smartphone brands, including Honor, OPPO, and ZTE. SU020
CU027 StepFun co-launched an elder-friendly 'senior AI assistant' (长辈AI助手) with ZTE in July 2025, adding accessibility features that let older users issue voice commands for tasks such as hospital registration and social-insurance payment. SU020
CU028 At StepFun's February 2025 'Step Up' ecosystem open day, the company announced deepened cooperation with Geely Auto Group, OPPO, and embodied-AI robotics maker Zhiyuan Robotics (智元机器人), alongside a second phase of its 'Starlight Program' (繁星计划) to fund and support third-party Agent-application developers. SU022
CU029 The Starlight Program is StepFun's structured developer-enablement mechanism, offering participating third-party builders access to models, compute, capital, and data as they build Agent applications on StepFun's platform. SU022
CU030 A January 2026 Xinmin Evening News report states StepFun's terminal Agent API call volume grew nearly 170% quarter-on-quarter for three consecutive quarters through late 2025, alongside a device install base of more than 42 million smartphones serving an estimated 20 million daily users. SU023
CU031 The same January 2026 report states the Geely Galaxy M9 sold nearly 40,000 units within three months of launch and that StepFun expects its models to be deployed ('上车') in more than one million vehicles during 2026, alongside an overseas expansion of the AgentOS cockpit. SU023
CU032 StepFun's official Yuewen product site states the app added strategic cooperation with Geely Auto and Shanghai United Media Group (上海报业集团) in March 2025, extending its multimodal capabilities into intelligent-driving and financial-news verticals beyond phones and cars. SU024
CU033 StepFun's Step 3.7 Flash was one of five major new models OpenRouter listed in a ten-day window ending June 4, 2026, alongside Claude Opus 4.8, MiniMax M3, Qwen3.7-Plus, and NVIDIA Nemotron 3 Ultra, indicating continued active third-party marketplace distribution in mid-2026. SU026
CU034 No source reviewed for this chapter discloses a total paying-customer count, API-account count, or per-customer/ARPU revenue figure for StepFun's API, device-licensing, or enterprise business.
CU035 No source reviewed for this chapter discloses net revenue retention, gross revenue retention, renewal rate, or churn for any StepFun customer segment (device, automotive, developer, or enterprise).
CU036 No source reviewed for this chapter discloses contract length, renewal terms, or procurement cycle detail for any specific StepFun OEM, automotive, or enterprise customer relationship.
CU037 StepFun's disclosed customer base concentrates on a small number of device and automotive relationships (three named phone OEMs, one primary automotive group, one hyperscaler cockpit partner) rather than a broad, independently-verifiable enterprise or developer customer roster. SU016, SU020, SU023
CU038 Several of StepFun's largest disclosed financing and distribution counterparties overlap: Tencent is both a multi-round shareholder and the named partner for AI cockpit agents, while Huaqin, Longcheer, OmniVision, and ZTE are simultaneously financing-round investors and the OEM/device customers StepFun cites as adoption proof. SU004, SU021
CU039 Because Qianli Technology -- StepFun's primary automotive distribution vehicle -- derives roughly a third to just under half of its own revenue from a single customer group (Geely) and is not yet consistently profitable, a slowdown or renegotiation at Geely would propagate financial pressure onto the AgentOS channel StepFun depends on for auto-sector proof. SU013, SU012
CU040 StepFun's device-partner customer proof rests on named product features (OPPO Find X8 Ultra, Honor Magic8, ZTE elder-assistant) rather than disclosed unit volumes, revenue, or renewal terms per OEM, so the 42-million-device and 60%-brand-share figures describe model install footprint, not verified commercial scale per account. SU018, SU019, SU020, SU023
CU041 StepFun's own ecosystem messaging (Step Up day, GELab-Zero launch, Yuewen product timeline) is the primary source for its developer-enablement and vertical-partnership claims, meaning several adoption claims in this chapter are company-disclosed rather than independently audited. SU022, SU024, SU020
CU042 Independent app-ranking data (AICPB) and independent app-store review data (Google Play) both show StepFun's own consumer app trailing in the market that its device-embedded distribution is supposed to feed, suggesting the C-end 'Yuewen' brand has not converted device-level reach into an independently measurable consumer audience. SU015, SU001
CU043 StepFun's adoption funnel runs from free open-weight or low-tier API access (Hugging Face, GitHub, developer console) through paid API tiers gated by cumulative recharge, into negotiated OEM/device or automotive integration agreements, and finally into named production features (OPPO, Honor, ZTE, Geely/Qianli) that StepFun cites as commercial proof, without a disclosed self-serve-to-enterprise conversion metric at any stage. SU002, SU003, SU018, SU019, SU023
CR001 StepFun is unwinding its offshore Cayman Islands 'red-chip' incorporation structure to pave the way for a planned Hong Kong IPO, per Reuters reporting. SR012, SR013
CR002 China's securities regulator (CSRC) said in March 2026 that it had instructed some offshore-registered red-chip companies to unwind their structures, citing low shareholder transparency and elevated compliance risk. SR012, SR015
CR003 Two sources told Reuters that StepFun judged an onshore corporate structure more appropriate because it is heavily backed by state capital, including Shanghai municipal and district government investment vehicles. SR012, SR013
CR004 StepFun completed its conversion into a joint-stock company around April 2, 2026, renaming to '上海阶跃星辰智能科技股份有限公司', with Qianli Technology chairman Yin Qi formally registered as StepFun's chairman. SR014, SR015
CR005 21st Century Business Herald reported StepFun planned to file its Hong Kong listing application by June 30, 2026, targeting roughly a $10 billion cornerstone valuation. SR014
CR006 Under China's Trial Measures effective March 31, 2023, indirect overseas listings via red-chip structures are subject to CSRC filing-based administration, requiring a filing within three working days of the overseas listing application and a 20-working-day CSRC review window. SR004, SR005
CR007 Law.asia reported that regulators had advised several companies planning Hong Kong listings to dismantle their red-chip structures, and that the CSRC responded that such notifications were part of routine regulatory requirements focused on shareholder-transparency and compliance-risk concerns. SR005
CR008 Analysts note that unwinding a red-chip structure can be legally complex and cost-prohibitive, creating real risk that affected companies delay or abandon their IPO plans. SR005, SR013
CR009 China's amended Cybersecurity Law, adopted October 28, 2025 and effective January 1, 2026, introduces a tiered penalty regime, adding new provisions supporting AI research while requiring AI ethics governance and enhanced risk monitoring. SR001, SR009
CR010 Under the amended Cybersecurity Law, critical-information-infrastructure operators face fines from RMB 500,000 up to RMB 10 million for serious cybersecurity-obligation failures, escalating further for violations causing major security harm. SR001, SR008
CR011 Hunton Andrews Kurth's analysis states the amended Cybersecurity Law's penalty restructuring is its primary focus, aligning liability provisions with China's Personal Information Protection Law and Data Security Law. SR008
CR012 China's AI-generated-content Labeling Measures and the mandatory GB 45438-2025 standard took effect September 1, 2025, imposing dual explicit/implicit labeling obligations on internet information service providers and content-distribution platforms. SR007, SR010
CR013 Linklaters reported that major Chinese platforms including Tencent, Douyin, RedNote, Weibo, and DeepSeek had already made public compliance statements on the AI-content labeling measures by their September 2025 effective date; no such public statement from StepFun was found. SR010
CR014 In March 2026, HKEX published a Phase 1 consultation proposing to lower the weighted-voting-rights (WVR) listing thresholds under Chapter 8A, including cutting the Test A market-capitalisation bar from HK$40 billion to HK$20 billion. SR006, SR023
CR015 RTHK reported the HKEX WVR reform would also expand the definition of 'innovative company' to cover business-model-driven applicants, not only technology-driven ones, and remained a proposal (not yet adopted) as of the March 2026 consultation. SR023
CR016 On May 31, 2026 the U.S. Bureau of Industry and Security issued guidance closing a loophole that had let China-headquartered firms buy advanced Nvidia/AMD chips through overseas subsidiaries in jurisdictions such as Singapore, Malaysia, and Taiwan without an export license. SR016, SR002
CR017 The closed loophole rests on an 'ultimate parent company' test embedded in the Export Administration Regulations since November 2023 (EAR §742.6(a)(6)(iii)(A)), creating retroactive compliance exposure for transactions dating back to that period. SR016
CR018 Tech Times reported that Trump administration officials were internally divided in June 2026 over whether the closed loophole had been an inadvertent policy gap, following a May 2025 decision to stop enforcing the Biden-era AI Diffusion Rule. SR016
CR019 BIS's June 2026 Entity List action added dozens of Chinese entities for allegedly developing advanced AI, supercomputing, and high-performance AI chips with ties to China's military-industrial complex. SR002
CR020 CNBC reported that as of June 17, 2026 the U.S. had held off formally blacklisting China's DeepSeek, memory-chipmaker CXMT, and more than 100 other companies already flagged as national-security risks by an interagency committee, to avoid escalating tensions with Beijing. SR017
CR021 CNBC reported the Pentagon's updated '1260H list' published June 9, 2026 added Alibaba, Baidu, and BYD to its roster of China military-linked companies, barring direct Defense Department contracting with listed firms. SR018
CR022 No public source reviewed in this chapter names StepFun directly on the U.S. Entity List or the Pentagon's 1260H list as of 2026-07-01, but the rapid expansion of both lists to other Chinese AI and tech companies signals adjacency risk for state-backed AI unicorns generally. SR017, SR018
CR023 StepFun co-founded the Model-Chip Ecosystem Innovation Alliance in July 2025 with roughly ten domestic chip and AI-infrastructure vendors including Huawei Ascend, MetaX, Biren, Enflame, Iluvatar CoreX, and Moore Threads, aiming to reduce reliance on export-restricted chips. SR025
CR024 Huawei's Ascend 950PR, launched around April 2026, delivers roughly 2.8x the FP4 inference throughput of Nvidia's export-approved H20 chip, but draws about 200W more power (600W vs. H20's roughly 400W). SR025
CR025 Independent analysis finds Huawei's Ascend line still trails Nvidia's Blackwell-class chips on per-chip efficiency, memory-stack density, and CUDA-ecosystem software maturity, relying instead on cluster-scale aggregation (e.g., a 384-chip CloudMatrix cluster) to match or exceed Nvidia's raw cluster throughput at nearly 4x the power draw. SR026
CR026 Deploying a 100-to-10,000 GPU training cluster in 2026 requires upfront capital ranging from roughly $5 million to over $500 million, with hardware alone representing only about 35% of total five-year cost of ownership once networking, power, cooling, and staffing are included. SR027, SR028
CR027 Industry analysis notes that neglecting non-hardware cost buckets (storage, networking, engineering overhead) commonly causes AI infrastructure budgets to overrun by 40-100% versus initial hardware-only estimates. SR028
CR028 KrASIA reported China's 'AI tigers' cohort -- 01.AI, Baichuan AI, MiniMax, Moonshot AI, StepFun, and Zhipu -- has begun moving toward public listings as private funding tightens, with Zhipu and MiniMax already trading in Hong Kong as of early 2026. SR021
CR029 Zhipu's on-premise revenue more than doubled in 2025 and cloud revenue rose 292%, reaching total 2025 revenue of RMB 724.3 million, up from RMB 312.4 million in 2024, a materially larger and more disclosed revenue base than StepFun's reported roughly RMB 500 million. SR021
CR030 A JPMorgan sector note (February 9, 2026) argues China's AI-model-developer field has consolidated from more than 200 viable developers to fewer than ten, with long-term winners determined by deployment scale and monetization efficiency rather than model-release cadence. SR022
CR031 JPMorgan initiated coverage on Zhipu AI and MiniMax with 'Overweight' ratings but did not extend equivalent public analyst coverage to StepFun, consistent with StepFun's still-private, less-disclosed status relative to its listed peers. SR022
CR032 ByteDance alone plans to spend more than RMB 160 billion (about $23.4 billion) on AI-related procurement in 2026, and Tencent's 2025 capital expenditure reached RMB 79.2 billion, an order of magnitude beyond the scale StepFun has disclosed for its own infrastructure spending. SR021
CR033 Qianli Technology's 2025 annual report shows sales to Geely-affiliated companies at 29.26% of total 2025 revenue, with the ratio ranging 30.8%-39.7% across 2022-2025, while purchases from Geely-affiliated suppliers reached as high as 50.4% in 2022 before easing to about 29.7% in H1 2025. SR029
CR034 On December 31, 2025, Qianli Technology's shareholders approved daily related-party-transaction limits with Geely-affiliated entities for both 2025 and 2026, with over 99% of voting shares in favor. SR019, SR003
CR035 A June 2025 cninfo filing shows Qianli Technology, Geely (via 浙江吉润汽车), Megvii-linked Yin Qi entities, and Chongqing state investment funds jointly structuring an intelligent-driving joint venture (重庆千里智驾有限公司) with RMB 1.5 billion committed capital, deepening the Qianli-Geely capital relationship beyond ordinary trading. SR003
CR036 Qianli Technology's core (non-recurring-items-excluded) net profit was negative for a third consecutive year in 2025 (roughly -RMB 255 million) despite 42.1% revenue growth to nearly RMB 10 billion, with the parent company's cumulative unrecovered losses reaching about RMB 948 million. SR029
CR037 Qianli Technology's core automotive-business gross margin was only 2.8% in 2025 (improved from -5.3% in 2023), while its newer intelligent-driving technology segment carried a 30.8% gross margin but contributed only 3.5% of total revenue. SR029
CR038 36Kr reported Qianli Technology -- formerly bankrupt motorcycle/fuel-vehicle maker Lifan -- was restructured in 2020 under Chongqing municipal government and Geely guidance, with Megvii co-founder Yin Qi leading its transformation into an 'AI + Mobility' intelligent-driving solutions provider. SR020
CR039 OPPO's public compliance statement commits the company to comply with local laws and regulations across the more than 60 countries where it operates, but discloses no StepFun-specific service-level or data-handling terms for their embedded-AI partnership. SR011
CR040 hubpy.io's 2026 profile of StepFun confirms partnerships with device makers Honor, OPPO, and ZTE for embedding StepFun's foundation models, consistent with the roughly 42-million-device distribution figure reported elsewhere, but adds no independent verification of contract terms. SR030
CR041 StepFun's chairman, Yin Qi, concurrently serves as chairman of Qianli Technology, the Geely-linked automotive supplier that is StepFun's primary automotive distribution counterparty, creating an interlocking-directorate structure rather than an arm's-length partnership. SR014, SR020
CR042 No audited 2025 financial statements or filed HKEX prospectus for StepFun were retrievable as of 2026-07-01, leaving its reported roughly RMB 500 million 2025 revenue and targeted $10-12 billion valuation without an independently audited baseline. SR014
CR043 StepFun's investor roster reported by Reuters includes Shanghai municipal and district government investment vehicles, Qiming Venture Partners, and Tencent Holdings, while its 2026 pre-IPO round separately drew participation from Hong Kong Investment Corporation (HKIC), a fund wholly owned by the Hong Kong government with an HK$62 billion mandate. SR012, SR032
CR044 HKIC's own disclosure describes itself as 'Patient Capital' seeking both financial return and enhancement of Hong Kong's long-term competitiveness, meaning its investment in StepFun carries an explicit public-policy objective alongside a commercial one. SR032
CR045 No public litigation, data-breach disclosure, or class-action record naming StepFun was found across the sources reviewed for this chapter as of 2026-07-01.
CR046 Independent analysis of 2.8 million LMArena model-comparison records found selective model submissions inflated leaderboard scores by up to 100 points through cherry-picking, with major labs submitting only their best-performing private-test variants. SR024
CR047 The benchmark-gaming pattern documented industry-wide is not StepFun-specific, but no independent, StepFun-authorized reproduction of its flagship agentic-benchmark claims was found among the sources reviewed for this chapter. SR024
CR048 China's Cyberspace Administration circulated a draft in April 2026 proposing additional mandatory ethics review, emergency procedures, and content-governance obligations specifically for AI services that simulate human interaction, with an expected effective date around July 2026. SR010
CR049 No independent security certification, bug-bounty disclosure, or public incident-response record for StepFun's platform, Step-Audio, or Step-Video was found among the sources reviewed for this chapter.
CR050 The amended Cybersecurity Law's penalty escalation for network-product/service providers (including malicious-program deployment and delayed vulnerability remediation) applies extraterritorially to any actions found to endanger China's cybersecurity, extending compliance exposure beyond StepFun's domestic operations to its overseas platform access. SR008, SR009
CV001 StepFun completed a share reform from limited liability company to joint stock company in late April 2026 and dismantled its offshore red-chip structure, a standard prerequisite for a Hong Kong listing. SV002, SV003
CV002 StepFun's nearly $2.5 billion pre-IPO funding round drew industrial investors Huaqin Technology, Longcheer Technology, OmniVision Technologies, and ZTE Corporation alongside the Hong Kong Investment Corporation (HKIC). SV001, SV002
CV003 The Hong Kong Investment Corporation's stake in StepFun is HKIC's only disclosed investment to date in a generative-AI foundation-model company. SV001
CV004 A February 2026 report described StepFun's pre-IPO financing as a two-tranche round with a first-tranche pre-money valuation of approximately $4 billion. SV005
CV005 Critics cited by The AI Chronicle argue that StepFun's roughly $12 billion IPO target valuation is inflated by 'national champion' sentiment and state-backed investment vehicles rather than by disclosed fundamentals. SV006
CV006 Analysts cited by the South China Morning Post, as reported by The AI Chronicle, warn that China's AI startup sector is approaching a valuation 'correction' because many billion-dollar valuations are not yet justified by revenue or profit. SV007
CV007 StepFun sits within a broader 'Hundred Model War' cohort where venture capital has funded dozens of Chinese LLM startups promising to become 'the OpenAI of China,' a saturation dynamic independent commentary ties to unsustainably high valuations. SV007
CV008 Wikipedia's Six AI Tigers entry lists StepFun alongside Z.ai (Zhipu AI), Moonshot AI, MiniMax, Baichuan AI, and 01.AI as the cohort of China-based foundation-model startups that rose to prominence between 2021 and 2024, and states the group's combined valuation was estimated to exceed RMB 100 billion by mid-2024. SV027
CV009 A second KrASIA analysis frames 2026 as the year public-market investors gained the ability to price China's AI-tiger cohort directly, following Zhipu's and MiniMax's Hong Kong listings. SV029
CV010 Zhipu AI (Z.ai, HKEX: 02513) listed on the Hong Kong Stock Exchange on January 8, 2026 at an issue price of HK$116.20, becoming the world's first publicly listed large language model company. SV015, SV013
CV011 Zhipu AI's FY2025 revenue reached RMB724 million, up 131.9%-132% year over year, alongside a net loss of RMB4.72 billion driven by R&D costs reported at roughly 4.4 times total revenue. SV012, SV014
CV012 According to 36Kr's analysis, Zhipu's market value of HK$495.8 billion against RMB724 million (about HK$780 million) of 2025 revenue implies a static price-to-sales ratio of about 576 times and an enterprise-value-to-revenue ratio of about 581 times. SV013
CV013 36Kr's analysis states that a price-to-sales ratio of 20 to 40 times is already considered an extreme valuation for top-tier global SaaS companies at peak growth, framing Zhipu's ~576x multiple as a further outlier even against that stretched benchmark. SV013
CV014 China Daily Brief reports Zhipu's market capitalization swung from about HK$57.8 billion to roughly HK$460 billion (about $59 billion) within three months of 2026, and 36Kr separately describes an interim peak near HK$880 billion before a sharp correction. SV012, SV013
CV015 Per 36Kr, only 84.8% of Zhipu's first-half-2025 revenue came from localized on-premises model deployment versus 15.2% from cloud/API calls, with mainland Chinese customers contributing 88.4% of the localized-deployment total. SV013
CV016 36Kr attributes part of Zhipu's June 2026 share-price resilience to its June 8, 2026 inclusion in the Hang Seng Tech Index and Stock Connect, which analysts at Morgan Stanley and JPMorgan estimated could force $1.25 billion to $4 billion of passive index-fund buying. SV013
CV017 36Kr reports that Zhipu's cornerstone shares unlocking on July 8, 2026 represent only 11.6% of Hong Kong-listed shares and are held mostly by patient, state-linked capital, versus MiniMax's July 9, 2026 unlock of 63% of its Hong Kong-listed shares, over a third held by early-stage VC/PE investors with an incentive to sell. SV013
CV018 China Daily Brief reports Zhipu's CEO has framed heavy R&D spending as an investment toward an Anthropic-like trajectory, targeting roughly $1 billion of annual recurring revenue by late 2026 after raising coding-plan prices about 30% following the GLM-5 launch. SV012
CV019 ChinaBiz Insider reports Zhipu's subscription plans have been repriced three times within a year and require daily queuing, alongside persistent overload on Kimi's and MiniMax's APIs, evidence cited as a compute-rationing signal rather than pure demand strength. SV008
CV020 MiniMax Group (HKEX: 0100) is a Cayman Islands-incorporated, weighted-voting-rights company that completed a Hong Kong global offering and listed on January 9, 2026, per its HKEXnews prospectus. SV018, SV016
CV021 MiniMax reported FY2025 revenue of $79 million, up 158.9% year over year from $30.5 million in 2024, with gross margin improving from 12.2% to 25.4% over the same period. SV031
CV022 MiniMax's FY2025 annual report showed a headline net loss of $1.87 billion, up 302% year over year, driven primarily by a roughly $1.59 billion non-cash fair-value loss on convertible preferred shares tied to its pre-IPO valuation step-up. SV031
CV023 Excluding the non-cash preferred-share remeasurement, MiniMax's adjusted (non-IFRS) net loss was about $251 million in 2025, only modestly wider than the $244 million adjusted loss reported for 2024. SV031
CV024 StockAnalysis.com put MiniMax's market capitalization at HKD130.79 billion (about $16.8 billion at prevailing HKD/USD rates) as of July 1, 2026, at a share price of HKD417.00, versus a 52-week high near HKD1,330. SV016, SV017
CV025 MiniMax's July 1, 2026 market value implies its shares have fallen more than two-thirds from their 2026 52-week high, and 36Kr separately reports 63% of its Hong Kong-listed shares unlock on July 9, 2026, more than a third of which sit with early-stage financial investors. SV016, SV013
CV026 Moonshot AI's post-money valuation rose from $4.3 billion at the end of 2025 to about $10 billion in early 2026 and doubled again to $20 billion in a $2 billion round led by Meituan's Long-Z Investment in May 2026. SV022, SV009
CV027 Moonshot AI's annualized recurring revenue topped $200 million in April 2026, implying its $20 billion valuation traded at roughly 100 times ARR at the round's close. SV022
CV028 Per AI Chat Daily, Moonshot AI raised a cumulative $3.9 billion over the six months preceding its May 2026 round, with backers including Alibaba, Tencent, HongShan, ZhenFund, IDG Capital, Tsinghua Capital, China Mobile, and CPE Yuanfeng. SV022
CV029 DeepSeek closed its first external funding round in June 2026, raising more than CNY 50 billion (about $7.4 billion) at a valuation exceeding CNY 330 billion (over $50 billion). SV019, SV020
CV030 DeepSeek's funding structure routes most outside capital through a limited partnership controlled by founder Liang Wenfeng, with investors subject to a five-year lock-up and no voting rights, except China's National Artificial Intelligence Industry Investment Fund, which invested directly and retains both voting rights and freedom from the lock-up. SV020
CV031 Reporting on DeepSeek's round places Anthropic's contemporaneous valuation at $965 billion following a $65 billion raise, and OpenAI's March 2026 valuation at $852 billion, giving frontier-lab reference points well above any disclosed Chinese-lab valuation in this comp set. SV019
CV032 Baichuan Intelligence raised a $300 million Series A1 round in October 2023 and a roughly RMB5 billion (about $691 million) Series A round in July 2024 from backers including Alibaba, Tencent, and Xiaomi, before pivoting in 2025 to a healthcare-AI vertical strategy. SV021
CV033 Baichuan's April 2026 Series B round reportedly kept its valuation near the roughly $2.7-2.8 billion level established at its prior round, materially below the disclosed 2026 valuations of Moonshot, DeepSeek, Zhipu, and MiniMax. SV021
CV034 Palantir traded at a market capitalization of about $279.69 billion against trailing-twelve-month revenue of about $5.22 billion as of July 2026, implying a price-to-sales ratio of about 53.5 times. SV024, SV025, SV026
CV035 Palantir's ~53.5x public-market revenue multiple sits far below Zhipu's ~576x static price-to-sales ratio, illustrating how much of the Hong Kong-listed Chinese AI cohort's pricing depends on scarcity and index-flow mechanics rather than a revenue multiple public investors would accept for an established AI-software platform. SV026, SV013
CV036 At StepFun's targeted $10-12 billion IPO valuation against its reported roughly $73.3 million of 2025 revenue, the implied trailing revenue multiple is approximately 137-164 times; against the reported roughly $175.9 million 2026 revenue projection, the implied forward multiple is approximately 57-68 times. SV001, SV006
CV037 StepFun's implied 57-68x forward revenue multiple sits below Zhipu's ~576x static multiple and below MiniMax's roughly 212x market-cap-to-revenue ratio (HKD130.79 billion market cap versus $79 million revenue), but remains far above Palantir's ~53.5x public-market benchmark and above the 20-40x ceiling 36Kr cites for extreme SaaS valuations. SV016, SV013, SV026
CV038 No public source reviewed for this chapter discloses an audited prospectus, cash position, burn rate, or gross margin for StepFun, so the valuation-support conclusions above rest on third-party-reported revenue and financing figures rather than filed financial statements. SV006
CV039 Whether StepFun's Hong Kong IPO ultimately prices toward the $10 billion or $12 billion end of its targeted range, and whether the roughly $500 million raise size holds, remained undecided as of the July 1, 2026 run date. SV001
CV040 An academic multi-method evaluation posted to arXiv in May 2026 frames the debate over AI-sector pricing as a question of whether markets are capitalizing a durable general-purpose technology or inflating a financial bubble, rather than treating either conclusion as settled. SV010
CV041 ChinaBiz Insider reports that Yann LeCun told CNBC on June 18, 2026 that an AI-sector bubble burst is not a distant scenario, arguing that advanced-AI product prices have risen while per-token production costs have not fallen fast enough to close the subsidy gap. SV008
CV042 Chinese frontier labs cut API prices repeatedly through the first half of 2026, with DeepSeek's V4-Pro and Moonshot's Kimi K2.6 both quoted at cache-hit floors below $0.10 per million tokens, a pricing environment that compresses the margin StepFun's own token-priced API business can earn even if usage grows. SV028
CV043 The 2026 Hong Kong IPO market saw cornerstone-investor subscriptions of HKD18.52 billion in January alone, a 13.3-fold increase year over year, with AI and semiconductor listings capturing a large share of that capital, evidence of the same liquidity-driven demand that has inflated Zhipu's and MiniMax's post-listing multiples. SV023
CV044 StepFun's cornerstone and pre-IPO investor base already includes Shanghai state-owned capital, HKIC, and industrial supply-chain investors, a profile that mirrors the state-linked cornerstone lineups 36Kr describes as central to Zhipu's post-listing price support. SV002, SV013
CV045 No comparable Chinese AI-tiger IPO in this dataset (Zhipu, MiniMax) has avoided a post-listing swing of at least 40-70% from its interim high, indicating that even a successful StepFun listing near $10-12 billion is unlikely to deliver price stability in its first two quarters of trading. SV013, SV016
CV046 StepFun's roughly $2.5 billion pre-IPO round and targeted $10-12 billion IPO valuation remain well below Moonshot's $20 billion and DeepSeek's above-$50 billion 2026 marks, positioning StepFun as mid-pack rather than the top of the disclosed Chinese foundation-model valuation tier. SV009, SV019
CV047 The Standard (Hong Kong) reports StepFun was named by MIT Technology Review as one of four Chinese AI startups worth watching beyond DeepSeek, and that founder Jiang Daxin, a former Microsoft vice president, described a team of over 500 staff, nearly 80% technical. SV001
CV048 KrASIA reports StepFun's pre-IPO round drew smartphone-ODM and component investors (Huaqin, Longcheer, OmniVision, ZTE) explicitly because model companies are competing for device-level distribution entry points, tying the company's device-embedded go-to-market to its capital raises rather than to disclosed standalone unit economics. SV002
来源
编号出版方标题引文
SO001 StepFun 阶跃星辰 | StepFun homepage
SO002 StepFun StepFun Platform -- Step 3.7 Flash
SO003 StepFun 阶跃AI (跃问) -- StepFun chat product
SO004 StepFun Step3: Cost-Effective Multimodal Intelligence Step3 is our cutting-edge multimodal reasoning model, built on a Mixture-of-Experts architecture with 321B total parameters and 38B active.
SO005 StepFun stepfun-ai (GitHub organization)
SO006 StepFun GitHub - stepfun-ai/Step3 Step3 is our cutting-edge multimodal reasoning model, built on a Mixture-of-Experts architecture with 321B total parameters and 38B active.
SO007 StepFun stepfun-ai/step3 (Hugging Face model card)
SO008 arXiv (StepFun Team) Step-3 is Large yet Affordable: Model-system Co-design for Cost-effective Decoding This paper introduces Step-3, a 321B-parameter VLM with hardware-aware model-system co-design optimized for minimizing decoding costs.
SO009 Wikipedia StepFun StepFun was founded in April 2023 by former Microsoft employees. Investors include Tencent, Qiming Venture Partners and Shanghai State-owned Capital Investment.
SO010 Wikipedia (Chinese) 阶跃星辰 阶跃星辰于2023年4月成立,总部位于上海市徐汇区,专注于大型语言模型及多模态模型的研发。公司由姜大昕创立并担任CEO。
SO011 Baidu Baike Shanghai Jieyue Xingchen Intelligent Technology Co., Ltd Shanghai Jieyue Xingchen Intelligent Technology Co., Ltd., commonly referred to as Stepfun, was established on April 6, 2023, with a registered capital of 20 million Chinese yuan.
SO012 TechNode Shanghai AI unicorn StepFun raises over $718 million in B+ round StepFun has completed a B+ financing round exceeding RMB 5 billion ($718 million), setting a new record for the largest single financing in China's large-model sector over the past 12 months.
SO013 KrASIA As AI consolidates, what makes StepFun worth a RMB 5 billion raise? Others want to be China's OpenAI or Anthropic. StepFun is closer to being China's xAI plus Tesla.
SO014 KrASIA China's investors double down on AI frontrunners as StepFun raises RMB 5 billion StepFun expects its models to be deployed in more than one million vehicles this year, according to PEdaily.
SO015 KrASIA StepFun nears USD 2.5 billion pre-IPO round as industrial investors join The investment platform, established by the Hong Kong government, has so far backed only one foundation model maker: StepFun.
SO016 Yicai Global Chinese AI Firm Stepfun Raises USD719 Mln for Model Development, AI Agent Rollout Stepfun, a Chinese artificial intelligence startup working closely with car giant Geely Automobile Holdings, has secured about CNY5 billion (USD719 million).
SO017 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.
SO018 Gasgoo Personnel Changes | Yin Qi Appointed Chairman of StepFun Yin Qi serves as chairman of Qianli Technology. He has long focused on the integration of AI and hardware.
SO019 Business Wire Geely Auto Group Teams Up with StepFun for a Joint Showcase at the 2025 World Artificial Intelligence Conference The Galaxy M9 -- highlighted as the world's first vehicle equipped with a Human-like AI agent -- powered by StepFun's end-to-end voice large model.
SO020 C114 (China Communications World) AI Model Developer StepFun Secures 5 Billion Yuan in Series B+ Funding This financing round has set a new record for the single largest funding amount in China's large-scale model sector.
SO021 The Standard (Hong Kong) Stepfun, China's AI Six Tigers, finishes new US$2.5b funding round for HK IPO China's AI Six Tigers include Zhipu, Moonshot AI, MiniMax, Baichuan Intelligence, StepFun, and 01.AI. With Zhipu and MiniMax listed in Hong Kong earlier in January this year.
SO022 FAQ (faq.com.tw) China's StepFun Files for $12 Billion Hong Kong IPO, Betting on AI Embedded in Every Device StepFun's flagship Step-1 large language model currently powers AI features in more than 40 million smartphones.
SO023 Reuters (via U.S. News) Chinese AI Startup StepFun to Unwind Offshore Structure to Pave Way for IPO, Sources Say Chinese AI agent StepFun is unwinding its offshore incorporation structure to pave the way for a planned Hong Kong initial public offering, three sources said, as Beijing tightens scrutiny of a system widely used for overseas fundraising.
SO024 BigGo Finance AI Unicorn StepFun Dismantles Red Chip Structure, Faces Timing Challenge in Hong Kong IPO Sprint StepFun's revenue for 2025 was close to CNY 500 million (approximately $73.3 million), with projected revenue for 2026 reaching approximately CNY 1.2 billion (approximately $175.9 million).
SO025 CnTechPost Chinese AI startup StepFun nears $2.5 billion funding ahead of Hong Kong IPO StepFun has been considering a Hong Kong IPO that could raise about $500 million, with a debut possible as early as this year, Bloomberg reported in February.
SO026 MIT Technology Review Four Chinese AI startups to watch beyond DeepSeek An elite group of companies known as the "Six Tigers" -- Stepfun, Zhipu, Minimax, Moonshot, 01.AI, and Baichuan -- are generally considered to be at the forefront of China's AI sector.
SM001 IMARC Group China Generative AI (AIGC) Market Size, Share, Trends and Forecast, 2026-2034
SM002 CNMO via Tencent News (news.qq.com) IDC发布2026中国手机市场十大洞察 AI手机占比将过半
SM003 Cyberspace Administration of China (CAC) 关于发布生成式人工智能服务已备案信息的公告(2026年3月至4月)
SM004 Xinhua via State Council Information Office China unveils guidelines to regulate, boost innovative development of AI agents
SM005 CNBC U.S. takes step to halt Nvidia AI chip shipments to Chinese firms outside China
SM006 Apidog The 2026 Chinese LLM Price War: Top 5 Frontier API Costs Compared
SM007 IDC China AI: Redefining Global Competition for the Agentic Future (IDC FutureScape 2026 excerpt)
SM008 StepFun (stepfun-ai) stepfun-ai/Step-3.5-Flash model card
SM009 China Biz Insider Chinese AI Models Lead OpenRouter Usage for 2nd Week
SM010 36Kr (EU edition) Global AI Developers' New Favorite: Jieyue Xingchen's Step 3.5 Flash Tops OpenRouter Trend List in Two Days
SM011 CnEVPost Geely unveils industry's first AI agent-powered car cockpit
SM012 China Daily (tech.chinadaily.com.cn) 占比49.5%!火山引擎领跑中国大模型公有云市场
SM013 中商产业研究院 via 中商情报网 (seccw.com) 2026年中国智能座舱市场规模及行业发展前景预测分析(图)
SM014 Sina Finance (finance.sina.com.cn) AI加速渗透下的新一轮变革—IDC 2026年中国PC市场十大洞察
SM015 State Council of the People's Republic of China (gov.cn) 生成式人工智能服务管理暂行办法
SM016 Cyberspace Administration of China (CAC) 关于印发《人工智能生成合成内容标识办法》的通知
SM017 Gartner Gartner Forecasts Worldwide AI Spending to Grow 47% in 2026
SM018 AutoTech News Cognitive Cockpits Take Center Stage as China's EV Makers Become AI Firms
SM019 State Council Information Office (chinavoices) China accelerates AI agent governance amid emerging security risks
SM020 StepFun (platform.stepfun.com) 定价与限速 - StepFun 开放平台文档中心
SM021 Firecat AI 每日资讯 阶跃星辰推出 Step Plan 包月订阅,OpenClaw 场景开发者享半价优惠
SM022 中关村在线 (ai.zol.com.cn) 全国生成式AI服务备案达868款,地方登记应用新增49款
SM023 TrendForce Chinese AI Models Reportedly Hit ~15% Global Share in Nov. 2025, Fueled by DeepSeek Open-Source Push
SM024 Forrester Forrester: Asia Pacific Tech Spending Expected To Grow 9.3% In 2026, But Rising Costs And Regulations Will Impact Real Growth
SM025 Sohu (sohu.com) 1.47亿台AI手机!IDC揭秘2026中国手机市场十大变局
SM026 China Daily Brief From Megvii's Ashes to a Geely-Backed Empire: Yin Qi's Bet on an AI+Car Closed Loop
SM027 FAQ (faq.com.tw) China's StepFun Files for $12 Billion Hong Kong IPO, Betting on AI Embedded in Every Device
SP001 StepFun (Official) Pricing and Rate Limits - StepFun Open Platform Docs (定价与限速) step-3.7-flash: input 1.35元 (cache-miss) / 0.27元 (cache-hit), output 8.1元 per 1M tokens
SP002 Z.ai (Zhipu AI, Official) Z.ai - Advanced AI Chatbot & Agent powered by GLM-5.2
SP003 Zhipu AI (Official) 智谱AI开放平台 (Zhipu AI Open Platform)
SP004 BigGo Finance Zhipu AI's First Annual Report: Revenue Doubles, Losses Widen, Yet Market Awards HK$400 Billion Valuation revenue surged 131.9% year-over-year to CNY 724 million...net loss expanded to CNY 4.72 billion...market capitalization above HK$400 billion
SP005 Moonshot AI (Official) Moonshot AI
SP006 Moonshot AI (Kimi, Official) Kimi AI with K2.6 | Better Coding, Smarter Agents
SP007 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...Moonshot was valued at $4.3 billion at the end of 2025
SP008 Forbes Chinese AI Model Developer Kimi Raising Funds Valuing It At $20 Billion
SP009 MiniMax (Official) MiniMax 自2022年初成立以来,我们以"与所有人共创智能"为使命,致力于推动人工智能科技前沿发展,实现通用人工智能(AGI)
SP010 WinBuzzer MiniMax Raises $619M in Hong Kong IPO as Chinese AI Startups Beat Silicon Valley to Public Markets MiniMax has raised HK$4.8 billion ($619 million) in its Hong Kong initial public offering...values the...company at approximately $6.5 billion
SP011 CNBC MiniMax doubles in Hong Kong debut, marking yet another Chinese AI listing
SP012 Baichuan Intelligence (Official) 百川大模型-百川智能 (Baichuan Large Model - Baichuan Intelligence) 百小医 你的 AI家庭医生 (Baixiaoyi, your AI family doctor)
SP013 TMTPost China's Baichuan Intelligence Launches Medical AI Model That Outperforms OpenAI in Key Benchmark Baichuan-M2 scored 60.1, surpassing OpenAI's gpt-oss120b (57.6)
SP014 Asia Tech Daily Alibaba-Backed Baichuan Secures $691 Million, Valued at $2.7 Billion State-backed investment entities such as the Beijing AI Industry Investment Fund, Shanghai AI Industry Investment Fund, and Shenzhen Capital Group also contributed to the funding.
SP015 DeepSeek (Official) DeepSeek | 深度求索 DeepSeek-V4 预览版本发布,具备世界顶级推理性能,Agent 能力大幅提高
SP016 DeepSeek (Official) Models & Pricing | DeepSeek API Docs 1M output tokens: $0.28 (V4-Flash) / $0.87 (V4-Pro); cache-hit input as low as $0.0028-$0.003625
SP017 TrendForce DeepSeek Completes First Funding Round, Raises Over CNY 50 Billion at Valuation Above CNY 330 Billion the sole exception is the National AI Industry Investment Fund, which invested directly in DeepSeek...retains voting rights
SP018 Alibaba (Qwen, Official) Qwen
SP019 Alibaba Cloud (Official) Alibaba Cloud Model Studio
SP020 Alibaba Cloud (Official) Alibaba Unveils Qwen3.6-Plus to Accelerate Agentic AI Deployment for Enterprises and Alibaba's AI Applications Qwen3.6-Plus will be integrated into Alibaba's ecosystem, including Wukong...and Qwen App, Alibaba's flagship AI application
SP021 Baidu (ERNIE, Official) ERNIE
SP022 Baidu (ERNIE, Official) ERNIE 5.1 Officially Released! Topping Multiple Leaderboards ERNIE 5.1...compresses total parameters to approximately one-third...using only about 6% of the pre-training cost of comparable models...4th place globally and 1st among Chinese models on the Arena Search leaderboard
SP023 ByteDance (Doubao, Official) Doubao (豆包)
SP024 Digital Applied Chinese AI Models Q2 2026: 10-Provider Landscape Report Step 3.5 Flash from StepFun...sits at #3 on the OpenRouter free-model leaderboard at 1.38T weekly tokens
SP025 AICPB China AI Rankings by App MAU — Issue 21 (Apr 2026 Edition) China AI Rankings for App are based on App Monthly Active Users (MAU) in Apr 2026
SP026 BenchLM Best Chinese AI Models (2026) — Ranked by Benchmark Data Qwen3.7 Max leads this ranking with a score of 90, followed by GLM-5.2 (90) and DeepSeek V4 Pro (Max) (87)
SP027 Geely (Official) Geely Unveils Full-Domain AI 2.0 Vision at Auto China 2026 Geely Auto Group presented its brands...alongside its latest group technologies
SP028 Apidog The 2026 Chinese LLM Price War: Top 5 Frontier API Costs Compared Chinese labs cut LLM API prices six times in the first half of 2026...DeepSeek V4-Pro now costs $0.87 per million output tokens
SP029 Forbes China's DeepSeek V4 And Qwen Reshape The Open-Source AI Race
SP030 Cybernews Chinese AI assistant reaches 200 million monthly active users Baidu AI assistant has reached 200 million monthly active users
SI001 StepFun 定价与限速 - StepFun 开放平台文档中心 (Pricing and rate limits) step-3.7-flash: 1M tokens 输入(缓存未命中)1.35元/输入(命中)0.27元/输出8.1元; step-3.5-flash 0.7/0.14/2.1元; step-1o-turbo-vision 2.5/0.5/8元。
SI002 AsiaICT / Investor China (投资界) Is StepFun's $10 Billion Valuation a Cure for AI Monetization or a New Source of Anxiety? According to Caijing magazine, StepFun's 2025 revenue approached RMB 500 million, with 2026 projected revenue around RMB 1.2 billion... revenue structure may rely heavily on a few leading smartphone and automotive manufacturers... revenue-sharing ratios often range from 30%-60%.
SI003 Caijing (财经) via Sina Finance 独家|阶跃星辰计划年内港股上市,2025年收入约5亿元 《财经》获悉,阶跃星辰2025年收入近5亿元,2026年预计收入约12亿元... 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。
SI004 Sina Finance 传阶跃星辰最快今日提交香港IPO申请 今年4月,公司完成股份制改造,注册资本由2394.36万元增至5626.81万元... 2027年年中或面临股份解禁... 智谱2025年营收7.24亿元,MiniMax同期营收约5.43亿元... MiniMax 2025年经调整净亏损约17.5亿元,智谱同期亏损约31.82亿元,毛利率分别为25.4%和41%。
SI005 BigGo Finance Stepfun Nears $2.5 Billion Funding, Supply-Chain Investors Bolster Hong Kong IPO Plans Stepfun is on the verge of closing a nearly $2.5 billion... Huaqin, Longcheer, OmniVision, and ZTE... HKIC has only backed Stepfun among foundation model companies.
SI006 Aiqicha (Baidu) 上海阶跃星辰智能科技股份有限公司 - 工商信息查询 - 爱企查 阶跃星辰于2023年4月成立... 已对外发布Step系列大模型... 在金融、知识服务、网络文学等领域与行业头部公司达成深度合作。
SI007 Tencent News (腾讯新闻) citing Qichacha 阶跃星辰注册资本增至5626万元 企查查APP显示,近日,上海阶跃星辰智能科技股份有限公司发生工商变更,注册资本由2394.36万元增至5626.81万元。
SI008 QCC (企查查) 阶跃星辰 company search result - Hong Kong non-Hong Kong company registration 上海階躍星辰智能科技股份有限公司 仍注册;商业登记号码:80414928;成立日期:2026-05-18;企业类型:注册非香港公司。
SI009 OpenRouter Step 3.5 Flash - API Pricing & Benchmarks | OpenRouter Step 3.5 Flash listed on OpenRouter's third-party developer marketplace alongside other StepFun models.
SI010 PricePerToken StepFun API Pricing (Updated 2026) - All Models & Token Costs 3 Total Models; $0.09 Cheapest Input (per 1M); 256K Max Context Length... Jan 2026 $0.09/$0.30; May 2026 $0.20/$1.15.
SI011 TheInsight.asia / IPO Early Insights StepFun nears $2.5bn fundraising as Hong Kong IPO plans advance Corporate records show the company completed a shareholding restructuring in April... HKIC... has so far invested in only one large-model company: StepFun.
SI012 China Daily Brief China's StepFun Taps Qianli Chief and Raises >RMB5bn to Push Foundation Models into Devices StepFun says the new capital will be channelled into base-model research and building a 'global top-tier base model' while fast-tracking commercial deployments of AI on phones, cars, wearables and embodied robotics.
SI013 CnTechPost Chinese AI startup StepFun reportedly plans Hong Kong IPO, potentially raising up to $500 million StepFun is considering an initial public offering (IPO) in Hong Kong that could raise up to $500 million, Bloomberg reported... recently closed a new funding round exceeding RMB 5 billion ($731 million).
SI014 Tracxn StepFun 2026 Company Profile: Valuation, Funding & Competitors StepFun has raised $718M in funding from Shanghai State-owned Capital Investment... has 172 active competitors, including 43 funded and 5 that have exited.
SI015 Lambda Finance GPU Supply 2026: NVIDIA $57B Quarter, HBM Bottleneck, Hyperscaler $131B Capex NVIDIA captures roughly 44 cents of every hyperscaler capex dollar going into AI infrastructure... NVIDIA DC Annualized Run-Rate $228B.
SI016 Dataquest India (citing TrendForce) CSP capex to soar past $520 billion in 2026, driven by GPU procurement and ASIC development TrendForce predicts that as rack-scale solutions like GB/VR systems grow in 2026, the total capex of the eight CSPs will hit a new peak, surpassing $520 billion with a 24% year-over-year rise.
SI017 Silicon Analysts NVIDIA GPU Prices 2026: B200 at $40K, H100 Dropping to $20K A single H100 GPU costs between $25,000 and $40,000... B200 procurement... 12 to 18-month wait times for volume orders.
SI018 Southern Metropolis Daily via Tencent News 阶跃星辰与"中国英伟达"们成立联盟,今年冲刺10亿元收入 阶跃星辰今年将冲刺10亿元人民币收入目标... 李璟表示:偏定制化交付的项目... 纯卖API的模式...这个模式注定是大厂的生意逻辑。
SI019 C114 Pro Tencent Joins Forces in StepFun's $2.5 Billion Funding Round, Culminating in Strategic Partnership Tencent Cloud inked a strategic cooperation pact with StepFun... developing AI cockpit agents to bridge... music, video, maps, payments, and travel services.
SI020 Gizmochina Nubia Z80 Ultra Becomes First Smartphone With Native OpenClaw AI Integration the Nubia Z80 Ultra will be the world's first smartphone to natively integrate OpenClaw... starts at 4,999 yuan (around $730).
SI021 BigGo Finance Tencent Doubles Down: Follows Stepfun's $2.5 Billion Funding Round, Accelerating AI Push Tencent Cloud and Stepfun signed a strategic cooperation agreement, integrating Stepfun's foundational large models with Tencent's content and application ecosystem.
SI022 Sohu (黑科技研究所) AI芯片"四大天王"聚首 阶跃星辰的大模型已应用于汽车、具身智能、手机等领域,服务超50%的头部国产手机品牌,上半年营收高速增长,冲刺全年10亿元目标。
SI023 GPULoans.com AI GPU Financing in 2026: Funding H100 and B200s Companies seeking to deploy 100 to 10,000 GPUs face upfront costs ranging from $5 million to over $500 million... asset-backed financing... GPUs themselves serve as collateral.
SI024 Baidu Baike 模芯生态创新联盟 (Model-Chip Ecosystem Innovation Alliance) 2026年6月,有行业评论指出,模芯生态创新联盟成立近一年来,其合作仍主要集中于硬件适配与基础调试,在推动规模化部署和吸引广大开发者迁移方面成果有限。
SI025 36Kr Research Institute 36氪研究院:2025年中国大模型行业发展研究报告 2024年中国大模型市场规模已达294.16亿元,预计到2026年将突破700亿元,其中多模态大模型市场规模为156.3亿元。
SI026 BigGo Finance Zhipu AI's First Annual Report: Revenue Doubles, Losses Widen, Yet Market Awards HK$400 Billion Valuation Zhipu generated revenue of approximately CNY 724 million, a year-over-year increase of 131.9%... gross margin declined from 56.3% the previous year to 41.0%... net loss widened to approximately CNY 4.72 billion.
SE001 StepFun (Official Platform) StepFun Open Platform homepage Step 3.7 Flash: 面向生产级 Agent 的高效率 Flash 模型 (production-oriented, high-efficiency Flash model for agents).
SE002 StepFun (Official Platform Docs) 定价与限速 (Pricing and rate limits) — StepFun Open Platform documentation step-3.7-flash: 1.35元 input(cache-miss)/0.27元(cache-hit)/8.1元 output per 1M tokens; step-2x-large 已于2026年06月12日结束限时免费.
SE003 Hugging Face / StepFun stepfun-ai/Step-3.7-Flash model card Step 3.7 Flash is a 198B-parameter sparse MoE vision-language model...activates approximately 11B parameters per token and delivers a throughput of up to 400 tokens per second.
SE004 StepFun (Official Blog) Step 3.7 Flash — A high-efficiency Flash model for Real-World Step 3.7 Flash is available on the StepFun Open Platform...OpenRouter, and NVIDIA NIM...supports flexible deployment across cloud, data center, and local environments.
SE005 GitHub / StepFun stepfun-ai GitHub organization 1.7k followers; Showing 10 of 32 repositories; Step-3.5-Flash 2.1k stars/85 forks; Step-3.7-Flash 257 stars/22 forks.
SE006 GitHub / StepFun stepfun-ai/Step-Audio2 repository Step-Audio 2 is an end-to-end multi-modal large language model designed for industry-strength audio understanding and speech conversation.
SE007 GitHub / StepFun stepfun-ai/Step-Video-T2V repository We release the inference code and model weights of Step-Video-T2V...Step-Video-TI2V, an image-to-video model based on Step-Video-T2V.
SE008 GitHub / StepFun stepfun-ai/Step1X-Edit repository A SOTA open-source image editing model, which aims to provide comparable performance against the closed-source models like GPT-4o and Gemini 2 Flash. 2,232 stars / 101 forks.
SE009 GitHub / StepFun stepfun-ai/gelab-zero (STEP-GUI) repository STEP-GUI: The top GUI agent solution in the galaxy. Developed by the StepFun-GELab team...2,209 stars / 196 forks.
SE010 arXiv / StepFun Step-Audio 2 Technical Report (arXiv:2507.16632) Step-Audio 2...integrates a latent audio encoder and reasoning-centric reinforcement learning (RL)...achieves state-of-the-art performance on various audio understanding and conversational benchmarks compared to other open-source and commercial solutions.
SE011 arXiv / StepFun Step-Video-T2V Technical Report (arXiv:2502.10248) Step-Video-T2V, a state-of-the-art text-to-video pre-trained model with 30B parameters...Video-VAE...16x16 spatial and 8x temporal compression...Video-DPO...to reduce artifacts.
SE012 arXiv / StepFun Step-3 is Large yet Affordable: Model-system Co-design for Cost-effective Decoding (arXiv:2507.19427) Step-3, a 321B-parameter VLM...Multi-Matrix Factorization Attention (MFA)...Attention-FFN Disaggregation (AFD)...decoding throughput of up to 4,039 tokens per second per GPU...higher than DeepSeek-V3's 2,324 in the same setup.
SE013 Digital Applied StepFun Step 3.7 Flash: 196B MoE Agentic Vision Model Most performance figures here are vendor-stated at release; we label them as such...the exact 9.3x multiple is not something we would underwrite without running the comparison on your own tasks.
SE014 Crafiq Step 3.7 Flash by StepFun: LLM Benchmarks, Rankings and Specs Intelligence #108...Coding #89...Context 262K...Parameters 198B (11B active).
SE015 OpenRouter Step 3.7 Flash - API Pricing and Benchmarks | OpenRouter
SE016 Flowtivity Step 3.7 Flash Review: We Tested StepFun's 198B Model on a DGX Spark We deployed the IQ4_XS quantization (105GB)...Our observed generation speed was approximately 27 tokens per second...requires at minimum 128GB of unified memory...Consumer laptops with 16-32GB of RAM cannot run this model locally.
SE017 Entireweb Status Is StepFun Down Right Now? Live Status, Outages and Service Issues StepFun is operating normally right now (1 July 2026)...In the past 24 hours, StepFun has received 0 user reports...Last 30 Days: 26 [problem reports].
SE018 State Council Information Office (chinavoices) China accelerates AI agent governance amid emerging security risks Chinese authorities have issued high-level warnings about vulnerabilities and counterfeit skill packages embedded with Trojan viruses on the OpenClaw agent platform.
SE019 Cyberspace Administration of China (CAC) 关于印发《人工智能生成合成内容标识办法》的通知 (Notice on AI-Generated Synthetic Content Labeling Measures) 国家互联网信息办公室、工业和信息化部、公安部、国家广播电视总局制定了《人工智能生成合成内容标识办法》.
SE020 PyPI step-audio2 package page Step-Audio 2...We release our vLLM backend and corresponding examples.
SE021 ComputerTech Step 3.5 Flash Review 2026: Better Than GPT-5? (We Tested It) The Hacker News crowd is buzzing about it. Developers are running it locally on Mac Studios...trades blows with GPT-5.2 and Claude Opus 4.5.
SE022 Hugging Face / StepFun stepfun-ai/Step-3.5-Flash model card Step 3.5 Flash: 196B MoE, 11B active parameters, 262K context window.
SE023 GitHub / StepFun stepfun-ai/Step3 repository Step3 is our cutting-edge multimodal reasoning model...321B total parameters and 38B active...Multi-Matrix Factorization Attention (MFA) and Attention-FFN Disaggregation (AFD).
SE024 Baidu Baike 模芯生态创新联盟 (Model-Chip Ecosystem Innovation Alliance) 2026年6月,有行业评论指出,模芯生态创新联盟成立近一年来,其合作仍主要集中于硬件适配与基础调试,在推动规模化部署和吸引广大开发者迁移方面成果有限。
SE025 StepFun (Official Consumer App) 阶跃AI (跃问) chat.stepfun.com
SE026 GitHub / StepFun stepfun-ai/Step-Realtime-Console repository 基于阶跃星辰实时语音 API 的前端演示项目...实时语音输入和输出,实现流畅的人机对话体验.
SU001 Google (Play Store) StepFun - Apps on Google Play there is no option for login through Google or TikTok etc, and also there are only few countries that can log in, what if our country isn't in that list, how we can login
SU002 Hugging Face stepfun-ai (StepFun) organization page stepfun-ai/Step-3.7-Flash-NVFP4 ... 309k downloads
SU003 GitHub stepfun-ai GitHub organization stepfun-ai/Step-3.7-Flash’s past year of commit activity ... 257 stars, 2 forks
SU004 BigGo Finance Hong Kong Sovereign Fund's Sole Generative AI Bet: StepFun Nears Record $2.5 Billion Pre-IPO Round The round is led by industry investors including Huaqin Technology, Longcheer Technology, OmniVision Group and ZTE.
SU005 Sohu (热点视听) 上海大模型企业阶跃星辰获50亿融资,推动AI+终端产业加速落地
SU006 GlobeNewswire (via ITNewsOnline) Accelerating the Evolution of Automotive Embodied Intelligence, Geely Auto Group Teams Up with StepFun for a Joint Showcase at WAIC 2025 Geely Auto Group partnered with its strategic tech ecosystem partner, StepFun, to jointly exhibit at the 2025 World Artificial Intelligence Conference (WAIC 2025).
SU007 Nasdaq (press release) ECARX Powers Global Launch of Geely Galaxy M9, Delivering a Next-Generation AI Cockpit Experience Orders for Galaxy M9 surpassed 40,000 within 24 hours after its pre-sales began on August 23, reflecting a strong demand for the flagship model.
SU008 Wilson Sonsini Goodrich & Rosati Wilson Sonsini Advises Huanxi Media on Strategic AI Cooperation with StepFun Huanxi Media Group Limited ... entered into a strategic framework agreement with Shanghai Stepfun Intelligent Technology Co., Ltd. to establish a joint venture aimed at leveraging artificial intelligence and large model technologies.
SU009 Hong Kong Exchanges and Clearing Limited (HKEXnews) Huanxi Media Group Limited — Voluntary Announcement: Framework Agreement for Strategic Cooperation on 11 November 2025, the Company and Shanghai Stepfun Intelligent Technology Co., Ltd. ... entered into a framework agreement for strategic cooperation
SU010 BigGo Nubia Z80 Ultra Claims World's First Native OpenClaw AI, But Security Concerns Linger the company launches a beta testing program for this feature, the tech community is watching to see if the benefits of deeply embedded AI can outweigh the potential risks
SU011 Shanghai Metals Market (SMM) Geely unveils next-gen intelligent cockpit Agent OS At WAIC 2025, Geely Auto, together with Qianli Technology and StepFun, launched the preview version of Agent OS—an AI-Agent-native next-generation intelligent cockpit.
SU012 Sina Finance 千里科技2倍溢价收购关联方资不抵债标的 关联交易额预计超200亿涨幅超80% 千里科技关联交易不断,一方面,公司溢价2倍收购关联方资不抵债标的;另一方面,公司较为依赖关联交易,吉利系多个财年为公司大客户与大供应商
SU013 Sina (手机新浪网) 千里科技招股书解读:净利降54.9% 净利率降295%,转型挑战大 2022 年至 2025 年上半年,前五大客户总收入占比分别为 48.1%、41.5%、41.9%及 42.4%,最大客户收入占比分别为 39.7%、33.6%、30.8%及 33.2%
SU014 The Wire China What is StepFun? Many of StepFun's latest models are multimodal
SU015 AI产品榜 (AICPB) AI产品榜·国内总榜(App) — 第21期(2026年4月版) AI产品榜 按月更新,由全球AI榜单标准制定者 AICPB 发布 ... 根据 App 月活跃用户排名 — 2026年4月
SU016 Sohu (车迷老司机) 阶跃星辰融资超50亿,AI与汽车的未来交汇点揭晓! 已与千里科技联合推出的AgentOS智能座舱,已在吉利银河M9等量产车型中得到应用
SU017 Edgen Tencent Bets $2.5B on StepFun to Build AI Cockpit Agents
SU018 NetEase (via 科创板日报) 阶跃星辰多模态大模型为OPPO新机提供技术支持 OPPO年度旗舰机型 Find X8 Ultra 全网首发的"一键闪记"功能,由阶跃星辰多模态提供技术支持
SU019 10100.com (大数跨境) 阶跃星辰以多模态助力荣耀全新旗舰 荣耀Magic8 系列,获最佳协作奖 阶跃星辰多模态大模型为 YOYO 智能体提供技术支持 ... 在荣耀 AI 分论坛上,阶跃星辰获得了「最佳协作奖」
SU020 Sina Technology 阶跃星辰开源GUI Agent套件,宣布合作超60%国产手机厂商 公司也已经与包括荣耀、OPPO、中兴在内超60%国产手机厂商达成合作 ... 早在今年7月,公司便联合中兴推出了"长辈AI助手"
SU021 AAStocks TENCENT Reportedly Joins StepFun's USD2.5B Financing Tencent Cloud and StepFun inking a strategic partnership. The collaboration will integrate StepFun’s foundational large models with the Tencent ecosystem ... to jointly develop AI cockpit agents.
SU022 Guancha (观察者网) 阶跃星辰发力智能终端Agent,将深化与OPPO、千里科技等合作 阶跃星辰宣布与吉利汽车集团、OPPO、智元机器人等企业深化合作关系 ... 发布「繁星计划」二期,旨在从模型、算力、资本、数据和企业孵化等方面全方位赋能开发者
SU023 Sina Finance (新民晚报) 阶跃星辰获超50亿元人民币B+轮融资,领跑大模型融资榜!印奇正式挂帅 终端Agent AP调用量连续三个季度增长近170% ... 国内60%头部手机品牌已与阶跃达成深度合作 ... 模型装机量超过4200万台,日均服务近2000万人次 ... 吉利银河M9上市3个月销量接近4万辆
SU024 StepFun (Yuewen official site) 跃问官网 (Yuewen product timeline) 跃问宣布与吉利汽车、上海报业集团等行业龙头达成战略合作,将多模态能力应用于智能驾驶、财经资讯等垂直场景
SU025 China Daily Brief From Megvii's Ashes to a Geely-Backed Empire: Yin Qi's Bet on an AI+Car Closed Loop Yin also controls Qianli Technology (formerly Lifan, now in the Geely orbit), a company positioning itself as an auto supplier of smart driving
SU026 Digital Applied OpenRouter June 2026: New Models, Pricing and Rankings OpenRouter listed five major new models — Claude Opus 4.8, Step 3.7 Flash, MiniMax M3, Qwen3.7-Plus, and NVIDIA Nemotron 3 Ultra
SU027 MarketScreener (S&P Capital IQ) Huanxi Media Group Limited and Shanghai Stepfun Intelligent Technology Co., Ltd. Enter into Framework Agreement for Strategic Cooperation Model and Agent Research and Development: Advance exploration and development of core technologies such as large models and Agents for the film and television sector
SR001 National People's Congress of China (npc.gov.cn) 全国人民代表大会常务委员会关于修改《中华人民共和国网络安全法》的决定 国家支持人工智能基础理论研究和算法等关键技术研发...完善人工智能伦理规范,加强风险监测评估和安全监管
SR002 U.S. Bureau of Industry and Security (bis.gov) Commerce Further Restricts China's Artificial Intelligence, Advanced Computing Capabilities Restrict the Chinese Communist Party's (CCP) ability to acquire and develop high-performance and exascale computing capabilities...for military applications
SR003 Shanghai Stock Exchange filing via cninfo.com.cn 重庆千里科技股份有限公司 关于对外投资暨关联交易的进展公告 (公告编号:2025-042) 重庆千里科技股份有限公司...关于对外投资暨关联交易的进展公告
SR004 ARC Group CSRC Requirements for Overseas Listing Companies are required to file with the CSRC within three working days after submitting a listing application
SR005 Law.asia Regulatory scrutiny over red-chip structures raises conundrum regulators had advised several companies planning Hong Kong listings to dismantle their red-chip structures
SR006 Law.asia Key proposed changes to HKEX's listing regime the HKEX published the Phase 1 consultation paper on its Listing Framework Competitiveness Review
SR007 Covington & Burling (InsidePrivacy) China Releases New Labeling Requirements for AI-Generated Content The Labeling Rules impose explicit and implicit labeling obligations on internet information service providers
SR008 Hunton Andrews Kurth CAC Issues Amendment to the Cybersecurity Law of China the latest amendments primarily focus on strengthening and restructuring the penalty regime
SR009 Hogan Lovells (HLC) Key updates on the amended cybersecurity law of China the revised Law establishes an additional tiered penalty regime featuring stricter fines for material cybersecurity violations
SR010 Linklaters (TechInsights) China: dual-track AIGC labelling and latest AI regulatory development Market players in China such as Tencent, Douyin, RedNote, Weibo and DeepSeek have already proactively made public statements
SR011 OPPO OPPO Statement on Compliance in every country and region where we operate, OPPO will comply with local laws and market regulations
SR012 Reuters (via Yahoo Finance) Chinese AI startup StepFun to unwind offshore structure to pave way for IPO, sources say Shanghai-based StepFun...has decided an onshore corporate structure would be more appropriate as it is heavily backed by state capital
SR013 Invezz Chinese AI startup StepFun to drop offshore structure for IPO Companies may face significant legal and financial hurdles in shifting their domicile back to China
SR014 21st Century Business Herald (21jingji.com) AI独角兽阶跃星辰,加速赴港IPO? 千里科技董事长印奇正式登记为董事长,公司创始人兼CEO姜大昕等人担任董事
SR015 NetEase (163.com) 阶跃星辰传拆除离岸注册架构,加速赴港IPO 部分红筹企业股权透明度低、合规风险相对较高,境内外监管机构和主管部门均较为关注此类合规风险
SR016 Tech Times BIS Closed China AI Chip Loophole: Trump Officials Dispute Whether Gap Ever Existed Any such company now faces retroactive exposure stretching back to November 2023
SR017 CNBC U.S. holds off blacklisting China's DeepSeek, more than 100 firms deemed security risks The U.S. has held off adding China's AI startup DeepSeek, memory chipmaker CXMT and more than 100 other companies flagged as national security risks
SR018 CNBC Pentagon expands list of China military-linked firms to include Alibaba, Baidu The Defense Department published an updated '1260H list'...a roster of companies the Pentagon considers affiliated with China's military or defense industrial base
SR019 FilingReader Chongqing Qianli Technology shareholders approve related party transactions for 2025-2026 Shareholders approved the 'Proposal on the Execution of Daily Related Party Transactions in 2025 and the Forecast...in 2026'
SR020 36Kr (EU) Yin Qi Fails to Rescue Qianli Technology Chongqing Qianli Technology officially submitted its prospectus to the Hong Kong Stock Exchange
SR021 KrASIA After tech giants, a new cohort of "AI tigers" finds footing in China Six companies are commonly included in this group: 01.AI, Baichuan AI, MiniMax, Moonshot AI, StepFun, and Zhipu
SR022 China Biz Insider JPMorgan Sees China's AI Race Narrowing as Zhipu AI and MiniMax Emerge From Industry Shakeout the number of viable model developers has shrunk from more than 200 to fewer than ten
SR023 RTHK Lower listing bar for companies with dual-class voting HKEX proposed on Friday that the bar for companies with weighted voting rights seeking to list in the SAR be lowered
SR024 UC Strategies AI Benchmarks Are a Game Now — And the Industry Is Cheating to Win selective model submissions inflated scores by up to 100 points through cherry-picking
SR025 Tech Insider Huawei Ascend 950PR: The 1.56 PFLOP AI Chip vs Nvidia Huawei has officially launched the Ascend 950PR, a 1.56-petaflop AI inference chip that delivers 2.8 times the FP4 performance of Nvidia's H20
SR026 Market Analysis Huawei vs. Nvidia in AI Silicon: Closing the Gap Through Scale, Still Trailing in Efficiency and Ecosystem Huawei's current strategy relies on 'parity by aggregation'...while Nvidia retains leadership in chip-level efficiency, density, and the software ecosystem
SR027 GPULoans AI GPU Financing in 2026: Funding H100 and B200s Companies seeking to deploy 100 to 10,000 GPUs face upfront costs ranging from $5 million to over $500 million
SR028 Lyceum Technology Total Cost of Ownership for a GPU Cluster in 2026 the silicon itself is often just 35 percent of the total infrastructure spend over a five-year lifecycle
SR029 Sohu (via Sina Finance) 千里科技二次递表:连年亏损毛利率仅10% 对吉利"双重依赖"困境难解 千里科技的扣非净利润已连续三年为负...母公司累计未弥补亏损金额高达9.48亿元
SR030 hubpy.io Stepfun (阶跃星辰) Guide 2026: The $718M AI Unicorn With a Trillion Parameters partnered with device makers Honor, Oppo, and ZTE
SR031 CoinUnited.io StepFun's Offshore Unwind Signals China's Next AI IPO Wave on HKEX Target HK IPO Size ~US$500M
SR032 Hong Kong Investment Corporation (HKIC) Hong Kong Investment Corporation Limited — About Us The Hong Kong Investment Corporation Limited...is wholly owned by the Government of the Hong Kong Special Administrative Region
SV001 The Standard (Hong Kong) Stepfun, China's AI Six Tigers, finishes new US$2.5b funding round for HK IPO Stepfun...has reportedly completed a new US$2.5 billion (HK$19.4 billion) funding round and dismantled its red-chip structure, accelerating its push toward a Hong Kong initial public offering.
SV002 KrASIA StepFun nears USD 2.5 billion pre-IPO round as industrial investors join StepFun has also dismantled its red-chip structure...several companies from the smartphone and consumer electronics supply chain, including Huaqin Technology, Longcheer Technology, OmniVision, and ZTE, have joined the round.
SV003 BigGo Finance China's Stepfun Nears $2.5 Billion Funding, Tearing Up VIE Structure as Chinese AI Model Startups Rush for Hong Kong IPOs Chinese AI model startup Stepfun is on the verge of closing a nearly $2.5 billion funding round. The company has also dismantled its variable interest entity (VIE) structure.
SV004 Pandaily StepFun Reportedly Restructures for Potential Hong Kong IPO StepFun Reportedly Restructures for Potential Hong Kong IPO.
SV005 TMTPost StepFun to Complete Pre-IPO Financing by April StepFun is in the Pre-IPO financing round, with a pre-investment valuation of about 4 billion U.S. dollars...conducted in two tranches.
SV006 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.
SV007 The AI Chronicle China vs US: AI Dominance and the Risk of Overvaluation According to analysts cited by the South China Morning Post, the market is on the verge of a 'correction.' Many companies currently valued at billions of dollars may collapse if they fail to prove they can turn their technology into a sustainable long-term business.
SV008 ChinaBiz Insider China AI Compute Crunch: Bubble Risk Grows in 2026 Yann LeCun...told CNBC that an AI bubble burst was not a distant scenario but an imminent one...Every major AI company is currently using investor capital to subsidize user consumption.
SV009 ChinaBiz Insider Kimi AI Raises $2B at $20B Valuation Moonshot AI's Kimi chatbot closes a $2 billion round led by Meituan at a $20B valuation, with ARR doubling to $200M in April on strong coding demand.
SV010 arXiv Boom, Bubble, or Buildout? A Multi-Method Evaluation of Whether Artificial Intelligence Is in an Ongoing Financial Bubble The rapid expansion of artificial intelligence (AI) investment has revived a recurrent question in financial economics: are AI-related assets experiencing a bubble, or is the market capitalizing a durable general-purpose technology.
SV011 PitchBook Zhipu 2026 Company Profile: Stock Performance & Earnings
SV012 China Daily Brief The Anthropic of the East: Deciphering the $60 Billion Valuation of China's Zhipu AI Financial disclosures for 2025 reveal a business defined by explosive top-line growth and equally aggressive capital destruction. While revenue surged by 132% to 724 million RMB, net losses widened to 4.7 billion RMB.
SV013 36Kr (EU) Unveiling the Secret Behind Zhipu AI's Nearly 600 Billion Market Valuation Based on the full-year revenue of 724 million yuan in 2025...Zhipu's corresponding static price-to-sales ratio at a market value of HK$495.8 billion is as high as 576.26 times, and the enterprise value-to-revenue ratio (EV/Revenue) is 581.52 times...The moats are real, but they cannot support a price-to-sales ratio of 576 times.
SV014 AsiaICT Stock Price Rockets by 30%! Zhipu AI Unveils Financial Report: Revenue Jumps 132%, API Usage Surges Threefold, Market Value Tops HK$400 Billion Zhipu AI (02513.HK)...released its first annual report after listing, showcasing doubled revenue, significantly increased gross profit.
SV015 Qiming Venture Partners China's AGI Pioneer and Leader Z.ai Listed on Hong Kong Stock Exchange Qiming Venture Partners' portfolio company Z.ai (formerly known as Zhipu AI)...successfully listed on the Hong Kong Stock Exchange on January 8th, 2026...becoming the world's first listed large language model company.
SV016 StockAnalysis.com MiniMax Group (HKG:0100) Market Cap & Net Worth MiniMax Group has a market cap or net worth of 130.79 billion as of July 1, 2026...Revenue 615.19M...Stock Price 417.00.
SV017 Yahoo Finance MiniMax Group Inc (0100.HK) Stock Price, News, Quote & History HKSE - Delayed Quote HKD 417.000 -12.400 (-2.89%) At close: June 30.
SV018 Hong Kong Exchanges and Clearing Limited (HKEXnews) MiniMax Group Inc. Global Offering Prospectus Global offering; stock code 0100; MiniMax Group Inc., incorporated in the Cayman Islands with a weighted voting rights structure.
SV019 Tech Xplore (AFP) DeepSeek valued at more than $50 bn after funding round: reports Investors have valued Chinese artificial intelligence startup DeepSeek at more than $50 billion in the company's first fundraising round...Anthropic is valued at $965 billion...OpenAI was valued at $852 billion in March.
SV020 Moneycontrol (Reuters) DeepSeek raises $7.4 billion at over $50 billion valuation, founder keeps control: Report Investors are subject to a five-year lock-up and will not have voting rights...China's National Artificial Intelligence Industry Investment Fund is the only exception, having invested directly...and retaining both voting rights and freedom from the lock-up.
SV021 EnkiAI Baichuan AI Infrastructure 2026: A Strategic Deep Dive Baichuan Intelligence secured over $1 billion in strategic funding...a $300 million Series A1 round in October 2023...followed by a massive RMB 5 billion (~$691 million) Series A round in July 2024.
SV022 AI Chat Daily Moonshot AI raises $2B at $20B valuation as Kimi climbs OpenRouter charts Moonshot's annual recurring revenue topped $200 million in April...Moonshot was valued at $4.3 billion at the end of 2025. By early 2026 that figure had more than doubled to $10 billion...Five months later, it has doubled again to $20 billion.
SV023 GMT Eight Hong Kong IPO Cornerstone Investments Surge: HKD 18.52 Billion In First Month, Up More Than 13 Times Year-On-Year In January 2026, all 13 new listings incorporated cornerstone investors, with total cornerstone subscriptions amounting to HKD 18.521 billion, a 13.3-fold increase from the same month a year earlier.
SV024 CompaniesMarketCap Palantir (PLTR) - Market capitalization As of July 2026 Palantir has a market cap of $279.69 Billion USD.
SV025 CompaniesMarketCap Palantir (PLTR) - Revenue Revenue in 2026 (TTM): $5.22 Billion USD.
SV026 CompaniesMarketCap Palantir (PLTR) - P/S ratio P/S ratio as of July 2026 (TTM): 53.5.
SV027 Wikipedia Six AI tigers The group consists of Z.ai (Zhipu AI), Moonshot AI, MiniMax, Baichuan AI, 01.AI, and StepFun...By mid-2024, the combined valuation of the companies was estimated to exceed RMB 100 billion.
SV028 Apidog The 2026 Chinese LLM Price War: Top 5 Frontier API Costs Compared Chinese labs cut LLM API prices six times in the first half of 2026...DeepSeek V4-Pro $0.87, Kimi K2.6 $0.07 cache-hit, GLM-5 $3.20.
SV029 KrASIA After tech giants, a new cohort of "AI tigers" finds footing in China What has changed in 2026 is not just the arrival of more large models or the companies building them, but the emergence of smaller players that public market investors can now price and evaluate.
SV030 Quartz Meet the 'Six Tigers' that dominate China's AI industry
SV031 OfficeChai China's MiniMax Reports $79M Revenue, 25.4% Margins In First Earnings As Public Company MiniMax reported full-year 2025 revenue of $79 million, up 158.9% year over year from $30.5 million in 2024...gross margins improved from 12.2% in 2024 to 25.4% in 2025.