初创公司尽调
尽调报告 AI / foundation models / healthcare AI late-stage private 2026-08-21

Baichuan AI

中国基础模型公司,医疗倾斜明显——产品和客户证据真实,收入质量仍不透明

Baichuan 既有真实医疗差异化、可信融资和可见商业化进展,但当前公开证据仍只支持继续观察,不足以按最后可见私募估值激进承保。

封面要素

最新披露估值 01
2899.11 USD M [CO022, CV001]
累计融资额 02
1038 USD M (tracker-reported) [CO021]
2025 年签约订单目标 03
1000-2000 RMB M [CV004]
2024 年 To B 收入代理指标 04
100 RMB M (reported) [CV005]
有名可查的儿科部署 05
Beijing Children's Hospital flagship reference [CO029, CO030]

公司概况

Baichuan AI 由 Wang Xiaochuan 于 2023 年在北京创立,如今已是中国知名度较高的私营基础模型创业公司之一。 公司最初押注更宽泛的中国 LLM 机会,之后把医疗 AI 打成最鲜明的公开切口,同时保留企业模型平台、开源开发者入口和企业工作流工具。 产品栈包括通用 Baichuan 模型、医疗 M 系列模型、Bai Xiaoyi 面向消费者和健康场景的入口,以及支持私有化部署和智能体式 工作流的企业 API / 平台层。公开护城河证据最强的地方在医疗,尤其是与 Beijing Children's Hospital 相关的 Futang·Baichuan 儿科工作;公司也在营销金融和教育用例。

官网
www.baichuan-ai.com
成立时间
2023-03-24
创始人
Wang Xiaochuan, Ru Liyun
创立地点
Beijing, China
总部
Beijing, China
产品
Baichuan 销售和分发通用基础模型、医疗专用模型、API 访问和企业工作流工具。公开材料显示,公司同时提供自助式开发者访问、企业咨询和医疗相关 部署入口,包括开源模型发布以及适合私有化部署的医疗产品。
客户
医院和医疗机构、企业 AI 团队、开发者 / 自托管用户,以及通过 Bai Xiaoyi 触达的部分消费者健康用户。
商业模式
按用量计费的 API 定价、企业部署和授权、医疗场景解决方案,以及相邻的工作流产品。
阶段
late-stage private
融资情况
据报道,2026 年 4 月 Series B 轮估值约 $2.8B-$2.9B;追踪数据表明累计融资约 $1.0B,战略投资方包括 37 Interactive、Xiaomi 和 Alibaba Cloud。
[CO001, CO006, CO013, CO021, CO022, CO025, CO029, CO032]

执行摘要

主要优势

  • Baichuan 在医疗场景切入真实,并有具名儿科部署作为证据,比许多泛化中国模型实验室更有差异化。
  • 公司产品面覆盖开源模型、API、企业工作流和面向健康的应用,不是只押一个狭窄入口。
  • 融资支持和战略投资人给 Baichuan 带来更强续航,强于资本单薄的前沿模型创业公司。
  • 据报道的订单目标和具名企业或医院引用显示,商业化已不只是零收入实验。

主要风险

  • 公开证据里,收入质量、留存、客户集中度和毛利率仍太薄,难以按最后报告估值形成高确信度。
  • 中国 API 价格压缩叠加医院级定制负担,会同时挤压估值倍数支撑和毛利率空间。
  • 算力政策波动和公司自身合规状态,仍是医疗相邻 AI 业务的重要估值压力。
  • Baichuan 战略可信度不错,但公开规模、分发和财务可见度仍落后于中国 AI 头部赢家。

未决问题

  • 没有经审计的公开财务包按产品线披露确认收入、毛利率或项目型与软件型收入结构。
  • 头部客户集中度、合同期限、续约和队列扩张指标仍未公开。
  • 最新股权结构表、清算优先权和任何投资人特定保护条款均未公开披露。
  • 面向公众上线服务的分产品 CAC 备案或安全评估状态,公开渠道看不到。

目录

Chapter 01

01公司概况

1.1 身份、产品范围和当下可见的运营足迹

Baichuan AI 现在的公开身份比「中国 LLM 创业公司」这个泛标签更具体。官网称,公司由前 Sogou CEO Wang Xiaochuan 于 2023-03-24 创立,使命是帮助公众获取世界知识和专业服务,核心团队来自 Sogou、Baidu、Huawei、Microsoft、ByteDance 和 Tencent。官网还把运营地点锚定在北京海淀区,列出百川智能科技有限公司这个法律品牌以及互联网服务、医疗信息备案号。这个细节重要,因为 Baichuan 不再只把自己呈现为模型实验室。首页主推 Baixiaoyi AI 家庭医生、Haina Baichuan 医疗 API 计划,以及围绕更安全、更低幻觉医疗使用的应用层承诺。同时,平台站点和 API 文档显示,Baichuan 仍在运营更宽的企业栈:通用 Baichuan4 模型、工具调用 API、企业智能体,以及金融和教育垂直定位。因此,一句话概括:Baichuan 是一家总部位于北京的私营模型公司,可见商业化层已经转向医疗,但没有放弃更大的平台野心。[CO001, CO002, CO003, CO004, CO005, CO006]

Baichuan AI 快照 KPI 表
指标数值 / 状态日期置信度缺口
公司 / 品牌Baichuan AI / 百川智能,Baixiaoyi 是当前官网首页露出的消费者健康入口2026-08-21
成立官网首页显示 2023-03-242023-03-24
创始人Wang Xiaochuan;公开记录也把 Ru Liyun 标识为资深联合创始人 / 总裁2024-12-24官方页面未完整列出确切法律层面的联合创始人名单
总部北京海淀区2026-08-21已审阅页面使用多个街道级地址,但城市一致
当前产品重心医疗 AI 加企业 / API 平台2026-08-21
最新报道估值CB Insights 显示 2026 年 4 月投后估值约 $2.9B2026-04-01公司未发布经审阅、含确切条款的融资新闻稿
累计融资CB Insights 显示累计约 $1.038B2026-04-01不同追踪器 / 来源对轮次金额和投资人名单略有差异
收入 / ARR / 员工数 / 客户数null2026-08-21已审阅公开来源未披露经审计收入、ARR、员工数或客户数

空值代表公开指标无支撑,而不是数值为零。

[CO001, CO006, CO021, CO022, CO035, CO036]
产品与渠道入口表
入口主要用户有证据支撑的能力当前角色注意点
Baixiaoyi消费者 / 家庭AI 家庭医生工作流,用于症状准备、结果解读和家庭健康管理首页应用品牌医疗建议免责声明仍限制其直接替代临床服务
Haina Baichuan 项目医疗服务机构面向有证据锚定医疗场景的免费 M3 Plus API生态播种和医疗渠道扩张限于获批真实服务场景和品牌规则
开放平台 / API开发者和企业授权模型 API,支持工具调用、JSON 模式和限速用量直接变现入口公开文档未披露实际价格或客户量
通用 Baichuan4 系列企业买方更高速、面向企业优化的通用模型让公司在医疗之外仍保持相关性官网营销说法来自公司自身,并非第三方审计
开放权重社区发布研究人员、开发者、自托管用户可通过 GitHub 和 Hugging Face 获取 Baichuan2、M1、M2、M3 及旧模型分发、信任和采用漏斗开放权重关注度能转化多少商业收入未披露

本表只映射已审阅官方和代码仓库材料中可见的用户入口;不推断未披露内部产品。

[CO005, CO008, CO009, CO025, CO026, CO031]
FO001: 公司里程碑时间线

Baichuan 在三年内从成立走到开放模型发布、战略融资,并形成医疗优先的商业化姿态。

公司未在所审来源发布带日期新闻稿时,里程碑日期采用最早审阅到的报道或追踪器轮次日期。

[CO001, CO004, CO017, CO019, CO021, CO022]

1.2 创始人与市场契合、领导层集中度和治理可见度

Baichuan 领导力叙事最强的一环,是 Wang Xiaochuan 的创始人与市场契合。TechCrunch 把公司直接连到 Wang 在 Sogou 长期积累的搜索和语言技术经验,也连到他公开呼吁中国需要自己的 OpenAI;官网和 Caixin 又把他与前 Sogou 运营负责人 Ru Liyun,以及来自中国和全球头部科技公司的团队放在一起。这个组合让 Baichuan 更像严肃的中国基础模型竞争者,而不是营销空壳。薄弱处在治理可见度。公开材料集中在 Wang,次之是 Ru;已审阅的官方页面没有公布董事会名单、委员会结构、财务负责人,也没有展开当前高管梯队。用户协议和隐私政策确实显示,公司在正式的中国法律框架内运营,包含海淀法院管辖条款、医疗使用免责声明和隐私响应承诺,但这些是合规界面,不是治理透明度。合适的阶段标签是商业化中的后期私营 AI 公司:它拿过多轮大额融资,有开放平台、清晰的行业包装和真实部署主张,但仍缺少上市软件公司投资者会期待的公开控制和披露深度。[CO007, CO010, CO011, CO012, CO013, CO014]

领导层与创始人表
人物 / 职能角色证据创始人-市场匹配或覆盖关键人依赖 / 缺口
Wang Xiaochuan创始人兼公开代表官方网站、TechCrunch、TMTPostSogou 创始人、前 CEO,具备搜索和语言 AI 背景公开叙事高度集中在一位创始人身上
Ru Liyun联合创始人 / 总裁Caixin 访谈与 TMTPost 提及Sogou 前 COO,也是可见的商业化操盘手官方入口对外可见度低于 Wang
核心技术团队招募自 Sogou、Baidu、Huawei、Microsoft、ByteDance、Tencent官网首页显示其能触达中国 AI 人才池并完成可信招聘未见经审阅的公开组织架构图或高管名单
董事会 / CFO / 委员会已审阅官方材料未公开呈现官方入口未找到董事会名单或财务负责人页面这是治理披露缺口,不是不存在的证明构成治理与控制权审查的重要尽调障碍

表中同时列出具名负责人和明确的治理覆盖缺口,因为公开披露集中在创始人身上。

[CO001, CO003, CO010, CO011, CO012, CO013]
治理、合规与未解决数据缺口表
主题已观察到的公开证据重要性状态具体尽调路径
董事会与委员会未见经审阅的公开董事会名单或委员会披露影响控制权、监督和投资人权利分析缺口索取股权结构表、董事会名单、委员会章程和观察员权利
财务负责人 / 经审计财务未见经审阅的 CFO 披露、经审计收入或利润率说明阻碍收入质量和资本效率分析缺口索取最新经审计财务和财务组织概览
Series B 募资额和条款CB Insights 给出轮次日期和投资人,但未给金额或优先权入场价格和稀释取决于确切条款部分获取轮次文件,或由管理层确认规模与优先股堆叠
客户数和员工数未找到经审阅的公开客户总数或员工总数限制对生产率和集中度的基准比较缺口索取分细分市场客户数及员工总数
地址规范化官网首页和政策文件显示不同海淀地址轻微法律 / 运营映射问题,但影响尽调记录部分核验注册地址、主要办公地址及搬迁历史
医疗合规入口官网首页、条款和隐私政策显示备案与免责声明信任姿态是医疗部署的核心部分索取底层备案证书、范围和续期状态

本表记录公开尽调尚无法核验的内容,并不证明公司缺少底层控制或指标。

[CO007, CO014, CO015, CO016, CO035, CO038]

1.3 融资路径、投资者基础,以及封面指标能支撑什么、不能支撑什么

Baichuan 的融资记录显示,这家仍未上市的中国模型公司得到了异常深的战略支持,但不同来源对具体轮次时间线并不完全一致。TechCrunch 称,公司在 2023 年成立后很快融资 $50 million。TMTPost 随后记录了 2023 年 Alibaba 和 Tencent 的战略支持,以及 2024 年 7 月由 Alibaba、Xiaomi 和 Tencent 领投、CICC 和国资 AI 基金参投的一轮更大融资。Tracxn 的时间线方向一致:2023 年 10 月 $300 million 轮、2024 年 4 月未披露金额轮,以及 2024 年 7 月按 $2.7 billion 投后估值完成的 $691 million 轮。CB Insights 把记录延伸到 2026 年 4 月 1 日,称 Baichuan 最新投后估值约 $2.9 billion,并明确列出 37 Interactive Entertainment 是 Series B 投资方。最稳妥的结论不是一份干净、由公司认证的股权结构表,而是一个有支撑的区间:Baichuan 累计融资约 $1.0 billion,估值从 2024 年中约 $2.7 billion 走到 2026 年 Series B 时约 $2.8-2.9 billion。同样重要的是缺口:已审阅的公开来源没有披露经审计收入、ARR、客户数、员工数、烧钱速度或 Series B 的确切募资额。[CO017, CO018, CO019, CO020, CO021, CO022]

利益相关方或投资人图谱
利益相关方角色控制权或经济重要性公开信号尽调问题
Wang Xiaochuan创始人 / 战略负责人对战略、招聘和外部叙事有关键人影响官方与媒体来源反复以其为中心所有权、投票控制权和继任深度
Alibaba Cloud / Alibaba Group战略投资人2023-2024 年多轮持续支持者,也可能是企业生态合作伙伴TMTPost、Tracxn、CB Insights 具名商业依赖或云渠道销售关联
Tencent战略投资人2023-2024 年反复出现的投资人,具备分发和生态价值TMTPost、Tracxn、CB Insights 具名是否存在排他性或渠道依赖
Xiaomi战略投资人参与早期和后续轮次,并在平台合作伙伴页面可见TMTPost、Tracxn 和平台网站具名设备或渠道合作深度
37 Interactive EntertainmentSeries B 投资人显示云巨头之外的后期战略资本进入CB Insights 在 2026 年 4 月 Series B 中明确具名确切出资额及董事会 / 投票条款
国资与财务投资人资本与政策支持层CICC 和多个地方 AI 基金拓宽资金基础TMTPost 和 Tracxn 具名优先股堆叠、跟投意愿和治理权利

投资人和利益相关方行把具名战略支持者与创始人放在一起,因为经济控制和商业杠杆都是公司画像的核心。

[CO017, CO018, CO019, CO020, CO021, CO022]
里程碑表
日期事件类型金额 / 估值 / 状态参与方含义
2023-03-24Baichuan 成立创立官方成立日期Wang Xiaochuan 和创始团队确定公司时间起点和北京身份
2023-04-10天使轮融资融资据报道 $50M天使投资人显示公司启动后投资人立刻有兴趣
2023-06-15发布 Baichuan-7B产品开放权重模型Baichuan 工程团队建立早期开放模型可信度
2023-07-11发布 Baichuan-13B,并获 TechCrunch 报道产品13B 双语模型Baichuan 与开发者社区提升其作为中国头部 LLM 创业公司的可见度
2023-10-17战略融资轮融资据报道 $300MAlibaba Cloud、Tencent、Xiaomi将中国主要科技支持者带入股权结构
2024-07-25大额 Series A 完成融资RMB5B / 约 $690M,估值约 $2.7BAlibaba、Xiaomi、Tencent、CICC、国资 AI 基金使 Baichuan 成为中国融资最充足的私营 AI 实验室之一
2025-03-20发布 Futang·Baichuan 儿科模型合作部署里程碑Baichuan、Beijing Children's Hospital、合作伙伴让医疗部署成为真实运营切口
2025-09-07推出 Baichuan-M2产品医疗推理模型Baichuan 工程团队加深医疗专业化和私有部署叙事
2026-04-01CB Insights 记录 Series B 轮融资投后估值约 $2.9B37 Interactive 与其他投资人显示相较 2024 年公开估值仅温和抬升
2026-08-21官网首页仍突出 Baixiaoyi 和 Haina Baichuan规模当前产品定位Baichuan确认运行日期时公开叙事仍以医疗为先

若未审阅公司新闻稿,日期采用首个已审阅发布日期或追踪器轮次日期,而不是经公证的交割日期。

[CO001, CO004, CO017, CO019, CO021, CO022]

1.4 医疗转向、里程碑节奏,以及后续章节需要继承的审慎点

后续章节需要承接的主要战略变化,是 Baichuan 从宽泛的前沿模型叙事转向以医疗为中心的应用栈。首页现在强调 Baixiaoyi 和 Haina Baichuan 计划,而不是泛化的消费者聊天机器人。TMTPost 称,Wang 明确把医疗基础模型称为 AI 研究的「皇冠明珠」;GitHub、Hugging Face 和官网则展示出模型弧线:从 Baichuan-7B、Baichuan-13B 到 Baichuan2、M1、M2、M3 和 Baichuan4。最具体的医疗里程碑,是 2025 年 3 月与 Beijing Children's Hospital 发布 Futang·Baichuan 儿科成果;ScienceNet 和北京软件行业协会把它描述为面向基层儿科照护的双医生部署路径。TMTPost 还补充,Baichuan-M2 可在相对轻量的硬件上私有化部署,HealthBench 得分 60.1,并面向真实临床环境构建。这些里程碑有分量,但没有消除核心尽调疑点。公司仍很少披露治理深度、商业质量或财务韧性;KrASIA 的更宽「AI 老虎」框架也暗示,Baichuan 仍暴露在商业模式压力下,压力来自更大的中国巨头和更显眼的前沿模型同行。[CO004, CO025, CO026, CO027, CO028, CO029]

Chapter 02

02市场分析

2.1 市场边界和规模测算:Baichuan 到底在卖什么

不能按整个中国 AI 或半导体经济给 Baichuan 做投资测算。公开页面显示,它的变现层窄得多,但仍然可观:国内模型 API、企业 / 智能体工作流、医疗部署,以及面向金融和教育的选择性垂直包装。官网和开放平台通过 Baixiaoyi、医疗 API 计划、通用企业模型,以及工具调用、结构化输出等 API 功能,把这个边界讲得很清楚。这把 Baichuan 放在 Forrester 所说更大基础设施热潮的下游,也放在医院和受监管企业的终端工作流预算上游。因此,公开市场口径应彼此拆开,不能硬塞进一个假的单一 TAM。IMARC 估计,中国生成式 AI 应用市场 2025 年约 $5.16 billion,到 2034 年增至 $19.56 billion。Gartner 和 Forrester 描述的是全球 AI 支出和中国 AI 基础设施中更大的相邻资金池;China Daily 和 IDC 则显示 token 消耗扩张速度远快于任何已经稳定的收入基础。对 Baichuan 来说,正确读法是分层市场:应用和模型用量真实存在,但最宽的支出池只有一部分归它,因为云基础设施、芯片和通用企业软件都在它的直接变现轨道之外。[CM001, CM002, CM003, CM005, CM006, CM007]

市场定义表
细分市场 / 类别纳入支出排除支出买方 / 付款方与 Baichuan 的相关性
国内模型 API 和智能体使用按 token 计费的模型调用、企业智能体工作流、托管推理原始 GPU 采购、通用云 IaaS、无关软件工具开发者、企业、服务机构直接对应 Baichuan 开放平台和 API 文档
医疗 AI 部署临床助手软件、私有部署集成、医院工作流工具药物研发、与 LLM 工作流软件无关的医疗器械医院、健康服务机构、数字健康合作伙伴当前首页战略和儿科部署证据都落在这里
受监管企业 AI金融、教育、客服和合规导向的 LLM 应用不含模型 / 应用层的广泛企业 IT 支出企业 IT、合规和业务部门预算Baichuan 平台上可见的次级垂直领域
消费者健康与家庭 AI症状准备、医疗结果解释、家庭健康管理通用社交 / 消息使用终端用户和家庭Baixiaoyi 是当前面向消费者的应用入口
排除的基础设施层None芯片、数据中心资本开支、存储、网络和通用云硬件基础设施买方作为成本输入影响 Baichuan,不是直接收入池

各行把 Baichuan 可能变现的应用、部署和工作流预算,与它只会消耗的更大基础设施和芯片层拆开。

[CM001, CM002, CM003, CM013, CM019]
TAM / SAM / SOM 或规模测算视角表
发布方年份地区数值CAGR / 采用率方法置信度局限
IMARC Group2025中国USD 5.16B 生成式 AI 应用市场 -> 2034 年 USD 19.56B15.96% CAGR订阅式市场报告应用口径排除了大部分基础设施,也可能低估垂直部署经济性
Frost & Sullivan 经 TMTPost 引述2023-2033中国2023 年 USD 1.2B AI 医疗市场 -> 2033 年 USD 42.5B43% CAGR媒体引用的第三方市场预测二手引用,不是经审阅的一手报告
Gartner2026全球USD 2.59T AI 支出47% YoY分析师支出预测全球,且远宽于 Baichuan 所在品类
Forrester2026中国> USD 70B AI 基础设施支出中国科技支出增长 7% 的一部分分析师支出预测仅基础设施层;不是 Baichuan 直接收入
IDC,经 China Daily2025-2026中国2025 年企业 / 公有云 MaaS token 1,944T -> 2026 年预测 40,000T2025 年同比约 16x,2026 年预测约 20x用量跟踪与预测Token 数是用量,不是收入
IDC FutureScape2027中国80% C1000 企业优先考虑 AI 主权采用目标分析师预测采用意向不等于马上形成软件支出

每个口径切的是市场的不同截面,单位也互不兼容,因此这些行应作为边界视角读取, 而不是相加得到市场规模。

[CM005, CM006, CM007, CM008, CM010, CM014]
现状替代方案与相邻品类表
替代方案买方为什么使用为什么与 Baichuan 竞争Baichuan 仍可能赢在哪里对 Baichuan 的限制
纯人工临床工作流受信任,且已嵌入医院避开模型风险和新增采购临床人员时间紧,AI 可在就诊前后帮忙临床买方可能更偏好人力加既有软件
传统医院或企业软件已签约并集成不用生成式 AI 也能满足基础工作流Baichuan 可加入推理、解释和智能体行为切换需要证明 ROI,也要投入集成
国产超大规模云厂商模型(Qwen、ERNIE、Doubao)分发广,云与合规栈完整API 形态相近,渠道更强Baichuan 可用医疗深度和私有化部署拉开差异既有巨头可在定价和 GTM 上砸更多钱
多供应商开放权重栈开发者可自托管并切换供应商降低锁定,利于控成本Baichuan 的医疗调优可能带来更好的医疗场景开箱效果开放权重也降低从 Baichuan 迁出的切换成本
通用消费者聊天机器人用户已经熟悉可承接轻量信息任务Baixiaoyi 的医疗健康指向更明确通用消费者工具仍可能抢走注意力和低风险使用

替代方案从无 AI 工作流延伸到更大的国产模型供应商和开放权重多供应商模式。

[CM019, CM020, CM026, CM028]

2.2 买方分层,以及 Baichuan 似乎想占住的医疗子市场

对 Baichuan 最重要的市场区分,不是抽象地区分消费者和企业,而是区分受监管的医疗需求和其他一切。Baichuan 官网、Haina Baichuan 医疗 API 计划和 Futang·Baichuan 儿科部署,都指向一个买方、用户和付款方被拆分在医院、医生、服务机构、患者以及部分公共卫生系统之间的市场。这种复杂性会拖慢商业化速度,但也创造了通用聊天应用难以自动占有的切口。TMTPost 引用 Frost & Sullivan 数据称,中国 AI 医疗市场将从 2023 年约 $1.2 billion 增至 2033 年 $42.5 billion;ScienceNet 和 BSIA 则显示,Baichuan 已把自己嵌入儿科工作流,而不是等待未来试点。医疗之外,Baichuan 仍服务金融和教育领域的企业团队,也服务通过 API 或开放权重版本接入的开发者和医疗科技 合作伙伴。这些相邻市场是真实的,但公开证据显示它们更像支撑性需求入口,不是首页身份。投资研判时,更窄但更能防守的框架是:Baichuan 瞄准的是中国医疗和受监管企业的 AI 预算,在这些场景里,私有化部署、低幻觉率和国内数据处理比原始消费者规模更重要。[CM003, CM012, CM014, CM015, CM016, CM017]

细分市场 / 买方图谱
细分市场买方用户付款方工作流预算负责人采用触发点
医院和诊所医院采购 / 科室负责人医生、护士、管理人员医院或公共卫生预算临床分诊、文档记录、患者教育、决策支持临床 IT / 科室预算需要私有化部署、缓解人手短缺,并拿到合规就绪工具
医疗健康服务机构数字健康运营方和服务平台医务人员和终端患者平台或服务运营方预算面向消费者的家庭健康或医生服务工作流产品 / 运营预算需要可规模化的中文医疗推理,并配套治理控制
金融 / 教育企业团队业务部门或 IT 负责人员工和终端客户业务部门预算知识工作、内容生成、服务自动化IT / 合规 / 业务线预算有压力部署经过行业调优的国产 AI
开发者和医疗科技伙伴开发者或合作伙伴产品团队前述用户加下游应用用户工程或创始人预算API 使用、原型验证、智能体搭建、私有化部署实验工程 / AI 产品预算低摩擦 API 或开放权重实验
消费者 / 家庭个人用户同上自付症状准备、报告解读、家庭健康管理自付需要易获取的中文 AI 健康指导

Baichuan 的市场是多边的:医疗买方、开发者、企业和消费者各有不同的采用触发点 和预算负责人。

[CM016, CM017, CM018, CM025, CM029]
医疗 AI 子市场表
子细分市场当前工作流问题AI 为什么有吸引力Baichuan 契合点未解预算 / 采用问题
就诊前症状准备患者就诊时信息不完整或噪声大LLM 可梳理症状并生成分诊问题Baixiaoyi 和儿科产品已在这里定位生产环境里谁付费:医院、保险方还是消费者?
就诊后结果解读患者难以理解报告和医嘱中文解释可减少困惑和后续错误Baixiaoyi 和医疗基础模型面向解释工作流打造服务方能承受多少临床责任?
临床决策支持数据和隐私约束下,医生需要更快检索和推理私有化部署和有证据锚定的输出有价值Baichuan 主打低幻觉医疗 API 和 M2 部署效率实际采购周期和测得 ROI 仍未披露
儿科和基层医疗顶级医院之外,专科医生短缺更尖锐双医生和儿科模型工作流可外延专家能力Futang·Baichuan 是 Baichuan 聚焦该方向的直接证据北京试点离开旗舰合作伙伴后,可复制性多强?
医疗健康服务平台需要在敏感数据不外流前提下获得可规模化 AI 辅助API 加私有化部署可加快产品上线Haina Baichuan 计划直接面向服务医务人员的机构收入分成和客户集中度仍未知

这些行聚焦 Baichuan 实际营销或已通过公开合作证据展示的医疗工作流, 而不是覆盖所有可能的医疗 AI 用例。

[CM003, CM014, CM015, CM016, CM025, CM029]
FM001: 买方 / 细分市场地图

Baichuan 几个主要市场切面的预算归属、监管摩擦和切换动态差异很大。

序数标签概括了各细分市场基于证据的匹配度与摩擦差异,并不声称衡量市场份额。

[CM016, CM017, CM018, CM019, CM025, CM026]

2.3 中国本土模型市场的增长驱动、约束和采购摩擦

有三股力量把需求拉向 Baichuan。第一,国家政策明确支持:2026 年 5 月 AI 智能体实施指引把 AI 智能体与更广的「AI+」行动直接绑定,并列出科研、产业、消费、民生和治理等 19 个目标应用场景。第二,IDC 中国 AI 展望称,到 2027 年,中国 1,000 强企业中 80% 将优先考虑 AI 主权,这让本土模型供应商相对受阻或政治敏感的美国产品更占优。第三,Baichuan 自己在私有化部署和医疗对齐上的技术立场,贴合一个重视本地数据控制的市场。但同一批驱动也带来摩擦。2023 年生成式 AI 办法意味着,备案、安全和内容治理不是一次性勾选,而是持续成本。医疗买方还会追加隐私和可靠性审查。美国对先进 AI 芯片收紧出口管制,增加供给风险,也抬高了模型效率或国产芯片兼容性的价值。2026 年中国 LLM 价格战还显示,即便需求在增长,愿意压低标价的竞争对手仍会挤压变现。这就是 Baichuan 的市场理论上诱人、实践中难做的原因:公司面对的是增长中的国内需求,但必须在全球政策塑形最重、利润率争夺最激烈的 AI 市场之一中执行。[CM004, CM010, CM011, CM020, CM021, CM022]

增长驱动因素与约束表
驱动因素 / 约束方向时点影响尽调问题
AI+ 与 AI 智能体政策支持正向当前 / 中期政府明确希望在民生和产业场景部署更多智能体Baichuan 哪些用例契合官方医疗和公共服务优先方向?
AI 主权需求正向中期企业偏好本地托管或政治风险更低的模型时,国产厂商受益Baichuan 管线中,有多少依赖主权敏感买方?
医疗工作流痛点与临床人员短缺正向当前 / 中期临床摘要、分诊和患者教育带来真实工作流需求哪些问题真正能报销或有预算支撑?
2023 年生成式 AI 合规负担负向当前备案、标识和内容治理工作抬高上线成本,也拖慢迭代Baichuan 每次部署的内部合规成本是多少?
高端芯片出口收紧负向当前 / 中期算力稀缺抬高模型服务风险,也利好高效模型或适配国产芯片的模型Baichuan 对受限 Nvidia 供应的依赖度,相比国产替代方案有多高?
中国 API 价格战负向当前需求增长仍可能转化为更低的实际单位经济性Baichuan 净合同价格相对标价是多少?

扩大市场的政策和需求力量,也会带来采购、合规和利润率摩擦。

[CM010, CM011, CM020, CM021, CM022, CM024]
监管与采购关口表
关口控制方要求为什么对 Baichuan 重要缺失证据
生成式 AI 服务规则CAC 及相关监管机构安全、内容治理和服务备案义务每个面向公众的模型或应用都必须通过这些义务Baichuan 各产品的备案状态公开披露不完整
AI 智能体落地指导CAC/NDRC/MIIT安全、标准化和应用对齐支持扩张,但也设定可控性预期产品路线图有多少直接绑定这 19 个场景?
医疗隐私与数据本地化医院、监管机构、隐私负责人敏感数据控制和安全部署架构强化私有化部署和本地部署选项需求详细架构和审计准备状态未公开
临床信任与责任医院和医疗专业人员低幻觉、证据锚定、监督、免责声明Baichuan 面向医疗而不只是通用聊天销售,因此至关重要独立错误率证据仍薄弱
出口管制与算力获取美国商务部规则加硬件供应链高端 GPU 获取可能受限可能利好高效模型或兼容国产芯片的模型,但会约束规模Baichuan 实际芯片组合和储备容量未披露

政策支持不会消除摩擦;它会把需求导入具体的备案、安全和基础设施关口, 进而影响成交速度和利润率。

[CM004, CM011, CM021, CM022, CM030, CM034]

2.4 现状替代品、相互矛盾的估算,以及 Baichuan 现实能拿下什么

只有忽略替代和口径冲突,Baichuan 的可服务市场才显得巨大。医院仍可以依赖医生、传统软件,或走一条没有生成式 AI 层的慢速数字化路径。企业可以选择 Qwen、ERNIE、Doubao 这样的超大规模云厂商栈,也可以叠加多个低价本土 API,切换摩擦有限。开发者可以在开放权重社区和兼容 chat-completions 的 API 之间多平台使用。与此同时,公开市场估算使用的单位完全不同:美元、基础设施支出、token 消耗、监管数量和采用率预测。它们不能相加,否则会重复计算。对 Baichuan 尤其如此,因为公司并不需要整个中国 AI 市场都可投资;它需要在医疗和受监管企业这些切片中拿到可信份额,而它的私有化部署和医疗导向主张正是在这些切片里有意义。到目前为止,证据支持市场机会真实存在,但没有证明 Baichuan 能把多少支出转化为持久的经常性收入。在采购深度、报销路径和客户集中度更清晰之前,正确立场是把中国 AI 和医疗顺风当作机会条件,而不是 Baichuan 最终拥有市场的证明。[CM019, CM023, CM026, CM028, CM031, CM036]

Chapter 03

03竞争格局

3.1 格局:Baichuan 位于中国模型领导者第一梯队之后

Baichuan 的竞争集合比「AI 老虎」叙事更宽,但比整个中国 AI 经济更窄。Digital Applied 的 2026 年 Q2 格局报告称,10 家供应商基本覆盖了所有有意义的中国 AI 输出,Alibaba、Z.ai、DeepSeek、Moonshot、MiniMax、ByteDance、Baidu、Tencent、Xiaomi 和 StepFun 吃掉大部分份额和能见度,Baichuan、Yi 等则落在第二梯队的利基带。这个框架有用,因为它把 Baichuan 与三类对手拆开。第一类是 Z.ai、Moonshot、MiniMax、DeepSeek 这样的纯创业公司同行,它们当前的规模信号都强于 Baichuan。第二类是 Qwen、ERNIE、Doubao 这样的内嵌大厂模型家族,背后有云、搜索或消费者生态,而 Baichuan 没有。第三类是开放权重和 API 生态,许多供应商提供相似技术接口,买方可以多平台使用。Baichuan 自己的公开材料让公司看起来比这些通才更窄:官网突出医疗和家庭健康工作流,而不是宽泛的消费者助手或云平台。因此,竞争分析从一个简单观察开始:Baichuan 不是在争做万能的中国模型领导者;它要证明的是,一个医疗优先的专科玩家能在更大的横向对手面前活下来。[CP001, CP002, CP003, CP004, CP005, CP006]

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相对 Baichuan 的限制或反之
Baichuan AI纯创业公司,垂直专家据报道估值约 $2.8-2.9B;私有公司医疗 AI、企业 API、开发者医疗模型、儿科证据、私有化部署角度资本、应用规模和渠道触达远窄于第一梯队
Z.ai / GLM 系列纯创业公司,企业优先国产企业市场形象强,资本市场可见度高企业部署和国资关联买方国产硬件和企业采购定位公开的医疗垂直切入点弱于 Baichuan
Moonshot / Kimi纯创业公司,通用前沿实验室$20B 估值,OpenRouter 用量信号强消费者和知识工作助手加 API基准测试和融资势头垂直专精弱于 Baichuan
MiniMax纯创业公司,已上市香港 IPO 融资约 $619M消费者应用和订阅,加智能体工作流公开市场融资渠道和已变现应用医疗健康叙事弱于 Baichuan
DeepSeek前沿实验室,开放权重经济性>CNY50B 融资和 >CNY330B 估值开发者、企业、自托管者开放权重可信度和更深资本没有明显医疗专精
Alibaba Qwen嵌入大科技公司的模型家族背靠 Alibaba 资产负债表和云栈云、企业、消费者、编程最广产品线和分发医疗健康品牌聚焦较弱
Baidu ERNIE嵌入大科技公司的模型家族背靠 Baidu 和主要国内分发搜索、企业、智能体应用庞大装机基础和低成本训练主张私有化医疗切入点不够明确
ByteDance Doubao嵌入大科技公司的消费者优先模型家族依托 ByteDance 入口拥有海量消费者分发消费者 AI 应用和企业 API最强注意力和分发能力医疗健康专精最弱

规模信号混合了私有估值、公开上市和母公司资产负债表支撑, 因此各行只能方向性比较,不能放在同一统一口径下精确对照。

[CP001, CP002, CP003, CP011, CP012, CP013]
融资 / 规模对比表
服务商公开规模信号重要性竞争判断局限
Baichuan私营初创公司,据报道估值约 $2.8-2.9B资本足够保持相关性,但不足以定义品类中型专业玩家,不具备顶级资产负债表实力公开财务披露有限
Moonshot2026 年估值 $20B、融资 $2B可大手笔投入模型、分发和人才资本差距明显仍是私营公司,透明度不充分
MiniMax香港 IPO 融资 $619M具备公开市场融资工具,商业化也更可见融资灵活性强于 Baichuan受公开市场波动影响
DeepSeek以 >CNY330B 估值融资 >CNY50B算力和商业化跑道很长规模比 Baichuan 大一个数量级治理结构非传统
Qwen / ERNIE / Doubao母公司资产负债表支持可以补贴价格和分发相比纯 AI 初创有结构性优势难拆分模型家族经济性

资本对比混合了私募轮估值、IPO 募资和母公司隐性支持,因此应理解为战略能力,而不是精确的同口径估值计算。

[CP011, CP012, CP013, CP014, CP015, CP016]

3.2 竞争者画像:融资、产品范围和战略方向

最强的纯创业公司同行,在公开资本和能见度上都领先 Baichuan。TechCrunch 称,Moonshot 在 2026 年 5 月按 $20 billion 估值融资 $2 billion;CNBC 称,MiniMax 在香港 IPO 中募资约 $619 million,大部分收入来自订阅和应用内购买。TrendForce 称,DeepSeek 以超过 CNY 330 billion 的估值融资超过 CNY 50 billion,瞬间拉开了相对 Baichuan 低得多的公开估值区间的资产负债表差距。Z.ai/GLM 围绕企业部署和国产硬件韧性定位;Alibaba 的 Qwen 则置于一个足够宽的云和应用生态中,可覆盖代码、多模态和企业工作流。Baidu 和 ByteDance 在分发上更难攻击,因为 ERNIE 可以借 Baidu 的巨大装机基础,Doubao 可以借消费者入口。Baichuan 明显更窄。它最广为宣传的资产是 M2/M3 等医疗模型、儿科部署证据,以及更便宜的医疗导向部署主张,而不是大众应用领导地位或全国基准统治力。这个策略是连贯的,但也意味着 Baichuan 选择的战场比领导者更小。[CP011, CP012, CP013, CP014, CP015, CP016]

功能 / 能力矩阵
能力BaichuanZ.aiKimiMiniMaxDeepSeekQwenERNIEDoubao
医疗健康特定公开定位有限有限有限有限有限有限有限
通用消费者助手规模有限中等偏强中等中等最强
企业云 / 采购栈中等中等中等中等最强
开放权重 / 自托管信号医疗 / 开放模型产品线强中等有限至中等有限有限
基准可见度小众 / 垂直最强最强中等
消费端分发护城河有限中等中等中等最强

单元格是基于已审阅公开来源给出的有证据支撑的顺序判断;它们概括的是定位,而不是给出数值评分。

[CP003, CP017, CP022, CP023, CP029, CP031]
开放权重与部署姿态表
服务商开放权重 / 自托管姿态部署信号战略含义相对 Baichuan 的判断
Baichuan开放模型线仍活跃,医疗模型尤其如此私有部署和医疗专属部署主张支撑受监管买方和开发者核心切口
Z.ai开放权重 / 企业姿态强国产硬件和采购适配吸引重视主权 / 自主可控的买方自托管企业交易中非常强的对手
DeepSeek开放权重经济性强API + 自托管 + 大资本基础对开发者和成本敏感型企业非常有吸引力最难打的开放权重对手
Qwen选择性开源支持叠加云包装Model Studio 和广泛生态触达平衡开放 / 社区与托管云覆盖最广的替代方案
Kimi性能强,但包装更偏自研闭源应用 + API + 融资光环主要靠性能和品牌取胜,而不是自托管控制威胁形态不同
ERNIE / Doubao更封闭,且由生态牵引分发优先于自托管灵活性在分发最关键的场景仍能主导自托管竞争不如 Qwen 或 DeepSeek 直接

在当下中国市场,开放权重和自托管姿态影响买方锁定与采购路径,其重要性不亚于模型质量。

[CP017, CP023, CP025, CP027, CP029, CP033]

3.3 能力、定价、分发,以及锁定效应为何难以维持

按当前基准测试观感,公开赢家在别处。BenchLM 的 2026 年 8 月切片由 Kimi K3 领先,并把 Qwen3.8 Max 列为最佳开放权重选项;Digital Applied 也显示 Qwen、MiniMax、Z.ai、DeepSeek 等领导者比 Baichuan 有更清晰的周 token 份额信号。定价也不是 Baichuan 的干净救命稻草。官网称 Baichuan4-Turbo 定价低于 GPT-4o,Baichuan4-Air 大幅降低推理成本,但 Apidog 的价格战对比显示,更广的中国市场已经在多个竞争对手之间压缩了标价。分发差异比定价还大。Qwen 搭载 Alibaba Cloud 和 Alibaba 应用,ERNIE 搭载 Baidu,Doubao 搭载 ByteDance,Kimi 有更好的基准和融资动能,MiniMax 和 Z.ai 已有公开市场货币,DeepSeek 则把开放权重可信度和更深资本结合起来。这些供应商大多提供 OpenAI 式 API 轨道或可下载模型,技术切换摩擦低。这对 Baichuan 有战略意义:公司必须靠医疗专属工作流契合、信任和部署经济性赢下买方,而不是靠泛化平台锁定效应。[CP023, CP024, CP025, CP026, CP027, CP028]

定价 / 套餐对比
服务商定价或套餐信号包含能力折扣 / 未知项含义
Baichuan官网称 Baichuan4-Turbo 约为 GPT-4o 定价的 80%,Baichuan4-Air 成本大幅更低企业优化、API 接入、医疗模型已审阅资料未清晰呈现完整公开 token 价目表靠效率竞争,但尚看不出是市场定价者
DeepSeek直接 API 文档和市场评论显示,它的 API 套餐低成本、可投入生产API 接入、开放权重背书、自托管吸引力企业实际折扣未知为市场压出低价参照
Qwen通过 Model Studio 提供广泛模型家族,部分层级可预览或免费使用代码、多模态、企业和智能体部署实际合同价格未披露不只靠价格,也靠产品宽度和云包装
Kimi价格竞争力叠加基准领先带动采用智能体编程和知识工作企业定价和渠道条款未完全公开Baichuan 不能只靠“最便宜”或“基准最强”赢
中国市场整体Apidog 显示 2026 年中国前沿模型 API 普遍降价聊天补全通道大体相似标价不等于净价整个品类定价权弱

各行只比较公开标价或套餐信号;不反映私下企业折扣或最低承诺量。

[CP023, CP024, CP025, CP026, CP028, CP029]
分发与采购对比
服务商主要分发入口买方获取路径竞争意义相对 Baichuan 的判断
Baichuan官网、医疗部署、API 文档、开放模型仓库垂直专家销售叠加开发者采用比起大众关注,更需要工作流验证窄,但若医疗转化可验证,可能守得住
QwenAlibaba Cloud + Alibaba 应用云牵引企业销售,叠加消费端曝光庞大装机基础和交叉销售能力最难打的企业渠道对手
ERNIEBaidu 生态搜索与平台牵引的企业 / 消费触达国内分发极强Baichuan 追不上默认入口
DoubaoByteDance 消费端入口消费端获客带来企业外溢注意力护城河最强Baichuan 不应正面打这场仗
Kimi消费应用 + 开发者热度 + 融资光环品牌牵引用量和 API 扩张势能会放大开发者采用Baichuan 在广泛心智上落后
Z.ai企业采购和国产硬件定位采购优先非常契合重视主权 / 自主可控的买方最直接的企业型纯 AI 公司对手

采购和分发能力的差异大于基准排名差异,因此横向领导者即使垂直适配并不明显更好,也能跑赢较窄的专业玩家。

[CP014, CP015, CP016, CP018, CP028, CP029]
FP001: 功能广度 / 能力图

Baichuan 在医疗专业化上最强;相较领先者,广泛分发、基准可见度和资本厚度较弱。

单元格是基于已审阅公开来源综合出的证据支持序数判断;并非正式基准分数。

[CP003, CP017, CP018, CP019, CP020, CP021]

3.4 护城河耐久性:医疗专精真实存在,但仍有条件

Baichuan 剩下的切口是具体的,而不是宽泛的。公司有医疗优先的公开品牌、可见的医疗模型家族、儿科部署证据,以及围绕低幻觉和私有化部署的明确主张。这比许多同行的泛泛「企业 AI」叙事更具运营细节。即便 Qwen、Kimi、DeepSeek 和 ERNIE 继续赢得通用叙事,它也给了 Baichuan 一个合理存在理由。但切口仍然有条件。Digital Applied 仍把 Baichuan 归入第二梯队利基供应商。KrASIA 认为,中国「AI 老虎」这一批公司仍需证明自己能在巨头支持的对手面前跑出持久经济性。开放权重和 API 兼容性帮助 Baichuan 触达开发者,也降低了切到资本更充足替代品的成本。医疗本身比消费者聊天更可能形成强护城河,但也是一个更慢、更受监管、更依赖信任的市场。因此,合理的竞争结论应保持均衡:Baichuan 不必处处击败每个中国模型实验室,但必须在横向对手认为同一垂直领域值得重兵投入之前,把医疗叙事转化为可重复的客户胜利。[CP003, CP034, CP035, CP036, CP037, CP038]

护城河耐久度 / 竞争风险登记表
护城河主张威胁严重性缓释手段 / Baichuan 必须证明什么尽调问题
医疗专业化Qwen、DeepSeek、ERNIE 或 Z.ai 也能切入医疗把垂直叙事转成可重复客户胜利和临床工作流证据收入或使用量中,真正医疗专属的占比是多少?
私有部署经济性对手也能降低推理成本或支持自托管中高证明部署更快、幻觉率更低,并适配国产芯片部署成本和价值实现周期有哪些客户证据?
开放模型生态可见度开放权重也会降低转向竞争对手的切换成本把开放模型关注度绑定到自有或高粘性工作流采用开放模型用户转为付费或企业部署的频率有多高?
垂直小众聚焦更小战场也可能意味着更低天花板深吃一个受监管用例,而不是追逐所有品类按细分市场看,现实可服务市场有多大?
目前第二梯队地位资本和应用规模领先者可以用钱和分发压过 Baichuan保持差异化,不与通用规模赛跑什么能阻止横向领导者复制这个切口?

风险登记表把 Baichuan 的专业化视为真实但有条件的优势,因为这个市场里的每项优势都容易被复制、被价格压力削弱,或被分发不对称冲掉。

[CP034, CP035, CP036, CP037, CP038, CP040]
Chapter 04

04财务情况

4.1 收入模式和价格界面:Baichuan 通过 API 用量、工具和企业项目变现,但实际收益率仍不透明

Baichuan 的公开材料让变现入口比损益表更容易看见。官方价格卡显示,通用模型、医疗模型、搜索插件、向量嵌入、文件存储和暂时免费的 Assistants API 都采用按用量付费。文档还显示,自助式开发者路径仍需要实名认证、充值和创建 API key;企业文档则把更大买方导向商务咨询流程。这意味着收入设计是混合型,而非纯 SaaS:边缘是自助 token 收入,中间是销售驱动的部署和集成,战略核心则通过 Hai Na Baichuan 计划培育医疗生态。难点在实际定价。公开标价无法说明大客户折扣、捆绑支持、收入确认,或 Baichuan 的医疗足迹有多少是在补贴换分发,而不是直接变现。财务上,价格界面是真实的;留存下来的收益率仍未公开。[CI001, CI002, CI003, CI004, CI005, CI006]

收入流表
收入流机制计费单位当前数值 / 状态质量尽调问题
通用 API 使用量Baichuan4 / 3 / 2 通用模型按 token 计费1,000 tokens官方价目表公开,且多个通用 SKU 仍在售标价可见度高,实际收入可见度低提供计费 tokens、企业折扣表以及按模型家族拆分的收入。
医疗 API 使用量Baichuan M 系列按 token 计费,通常叠加医疗搜索调用1,000 tokens + 搜索调用官方价目表公开;Hai Na 项目可让符合条件的机构实际价格降至零需求面可见度中等,净变现可见度低提供付费与免费医疗流量、免费项目向付费合同的转化,以及医疗模型毛利率。
搜索与检索附加项在模型使用量之上叠加网络搜索或医疗搜索调用每次调用官方价目表列出每次搜索调用 RMB0.03费率可见度高,附加率可见度低提供每个客户工作流的平均搜索调用量,以及每次会话的有效混合收入。
向量嵌入与知识库存储为检索工作流提供向量嵌入调用和托管文件存储1,000 tokens 和 GB/day官方价目表列出向量嵌入和存储收费标价可见度高,采用结构可见度低提供存储增长、平均知识库规模以及企业部署附加率。
企业部署与集成私有部署、工作流搭建以及医疗 / 企业集成,通过商务咨询销售定制合同官方文档显示有销售辅助入口和行业工作流,但没有合同价目表存在性可见度中等,经济性可见度低提供 MSA 样本、部署费用、支持义务,以及服务与软件收入占比。
医疗生态培育面向符合条件的医疗服务机构提供战略性免费访问,隶属 Hai Na Baichuan项目 / 机构符合条件用户可永久免费使用 M3-Plus,这一点已公开短期收入低,战略意图强提供免费生态用户队列转为经常性付费合同或下游私有部署的转化情况。

本表把可见收费面与实际经济性拆开。免费或补贴的医疗访问即使不立即最大化收入,也可能具备战略价值。

[CI001, CI002, CI003, CI004, CI005, CI006]
定价 / 变现表
SKU 或合同价格 / 单位 / 合同标价与实际价格折扣 / 未知项来源
Baichuan4 旗舰版RMB0.1 / 1K tokens公开标价企业折扣和承诺用量价格未公开官方价目表
Baichuan4-TurboRMB0.015 / 1K tokens公开标价官方页面称约为 GPT-4o 定价的 80%,但实际合同条款未知官网 + 官方价目表
Baichuan4-AirRMB0.00098 / 1K tokens公开标价接近地板价的定价可能反映竞争压力,而非持久利润率官网 + 官方价目表
医疗 M 系列M3-Plus 输入 RMB0.005 / 输出 RMB0.009;M3 输入 RMB0.01 / 输出 RMB0.03;M2 输入 RMB0.002 / 输出 RMB0.02,均按 1K tokens公开标价付费使用可能被部分机构的免费项目访问抵消官方价目表
向量嵌入和存储向量嵌入 RMB0.0005 / 1K tokens;文件存储 RMB1.5 / GB/day公开标价实际支出取决于知识库采用和留存官方价目表
Assistants / Hai Na BaichuanAssistants API 暂时免费;符合条件的医疗服务机构可永久免费使用 Hai Na M3-Plus已公布的促销 / 项目定价免费使用能否转为付费企业服务或私有部署仍未知官方价目表 + 官网

作为一家中国私营模型公司,Baichuan 对标价的透明度异常高;但公开费率可见,不等于实际变现可见。

[CI002, CI003, CI004, CI005, CI006, CI009]
FI001: 收入模型桥

Baichuan 把模型使用量、搜索调用、向量嵌入和企业项目转成总收入,但实际毛利取决于折扣、免费计划使用量,以及未公开的重算力和服务层。

这是一条定性桥,依据官方价目表、文档和企业端页面搭建。Baichuan 未公布分部收入瀑布图或毛利率桥。

[CI001, CI002, CI003, CI004, CI005, CI006]

4.2 商业化代理指标:公开订单和客户信号存在,但仍不足以形成清晰的经常性收入视图

Baichuan 已不再只是没有收入证据的研究实验室。Caixin 报道称,联合创始人 Ru Liyun 预计 2025 年签约订单额为 RMB1-2 billion;36Kr 的反向报道则称,公司 To B 业务 2024 年收入接近 RMB100 million。同一篇 36Kr 文章列出 Bank of China、NE Digital、China Merchants Bank、Xinyada 和 Tiankai Group 等客户和合作伙伴;ScienceNet 和 BSIA 则记录了儿科和医院部署。这些是有意义的需求信号,因为它们显示 Baichuan 已从模型发布走进真实企业和医疗工作流。但证据仍有大洞。签约订单不等于确认收入,一次性定制项目也不等于持久的经常性软件利润率。公开记录同样没有拆分当前需求中有多少来自付费推理、付费部署,或免费向医疗机构开放 M3-Plus 这类受补贴战略项目。[CI009, CI011, CI012, CI014, CI015, CI021]

单位经济模型表
指标数值 / 空缺置信度重要性尽调问题
2024 年 To B 收入2024 年接近 RMB100M(第三方报道)说明 Baichuan 已产生可观 B2B 收入,但无法判断是否经常性、是否偏服务提供经审计的 2024 年按产品线拆分收入及毛利贡献。
2025 年签约订单目标联合创始人 Ru Liyun 提到 RMB1-2B 签约订单目标(第三方报道)表明商业野心和管线规模,但签约订单不等于确认收入提供积压订单、转收入比例、取消率和回款时间。
2025 年前商业化进展Caixin 标题称半年签约订单已达数亿元 RMB支持收入生成是真实的,而非纯愿景提供签约订单到确认收入和现金回款的桥接。
免费项目医疗流量未公开拆分免费 M3-Plus 使用可能带来线索;若转化弱,也可能稀释变现提供符合条件的机构数量、活跃使用量和一段时间内的付费转化。
毛利率未公开披露没有毛利率,就无法判断低价到底是战略上聪明,还是财务上自毁按模型 / 产品线提供算力、支持和服务成本桥接。
CAC / 回本周期 / NRR未公开披露客户效率指标是区分软件型收入与定制项目收入的必要证据按垂直行业提供销售周期、获客成本、扩张率和留存队列。
账面现金估计36Kr 引述的一位行业人士估计,Baichuan 手上仍有 >RMB3B有方向性参考价值,但不是经审计的流动性证据提供不受限现金、短期投资,以及按法人实体列示的现金分布。

该表有意把已观察指标和明确空值混在一起。公开证据在商业化信号上最强,在可重复的单位经济上最弱。

[CI011, CI012, CI014, CI015, CI016, CI039]
FI002: 单位经济性桥

公开证据在定价和订单信号层最强,在收入确认层较弱,在毛利率、CAC 和经常性效率指标上最弱。

这条桥刻意停在公开证据的边界。由于 Baichuan 未披露管理层级的留存或利润率数据,下游单位经济性输出仍未解决。

[CI011, CI012, CI014, CI015, CI024, CI025]

4.3 成本结构和资本强度:医疗专精可能改善产品契合,但不会让模型业务变便宜

Baichuan 的公开成本图景碎片化,但方向清楚。公司仍在运营基础模型栈,依赖昂贵算力、频繁模型更新和专门医疗适配。美国限制 Nvidia H20 出货,提高了中国实验室面对更高有效基础设施成本和更复杂供应路径的概率。产品层面,Baichuan 自己的代码库称 Baichuan-M2 可在单张 RTX 4090 上运行,这有助于部分买方的私有化部署经济性,但它只降低边缘推理摩擦;它没有披露 Baichuan 自身的训练和平台账单。2026 年中国价格战是另一项主要成本侧问题。如果 Baichuan 必须匹配一个反复降价并已固化为常态的市场,除非医疗专精真的提高支付意愿,否则利润率扩张会更难。医疗可以形成比通用聊天更好的护城河,但 36Kr 的报道认为,同一垂直领域也可能高度定制、扩张缓慢,而且以当前 To B 形态看并不显然盈利。[CI015, CI024, CI025, CI027, CI028, CI029]

FI004: 资本强度 / 现金流图

Baichuan 有强外部资本支持,商业化活动也可见,但医疗 To B 和免费计划变现的成本可见度与投资测算置信度仍最弱。

该矩阵把本章证据转成投资测算可见度类别,而不是硬评分的财务输出。Baichuan 未发布足够报表来支撑定量现金流模型。

[CI015, CI016, CI018, CI020, CI023, CI024]

4.4 资本充足性和融资依赖:Baichuan 背后资金强,但公开流动性可见度仍弱

公开记录中支撑最强的财务优势是融资能力,不是现金生成。Tracxn 称 Baichuan 已完成 4 轮融资、累计约 $1.04 billion;CB Insights 显示累计融资 $1.038 billion,2026 年 4 月估值接近 $2.9 billion,且 Series B 包括 37 Interactive。TMTPost、Yahoo/SCMP 和 36Kr 也印证了 2024 年 7 月由中国头部战略方和国资相关基金领投的大额融资。这让 Baichuan 在中国当前 AI 融资体系内看起来可融资。但资本充足性不同于资本可得性。公开材料仍没有经审计现金余额、月度烧钱披露、债务时间表或现金跑道模型。36Kr 的反向文章引用一位未具名行业人士估计称,Baichuan 手头仍有超过 RMB3 billion,但这不是管理层级别证据。实践结论是,Baichuan 的短期偿付能力因为股东基础够深而看起来合理,但投资判断仍依赖私下流动性数据,而不是公开资产负债表证明。[CI016, CI017, CI018, CI019, CI020, CI021]

资本充足性表
指标公开数值 / 状态置信度重要性尽调要求
手头现金没有官方公开现金披露;一位第三方来源估计剩余 >RMB3B缺资产负债表,流动性就无法承销提供最新现金余额、受限现金和短期投资。
月度烧钱速度未公开披露烧钱速度决定 Baichuan 多快必须证明商业化,或再次融资提供月度运营现金消耗和算力支出趋势。
现金跑道月数未公开披露现金跑道是连接当前现金和下一轮融资压力的关键桥梁提供基准、下行和增长情景的现金跑道模型。
计划资金用途融资报道指向继续投入模型研发、应用和垂直部署说明资金仍在用于扩张,而不是资本回流或稳态优化提供董事会批准的资金用途计划和资本配置优先级。
下一轮融资触发条件未正式披露;可能取决于商业化证明和持续模型投入需求投资人需要知道哪些条件会迫使公司再融资提供融资协议约束阈值、最低现金目标和融资决策规则。
债务 / 项目融资义务在来源集中未识别出公开债务时间表或项目融资敞口隐性义务可能实质改变真实现金跑道和风险提供债务额度、云资源预留、担保和表外承诺。
外部资本渠道公开来源一致显示,公司已融资 >$1.0B,并获得战略和国资关联投资人支持即便收入指标仍弱,强投资人支持也能提高存活率提供当前股权结构表、优先股堆叠,以及绑定后续融资的投资人权利。

该表聚焦前瞻资本充足性,而不是重复报告其他部分已经建立的完整融资时间线。

[CI016, CI017, CI018, CI019, CI020, CI021]
FI003: 财务估计区间

公开报道覆盖了有意义的收入代理指标、超过 $1 billion 的累计融资,以及二十亿美元高段估值,但许多点位仍来自媒体报道或市场数据估计,而非经审计披露。

USD 和 RMB 数值沿用来源使用的单位。图表混合了公司 / 市场数据直接点位和第三方报道,因为 Baichuan 不发布经审计财务报表。

[CI012, CI014, CI016, CI017, CI018, CI019]

4.5 财务阻碍和结论:商业化可期,收入质量未解

Baichuan 的正向逻辑很直接:公开价格卡、可见客户和医院部署、可信投资者支持,以及独立报道显示商业化不再为零。负向逻辑同样重要。公司没有公布判断收入究竟像软件、偏服务还是靠战略补贴所需的合同条款、留存曲线、成本栈、折扣规则或利润率桥。若干最佳收入数据点也来自媒体报道,而非审计。与此同时,医疗优先转向和国内 API 价格战形成张力:专精可能改善产品契合,但也可能把公司困在长销售周期、高定制化和预算有限的买方中。截至报告日期,Baichuan 看起来是一家严肃的商业化阶段私营 AI 公司,有足够资本支持继续建设;但仅凭公开证据,还不能干净地研判收入质量、利润率路径或现金跑道。[CI012, CI014, CI015, CI016, CI020, CI024]

公开财务缺口表
缺失的私有指标影响具体尽调路径
经审计财务报表没有经审计报表,订单和收入轶事就无法核对到确认收入或毛利润索取最新经审计年度和中期报表,以及按产品线拆分的收入确认政策。
现金、烧钱速度和现金跑道资本充足性只能推断,不能衡量索取月度现金桥、董事会现金跑道材料,以及云 / 资本开支承诺。
实际成交价格和折扣政策如果大买家拿到深度折扣,标价会明显高估收入质量审阅企业合同样本、折扣审批和已计费使用量导出。
客户集中度和续约少数具名客户胜利不能证明耐久的经常性收入索取头部客户集中度、续约曲线、NRR 和积压订单账龄。
服务与软件收入结构高定制化会压低毛利,也让收入更难复现分别拆出部署、集成、支持、推理和许可证收入。
临床责任和变现约束医疗免责声明、医院预算和隐私限制会拖慢收入放大并推高成本索取医院采购模型、赔偿立场、隐私架构和临床风险治理备忘录。

截至 2026-08-21,这些是阻碍仅凭公开信息对 Baichuan AI 作出清晰投资判断的最高影响拦路虎。

[CI015, CI016, CI024, CI025, CI031, CI039]
Chapter 05

05产品与技术

5.1 产品界面和模块地图:Baichuan 交付的产品家族比医疗优先品牌最初看起来更宽

Baichuan 当前的公开品牌主打医疗,但底层产品地图明显更宽。官网、价格页和平台首页合在一起,至少展示了四个可见产品层:Baichuan4-Turbo、Baichuan4-Air 等通用基础模型;面向医疗、金融、教育和角色 / NPC 用途的领域增强线;带知识库和工具调用功能的企业智能体平台;以及通过 GitHub、Hugging Face 和研究论文分发的开放权重模型。Bai Xiaoyi 把医疗栈变成面向用户的家庭医生工作流;NPC 领域线则显示,Baichuan 仍支持角色型智能体,而不只做医院或临床软件。实践中,这意味着 Baichuan 应被理解为一个平台:它可以以 API、私有化部署包、垂直模型家族,或类似消费者产品的医疗互动层出现。产品界面因此是连贯的,但并不简单:同一套栈通过多种交付模式暴露给不同客户和不同信任门槛。[CE001, CE002, CE003, CE004, CE005, CE006]

产品模块 / 资产矩阵
模块 / 资产主要用户状态 / 成熟度差异化尽调缺口
Baichuan4 / 4-Turbo / 4-Air企业开发者和 API 买家当前 / 积极推广面向企业优化、成本定位激进的通用中文模型独立基准和可靠性证据仍有限
医疗 M 系列(M1、M2、M3、M4)医疗机构、健康服务提供方、临床 AI 开发者当前 / 旗舰垂直线医疗问询、证据检索、低幻觉定位和领域对齐许多性能声明来自公司自身,生产治理仍不透明
Bai Xiaoyi患者、家庭、医疗健康服务运营方当前 / 可见产品形态把医疗技术栈落到诊前、诊后和家庭健康工作流缺少活跃用户、留存和临床监督模式的公开指标
角色 / NPC 模型线游戏、娱乐和角色智能体开发者当前 / 范围更窄但已上线角色知识库、记忆和可定制角色设置采用情况和安全控制的公开证据薄弱
智能体平台 + 知识库 + 工具调用企业客户和集成商当前 / 核心平台层把检索、工具使用、规划和 API 集成打包进一个平台支持义务和生产 SLA 缺少充分公开文档
开放权重仓库和模型卡开发者、自托管用户、研究社区当前 / 成熟分发路径GitHub、Hugging Face、arXiv、vLLM 和 SGLang 兼容性降低采用摩擦开放分发也降低切换成本,并暴露对外部框架的设计依赖

各行汇总官方页面、文档、仓库和论文中可见的主要客户侧产品形态。矩阵聚焦每项资产在运营上实际做什么,而不是重复营销口号。

[CE001, CE002, CE003, CE004, CE005, CE006]
工作流 / 使用场景表
用户任务当前工作流Baichuan 方案可衡量收益限制
通用企业助手开发者或企业团队想要低成本中文 LLM 接入通过 API 或平台使用 Baichuan4-Turbo / 4-Air官方平台强调速度、更低部署成本和工具集成正常运行时间和企业规模可靠性的独立证据稀少
医疗咨询和随访用户询问症状、就医和诊后解读Bai Xiaoyi + M 系列医疗模型公开材料展示诊前分诊、诊后解释和家庭健康支持输出明确定位为辅助,不能替代专业诊断
临床决策支持 / 医院部署机构需要把有证据支撑的医疗 AI 嵌入工作流带检索、记忆和多模态工具的 M3/M4 技术栈医疗专用推理和医院部署证明,使它区别于通用聊天机器人部署经济性、合规开销和错误处理尚未完全公开
私有或边缘部署买家想自托管,或在受限硬件上部署开放权重 M1/M2/M3 资产,加 vLLM/SGLang 或本地量化配置单张 4090 跑 M2 的路径和 M3 量化选项,降低部分工作负载的进入门槛真实吞吐、安全性和集成工作量取决于客户环境
角色或领域智能体开发者想要基于角色或领域条件约束的智能体行为NPC 模型线 + 知识库 + 工具调用公开材料展示角色记忆、自定义设置,以及基于背景知识的事实性控制公开记录没有清楚区分成熟产品和实验性功能集

这些行把公开产品栈翻译成具体用户任务。表格有意以工作流为中心,因为 Baichuan 当前产品形态横跨 API、部署和终端用户交互模式。

[CE001, CE005, CE008, CE016, CE019, CE021]
FE001: 产品架构图

Baichuan 通过通用模型、医疗模型、角色模型、企业平台和开放权重分发,把同一套底层模型与智能体栈对外露出。

这套栈由公开页面、模型卡和论文拼出,而非来自内部系统图。图中强调客户可见界面如何映射回共享平台能力。

[CE001, CE004, CE005, CE008, CE010, CE021]

5.2 模型和智能体架构:医疗线越来越像受约束的智能体系统,而不只是聊天模型

公开记录中最深的技术证据,位于医疗模型谱系内部。Baichuan-M1 描述了一个从零训练的医疗模型,训练数据混合医疗和通用语料,并针对长上下文和临床推理任务做了明确架构调整。Baichuan-M2 随后把这套栈推向真实世界推理:Large Verifier System、患者模拟器、中期训练医疗适配、多阶段强化学习,同时仍宣传某些配置可用单张 4090 部署。Baichuan-M3 通过主动临床问询、事实感知强化学习、分段工作流奖励、OpenAI 兼容服务、推测解码和量化部署,进一步提升运行模式复杂度。M4 论文给出最清晰的架构转向:它明确界定 Baichuan-Harness、核心推理模型和临床工具层,并加入长期患者记忆、循证检索、多模态感知、动作约束和子智能体分派。这很重要,因为它说明 Baichuan 正尝试把医疗决策支持产品化为一套受管理的交互系统,而不只是无约束的下一个 token 生成。[CE010, CE011, CE012, CE013, CE014, CE015]

技术 / 运营架构表
层 / 流程 / 组件作用依赖风险
基础模型和继承骨干提供从通用到医疗的底层模型族M2 基于 Qwen2.5-32B;M3 基于 Qwen3;通用线沿用早期 Baichuan/Baichuan2 系列依赖外部基座会收窄自研护城河,并产生适配债务
医疗训练和验证器系统让模型对齐临床问询和更安全的推理患者模拟器、验证器系统、医疗数据整理、多阶段 RL质量取决于内部评测设计,外部买家不容易审计
智能体运行时 / Harness 层在工具使用和约束上,让训练行为与部署行为保持一致Baichuan-Harness、动作护栏、记忆系统、子智能体调度复杂运行时扩大安全面,也增加运营调试负担
检索和知识库技术栈用证据和企业文档约束输出嵌入模型、向量检索、稀疏检索、证据排序检索质量和来源治理会实质影响幻觉与信任
推理和服务层把权重变成可部署 API 或私有端点Transformers、vLLM、SGLang、量化、推测解码框架兼容性和 GPU 配置可能变成支持瓶颈
多模态感知工具处理 OCR、X 光、皮肤科和临床文档解析视觉语言模型、OCR 管线、图像分析工具图像偏差、格式差异和边缘案例失败会制造临床风险
企业适配工具链为客户领域定制模型数据处理、增量预训练、微调、PPO/DPO RL、评测、压缩、部署服务占比高的适配会抬高实施成本,削弱可重复性

架构栈由官方平台页面、仓库、模型卡和论文拼出。公司不披露实现细节时,行内聚焦作用和依赖,而不是底层内部实现。

[CE007, CE008, CE011, CE012, CE013, CE014]
FE002: 客户工作流 / 运行流程

Baichuan 的医疗工作流先从问询进入受检索约束的推理,再给出辅助指导和随访记忆,而不是自主临床行动。

图中把几代模型简化为一条运行流程,因为 Baichuan 公开医疗栈越来越汇聚到检索、推理、工具使用和受控输出上。

[CE018, CE021, CE022, CE025, CE026, CE027]

5.3 部署、集成和工作流:Baichuan 正在从 API 自助到辅助私有化部署之间优化多种运行模式

Baichuan 的部署叙事相对于私营中国模型实验室而言格外明确。API 文档展示了常规自助流程:认证、充值、API key 设置和 JSON 端点;平台页和 NPC 页面则反复把更大客户导向商务咨询和辅助上手。企业平台页还列出内部工作流栈:数据处理、增量预训练、微调、PPO/DPO 强化学习、评估、压缩、模型部署、知识库和工具调用。这说明 Baichuan 卖的不只是原始推理;它把自己定位为可在客户工作流中适配和部署模型的垂直赋能层。开放权重侧,GitHub 和 Hugging Face 材料给出基于 transformers、vLLM 和 SGLang 的具体服务指南。医疗侧,ScienceNet 和 BSIA 显示,产品可以嵌入儿科和医院工作流,而不是停留在演示。缺点是,每种运行模式都会带来支持负担、集成风险,以及不同买方之间变化的信任要求。[CE007, CE008, CE009, CE016, CE019, CE028]

信任 / 质量 / 合规表
控制 / 认证 / 质量指标状态范围缺口
医疗使用免责声明已确认用户协议、M3 模型卡和 M4 论文都称,输出仅作辅助,不能替代医疗诊断 / 治疗公开材料没有展示覆盖所有部署场景的详细临床治理运营模型
动作约束和护栏声称M4 Harness 论文称,运行时会验证工具动作、数据访问和护理路径合规真实场景护栏表现缺少公开独立验证
隐私和安全政策政策层面已确认官方隐私和用户协议页面描述隐私义务和事件响应姿态审计报告、正常运行时间历史或泄露防护细节等公开运营证据有限
基于证据的检索已声明并有技术描述M4 和 M3 材料把检索定位为主要抗幻觉控制来源治理、新鲜度和检索失败处理在来源集中未经外部审计
多模态安全限制技术论文已确认M4 论文明确标出罕见病限制和图像偏差风险未找到公开部署级错误率看板或上市后监测数据

该表聚焦运营性信任控制,而不是泛泛的 AI 伦理声明。若干控制在概念上可见,但仍缺少独立生产证据。

[CE020, CE023, CE024, CE025, CE030, CE036]
路线图 / 发布 / 发展阶段表
日期 / 阶段功能 / 里程碑状态含义来源
2023 启动阶段Baichuan-7B 和 13B 开放模型;Baichuan2 多语言后续版本历史 / 已发布早期建立开放权重和多语言基础模型可信度Baichuan2 技术报告 + GitHub 仓库
2025 医疗基座建设Baichuan-M1 开放医疗模型,采用新架构和 20T 级混合语料已发布显示公司愿意做深度领域适配,而不是浅层提示词包装M1 仓库 + arXiv
2025 推理升级Baichuan-M2,配大型验证器系统、患者模拟器和 4090 部署选项已发布把技术栈推向真实世界推理和可部署性M2 仓库
2026 临床问询升级Baichuan-M3,带事实感知 RL、推测解码和兼容 OpenAI 的服务已发布提升生产就绪度和有证据托底的医疗推理能力M3 Hugging Face + arXiv 资料
2026 持续照护智能体系统Baichuan-M4,带 Harness、长期记忆、检索、多模态工具和更低幻觉率声明已发布 / 新近宣布把 Baichuan 从问答模型推向托管式临床智能体系统M4 论文 + AIBase
当前平台状态企业平台宣传智能体工作流、领域增强工具链和多行业解决方案当前 / 已上线暗示 Baichuan 想要可重复部署工具,而不只是一次性模型发布平台首页 + 官方网站

路线图根据来源集中可见的已发布公开资产和上线信息推断。它追踪有意义的能力变化,而不是每个小版本。

[CE011, CE014, CE016, CE018, CE021, CE028]
FE003: 关键依赖图

Baichuan 技术栈依赖外部基座模型、公开推理服务框架、领域数据和检索质量,也依赖医院与企业部署纪律。

DAG 标出会实质影响产品成熟度和护城河的依赖项。由于 Baichuan 未公布完整基础设施示意图,它刻意比模型图更高层。

[CE014, CE016, CE018, CE019, CE021, CE026]

5.4 差异化、数据和开发栈:Baichuan 最强切口,是建立在开放生态上的医疗专属工作流贴合度

Baichuan 对差异化的宽泛主张,并不是拥有端到端全栈。多份公开材料清楚显示,公司常常搭建在外部基础和流行推理框架之上。更强的差异化逻辑更窄,也更偏运营:以医疗为中心的训练和部署纪律,把医疗数据整理、验证器系统、临床问询建模、证据检索、记忆管理、多模态医疗感知,以及可在更受限环境中运行的部署包组合起来。M1 强调医疗数据规模和定制架构调优;M2 强调验证器引导推理和轻量部署;M3 强调低幻觉临床问询和推测解码;M4 则把这些模块重构为连续照护智能体系统。平台首页还指向一套端到端垂直增强工具链,包括数据处理、增量预训练、模型微调、强化学习、评估、压缩和部署。如果这套运行栈能证明可重复,Baichuan 就有真实切口。否则,它可能沦为跑在更快开放生态之上的一层包装。[CE011, CE012, CE013, CE016, CE018, CE021]

FE004: 产品成熟度 / 能力图

Baichuan 医疗栈在领域推理和部署针对性上最强,但独立验证和治理证据仍偏弱。

该矩阵把本章证据转换为成熟度分类,而不是数值评分。它区分公开技术丰富度和独立运营验证。

[CE005, CE007, CE016, CE018, CE021, CE028]

5.5 信任、安全、合规和技术风险:产品明确定位为辅助工具,这很谨慎,但也定义了当前上限

Baichuan 的公开医疗材料对安全边界说得格外直接。用户协议称平台不构成诊断或处方;M3 卡片称它用于研究 / 参考,并应在专业指导下使用;M4 论文称,输出不得替代医生判断,也不得直接用于最终诊断、治疗计划、处方或紧急关键决策。这是严肃医疗 AI 栈应有的姿态,但也说明 Baichuan 卖的是受控决策支持,而不是自主医疗。同一批文件还暴露了更深的技术风险:罕见病局限、图像偏差风险、长尾泛化问题,以及对可靠检索、记忆管理和工具约束的高度依赖。隐私和安全控制主要在政策层面讨论,但公开的运行证据在正常运行时间、事故报告、外部审计和真实世界错误率上仍然很薄。结果是,这套产品栈看起来技术野心强,也比通用聊天部署更安全,但仍需要在生产可靠性、合规运营和临床治理上拿出更强的独立证明。[CE020, CE023, CE024, CE030, CE036, CE040]

Chapter 06

06客户情况

6.1 客户分层和触达界面

Baichuan 的客户地图比单纯医疗优先叙事更宽。官方页面至少展示了五类对外用户群:与 Bai Xiaoyi 互动的个人 / 家庭用户;评估或部署 M 系列产品的医院和医疗服务机构;企业 API 买方和智能体平台客户;下载开放权重的开发者和自托管用户;以及银行、保险、教育、零售、制造和终端设备场景中的非医疗垂直买方。平台首页列出许多可识别企业 标识和行业解决方案通道,强化了这一判断;合作页面则邀请开发者、企业和组织进入销售辅助漏斗。从财务和战略角度看,这些并不是平等分层。医院提供高信任证明和垂直可信度。企业提供更大合同路径。开发者扩大分发和采用,但不自动等于付费账户。Bai Xiaoyi 等消费者医疗体验制造认知,也提供潜在 To C 路径,但在公开记录中,其商业耐久性最不透明。[CU001, CU002, CU003, CU004, CU014, CU015]

客户分群表
分群买方 / 用户 / 付款方使用场景规模收入 / 战略价值缺口
Bai Xiaoyi 个人用户个人 / 个人或家庭 / 目前主要是用户诊前分诊、诊后解释、家庭健康管理官方可见的产品形态,但没有公开 MAU 或付费方数量重要的 To C 心智入口,也是未来导入医疗健康服务的漏斗无公开留存、付费转化或地域拆分
医院和医疗机构医院管理层 / 临床医生 / 机构AI 儿科医生、医疗问询、临床决策支持、私有部署北京儿童医院和其他医疗机构有具名证明信任度最高的部署证明,也是最强垂直切入点合同规模、续约和生产覆盖广度仍未公开
使用 Hai Na / M 系列 API 的医疗健康服务提供方医疗服务提供方 / 临床医生或终端用户 / 机构有证据支撑的医疗 API 使用和嵌入式工作流基于资质的免费和付费接入公开,但机构数量未公开带来生态触达,也可能形成先落地再扩张的路径免费使用与付费转化未披露
企业 API 和智能体平台账户IT / 运营负责人 / 内部用户 / 公司摘要、智能体工作流、知识库、垂直自动化平台标识加上媒体报道的金融 / 企业客户和伙伴医疗之外最可能规模化的 B2B 合同层标识和具名客户看不出支出或留存
开发者和自托管社区开发者 / 开发者 / 自费或雇主付费下载、基准测试、自托管、原型开发、微调开放的 GitHub/HF 分发,以及据报道数十万次下载拓宽分发,也降低买家评估风险开发者使用不等于经常性收入
非医疗垂直买家银行、保险、教育、零售、制造业买家 / 员工 / 公司领域增强 ToB 工作流官方平台展示多条垂直赛道和企业标识支撑 Baichuan 能卖到医疗之外的论点具体生产案例稀少

分群拆开用户 / 买方 / 付款方角色,避免把开源触达、医院证明和企业销售压成一个客户叙事。

[CU001, CU002, CU003, CU004, CU005, CU010]
渠道 / 合作伙伴依赖表
渠道或合作伙伴类型依赖证据给 Baichuan 的价值风险
医院合作伙伴具名儿科和肿瘤科部署高信任客户证据和领域反馈回路采购慢、定制集成、隐私驱动的碎片化
企业平台渠道logo、业务咨询和 To B 客户报道医疗之外的收入扩张和更广工作流覆盖logo 证据可能跑在真实生产深度前面
开源分发GitHub 和 Hugging Face 工件开发者认知、全球基准测试、自托管采用评估容易,也会降低切换成本
消费者 / App 分发Bai Xiaoyi 和公开消费者配额说法建立直接终端用户触达和潜在数据 / 反馈回路这里的公开留存和变现最不透明

Baichuan 通过几条不同渠道触达客户,每条渠道的经济性和证据质量都不同。

[CU001, CU010, CU016, CU018, CU019, CU023]
FU001: 客户旅程图

Baichuan 的客户旅程因客群而异,但大多数路径都是先低门槛评估;只有领域匹配跑通后,才进入更高信任或更高价值的部署。

该旅程图根据公开产品和客户触点推断,而不是来自公司披露的内部漏斗。

[CU001, CU005, CU010, CU016, CU018, CU023]

6.2 有名可查的客户证明和生产成熟度

Baichuan 最好的有名客户证明在医疗。BSIA 和 ScienceNet 都记录了与 Beijing Children's Hospital 合作的 Futang-Baichuan 儿科模型;TMTPost 还补充说,Baichuan 已在那里部署 AI 儿科医生,同时也引用了 Cancer Hospital of the Chinese Academy of Medical Sciences 的使用案例。这些不是泛泛标识;它们是与具名机构绑定、指向具体工作流的医疗参考。在更宽企业侧,证据质量低一档,但仍有意义。36Kr 称,Baichuan 2024 年商业服务部门带来了 Bank of China、NE Digital、China Merchants Bank、Xinyada 和 Tiankai Group 等客户和合作伙伴。平台首页还展示 Tencent、Xiaomi、Intel、Didi、iQiyi、58.com 等标识,但单凭标识看不出生产深度、支出或留存。正确读法是:Baichuan 有真实部署和账户证明,但生产成熟度因分层而异;医院侧远强于一般企业披露。[CU003, CU004, CU005, CU006, CU011, CU012]

具名客户证据表
客户细分领域部署 / 用例生产环境与试点成果局限
北京儿童医院医院 / 儿科诊疗Futang-Baichuan 儿科模型和 AI 儿科医生工作流看起来已越过试点;已有公开发布和部署证据源材料中最强的具名医疗引用商业条款、使用深度和续约状态未公开
中国医学科学院肿瘤医院医院 / 肿瘤科TMTPost 提及的医疗模型部署据报道已部署显示 Baichuan 的医疗栈进入另一家头部机构证据质量来自一家强媒体,而非医院署名案例
Bank of China(中国银行)企业 / 金融2024 年商业服务客户或合作伙伴据报道存在商业关系证明 Baichuan 触达医疗之外的大型企业买家未披露成果指标或当前生产状态
China Merchants Bank企业 / 金融2024 年商业服务客户或合作伙伴据报道存在商业关系支持金融行业客户触达未披露合同规模、续约或部署深度
Tencent / Xiaomi / Didi / Intel / iQiyi 等企业生态 / 混合Baichuan 平台上的 logo 证据仅凭 logo 无法判断拓宽可见商业生态单靠 logo 不证明生产使用、成果或留存

具名证据最强之处在于部署场景清楚。没有工作流细节的 logo 和合作伙伴名称,与医院证据分开处理。

[CU003, CU004, CU005, CU006, CU015]
FU003: 客户证明矩阵

医疗部署是 Baichuan 最强的具名证明;广泛的平台标识和开发者分发提供规模表面,但留存可见度更弱。

该矩阵排序的是证明质量,而不是客户价值。它区分真正具名的工作流证据和宽泛但浅层的营销材料。

[CU003, CU004, CU005, CU006, CU015, CU020]

6.3 采用和使用信号

把用户群分开看时,公开采用图景最强。Bai Xiaoyi 和更宽医疗应用层显示,Baichuan 并不只是后端 API 供应商;它试图在诊前和诊后照护流程中触达终端用户。ToolChase 独立描述这款消费者产品免费但有每日配额,同时称开源模型家族受到开发者和研究者欢迎。M3 技术博客又在研究和客户证明之间搭了一座重要桥梁:M2 发布后吸引了数十万次下载。企业和机构使用方面,最好的规模线索更偏定性:可见 标识、具名医院部署,以及媒体报道中的具名 To B 客户。Caixin 的 2025 年订单目标和财务章节的收入代理指标说明,这些客户界面并非纯实验。但公开采用数据仍缺少活跃账户数、月活用户或少数具名医疗案例之外的已安装站点数量。[CU001, CU007, CU008, CU009, CU010, CU017]

客户增长 / 采用轨迹表
指标数值日期来源置信度含义缺失分母
2025 签约订单目标RMB1B-RMB2B2024-12-24 采访中的 2025 目标Caixin表明商业化野心明显高于零收入创业阶段未披露合同数量、细分构成或转化为收入的桥梁
2024 年 To B 收入代理指标~RMB100M2025 年报道指向 2024 年36Kr说明企业 / 商业服务已有真实付费客户活动未按产品、服务或经常性使用拆分
M2 下载牵引力数十万次下载2026 年 M3 博客回顾Baichuan M3 博客显示开发者 / 研究者漏斗顶部需求强劲下载量不等于活跃生产部署
医院部署证据Beijing Children's Hospital 及其他医院引用2025-2026BSIA / ScienceNet / TMTPost验证纯模型发布之外已有具名医疗采用未披露医院数量或续约细节
企业 logo 覆盖平台首页 20+ 个可识别 logo2026-08-21 访问Baichuan 平台首页显示跨行业商业拓展面较广logo 出现不证明付费生产使用
2024 年具名 To B 客户 / 合作伙伴Bank of China、NE Digital、China Merchants Bank、Xinyada、Tiankai Group 等企业客户2025 年报道指向 2024 年36Kr给出具体企业证据,而不是泛泛的平台话术客户数量、ACV 和当前状态未知

该表拆分收入类、下载类、部署类和 logo 类采用信号,因为它们衡量的客户质量并不相同。

[CU004, CU007, CU008, CU015, CU017]
FU002: 采用 / 部署漏斗

Baichuan 的采用漏斗顶部可见,但底部缺少衡量;留存和扩张本应在这里得到证明。

该流程故意止于一个衡量缺口,因为公开记录更强的是漏斗顶部发现,而不是漏斗后段持久性。

[CU007, CU008, CU010, CU017, CU024, CU025]

6.4 耐久性、重复使用和扩张

客户故事在耐久性上变得不完整得多。公开证据足以显示采用路径:消费者 / 应用界面、医院试点或部署、企业 API 和私有化部署通道,以及开源开发者分发。也存在一条合理的先落地再扩张逻辑:免费或低摩擦的开放权重使用可以引向 API 使用;医疗机构可以先评估或使用补贴访问,再转向私有化部署;企业智能体平台项目可以从一个工作流扩到多个工作流。但这条逻辑大多仍是推断。Baichuan 没有公开披露 NRR、GRR、流失率、续约率、合同期限、席位增长、队列曲线、CSAT 或客户集中度。即便最强的医疗证明,也没有说明部署是否付费、是否经常性、是否扩张。开发者下载显示的是漏斗顶部热度,不是已变现留存。因此,本章可以支持可重复采用路径,但还不能以机构级信心支持持久客户经济性。[CU007, CU008, CU016, CU017, CU023, CU024]

留存 / 重复使用 / 满意度表
指标值 / 空值客户段置信度尽调要求
NRR / GRR未公开披露企业和医院账户提供主要细分和头部客户的净留存与总留存。
合同期限 / 续约周期未公开披露医院和企业账户提供标准期限、续约节奏和试点转生产转化率。
消费者重复使用未找到 Bai Xiaoyi 公开 MAU/DAU/队列数据个人用户提供月活用户、30/90 天留存和分渠道付费转化。
开发者重复参与M2 据报道有数十万次下载,但没有活跃用户或生产转化分母开发者 / 自托管者提供活跃安装量、重复拉取量,以及转化为付费 API 或企业交易的情况。
满意度 / CSAT / NPS未找到公开满意度 KPI所有细分按细分和主要投诉类别提供 NPS/CSAT。

多数耐久性字段仍为空,因为公开记录更清楚地显示采用触点,而不是重复变现或续约行为。

[CU017, CU024, CU025]

6.5 集中度、采购和客户风险

客户质量受到几项可见风险约束。第一,最有说服力的客户证明集中在医疗和一组有限企业参考上,由此带来公开记录无法回答的集中度问题。第二,36Kr 认为,医疗 To B 商业化很难,因为医院数据彼此割裂,部署高度定制,百万元级年度项目成本对并非以利润最大化为目标的医院来说仍可能显得昂贵。第三,Baichuan 自己的法律和产品免责声明把医疗系统保持在辅助姿态,而不是完全自主照护角色;这很谨慎,但也收窄了能快速转化为高信任收入的工作流类别。最后,平台 标识和合作伙伴名称不等于可留存的高价值合同。因此,客户结论应保持均衡:Baichuan 比许多私营模型实验室有更强的具名证明,尤其在医疗,但采购摩擦、稀疏留存指标和不确定集中度,仍挡住了干净的客户质量研判。[CU004, CU011, CU021, CU025, CU026, CU027]

扩张与集中度风险表
扩张驱动集中度风险影响尽调路径
医院证据向更广医疗铺开客户证据可能仍集中在少数明星医院医疗护城河可能看起来强于实际装机基础要求提供完整医院客户名单、上线现场数量和生产扩张状态。
开发者下载转向付费企业使用开源热度未必顺畅转化为 API 或部署收入漏斗顶部很大,变现仍可能偏弱要求提供下载或评估转化为付费账户及 ACV 分层的数据。
平台 logo 转为真实合同logo 可能夸大生产成熟度或当前支出营销材料容易误读商业牵引力要求按 logo 映射试点、生产和流失账户。
医疗 To B 定制化项目可能过度定制,难以高效扩张客户增长带来的服务负担可能快于留存软件毛利审查实施工时、支持成本和标准化部署包使用情况。
消费者 Bai Xiaoyi 路线转向 To C 可能扩大触达,但稀释焦点或变现纪律消费者和机构客户桶并存,客户组合可能更难评估要求按产品线提供收入结构、重复使用和获客成本。

扩张逻辑存在,但每条路径的转化和集中度风险不同。

[CU021, CU022, CU026, CU027, CU028, CU029]
Chapter 07

07风险

7.1 监管、法律和合规风险是最清晰的硬性外部约束

Baichuan 处在中国监管体系下;这个体系一边支持本土 AI 发展,一边对服务提供方要求很高。2023 年 7 月《生成式人工智能服务管理暂行办法》明确规定,面向国内公众提供服务的提供方必须管理合法训练数据、个人信息处理、禁止内容控制、投诉机制、服务协议,以及——当服务具有舆论属性或社会动员能力时——安全评估和算法备案义务。这些不是抽象政策主题,而是运营要求,会影响新产品上线速度、模型训练方式,以及投资者应索取什么证据。Baichuan 自己的法律界面也强化了这不是泛化风险。公司用户协议称,产品提供咨询而非医疗诊断;隐私政策强调事故响应和数据保护义务。对一家医疗导向模型公司来说,合规负担不仅落在前沿模型治理上,也落在患者相邻信任、投诉处理和个人信息纪律上。公开记录没有展示 Baichuan 的具体 CAC 备案状态、安全评估结果或经审计合规控制,因此谨慎读法是:即便今天没有已知执法行动,监管暴露仍然重大。[CR001, CR002, CR003, CR004, CR005, CR006]

监管 / 法律风险登记表
风险 / 规则 / 问题辖区 / 买方集合当前信号可能性严重性缓释措施剩余暴露尽调路径
生成式 AI 暂行办法合规中国境内公共 AI 服务CAC 牵头的正式规则自 2023 年 8 月起生效服务协议、内容控制、隐私义务和内部合规流程高 — 要求是持续性的,不是一次性事项要求提供 CAC 备案状态、安全评估结果和内部合规责任图
算法备案 / 安全评估义务中国境内具备舆论属性或社会动员能力的服务官方办法要求特定服务做安全评估和算法备案中高可能收窄公开面向用户的范围,并分阶段发布中高 — 看不到 Baichuan 状态的公开证据要求提供备案编号、评估结果和逐产品适用范围分析
个人信息和使用记录处理医疗和消费者用户规则和 Baichuan 隐私材料强调个人信息义务隐私政策、事件响应和访问控制高 — 医疗相邻数据纪律关系到信任审查留存政策、日志处理、删除工作流和第三方数据流
医疗用途免责声明和范围控制消费者和医院用户Baichuan 称平台提供咨询,而非诊断辅助定位、人机协同使用和工作流免责声明中 — 误用或预期漂移仍可能引发争议要求提供临床治理政策、升级规则和不良事件处理流程
投诉处理和下架响应所有公共用户暂行办法要求建立投诉和举报机制已发布的服务和隐私条款意味着投诉受理和响应义务中 — 失败可能引发监管或声誉外溢要求提供投诉量、处理 SLA 和监管沟通流程

法律风险来自明确的中国 AI 治理义务和医疗相邻信任责任,而不是当前执法案件传闻。

[CR001, CR002, CR003, CR004, CR005, CR006]
FR001: 风险热力图

Baichuan 最高的剩余风险集中在算力政策、医疗经济性和监管合规,而不是单纯产品无关紧要。

该热力图是有序等级,并有来源支持。它排序剩余风险,而不是编造概率。

[CR001, CR010, CR018, CR021, CR027, CR032]

7.2 算力、出口管制和技术运营风险可能快速变化

Baichuan 是一家中国前沿模型公司,所处市场即便国产替代在改善,仍高度依赖 Nvidia 和 CUDA 兼容生态。这带来一个 Baichuan 无法直接控制的风险:美国出口规则的政策摇摆,以及随后中国侧的反向压力。CNBC 报道称,美国采取行动关闭一个漏洞,这个漏洞可能允许中国企业的海外子公司无需许可证获取先进芯片。Benzinga 基于 Reuters 的报道称,H20 需要出口许可证,且中国主要买方下单超过 $16 billion,显示中国 AI 需求仍多么暴露于这条特定产品线。Yahoo 的 Reuters 报道后来显示许可证恢复,但 Business Standard 随后描述了中国围绕 H20 出货的安全担忧和生产暂停。Baichuan 得到的教训不只是「芯片稀缺」。真正风险在于,硬件可得性、价格和部署规划都可能在几乎没有预警的情况下双向摆动。产品效率工作、更小的私有化部署占用和国产芯片适配或许能缓冲冲击,但无法消除算力受限政策环境带来的运营不确定性。[CR010, CR011, CR012, CR013, CR014, CR015]

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暴露未解决缺口
先进 GPU 出口管制反复低中 — Baichuan 可以优化模型,但控制不了美国许可政策高 — 算力规划和成本假设可能突然变化Baichuan 未公开披露 GPU 库存、国产芯片占比或云合同
中国监管方对进口 AI 芯片的担忧低中 — 替代采购和国产替代可能有帮助中高 — 即便美国批准,中国侧限制也可能扰乱供应Baichuan 未按芯片系列公开供应商暴露
医疗幻觉 / 可靠性风险极高中 — 医疗专用模型和辅助性免责声明有帮助高 — 一个糟糕的临床相邻工作流就可能过度伤害信任未按医疗任务、医院或工作流公开质量 KPI
医院部署中的重服务定制负担低中 — 私有部署和更窄用例可能降低复杂度高 — 交付人力可能跑赢软件毛利未披露部署时间、服务毛利或实施成本
模型 / API 商品化与切换低中 — 医疗聚焦和具名部署是部分防线高 — 开放权重和 API 兼容性降低客户锁定未公开相较 Qwen、DeepSeek 或 Z.ai 的赢单率或续约数据

运营风险最高的地方,是产品复杂度、算力依赖和医疗信任交汇处。

[CR010, CR011, CR012, CR013, CR014, CR015]
FR002: 风险传导图

Baichuan 的主要风险通过合规负担、算力波动和医疗经济性传导到利润率、增长、融资和估值。

该图展示公开证据暗示的因果方向,不是概率模型。

[CR002, CR011, CR019, CR028, CR031, CR035]

7.3 竞争、价格战和依赖风险更威胁利润率,而不是需求相关性

Baichuan 最可能的失败方式不是马上失去相关性,而是经济性被挤压。公司有足够的产品深度,也拿出了医疗场景证据,仍然值得被关注;但它卖进的是中国大语言模型市场,定价、分发和基准领先权已经被 Qwen、DeepSeek、Doubao、Z.ai、Moonshot、MiniMax 等资金更厚或渠道更强的对手塑形。Apidog 2026 年价格战调查、Digital Applied 的服务商份额报告,以及 KrASIA 对「AI 老虎」的表述,都指向同一个结论:中国大语言模型市场很活跃,但并不宽容。开放权重和 OpenAI 式 API 通道降低客户评估摩擦,也降低切换成本。战略投资人和渠道伙伴能提高生存概率,但 Baichuan 缺少 Alibaba、Baidu、ByteDance 那种内置云、搜索或社交分发引擎。公司因此更依赖产品专业化、销售执行,以及医院或企业场景验证。利润率风险尤其突出:通用模型价格持续下压,医疗部署却可能更慢、更定制化,也更难标准化。[CR018, CR019, CR020, CR021, CR022, CR023]

合作伙伴 / 依赖风险登记表
依赖交易对手 / 生态角色集中度失败场景严重性缓释措施剩余暴露
Nvidia / CUDA 生态Nvidia 和 GPU/云供应链训练和高性能推理经济性政策或采购冲击抬高成本,或拖延产品路线图极高模型效率工作、国产适配、更小部署足迹
医院和医疗合作伙伴Beijing Children’s Hospital 和其他医疗机构证据、工作流集成和领域可信度中高试点或明星部署没有扩展成广泛付费采用多个医疗产品和更广企业栈中高
战略投资者和渠道盟友Xiaomi、Alibaba Cloud、37 Interactive、企业合作伙伴资本、云或生态入口支持停留在财务层面,没有转化成耐久客户分发分散客户触点,保留直接产品驱动分发中高
企业平台潜在客户具名企业客户和 logo 生态医疗之外的商业扩张logo 夸大生产深度或当前支出将 logo 绑定到可衡量部署和续约中高
开源社区GitHub、Hugging Face、自托管者评估、认知和漏斗顶部需求开发者拿走模型,但把变现使用切到别处把开源评估桥接到 API 和企业工作流

Baichuan 的依赖图表面上多元,但若干节点披露很浅,因此很难承销。

[CR018, CR019, CR020, CR021, CR022, CR023]
FR003: 依赖关系图

Baichuan 的依赖网络穿过外部算力、医疗伙伴、企业客户证明和资本支持,而不是压在单一主导节点上。

依赖关系图强调交易对手和生态杠杆点,而不是所有权结构。

[CR010, CR018, CR020, CR023, CR024, CR025]

7.4 医疗商业化与执行风险,仍是最难穿透的内部投资判断题

Baichuan 的医疗重心既是战略差异化,也是一项高要求的经营选择。已有具名医院案例,M 系列加上 Futang-Baichuan 叙事也是实打实的优势。但关于公司的最佳反向报道指出,医疗 To B 项目可能高度定制,数据接入分散,医院采购慢,百万元人民币级年度项目成本对客户仍可能偏贵。这让医疗切口两面性很强。它能提高产品匹配度和信任,也可能把公司困在服务含量高的实施、续约质量不清晰、低毛利定制里。Caixin 报道的 2025 年签约目标和财务章节的收入信号说明商业化在推进,但仍无法回答业务中有多少是经常性软件收入,多少是项目制部署。与此同时,公开记录对管理层梯队深度、客户集中度,以及创始人 Wang Xiaochuan 个人影响之外的组织韧性披露仍薄。实际风险在于,Baichuan 在医疗方向可能判断正确,却要同时扛住前沿模型研发、监管义务、企业销售和医院级交付,运营上被拉得过紧。[CR027, CR028, CR029, CR030, CR031, CR032]

人员 / 执行 / 财务模型风险登记表
角色 / 职能或问题依赖或缺口可能性严重性缓释措施尽调路径
创始人主导战略公开叙事仍高度围绕 Wang Xiaochuan 和少数可见高管即便未完全公开,联合创始人和资深商业化负责人可能仍然存在要求提供组织架构图、继任计划,以及产品、医疗、销售之间的权限划分
医疗转向执行需要把前沿模型研发、场景部署和临床信任拼起来M 系列聚焦和医院合作提高医疗场景契合度审查实施待办、部署周期和医院使用率指标
订单到收入转化质量公开订单和收入代理指标可见,但合同经济性仍不透明资本支持为商业化迭代争取时间审查 ARR 与项目制收入组合、按产品划分的毛利率和递延收入
烧钱速度和现金跑道不透明公开估值和融资可见;看不到经审计的运营烧钱中高深厚投资人基础和战略资本降低短期资不抵债风险要求提供月度烧钱、GPU 承诺、招聘计划和下行情景现金跑道模型
人才竞争中国 AI 头部公司和大科技巨头能在算力和人才上出价压过创业公司使命感和医疗切入点有助于招聘要求按技术职能提供流失率,并给出相对头部同业的薪酬基准

执行风险不只看产品质量;更在于 Baichuan 能否把医疗主导的模型公司规模化,同时避免过度依赖创始人、定制项目或补贴式增长。

[CR027, CR028, CR029, CR030, CR031, CR032]

7.5 风险结论与投资逻辑破裂触发点

因此,投资问题变成:Baichuan 能否在市场围绕它商品化之前,把一个受政策塑形、资本密集、价格受压的机会,转成以医疗为核心的、可持续的企业经济性。现有公开证据支持一个平衡但谨慎的答案。没有单一来源证明公司已经失灵;但公开证据也不足以把主要风险轻描淡写成普通创业公司的成长阵痛。正确的尽调姿态是盯住风险之间的相互作用。医疗场景高度依赖信任,监管审视因此更重要。通用模型价格已经承压,出口管制波动因此更重要。医院部署看起来是战略核心,客户不透明因此更重要。公司仍是私营且披露有限,创始人或执行集中度因此更重要。投资逻辑不是被一条新闻打断,而是在其中两三项风险相互强化时破裂——比如,芯片政策再次收紧,同时价格压缩加深,医疗部署又无法转成可重复、更高毛利的客户账户。[CR035, CR036, CR037, CR038, CR039, CR040]

风险缓释与止损标准表
风险可监控触发项阈值 / 事件行动含义
CAC / AI 治理不合规监管备案缺口、调查通知或产品暂停任何正式执法行动、上线公开产品缺少强制备案,或被迫回滚服务暂停评估医疗主导扩张,直到合规证据补齐
算力供应冲击美国新限制、中国侧采购停摆或成本明显飙升连续 2+ 个季度失去合规 Nvidia 供应,或发生重大被迫迁移成本上调基础设施成本假设,并大幅下调增长 / 毛利率预期
医疗部署经济性失效具名医院验证不能转化为软件式经常性收入连续两个重大医疗部署仍偏项目制,且缺少扩张或续约证据将医疗切入点视为信任背书,而非变现护城河
国内价格战加深竞争对手前沿 API 价格再次下探,而 Baichuan 无法证明溢价定价市场再次广泛重定价,且没有企业锁定效应或垂直行业付费意愿抵消压缩估值倍数;新资金评估前要求毛利率桥接证据
客户集中 / 不透明持续对头部客户、续约或分部组合没有可用披露尽调后期资料室仍缺少集中度、NRR 和合同层级可见度因经济性仍无法支撑投资判断,从「可跟踪风险」移向「投资逻辑破裂」区间

关键投资逻辑破裂触发因素来自政策动作、合同证据和部署经济性,而不是模糊叙事恶化。

[CR003, CR010, CR019, CR027, CR035, CR036]
Chapter 08

08估值

8.1 定价背景与当前估值标记

Baichuan 的估值故事有一个硬锚,也有几个软锚。硬锚是 2026 年 4 月 Series B 融资,据报道估值约 $2.8B,累计融资处于高数亿美元区间,战略投资人包括 37 Interactive、Xiaomi 和 Alibaba Cloud。这不是种子期或科研项目估值,而是商业化阶段 AI 公司的估值标记。但公司仍处在私募市场里一个典型区间:估值可以跑在披露前面。公开证据显示,公司有真实的收入表面、具名医院案例和企业牵引信号;但投资人通常用来支撑私募估值的经审计输入——确认收入、续约质量、分部利润率、项目制与软件收入结构、客户集中度、优先权条款——仍然缺位。实际含义是,Baichuan 当前估值不能单独读成低估或高估的清晰证明。更合理的读法是:市场为中国医疗 AI 与更广泛企业模型平台的战略可选性协商出的价格。公开证据足以支撑继续关注,但不足以支撑不计价格买入。[CV001, CV002, CV003, CV004, CV005, CV006]

建议摘要表
维度当前判断证据质量决策含义
建议继续研究 / 观察不要把最近可见的 $2.8B 估值本身当作买入信号。
信心中低公司质量看得见;估值质量证据仍不完整。
风险评级价格战、医疗交付复杂度、算力 / 合规风险同时重要。
估值立场$2.8B 下合理偏满;大约低于 $2.3B 更有吸引力需要更低入场价,或明显更强的收入质量证据。
持有 / 退出框架新资金若以 $2.8B 进入,稀释前需要约 $4.2B-$5.6B 才有 1.5-2x,约 $5.6B-$8.4B 才有 2-3x中低回报算术并非不可能,但没有头部 AI 赛道叙事看起来那么宽松。

建议取决于价格和证据。Baichuan 可以是一家好公司,但在最后披露的私人估值下,不一定是有吸引力的新进场机会。

[CV001, CV004, CV006, CV010, CV027, CV031]
投资逻辑 / 反向逻辑表
支柱投资逻辑反向逻辑哪些证据会改变判断
医疗差异化Baichuan 拥有真实的医疗模型和医院验证,这是许多通用型对手缺少的。医疗验证仍可能偏服务、推进慢,且难以用软件式毛利率变现。来自多个上线医疗部署的续约、扩张和毛利率数据。
商业化动能订单目标、具名客户和可见价格界面显示,公司已不再是零收入实验。公开信号尚未证明收入质量、留存或客户集中纪律。经审计收入桥接和分部留存。
相对同业创业公司的定价Baichuan 定价远低于 Moonshot 和 DeepSeek 这类突破型私人市场龙头。那些估值更高的同业也有更强的公开规模、消费端分发或资本市场证明。证据需显示:即便可见度较低,Baichuan 的医疗切入点也能带来更强韧性。
相对上市巨头的溢价如果 Baichuan 复合增长更快,相对 Baidu 或 Alibaba 享有私人 AI 成长溢价可以成立。上市巨头已经披露规模,交易倍数接近低个位数 P/S;Baichuan 需要证明更强的增长溢价,但目前证据还不够。更清晰的增长、毛利率和可复制性证据。
战略可选性医疗、企业 AI 和开源分发叠加,给公司留下多条胜出路径。战线太多也可能意味着焦点分散、烧钱更高、变现更模糊。更细的产品线经济性和资源配置披露。

这张表把公司质量和特定价格下的可投资性拆开;这正是核心估值问题。

[CV007, CV011, CV013, CV014, CV019, CV023]
FV001: 建议逻辑

建议来自一条简单链条:真实产品和客户证明足以支撑持续关注,但收入质量和风险证据还不足以在最后可见估值上给出买入判断。

[CV001, CV004, CV011, CV027, CV031, CV034]
FV004: 投资 KPI

当前 Baichuan 估值判断的快速评分卡。

[CV001, CV004, CV005, CV024, CV031, CV032]

8.2 可比框架与倍数读法

看 Baichuan 估值,最有用的镜头不是单一直接可比公司,因为这样的公司不存在。Alibaba、Baidu 等中国上市平台巨头按低个位数市销率交易,反映的是更成熟、增长更慢但披露强得多的业务。Nvidia 的倍数高得多,但它是全球算力平台,收入规模和公开市场流动性都异常强。中国私营 AI 创业公司的可比样本又讲了另一套故事:Moonshot 据报道的 $20B 融资、DeepSeek 估值超过 CNY330B 的巨额融资,以及 MiniMax 的公开融资或 IPO 路径,都说明市场愿意为可见规模、基准动能或消费者分发支付多高溢价。Baichuan 还没有这一层级的公开证明。Digital Applied 的服务商份额研究和竞争章节都把 Baichuan 放在二线但仍可信的专业玩家位置,而不是通用模型赢家组的顶部。这意味着 Baichuan 相比上市在位企业应拿到一定私募 AI 成长溢价,但也应相对资金最足的突围实验室折价。因此,当前 $2.8B 估值按相对位置看可以理解;但把客户不透明、医疗项目风险和价格压缩纳入后,并不明显便宜。[CV011, CV012, CV013, CV014, CV015, CV016]

可比估值表
可比对象指标倍数 / 估值 / 状态参考意义局限性
Alibaba公开市场市值和 P/S市值 ~$312.9B;P/S ~2.15x显示一家已披露的大型中国 AI / 云平台在公开市场的交易水平业务太成熟且多元,不能直接对标创业公司
Baidu公开市场市值和 P/S市值 ~$31.2B;P/S ~1.71x中国 AI 平台参考对象,有直接模型敞口和公开披露业务仍比 Baichuan 宽得多,也成熟得多
Nvidia公开市场市值和 P/S市值 ~$5.25T;P/S ~20.7x展示极致 AI 基础设施规模能拿到多高倍数不是中国私人模型创业公司的直接经营可比
Moonshot AI私人轮估值May 2026 据报道估值 ~$20B显示资本给消费端可见度更强的突破型中国 AI 创业公司多高溢价据报道的私人估值,公开财务披露有限
DeepSeek私人融资估值2026 融资轮据报道估值 >CNY330B显示资本市场能给突破型技术赢家多高溢价极端异常值;不是 Baichuan 的干净基准
MiniMax公开融资 / IPO 信号Hong Kong IPO 募资 ~$619M;市场可见度强于 Baichuan表明知名度更高的中国 AI 同业能进入资本市场IPO 募资额不等于企业价值或持久经济性
Baichuan AI 当前估值最新披露的私人估值April 2026 据报道 Series B 估值 ~$2.8B当前投资评估和回报算术的锚点需要更深入的股权结构表、优先权和收入质量背景

可比组有意混合:上市巨头用于底部倍数,前沿上市 AI 基础设施用于溢价背景,中国 AI 私人估值或融资记录用于相对创业公司定位。

[CV001, CV011, CV012, CV013, CV014, CV015]

8.3 乐观、基准、悲观情景与建议

合理的乐观情景存在。它要求 Baichuan 把医疗可信度转成可重复的企业经济性,证明 2025 年签约是真实且可转化的业务,而不主要是试点阶段承诺,并守住足够定价权,让国内 API 价格战不至于抹掉毛利潜力。在这个世界里,今天的估值标记可能显得保守,估值升至 $3B 中段到 $4B 中段变得说得通。基准情景更谨慎。公开证据最强的是战略方向和产品相关性,不是经审计经济性。公司真实、有资金、也有差异化,但仍难以高确信度承销。这把最可辩护的现值区间压在约 $2.0B-$2.6B,也让 2026 年 4 月估值显得合理偏满。悲观情景不是破产,而是商业化不达预期。如果医疗 To B 仍然定制化且受预算约束,如果价格竞争加深,或者硬件和合规摩擦拖慢部署,Baichuan 的公允价值可能跌向约 $1.0B-$1.6B。这样的分布支持继续研究 / 观察建议,而不是按最后公开价格给出买入判断。[CV023, CV024, CV025, CV026, CV027, CV028]

乐观 / 基准 / 悲观情景表
指标乐观情景(25%)基准情景(50%)悲观情景(25%)
2026/27 收入运行率框架RMB1.5B-RMB2.0B 已确认收入或清晰合同收入路径RMB0.8B-RMB1.2B 收入质量路径,项目制 / 软件经济性混合RMB0.3B-RMB0.6B 路径,转化慢且定价权弱
毛利率方向医疗和企业部署标准化后,软件式收入组合改善服务和折扣仍占较大比重,毛利率改善缓慢价格战和定制化让利润率维持低位
客户质量具名医院和企业验证扩展为多个垂直行业的重复账户有一些强账户,但集中度和续约仍只得到部分证明试点型或定制型账户占主导
估值倍数逻辑高增长私人 AI 专精公司溢价仍有支撑中等私人市场成长溢价,但低于突破型实验室折价至较低私人市场或上市科技公司倍数区间
隐含当前估值区间~$3.4B-$4.5B~$2.0B-$2.6B~$1.0B-$1.6B
以 $2.8B 入场的新资金回报逻辑按当前价值为 1.2x-1.6x;若下一轮估值或退出成功,2x+ 有可能只有执行改善且稀释保持合理,新资金才大致跑得通下轮降估值或低回报结果变得可能
关键触发因素医疗切入点证明可复制且能收取溢价商业化延续,但质量参差医疗 To B 停滞、价格战加深,或合规 / 算力冲击打到执行

情景区间是启发式、由证据支撑的判断,不是伪精确 DCF 输出。公开记录太薄,无法建立干净的经审计现金流模型。

[CV023, CV024, CV025, CV026, CV027, CV028]
FV002: 估值敏感性

估值对不同情景下收入质量、倍数和当前价值区间的敏感性。

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

围绕 Baichuan 当前估值标记、基准情景公允价值和情景上下行的区间视图。

[CV001, CV023, CV024, CV025, CV027, CV028]

8.4 最终尽调问题与投资逻辑破裂触发点

缺失的尽调工作异常具体。投资人主要不需要更多 Baichuan 雄心叙事;他们需要能把雄心转成可承销价值的经济性。最高优先级问题包括分部收入结构、头部客户集中度、医疗与通用企业客户的续约或扩张、项目服务负担、毛利率桥、GPU 和云承诺,以及最新一轮附带的任何优先权条款或附函。同样重要的是,投资人需要公司层面证据,确认面向公众的线上产品是否取得 CAC 或备案状态,因为医疗 AI 是一个合规错误能迅速污染估值的领域。若两件事同时发生,投资逻辑就会破裂:商业化证据持续不透明,而外部压力恶化。实际表现可能是国内价格重置更剧烈、医疗部署转化更慢、算力政策冲击,或有证据显示当前估值假设的收入质量强于公司能够证明的水平。除非这些缺口闭合,正确姿态是纪律性跟踪,而不是热情承销。[CV034, CV035, CV036, CV037, CV038, CV039]

投资逻辑破裂和止损触发因素表
触发因素阈值对投资逻辑的传导行动含义
医疗验证不能变现下一轮复盘周期内,具名医疗部署仍像试点或偏项目制击穿溢价差异化的核心论据将公允价值推向悲观区间,并把医疗视为信用背书而非护城河
国内价格重置再次加深中国前沿 API 价格大范围明显下探,且没有足够溢价需求抵消同时压缩毛利率和倍数假设下调估值区间,并要求证明溢价留存
算力政策冲击美国新限制或中国侧 GPU 停摆明显损害供应规划放慢模型路线图并抬高基础设施成本下调增长假设并提高下行概率
合规缺口浮现产品缺少必要备案 / 评估证据,或面临正式行动把可管理风险变成直接估值损伤暂停新的投资评估,直到修复
收入质量不透明持续资料室仍缺少分部毛利率、集中度、续约和优先权清晰度即使表面估值看似合理,也无法支撑投资判断保持观察 / 继续研究,而不是上调

关键估值杀手是组合事件:同时削弱经济性和信心。

[CV027, CV031, CV035, CV036, CV037, CV038]
最终尽调问题表
主题缺失证据重要性负责人 / 尽调路径
已确认收入和毛利率桥接没有公开的经审计分部损益或毛利率桥接估值取决于 Baichuan 到底像软件公司、偏服务,还是靠战略补贴支撑要求按 API、企业部署和医疗产品提供月度或季度收入,并按产品线提供毛利率
客户集中度和续约没有公开 NRR/GRR、头部账户敞口或合同期限披露具名客户如果没有续约和集中度数据,不能证明持久价值要求提供前 10 大客户、续约率、队列扩张和试点到生产转化
股权结构表和优先权条款上一轮估值公开,优先股堆叠不公开回报算术取决于清算优先权、按比例跟投权和附函要求提供最新股权结构表、优先权条款和任何投资人特定保护
GPU / 云承诺没有公开算力合同或供应商敞口拆分基础设施成本和路线图风险是下行情景核心要求提供云支出、GPU 库存、供应商集中度和备选情景
按产品划分的监管状态Baichuan 具体 CAC 备案 / 评估证据未公开合规滑坡会同时损害增长和估值信心要求提供上线产品的备案编号、评估结果和内部合规签字

这些问题是把 Baichuan 从有意思的公司推进到可大额承做投资所需的最低材料包。

[CV004, CV005, CV009, CV034, CV035, CV036]

免责声明

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

证据索引

结论
编号陈述可信度来源
CO001 Baichuan AI's official homepage states that the company was founded on 2023-03-24 by Wang Xiaochuan. SO001
CO002 The official mission is to help the public access world knowledge and professional services and to build a leading Chinese model base through language AI. SO001
CO003 Baichuan says its core team came from Sogou, Baidu, Huawei, Microsoft, ByteDance, and Tencent. SO001
CO004 The official homepage says Baichuan released Baichuan-7B and Baichuan-13B within 100 days of founding and that downloads exceeded one million. SO001
CO005 Baichuan's current homepage foregrounds the Baixiaoyi AI family doctor and the Haina Baichuan medical API program rather than a generic chatbot brand. SO001
CO006 The homepage footer gives a Haidian District, Beijing contact address at Zhongguancun East Road 1, Building 8, 15th Floor, B1701. SO001
CO007 Baichuan's terms and privacy policy identify Baichuan Intelligent Technology Co., Ltd. as the service provider and place disputes under PRC law and Haidian court jurisdiction. SO005, SO006
CO008 The open-platform homepage markets enterprise agents, industry workflows, and vertical models for medical, finance, and education use cases. SO002
CO009 Baichuan's API documentation exposes authenticated chat-completions APIs, tool-calling support, JSON output mode, model IDs, and rate limits, indicating an active developer-platform commercialization layer. SO003
CO010 TechCrunch reported that Wang Xiaochuan stepped down from Sogou in late 2021, launched Baichuan in April 2023, and had earlier said China needed its own OpenAI. SO007
CO011 Wang Xiaochuan's public founder-market fit comes from search, language technology, and Sogou operating history rather than from a previously disclosed cloud-platform or listed-software CFO bench. SO001, SO007
CO012 The reviewed official surfaces do not publish a board roster, committee structure, finance chief, or broad executive bench beyond the founder-centered narrative. SO001, SO005, SO006
CO013 Caixin identifies Ru Liyun, former Sogou COO, as Baichuan's cofounder and president and says the company focuses on financial, education, and healthcare deployment. SO011
CO014 Baichuan's user agreement says outputs in medical, legal, education, news, and investment contexts are not substitutes for professional advice. SO005
CO015 Baichuan's privacy policy says the company operates under PRC cybersecurity and personal-information law and maintains a dedicated emergency-response team for data incidents. SO006
CO016 The official homepage displays ICP, B2 telecom, public-security, and medical network-information filing references, showing a visible compliance surface for an internet and healthcare-adjacent service. SO001
CO017 TechCrunch reported that Baichuan quickly pocketed $50 million in financing from angel investors after launch. SO007
CO018 TMTPost reported that Alibaba and Tencent joined Baichuan's funding in October 2023. SO008
CO019 TMTPost reported that Baichuan raised RMB5 billion, about $690 million, in July 2024 from Alibaba, Xiaomi, Tencent, Asia Investment Capital, CICC, and state AI funds. SO009
CO020 Tracxn lists Baichuan funding rounds on 2023-04-10, 2023-10-17, 2024-04-16, and 2024-07-25, including a $691 million July 2024 round at a $2.7 billion post-money valuation. SO013
CO021 CB Insights says Baichuan has raised $1.038 billion over seven rounds. SO012
CO022 CB Insights says Baichuan's latest post-money valuation was $2,899.11 million in April 2026. SO012
CO023 CB Insights explicitly names 37 Interactive Entertainment as an investor in Baichuan's April 1, 2026 Series B and lists 13 investors overall. SO012
CO024 TMTPost reported in September 2025 that Baichuan was preparing a Series B around a $2.75 billion valuation. SO010
CO025 Official and repository materials show that Baichuan maintains both general-purpose and healthcare-specialized model families, spanning Baichuan-7B, Baichuan-13B, Baichuan2, M1, M2, M3, and Baichuan4. SO001, SO016, SO017, SO020, SO025
CO026 The Baichuan-M2 repository describes M2 as Baichuan's second medical-enhanced model, built for real-world medical reasoning with a large verifier system. SO015
CO027 TMTPost reported that Baichuan-M2 can be deployed on a single RTX 4090 after quantization and is compatible with mainstream domestic chips for hospital rollout. SO010
CO028 TMTPost reported that Baichuan-M2 scored 60.1 on HealthBench, above OpenAI's gpt-oss120b at 57.6. SO010
CO029 ScienceNet and the Beijing software-industry association reported that Baichuan and Beijing Children's Hospital released the Futang·Baichuan pediatric model in March 2025. SO022, SO023
CO030 TMTPost reported that Baichuan had already deployed AI pediatricians in Beijing Children's Hospital by the time of the Baichuan-M2 launch. SO010
CO031 Baichuan's Haina Baichuan program offers free M3 Plus API access to organizations serving medical workers, subject to real-service-scene restrictions and powered-by branding requirements. SO001
CO032 The platform homepage still markets finance and education vertical models alongside healthcare, showing that the company has not fully abandoned a broader enterprise story. SO002, SO011
CO033 KrASIA groups Baichuan among China's AI tigers but implies the cohort still has to prove sustainable business models against larger tech incumbents and crowded market economics. SO021
CO034 Baichuan's visible licensing, terms, privacy policy, and medical disclaimers show that trust and compliance are core parts of the company profile rather than side notes. SO001, SO005, SO006
CO035 No reviewed public source disclosed audited revenue, ARR, customer count, or headcount for Baichuan as of the run date. SO001, SO010, SO011, SO012
CO036 Taken together, Tracxn's $2.7 billion July 2024 valuation and CB Insights' roughly $2.9 billion April 2026 valuation imply only a modest public step-up into the Series B period. SO012, SO013
CO037 Baichuan's 2024 public ambition for a much larger Series B valuation did not become the main public benchmark in later reviewed sources, which cluster nearer $2.75-$2.9 billion. SO009, SO010, SO012
CO038 Public materials reviewed for this chapter do not disclose board composition, a CFO, or formal governance committees, making governance transparency a material diligence gap. SO001, SO005, SO006
CO039 The homepage and policy documents consistently place Baichuan in Beijing's Haidian district but use different street-level addresses, so city-level headquarters is clear while exact office normalization remains slightly messy. SO001, SO005, SO006
CO040 GitHub and Hugging Face releases keep Baichuan active in open-source developer ecosystems even though the main homepage has pivoted toward healthcare applications. SO015, SO016, SO017, SO020, SO024, SO025
CO041 The presence of 37 Interactive as a named Series B investor suggests Baichuan's later funding base broadened beyond cloud and telecom-adjacent strategic capital into other listed Chinese corporates. SO012, SO014
CM001 Baichuan competes in China's domestic foundation-model application, deployment, and healthcare-workflow layer rather than in the full AI infrastructure economy. SM001, SM002, SM003
CM002 Baichuan's public platform monetization surface is API and agent usage across general and vertical models, not chip sales or generic cloud infrastructure. SM002, SM003
CM003 Baichuan's homepage, Haina Baichuan program, and pediatric deployment evidence make healthcare AI a core part of the company's addressable market, not a side experiment. SM001, SM018, SM019
CM004 China's 2023 generative-AI rules make compliance a recurring cost of doing business for Baichuan's public-facing services. SM015, SM025
CM005 IMARC estimates China's generative-AI applications market at USD 5.16082 billion in 2025 and USD 19.56 billion by 2034. SM007
CM006 Gartner forecasts worldwide AI spending of USD 2.59 trillion in 2026, up 47% year over year. SM013
CM007 Forrester says China's AI infrastructure spending will exceed USD 70 billion in 2026, which is adjacent to but not directly the same as Baichuan's revenue layer. SM014
CM008 China Daily, citing IDC, says enterprise or public-cloud MaaS token usage in China rose from 114 trillion tokens in 2024 to 1,944 trillion in 2025 and could reach 40,000 trillion in 2026. SM011
CM009 TrendForce reported that Chinese AI models reached about 15% global share in November 2025, up sharply from about 1% a year earlier. SM012
CM010 IDC FutureScape says 80% of China's top-1,000 enterprises will prioritize AI sovereignty by 2027. SM010
CM011 The May 2026 AI-agent implementation guidelines identify 19 application scenarios across scientific research, industry, consumption, public well-being, and social governance. SM008
CM012 IDC's China AI agent market map includes healthcare, finance, retail, automobile, and government as target industries around model, application, and agent development platforms. SM010
CM013 Because Baichuan visibly targets healthcare, finance, education, and API developers, its SAM is narrower than full China AI but broader than a single hospital-assistant application. SM001, SM002, SM006
CM014 TMTPost cites Frost & Sullivan projections that China's AI healthcare market could expand from USD 1.2 billion in 2023 to USD 42.5 billion in 2033. SM004
CM015 ScienceNet and BSIA show that Baichuan's pediatric partnership is already tied to real pediatric-care deployment narratives rather than a purely conceptual healthcare use case. SM018, SM019
CM016 Healthcare buyers in Baichuan's market care about private deployment, low hallucination risk, Chinese clinical context, and clear operational boundaries. SM001, SM004, SM022, SM025
CM017 Enterprise teams in finance and education are secondary Baichuan buyer groups because the platform still markets those verticals even though healthcare leads the public narrative. SM002, SM006
CM018 Developers and healthtech partners form another buyer class through Baichuan's API docs, GitHub repositories, and Hugging Face model cards. SM003, SM022, SM023
CM019 Status-quo substitutes for Baichuan range from human-only clinical workflows and legacy enterprise software to larger domestic model vendors and open-weight self-hosting stacks. SM001, SM002, SM010, SM017
CM020 Demand for domestic models in China is shaped by AI sovereignty and local compliance requirements, not just by low list prices. SM008, SM010, SM015
CM021 Healthcare and regulated-enterprise deployment in China face registration, security, and content-governance obligations that raise launch and support costs. SM008, SM015, SM025
CM022 CNBC reported that the U.S. closed a loophole that had allowed Chinese firms to buy advanced Nvidia and AMD AI chips through overseas subsidiaries without a license. SM009
CM023 IDC's China foundational-model player map explicitly includes Baichuan and Baichuan-M2 among the country's tracked LLM and reasoning-model players. SM010
CM024 Apidog's 2026 Chinese LLM price-war comparison shows that frontier-model list prices in China have compressed into a tightly competitive band. SM016
CM025 Baichuan's official materials position enterprise optimization, evidence anchoring, and private medical deployment as market differentiators rather than pure consumer scale. SM001, SM002, SM022
CM026 Because Chinese model APIs expose similar chat-completions interfaces and public price sheets, switching friction for many developer and enterprise workloads is lower than it would be in a proprietary application market. SM003, SM016
CM027 China's model market is scaling on usage faster than on proven revenue quality, which means rising token volume cannot be treated as the same thing as durable software economics. SM011, SM013
CM028 Baichuan's healthcare-first public positioning likely narrows headline TAM versus consumer-chat leaders but may improve willingness to pay in privacy-sensitive and workflow-critical settings. SM001, SM004, SM018
CM029 Healthcare AI adoption creates a four-sided buyer map in which hospitals, clinicians, patients, and service operators do not always share the same incentives or budgets. SM001, SM018, SM019
CM030 The May 2026 AI-agent guidance treats public well-being as a targeted application area, which is supportive context for healthcare-oriented AI vendors. SM008
CM031 KrASIA's AI-tigers framing suggests independent Chinese model startups still must prove sustainable business models against better-capitalized tech giants. SM017
CM032 Forrester warns that rising costs, hardware volatility, and sovereignty mandates will erode purchasing power across Asia Pacific, which matters for Chinese enterprise AI budgets too. SM014
CM033 Gartner expects enterprises to expand their use of both embedded and model-layer AI in 2026, supporting the view that enterprise demand is broadening even if winners are unsettled. SM013
CM034 Baichuan's Haina Baichuan program requires real-service usage, powered-by branding, and no accuracy-degrading output modifications, showing that deployment governance is part of the product offer. SM001
CM035 Baichuan's medical-market pitch is strengthened by the claim that M2 can run on a single RTX 4090 after quantization and on mainstream domestic chips, lowering private-deployment friction for hospitals. SM004, SM022
CM036 The public market lenses for Baichuan use incompatible units such as dollars, infrastructure spend, token volume, and policy/adoption counts, so they should be treated as non-additive bounds rather than one TAM. SM007, SM010, SM011, SM013, SM014
CM037 Baichuan's realistic SOM is probably concentrated in Chinese healthcare and regulated-enterprise workflows instead of the mass-market consumer AI category. SM001, SM004, SM006, SM018
CM038 Public sources do not quantify hospital procurement depth, reimbursement pathways, or realized budget capture for Baichuan's healthcare deployments. SM018, SM019, SM021
CP001 Digital Applied says ten providers cover essentially all meaningful Chinese AI output in Q2 2026, while Baichuan sits in a second-tier niche band behind the top group. SP004
CP002 Baichuan is a pure-play startup competitor rather than a big-tech-embedded model family. SP001, SP004
CP003 Baichuan's official public wedge is healthcare and family-health workflows rather than a broad consumer assistant or cloud-platform identity. SP001, SP002
CP004 Z.ai publicly presents itself as an advanced chatbot and agent powered by GLM-5.2. SP007
CP005 BigModel surfaces Zhipu's platform identity and reinforces its enterprise-oriented positioning. SP008
CP006 Moonshot and Kimi publicly market a general-purpose knowledge-work and agentic-coding experience rather than a narrow vertical solution. SP009, SP010
CP007 MiniMax's official site presents a broad multimodal and agentic product scope. SP012
CP008 DeepSeek's official surfaces span web, app, API, and a frequent release cadence. SP014
CP009 Alibaba's Qwen and Baidu's ERNIE are embedded inside broader incumbent ecosystems rather than standing alone as startups. SP018, SP020
CP010 Doubao is best understood as a consumer-first ByteDance AI surface, not a healthcare specialist. SP022
CP011 TechCrunch reported that Moonshot raised $2 billion at a $20 billion valuation in May 2026. SP011
CP012 TechCrunch also reported that Kimi K2.6 was the second-most used LLM on OpenRouter at the time of Moonshot's raise. SP011
CP013 CNBC reported that MiniMax raised about $619 million in a Hong Kong IPO and that its shares doubled on debut. SP013
CP014 TrendForce reported that DeepSeek raised more than CNY 50 billion at a valuation above CNY 330 billion in its first external funding round. SP016
CP015 Alibaba Cloud says Qwen3.6-Plus deploys through Model Studio and feeds Alibaba's own AI applications. SP018, SP019
CP016 ERNIE 5.1's official release claims top-tier reasoning and agentic capability while using about 6% of the pre-training cost of comparable models. SP021
CP017 BenchLM's August 2026 Chinese-model ranking leads with Kimi K3 and places Qwen3.8 Max as the best open-weight option. SP005
CP018 Digital Applied shows Alibaba Qwen at 13.9% provider share and 2.77T weekly tokens in its Q2 2026 landscape. SP004
CP019 Digital Applied shows MiniMax at 8.1% provider share and 1.62T weekly tokens in the same landscape. SP004
CP020 Digital Applied shows Z.ai at 5.6% provider share and 1.12T weekly tokens. SP004
CP021 Digital Applied shows DeepSeek at 5.6% provider share and 1.11T weekly tokens. SP004
CP022 The same Digital Applied report explicitly treats Baichuan as a second-tier niche player rather than a top-share provider. SP004
CP023 Baichuan's homepage says Baichuan4-Turbo is priced at about 80% of GPT-4o and that Baichuan4-Air lowers inference cost sharply. SP001
CP024 Apidog's 2026 Chinese price-war comparison indicates frontier Chinese API pricing has compressed across multiple major providers. SP006
CP025 DeepSeek's direct API docs show it is productized enough to publish public pricing and usage materials. SP015
CP026 Most Chinese model leaders expose compatible API rails or downloadable models, which makes technical multi-homing relatively easy for buyers. SP006, SP015, SP018, SP025
CP027 Z.ai's mix of chat surface, platform identity, and enterprise reputation makes it a particularly direct rival for sovereignty-sensitive enterprise deals. SP004, SP007, SP008
CP028 Moonshot combines benchmark and distribution momentum with a better funding position than Baichuan. SP010, SP011, SP005
CP029 Qwen has the broadest cross-channel packaging in the current field because it combines cloud access, enterprise deployment, app integration, and open-community support. SP018, SP019, SP005
CP030 ERNIE's official 5.1 release positions Baidu as a serious reasoning and agent rival with notable cost-efficiency claims. SP020, SP021
CP031 AICPB and Cybernews show that Baidu and other consumer AI leaders operate at monthly active user scale that Baichuan has not publicly matched. SP023, SP024
CP032 No reviewed public source disclosed comparable Baichuan consumer app scale, cloud distribution scale, or broad app MAU leadership. SP001, SP023, SP024
CP033 Baichuan still operates as a model-platform alternative because it maintains APIs plus open-weight medical models, even if its public-facing wedge is narrower than peers. SP001, SP002, SP025
CP034 Baichuan's healthcare specialization gives it a more explicit regulated-workflow identity than most generalist Chinese frontier labs. SP001, SP002, SP003
CP035 That healthcare wedge also implies a smaller reachable battlefield than the broader consumer or cloud markets led by Qwen, Kimi, ERNIE, or Doubao. SP001, SP004, SP023
CP036 KrASIA argues that China's AI-tigers cohort still has to prove durable business models against giant-backed rivals, which applies directly to Baichuan. SP003
CP037 API compatibility and open-weight access lower lock-in across the field, making Baichuan's moat more dependent on vertical workflow fit than on switching friction. SP006, SP015, SP025
CP038 Baichuan's moat depends on turning medical specialization into repeatable customer capture before horizontal leaders decide to attack the same vertical more aggressively. SP001, SP002, SP003, SP004
CP039 Public rankings, share tables, and scale disclosures do not show Baichuan as the current category leader on benchmark performance, broad usage, or distribution. SP004, SP005, SP023
CP040 The competitive conclusion is that Baichuan is more credible today as a track-worthy healthcare specialist than as a broad Chinese model winner. SP001, SP002, SP003, SP004, SP005
CI001 Baichuan's official price card says usage is billed by actual volume per 1,000 tokens and notes that one token is roughly equivalent to 1.5 Chinese characters. SI001
CI002 The official price card lists split input/output pricing for medical models, including Baichuan-M3-Plus at RMB0.005 input and RMB0.009 output per 1,000 tokens, Baichuan-M3 at RMB0.01 / RMB0.03, and Baichuan-M2 at RMB0.002 / RMB0.02. SI001
CI003 The same official price card lists general-model prices including Baichuan4-Turbo at RMB0.015 per 1,000 tokens, Baichuan4-Air at RMB0.00098, Baichuan4 at RMB0.1, Baichuan3-Turbo at RMB0.012, and Baichuan2-Turbo at RMB0.008. SI001
CI004 Baichuan publicly lists both search-enhancement and medical-search add-ons at RMB0.03 per call. SI001
CI005 Baichuan publicly lists Baichuan-Text-Embedding at RMB0.0005 per 1,000 tokens and knowledge-base file storage at RMB1.5 per GB per day. SI001
CI006 Baichuan's official price card says the Assistants API is currently free and that new users received promotional free-credit balances. SI001
CI007 Baichuan's API docs say users must complete real-name verification, recharge, and create an API key before using the platform. SI002
CI008 Baichuan's embedding API docs describe a synchronous POST endpoint with 1024-dimensional output vectors and public rate limits of 120 requests per minute for enterprise-certified accounts versus 60 for non-enterprise accounts. SI003
CI009 Baichuan's official homepage says the Hai Na Baichuan program permanently provides Baichuan-M3 Plus API access free to institutions that serve medical workers, subject to qualification rules and use restrictions. SI004
CI010 The official homepage positions Baichuan4-Turbo as enterprise-optimized and says its price is about 80% of GPT-4o, while Baichuan4-Air is marketed as an extreme cost-performance model with a call price of 0.98 li per 1,000 tokens. SI004
CI011 Baichuan's official homepage presents Bai Xiaoyi as an AI family-doctor app and bot spanning symptom triage, post-visit explanation, and household health management workflows. SI004
CI012 Caixin reported that co-founder Ru Liyun expected Baichuan's 2025 signed-order bookings to reach RMB1-2 billion. SI012
CI013 The same Caixin interview said Baichuan completed a RMB5 billion Series A in July 2024 and was valued at RMB20 billion at that time. SI012, SI014
CI014 A 36Kr report said Baichuan's To B business generated nearly RMB100 million of revenue in 2024. SI013
CI015 The same 36Kr report argued that Baichuan's current medical To B business was not making money because customization costs were high and hospital buyers were difficult to scale profitably. SI013
CI016 36Kr cited an industry source estimating that Baichuan still had more than RMB3 billion on hand, but this was not presented as an audited or official company disclosure. SI013
CI017 TMTPost reported that Baichuan completed a $300 million A1 strategic funding round in October 2023 and had accumulated $350 million including the angel round. SI010
CI018 TMTPost and Tracxn both support a very large July 25 2024 financing round of about $690-$691 million / RMB5 billion. SI005, SI008, SI011
CI019 Tracxn says Baichuan has raised about $1.04 billion over four funding rounds. SI008
CI020 CB Insights says Baichuan has raised $1.038 billion, records an April 1 2026 valuation of $2,899.11 million, and identifies 37 Interactive as a Series B investor. SI007
CI021 Yahoo Finance carrying SCMP reporting said Baichuan raised about RMB5 billion in a round backed by Alibaba, Tencent, Xiaomi, and state-linked funds at a valuation above RMB20 billion. SI005, SI014
CI022 36Kr said Baichuan had completed a RMB5 billion A round and planned to open a Series B at a RMB20 billion valuation, illustrating continuing dependence on external financing markets. SI014
CI023 37 Interactive's 2025 annual report lists Baichuan AI in its strategic investment portfolio, confirming strategic-backer alignment through a public filing. SI009
CI024 Apidog reported that Chinese labs cut LLM API prices six times in the first half of 2026 and that several cuts were made permanent, reinforcing the risk of continued price compression in Baichuan's domestic API market. SI015
CI025 CNBC reported that the US moved to halt Nvidia H20 AI chip shipments to Chinese firms outside China, which implies higher infrastructure and sourcing pressure for Chinese model developers such as Baichuan. SI016
CI026 ScienceNet and BSIA both document Baichuan's pediatric-model work with Beijing Children's Hospital, showing that healthcare deployments have moved beyond abstract marketing claims. SI017, SI018
CI027 Baichuan's GitHub page for Baichuan-M2-32B says the model can be deployed on a single RTX 4090 card, which lowers inference friction for some private deployments. SI019
CI028 Baichuan's public model pages present M3 as a 235B-parameter healthcare-enhanced model line and M1 as the first medical-enhanced model, indicating ongoing specialized R&D spending rather than a lightweight wrapper business. SI004, SI020, SI021
CI029 The Baichuan-M1 paper says the company built the model on a 20T medical-enhanced corpus, reinforcing that the healthcare stack depends on substantial domain-specific data work. SI021
CI030 Baichuan's platform homepage says the open platform includes rich industry workflows and can efficiently build enterprise-specific agents, supporting a sales motion beyond commodity raw inference. SI022
CI031 Baichuan's user agreement says the platform provides consultation only and does not constitute medical diagnosis, academic opinion, prescription, or electronic prescription. SI023
CI032 The 36Kr adverse report says Baichuan's 2024 business-services arm landed customers and partners including Bank of China, NE Digital, China Merchants Bank, Xinyada, and Tiankai Group. SI013
CI033 UsagePricing characterizes Baichuan's healthcare pivot as a deliberate narrowing in which the company uses free M3-Plus access to own the clinical ecosystem rather than maximize per-token monetization. SI006, SI004
CI034 CB Insights shows a latest round marker labeled as a rumored IPO in July 2026 while also leaving revenue fields undefined, which underscores how incomplete Baichuan's public financial record still is. SI007
CI035 Second Talent's 2026 China AI investment statistics support the view that external AI funding remains large in China, even though the market is crowded and competitive. SI025
CI036 Baichuan's docs direct larger users toward business-consultation and support channels, indicating that enterprise monetization is at least partly sales-led rather than purely self-serve. SI002, SI022
CI037 Baichuan's pricing pages show price discrimination across model classes and time windows, including a higher daytime price for Baichuan2-53B and different billing logic for general versus medical models. SI001
CI038 Baichuan's price card says calls to Baichuan-M3-Plus and Baichuan-M2-Plus automatically trigger medical search, meaning some healthcare usage carries an additional non-token cost layer. SI001
CI039 The 36Kr adverse report argues that hospital budgets, privacy barriers, isolated datasets, and high customization make medical To B a difficult commercialization segment even when product fit is strong. SI013
CI040 Taken together, Baichuan's free medical program and its user-agreement restrictions imply that some of the most strategically important healthcare workflows may prioritize ecosystem capture and controlled deployment over immediate high-margin monetization. SI004, SI023
CE001 Baichuan's official /pricing and /baixiaoyi pages frame Bai Xiaoyi as an AI family-doctor workflow covering symptom sorting before visits, post-visit explanation, and ongoing household health management. SE001, SE003
CE002 Baichuan's official pages show the Baichuan4 family remains an active general-purpose product line even after the healthcare pivot. SE001, SE017
CE003 The platform homepage describes Baichuan4-Air as a PRI-architecture MoE model aimed at low-cost, enterprise-oriented deployment. SE018
CE004 Baichuan's official pages list Baichuan-M3, M2, and M1 directly alongside the general line, making the medical family a first-class product surface rather than a side experiment. SE002, SE017
CE005 The platform homepage shows Baichuan positioning itself as an enterprise platform with intelligent agents, knowledge base, tool calling, and multi-industry solution templates. SE018
CE006 The same platform page still advertises finance, education, and role/NPC domain-enhanced models, indicating that Baichuan's stack remains broader than a pure medical-only product. SE018
CE007 Baichuan publicly describes a full domain-enhancement toolchain including data processing, incremental pretraining, model fine-tuning, PPO/DPO reinforcement learning, evaluation, compression, and deployment. SE018
CE008 Baichuan's platform page says the agent layer includes knowledge base, tool calling, web search, image generation, code interpreter, and file parsing. SE018
CE009 Baichuan's API docs and NPC page show a mixed delivery motion in which larger users are routed toward sales-assisted consultation rather than only self-serve access. SE004, SE009
CE010 Baichuan-M1-14B is described as an open-source model built from scratch specifically for medical scenarios rather than a lightly adapted generic chatbot. SE011, SE013
CE011 The M1 repo says Baichuan-M1 was trained on 20T tokens of mixed medical and general data, including medical corpora, multilingual general data, and fine-grained specialty coverage. SE011
CE012 The M1 repo describes architectural changes including short-convolution attention, sliding-window attention in some layers, and attention-head dimensional adjustments to improve long-context behavior and efficiency. SE011
CE013 The M1 materials describe a multi-stage training and alignment regime including stage-wise curriculum, ELO, TDPO, and PPO, indicating that Baichuan's medical line is trained through layered post-training rather than only prompt engineering. SE011
CE014 The Baichuan2 technical report says Baichuan2-7B and 13B were trained from scratch on 2.6 trillion multilingual tokens and include architecture/tokenizer modifications beyond a vanilla Transformer. SE014
CE015 The Baichuan2 report also says the family performs strongly on medicine and law benchmarks and was released openly to support research and commercial use. SE014, SE025
CE016 The Baichuan-M2 repo says M2 is built on a Qwen2.5-32B base and adds a Large Verifier System, patient simulator, medical mid-training, and multi-stage reinforcement learning. SE010
CE017 The same M2 repo says the model can run with 4-bit quantization on a single RTX 4090 and that an MTP version improves single-user token throughput by 58.5%. SE010
CE018 The M3 model card and arXiv paper describe M3 as a clinical-decision model that uses proactive inquiry, fact-aware reinforcement learning, and segmented workflow rewards to reduce hallucination and improve real-world medical reasoning. SE012, SE021
CE019 The M3 model card provides concrete serving guidance through transformers, vLLM, and SGLang, including an OpenAI-compatible deployment path and an 8×H20 speculative-decoding example. SE012
CE020 The M3 model card says M3 is for research and reference only, cannot replace diagnosis or treatment, and is intended for medical education, health consultation, and clinical decision support under professional guidance. SE012
CE021 The M4 paper defines Baichuan-M4 as a coordinated medical agent system built around Baichuan-Harness, a core reasoning model, and a clinical tool layer. SE005, SE006
CE022 The M4 paper says the system is designed for continuous-care tasks such as pre-visit triage, initial consultation, follow-up, chronic-disease management, document parsing, and multimodal image support. SE006
CE023 The M4 paper explicitly says outputs must not replace licensed physician judgment or be used directly for final diagnosis, treatment planning, prescription issuance, or emergency and critical care decisions. SE005, SE006
CE024 The M4 paper explicitly discloses limitations around rare diseases and possible bias or error in multimodal image analysis due to image quality and model generalization limits. SE006
CE025 The M4 paper says Baichuan-Harness supports subagent dispatch, dynamic role switching, long-term patient memory, and action-space constraints that validate tool usage and unauthorized actions. SE006
CE026 The M4 paper describes an evidence-retrieval layer built around a six-level pyramid of authoritative medical evidence plus PICO-based query decomposition and verifier-aligned retrieval quality. SE006
CE027 The same paper describes a multimodal clinical tool layer spanning OCR for medical documents, X-ray understanding, and an evidence-driven dermatology agent. SE006
CE028 AIBase reported that Baichuan paired M4 with a Bai Xiaoyi medical system positioned as moving from consultation into broader practice-management style workflows. SE007
CE029 ScienceNet, BSIA, and TMTPost all provide independent deployment proof that Baichuan's medical stack reached pediatric or hospital settings including Beijing Children's Hospital. SE015, SE016, SE024
CE030 Baichuan's user agreement says the platform provides consultation services only and does not constitute diagnosis, academic opinion, prescription, or electronic prescription. SE019
CE031 The platform homepage shows an ecosystem of recognizable enterprise logos and industry-solution tiles, suggesting Baichuan wants the product understood as deployable across many enterprise contexts rather than only as a lab demo. SE018
CE032 Baichuan's platform pages describe a role/NPC model with character knowledge base, multi-turn memory, customizable settings, and factual adherence to uploaded background knowledge. SE018
CE033 Baichuan claims its full-chain domain enhancement toolchain can push multi-scenario enterprise usability to 96%, underscoring a repeatable adaptation narrative even if the metric itself is company-authored. SE018
CE034 Across M1, M2, M3, and M4, the deepest product differentiation appears to come from healthcare-specific data, verifier systems, retrieval, memory, and workflow logic rather than from a wholly independent base-model ecosystem. SE010, SE011, SE012, SE006
CE035 The public stack also shows clear ecosystem dependence: M2 cites a Qwen2.5 base, the M3 card names a Qwen3 base plus vLLM/SGLang/verl, and the deployment path relies heavily on popular open-source serving frameworks. SE010, SE012
CE036 Taken together, the user agreement, M3 card, and M4 paper show a consistent assistive-product posture that keeps Baichuan inside controlled decision-support boundaries rather than autonomous clinical action. SE019, SE012, SE006
CE037 Baichuan's official surfaces still display finance, education, and role models alongside the healthcare-first brand, suggesting the company preserves a broader product option set even as public messaging narrows. SE001, SE002, SE018
CE038 The combination of API docs, platform toolchain, and open-weight artifacts shows Baichuan optimizing for multiple adoption paths: hosted API, sales-assisted integration, and self-hosted or semi-private deployment. SE009, SE012, SE018
CE039 Because so much of Baichuan's product value sits in retrieval, adaptation, hospital integration, and runtime controls, implementation quality may matter as much as raw benchmark quality for buyers. SE006, SE018, SE024
CE040 Public policy and product documents show awareness of privacy, safety, and clinical scope, but they do not yet provide strong independent evidence on uptime, incident reporting, external audits, or deployment-level post-market monitoring. SE019, SE020, SE006
CU001 Baichuan's official Bai Xiaoyi pages and archived Ying surface position the product as a consumer-facing AI family-doctor workflow for pre-visit triage, post-visit explanation, and household health management. SU001, SU013
CU002 Baichuan's cooperation pages explicitly invite developers, enterprises, and organizations into its commercial funnel, showing that the company does not rely on a single customer type. SU004, SU020
CU003 The Baichuan platform homepage shows a broad enterprise ecosystem of recognizable logos spanning Tencent, Xiaomi, Intel, Didi, iQiyi, 58.com, and other brands. SU012
CU004 36Kr reported that Baichuan's 2024 commercial-services effort brought in customers and partners including Bank of China, NE Digital, China Merchants Bank, Xinyada, and Tiankai Group. SU010
CU005 BSIA, ScienceNet, and TMTPost all support Baichuan's deployment proof at Beijing Children's Hospital and the Futang-Baichuan pediatric model. SU008, SU009, SU014
CU006 TMTPost also says Baichuan had deployed medical models at the Cancer Hospital of the Chinese Academy of Medical Sciences, broadening the named hospital set beyond pediatrics. SU014
CU007 Caixin reported that co-founder Ru Liyun expected 2025 signed-order bookings to reach RMB1-2 billion. SU011
CU008 36Kr reported that Baichuan's To B business generated about RMB100 million of revenue in 2024. SU010
CU009 ToolChase describes Baichuan as best suited to developers, researchers, and Chinese healthcare teams rather than mainstream consumers seeking the most polished chatbot experience. SU007
CU010 Open-weight distribution through GitHub, Hugging Face, and Baichuan's own model pages makes developers and self-hosters a real customer or user segment even when they are not all direct API payers. SU015, SU016, SU024
CU011 TMTPost said Baichuan-M2 was designed for private deployment in clinical settings and had already deployed AI pediatricians in Beijing Children's Hospital. SU014
CU012 BSIA described the Beijing Children's Hospital collaboration as a “dual-doctor” model meant to upgrade grassroots medical service quality, giving Baichuan a concrete user-value narrative rather than a generic AI label. SU008
CU013 ScienceNet described Futang-Baichuan as the first domestic pediatric foundation model, reinforcing the strategic importance of that customer reference. SU009
CU014 Baichuan's platform surfaces show customer-facing solution lanes in healthcare, banking, insurance, education, retail, manufacturing, and intelligent terminals. SU003, SU012
CU015 Baichuan's platform ecosystem implies a wide enterprise prospect set, but logos alone do not confirm production depth, current spend, or retention. SU012
CU016 Baichuan's customer acquisition path is at least partly sales-assisted, with consultation and onboarding surfaces visible on the business and NPC pages. SU004, SU025
CU017 Baichuan's M3 technical blog says the prior M2 release attracted hundreds of thousands of model downloads, which is meaningful developer-demand proof even if it is not equivalent to paying account count. SU002, SU007
CU018 ToolChase says Baichuan offers free consumer chat and free open weights alongside paid API and enterprise contracts, suggesting a low-friction top-of-funnel feeding into higher-value accounts. SU007
CU019 The official Bai Xiaoyi, pricing, and root pages all reinforce that Baichuan maintains a direct end-user product surface rather than only a developer-facing API. SU013, SU017, SU023
CU020 AIBase and Houdao both describe Bai Xiaoyi as moving toward broader medical-system or hospital-linked use rather than a pure consumer chatbot. SU005, SU006
CU021 36Kr argues that medical To B is difficult because projects are customized, hospitals face tight budgets, and million-yuan annual model spending can still be hard to justify. SU010
CU022 The same 36Kr report said Baichuan was discussing combining medical capabilities with BaiXiaoying/Bai Xiaoyi to develop a C-end medical product, indicating an active To B / To C mix question in customer strategy. SU010
CU023 Baichuan's platform and API pages imply a land-and-expand path from evaluation and open experimentation into agent workflows, enterprise integration, and possibly private deployment. SU003, SU012, SU020
CU024 The public record shows adoption surfaces more clearly than active account or repeat-usage denominators, especially for Bai Xiaoyi and enterprise customers. SU007, SU013, SU017
CU025 Baichuan does not publicly disclose NRR, GRR, churn, contract length, or renewal rate in the reviewed source set. SU010, SU011, SU018
CU026 Because the strongest customer proof clusters in hospitals and a few named enterprise references, concentration risk cannot be ruled out from public evidence alone. SU005, SU010, SU014
CU027 Baichuan's own disclaimers keep the medical stack in an assistive, consultation-oriented posture, which likely slows conversion into fully autonomous or high-liability clinical workflows. SU018, SU019
CU028 Platform logos, industry lanes, and solution language show enterprise breadth, but public evidence does not separate pilots from large recurring production accounts. SU012
CU029 TMTPost's description of lightweight, lower-cost hospital deployment suggests Baichuan is consciously optimizing product delivery for procurement-sensitive medical buyers. SU014
CU030 The overall customer picture is therefore stronger on access and named deployment proof than on durability, satisfaction, or concentration disclosure. SU005, SU010, SU011, SU012
CU031 The platform homepage positions Baichuan's general and domain models as suitable for enterprise ToB scenarios in multiple industries, not just research users. SU012
CU032 Baichuan's model pages repeatedly route users into open-source communities, which supports the view that developer ecosystems are part of the company's customer acquisition surface. SU024
CU033 The combination of GitHub, Hugging Face, and price-card/API surfaces means developers can move from evaluation into paid API or self-hosted deployment without abandoning the Baichuan ecosystem immediately. SU015, SU016, SU021
CU034 ToolChase's independent review says the open-source commitment made Baichuan popular with international researchers, expanding customer reach beyond domestic China buyers. SU007
CU035 China AI Atlas describes Baichuan as pivoted hard to healthcare, with Futang Baichuan and Beijing Children's Hospital at the center of its current public customer narrative. SU022
CU036 The official pricing and app-entry pages reinforce that Bai Xiaoyi can serve as both an engagement product and an API-acquisition surface, not only a standalone chat experience. SU021, SU023
CU037 The official customer evidence is freshest where it ties a named institution to a named workflow, such as Futang-Baichuan in pediatrics, rather than when it shows only generic logos or broad solution categories. SU008, SU009, SU012
CU038 Baichuan's direct-consumer product route may help feedback collection and awareness, but public monetization and retention disclosure are far weaker there than for the named hospital references. SU001, SU013, SU007
CU039 The combination of customer-proof articles and company platform collateral is enough to show real adoption, but not enough to benchmark Baichuan against mature enterprise-software disclosure standards. SU010, SU012, SU014
CU040 As of the run date, Baichuan looks like a company with genuine hospital, enterprise, developer, and consumer customer surfaces, but still with sparse public evidence on renewal and concentration. SU005, SU007, SU010, SU011, SU014
CR001 The July 2023 CAC-led Interim Measures impose formal obligations on generative-AI providers serving the Chinese public, including lawful data use, content controls, and privacy duties. SR001, SR002
CR002 The same official measures contemplate security assessment and algorithm-filing obligations for some services with public-opinion or social-mobilization attributes. SR001
CR003 Baichuan therefore faces a live compliance burden that is ongoing and operational, not a one-time licensing box. SR001, SR002, SR016
CR004 Baichuan’s user agreement says the platform provides consultation rather than medical diagnosis, confirming deliberate scope limitation in health-related use cases. SR014
CR005 Baichuan’s privacy policy says the company has a dedicated incident-response team and security plans for different incidents, indicating privacy and security are explicit governance responsibilities. SR015
CR006 The public record reviewed here does not disclose Baichuan’s detailed CAC filing number, security-assessment outcome, or product-by-product regulatory status. SR001, SR014, SR015
CR007 The Interim Measures also require providers to sign service agreements with users and protect user input information and usage records. SR001
CR008 For a healthcare-adjacent model company, Chinese AI regulation and Baichuan’s own legal language together raise the cost of mistakes in privacy, labeling, and scope control. SR001, SR014, SR015
CR009 No current public enforcement action was found against Baichuan in the reviewed sources, but absence of public enforcement is not equivalent to low compliance risk. SR001, SR006, SR014
CR010 Chinese frontier-model companies remain exposed to GPU policy because advanced Nvidia access can change with U.S. export licensing rules. SR003, SR004, SR005
CR011 CNBC reported the U.S. moved to close a loophole that may have allowed overseas subsidiaries of Chinese entities to obtain advanced chips without a license. SR003
CR012 Benzinga’s Reuters-based reporting said Nvidia was told the H20 required an export license for China sales and that major Chinese customers had built large order books around the chip. SR004, SR032
CR013 Yahoo’s Reuters report later said the Commerce Department had started issuing licenses again, showing how quickly the operating environment can reverse. SR005, SR031
CR014 Business Standard reported Chinese regulators then raised security concerns around H20 shipments and related production, showing that China-side policy can also affect imported compute. SR006
CR015 Built In’s timeline frames April-to-July 2025 as a clear episode of U.S. policy whiplash rather than a stable ruleset. SR007
CR016 LevelFields argues that China’s AI sector remains deeply dependent on American GPUs, especially Nvidia, even as domestic substitution improves. SR008
CR017 Baichuan’s smaller-footprint private-deployment and efficiency messaging may reduce some inference friction for customers, but it does not remove the company from the larger compute-policy regime. SR011, SR017, SR021
CR018 The Chinese LLM price war means Baichuan’s principal economic risk is margin compression rather than simple lack of demand. SR024, SR025, SR026
CR019 Apidog reported that major Chinese labs cut API prices repeatedly in 2026 and that several cuts became permanent, supporting a harsh pricing backdrop. SR024
CR020 Digital Applied and KrASIA both support the view that a small set of larger or better-distributed Chinese AI providers now dominate visibility, token share, or strategic mindshare. SR025, SR026
CR021 Official competitor surfaces such as Alibaba Cloud’s Model Studio and DeepSeek’s public pricing docs show that rivals already package models as accessible enterprise or developer products, not research curiosities. SR027, SR028
CR022 Because many Chinese model vendors expose OpenAI-like APIs or downloadable weights, switching and multi-homing risk remain structurally high for Baichuan. SR021, SR028, SR030
CR023 Strategic investors and enterprise logos help Baichuan survive, but they do not automatically solve distribution the way Alibaba, Baidu, or ByteDance ecosystems do for their own model families. SR016, SR026, SR027
CR024 Baichuan’s open-source and developer surfaces are a two-sided risk: they increase reach and evaluation, but they also lower technical lock-in. SR020, SR021, SR030
CR025 TrendForce’s reporting on DeepSeek’s enormous funding round underscores how heavily capitalized some rivals now are relative to Baichuan. SR029
CR026 The residual competitive conclusion is that Baichuan still looks relevant, but its room for pricing mistakes is narrow. SR024, SR025, SR026, SR029
CR027 Baichuan’s healthcare wedge is strategically differentiated, but the best adverse reporting says medical To B deployments can be customized, slow, and budget-constrained. SR009, SR011, SR022
CR028 36Kr specifically argues that medical To B work is highly customized, hard to scale through fragmented data, and expensive for hospital buyers. SR009
CR029 Named hospital proof at Beijing Children’s Hospital strengthens credibility, but it does not by itself prove renewal quality, breadth of paid deployment, or attractive margin. SR011, SR012, SR013
CR030 Caixin’s reported 2025 bookings target implies meaningful commercialization ambition, yet it does not reveal contract quality, revenue recognition mix, or gross margin durability. SR010
CR031 Baichuan’s public legal and product materials reinforce assistive positioning, which reduces some medical-liability exposure but may also narrow which workflows can monetize quickly. SR014, SR019, SR022
CR032 The public evidence base remains thin on customer concentration, retention, NRR, and renewal by hospital or enterprise segment. SR009, SR010, SR016
CR033 The public narrative still centers heavily on founder Wang Xiaochuan and a relatively thin visible management bench, so leadership-concentration risk cannot be ruled out. SR019, SR023, SR010
CR034 Talent competition is likely intense because Baichuan competes for the same model, systems, and commercialization talent pools as much larger Chinese AI and big-tech players. SR025, SR026, SR029
CR035 Baichuan’s risk profile is best understood as stacked risk: regulation, compute volatility, pricing pressure, and healthcare economics can reinforce one another. SR001, SR010, SR024, SR025
CR036 A fresh GPU-policy shock during an active domestic price war would likely pressure both margin and growth simultaneously. SR003, SR004, SR024
CR037 The healthcare thesis breaks if marquee hospital proof fails to convert into recurring, expandable, software-like revenue rather than bespoke projects. SR009, SR011, SR029
CR038 The compliance thesis breaks if a live Baichuan product is later shown to lack required filings, safety review, or data-governance controls for its use case. SR001, SR014, SR015
CR039 The valuation-thesis risk is highest when economics remain opaque even as capital intensity and competition stay high. SR010, SR024, SR029
CR040 Bottom line: Baichuan’s risks look manageable only if compliance stays clean, compute access remains workable, and healthcare deployments mature into repeatable economics before commoditization accelerates further. SR001, SR005, SR009, SR024, SR025
CV001 The last well-supported Baichuan private valuation mark is roughly $2.8 billion from the reported April 2026 Series B round. SV001, SV002, SV003
CV002 Public trackers and reporting suggest Baichuan’s cumulative funding is in the high-hundreds-of-millions range, with strategic investors including 37 Interactive, Xiaomi, and Alibaba Cloud. SV001, SV002, SV003, SV004
CV003 The April 2026 mark should be treated as a negotiated private price, not a clean intrinsic value signal. SV001, SV003, SV004
CV004 Caixin reported that Baichuan expected 2025 signed-order bookings in the RMB1B-RMB2B range. SV005
CV005 36Kr reported that Baichuan’s 2024 To B business generated roughly RMB100 million of revenue. SV006
CV006 Because public evidence shows only partial revenue quality, the current $2.8B mark cannot be confidently defended from public information alone. SV001, SV005, SV006
CV007 Baichuan has enough public proof of product, customers, and funding to justify serious investor attention rather than a speculative-dismissal stance. SV001, SV021, SV022, SV024
CV008 The company’s public product and customer evidence imply strategic optionality across healthcare AI, enterprise workflows, and developer distribution. SV021, SV022, SV027, SV028, SV029
CV009 However, public evidence on retention, concentration, project burden, and cap-table preferences remains sparse. SV005, SV006, SV022
CV010 That opacity makes valuation discipline more important than category enthusiasm. SV001, SV005, SV006, SV007
CV011 Digital Applied’s Q2 2026 landscape places Baichuan closer to a credible specialist or second-tier niche player than to the very top share leaders. SV011
CV012 As of August 2026 Alibaba’s public P/S ratio is about 2.15 on roughly $145.39B of TTM revenue. SV013, SV014
CV013 As of August 2026 Baidu’s public P/S ratio is about 1.71 on roughly $18.27B of TTM revenue. SV016, SV017
CV014 These public Chinese platform multiples provide a discipline floor, but they are not direct startup comps because both companies are larger, slower, and more diversified. SV012, SV013, SV015, SV016
CV015 As of August 2026 Nvidia’s public P/S ratio is about 20.7 on roughly $253.49B of TTM revenue and a $5.252T market cap. SV018, SV019, SV020
CV016 Nvidia is useful only as a reminder of how richly the market rewards exceptional AI scale and platform leverage; it is not a direct Baichuan multiple anchor. SV018, SV019, SV020
CV017 Reported 2026 private marks for Moonshot, DeepSeek, and MiniMax show that the market assigns much larger premia to Chinese AI companies with stronger visibility, scale, or public-market access. SV008, SV009, SV010
CV018 Baichuan’s $2.8B mark is therefore modest relative to the breakout-winner cohort, but modest relative pricing alone does not make it attractive. SV001, SV008, SV009, SV010
CV019 Baichuan deserves some premium to slow-growth public China tech because it still carries higher growth optionality and private-AI upside. SV011, SV012, SV015, SV021
CV020 Baichuan also deserves a discount to the breakout private leaders because its public proof on consumer scale, benchmark leadership, and financial visibility is materially weaker. SV008, SV009, SV010, SV011
CV021 The Chinese API price war is a direct valuation headwind because it compresses the multiple investors should pay for uncertain future revenue. SV007, SV011, SV023
CV022 A broad enterprise and developer surface exists, but those routes also lower switching costs and reduce confidence in long-term pricing power. SV022, SV027, SV028, SV029, SV030
CV023 The bull case requires Baichuan to show that healthcare credibility translates into repeatable, scalable, and at least partly premium-priced revenue. SV024, SV025, SV026, SV005
CV024 The most defensible base case from public evidence is that Baichuan is real and promising, but still not disclosed well enough to justify paying above the last round with conviction. SV001, SV005, SV006, SV011
CV025 The bear case is driven by commercialization disappointment rather than technological collapse: medical To B remains bespoke, budget-constrained, and hard to scale. SV006, SV007, SV024
CV026 A scenario-based range is more appropriate than a single-point DCF because audited revenue, margin, and cash-flow inputs are not public. SV001, SV005, SV006
CV027 The best-supported present-value range is roughly $2.0B-$2.6B, which places the April 2026 mark at the fair-to-full end of reasonable public-evidence underwriting. SV001, SV005, SV006, SV011
CV028 A more attractive new-money entry would likely sit below roughly $2.3B unless stronger revenue-quality evidence emerges. SV001, SV006, SV011
CV029 A plausible bull-case range of roughly $3.4B-$4.5B exists if bookings convert, healthcare deployments scale, and the risk discount narrows. SV005, SV024, SV025, SV026
CV030 A plausible bear-case range of roughly $1.0B-$1.6B exists if medical To B underdelivers and price-war dynamics continue to weaken revenue quality. SV006, SV007, SV011
CV031 The supportable risk rating at the current mark is high because pricing, compute, compliance, and customer-quality risks all matter simultaneously. SV006, SV007, SV011, SV024
CV032 The supportable recommendation at the current mark is research-more / track rather than buy. SV001, SV005, SV006, SV011, SV024
CV033 New money at $2.8B requires roughly $4.2B-$5.6B for a 1.5-2x outcome and roughly $5.6B-$8.4B for a 2-3x outcome before dilution. SV001
CV034 The highest-priority diligence asks are segment revenue mix, retention by customer type, margin by product line, concentration, and cap-table preference terms. SV005, SV006, SV011
CV035 Company-specific CAC filing or assessment evidence for live products is also a valuation-critical diligence item because compliance failures would directly damage both growth and exit readiness. SV021, SV022, SV024
CV036 The thesis breaks if healthcare proof fails to convert into repeatable software-like economics despite continued product investment. SV006, SV024, SV025
CV037 The thesis also breaks if a new compute-policy or compliance shock lands before Baichuan demonstrates stronger revenue quality. SV007, SV021, SV022
CV038 Persistent opacity on concentration, renewal, and preferences is itself a reason not to upgrade the recommendation. SV001, SV005, SV006
CV039 Exit readiness is not yet something investors should assume from category momentum alone; it depends on cleaner economics and governance-grade disclosure. SV001, SV004, SV011
CV040 Bottom line: Baichuan is interesting and real, but the current public evidence supports disciplined tracking more than aggressive underwriting at the last visible price. SV001, SV005, SV006, SV011, SV024
来源
编号出版方标题引文
SO001 Baichuan AI 百川大模型-百川智能 百川智能成立于2023年3月24日,由前搜狗公司CEO王小川创立。
SO002 Baichuan AI 百川大模型开放平台首页 百川大模型开放平台内置丰富行业工作流,高效构建企业专属智能体。
SO003 Baichuan AI 百川开放平台 API 文档 / pricing 使用的模型ID列表包括 Baichuan4-Turbo、Baichuan4-Air、Baichuan4、Baichuan3-Turbo、Baichuan2-Turbo。
SO004 Baichuan AI 开源模型商业授权申请 欢迎各位开发者、企业和机构与我们一起开创人工智能新时代。
SO005 Baichuan AI 用户协议 本平台仅提供咨询服务...不代表医疗诊断等专业意见,也不属于学术观点,更不构成处方或电子处方。
SO006 Baichuan AI 隐私政策 我们建立了专门的应急响应团队,针对不同安全事件启动安全预案。
SO007 TechCrunch China's search engine pioneer unveils open source large language model to rival OpenAI Wang stepped down from Sogou in late 2021... launched Baichuan in April and quickly pocketed $50 million in financing.
SO008 TMTPost Alibaba and Tencent Join in AI Startup Baichuan's Funding as Baidu Releases GPT-4 Rival Model Alibaba and Tencent joined in Baichuan's funding as Baidu released GPT-4 rival model.
SO009 TMTPost Chinese AI Startup Baichuan AI Raises $5 Billion in Funding Baichuan AI raised 5 billion yuan ($690 million) in Series A round from major investors including Alibaba, Xiaomi and Tencent.
SO010 TMTPost China's Baichuan Intelligence Launches Medical AI Model That Outperforms OpenAI in Key Benchmark Founded in 2023, Baichuan has raised three funding rounds... and is preparing a Series B at a $2.75 billion valuation.
SO011 Caixin GPT革命|专访百川智能茹立云:创业需避开互联网厂商 商业化半年签约订单数亿元 百川智能是国内六家头部大模型创业公司之一,由前搜狗创始人王小川、前搜狗COO茹立云等人于2023年4月成立。
SO012 CB Insights Baichuan AI Stock Price, Funding, Valuation, Revenue & Financial Statements Baichuan AI's valuation in April 2026 was $2,899.11M.
SO013 Tracxn Baichuan funding and investors Baichuan has raised a total of $1.04B over 4 funding rounds.
SO014 InvestGame 37 Interactive Entertainment FY2025 Annual Report 37 Interactive Entertainment Network Technology Group Co., Ltd. is a listed company (stock code 002555).
SO015 GitHub / baichuan-inc Baichuan-M2-32B repository Baichuan-M2-32B 是百川智能推出的医疗增强推理模型,这是百川开源发布的第二个医疗增强模型。
SO016 GitHub / baichuan-inc Baichuan2 repository A series of large language models developed by Baichuan Intelligent Technology.
SO017 GitHub / baichuan-inc Baichuan-13B repository A 13B large language model developed by Baichuan Intelligent Technology.
SO018 arXiv Baichuan-M1: Pushing the Medical Capability of Large Language Models Baichuan-M1: Pushing the Medical Capability of Large Language Models.
SO019 Baichuan AI Baichuan2 technical report PDF Baichuan 2 technical report documents bilingual and multilingual evaluation results.
SO020 Hugging Face / baichuan-inc Baichuan-M3-235B model card baichuan-inc/Baichuan-M3-235B model card keeps the healthcare-enhanced line visible to developers.
SO021 KrASIA After tech giants, a new cohort of “AI tigers” finds footing in China The new cohort of AI tigers still has to find durable footing against Chinese tech giants and crowded market economics.
SO022 ScienceNet 国内首个儿科大模型“福棠·百川”发布 国内首个儿科大模型“福棠·百川”发布。
SO023 Beijing Software and Information Service Industry Association 百川智能携手北京儿童医院发布全球儿科大模型,双医模式助力基层医疗升级 百川智能携手北京儿童医院发布全球儿科大模型,双医模式助力基层医疗升级。
SO024 GitHub / baichuan-inc Baichuan-M1-14B repository Baichuan-M1-14B is described as the first medical-enhanced model in the Baichuan line.
SO025 GitHub / baichuan-inc Baichuan-7B repository A large-scale 7B pretraining language model developed by BaiChuan-Inc.
SM001 Baichuan AI 百川大模型-百川智能
SM002 Baichuan AI 百川大模型开放平台首页
SM003 Baichuan AI 百川开放平台 API 文档 / pricing
SM004 TMTPost China's Baichuan Intelligence Launches Medical AI Model That Outperforms OpenAI in Key Benchmark
SM005 TMTPost Chinese AI Startup Baichuan AI Raises $5 Billion in Funding
SM006 Caixin GPT革命|专访百川智能茹立云:创业需避开互联网厂商 商业化半年签约订单数亿元
SM007 IMARC Group China Generative AI (AIGC) Market Size, Share, Trends and Forecast 2026-2034
SM008 Xinhua / State Council Information Office China unveils guidelines to regulate, boost innovative development of AI agents
SM009 CNBC U.S. takes step to halt Nvidia AI chip shipments to Chinese firms outside China
SM010 IDC IDC FutureScape 2026 China AI excerpt
SM011 China Daily Volcano Engine leads China public-cloud large-model market
SM012 TrendForce Chinese AI models reportedly hit ~15% global share in Nov. 2025
SM013 Gartner Gartner Forecasts Worldwide AI Spending to Grow 47% in 2026
SM014 Forrester Asia Pacific Tech Spending Expected To Grow 9.3% In 2026
SM015 State Council / CAC 生成式人工智能服务管理暂行办法
SM016 Apidog The 2026 Chinese LLM Price War: Top 5 Frontier API Costs Compared
SM017 KrASIA After tech giants, a new cohort of “AI tigers” finds footing in China
SM018 ScienceNet 国内首个儿科大模型“福棠·百川”发布
SM019 Beijing Software and Information Service Industry Association 百川智能携手北京儿童医院发布全球儿科大模型,双医模式助力基层医疗升级
SM020 TMTPost Alibaba and Tencent Join in AI Startup Baichuan's Funding as Baidu Releases GPT-4 Rival Model
SM021 CB Insights Baichuan AI Stock Price, Funding, Valuation, Revenue & Financial Statements
SM022 GitHub / baichuan-inc Baichuan-M2-32B repository
SM023 Hugging Face / baichuan-inc Baichuan-M3-235B model card
SM024 Tracxn Baichuan funding and investors
SM025 Baichuan AI 用户协议
SP001 Baichuan AI 百川大模型-百川智能
SP002 TMTPost China's Baichuan Intelligence Launches Medical AI Model That Outperforms OpenAI in Key Benchmark
SP003 KrASIA After tech giants, a new cohort of “AI tigers” finds footing in China
SP004 Digital Applied Chinese AI Models Q2 2026: 10-Provider Landscape Report
SP005 BenchLM Best Chinese AI Models (August 2026): Kimi K3 Leads
SP006 Apidog The 2026 Chinese LLM Price War: Top 5 Frontier API Costs Compared
SP007 Z.ai Z.ai - Advanced AI Chatbot & Agent powered by GLM-5.2
SP008 BigModel / 智谱 智谱丨BigModel 平台
SP009 Moonshot AI Moonshot AI homepage
SP010 Kimi Kimi AI with K3
SP011 TechCrunch China's Moonshot AI raises $2B at $20B valuation as demand for open source AI skyrockets
SP012 MiniMax MiniMax homepage
SP013 CNBC MiniMax doubles in Hong Kong debut, marking yet another Chinese AI listing
SP014 DeepSeek DeepSeek | Into the Unknown
SP015 DeepSeek DeepSeek API Docs pricing
SP016 TrendForce DeepSeek completes first funding round at valuation above CNY 330 billion
SP017 Qwen Qwen homepage
SP018 Alibaba Cloud Alibaba Cloud Model Studio
SP019 Alibaba Cloud Alibaba Unveils Qwen3.6-Plus to Accelerate Agentic AI Deployment
SP020 Baidu ERNIE ERNIE homepage
SP021 Baidu ERNIE ERNIE 5.1 Officially Released
SP022 ByteDance / Doubao Doubao homepage
SP023 AICPB China AI Rankings by App MAU — Issue 23 (Jun 2026 Edition)
SP024 Cybernews Chinese AI assistant reaches 200 million monthly active users
SP025 Baichuan AI 百川开放平台 API 文档 / pricing
SI001 Baichuan AI 百川大模型价格说明 按照实际使用的数据量(千tokens)收费。一般情况下百川大模型1个token约等于1.5个中文汉字。
SI002 Baichuan AI 百川开放平台 API 使用指南 由百川提供,API 开放平台完成实名认证、充值、创建 APIkey 等流程。
SI003 Baichuan AI Text Embedding API 文档 当前企业认证账号限制 120 记录/分钟,非企业认证账号为 60 记录/分钟。
SI004 Baichuan AI 百川大模型-百川智能 Free 永久免费调用 Baichuan-M3 Plus API。
SI005 Yahoo Finance / South China Morning Post Chinese AI start-up Baichuan raises US$700 million from Alibaba, Tencent, Xiaomi Baichuan AI ... raised about 5 billion yuan (US$687.6 million) in a new funding round that valued the start-up at more than 20 billion yuan.
SI006 UsagePricing Baichuan AI Pricing Its 海纳百川 program now gives the Baichuan-M3-Plus medical API away free, permanently, to institutions serving healthcare workers.
SI007 CB Insights Baichuan AI Stock Price, Funding, Valuation, Revenue & Financial Statements Baichuan AI has raised $1.038B over 7 rounds. Baichuan AI's valuation in April 2026 was $2,899.11M.
SI008 Tracxn Baichuan funding and investors Baichuan has raised a total of $1.04B over 4 funding rounds.
SI009 37 Interactive Entertainment 2025 Annual Report (Summary) The portfolio includes Zhipu AI, Moonshot AI, Baichuan AI...
SI010 TMTPost Alibaba and Tencent Join in AI Startup Baichuan's Funding as Baidu Releases GPT-4 Rival Model Baichuan Intelligent Technology said it has completed A1 strategic funding round of $300 million.
SI011 TMTPost Chinese AI Startup Baichuan AI Raises $5 Billion in Funding Baichuan AI ... secured 50 billion yuan ($5 billion) in Series A round from major investors including Alibaba, Xiaomi and Tencent.
SI012 Caixin GPT革命|专访百川智能茹立云:创业需避开互联网厂商 商业化半年签约订单数亿元 茹立云预计2025年订单签约额将在10-20亿元的水平。
SI013 36Kr / 职场Bonus Bonus独家 | 百川智能急刹车,调整医疗ToB,基础研发停摆 To B业务是百川造血等心脏,在2024年为百川带来了近1亿的收入。
SI014 36Kr / 创投日报 百川智能晋升200亿大模型独角兽 公司确已于近期完成了A轮融资,总融资金额达50亿元人民币,并且将以200亿估值开启B轮融资。
SI015 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, and three of those cuts were declared permanent.
SI016 CNBC U.S. takes step to halt Nvidia AI chip shipments to Chinese firms outside China The U.S. has taken a step to halt shipments of Nvidia H20 AI chips to Chinese firms outside China.
SI017 ScienceNet 国内首个儿科大模型“福棠·百川”发布 国内首个儿科大模型“福棠·百川”发布。
SI018 Beijing Software and Information Service Industry Association 百川智能携手北京儿童医院发布全球儿科大模型,双医模式助力基层医疗升级 百川智能携手北京儿童医院发布全球儿科大模型。
SI019 GitHub / baichuan-inc Baichuan-M2-32B repository 4090单卡可部署。
SI020 Hugging Face / baichuan-inc Baichuan-M3-235B model card Baichuan-M3-235B keeps the healthcare-enhanced flagship line visible to developers.
SI021 arXiv Baichuan-M1: Pushing the Medical Capability of Large Language Models Baichuan-M1: Pushing the Medical Capability of Large Language Models.
SI022 Baichuan AI 百川大模型开放平台首页 百川大模型开放平台内置丰富行业工作流,高效构建企业专属智能体。
SI023 Baichuan AI 用户协议 本平台仅提供咨询服务...不代表医疗诊断等专业意见,更不构成处方或电子处方。
SI024 TMTPost China's Baichuan Intelligence Launches Medical AI Model That Outperforms OpenAI in Key Benchmark Baichuan ... is preparing a Series B at a $2.75 billion valuation.
SI025 Second Talent Top 50+ Chinese AI Investment Statistics [2026] China's AI market received massive investment inflows in 2026, underscoring a deep but crowded financing environment.
SE001 Baichuan AI 百川大模型-百川智能 (/pricing) 百小医 你的 AI家庭医生。
SE002 Baichuan AI 百川大模型-百川智能 (/models) Baichuan-M3、Baichuan-M2、Baichuan-M1 等模型在页面中被直接列出。
SE003 Baichuan AI 百川大模型-百川智能 (/baixiaoyi) 看病前帮你梳理症状,准备就医;看病后帮你分析病情,解读医嘱。
SE004 Baichuan AI NPC product page 请填写合作咨询申请体验测试,我们会安排商务尽快跟进。
SE005 arXiv Baichuan-M4: A Clinical-Grade Medical Agent System for Continuous Care Baichuan-M4 is Baichuan Intelligence's clinical-grade medical large model, designed for continuous care rather than single-turn medical question answering.
SE006 arXiv Baichuan-M4 HTML full text It is built as a coordinated medical agent system around three pillars: Baichuan-Harness, a core reasoning model, and a clinical tool layer.
SE007 AIBase BaiChuan Intelligence Launches Baichuan-M4 Large Model and BaiXiaoYi AI Medical System BaiXiaoYi AI Medical System transitions from consultation to general practice management.
SE008 Baichuan AI Text Embedding API 文档 模型名称,目前仅支持“Baichuan-Text-Embedding”,输入最长是 512 个 token,输出 1024 维。
SE009 Baichuan AI API 使用指南 申请加入海纳百川计划·免费使用M3Plus API。
SE010 GitHub / baichuan-inc Baichuan-M2-32B repository Baichuan-M2采用了三个核心技术创新:大型验证器系统、医疗领域适应性增强的中期训练、多阶段强化学习策略。
SE011 GitHub / baichuan-inc Baichuan-M1-14B repository Baichuan-14B-M1 是业界首款从零开始专为医疗场景优化的开源大语言模型。
SE012 Hugging Face / baichuan-inc Baichuan-M3-235B model card Baichuan-M3 is trained to explicitly model the clinical decision-making process, aiming to improve usability and reliability in real-world medical practice.
SE013 arXiv Baichuan-M1: Pushing the Medical Capability of Large Language Models Baichuan-M1: Pushing the Medical Capability of Large Language Models.
SE014 Baichuan AI Baichuan2 technical report PDF Baichuan 2 is a series of large-scale multilingual language models trained from scratch on 2.6 trillion tokens.
SE015 ScienceNet 国内首个儿科大模型“福棠·百川”发布 国内首个儿科大模型“福棠·百川”发布。
SE016 Beijing Software and Information Service Industry Association 百川智能携手北京儿童医院发布全球儿科大模型,双医模式助力基层医疗升级 百川智能携手北京儿童医院发布全球儿科大模型,双医模式助力基层医疗升级。
SE017 Baichuan AI 百川大模型-百川智能 Baichuan4-Turbo、Baichuan4-Air、Baichuan-M3、Baichuan-M2 等产品在主页被直接展示。
SE018 Baichuan AI 百川大模型开放平台首页 全链路领域增强工具链+优质通用训练数据,企业多元场景达到可用率96%。
SE019 Baichuan AI 用户协议 本平台仅提供咨询服务...不代表医疗诊断等专业意见,也不构成处方或电子处方。
SE020 Baichuan AI 隐私政策 我们建立了专门的应急响应团队,针对不同安全事件启动安全预案。
SE021 arXiv Baichuan-M3: Modeling Clinical Inquiry for Reliable Medical Decision-Making Baichuan-M3 models the clinical decision-making process for reliable medical decision support.
SE022 China AI Atlas Baichuan AI (百川智能) lab profile Pivoted to healthcare AI mid-2024; flagship is Futang Baichuan pediatric AI model with Beijing Children's Hospital.
SE023 Pandaily Baichuan releases “M2Plus”, an Evidence-Augmented Medical Model Billed as a “ChatGPT for Doctors” Baichuan releases “M2Plus”, an Evidence-Augmented Medical Model Billed as a “ChatGPT for Doctors”.
SE024 TMTPost China's Baichuan Intelligence Launches Medical AI Model That Outperforms OpenAI in Key Benchmark Baichuan-M2 is already deployed at Beijing Children's Hospital and the Cancer Hospital of the Chinese Academy of Medical Sciences.
SE025 GitHub / baichuan-inc Baichuan-7B repository A large-scale 7B pretraining language model developed by BaiChuan-Inc.
SU001 Ying.ai / Baichuan ying.ai archived homepage AI-powered medical inquiry experience is surfaced through the Ying / Bai Xiaoyi path.
SU002 Baichuan AI Baichuan-M3 blog post Baichuan-M2 ... attracted hundreds of thousands of model downloads.
SU003 Baichuan AI platform.baichuan-ai.com root 百川大模型开放平台。
SU004 Baichuan AI 合作咨询 business proposal 欢迎各位开发者、企业和机构与我们一起开创人工智能新时代。
SU005 Houdao AI Baichuan Intelligence Launches Medical LLM BaiXiaoYi Partners with top hospitals to create doctors.
SU006 AIBase BaiChuan Intelligence Launches Baichuan-M4 Large Model and BaiXiaoYi AI Medical System BaiXiaoYi AI Medical System transitions from consultation to general practice management.
SU007 ToolChase Baichuan Intelligence Review 2026 Best for developers and researchers wanting clean open-source Chinese LLMs, plus Chinese healthcare teams exploring clinical AI.
SU008 Beijing Software and Information Service Industry Association 百川智能携手北京儿童医院发布全球儿科大模型 百川智能携手北京儿童医院发布全球儿科大模型,双医模式助力基层医疗升级。
SU009 ScienceNet 国内首个儿科大模型“福棠·百川”发布 国内首个儿科大模型“福棠·百川”发布。
SU010 36Kr / 职场Bonus Bonus独家 | 百川智能急刹车,调整医疗ToB,基础研发停摆 百川智能商业服务部门2024年给百川带来了中国银行、北电数智、招商银行、信雅达、天开集团等多个头部企业客户和合作伙伴。
SU011 Caixin GPT革命|专访百川智能茹立云:创业需避开互联网厂商 商业化半年签约订单数亿元 茹立云预计2025年订单签约额将在10-20亿元的水平。
SU012 Baichuan AI 百川大模型开放平台首页 内置丰富行业工作流,高效构建企业专属智能体。
SU013 Baichuan AI 百小医页面 看病前帮你梳理症状,准备就医;看病后帮你分析病情,解读医嘱。
SU014 TMTPost China's Baichuan Intelligence Launches Medical AI Model That Outperforms OpenAI in Key Benchmark The company has already deployed AI pediatricians in Beijing Children's Hospital and launched the “Futang·Baichuan” pediatric foundation model.
SU015 GitHub / baichuan-inc Baichuan-M3-235B repository Baichuan-M3 Modeling Clinical Inquiry for Reliable Medical Decision-Making.
SU016 Hugging Face / baichuan-inc Baichuan-M3-235B model card Create an OpenAI-compatible API endpoint using sglang or vllm.
SU017 Baichuan AI 百川大模型官网 百小医 你的 AI家庭医生。
SU018 Baichuan AI 用户协议 本平台仅提供咨询服务...不代表医疗诊断等专业意见。
SU019 Baichuan AI 隐私政策 我们建立了专门的应急响应团队,针对不同安全事件启动安全预案。
SU020 Baichuan AI API docs 申请加入海纳百川计划·免费使用M3Plus API。
SU021 Baichuan AI 价格说明 Assistants API 具体价格如下:限时免费。
SU022 China AI Atlas Baichuan AI lab profile Pivoted to healthcare AI mid-2024; flagship is Futang Baichuan pediatric AI model with Beijing Children's Hospital.
SU023 Baichuan AI Pricing page / customer surface 立即体验 获取API。
SU024 Baichuan AI Models page 进入开源社区。
SU025 Baichuan AI NPC page 请填写合作咨询申请体验测试。
SR001 CAC 生成式人工智能服务管理暂行办法 《生成式人工智能服务管理暂行办法》...自2023年8月15日起施行。
SR002 State Council / Gov.cn 生成式人工智能服务管理暂行办法政策库页面 政府政策库页面同步发布生成式人工智能服务管理暂行办法。
SR003 CNBC U.S. takes step to halt Nvidia AI chip shipments to Chinese firms outside China The U.S. Department of Commerce on Sunday moved to close a year-old potential loophole... to Chinese entities located outside China.
SR004 Benzinga / Reuters Nvidia's China Sales Face Setback As H20 Chip Restrictions Catch Key Buyers Off Guard: Reuters The Trump administration informed it on April 9 that its H20 chip would require an export license for sales to China.
SR005 Yahoo Finance / Reuters US licenses Nvidia to export chips to China, official says The commerce department has started issuing licenses to Nvidia to export its H20 chips to China.
SR006 Business Standard Nvidia halts H20 chip production after China cites security concerns Chinese regulators have flagged security concerns, prompting suspension of production and deliveries.
SR007 Built In Trump Lifted the AI Chip Ban on China, Clearing Nvidia and AMD to Resume Sales: Now What? After halting AI chip exports to China in April 2025, the Trump administration quietly reversed its policy just three months later in July 2025.
SR008 LevelFields NVIDIA to Resumes H20 GPU Sales to China Amid Trade Policy Shift China remains deeply dependent on American GPUs, especially those made by Nvidia.
SR009 36Kr / 职场Bonus Bonus独家 | 百川智能急刹车,调整医疗ToB,基础研发停摆 医疗ToB业务定制化程度高、数据难以打通、年成本上百万元仍然昂贵。
SR010 Caixin GPT革命|专访百川智能茹立云:创业需避开互联网厂商 商业化半年签约订单数亿元 茹立云预计2025年订单签约额将在10-20亿元的水平。
SR011 TMTPost China's Baichuan Intelligence Launches Medical AI Model That Outperforms OpenAI in Key Benchmark The company has already deployed AI pediatricians in Beijing Children's Hospital.
SR012 BSIA 百川智能携手北京儿童医院发布全球儿科大模型 双医模式助力基层医疗升级。
SR013 ScienceNet 国内首个儿科大模型“福棠·百川”发布 国内首个儿科大模型“福棠·百川”发布。
SR014 Baichuan AI 用户协议 本平台仅提供咨询服务...不代表医疗诊断等专业意见。
SR015 Baichuan AI 隐私政策 我们建立了专门的应急响应团队,针对不同安全事件启动安全预案。
SR016 Baichuan AI 百川大模型开放平台首页 内置丰富行业工作流,高效构建企业专属智能体。
SR017 Baichuan AI API docs 申请加入海纳百川计划·免费使用M3Plus API。
SR018 Baichuan AI 价格说明 Assistants API 具体价格如下:限时免费。
SR019 Baichuan AI 百川大模型官网 百小医 你的 AI家庭医生。
SR020 GitHub / baichuan-inc Baichuan-M3-235B repository Baichuan-M3 Modeling Clinical Inquiry for Reliable Medical Decision-Making.
SR021 Hugging Face / baichuan-inc Baichuan-M3-235B model card Create an OpenAI-compatible API endpoint using sglang or vllm.
SR022 AIBase BaiChuan Intelligence Launches Baichuan-M4 Large Model and BaiXiaoYi AI Medical System BaiXiaoYi AI Medical System transitions from consultation to general practice management.
SR023 China AI Atlas Baichuan AI lab profile Pivoted to healthcare AI mid-2024.
SR024 Apidog Chinese LLM price war 2026 Chinese labs cut LLM API prices six times in the first half of 2026 and several cuts were made permanent.
SR025 Digital Applied Chinese AI models Q2 2026 market share report Ten providers now absorb most meaningful Chinese AI output and token share.
SR026 KrASIA After tech giants, a new cohort of AI tigers finds footing in China China’s new AI tigers still need to prove durable economics against giant-backed rivals.
SR027 Alibaba Cloud Model Studio / Qwen enterprise platform Qwen deploys through Alibaba Cloud’s Model Studio.
SR028 DeepSeek API pricing docs DeepSeek publishes direct API pricing and usage materials.
SR029 TrendForce DeepSeek completes first funding round at valuation above CNY 330 billion DeepSeek raised over CNY 50 billion at a valuation above CNY 330 billion.
SR030 ToolChase Baichuan Intelligence Review 2026 Best for developers and researchers wanting clean open-source Chinese LLMs, plus Chinese healthcare teams exploring clinical AI.
SR031 Reuters Nvidia says US to limit H20 chip exports to China Reuters headline: Nvidia says US to limit H20 chip exports to China.
SR032 Reuters China tech giants had ordered at least $16 billion of Nvidia H20 server chips in first quarter 2025 Reuters headline: China tech giants had ordered at least $16 billion of Nvidia H20 server chips in first quarter 2025.
SV001 Yahoo Finance / SCMP syndication Chinese AI start-up Baichuan raises $694 million in round valuing it at $2.8 billion Baichuan raises $694 million... valuing it at $2.8 billion.
SV002 CB Insights Baichuan financials profile CB Insights lists Baichuan financing history and company profile context.
SV003 Tracxn Baichuan funding and investors Tracxn tracks Baichuan investors and funding rounds.
SV004 37 Interactive / InvestGame 37 Interactive Entertainment FY2025 Annual Report The Board of Directors ... guarantee the truthfulness, accuracy and completeness of the contents of this Report.
SV005 Caixin GPT革命|专访百川智能茹立云:创业需避开互联网厂商 商业化半年签约订单数亿元 茹立云预计2025年订单签约额将在10-20亿元的水平。
SV006 36Kr / 职场Bonus Bonus独家 | 百川智能急刹车,调整医疗ToB,基础研发停摆 2024年百川ToB业务收入近亿元,但医疗ToB仍面临定制化和预算压力。
SV007 Apidog Chinese LLM price war 2026 Chinese labs cut LLM API prices six times in the first half of 2026 and several cuts were made permanent.
SV008 CNBC MiniMax Hong Kong IPO AI tigers Zhipu CNBC reported that MiniMax raised about $619 million in a Hong Kong IPO.
SV009 TrendForce DeepSeek completes first funding round at valuation above CNY 330 billion DeepSeek raised over CNY 50 billion at a valuation above CNY 330 billion.
SV010 TechCrunch China's Moonshot AI raises $2B at $20B valuation as demand for open-source AI skyrockets Moonshot raised $2 billion at a $20 billion valuation.
SV011 Digital Applied Chinese AI models Q2 2026 market share report Ten providers now absorb most meaningful Chinese AI output and token share.
SV012 CompaniesMarketCap Alibaba market capitalization As of August 2026 Alibaba has a market cap of $312.86 Billion USD.
SV013 CompaniesMarketCap Alibaba revenue Revenue in 2026 (TTM): $145.39 Billion USD.
SV014 CompaniesMarketCap Alibaba P/S ratio P/S ratio as of August 2026 (TTM): 2.15.
SV015 CompaniesMarketCap Baidu market capitalization As of August 2026 Baidu has a market cap of $31.19 Billion USD.
SV016 CompaniesMarketCap Baidu revenue Revenue in 2026 (TTM): $18.27 Billion USD.
SV017 CompaniesMarketCap Baidu P/S ratio P/S ratio as of August 2026 (TTM): 1.71.
SV018 CompaniesMarketCap NVIDIA market capitalization As of August 2026 NVIDIA has a market cap of $5.252 Trillion USD.
SV019 CompaniesMarketCap NVIDIA revenue Revenue in 2026 (TTM): $253.49 Billion USD.
SV020 CompaniesMarketCap NVIDIA P/S ratio P/S ratio as of August 2026 (TTM): 20.7.
SV021 Baichuan AI 百川大模型官网 百小医 你的 AI家庭医生。
SV022 Baichuan AI 百川大模型开放平台首页 内置丰富行业工作流,高效构建企业专属智能体。
SV023 Baichuan AI 价格说明 Assistants API 具体价格如下:限时免费。
SV024 TMTPost China's Baichuan Intelligence Launches Medical AI Model That Outperforms OpenAI in Key Benchmark The company has already deployed AI pediatricians in Beijing Children's Hospital.
SV025 BSIA 百川智能携手北京儿童医院发布全球儿科大模型 双医模式助力基层医疗升级。
SV026 ScienceNet 国内首个儿科大模型“福棠·百川”发布 国内首个儿科大模型“福棠·百川”发布。
SV027 ToolChase Baichuan Intelligence Review 2026 Best for developers and researchers wanting clean open-source Chinese LLMs, plus Chinese healthcare teams exploring clinical AI.
SV028 GitHub / baichuan-inc Baichuan-M3-235B repository Baichuan-M3 Modeling Clinical Inquiry for Reliable Medical Decision-Making.
SV029 Hugging Face / baichuan-inc Baichuan-M3-235B model card Create an OpenAI-compatible API endpoint using sglang or vllm.
SV030 Alibaba Cloud Model Studio / Qwen enterprise platform Qwen deploys through Alibaba Cloud's Model Studio.