初创公司尽调
尽调报告 On-device / edge foundation models (China) Growth-stage private company at unicorn threshold after April 2026 financing 2026-07-02

ModelBest

端侧模型已有真实牵引力,也拿到战略资本;但作为独角兽标的,披露仍不足

ModelBest 像是一家值得继续研究的边缘 AI 公司,开源牵引力和战略投资人支持都真实存在;但当前接近独角兽门槛的价格,在收入、定价和证券条款上仍缺少披露支撑。

封面要素

最新估值信号 01
1000 USD M+ (unicorn-threshold after Apr 2026 round; exact post-money undisclosed) [CO017, CV006]
2026 年 Q1 融资 02
1000 RMB M+ cumulative across China Telecom and Apr 2026 rounds [CO016]
据报道 MiniCPM 下载量 03
24 M+ cumulative across GitHub and Hugging Face (company-reported via press) [CO026, CV012]
开源足迹 04
158 models on OpenBMB Hugging Face org snapshot [CO021]
成立时间 05
Aug 2022 [CO002]
总部 06
Beijing, China [CO005]

公司概况

ModelBest(面壁智能),法定名称为北京面壁智能科技有限责任公司,是一家北京端侧模型创业公司,2022 年 8 月围绕 Li Dahai、Liu Zhiyuan 和 Zeng Guoyang 创立。公司脱胎于清华大学 NLP / OpenBMB 生态,把 MiniCPM 系列做成紧凑、面向部署的模型家族,用来替代只能跑在云端的前沿模型。公开来源显示,它的打法先靠 OpenBMB、GitHub 和 Hugging Face 做开源分发,再进入手机、AI PC、智能座舱、具身机器人、可穿戴设备和法律 AI 工作流等端侧与嵌入式部署场景。2024-2026 年连续融资后,2026 年 4 月的一轮融资把 ModelBest 推过独角兽门槛,但收入质量、定价结构、治理和具体投资条款仍缺公开披露。

官网
www.modelbest.cn
创始人
Li Dahai, Liu Zhiyuan, Zeng Guoyang
创立地点
Beijing, China
总部
Beijing, China
产品
MiniCPM 是一组紧凑型语言、多模态和全模态模型,面向主流芯片、本地或端侧部署优化;分发路径是官方文档、开源仓库、Hugging Face 发布和部署演示,而不是清晰可见的公开 API 下单流程。
客户
设备 OEM 与车企项目嵌入端侧 AI,企业和公共部门团队部署本地推理,广泛的开发者 / 开源社区则在手机、PC、机器人和嵌入式工作流上评估 MiniCPM。
商业模式
最可能的收入组合是 OEM 嵌入、企业 / 本地部署、方案集成,以及从开源走向商业转化,而不是完全公开的自助式 API 定价模型;具体合同结构和实际成交价格仍未披露。
阶段
Growth-stage private company / unicorn threshold
融资情况
2024-2025 年融资由 Primavera / Hubble、Loongson / SAIF 关联方、Hongtai / Gozone 等轮次逐步堆起;2026 年先有 China Telecom 领投,随后 4 月 Shenzhen Capital Group 和 Huichuan 领投,据报道把第一季度累计融资推到 RMB1 billion 以上,并把公司推过 $1 billion 估值门槛。
[CO001, CO002, CO005, CO007, CO008, CO009, CO015, CO016]

执行摘要

主要优势

  • 切入点贴合真实市场缺口:面向端侧和边缘部署的紧凑多模态模型,而不是泛化的纯云端 LLM 供给。
  • OpenBMB、GitHub 和 Hugging Face 带来可见生态触达,包括 158 个托管模型,以及据称 24M+ 累计下载量。
  • 战略投资人覆盖电信、国资背景创投和工业自动化,不只补资本,也改善渠道入口。
  • 电信合作和汽车项目里已有具名落地证据,让故事比纯研究实验室叙事更有实质。

主要风险

  • 公开资料没有披露收入、ARR、价格表、毛利率或客户数锚点,当前价格仍难以承保。
  • 据报道 $1B+ 的估值,更像由战略稀缺性和投资人联盟质量支撑,而不是由已验证变现能力支撑。
  • 汽车和设备渠道证据集中在少数具名项目和伙伴上,抬高了集中度和执行风险。
  • 如果 OEM 和企业付费转化率仍弱或不透明,开源热度未必能转成付费部署经济性。

未决问题

  • 经审计收入、毛利率和月度烧钱桥接仍未披露。
  • 2026 年 4 月准确投后估值、稀释和优先股堆叠未公开。
  • 公开资料没有披露付费客户数、续约行为或头部客户集中度。
  • 已抓取公开记录中的董事会构成和股权结构持股仍不完整。

目录

Chapter 01

01公司概况

1.1 身份、起源与端侧模型定位

ModelBest 以北京面壁智能科技有限责任公司为法律主体,公开来源把成立时间放在 2022 年 8 月。抓取到的 2026 年融资资料一致把它描述为一家北京创业公司,源自清华大学自然语言处理生态;OpenBMB 自己的材料也称,开源社区由清华 NLP Lab 和 ModelBest 联合发起。公司并不把自己包装成通用前沿模型实验室:官网称,它想把大模型放到最贴近用户的地方,覆盖手机、AI PC、智能座舱、具身机器人、可穿戴设备和法律 AI 工作流。官网、GitHub 仓库和 Hugging Face 组织页共同指向 MiniCPM:一个紧凑、面向部署的模型家族,靠知识密度和设备效率取胜,而不是纯拼参数规模。[CO001, CO002, CO003, CO004, CO005, CO006]

快照 KPI 表
指标值 / 状态日期置信度缺口
法律实体北京面壁智能科技有限责任公司
成立时间August 20222022-08-12
总部 / 模型研发基地北京2026-04-09审阅的公开资料没有提供清晰的当前多地办公室列表,也没有给出精确的公开总部地址。
当前阶段2026 年 4 月报道中跨过独角兽门槛的私营创业公司2026-04-09公开报道验证了跨过门槛,但未验证精确投后估值。
最新披露轮次由 Shenzhen Capital Group 和 Huichuan Capital 领投,数亿元人民币2026-04-07
2026 年累计融资2026 年 Q1,China Telecom 轮和 4 月轮合计超过 RMB 1 billion2026-04-09
核心产品家族MiniCPM 端侧 / 边缘模型家族,包含语言、多模态和全模态变体2026-07-02
开源足迹OpenBMB GitHub 加 Hugging Face 分发;Hugging Face 组织页显示 150+ 个模型和多个演示空间2026-07-02开发者信号指标显示公开活跃度,但不显示收入或付费用户。
披露的 MiniCPM 下载量GitHub 和 Hugging Face 累计下载 24M+2026-04-09这是媒体引用的公司口径指标,不是独立审计遥测数据。
具名部署证据公开文章提到 Changan Mazda EZ-60、Geely Galaxy M9,以及手机 / AIPC / 智能家居项目2026-04-09审阅来源给出了例子,但没有量化总出货量或付费 OEM 合同。
公开董事会名单抓取的公开来源没有提供完整现任董事会名单或完全稀释后股权拆分。

融资、估值、下载和部署行混合了公司官方定位与第三方报道。空值表示抓取的公开记录不足以为第一章事实条精确验证的信息。

[CO001, CO002, CO003, CO004, CO015, CO016]
FO002: 公司快照逻辑

ModelBest 把清华 / OpenBMB 研究根基接到 MiniCPM 开源分发,再试图把它转成设备部署和产业资本,同时处理生态依赖。

[CO005, CO006, CO020, CO024, CO029, CO033]

1.2 领导层、治理与关键人依赖

公开领导层记录聚焦创始三人组,而不是完整披露的高管梯队。Li Dahai 是外部可见的 CEO,也是公司智能密度法则和开源定位的主要发声者;抓取到的融资资料还称,他在创立 ModelBest 前曾任 Zhihu CTO / 高管。可访问的学术和人物资料显示,Liu Zhiyuan 是联合创始人兼首席科学家,同时仍在清华任教;Zeng Guoyang 是联合创始人兼 CTO,直接负责 MiniCPM 工程。这个组合让 ModelBest 在研究、开源工具和产品化上有很强的创始人-市场匹配,但也带来集中风险:本次审阅的公开记录仍没有完整董事会名单或清晰持股拆分,后续治理分析需要私密材料,不能只靠媒体报道。[CO007, CO008, CO009, CO010, CO032, CO033]

领导层和创始人表
人物职务背景创始人市场匹配或职能覆盖关键人物依赖
Li Dahai联合创始人兼 CEO前 Zhihu CTO / 高管把研究资产转化为资本市场、产品和商业化叙事;密度定律策略的公开代表
Liu Zhiyuan联合创始人兼首席科学家清华计算机科学教员;长期 THUNLP 研究员支撑研究可信度、清华网络入口和 OpenBMB 生态合法性
Zeng Guoyang联合创始人兼 CTOTHUNLP 培养的工程师,创立 ModelBest 前参与 WuDao / Wenyuan 相关工作负责 MiniCPM 工程执行和跨端侧模型发布的产品化

该表列出抓取公开记录中明确具名的创始人领导者。审阅材料没有披露完整高管名单或董事会组成,因此覆盖面有意保持部分。

[CO007, CO008, CO009, CO010, CO032, CO033]

1.3 融资历史、投资人联盟与当前阶段

ModelBest 的融资史显示,这家公司一边高频融资,一边扩大政治与产业赞助。Caixin 记录了 2024 年 Primavera / Hubble 领投轮、2024 年 12 月 Loongson-Dinghui-Zhongguancun / SAIF 轮,以及 2025 年 5 月 Hongtai、Gozone、Qingkong Jinxin 和 Moutai Fund 参与的轮次。2026 年节奏再次加快:新华社体系报道了 2 月 China Telecom Investment 领投轮,随后一组 2026 年 4 月来源称,Shenzhen Capital Group 和 Huichuan Capital 又领投了数亿元人民币新一轮融资。同一批 4 月报道称,2026 年第一季度累计融资超过 RMB 1 billion,最新一轮把 ModelBest 推过独角兽门槛。公开报道仍未披露的是 4 月投后估值的精确 RMB 或 USD 数值,因此本章只把 >US$1 billion 视为方向性已核实,精确数字仍待解决。[CO011, CO012, CO013, CO014, CO015, CO016]

利益相关方或投资人地图
利益相关方角色控制权或经济重要性尽调问题
Zhihu天使 / 战略股东2023 年 4 月天使轮锚定方和早期战略支持者,与 Li Dahai 有创始人关系确认当前持股比例,以及任何商业合作或信息权。
Beijing AI Industry Investment Fund 等投资方北京国资支持方多次出现在 2024 年轮次中,代表围绕北京研发基地的地方政府支持要求披露董事会 / 观察员权利,以及任何政策挂钩商业化条件。
Primavera Capital + Huawei Hubble2024 年 A 轮领投方传递对端侧模型投资逻辑的早期机构和产业信心厘清任一投资方是否持有与未来融资或战略合作绑定的特殊权利。
Loongson Venture 及相关 2024 芯片生态支持方国产算力生态投资人把 ModelBest 更紧密地连接到中国半导体和基础设施生态梳理这些投资人是否推动部署、联合营销或技术适配承诺。
China Telecom Investment2026 年战略投资人带来云、网络、政企渠道和企业分发杠杆要求披露所称云 / 网络 / 终端协作背后的商业条款。
Shenzhen Capital Group (深创投)2026 年 4 月领投方增加国家级硬科技国资,并强化深圳侧影响力确认深圳支持是否改变注册地、城市激励或后续融资预期。
Huichuan Capital / 汇川产投2026 年 4 月产业领投方连接 ModelBest 与工业自动化、机器人和具身 AI 商业化渠道厘清投资附带哪些具体产品、工厂或机器人管线入口。
QCC 列出的广泛股东扩大的股东基础,包括 Zhipu、Hubble、Guotai Junan 和多个基金载体显示股权表拥挤,战略和财务利益方很多获取完整股权表、优先权和治理权利清单,而不是只依赖股东名称快照。

这是公开利益相关方地图,不是权威所有权台账。它优先纳入抓取报道和注册信息来源中反复出现的具名投资人和战略利益相关方。

[CO011, CO012, CO013, CO014, CO015, CO019]

1.4 里程碑、开源牵引力与部署证据

ModelBest 第一章最有辨识度的资产在于,技术定位不只写在投资人材料里,也真实出现在公开开发者渠道。OpenBMB、GitHub 和 Hugging Face 都展示了围绕 MiniCPM 组织的生态,主仓库也明确主打端侧、移动端和多模态部署。ThePaper 2026 年 4 月报道称,公司在 2023 年底发布 XAgent、AgentVerse 和 ChatDev,2024 年初更重押 MiniCPM;Tencent News 随后报道,2026 年开源的 MiniCPM-o 4.5 是一个 9B 全双工全模态模型。多篇 2026 年 4 月融资报道进一步称,MiniCPM 在 GitHub 和 Hugging Face 的累计下载量超过 24 million,部署已经进入 Changan Mazda EZ-60、Geely Galaxy M9 等具名汽车项目。这些牵引力数字很重要,但仍是公司通过媒体渠道披露的数据,不是独立审计的遥测。[CO020, CO021, CO022, CO023, CO024, CO025]

里程碑表
日期事件类型金额 / 估值 / 状态参与方含义
2022-08-12ModelBest 成立创立公司在北京成立Li Dahai、Liu Zhiyuan、Zeng Guoyang 三位创始人为后续章节建立可复用的创立锚点。
2023-04-01后续融资报道提到 Zhihu 支持的天使轮融资回溯披露的天使轮Zhihu; ModelBest显示公司早期就有战略股东支持。
2023-12-01智能体工具浪潮:XAgent、AgentVerse、ChatDev 及相关 AI 开发工具产品2023 年末多智能体工具推进ModelBest; OpenBMB 生态显示公司在更集中转向 MiniCPM 前,曾尝试智能体软件。
2024-01-15MiniCPM 成为核心产品方向产品2024 年初转向端侧模型的战略调整ModelBest把后续端侧重点框定为有意策略,而不是后期转向。
2024-04-112024 年春季融资轮完成融资数亿元人民币Primavera Capital、Huawei Hubble、Beijing AI Industry Investment Fund、Zhihu 等投资方标志着抓取到的第一轮明确大型机构融资。
2024-12-01另一轮大额融资完成融资数亿元人民币Loongson Venture、Dinghui Bafu、Zhongguancun Science City Fund、SAIF 等投资方加深芯片、国资和成长资本支持。
2025-05-21追加成长轮完成融资数亿元人民币;确切估值未披露Hongtai Fund、Gozone Capital、Qingkong Jinxin、Moutai Fund 等投资方显示即便此前已多次融资,公司仍持续依赖资本。
2025-08-15三周年公开信发布治理Li Dahai 重申“端侧第一”和长期 AGI 叙事Li Dahai 和管理团队提供扩张周期中段的战略自画像。
2026-02-04MiniCPM-o 4.5 开源产品9B 全双工全模态模型ModelBest; OpenBMB用明确的 2026 年发布里程碑更新产品前沿。
2026-02-28China Telecom 领投融资完成融资数亿元人民币China Telecom Investment、CITIC Jingshi、CITIC Private Equity 等投资方把战略电信投资人带入股权表和企业渠道。
2026-04-07Shenzhen Capital / Huichuan 轮完成,ModelBest 被认定为达到独角兽门槛的公司融资数亿元人民币;2026 年 Q1 累计融资 >RMB1B;投后 >~US$1B 门槛Shenzhen Capital Group、Huichuan Capital、Daohe、Guotai Junan Innovation、Wuyuefeng 等投资方确认资本获取和外部阶段认知都上了一个台阶。
2026-06-13CEO 公开把边缘 AI 瓶颈描述为生态问题反向芯片 / 内存 / 软件共建仍有必要Li Dahai; Tencent News把商业化摩擦转化为创始人明确承认的风险因素。

该表是第一章唯一记录的时间线,混合公司、媒体和面向开发者的里程碑。2023-04-01、2024-01-15 等日期是在抓取来源集未披露精确日期时使用的月份级锚点。

[CO002, CO011, CO012, CO013, CO014, CO015]
FO001: 公司里程碑时间线

ModelBest 从 2022 年清华关联创立,走到 2024-2026 年连续融资、以 MiniCPM 为中心的产品收束,以及 2026 年边缘 AI 扩张争论。

[CO002, CO011, CO013, CO014, CO015, CO025]
FO003: 快照 KPI

公开 KPI 图景指向一家快速融资的边缘模型创业公司:开源牵引力真实,披露保留也真实。

[CO016, CO017, CO018, CO026, CO027, CO028]

1.5 反向信号与仍未披露的问题

第一章的反向记录不是已知诉讼或处罚,而是商业化压力、生态依赖和披露不完整。Caixin 2024 年 4 月报道在大额融资后仍把商业化列为核心问题,2025 年 5 月跟进又强调 ModelBest 已在一年多内完成三轮融资。Tencent 2025 年 12 月的质疑性报道认为,汽车座舱机会已经挤满芯片、OS 和汽车软件栈的既有玩家;Li Dahai 2026 年 6 月采访也承认,端侧 AI 进展现在靠芯片、内存、带宽和软件生态共同建设,不能只靠模型质量。再叠加尚未披露的董事会名单、股权结构表、4 月投后精确估值,以及可独立审计的 24 million 下载遥测,ModelBest 在公司概况阶段看起来有前景,但尽调负担仍重。[CO010, CO018, CO027, CO034, CO036, CO037]

1.6 图表

Chapter 02

02市场分析

2.1 市场边界:ModelBest 位于端侧模型层,不等于整个 AI 经济

ModelBest 自己的材料并没有描述一个宽泛的云端模型平台;它描述的是一家公司,试图把大模型放到最贴近用户的手机、AI PC、智能座舱、具身机器人、可穿戴设备和其他终端上。MiniCPM 仓库也强化了这个框架,强调适合本地部署和资源受限环境的紧凑模型。36Kr 和 Gasgoo 的独立报道指向同一方向:ModelBest 被定位为早期边缘智能厂商,商业路径穿过 OEM、芯片伙伴和终端部署,而不是纯公共云流量。因此,真正的市场包括端侧模型集成、本地推理软件、座舱 AI,以及对隐私敏感的本地部署;它排除了大部分上游云训练资本开支,也排除了相当一部分泛应用经济支出。换句话说,ModelBest 竞争的地方,是 NPU 能力、模型压缩、OEM 集成以及隐私或时延约束比模型参数规模头条更重要的地方。[CM001, CM002, CM003, CM005, CM006, CM007]

市场定义表
细分 / 类别纳入支出排除支出主要买方 / 付款方为什么对 ModelBest 重要
AI 手机模型层端侧模型集成、本地助手、多模态设备功能,以及经 NPU 调优的手机 AI无法归因于 AI 的通用手机硬件价值、运营商服务收入、无关 app 支出智能手机 OEM 和平台团队IDC 和 ModelBest 都把品类框定在设备端 GenAI 能力,而不是纯云端流量
AI PC / 本地智能体层本地推理软件、捆绑 copilot 工具、NPU 优化模型部署、高端设备差异化没有 AI 差异化的大宗 PC 硬件、通用办公软件收入PC OEM、芯片伙伴、企业设备团队ModelBest 公开营销 AI-PC 相关性,AI-PC 渗透率也在快速上升
智能座舱 / 车端边缘 AI座舱多模态交互、本地语音 / 视觉模型、舱内智能体工作流、车端模型集成电池、动力总成和非 AI 车辆 BOM汽车 OEM 和座舱项目负责人ModelBest 已披露车辆项目和座舱相关性,让汽车成为有证据支撑的需求面
受监管本地部署在法律、工业和公共部门工作流中进行本地或混合部署,适用于隐私或时延重要的场景广义云训练资本开支和无关企业 SaaS 预算企业 IT、合规或业务单元负责人中国监管越来越奖励能在敏感工作流中管理备案、标识和数据处理的供应商
排除的上层云 / 训练经济无直接纳入;这一层主要是输入成本或竞争参照基准超大云厂商云训练支出、GPU 集群、通用公有云模型服务收入云平台和算力所有者ModelBest 会通过芯片成本和模型价格压力受这一层影响,但这里不是计算公司自身可寻址收入的最干净口径

边界遵循 ModelBest 自身终端优先定位,以及关于其 OEM 驱动部署的独立报道。它有意排除大部分上游云和训练支出;这些支出更适合作为输入成本、伙伴杠杆或竞争压力处理。

[CM001, CM002, CM003, CM006, CM021, CM048]

2.2 市场测算需要多副镜头,因为公开估算描述的是不同层

没有一个公开数字能干净衡量 ModelBest 的可服务市场,所以本章并列保留多副镜头,而不是硬凑一个 TAM。设备普及是最清晰的近期镜头:IDC 的全球预测把 GenAI 手机放在 2024 年 234.2 million 台、2028 年 912 million 台;Counterpoint 预计,2026 年具备 GenAI 能力的设备将达到全球智能手机出货量的 45%。中国需求也很大:IDC 关联报道预计,2026 年中国智能手机出货约 278 million 台,其中 147 million 台是下一代 AI 手机,渗透率为 53%。AI PC 是第二副镜头,Counterpoint 称具备 AI 或 GenAI 能力的笔记本在 2024 年已拿下 27% 市场,2025 年可能接近 60%。第三副镜头是更宽的端侧模型经济,但这些数字彼此冲突,因为范围并不相同:Grand View 把 2025 年端侧 AI 市场估为 USD 10.76 billion,BCC 把边缘 AI 估为 USD 11.8 billion,MarketsandMarkets 则把边缘 AI 硬件估为 USD 26.14 billion。[CM007, CM008, CM010, CM012, CM014, CM015]

TAM / SAM / SOM 或规模测算镜头表
发布方镜头地理范围当前值增长 / 采用信号数字实际衡量什么局限
IDCGenAI 智能手机出货量全球2024 年 234.2M 台2028 年 912M;78.4% CAGR按 >30 TOPS 端侧 GenAI 能力定义的下一代 AI 智能手机出货量全球出货量口径,不是中国收入或 ModelBest 份额
Counterpoint具备 GenAI 能力的智能手机占比全球2026 年出货量 45%2027 年 52%;2026 年智能手机总出货量降至 1.08B具备 GenAI 能力的智能手机出货占比追踪的是具备能力的设备,不是任何单一模型供应商的变现
IDC,经 Tencent/CNMO 转引AI 手机采用中国2026 年智能手机总出货量 278M下一代 AI 手机 147M;渗透率 53%中国设备换新口径,覆盖最相关的终端品类计算设备台数,不计算 ModelBest 在这些设备中的绑定率
CounterpointAI 笔记本渗透率全球2024 年笔记本市场 27%2025 年接近 60%AI PC 成为标准出货组合的速度全球笔记本口径,不是中国 AI-PC 收入池
Grand View Research端侧 AI 市场全球2025 年 USD 10.76B2030 年 USD 36.64B;27.8% CAGR覆盖组件、设备和垂直行业的广义端侧 AI 收入估计范围比 ModelBest 更宽,也不同于边缘 AI 或纯硬件估算
BCC Research边缘 AI 市场全球2025 年 USD 11.8B2030 年 USD 56.8B;36.9% CAGR覆盖产品和行业的全球边缘 AI不宜直接对比只覆盖端侧 AI 或只覆盖硬件的估算
MarketsandMarkets边缘 AI 硬件市场全球2025 年 USD 26.14B2030 年 USD 58.90B;17.6% CAGR仅硬件口径的边缘 AI 市场,包含智能手机、可穿戴、汽车和边缘服务器口径更大,因为只切硬件,不含软件或混合市场层
Counterpoint边缘 AI 可穿戴全球2025 年出货渗透率 30%到 2032 年接近 80%;边缘 AI 占累计 USD 1T 收入机会的约 75%从相邻市场看使用本地 AI 的小型常开设备预测周期较长,也不对应 ModelBest 当前收入结构
TrendForce汽车半导体市场全球2024 年 USD 67.7B2029 年 USD 96.9B;逻辑处理器 8.6% CAGR支撑智能座舱的高算力汽车硬件层属于硬件使能层,不是直接软件 / 模型收入

这些行刻意不做同质化:单位、地域、层级都不同,只用来勾勒 ModelBest 品类周边的需求包络,不能相加成一个总 TAM。

[CM008, CM010, CM014, CM015, CM016, CM017]
FM001: 市场规模测算视角

证据受限的金字塔:从全国 AI 终端采用的宽口径,收窄到 ModelBest 已披露活动的更窄场景。

该图刻意混合政策渗透目标、设备份额和收入代理,因为已审阅公开来源没有单独切出一个干净的中国端侧 AI TAM,能精确对应 ModelBest 的品类。

[CM001, CM014, CM015, CM016, CM017, CM022]
FM002: 市场估计区间

围绕端侧和边缘 AI 的 2025 与 2030 年市场规模估计,保留已发布低 - 高区间和口径漂移,而不是强行压成一个 TAM。

低值来自端侧 AI,中间公开估计来自更宽的边缘 AI,高值来自边缘 AI 硬件。三者都成立但口径不匹配,因此这个区间应被理解为分散度,不是精确值。

[CM017, CM018, CM019, CM020]

2.3 买方地图:设计导入和本地工作流批准,比独立应用份额更关键

ModelBest 的市场路径与消费者聊天机器人公司结构上不同。公司公开披露了与车企、手机或 PC 生态、芯片厂商、机器人硬件玩家的合作,Gasgoo 也把其商业化直接连到汽车、智能手机、PC 和智能家居部署。这意味着真正买方通常是 OEM 产品团队、平台方、芯片伙伴,或企业和公共部门部署方,而不是终端用户。在智能手机和 AI PC 里,实际决策者是 OEM 或平台团队,他们判断本地 AI 功能是否值得付出内存、NPU 和集成成本。在汽车里,买方是掌握座舱架构和上市节奏的车型项目负责人。在受监管的本地部署里,买方是看重隐私、时延和本地控制的 IT 或合规组织。这种结构把价值捕获从零售应用使用量,转向设计导入、产品周期节奏和伙伴集中度;因此,即便设备级 AI 普及加速,议价权仍可能很高。[CM025, CM026, CM027, CM028, CM029, CM030]

细分市场 / 买方地图
细分市场买方用户付费方工作流预算负责人采用触发因素
AI 手机集成手机 OEM 和移动平台团队智能手机终端用户OEM 软件 / 产品预算本地助手、端侧多模态 AI、隐私敏感推理手机业务 GM 或软件平台负责人旗舰或高端产品周期差异化
AI PC / 本地生产力PC OEM、芯片伙伴或企业设备团队知识工作者、开发者或高阶消费者设备捆绑预算或企业更新预算端侧 Copilot、编程、生产力、离线或低延迟任务品类 GM、渠道负责人或 CIOWindows 10 换机、NPU 可用性和高端 AI PC 加价销售
智能座舱汽车 OEM 和座舱架构团队驾驶员和乘客整车平台研发预算语音、多模态 HMI、导航和舱内智能体工作流车型项目负责人新车型发布和座舱 SoC 集成周期
机器人 / 可穿戴 / 智能家居硬件 OEM终端设备所有者硬件研发和产品预算本地控制、常开感知、低功耗交互设备品类负责人电池、延迟和形态约束让依赖云端的方案吃亏
受监管本地部署企业或公共部门 IT / 合规团队员工、法官、操作员或公民IT、转型或合规预算用本地或混合推理辅助敏感工作流CIO、合规负责人或业务单元负责人数据敏感性、备案负担或工作流延迟要求

这张买方地图看的是采购现实,不是简单用户数。ModelBest 的关键交易对手,往往是掌握硬件发布、渠道入口或合规审批的一方;他们比直接零售应用需求更重要。

[CM025, CM026, CM027, CM028, CM029, CM048]
FM003: 买方 / 细分市场地图

这条流展示 ModelBest 的市场路径:从芯片和模型优化,进入 OEM 项目,再到受监管或消费者端使用。

该流程简化了多条并行渠道。真实项目往往在芯片调优、OEM 测试、备案和重新发布之间来回循环,而不是严格线性推进。

[CM025, CM026, CM027, CM029, CM031, CM035]

2.4 政策顺风与商业化关口同时放大市场、筛选玩家

中国政策组合一边扩大机会,一边抬高参与门槛。2023 年 CAC 和 MIIT 生成式 AI 措施把面向公众的服务纳入监管,要求内容治理、用户数据保护,并且对部分服务要求安全评估和备案。2025 年标识规则又在生成和分发链条上加入显式与隐式标识要求。与此同时,国务院 2025 年 AI+ 行动设定了激进的国家目标:到 2027 年,新一代智能终端和智能体渗透率超过 70%;到 2030 年超过 90%;并明确优先发展 AI 手机、AI 电脑、网联汽车、可穿戴设备、智能家居和智能体应用。2026 年 5 月 AI 智能体指引又增加了运营层,定义智能体并识别 19 个应用场景。CAC 2026 年 4 月备案总数也显示了规模:截至 4 月 30 日,已有 868 项生成式 AI 服务备案、530 个应用或功能完成登记。对 ModelBest 来说,这是顺风,也是门槛:政策让端侧部署更合法,却把备案、标识和可追溯性变成反复要做的商业化工作。[CM031, CM032, CM033, CM034, CM035, CM036]

增长驱动与约束表
驱动 / 约束方向证据时间估值含义尽调问题
隐私、离线可用性和低延迟推理驱动36Kr 和 Grand View 都称,本地处理比纯云端推理更快,也更能保护隐私当前网络质量或数据敏感性重要时,支撑边缘模型采用ModelBest 部署中,有多大比例主要由隐私或延迟经济性支撑?
AI+ 终端和智能体目标驱动国务院 AI+ 计划提出,终端 / 智能体渗透率到 2027 年超过 70%,到 2030 年超过 90%2025-2030智能终端和智能体式产品获得强政策顺风哪些被点名的重点品类已经能为 ModelBest 变现?
2026 年 AI 智能体指引和 19 个场景驱动国新办 / 新华社称,已在科研、产业、消费、民生和治理等领域明确 19 个应用场景近期扩大智能体式产品的合规用例哪些场景匹配 ModelBest 当前产品能力和合规准备度?
备案和标识制度驱动兼约束2023 年办法叠加 2025 年标识规则,带来审批、披露和可追溯义务持续准备充分的供应商可把它变成护城河,但也推高持续合规成本ModelBest 哪些产品已经备案、打标,或通过注册主体分发?
内存涨价和手机需求疲软约束IDC 关联的中国预测和 Counterpoint 都提到内存成本上升、中国设备需求放缓2026可能拖慢出货,也压缩 OEM 为模型差异化付费的意愿如果 OEM 重新收缩中端机发布,ModelBest 的目标项目有多抗压?
中国 LLM 价格战约束KrASIA 和 DigWatch 显示,市场出现极端降价、免费档和很低的 token 价格持续边缘效率因此更具战略价值,但按 token 或功能变现可能塌陷ModelBest 按设备、API 或混合部署拆开的真实毛利率是多少?
美国芯片出口收紧叠加国产替代约束兼驱动CNBC 报道 2026 年 5 月堵上漏洞;CSIS 认为管制会加速中国替代即期和中期它抬高算力不确定性,也让适配国产栈的高效模型更值钱训练端多依赖进口算力?推理端多能落到国产友好栈?
OEM 和渠道议价权约束中国智能手机份额仍集中在头部 OEM,ModelBest 的上市路径也以合作伙伴为主结构性较小的模型供应商,拿到的价值可能少于掌握出货规模和 UX 入口的 OEMModelBest 收入在 OEM、品类和直接软件合同之间有多分散?

几行刻意写成双刃剑。政策支持、硬件进步、本地化压力让端侧 AI 更容易被采用,也会强化 OEM 和监管方对模型供应商的议价权。

[CM029, CM030, CM031, CM034, CM036, CM038]
FM004: 采用漏斗或价值链图

按指数展示端侧 AI 项目的商业化漏斗,从硬件就绪到规模化部署。

数值是相对指数点,以硬件就绪度 = 100 为基准,不是真实转化率;图中只展示公开证据暗示摩擦会在哪些环节累积:内存成本、集成、合规和 OEM 议价能力。

[CM007, CM016, CM031, CM034, CM039, CM040]

2.5 硬约束:内存成本、价格战、出口管制和缺失的份额数据

市场方向有吸引力,但经济账仍难。内存成本上涨后,Counterpoint 对中国智能手机的展望恶化:其 2025 年 12 月观点还预计 2026 年下滑 5%,到 2026 年 4 月更新时已恶化至 9%,OEM 已开始涨价,并更多依赖高端 AI 功能刺激换机需求。API 层也承压。KrASIA 和 DigWatch 显示,中国模型价格战中,Alibaba 将 Qwen-Long 定价下调 97%,Baidu 向部分企业用户免费开放模型,ByteDance 早已重置价格地板。出口管制是另一道约束:CNBC 报道,美国在 2026 年 5 月堵上一个漏洞,此前总部在中国的公司可在海外无需许可证获得先进芯片;CSIS 则认为同一批管制也会加速中国替代。最后,本次审阅的公开来源没有一个披露 ModelBest 在 AI 手机、AI PC 或智能座舱中可归因的付费份额,所以本章能展示需求扩张,却还不能证明这些价值有多少会落到 ModelBest,而不是它的 OEM 伙伴手里。[CM011, CM013, CM040, CM041, CM042, CM043]

2.6 图表

Chapter 03

03竞争格局

3.1 竞争版图:直接端侧模型同业、平台巨头与工作流替代品

ModelBest 同时在不止一个市场竞争。最窄的一层,MiniCPM 对打的是同样承诺本地推理、移动部署或更优单位参数智能水平的小模型与效率型模型家族:Microsoft Phi、Google Gemma、Meta Llama 3.2、Mistral 的 Ministral 系列,以及 Alibaba 较小的 Qwen 变体,都属于这一类。OEM、应用开发者或企业团队决定把哪种开放或半开放基础模型跑在用户身边,而不是跑在超大规模云里时,这些是最干净的“同一任务”比较。更宽的一层,只要买方更在意代码、长上下文、工具使用或打包生态服务,而不是纯粹的端侧极简,ModelBest 也会和 Moonshot Kimi、MiniMax、Z.ai / GLM、Baidu ERNIE 等更大的中国智能体模型厂商竞争。 替代品集合还要更宽。Apple Intelligence 不是作为开放模型出售,但它是强力嵌入式替代,因为 Apple 让任何应用都能离线、免费调用端侧 Apple Intelligence 功能。OpenAI 的 ChatGPT 是另一种替代:它云优先而不是端侧优先,但开发者和用户仍可能直接采用默认助手,而不是接一套本地栈。内部自建也是现实选项,因为许多竞争对手提供开放权重、标准 transformer 接口或 OpenAI 兼容 API,让买方更容易组装私有栈,而不是押注单一供应商。[CP001, CP002, CP006, CP009, CP012, CP016]

竞争对手画像表
竞争对手类别规模 / 包装信号目标细分市场差异化局限
ModelBest / MiniCPM直接边缘模型初创公司核心文本栈采用 Apache 许可;多模态边缘模型;OEM 和受监管行业合作中国 OEM、应用开发者、受监管企业、边缘设备厂商跨手机、PC、座舱、机器人、可穿戴的多模态效率分发靠伙伴驱动,不是消费者驱动;搭载率经济性仍未披露
Microsoft Phi全球直接 SLM 竞争者MIT 许可的 SLM 家族;Azure Foundry 分发;端侧定位需要低成本、低延迟多语言推理的企业开发者端侧叙事清晰,企业工具和安全定位强分发强度依托 Azure;不是中国原生 OEM 切入口
Google Gemma全球直接 SLM 竞争者开放权重;披露移动端内存占用;DeepMind / Google 开发者生态移动、浏览器、IoT 和本地应用开发者小模型打包和移动端优化文档充分Google 生态触达强,但设备端商业集成要看开发者执行
Apple Intelligence平台替代者 / 既有玩家OS 级端侧模型;Foundation Models 框架可离线使用,按请求不收费Apple 平台应用开发者和终端用户默认分发到 iPhone、iPad 和 Mac,并带隐私品牌背书生态封闭;不能作为通用跨平台开放模型使用
Meta Llama 3.2全球直接 SLM 竞争者1B/3B 家族,128K 上下文,社区分发规模巨大重视生态、社区工具和量化本地使用的开发者开源心智很强,智能体 / 检索用例丰富自定义社区许可不如 MIT / Apache 同行宽松
Mistral / Ministral 3全球直接 SLM 竞争者Apache 2.0 小型密集模型家族,覆盖 3B/8B/14B,并主打自托管 / 本地部署叙事边缘、企业和主权部署许可宽松,自托管叙事强面向中国 OEM 渠道的优化不如 ModelBest 明确
Alibaba Qwen中国平台既有玩家兼直接模型对手开放权重密集模型下探到 0.6B;8B+ 支持 128K;Alibaba 生态工具需要多语言或智能体模型的中国及全球开发者模型梯队宽,本地工具和应用生态触达强开放代码并不能消除对 Alibaba 控制产品入口的依赖
Moonshot Kimi K2.6中国工作流替代者开源编程 / 智能体模型;免费用户档加付费计划;256K API优先看重编程、长周期执行和智能体集群的开发者智能体工作流卖点强,并提供 OpenAI 兼容 API主要优化方向不是手机级推理或 OEM 嵌入
MiniMax M3中国工作流替代者1M 上下文多模态模型,宣称前沿编程 / 智能体能力;按 token 计划包装需要长上下文本地或托管智能体工作流的开发者把长上下文、多模态和编程能力打成一个包公开包装对独立定价和移动端占用披露较少
Z.ai / GLM-5.1 模型中国工作流替代者智能体 / 聊天品牌,加上长周期编程文档和 OpenAI 式 API需要长时间自主执行的企业或开发者主打 8 小时任务,编程导向强小型、移动优先占用的定位不如 MiniCPM 或 Phi 清晰
Baidu ERNIE 5.1中国既有玩家替代者国内旗舰模型,强调成本效率,并有既有搜索 / 企业基础设施中国企业、搜索和国产基础设施买方性价比叠加国内生态可信度开放权重可迁移性弱于 ModelBest、Phi、Gemma 或 Mistral
OpenAI / ChatGPT现状云端替代方案云优先对话助手,默认心智很强愿意用云端助手而非本地模型的开发者或用户不需要离线、主权或逐设备推理时,它是最快默认路径既不是端侧开放模型,也不是中国原生 OEM 方案

表中把直接边缘模型同行和更宽的工作流、平台替代者放在一起,因为 OEM 和企业买方可以用不止一种方式完成同一任务;未知项保持明示,不用厂商营销材料倒填。

[CP001, CP004, CP006, CP007, CP010, CP012]
FP001: 竞争定位图

按开放 / 边缘可达性和默认分发能力,对选定竞争者做序数定位。

坐标轴依据已检索的许可、部署和分发信号,综合出有证据支撑的序数评分;并非标准化基准指数。

[CP006, CP010, CP012, CP016, CP017, CP019]

3.2 能力、开放度与打包:ModelBest 很强,但并不孤单

ModelBest 的正面案例成立。公司把 MiniCPM 定位到手机、AIPC、座舱、机器人和可穿戴设备;MiniCPM 仓库把 MiniCPM5-1B 描述为 1B 级端侧 SOTA 模型;MiniCPM4.1 强调 128K 上下文下的稀疏注意力提速;MiniCPM-V 系列则声称每参数多模态质量异常强,并支持 iOS、Android 和 HarmonyOS 移动部署。相比只能跑在云端的代码模型,这更适合 OEM 和中国端侧设备建设者。 但竞争集合已经向同一个设计中心收敛。Microsoft 明确把 Phi 营销为可在无云连接情况下端侧使用的开源 SLM,Phi-4-mini 有 3.8B 参数、128K 上下文和函数调用。Google 现在提供 Gemma 4 E2B / E4B,面向移动和 IoT 部署,并公布移动内存占用。Meta 的 Llama 3.2 有 1B 和 3B 尺寸,支持量化端侧用例,但采用自定义社区许可,而不是 MIT 或 Apache。Mistral 发布了 3B、8B 和 14B 尺寸的 Ministral 3 模型,按 Apache 2.0 用于端侧和本地部署。Qwen3 将稠密开放权重模型下探到 0.6B;Moonshot、MiniMax、GLM 和 ERNIE 也越来越围绕智能体编码、多模态或长程执行优化。含义很直接:ModelBest 并不独占“高效开放模型”;它必须在跨平台部署、多模态效率和中国本土商业集成等执行细节上赢。[CP003, CP004, CP005, CP007, CP010, CP011]

功能 / 能力矩阵
能力ModelBestPhi-4-miniGemma 4 smallApple IntelligenceLlama 3.2Ministral 3Qwen3 smallKimi K2.6MiniMax M3GLM-5.1ERNIE 5.1
开放权重 / 代码访问强(Apache-2.0 核心仓库)强(MIT)强(开放权重)弱(仅封闭框架访问)中(开放访问,自定义许可)强(Apache 2.0)强(密集 Qwen3 采用 Apache 2.0)强(开源)强 / 仍在演进(官方开源推出中)中(文档 + 平台,包装不够宽松)弱至中(官方平台优先)
手机 / 设备优先定位Apple 设备内最强弱至中弱至中弱至中
检索页面显示的原生多模态中(家族包含多模态)小文本模型线较弱检索文档显示较低检索发布显示较低
长上下文 / 长周期智能体执行中至强强(128K)强(小模型支持 128K)强(128K)强(256K)最强(1M)强(主打 8 小时任务)
跨平台本地部署工具中至强仅 Apple社区强
监管 / 信任叙事在中国 OEM / 法律细分场景强企业安全卖点强隐私 / PCC 强政策 / 社区防护中等自托管 / EU 托管强国产性价比强
默认分发能力中(通过伙伴)强(依托 Azure)强(依托 Google 渠道)很强(依托 Apple OS)强(依托开源生态)强(依托 Alibaba 生态)强(依托 Baidu 生态)

各单元格是基于证据的定性判断,综合官方产品、仓库和文档页面得出。“弱”或“中”表示检索到的来源不足以支撑更强表述,并不等于竞争对手一定缺少该能力。

[CP004, CP006, CP007, CP009, CP010, CP012]
定价 / 包装对比
提供方价格 / 计费单位打包方式包含内容未知项 / 含义
ModelBest / MiniCPM开放权重本地使用;公开页面没有给出统一 API 价格表Apache 许可仓库加 OEM / 伙伴部署文本与多模态端侧模型、框架支持、OEM / 渠道集成经济性取决于伙伴合同,而不只是公开模型访问
Microsoft Phi检索到的模型页面没有发布推理标价;模型可通过 Azure Foundry 和 MaaS 使用MIT 许可模型加 Azure 托管目录3.8B 小模型、128K 上下文、函数调用、多语言支持即使没有透明的公开端侧设备单价,企业买家侧的打包仍很强
Google Gemma开放权重;本地推理经济性取决于硬件占用,而不是 token 价格通过开发者生态和移动优化变体发布开放权重2B/4B 小型移动模型、量化移动路径、多模态自托管用户切换成本低,但设备总成本取决于硬件约束
Apple Intelligence应用开发者使用端侧模型时没有按请求收费OS 原生框架和 API离线端侧推理,加上升级调用 Private Cloud Compute在 Apple 自有生态内,第三方很难在边际成本上打赢
Meta Llama 3.2模型访问需接受许可;本地成本取决于所选量化方案和硬件开放访问下载,附自定义社区许可1B/3B 模型、128K 上下文、量化端侧用例尽管开发者访问面广,商业打包不如 MIT / Apache 同行顺滑
Mistral / Ministral 3API 定价存在,但检索到的页面更强调开放权重和部署选项,而不是醒目标价Apache 2.0 权重加 Mistral 平台 / 云伙伴3B/8B/14B 端侧模型、自托管、企业支持主要拼部署可迁移性,而不是简单的按 token 价格表
Alibaba Qwen开放权重 Qwen3 稠密模型;托管访问可通过 Alibaba 工具路由Apache 2.0 小型稠密模型加云 / 应用生态0.6B-32B 稠密范围、本地框架、智能体支持打包优势来自 Alibaba 分发,而不只是原始模型价格
Moonshot Kimi K2.6用户端免费访问并提供付费计划;API 文档指向单独定价文档开源模型加 OpenAI 兼容 API256K 上下文、多模态、工具使用、编码 / 智能体工作流开发者打包不错,但检索到的页面只给出部分价格透明度
MiniMax M3Token Plan 用户价格不变,同时获得更强性能模型页面、Token Plan 打包、开源 / 本地路线图1M 上下文、多模态、编码 / 智能体流程打包有吸引力,但公开单位经济性仍不如 Apple 或纯开放权重模型明确
Z.ai / GLM-5.1 模型公开文档强调 API 用法,而不是简单标价智能体 / 聊天品牌,加面向编码的 API 文档推理模式、65,536 最大输出 token、长周期编码即使价格透明度有限,类 OpenAI 接口也会降低迁移成本
Baidu ERNIE 5.1检索到的发布内容强调 6% 相对训练成本,而不是公布客户价格国内平台加旗舰模型发布智能体 RL、成本效率、榜单定位Baidu 更靠捆绑生态经济性竞争,而不是公开开放权重打包

本表比较商业打包经济性、开放程度和买方摩擦,而不是假装每个竞争者都披露同一套按 token 标价。Apple 的“按请求零成本”和开放权重访问,是已检索来源集中最可比的公开经济性指标。

[CP004, CP006, CP009, CP012, CP014, CP018]
FP002: 功能广度 / 能力图

综合评分图,覆盖开放性、边缘效率、多模态、长上下文智能体能力和生态触达。

本图有意把异质证据压缩成定性档位;TP002 则保留各标准的具体措辞和限制说明。

[CP010, CP012, CP014, CP016, CP019, CP020]

3.3 分发、信任姿态和切换成本更偏向既有巨头,而不是 ModelBest

分发力让这个领域变得不对称。Apple 可通过 iPhone、iPad 和 Mac 装机量推送端侧模型;Microsoft 通过 Azure Foundry 和更宽的企业栈分发 Phi;Meta 受益于庞大开源社区和数亿下载;中国消费互联网巨头则控制巨大的 AI 应用表面。AICPB 2026 年 4 月排名显示分发有多集中:Doubao 以 336.04 million MAU 领先,Qwen 为 220.35 million,Quark 为 162.30 million,DeepSeek 为 138.98 million,Kimi 为 25.33 million。ModelBest 未进入前 50 名,因此它走向重要性的路径是伙伴嵌入,而不是终端用户拉动。 信任姿态更混合。ModelBest 有可信的本地信号——法院网络部署、北京重点实验室、ASPICE L2 汽车过程认证——这些对受监管买方和 OEM 买方重要。但竞争对手也以不同形式提供信任:Apple 强调隐私和 Private Cloud Compute;Mistral 强调自托管和欧盟托管选项;Microsoft 称 Phi 已经过安全后训练,可用于生产;Baidu 则把 ERNIE 包装成已整合进国内基础设施的高性价比旗舰。模型层切换成本看起来偏低,因为 Kimi 和 GLM 暴露 OpenAI 兼容 API,MiniCPM、Gemma、Mistral、Llama 和 Qwen 都支持标准开源推理栈。因此,锁定效应如果存在,更可能来自 OEM 渠道、操作系统、应用分发或企业采购,而不是模型 API 或权重格式本身。[CP008, CP013, CP015, CP027, CP028, CP029]

护城河持续性 / 竞争风险清单
护城河主张竞争威胁严重性证据缓释 / 尽调要求
跨平台多模态端侧效率Phi、Gemma、Ministral 和 Qwen 现在都在主打高效开放模型;Apple 在自有 OS 内提供免费端侧 APIMiniCPM-V 4.6 移动部署主张,对照 Gemma/Qwen/Ministral,加上 Apple Foundation Models 零成本打包要求提供前五大端侧竞争对手的归一化设备基准、时延和功耗
开源战略和开发者好感开源赛道已很拥挤:Phi(MIT)、Mistral(Apache)、Gemma(开放权重)、Qwen(Apache)、Kimi 和 MiniMax 都在争同一批心智各同行的公开许可和仓库证据要求 ModelBest 按模型家族提供真实下游采用、stars、下载量和商业转化
中国 OEM 和受监管行业准入伙伴主导的分发可以很强,但也会把议价权集中到 Huawei、Geely、Baidu Cloud 等更大交易方手中中高ModelBest 伙伴名单很长,但经济性未披露要求提供合同期限、续约机制、排他性、附加率和按伙伴划分的收入集中度
信任与合规叙事Apple 隐私、Microsoft 企业安全、Mistral 自托管和 Baidu 国内基础设施,都提供了替代信任锚信任叙事不同,削弱 ModelBest 的独特性询问客户为何选择 MiniCPM 而非这些其他信任模型,以及该选择是否粘性足够
API / 模型层切换成本低OpenAI 兼容 API 和标准开源推理栈让多栖部署很容易Kimi 和 GLM API 兼容性,加上 MiniCPM、Gemma、Llama、Mistral 共用框架用代表性应用测试 MiniCPM、Kimi、GLM 和 Gemma 之间的迁移时间
消费者或开发者默认入口AICPB 排名显示,中国头部 AI 应用以及 Apple / Meta 平台的默认触达远大于 ModelBestModelBest 缺席前 50 应用排名;在 MAU 或生态分发上,现有巨头占优要求提供直接用户遥测,并证明伙伴嵌入带来的是经常性使用,而不是被动安装

严重性衡量的是 ModelBest 竞争耐久度的风险,而不是整体端侧 AI 需求的风险。关键区别在于:技术很强的产品家族,是否真的变成商业粘性护城河。

[CP029, CP030, CP031, CP037, CP038, CP039]
FP003: 护城河 / 就绪度 KPI

精简评分卡,概括最能帮助或伤害 ModelBest 竞争耐久性的公开信号。

数值概括对引用证据的判断,并非管理层提供的 KPI。

[CP031, CP037, CP040, CP041, CP042, CP043]

3.4 护城河耐久性:开源有帮助,但前提是 ModelBest 把效率转成粘性渠道

反向证据很直接。开源可得性现在很常见,不再稀缺。Phi 使用 MIT 许可,Mistral 新小模型使用 Apache 2.0,Gemma 是开放权重并允许负责任商业使用,Qwen3 稠密模型是 Apache 2.0,Kimi K2.6 和 MiniMax M3 强调开放权重或即将开放权重分发,Llama 3.2 虽许可更受限,却拥有庞大社区。同时,Apple 的 Foundation Models 框架让任何 iOS 开发者可按每次请求免费使用端侧 AI,中国消费端领先者也已经掌握远大于 ModelBest 的应用分发。上述事实削弱了任何“MiniCPM 只靠开放性或基础端侧效率就能自我防御”的说法。 ModelBest 仍有可成立的护城河,但比营销叙事更窄、条件更多。MiniCPM 最好的证据是跨平台多模态效率:MiniCPM-V 4.6 声称在微小规模下,相比 Gemma、Qwen 和 Ministral 有更好的视觉语言效率;公司在中国也有真实的 OEM、云和受监管行业关系。买方如果想要本地部署、中国生态适配,并且不想被单一外国云或 OS 供应商绑定,这个组合可能重要。缺失的证明是商业粘性:搭载率、续约条款、实际收入,以及 Geely、Huawei、Baidu Cloud 等伙伴究竟把 ModelBest 视作核心基础设施,还是众多可替换模型中的一个。证据出现以前,MiniCPM 作为强产品家族看起来有耐久性;作为独立护城河,耐久性最多中等。[CP033, CP034, CP041, CP042, CP043, CP044]

3.5 图表

Chapter 04

04财务情况

4.1 收入模型和货币化:部署可见,标价不可见

ModelBest 的公开表面不像传统自助式 SaaS 漏斗。官网、Feishu 文档、GitHub 仓库和原始 README 都强调端侧部署、本地推理、多模态演示和集成工具链,而不是公开 API 下单流程或 token 价格表。这把最可能的货币化组合推向:手机、汽车、AIPC、可穿戴设备和机器人里的 OEM 嵌入;法律和工业场景中的企业与本地部署;以及通过集成、支持和定制把开源转成企业收入。同一批证据也解释了为什么投资核验困难:公司看起来商业上真实,但公开材料没有标价到实际成交价的桥,没有公开合同排期,也没有披露 OEM 经济性到底是授权费、分成、打包支持,还是战略性交叉补贴。开源分发显然帮助采用,但也意味着收入质量取决于合同结构,而不是公开用量定价。[CI001, CI002, CI004, CI014, CI015, CI016]

收入流表
收入流机制单位当前数值 / 状态质量尽调要求
OEM / 设备嵌入与手机、汽车、AIPC、可穿戴设备、机器人绑定的授权、集成或收入分成按设备 / 合同公开证据显示设备侧聚焦很广,并有报道称已在汽车 / 手机落地,但没有公开收费表渠道存在的置信度中,经济性的置信度低提供头部 OEM 合同,包括版税基础、最低承诺和出货量假设。
企业 / 本地私有部署面向法律和其他企业工作流的项目、订阅或支持合同按部署 / 年度合同官网确认法院网络部署和私有场景定位;计费模型未披露用例存在的置信度中,变现细节置信度低提供合同模板、续约条款和支持人员配置假设。
工业 / 机器人合作通过工业自动化、机器人和物理世界 AI 伙伴关系商业化定制商业协议投资人评论把 ModelBest 直接连接到工业自动化和机器人用例战略契合度置信度中,已确认收入可见度低披露收入来自软件授权、联合解决方案、硬件附加,还是里程碑付款。
开源到企业转化免费 / 开放模型分发,把需求导向付费集成、调优、托管或合规工作支持 / 服务 / 企业许可强开源生态和部署指南暗示有转化潜力,但没有公开转化率漏斗存在置信度中,变现转化置信度低提供下载到销售管线的转化、付费账户数和支持服务附加率。
多模态演示 / 托管服务层围绕 MiniCPM-o 和相关多模态产品的 GPU 支撑演示或服务运营用量 / 项目 / 托管服务官方演示文档显示真实的托管基础设施要求,但没有公开价目表低到中披露托管多模态演示只是内部赋能,还是可计费服务线。

各行把有直接证据的商业接触面,与仍未量化的变现机制分开。本章审阅的公开来源均未披露实际定价、收入分成比例或分部收入结构。

[CI014, CI015, CI016, CI017, CI018, CI019]
定价 / 变现表
接触面价格 / 单位标价 vs 实际成交价折扣 / 未知项来源
公开 API / token 定价在已审阅的官方界面上未找到公开标价未知是没有公开 API,还是仅按询价定价官网、Feishu 文档、GitHub 仓库、原始 README
OEM 嵌入 / 设备交易按询价 / 未披露很可能只有实际成交价未知费用按设备、按模型、收入分成,还是捆绑支持2026 年 4 月独立融资报道 + 官方设备定位
企业 / 本地部署按询价 / 未披露很可能只有实际成交价未知是经常性订阅、一次性部署,还是混合支持合同官方法律工作流定位 + 部署文档
工业 / 机器人协作按询价 / 未披露很可能只有实际成交价未知软件、解决方案集成或硬件附加收入哪一项占主导Eastmoney / QQ 投资逻辑报道 + Inovance 简介
开源模型访问免费 / 开放分发公开可见的是分发,不是公开变现免费用户转为付费企业工作的比例未知OpenBMB 主页、GitHub 组织、Hugging Face 组织
托管多模态演示 / 服务未披露无公开实际成交价未知托管产品是向客户计费,还是内部赋能资产MiniCPM-o 官方演示 README

本表有意区分公开可用性和变现。公开价格为空本身就是发现:ModelBest 展示技术访问和部署材料,却没有公开商业费率表。

[CI014, CI015, CI018, CI025, CI039, CI046]
FI001: 收入模型桥接图

ModelBest 的公开漏斗从开源和文档带来的采用,延伸到 OEM 嵌入、企业 / 本地部署和工业解决方案,但已确认收入仍藏在未披露的合同结构之后。

定性桥接图基于官方产品触点、部署文档和独立融资报道;不是公司披露的分部收入瀑布。

[CI015, CI016, CI017, CI018, CI019, CI030]

4.2 商业化路径和部署经济性:B2B2D 渠道可见,单位经济模型不可见

最强商业信号是渠道和部署信号。官方页面把 ModelBest 放在手机、AIPC、智能座舱、具身机器人、可穿戴设备和法律工作流周围;第三方 4 月融资报道称,MiniCPM 已经落地汽车、智能手机、AIPC 和智能家居,GitHub 与 Hugging Face 累计下载超过 24 million。文档栈走得更远:MiniCPM-V-Apps 设计成可在 iOS、Android 和 HarmonyOS 上完全端侧运行;MiniCPM-o 部署材料则仍假设为了更丰富的多模态体验,需要 GPU 支撑的 worker 基础设施。放在一起,这些信号指向一个 B2B2D 商业化路径:开源采用和开发者热情向 OEM 与企业管线输送线索,但最终货币化仍取决于集成和伙伴合同。投资人真正需要的部分仍缺席:客户数、合同期限、OEM 分成率、续约行为、CAC、回本周期,或软件价值与系统集成工作量之间的拆分。[CI003, CI004, CI005, CI006, CI007, CI008]

单位经济性表
指标数值 / null置信度重要性尽调要求
公开生态下载量24M+ 累计下载量(报道)显示分发触达和漏斗规模,不代表变现质量把下载量拆成活跃企业评估者、OEM 伙伴和非商业开源用户。
MiniCPM-o 部署每个工作进程的 VRAM~21.5 GB,初始化后说明即便是端侧导向的商业化,仍可能需要可观后端 GPU 资源按产品线提供生产架构,以及每位客户或每个 OEM 项目的混合 GPU-hours。
2026 年推理占 AI 计算的份额66%(Deloitte,经 QQ 报道)暗示未来成本压力会落在服务推理,而不只是训练提供管理层判断:ModelBest 计算预算中训练 vs 推理各占多少。
许多在线推理场景中观察到的训推芯片利用率5%-10%(报告基准范围)如果容量规划差,低利用率会压垮毛利按集群 / 工作负载提供利用率,以及端侧卸载比例。
毛利率检验开源端侧采用能否形成耐久软件经济性的核心指标提供产品线毛利率和主要 COGS 项,包括计算和支持。
CAC / 销售周期 / 回本期判断 OEM 和企业渠道能否高效扩张所必需按细分市场提供顶部漏斗转化、赢率、销售周期,以及按渠道的回本期。
NRR / 续约行为决定任何经常性企业收入基础的质量按产品 / 客户类型提供队列留存、增购和流失。

公开记录止步处,null 是有意保留。纳入报告的生态或基础设施数字仅作代理指标,不应被误读为公司级盈利能力指标。

[CI008, CI009, CI010, CI030, CI036, CI038]
FI002: 单位经济性桥接图

公开证据在分发和部署层最强,合同经济性较弱,毛利和留存层最弱。

桥接图有意停在公开披露停止的位置;下游节点仍是尽调问题,不是已验证指标。

[CI006, CI007, CI008, CI030, CI031, CI038]

4.3 资本可得性和充足性:融资势头强,现金能见度缺失

公开材料里最清楚的财务事实不是收入,而是融资可得性。多篇独立 2026 年 4 月报道称,ModelBest 完成由 Shenzhen Capital Group 和 Huichuan / Inovance 关联产业资本领投的数亿元人民币新一轮融资;此前据报道 2026 年 2 月 China Telecom 领投一轮,使第一季度累计融资超过 RMB1 billion,并把公司推入独角兽估值门槛。这些投资人的战略逻辑也异常明确:Shenzhen Capital 把 ModelBest 的智能密度路径与更低部署和调用成本联系起来,Huichuan / Inovance 把公司视为工业自动化和机器人领域的契合对象,China Telecom 则被描述为云-网-端分发盟友。这是一组真实资本栈,也可能降低近期融资脆弱性。但它仍没有回答投资核验问题。抓取来源没有提供现金余额、月度烧钱速度、现金跑道、债务或已承诺算力支出,也没有可访问的官方投资人公告或备案能超越媒体报道,把证据链闭合。[CI020, CI021, CI022, CI023, CI024, CI025]

资本充足性表
项目公开数值 / 状态置信度含义尽调要求
2026 年 4 月融资轮数亿元人民币;SZVC 与 Huichuan/Inovance 关联工业资本领投确认新外部资本进入,战略投资人有兴趣获取签署版轮次交割备忘录、股权结构表影响和募集资金用途计划。
2026 年 Q1 累计融资> RMB1.0B 披露下限暗示近期资本获取能力强确认准确总额、现金到账日期,以及受限 vs 非受限资金。
最新估值独角兽门槛 / ~$1.0B 下限(报道)支撑融资势头,但不代表收入质量要求提供投后估值、清算优先栈和任何业绩棘轮条款。
投资人组合电信 + 国资关联 VC + 工业自动化资本降低单一投资人依赖,并可能带来渠道支持提供董事会观察员权利、商业承诺和任何投资人关联采购目标。
账面现金缺少当前现金,无法估算跑道提供月末现金、短期投资和受限现金。
月度烧钱 / 跑道判断下一轮融资依赖的关键缺失变量按 R&D、计算、销售和 G&A 提供月度现金消耗,以及基准 / 乐观 / 悲观跑道。
债务 / 项目融资义务未找到公开披露未知计算或硬件扩张是否有债务支持披露租赁、供应商融资、云承诺和任何保底采购义务。

公开证据证明的是融资势头,不是资本充足性。本表区分已披露融资事实,以及承销所需但仍属私有的现金和跑道数据。

[CI020, CI021, CI022, CI023, CI024, CI025]
FI003: 财务估算区间

公开区间主要由融资和成本底线指标支撑,而不是收入或利润率指标。

各行单位不同,并写在 displayValue 和 detail 中。已披露融资和估值条目是下限 / 门槛,不是完整审计总额。

[CI008, CI009, CI010, CI022, CI023, CI030]

4.4 成本结构和定价压力:端侧推理有帮助,但算力和开源仍压缩利润率

ModelBest 的端侧优先策略在经济上直观:如果更多推理跑在用户设备或 OEM 控制的硬件上,一部分集中式推理服务成本就能从公司损益表中移走。原始文档和应用仓库支撑这个逻辑,投资人评论也明确把更低部署和调用成本列为吸引力之一。问题是,公开记录同样显示,这还不足以推导出强利润率。MiniCPM-o 官方演示架构仍假设 gateway、worker、backend、挂载权重、专用 GPU,以及每个初始化 worker 约 21.5 GB VRAM。更宽的中国 AI 基础设施报道称,推理已成为主导商业工作负载,许多通用训推芯片在在线推理中利用率很差,成本控制仍是核心瓶颈。再叠加 ModelBest 的完全开放分发,利润率风险就很清楚:采用可以快速放大,但除非 OEM 和企业合同足够自律,价格竞争和支持成本仍可能跑在货币化前面。[CI007, CI008, CI009, CI010, CI011, CI013]

FI004: 资本强度 / 现金流图

各渠道收入可见度不一,但成本可见度都弱,外部融资仍在为模式提供资金支撑。

矩阵标签只是按档位总结证据可见度,不是公司内部指标。

[CI025, CI026, CI027, CI037, CI038, CI040]

4.5 财务结论和尽调阻碍:故事有融资能力,但还不能被严肃投资核验

ModelBest 的正面案例比收入模型更容易辩护。公司有可见的产品-市场方向、活跃开源生态、可信行业部署,以及战略激励与端侧 AI 采用相匹配的投资人。这个组合让 ModelBest 继续融资、把技术声誉转成嵌入式和本地部署渠道,变得可信。负面案例也同样具体。公开证据仍缺收入、ARR、毛利率、烧钱速度、现金跑道、实际成交价、客户集中度和合同结构;就连备案/登记核验也不完整,因为可访问的官方门户在本次审阅中没有露出干净的公司特定记录。换句话说,市场已经为战略叙事买单,但外部投资人仍没法把损益表算清楚。在管理层提供合同级定价、客户集中度、算力成本证据,以及经审计或至少董事会级财务报表以前,正确的财务立场应保持谨慎:商业可信、战略融资充足,但披露仍轻,利润率仍不确定。[CI020, CI023, CI036, CI037, CI038, CI039]

公开财务缺口表
缺失的私有指标对判断的影响精确尽调路径
收入 / ARR / 分部结构无法检验开源热度是否转化为经常性付费收入要求按 OEM、企业 / 本地部署、支持 / 服务及任何托管用量线提供月度收入桥表。
标价 vs 实际成交价和折扣政策无法把部署可见度翻译成变现质量要求提供前 10 大合同摘要、折扣区间和 OEM 收入分成公式。
毛利率和计算 COGS无法判断端侧经济性是否真的更优要求提供计算账单、模型服务架构和产品线毛利率。
现金余额、烧钱、跑道和债务无法评估融资依赖和下一轮时间点要求提供最新管理账、现金瀑布和债务 / 云承诺明细。
客户集中度和出货暴露无法判断收入有多少依赖少数 OEM 发布要求提供头部客户集中度、按 OEM 划分的上线时间,以及积压订单 / 已签管线。
续约、留存和支持负担无法区分可扩展软件收入和重人力项目工作要求提供续约队列、按客户划分支持人数和 SLA 盈利能力。

截至 2026-07-02,这些是阻碍仅凭公开信息承销 ModelBest 的最低缺失项。

[CI036, CI037, CI038, CI039, CI046, CI047]

4.6 图表

Chapter 05

05产品与技术

5.1 产品线:一个端侧模型产品族,覆盖文本、多模态、全模态语音和打包移动应用

ModelBest 交付的产品不是单一托管模型端点,而是一套分层 MiniCPM 产品族,能映射到具体终端用户任务。公司官网把这个家族放在手机、AI PC、智能座舱、具身机器人、可穿戴设备、GUI 智能体体验和法律工作流周围,这和泛聊天机器人定位有实质差异。基础文本线仍重要:公开 MiniCPM 历史从 MiniCPM-2B,到 2B-128k、MoE 和 1B 变体,再进入 MiniCPM3-4B 以及更新的 4.x 和 5-1B 系列。在这个文本骨干之上,ModelBest 叠加 MiniCPM-V 用于图像、视频、OCR 和文档解析;MiniCPM-o 用于实时语音和实时多模态流;MiniCPM-V-Apps 则用于完全本地的 iOS、Android 和 HarmonyOS 打包。这种宽度是本章关键产品事实:ModelBest 卖的是一套可复用的端侧部署栈,服务多个任务,而不只是一张基准测试表。限制在于,公开证据对公司能打包和演示什么更强,对这些表面已有多少进入规模化生产项目则弱得多。[CE001, CE003, CE004, CE005, CE009, CE013]

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
MiniCPM 文本家族 (2B, 2.4B, 3-4B, 4.x, 5-1B)设备端、本地智能体和推理开发者开放发布家族;当前旗舰文本线是 MiniCPM5-1B小模型效率,加混合推理和工具使用聚焦各版本商业部署数量未披露
MiniCPM-V 家族 (2.6, 4.5, 4.6)多模态应用和 OEM 团队持续发布并有文档端侧设备上的高 token 密度 OCR、视频和文档理解独立复现证据比厂商基准覆盖更薄
MiniCPM-o 家族 (2.6, 4.5)语音和全模态助手构建者开放且快速迭代实时语音对话加实时多模态流官方文档仍承认全模态语音和基础能力限制
MiniCPM-V-Apps移动 OEM 和本地应用开发者iOS、Android、HarmonyOS 上的可运行源码演示多模态聊天的完全离线打包应用采用、留存和企业支持条款未公开
GGUF / 量化封装使用 llama.cpp 或合作伙伴平台的本地部署者已覆盖主要版本降低硬件门槛,支持离线安装不同量化等级的质量取舍尚无统一基准表
VoxCPM2TTS 与语音设计开发者纳入 V-Apps 栈给端侧栈补上多语种 TTS 和声纹克隆克隆场景的滥用控制和审核机制未公开详述
法律 AI 专网工作流法院与法律机构公司称已落地于专业司法场景本地部署或专网定位贴合受监管工作流具名客户数量和续约证据未公开
GUI / 座舱 / 具身场景OEM 与智能体框架开发者已公开展示,但各模块成熟度不一把 MiniCPM 从聊天推向设备原生动作闭环展示面广,比独立验证的大规模生产证据更清楚

各行区分公开封装覆盖面与已验证的大规模生产落地。「状态 / 成熟度」反映截至 2026-07-02 官方公开渠道可见的信息,而非私有部署数量。

[CE001, CE004, CE005, CE009, CE013, CE016]
工作流 / 使用场景表
用户任务现有工作流ModelBest 方案可衡量收益限制
在手机或平板上运行私有多模态助手云端聊天机器人或服务器托管的视觉 APIV-Apps 中的 MiniCPM-V,或本地 llama.cpp 封装按版本不同,可在 6-8GB 级设备目标上本地运行老旧低内存设备仍可能换页或性能下降,V 2.6 尤其明显
在端侧解析密集文档、图像或视频OCR 加独立 VLM,或往返服务器MiniCPM-V 2.6 / 4.5V2.6 用 640 token 编码图像,V4.5 实现 96x 视频 token 压缩主要精度卖点多由厂商发布,而非中立基准
构建双语实时语音助手ASR + LLM + TTS 级联,或云端实时 APIMiniCPM-o 2.6 / 4.5一个集成语音与多模态的栈,支持可配置音色官方文档承认全模态语音仍可能不稳定或读错音
发布全离线多模态移动应用围绕多个运行时做定制应用集成MiniCPM-V-Apps 源代码库iOS、Android、HarmonyOS 参考应用降低集成摩擦仍需要各平台专用工具链和构建步骤
运行本地代码或工具助手小型本地文本模型或前沿 API 智能体MiniCPM5-1B1B 级端侧模型,具备工具使用和推理姿态当前公开证据偏开发者场景,没有企业 SLA 背书
支持座舱、机器人或 GUI 智能体工作流带嵌入式感知与动作能力的定制 OEM 栈MiniCPM 加官网展示的座舱、GUI 智能体与具身场景边缘原生延迟和设备集成叙事贴合这些场景公开证据以演示和合作方提及最强,大规模 KPI 较弱
在专网上部署受监管法律工作流隔离系统内的人工研究与文档辅助ModelBest 法律 AI 栈部署在法院专网让数据和推理更靠近受监管用户官方渠道未披露客户覆盖面、正常运行时间或审计控制

收益按可直接观察的部署或封装优势表述,不等同于已审计的 ROI 结果。

[CE001, CE004, CE010, CE012, CE014, CE017]

5.2 架构和部署:MiniCPM 靠压缩算力、激进打包并把推理搬到端侧取胜

技术差异化始终围绕效率,而不是打造最大的基础模型。最早的 MiniCPM 论文认为,1.2B 和 2.4B 模型可以凭更好的规模化纪律与大得多的系统竞争;当前仓库历史又加入 MiniCPM4.1、MiniCPM-SALA 和 BitCPM 等稀疏推理与量化分支。在多模态工作负载中,最具体的机制是视觉 token 压缩:MiniCPM-V 2.6 称,它只用 640 个视觉 token 就能处理 1.8M-pixel 图像;MiniCPM-V 4.5 又用统一 3D-Resampler,把 6 个 448x448 视频帧压缩为 64 个 token。部署对护城河故事同样重要。模型卡和 V-Apps README 共同显示,它支持 Transformers、llama.cpp、vLLM、SGLang、GGUF 打包和完全离线移动应用,公开 RAM 下限从 MiniCPM5-1B 约 4GB 到 MiniCPM-V 2.6 的 8GB 不等。因此,当买方想在严格时延、隐私或设备成本限制下获得像样的多模态能力时,这个产品最强。[CE006, CE007, CE008, CE010, CE011, CE015]

技术 / 运营架构表
层 / 组件作用依赖风险
MiniCPM 稠密文本骨干提供基础语言、代码和推理能力MiniCPM 训练配方,加兼容 Transformers 的封装与更大同类模型相比的能力主张,多来自公司或论文作者
稀疏推理 / 效率变体(MiniCPM4.1、SALA、BitCPM)拉高推理深度,并降低每个有效 token 的计算成本定制稀疏注意力、低比特量化和模型专属训练选择版本碎片化可能加重跨运行时维护负担
MiniCPM-V 视觉栈图像、OCR、文档与视频感知SigLIP 家族编码器,加 token 压缩策略感知质量和基准泛化仍取决于具体版本
MiniCPM-o 全模态栈语音、实时视频 / 音频流和主动交互Whisper 或语音模块、时序对齐和流式推理厂商公开承认基座与语音稳定性仍有短板
GGUF / 量化封装支持本地安装并降低内存占用兼容 llama.cpp 的权重和转换流程相比全精度的质量损失没有在统一基准矩阵中披露
运行时适配器在 Transformers、vLLM、SGLang、llama.cpp 和 Ollama 类路径中服务模型开源推理引擎和后端兼容工作部分路径依赖模型专属分支或功能,各模态成熟度不均
V-Apps 外壳与移动构建系统把权重和运行时打包成可安装应用Apple、Android、HarmonyOS 工具链,加 llama.cpp-omni 子模块想从源码构建的开发者仍要处理不低的平台配置门槛
分发平台让同一模型家族覆盖 GitHub、Hugging Face、ModelScope 和不同包格式第三方模型平台和合作伙伴镜像平台门禁、重 JS 页面或政策变化,可能在核心代码库之外增加摩擦

ModelBest 的公开产品故事同时依赖模型架构、封装和运营层,因此表格把三者放在一起。

[CE005, CE008, CE010, CE013, CE017, CE020]
FE001: 产品架构图

分层展示从用户工作流,到模型家族、效率机制和运行时封装。

[CE001, CE005, CE010, CE017, CE023, CE024]
FE002: 客户工作流 / 运营流程

开发者或 OEM 如何从目标用户任务推进到设备端或服务器上的 MiniCPM 运行时。

[CE011, CE015, CE023, CE024, CE026, CE028]
FE003: 关键依赖图

关键依赖决定 MiniCPM 是继续保持差异化边缘部署栈,还是沦为又一个开放权重家族。

[CE023, CE024, CE028, CE036, CE037, CE039]

5.3 分发和路线图:家族异常开放、迭代很快,但版本纪律很重要

开发者分发是最强的公开证据点之一。OpenBMB 的 GitHub 和 Hugging Face 表面显示,MiniCPM 被维护成一个宽发布家族,而不是一次性废弃的研究发布;仓库也有实质社区规模:访问日 MiniCPM 仓库超过 9.5k GitHub stars,MiniCPM-V 仓库超过 25k,V-Apps 仓库一直活跃到 2026 年 7 月。Hugging Face 组织页也显示了很深的目录,ModelScope 至少为 GGUF 打包增加了一条伙伴分发路径。路线图节奏同样真实。公开历史显示,MiniCPM-2B 在 2024 年初发布,2B-128k 和 MoE 后续在 2024 年 4 月,MiniCPM3-4B 在 2024 年 9 月,MiniCPM-o 2.6 在 2025 年 1 月,MiniCPM-V 4.5 在 2025 年,MiniCPM5-1B 加 MiniCPM-V 4.6 在 2026 年。这个速度有利于保持相关性,但也意味着部署兼容性往往取决于模型特定分支、GGUF 转换或运行时版本协调,而不是一个冻结的企业平台。[CE005, CE017, CE018, CE023, CE027, CE029]

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
2024-02MiniCPM-2B 发布历史发布早期就确立了该家族的小模型效率论点代码库 + arXiv
2024-04MiniCPM-2B-128k、MiniCPM-MoE-8x2B、MiniCPM-1B 发布历史发布显示团队早期就在试长上下文、MoE 和更小占用代码库
2024-08MiniCPM-V 2.6 发布,官方随后宣布支持 llama.cpp历史发布让多模态线开始适配本地 OCR 和视频工作负载HF 模型卡 + 代码库
2024-09MiniCPM3-4B 发布历史发布补齐中等规模文本档位,支撑更强本地推理代码库
2025-01MiniCPM-o 2.6 开源历史发布把模型家族从多模态视觉扩到实时语音和流式处理HF 模型卡 + 代码库
2025-08 至 2025-10MiniCPM-V 4.5,以及 Feishu 发布的 V4.0 / 4.1 更新已发布 / 已推广表明节奏已转向快速多模态与稀疏推理迭代,而非停在 2.x 家族HF 模型卡 + Feishu
2026-02MiniCPM-o 4.5,以及 Mac 或 GPU 本地实时演示路径当前代际推动全模态栈走向本地实时交互代码库 + arXiv
2026-05MiniCPM5-1B 发布当前代际刷新基础文本线,服务本地助手和代码智能体代码库
2026-05 至 2026-06MiniCPM-V 4.6,以及 API 和 Ollama 库分发通知当前代际表明 ModelBest 在 4.5 跳升后仍继续改进封装和伙伴分发MiniCPM-V 代码库页面

路线图表把历史里程碑和仍在使用的当前家族放在一起,因为公开采用取决于向后兼容和包可用性,不只看最新旗舰。

[CE005, CE009, CE013, CE016, CE023, CE027]
FE004: 产品成熟度 / 能力图

综合公开成熟度、部署、信任披露和社区牵引,横向比较 MiniCPM 主要层级。

[CE001, CE011, CE018, CE023, CE031, CE032]

5.4 信任、安全与合规:开放是加分项,但公开控制面披露仍薄

公开信任图景有真正正面因素。MiniCPM 和 MiniCPM-V 采用 Apache 2.0 许可,代码和部署示例开放,ModelBest 还称已经通过 ASPICE L2 汽车过程成熟度认证。相比没有公开工程表面的纯闭源模型厂商,这些信号实质更好。不过,公开控制面落后于能力面。MiniCPM-o 自己的文档承认,全双工全模态能力仍需改进,omni 模式下语音输出可能读错音;通用 vLLM 文档也表明,广泛多模态兼容仍取决于后端细节,而不是普遍开箱即用。更重要的是,本次审阅的官方表面没有公开安全信任中心、SOC 2 或 ISO 27001 认证页,也没有公开可用性和事件响应计划。这不证明控制弱,但确实意味着安全、内容治理和生产可靠性尽调必须从公开网页证据转入私密材料核验。[CE034, CE036, CE039, CE040, CE041, CE042]

信任 / 质量 / 合规表
控制 / 信号状态范围缺口
核心代码库采用 Apache 2.0 许可公开且已核验MiniCPM 与 MiniCPM-V 的开源权重或代码入口宽松许可本身不能证明企业支持或安全控制
开放代码与部署示例公开且覆盖广代码库、README、模型卡、V-Apps 构建路径代码覆盖面大于独立生产落地证据覆盖面
ASPICE L2 流程主张公司公开主张汽车工程流程成熟度叙事已审材料中未发现独立公开审计报告或客户案例细节
已知限制披露公开但不完整MiniCPM-o 承认基座和语音短板披露只针对具体模型,不是完整信任中心项目
平台访问与封装治理好坏参半GitHub 代码库开放,但部分模型平台页面有门禁或重 JS开发者要可复现权重时,分发摩擦可能正好出现
公开安全信任入口未找到SOC 2、ISO 27001、公开正常运行时间、事故历史、信任中心公开网页信息薄,需要私下尽调
语音与内容治理细节已审公开渠道未找到克隆、多模态流式处理和法律工作流公开的审核、水印或滥用响应控制仍不清楚

「未找到」指该控制在已审公开网页渠道不可见;不能证明其内部不存在。

[CE034, CE039, CE040, CE041, CE042, CE044]

5.5 图表

Chapter 06

06客户情况

6.1 客户分层:用户表面很宽,公开付款方披露很窄

ModelBest 的公开材料描述了很宽的用户地图,但披露的付款方地图窄得多。官方产品页把 MiniCPM 放进手机、AIPC 设备、智能座舱、具身机器人和可穿戴设备,法律板块则声称有法院专网工作流部署。这意味着本章至少有五类有意义用户:嵌入模型的 OEM 和设备制造商,购买私有或端侧部署的企业与公共部门部署方,把 MiniCPM 打包进本地系统的机器人和边缘集成商,自托管或适配开放模型的开发者,以及通过 OpenBMB 和清华关联开放发布互动的学术或研究用户。关键分析切分是买方、用户和付款方。终端用户可能通过汽车、手机或工具接触 MiniCPM,却不直接向 ModelBest 付费;可见付款方很可能是小得多的一组 OEM、电信和企业账户。因此,公开证据支撑强分层宽度,但不支撑广泛的账户级货币化披露。[CU001, CU002, CU003, CU004, CU025, CU026]

客户分群表
分群买方 / 用户 / 付款方主要使用场景公开规模信号收入 / 战略价值关键缺口
OEM / 设备厂商OEM 或终端厂商采购 / 集成,终端用户使用嵌入式 AI手机、PC、可穿戴设备、智能家居、汽车座舱官网列出手机、AIPC、座舱、可穿戴设备;Sina 提到 Geely、Changan、Volkswagen、Huawei设备集成可规模化分发,直接商业价值可能最高未公开授权费、ASP、合同期限或续约数据
企业 / 电信 / 法律部署方企业或电信客户付款;专业人员使用本地 / 私有部署专网法律工作流、电信边缘部署、行业 AI官方法院工作流主张,加具名 China Telecom 合作可能支撑更高价值的私有部署和服务合同未披露具名法院 / 客户单位、收入贡献或留存
机器人 / 边缘集成商集成商或平台伙伴付款;下游运营方使用模型具身机器人、边缘 AI 模块、智能座舱栈官网明确列出具身机器人和智能座舱高价值垂直解决方案的战略扩张渠道具名机器人集成商和合同范围未公开
开发者 / 开源社区多为自助用户;付款方占比未披露本地部署、微调、框架集成、演示应用24M+ 累计下载;HF 组织有 158 个模型和 11 个 Spaces;支持 Ollama / OpenVINO强获客漏斗入口和生态撬动未披露下载转付费账户转化率
学术 / 研究用户研究实验室或学术团队使用开放版本;通常无直接付款方或只间接付费基准测试、多模态研究、边缘模型实验OpenBMB + Tsinghua 合作和 arXiv 技术报告贡献战略信誉和反馈回路,而非即时收入未公开学术使用与商业使用拆分

分群较宽,但只有部分分群具备具名生产级证据。公开材料更清楚地标出用户入口,而不是付款方。

[CU001, CU003, CU023, CU026, CU032, CU046]
FU001: 客户旅程图

呈现 ModelBest 如何从广泛的社区发现走向少数具名商业证明;证据最密集的环节在开发者和边缘评估阶段。

[CU001, CU011, CU013, CU014, CU033, CU034]

6.2 具名证明存在,但主要来自渠道和伙伴公告

最清晰的具名商业证明由渠道带动,而不是传统 SaaS 式案例研究。China Telecom 领投了 2026 年一轮融资,CnTechPost 和 Gasgoo 都称,它计划借助云和算力网络与 ModelBest 深度业务协作。汽车方面,Gasgoo 和 Sina Finance 都称 ModelBest 的 MiniCPM 已集成到 Geely 和 Changan Mazda 的车辆中;Sina Finance 还点名 Volkswagen 和 Huawei 是深度合作伙伴,Gasgoo 则称 Geely Galaxy M9 搭载 MiniCPM 多模态模型,Changan Mazda EZ-60 是首款量产端侧大模型汽车。这些信号有意义,因为它们识别了真实分发渠道和产品表面,而不只是未标注标识。但它们仍低于成长投资人通常希望在客户章节看到的证据:没有合同金额,没有伙伴收入贡献,没有单位经济模型,也没有续约历史。因此,本章把具名部署视为接近生产的证明,而不是已披露经常性客户数据的替代品。[CU017, CU018, CU019, CU020, CU021, CU022]

客户增长 / 采用轨迹表
指标日期来源置信度含义缺失分母
MiniCPM 在主要平台累计下载量24M+2026-02CnTechPost;Gasgoo开源分发在全球拥有较大的漏斗入口覆盖未拆分试用、研究、生产或付费用途
OpenBMB Hugging Face 组织规模158 个模型 / 18 个集合 / 11 个 Spaces2026-06 快照Hugging Face 组织页显示持续的模型发布节奏和社区维护平台流量不等于客户数量
MiniCPM5-1B Hugging Face 下载量~322k2026-06 快照Hugging Face 组织页1B 端侧模型的单模型开发者需求强下载量不透露企业部署情况
MiniCPM-V-4.6 Hugging Face 下载量~802k2026-06 快照Hugging Face 组织页在截取快照中,这是当前多模态拉力最强的可见信号无重复使用、付费 API 或转化信息
MiniCPM-o 4.5 Hugging Face 下载量25万+ (>250k)2026-02 至 2026-07Sina Tech 文章新模型上线后仍持续吸引用户的新证据指标是下载量,不是活跃用户或客户
China Telecom 战略合作官宣2026-02-28CnTechPost / Gasgoo具名企业渠道,可扩展私有和边缘部署未披露合同规模、期限或在线客户数量
公开具名且包含 MiniCPM 细节的汽车项目2 个核心车型项目(Geely Galaxy M9;Changan Mazda EZ-60)2025-12 至 2026-02 公开报道Gasgoo / Sina Finance / Mazda与许多开源同类不同,存在真实生产式 OEM 证据公开具名的 OEM 项目数量很少

轨迹表把平台遥测、具名部署事件和合作公告放在一起,因为 ModelBest 不发布常规客户 KPI 看板。

[CU006, CU007, CU008, CU010, CU017, CU019]
具名客户验证表
客户 / 合作伙伴分群部署 / 使用场景生产 / 试点结果 / 证据质量限制
China Telecom企业 / 电信渠道面向行业部署的云、算力网络和边缘合作已官宣战略合作;商业深度未披露两家独立新闻源将 China Telecom 列为战略投资方和业务合作伙伴未披露合同金额、付费席位数或续约历史
Geely Galaxy M9汽车 OEM / 智能座舱MiniCPM 多模态模型用于旗舰六座 SUV 座舱体验生产式具名部署Gasgoo 和 Sina Finance 均点名模型和车型未披露单位经济、ModelBest 对应出货量或续约条款
Changan Mazda EZ-60汽车 OEM / 智能座舱与 Wutong / TINNOVE 开发、搭载边缘端模型的量产车生产式具名部署Gasgoo 和 Sina Finance 均称其为首款搭载边缘端模型的量产车;Mazda 确认该车型项目存在Mazda 未直接点名 ModelBest;商业范围仍未披露
Huawei设备 / 生态合作伙伴Sina Finance 融资报道提到深度合作;可能指向终端或生态商业化具名合作;具体部署场景未披露公开媒体将 Huawei 列为深度合作伙伴,证明其与设备厂商细分市场相关未公开具体 SKU、合同金额或实际客户指标
Volkswagen汽车 / 设备生态伙伴Sina Finance 融资报道点名深度合作具名合作;具体部署场景未披露让公开伙伴名单里又多了一个全球汽车品牌未公开产品细节、合同金额或续约数据
Intel OpenVINO 生态边缘集成商 / 开发者渠道MiniCPM-V-2 在 Intel 推理栈上的官方优化指南已可支持生产部署,但不能证明直接收入有助于证明第三方生态愿意为 MiniCPM 部署投入工程资源这是生态采用,不是 Intel 付费客户证明

该表有意混合潜在付费渠道和生态证明,因为公开证据稀疏。各行应解读为对采用或商业相关性的点名证明,而不是统一证明已披露的经常性收入。

[CU017, CU019, CU020, CU021, CU022, CU023]
FU002: 采用 / 部署漏斗

展示从广泛社区触达到少量具名商业项目,再到续约阶段仍未解的问题。

[CU006, CU014, CU017, CU019, CU020, CU021]

6.3 开发者、学术和端侧集成商采用,是最强公开证据集

ModelBest 最可见的用户基础是开源和端侧部署社区。OpenBMB 的 Hugging Face 组织显示出很大的活跃足迹:在抓取的 2026 年 6 月快照中,有 158 个模型、18 个合集、11 个 Spaces,MiniCPM5-1B 约 322k 下载,MiniCPM-V-4.6 约 802k 下载。GitHub 仓库也讲同一个故事:MiniCPM 为端侧和资源受限使用而构建;MiniCPM-V 记录了对 llama.cpp、vLLM、Ollama 和移动部署的支持;MiniCPM-V-Apps 提供 iOS、Android 和 HarmonyOS NEXT 的完全离线演示。Intel 的 OpenVINO 团队发布了 MiniCPM-V 优化指南,Ollama 托管官方 openbmb/minicpm5 条目,说明外部生态正在替 ModelBest 做真实适配支持。这在战略上重要,因为它降低采用摩擦,也扩大了全球开发者触达。它也制造了测量问题:这些信号证明兴趣、实验和技术兼容性,但仍未披露其中哪些用户会变成付费客户。[CU006, CU007, CU008, CU009, CU011, CU012]

开发者与学术采用场景
场景公开证明证明了什么未证明什么渠道含义
GitHub MiniCPM / MiniCPM-V 仓库端侧定位、部署指南、框架引用开发者能够发现、评估并在本地部署模型没有付费转化或企业账户数量GitHub 是主要漏斗入口渠道
Hugging Face OpenBMB 组织158 个模型、18 个集合、11 个空间、按模型统计的下载量活跃模型目录庞大,开发者兴趣可衡量下载总量不等于付费客户或留存部署Hugging Face 是主要分发和背书渠道
MiniCPM-V-Apps / HarmonyOS 演示通过 llama.cpp 运行的离线 iOS、Android 和 HarmonyOS NEXT 演示消费设备上存在真实应用级部署路径没有证据显示演示用户会转为企业客户移动端演示渠道扩大了试验场景
Intel OpenVINO官方 MiniCPM-V 优化指南和设备运行时文档第三方生态围绕 MiniCPM 部署投入工程资源没有证据证明 Intel 相关部署带来直接收入OpenVINO 可培育企业和边缘集成商采用
Ollama / 本地桌面场景官方 openbmb/minicpm5 条目和桌面宠物引用MiniCPM 进入热门本地 LLM 分发渠道看不到重复使用、付费支持或账户归属Ollama 扩大了在爱好者和准专业开发者中的触达
OpenBMB / arXiv / Tsinghua 研究渠道OpenBMB 网站更新节奏和 MiniCPM4 技术报告研究用户与学术可信度形成持续循环学术关注可能永远无法直接变现研究声誉增强上游认知,也带动下游招聘 / 贡献循环

这张额外表有意聚焦渠道,而不是收入。它帮助区分可见生态采用和不可见付费客户转化。

[CU007, CU011, CU013, CU014, CU015, CU016]

6.4 留存和重复使用:几乎没有公开的持久性披露

本章审阅的资料没有披露能把采用热度转化为持久客户信心的指标。围绕 China Telecom、Geely、Changan Mazda 或任何法律行业部署,公开资料都没有付费客户数、ARR、NRR 或 GRR、流失率、合同期限。公开可见的复用代理指标只能弱一些:累计下载量、反复发布模型、MiniCPM-V 4.6 的免费公开 API key、免费的 MiniCPM-o 4.5 演示和 API,以及降低试用门槛的社区基础设施。下载、模型发布、免费 API 和社区工具能说明注意力持续存在,却不能证明商业留存。本章最强的反向证据来自 MiniCPM-o 4.5 报道本身:早期半双工交互体验差,让不少用户失去耐心,也拖慢了多模态落地。ModelBest 的产品方向可能缓解这个摩擦,但截至目前公开资料仍缺少耐久性证据。[CU010, CU027, CU028, CU029, CU031, CU038]

留存 / 重复使用 / 满意度表
指标数值 / null细分市场置信度尽调问题
付费客户数量全部细分市场n/a — 未披露索取付费客户总数,并按 OEM、企业、开发者 / API 账户拆分
净收入留存率(NRR)企业 / OEMn/a — 未披露索取过去四个季度按主要商业细分市场拆分的 NRR
毛收入留存率 / 流失率企业 / OEMn/a — 未披露索取前十大账户或渠道的续约和流失历史
合同期限 / 续约结构China Telecom 与汽车 OEM 项目n/a — 未披露索取样本合同条款、续约日期和任何最低用量承诺
社区重复使用代理指标累计下载 24M+、MiniCPM-o 4.5 下载 250k+、公开免费 API / 演示开发者 / 开源将重复 API 用户、活跃企业与一次性下载区分开
采用摩擦信号公开来源称,半双工交互薄弱会让用户失去耐心,也可能拖慢多模态落地消费者 / 多模态用户衡量 MiniCPM-o 4.5 界面的演示重复使用、会话深度和 30 天留存

null 值是有意保留:本章审阅的公开来源没有披露正式留存或续约指标。最后两行只是较弱代理指标,不应替代队列数据。

[CU010, CU027, CU028, CU029, CU031, CU038]

6.5 扩张飞轮已经成形,但集中度和渠道依赖仍是核心风险

ModelBest 的扩张逻辑在公开资料里讲得通。开源发布先带来认知和技术信任;部署手册和框架支持把开发者推到本地评估;设备演示与更适合边缘的版本降低集成摩擦;随后,少数 OEM 或企业渠道才以具名生产案例露出。这个循环在手机、汽车座舱、私有化部署和机器人场景里可能跑通,尤其是端侧推理改善隐私和时延。但同一循环也把风险集中起来。公开具名的商业证明仍围绕 China Telecom 和少数汽车品牌,开发者发现则高度依赖 GitHub、Hugging Face、Ollama、OpenVINO 等第三方平台。法律 / 法院场景在战略上有吸引力,但客户未具名,运营细节不透明。换句话说,ModelBest 有宽的漏斗入口和可信的先落地再扩张路径,但公开可见的具名、变现、可续约账户基础仍然很薄。[CU032, CU033, CU034, CU035, CU036, CU037]

扩张与集中度风险表
扩张驱动因素集中度 / 依赖风险潜在影响尽调路径
开源发布 -> 本地评估 -> OEM / 企业集成循环转化依赖第三方渠道(GitHub、Hugging Face、Ollama、OpenVINO),公开资料未把它与直接销售漏斗相连ModelBest 可以拥有巨大的技术触达,却仍无法证明付费客户深度索取线索来源数据,以及社区用户转为付费部署的转化率
China Telecom 合作具名电信渠道具有战略重要性,但商业情况不透明如果 China Telecom 是主要企业分销路径,集中度可能高于公开材料呈现索取与 China Telecom 有关的已签约收入、销售管线归因和排他条款
汽车座舱部署公开点名的车辆证明集中在 Geely 和 Changan Mazda,更多品牌名称缺少部署细节少数汽车渠道流失或延迟,会明显削弱可见客户叙事索取逐 OEM 收入、上线时间表和续约状态
法律 / 法院私有化部署官方法律场景主张具有战略吸引力,但终端客户未具名未具名公共部门证明无法支撑强留存或采购假设索取具体法院、合同阶段、采购结构和部署后使用数据
框架与生态支持Intel、Ollama 和应用演示仓库的外部适配有助于采用,但也让 ModelBest 依赖外部维护者和平台如果直接变现薄弱,渠道权力可能留在生态方,而不是 ModelBest 手中索取各生态入口贡献的活跃用户比例,以及任何联合营销或收入分成条款

集中度风险来自狭窄的具名商业引用集,而不是已披露的收入集中度表;由于没有这类公开表,本表只能基于外部证据推断。

[CU017, CU030, CU033, CU035, CU036, CU037]
FU003: 客户证明矩阵

比较各渠道的公开证明质量:汽车领域的生产部署具体性最好,留存和收入可见性则全线偏弱。

定性评分总结的是公开证据质量,不代表内部销售状态;目的在于显示证明质量哪里最强、哪里最弱,而不是量化客户满意度。

[CU024, CU025, CU029, CU030, CU031, CU036]

6.6 图表

Chapter 07

07风险

7.1 中国监管和法律场景:备案、标识、网络安全和标准现在必须同时处理

ModelBest 的第一层风险在于,产品表面已经足够宽,会同时落入中国多个相互重叠的 AI 管控制度。公司官网把 MiniCPM 放在手机、AI PC、智能座舱、具身机器人、可穿戴设备和法院网络法律工作流中;MiniCPM-o 材料展示的是全双工语音和多模态交互,而不是简单文本模型。关键在这里:中国《生成式人工智能服务管理暂行办法》要求具有舆论属性或社会动员能力的服务完成安全评估和算法备案;2025 年 3 月的 AI 生成内容标识规则要求显式和隐式标识、在用户协议中披露标识方法,并在提供未标识输出时留存 6 个月日志。CAC 2026 年 5 月的备案公告说明规则并未沉睡:2026 年 3–4 月又有 72 项服务完成备案,已上线应用 还需要展示模型名称、备案号或上线编号。2026-01-01 生效的修订版《网络安全法》进一步抬高下行风险,把 AI 服务运营与更强的个人信息和关键信息基础设施执法挂钩;严重情形下,罚款可升至 RMB10 million,并可能触发停业整顿式措施。在这个背景下,本章审阅的公开页面没有露出 ModelBest 的备案号、公开标识声明或隐私政策链路,因此合规负担是具体、可跟踪的,不是理论风险。[CR001, CR003, CR007, CR011, CR012, CR013]

监管 / 法律风险登记表
规则 / 许可 / 案件司法辖区状态可能性严重性缓释措施剩余暴露尽调路径
《暂行办法》下的生成式 AI 备案与安全评估要求中国(CAC / MIIT)生效;符合条件的公共服务需要服务级备案 / 评估ModelBest 可通过记录每个公开服务界面是否纳入监管范围,并出具备案 / 评估材料来缓释本章审阅的公开来源未显示 ModelBest 专属备案号或评估状态索取 ModelBest 的 CAC 备案 / 注册编号、纳入范围的产品清单,以及按产品界面拆分的安全评估记录。
AI 生成合成内容标识规则中国(CAC / MIIT / MPS / NRTA)自 2025-09-01 起生效在文本 / 音频 / 图像 / 视频输出中落地显式与隐式标识、用户协议措辞和日志留存本章审阅的 ModelBest 公开界面没有显示可见标识声明或用户协议轨迹索取 MiniCPM-o 和面向公众服务的标识实现截图、元数据样本和用户协议文本。
修订后的《网络安全法》中国自 2026-01-01 起生效合规计划、数据治理控制,以及面向法律 / 座舱 / 消费者场景的行业专项处理严重情形下,处罚可升至 RMB10 million,并可暂停服务、关闭应用或吊销许可证索取任何 CIIO 相关服务的内部分类、PII 处理流程和合规负责人。
GB/T 45654-2025 生成式 AI 安全基线中国自 2025-11-01 起实施记录训练数据来源、违法内容过滤、个人信息处理法律依据和 IP 风险披露ModelBest 的开源和敏感行业用例带来的证明义务,超过当前公开记录能显示的范围索取最新版生成式 AI 安全自评包、数据来源控制,以及红队 / 抽样程序。
开源许可与 IP / 来源合规跨境 / 多司法辖区Apache 2.0 代码公开;下游数据、模型和输出权利仍需按事实判断将代码许可合规与模型 / 数据 / 输出权利治理拆开,并记录企业使用条款可分叉性和下游权利不清会压缩定价,并制造企业采购摩擦索取企业条款、训练数据来源政策,以及侵权 / 下架处理流程。
公开诉讼 / 执法记录中国 / 其他可触达公开渠道留存来源未确认重大公开案件;核验仍不完整持续监测案卷和监管机构没找到案件,并不证明法律记录干净在中国法院、监管机构和商业风险数据库重复检索公司名称;向管理层索取诉讼清单。

按严重性排序。覆盖不完整:本表优先列出最可能打破投资假设的规则和法律检查,而不是尝试逐司法辖区做完整法律清单。

[CR006, CR011, CR012, CR013, CR014, CR015]

7.2 开源、数据来源和安全风险:分发优势也会降低护城河、抬高合规负担

ModelBest 的第二大风险在于,同一套开源和部署优先策略一边加速采用,一边放大合规和变现暴露。OpenBMB 称社区由 Tsinghua NLP Lab 与 ModelBest 共同发起,MiniCPM 仓库可公开下载,代码采用 Apache 2.0 许可,README 明确强调端侧部署和全双工多模态用例。这样做扩大了生态触达,但也让技术栈更容易被分叉、跑基准测试和重新封装,削弱靠保密支撑的定价权,并把价值捕获推向合同、集成或专有数据权。新的 TC260 基线标准进一步加剧这个问题:它要求干净的数据来源、个人信息同意或其他合法基础、在用户协议中披露 IP 风险,并拒绝非法内容占比超过 5% 的训练来源。ModelBest 自有公开材料又声称已在法院网络法律场景部署,把个人信息、文档来源和输出可靠性等敏感要求一并纳入负担。可是本章抓取的公开页面没有暴露公开隐私政策、用户协议、安全认证或事件响应页面;数据处理、模型治理控制和公开安全姿态仍是尽调空白。[CR003, CR004, CR005, CR006, CR007, CR021]

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
尽管存在多模态和社会影响力用例,已审阅界面仍缺少公开标识 / 备案证据低;规则清晰,但留存来源没有公开公司级证据一次合规失误可能同时触发多个界面的下架、暂停或被迫改版需要逐产品映射备案范围、标识位置和元数据实现。
开源栈容易被分叉和基准测试,削弱基于保密的护城河中;ModelBest 可在集成、密度和伙伴执行上竞争,而不是靠保密商业价值可能流向 OEM 集成或云绑定,而不只留在模型许可需要 MiniCPM 部署的付费转化、企业绑定和合同续约证据。
训练数据、IP 和个人信息来源义务超过公开披露低到中;TC260 给出了蓝图,但 ModelBest 内部控制未公开如果来源或同意控制薄弱,敏感行业部署可能过不了尽调需要数据来源治理、同意 / 法律依据记录,以及 IP / 下架流程文件。
留存来源没有公开记录安全 / 事件响应姿态低;本次审阅未找到公开隐私政策、安全白皮书或事件页面如果看不到治理或事件响应姿态,公共部门买方可能放慢采购需要安全负责人、披露政策、测试节奏,以及事件日志 / 漏洞奖励姿态。
混合边缘云架构在部分任务上仍依赖云端重推理中;端侧定位降低但没有消除集中式算力需求如果高价值工作流的推理或编排仍重度依赖云端,单位经济性仍可能恶化需要端侧与云端工作负载拆分,以及按部署类型拆分的毛利率假设。

严重性和可能性是基于引用的技术、法律和新闻证据作出的序数判断。本登记表聚焦从公开证据中仍可见的运营风险,而不是所有可能私下存在的内部安全控制。

[CR003, CR005, CR006, CR007, CR018, CR021]

7.3 算力、伙伴和竞争风险:出口管制、国产芯片生态不成熟、OEM 集中和价格战相互强化

ModelBest 的第三组风险不在模型权重本身,而在算力获取、伙伴集中和商品化的市场结构。BIS 2026 年 5 月 31 日指引和 6 月 17 日 FAQ 明确,先进计算许可要求仍适用于总部位于 Country Group D:5 或最终母公司位于该组别的实体,即使采购主体设在中国境外;CNBC 报道称,这条规则针对通过 Malaysia 等地海外子公司流转的路径,出口商仍应申请许可。与此同时,Tencent 2026 年 6 月对 Li Dahai 的采访称,ModelBest 当前瓶颈是芯片、内存和带宽的共建,国产软件生态仍显著不如 Nvidia 完整。CNBC 和 CSIS 描述的中国解法,是用廉价电力和补贴支撑大规模 Huawei 芯片集群,但这条路比 Nvidia 系统消耗更多芯片和电力。下游公开部署记录也很集中:Gasgoo 和 Tencent 主要指向 Geely、Volkswagen、Changan、GAC 以及 Geely Galaxy M9 座舱项目等汽车和设备伙伴。这个集中风险发生在中国模型市场重新定价的同一时点。KrASIA 和 Digital Watch 显示 Alibaba、Baidu 把 API 价格压到接近零边际水平;消息源称,超大云厂商或许还能靠云服务绑定变现,但对小型创业公司明显更难。因此 ModelBest 面临复合风险:供应约束可能拖慢进展,而定价权和伙伴议价力正同时转向不利于独立实验室的方向。[CR024, CR025, CR026, CR027, CR028, CR029]

合作伙伴 / 依赖风险登记表
依赖项交易对手角色集中度失效场景严重性缓释措施剩余暴露
汽车设计定点与座舱部署Geely / Volkswagen / Changan / GAC 及相关集成商部署证明和收入渠道伙伴公开证据集中在少数具名汽车 / 设备关系上少数旗舰 OEM 项目流失、延迟或变现薄弱,会削弱边缘商业化叙事扩大 OEM 集合,并披露从设计定点到付费量产项目的转化当前公开证明更强在具名伙伴,而不是合同经济性或续约深度。
设备与端点生态采用公开材料点名的手机 / AI PC / 智能家居伙伴端侧 AI 分发的规模化路径公开说法显示覆盖面广,但几乎不披露单个伙伴的出货量或收入头部触达不等于持久收入,也可能被强势 OEM 重新定价谈下更黏的集成、支持和软件更新经济性在定价和续约证据出现前,剩余议价力仍在大型 OEM 手中。
国产芯片与软件生态Huawei 相关和其他国产算力栈训练 / 推理适配,以及抵御出口管制的韧性管理层说法显示技术依赖度高工具链薄弱或供给不足会拖慢模型迭代,并迫使产品妥协继续多供应商适配,并保持云 / 边缘工作负载拆分灵活国产替代降低单一路径风险,但不能抹掉性能和软件差距暴露。
战略资本基础China Telecom、Shenzhen Capital、Beijing AI funds、Huichuan 相关资本、其他国资 / 产业投资方资金续航和渠道入口资本结构很宽,但带有明确战略取向未来融资或渠道决策可能优先服务生态或政策目标,而不是纯粹股东经济性澄清投资者权利、治理权利和商业对价安排融资实力降低短期资金续航风险,同时抬高治理复杂度风险。
开源分发平台与社区外溢OpenBMB / GitHub / 下游开发者采用漏斗和声誉引擎设计上就是广泛且公开分叉版本或社区替代品吸收采用,却不向 ModelBest 付费通过集成、企业支持或专有部署诀窍变现开放分发仍是一把双刃剑:增长渠道加上护城河泄漏。

各行按严重性排序,捕捉可从公司外部监测的依赖。本登记表对可见交易对手把握更强,对私下合同保护把握较弱,因为这些保护没有在留存证据中公开。

[CR002, CR005, CR024, CR025, CR027, CR028]
FR003: 依赖关系图

ModelBest 处在监管、算力生态、开源渠道、OEM 和战略性资本的下游。

[CR024, CR028, CR030, CR032, CR038, CR041]

7.4 人员、治理和变现执行:学术背景和充足资本不能替代已披露经济性

ModelBest 的第四项风险,是在经济性不透明的情况下同时推进太多执行面。公司受益于清华系技术深度和异常强的资本联盟,但公开记录也显示,这种优势可能制造复杂性,而不是清晰度。QCC 列出的投资人阵容拥挤,涵盖电信、国资、产业和战略投资方,包括 China Telecom Investment、北京 AI 产业基金、Shenzhen Capital、Hubble 相关资本,甚至竞争对手 Zhipu。Caixin 2024 年画像已经把商业化列为关键问题,2025 年后续报道又称 ModelBest 一年多完成三轮融资,却仍未披露该轮确切金额或估值。InfoQ 随后报道,2026 年一季度融资超过 RMB1 billion,ModelBest 正从模型延伸到 Pinea Pi、EdgeClaw Box 等硬件产品。这证明公司有野心,但也意味着投资人要同时承销模型研发、芯片适配、产品化和硬件系统执行,而且没有公开收入、定价或客户集中度披露。由此产生的风险,不是 ModelBest 缺少资本或伙伴兴趣;而是公开证据仍无法说明这些关系能否产出持久毛利、续约行为或简单治理结构。[CR004, CR008, CR009, CR010, CR032, CR033]

人员 / 执行风险登记表
角色 / 职能依赖或缺口可能性严重性缓释措施尽调路径
创始人 / 学术领导梯队Tsinghua 出身的技术身份是优势,但也把可信度和路线图解释权集中在小型创始团队在创始科学家叙事之外,扩大已披露运营梯队和接班深度索取组织架构、接班计划,以及研究、产品、合规和销售的角色覆盖。
学术到商业转化尽管融资和部署公告不断,Caixin 仍把商业化表述为开放问题将技术里程碑绑定到付费合同,以及已披露的续约 / 毛利率指标索取已实现经济性和续约行为的付费 OEM / 企业项目队列。
股权结构和治理复杂度QCC 显示投资方名单庞大,覆盖电信、国资、产业和同业 AI 资本,包括 Zhipu简化董事会 / 观察员权利,并披露相关战略约束索取股权结构、投票 / 否决权、董事席位和信息权。
执行边界蔓延ModelBest 现在横跨模型、芯片适配、硬件产品、法律 AI、座舱和消费者多模态界面优先推进少数可变现产品线,并谨慎安排硬件节奏索取未来 12-18 个月按业务线拆分的产品优先级和资本配置计划。
披露纪律频繁融资头条仍没有带来公开定价、收入或估值透明度把投资判断从叙事驱动转向指标驱动索取董事会级收入、定价、客户集中度和资金续航材料。

本登记表强调公开证据尚无法关闭的执行和治理问题。严重性最高之处,是资本或技术实力仍可能无法转化为透明经济性或可管理治理。

[CR004, CR008, CR009, CR010, CR032, CR033]

7.5 缓释、可监控触发器和否决标准:下一轮尽调应由事件驱动

承销 ModelBest 风险,不能停留在泛泛的 AI 谨慎,而要盯住具体、外部可观察的触发器。部分缓释因素确实存在:公司有战略支持者、多个具名部署渠道、端侧优先的技术身份,以及能降低采用摩擦的开源生态。但这些缓释并不完整。战略资本不是收入质量证明,开源不是护城河证明,国产芯片适配也不等于对出口管制具备安全韧性。因此,本章把优先级最高的否决标准定义为公开事件:披露 ModelBest 的 CAC 备案或标识声明,出现任何出口管制指定或明确采购阻断,有更多证据显示价格下跌快于绑定价值增长,汽车 / 设备渠道出现伙伴集中裂缝,以及公开治理或法律披露澄清股权结构和敏感行业部署是否可管理。下一次刷新时,如果这些触发器没有改善,剩余暴露应视为打破投资逻辑,而不是普通创业公司不确定性。[CR011, CR017, CR018, CR024, CR025, CR028]

缓释措施与否决标准表
风险可监测触发因素阈值 / 事件行动含义
ModelBest 专属 CAC 合规可见度公开产品页、应用详情页或公司披露到下一次尽调更新,仍看不到可见模型名称 + 备案 / 上线编号或标签声明将其上调为合规阻断项;在承销消费者或具社会影响力的场景前,必须先拿到证据。
AI 标识执行观察到的产品输出或公司合规材料多模态服务已上线,但范围内输出没有明确标识、没有元数据证据,或用户协议没有相关语言将其视为中国消费者端部署准备度的论点破坏项。
先进算力采购韧性BIS / 公司采购更新 / 供应商通知出口许可被拒、无法采购所需先进计算物项,或明显依赖受阻路径按更慢的训练 / 推理服务进展,重设发布节奏和估值。
OEM 集中度和变现证明新客户披露或合同证据没有扩展到当前具名渠道之外,或证据显示旗舰项目仍是试点,未形成付费规模下调设备 / 座舱收入假设,只把重点放在有证据支撑的垂直场景。
价格战压力行业 API 降价公告和公司定价姿态市场继续普遍降价,而 ModelBest 没有披露附着价值或利润率保护把业务建模为集成 / 服务驱动,而不是模型授权驱动。
治理 / 股权结构复杂性董事会、融资或权利披露战略投资者权利实质限制商业灵活性,或制造关联方冲突要求治理条件,或把未来融资条款视为结构性更高风险。
安全 / 隐私治理公开政策页、事件披露或尽调室文件公司增加更多公开触点时,仍没有隐私政策、安全负责人或事件处理流程证据在治理控制得到证明前,冻结消费者端场景承销。

这些触发器刻意设计成可从公司外部监测,方便下一次更新由事件驱动。它们不是泛泛的创业公司警示;每个触发器都对应本章已有证据中的特定合规、变现、合作伙伴或算力失效模式。

[CR011, CR015, CR017, CR018, CR024, CR025]
FR001: 风险热力图

ModelBest 的剩余风险集中在中国合规可见性、与出口管制相关的算力依赖、OEM 商业化集中度,以及从学术走向商业的执行力。

单元格是基于保留公开证据的序数判断,不是来自已披露的内部风险模型。

[CR011, CR015, CR020, CR024, CR028, CR030]
FR002: 风险传导图

合规、算力、定价和治理风险最终都压到商业化质量和融资杠杆上。

[CR011, CR018, CR024, CR028, CR030, CR035]

7.6 图表

Chapter 08

08估值

8.1 融资背景和当前估值标记

ModelBest 当前最有支撑的估值锚点仍是底线,而不是完整定价的估值标记。多篇 2026 年 4 月报道称,公司完成 2026 年第二轮融资,一季度累计融资超过 RMB1 billion,并跨过大模型独角兽门槛。ThePaper 进一步把措辞收紧,称公司在 4 月融资后进入 "$1 billion 级别"。这些信号足以把约 $1 billion,或简写的 RMB7 billion-plus,视为可信的当前头部估值。 关键限制在于公开证据到此为止。同样验证独角兽门槛的 4 月报道,并未披露精确投后估值、完整稀释图景或投资人权利条款。CB Insights 公开公司页进一步强化不透明:它展示 2026 年 4 月 7 日融资和投资人名单,但把估值遮为 $XXM,并在公开页把收入列为 0 FY undefined。换句话说,市场已经方向性地透露 ModelBest 不再是小规模种子公司;但还没有给出足够信息,让人有信心承销确切证券价格。[CV001, CV002, CV003, CV004, CV005, CV006]

推荐结论汇总表
决策字段当前观点支撑证据决策含义
推荐继续研究公开证据能支持 $1B 级估值标记,但变现细节还不够,无法干净承销把公司留在漏斗中;缺少更多尽调前,不要按当前价格放行投资
置信度融资、牵引力和可比样本都是真实的,但准确估值、收入和条款仍不透明保持覆盖,而不是直接给出硬性放弃或买入结论
风险评级变现不透明、合同经济性不确定、资本市场敏感性仍然重要把任何交易都视为尽调强度高、且对条款清单敏感
估值立场偏高当前标记更多靠战略稀缺性和边缘 AI 期权价值,而不是公开财务证明要么拿到更好证据,要么要求更好价格
价格纪律没有审计收入和干净条款,就不在当前 $1B 级或更高估值投新钱公开收入或优先权层级都不足以支撑不计价格的买入结论只有数据室补齐主要缺口,才按当前价格重新评估
上调路径更好证据或更低入场价审计后的收入转化、OEM 抽成率和股权结构清晰度,比更多叙事动能更能快速改变结论在承诺投资前,跟踪下一轮融资、客户证明和尽调包

这张表刻意对价格敏感:它总结当前估值已经假设了什么,而不是泛泛评价公司质量。

[CV001, CV002, CV013, CV039, CV043, CV045]
FV001: 投资建议逻辑

当前建议来自一个真实的边缘 AI 切入点,但它撞上了商业化证明缺失和证券条款不完整。

该流程是定性而非概率模型;它映射的是保留公开证据所暗示的当前决策链。

[CV004, CV007, CV011, CV013, CV018, CV039]

8.2 战略价值真实存在,但变现证明仍是缺口

为什么一家没有公开收入披露的公司也能有 $1B 级估值标记?因为 ModelBest 确实有投资人看得见的资产。官方和面向开发者的来源显示,它拥有真实的端侧模型资产:公司把 MiniCPM 放在手机、AIPC、智能座舱、机器人、可穿戴设备和法律工作流里;OpenBMB 把公司与清华根基的开源生态绑定;GitHub 和 Hugging Face 则显示 OpenBMB 组织页上有大量活跃触点,包括 MiniCPM 仓库、演示、合集和 158 个托管模型。媒体报道也持续引用累计下载超过 24 million 和具名汽车部署。 但这些都不是变现证明。24 million 这个数字来自公司经媒体转述,而不是经审计遥测数据。审阅的公开资料没有披露 ModelBest 收入、ARR、客户数、价目表、OEM 分成、毛利率桥或优先权结构。Caixin 和 Tencent 报道把反向因素讲得很清楚:商业化仍是未解测试,反复融资仍重要,汽车座舱赛道拥挤到即便技术故事不错也可能无法放大。Li Dahai 2026 年 6 月的表态也再次说明,难点是芯片、内存、带宽和软件的生态执行,不只是模型质量。这让估值讨论仍围绕战略期权价值,而不是已经验证的损益表。[CV007, CV008, CV009, CV010, CV011, CV012]

投资论点 / 反论点表
论点方向支撑依据什么会改变观点
ModelBest 的边缘 AI 切口是真实业务,不是 PPT 故事。正向官方、GitHub 和 Hugging Face 资料显示,MiniCPM 生态已在多类设备上跑起来。更多独立客户和付费部署证据,会把这条线索强化成完整商业化论点。
当前估值更多由战略投资者胃口支撑,而不是公开变现披露。反向2026 年 4 月融资报道验证了价格底线,但公开收入和定价仍缺席。审计后的收入桥和合同经济性,会降低对战略溢价逻辑的依赖。
开源牵引力是漏斗,本身不是收入模型。反向24M 下载量和 158 个托管模型显示触达,但来源没有披露转化、付费席位或 OEM 抽成率。经验证的下载和部署向经常性收入转化,会显著改善观点。
中国和香港资本市场当前正在奖励 AI 稀缺性。正向CB Insights 和 KPMG 都显示 AI 融资集中,香港专科科技上市窗口也很热。IPO 市场降温或下一轮走弱,会迅速削弱这种外部估值支撑。
ModelBest 比 Moonshot 或 DeepSeek 便宜,但变现也更不透明。反向Moonshot 和 DeepSeek 的头部估值更大,但 Moonshot 至少有公开 ARR 锚点,DeepSeek 也拥有前沿稀缺地位。如果 ModelBest 拿出可比的变现证据,相对估值会更有吸引力。
证券质量可能落后于公司质量。反向没有公开股权结构、优先权层级、清算瀑布或投资者权利包。清晰的条款披露,比更多产品头条更快改善当前估值的可投性。

这些论点把公司质量与证券质量分开;反论点主要指向定价、变现和条款,而不是产品不存在。

[CV007, CV010, CV013, CV014, CV034, CV036]
FV004: 投资 KPI

ModelBest 在战略价值和产品方向上得分不错,但经济性可见性和估值舒适度较弱。

评分是截至运行日期、基于保留证据推导的 IC 式启发,不是管理层 KPI。

[CV007, CV010, CV011, CV013, CV039, CV043]

8.3 可比公司组与中国资本市场背景

合适的可比框架不是单一整洁倍数,而是一组区间。站在中国私有模型稀缺性的高端,Moonshot 以 $20 billion 估值融资约 $2 billion;公开来源称其 4 月 ARR 超过 $200 million。DeepSeek 首轮外部融资据报估值超过 $50 billion;StepFun 2026 年 pre-IPO 融资传闻从投前 $4-6 billion 到预期 $10 billion 基石标记不等。公开市场侧,Zhipu 和 MiniMax 说明,即便大模型发行人仍亏损,香港投资人也愿意支付有意义的稀缺溢价;Palantir 和 C3.ai 则提供上下沿公开市场倍数锚点,且收入账本透明得多。 中国 2026 年资本市场背景很重要,因为它解释了 ModelBest 头部价格为什么并非天然荒谬。CB Insights 称 2026 年 Q1 AI 融资异常集中在超级轮次领头公司;KPMG 称香港 2026 年 Q1 IPO 募资额全球第一,录得 6 家专门科技上市和 366 个有效申请。Reuters 1 月 MiniMax IPO 报道补充了短期情绪信号:香港刚经历 2021 年以来最强 IPO 年,AI / 芯片发行人正迎着强劲投资人需求入场。这个环境能支撑 ModelBest 当前头部估值标记。但环境本身不能让这个价格具备吸引力,因为 ModelBest 披露出的变现成熟度远落后于 Moonshot,也远不如任何公开可比公司透明。[CV021, CV022, CV023, CV024, CV025, CV026]

可比估值表
可比对象关键指标倍数 / 估值 / 状态相关性局限
ModelBest(私有,2026 年 4 月)无公开收入;报道下载量 24M+;2026 年第二轮融资约 $1B 级 / 独角兽门槛估值;具体投后未披露本次直接审查标的没有公开收入、定价或条款清单,无法把名义估值转成经济性
Moonshot AI(私有,2026 年 5 月)ARR 在 2026 年 4 月超过 $200M$20B 估值;按披露下限约 100x ARR带有最新估值数据的中国开放权重变现最佳可比变现基础大得多,API / 订阅面也比 ModelBest 更宽
Zhipu AI(香港上市,2026 年 1 月)H1 2025 收入 CNY190.9M;H1 净亏损 CNY2.4B2026 年首日收盘市值 $7.4B说明即便亏损,香港也会给 AI 稀缺性定价IPO 定价和早期流通盘动态会高估稳态价值
MiniMax(香港上市,2026 年 1 月)首日市值 >$11.5B;9M25 净亏损 $512M公开市场接受了一家市值 $11.5B+ 的亏损 AI 发行人有用的中国 AI 公开发行先例首日估值波动大,业务组合也比 ModelBest 更全球化
StepFun(IPO 前,2026 年)$4B 第一批投前;$5-6B 第二批;预期约 $10B 基石估值后期私募 / IPO 过渡期定价另一家偏设备与分发的中国 AI 创业公司IPO 前条款仍未官方确认,并且继续依赖上市窗口
DeepSeek(私有,2026 年 6 月)未披露公开收入;首轮外部融资估值 >$50B> $50B 估值,并带有不寻常的控制权和锁定条款中国前沿模型稀缺性的上界控制权结构和缺少收入披露,削弱直接可比性
Palantir(公开,2026 年 7 月)TTM 收入 $5.22B市值 $309.97B;59.25x P/S;57.76x EV/Sales有真实收入和盈利能力的成熟 AI 平台,可作为公开市场上界比 ModelBest 成熟、全球化、财务透明得多
C3.ai(公开,2026 年 7 月)FY2026 收入 $250.3M;31% GAAP 毛利率市值 $1.41B;5.64x P/S;3.57x EV/Sales执行承压 AI 软件公司的较低公开基准不是边缘模型平台,公开市场情绪也明显弱于中国稀缺性可比对象

这是一组样本式可比对象,用来框定估值逻辑;不是穷尽每一家 AI 发行人或私募轮次。

[CV001, CV021, CV023, CV025, CV026, CV028]
FV002: 估值敏感性

当前估值对商业化证明最敏感,而不是对又一条增量牵引力新闻最敏感。

取值是 1 到 5 的序数影响评分,不是概率估计,也不是机械定价模型。

[CV012, CV013, CV014, CV034, CV036, CV043]

8.4 牛 / 基准 / 熊情景逻辑与建议

以目前证据水平,正确建议是继续研究,估值立场是偏高。原因不是 ModelBest 缺少战略价值,而是 $1B 级 入场价已经要求投资人相信,开源牵引力和伙伴部署会在中国 AI 稀缺性热度降温前转化为有意义、可审计收入。没有公开收入基底时,合适方法是里程碑和情景分析,不是假精确 DCF 或单一收入倍数。 牛市情景下,ModelBest 把下载量和 OEM 布点转化为真正的经常性软件经济性,证明清晰收入桥,并继续受益于中国 AI 融资和 IPO 市场支持;这可以支撑 $1.3-1.8 billion 区间。基准情景下,部署证明继续增加,但变现仍只部分披露,当前 $0.9-1.1 billion 区域更接近公允,而非便宜。熊市情景下,开源使用主要停留在社区,OEM 交易更像战略试点而不是持久软件收入,或资本市场不再奖励稀缺性,价值可能滑回约 $0.6-0.8 billion。升级路径因此取决于证据或价格:要么有更好的经审计证明,要么有更好的入场价。[CV039, CV041, CV042, CV043, CV044, CV045]

乐观 / 基准 / 悲观情景表
情景假设估值 / 回报逻辑概率信号关键风险
乐观审计后的年化收入出现,OEM 和企业合同显示真实经常性软件经济性,中国 AI 稀缺资产仍有强买盘。$1.3B-$1.8B;只有变现迅速追上开源和部署牵引力,当前估值标记才算可接受。低-中需要快速拿出商业化证明,而不只是再融一轮战略资金。
基准部署证明继续扩大,但收入披露仍然只是局部,市场仍给边缘 AI 稀缺性溢价。$0.9B-$1.1B;当前估值大致合理,但安全边际很薄。新投资者仍暴露在稀释、优先权和市场窗口风险下。
悲观下载主要停留在社区使用,OEM 交易证实以试点为主,或中国 AI 资本市场在变现可见前降温。$0.6B-$0.8B;当前估值标记会向非独角兽区间下修。中-高开源变现失败和溢价压缩可能同时发生。

情景区间是示意性的里程碑估值,不是 IRR 输出,因为公开记录没有披露收入质量、清算条款或稀释机制。

[CV039, CV041, CV042, CV043, CV044, CV049]
FV003: 估值 / 回报区间

用情景区间比单点目标更站得住,因为 ModelBest 有估值底部,却没有公开损益表。

区间是基于可比公司范围、资本市场状况和披露质量做出的情景判断,不是 DCF。

[CV034, CV036, CV039, CV042, CV043, CV049]

8.5 最终尽调问题和打破投资逻辑的触发器

估值缺口异常集中在少数可以补齐的尽调事项上。投资人需要月度收入和毛利率桥、真实 OEM 和企业定价结构、股权表和清算瀑布、活跃付费部署相对开源下载的独立遥测数据,以及任何可能让头部独角兽估值在经济上弱于表面的条款。没有这些,当前价格更像战略叙事,而不是经过承销的证券。 打破投资逻辑的触发器也同样具体。如果新一轮融资低于当前头部估值,说明估值支撑比表面更薄。如果披露的收入桥相对装机基础仍显得太小,说明开源牵引力没有转化为变现。如果优先权结构 或治理结构大幅压低新投资人优先级,即便公司质量仍强,证券质量也会改变。若 ModelBest 证明收入之前,中国 AI 融资或香港专门科技窗口转弱,当前估值标记内嵌的战略溢价可能很快压缩。[CV002, CV012, CV013, CV014, CV017, CV029]

论点失效与否决触发器表
触发器阈值对论点的传导行动含义
低价轮,或低于当前名义估值的结构化内部人主导融资新资金定价明显低于当前独角兽门槛,或加入很重的高级优先权说明当前价格的外部支撑被高估下调至回避 / 不投,直到条款和市场支撑重置
审计变现桥薄弱已确认的年化收入相对于装机牵引力仍太小,或看起来主要偏服务打破开源和 OEM 牵引力正转化为耐久软件经济性的论点把当前估值视为过度拉伸
合同经济性不利OEM 或企业交易证实以试点为主、靠补贴,或主要是集成劳务而非经常性软件即使部署是真实的,也会削弱边缘 AI 变现论点将估值重估到更接近非独角兽区间
股权结构 / 治理悬置风险优先权层级、清算瀑布或锁定条款让新投资者明显处于劣后把一家看起来成立的公司,变成条款弱的证券要求大幅折价,否则离场
中国 AI 市场窗口收缩香港专科科技需求或中国 AI 融资胃口,在 ModelBest 证明收入前降温拿掉估值论点中的稀缺性支撑从继续研究转为跟踪 / 等待重新定价

这些是与入场价和证据质量绑定的估值否决标准,不是对更广泛风险登记表的重复。

[CV029, CV034, CV036, CV037, CV043, CV044]
最终尽调问题表
主题缺失证据重要性负责人 / 尽调路径
收入桥月度确认收入、ARR 等效运行率、递延收入和渠道组合判断开源和部署牵引力是否真正支撑当前估值标记财务团队、审计资料包和董事会材料
定价和合同结构OEM 分成、单设备定价、企业订阅条款、支持或服务组合、续约条款区分真实软件价值与补贴试点或劳务密集集成管理层数据室 + 头部客户合同审查
股权结构和优先权层级完整股本结构表、清算瀑布、反稀释、ROFR 和投资者权利定义当前名义估值下的实际下行保护和证券质量公司法律顾问和融资文件
毛利率 / 算力成本桥托管与边缘推理成本分摊、GPU 承诺和支持人员负担测试边缘部署是否真正改善单位经济性FP&A 与基础设施运营审查
牵引力验证活跃付费部署、OEM 出货量和下载到付费转化的独立遥测把社区热度与可变现采用区分开合作伙伴确认、账单数据和平台看板
下一轮或退出路径下一轮融资、二级交易或上市流程的时间和结构当前估值部分依赖仍然打开的中国资本市场窗口董事会、领投方和投行顾问

这些问题按它们能多快改变推荐结论或可接受入场价排序,而不是按获取难度排序。

[CV002, CV012, CV013, CV014, CV015, CV017]

8.6 图表

免责声明

本报告仅供参考,反映截至 2026-07-02 的公开来源尽调。ModelBest 是私营公司;许多商业和治理细节仍未经审计或未披露,任何投资决策前都应独立核验。

证据索引

结论
编号陈述可信度来源
CO001 ModelBest operates under the legal entity name 北京面壁智能科技有限责任公司. SO002, SO003
CO002 Reviewed public sources place ModelBest's founding in August 2022. SO003, SO015, SO017
CO003 ModelBest's homepage says it wants to put large models closest to users across phones, AI PCs, smart cockpits, embodied robots, and wearables. SO001
CO004 ModelBest's public flagship family is MiniCPM, which official sources describe as an efficient end-side model series built for mainstream chips and system platforms. SO001, SO007
CO005 Multiple 2026 funding profiles describe ModelBest as a Tsinghua NLP lab-derived startup with its model R&D and training base in Beijing. SO010, SO017, SO027
CO006 OpenBMB's about page says the community was jointly initiated by Tsinghua University's NLP lab and ModelBest. SO005
CO007 Li Dahai is publicly identified as ModelBest's co-founder and CEO and previously served as a senior Zhihu technology executive. SO010, SO015, SO017
CO008 Accessible academic and profile sources identify Liu Zhiyuan as ModelBest's co-founder and chief scientist while he remains Tsinghua CS faculty. SO004, SO022
CO009 Public profile sources identify Zeng Guoyang as ModelBest's co-founder and CTO and connect him to earlier WuDao/Wenyuan model work. SO023
CO010 Reviewed public sources do not disclose a complete current board roster or ownership split for ModelBest. SO002, SO010, SO015
CO011 Caixin reported in April 2024 that ModelBest closed a hundreds-of-millions RMB round led by Primavera Capital and Huawei Hubble, with Beijing AI Industry Investment Fund and Zhihu following. SO018
CO012 Caixin reported that ModelBest then completed a December 2024 round led by Loongson Venture, Dinghui Bafu, Zhongguancun Science City Fund, and SAIF, with Beijing AI Industry Investment Fund and Qingke following. SO019
CO013 Caixin reported a further May 2025 round funded by Hongtai Fund, Gozone Capital, Qingkong Jinxin, and Moutai Fund, without disclosing the exact amount or valuation. SO019
CO014 Xinhua-affiliated coverage reported that ModelBest raised another hundreds-of-millions RMB round in February 2026 led by China Telecom Investment, with CITIC Jingshi and CITIC Private Equity participating. SO014
CO015 April 2026 coverage consistently says ModelBest completed a new hundreds-of-millions RMB round led by Shenzhen Capital Group and Huichuan Capital, with Daohe, Guotai Junan Innovation, and Wuyuefeng following. SO010, SO011, SO013, SO016, SO026, SO027
CO016 Multiple April 2026 reports say ModelBest's cumulative first-quarter 2026 financing exceeded RMB 1 billion once the China Telecom and Shenzhen/Huichuan rounds were combined. SO010, SO011, SO012, SO013, SO026, SO027
CO017 April 2026 reports say the latest financing pushed ModelBest's post-money valuation over the unicorn threshold, i.e. above roughly US$1 billion. SO010, SO012, SO017, SO027
CO018 The reviewed public round stories stop at unicorn-threshold language and do not disclose a precise April 2026 post-money figure in RMB or USD. SO010, SO017, SO027
CO019 QCC's shareholder roster shows a wide cap table spanning Zhihu, Beijing state funds, China Telecom Investment, Zhipu, Hubble, Loongson-related vehicles, Shenzhen Capital affiliates, and Huichuan-linked funds. SO002
CO020 ModelBest's company and community surfaces show that open-source distribution via OpenBMB, GitHub, and Hugging Face is central to its ecosystem strategy. SO005, SO006, SO009
CO021 The OpenBMB Hugging Face organization page shows a large public footprint with more than 150 hosted models and multiple MiniCPM demo spaces. SO009
CO022 The main MiniCPM GitHub repository describes MiniCPM5-1B as a dense 1B model built for on-device, local deployment, and resource-constrained scenarios. SO007
CO023 The MiniCPM-V repository describes the MiniCPM-V and MiniCPM-o lines as efficient multimodal families deployable across iOS, Android, and HarmonyOS. SO008
CO024 ModelBest's homepage says its MiniCPM series can be applied across phones, AI PCs, smart cockpits, embodied robots, wearables, and legal-AI workflows. SO001
CO025 Tencent News reported that ModelBest open-sourced MiniCPM-o 4.5 on 2026-02-04 as a 9B full-duplex omni-modal model for real-time interaction. SO025, SO008
CO026 Several April 2026 reports say cumulative MiniCPM downloads on GitHub and Hugging Face exceeded 24 million. SO012, SO017, SO026, SO027
CO027 Those 24 million download figures are company-reported through press coverage rather than independently audited platform telemetry. SO012, SO017, SO026, SO027
CO028 Independent and company-amplified coverage says MiniCPM has already been deployed in automotive programs including the Changan Mazda EZ-60 and Geely Galaxy M9. SO012, SO017, SO021
CO029 The February 2026 Xinhua-affiliated report says ModelBest had already partnered with Geely, Volkswagen, Changan, and GAC, and also participated in legal-AI infrastructure through the FAXIN digital-courts effort. SO014
CO030 ModelBest's August 2025 anniversary letter quoted Li Dahai calling the company both a leading legal-model provider and 端侧第一 in edge models. SO015
CO031 ThePaper's April 2026 profile says ModelBest released XAgent, AgentVerse, ChatDev, and an AI development tool in late 2023 before shifting harder into MiniCPM in early 2024. SO024
CO032 Li Dahai said after the April 2026 round that ModelBest would continue to use density law and open-source as first principles so that every terminal can access AGI capabilities. SO011, SO026
CO033 Funding and profile coverage consistently describes ModelBest's density-law strategy as pursuing stronger capability with fewer parameters under compute constraints rather than chasing sheer model scale. SO010, SO014, SO021, SO024
CO034 In June 2026 Li Dahai said the main bottleneck for edge models is not model capability itself but co-building with chips, memory, bandwidth, and software ecosystems. SO021
CO035 April 2026 funding coverage says ModelBest planned to launch hardware products such as the Songguo Pai development board and EdgeClaw Box in mid-2026 to connect models with applications. SO010, SO027
CO036 Caixin's April 2024 company profile framed commercialization as the central open question even as ModelBest announced another large financing round. SO018
CO037 Caixin's May 2025 report highlighted that ModelBest had already completed three rounds in a little over a year, signaling continued dependence on outside capital to fund expansion. SO019
CO038 Tencent's December 2025 skeptical profile argued that ModelBest's automotive-cockpit opportunity faces crowded competition from chip, OS, and auto-stack incumbents, making scale-up difficult. SO020, SO024
CO039 The chapter-one evidence base supports treating exact board disclosure, exact April 2026 post-money valuation, and independently auditable 24 million download telemetry as unresolved diligence items rather than settled facts. SO010, SO012, SO015, SO017, SO027
CM001 ModelBest’s homepage frames the company around putting large models close to users across phones, AI PCs, smart cockpits, embodied robots, and wearables. SM001
CM002 ModelBest says the MiniCPM family is designed to run across mainstream consumer electronics and different chip and system platforms with smaller parameter counts and faster inference. SM001
CM003 The MiniCPM repository describes MiniCPM5-1B as a dense 1B model built for on-device, local deployment, and resource-constrained scenarios. SM002
CM004 The MiniCPM repository says the MiniCPM4 series delivered more than 5x generation acceleration on typical edge chips. SM002
CM005 36Kr describes ModelBest’s strategic turn toward edge AI and smaller high-density models rather than chasing scale alone. SM003
CM006 36Kr highlights ModelBest’s partner set around Huawei, MediaTek, Lenovo, Intel, and Great Wall while discussing its edge-AI push. SM003
CM007 Gasgoo says ModelBest presents itself as an early edge-intelligence vendor with MiniCPM deployed at scale across automobiles, smartphones, PCs, and smart-home devices. SM004
CM008 IDC defines next-gen AI smartphones as devices with NPUs above 30 TOPS that can run on-device generative AI more efficiently. SM005
CM009 IDC forecasts global GenAI smartphone shipments at 234.2 million units in 2024 and 912 million units by 2028. SM005
CM010 IDC says next-gen AI smartphones face higher bill-of-materials pressure because 16GB of memory is already treated as a minimum configuration. SM005
CM011 Counterpoint forecasts GenAI-capable smartphones will account for 45% of global shipments in 2026 and 52% in 2027. SM006
CM012 Counterpoint also expects total global smartphone shipments to fall 13.9% year over year to 1.08 billion units in 2026 because of the memory supply crisis. SM006
CM013 TechNode, citing Counterpoint, says GenAI smartphone shipments should surpass 400 million units and about 30% of global smartphone shipments in 2025. SM027
CM014 TechNode says Chinese smartphone brands are not expected to bring GenAI features to mid- and low-end models until 2026 or 2027. SM027
CM015 IDC-linked reporting expects roughly 278 million smartphone shipments in China in 2026, down 2.2% year over year. SM028
CM016 IDC-linked reporting expects 147 million next-generation AI-phone shipments in China in 2026, or 53% of the market. SM028
CM017 Counterpoint says AI and GenAI-capable laptops already captured 27% of the global laptop market in 2024 and could approach 60% in 2025. SM007
CM018 Grand View sizes the global on-device AI market at USD 10.76 billion in 2025 and USD 36.64 billion in 2030. SM008
CM019 BCC Research sizes the global edge AI market at USD 11.8 billion in 2025 and USD 56.8 billion in 2030. SM009
CM020 MarketsandMarkets sizes the edge AI hardware market at USD 26.14 billion in 2025 and USD 58.90 billion in 2030. SM010
CM021 The reviewed market estimates are scope-mismatched because they describe on-device AI, broader edge AI, and edge-AI hardware rather than one identical market. SM008, SM009, SM010
CM022 MarketsandMarkets says smartphones represented 80.5% of edge AI hardware volume in 2024 and consumer electronics represented 81.3% of vertical volume. SM010
CM023 Counterpoint projects edge AI wearables will rise from 30% of shipments in 2025 to nearly 80% by 2032. SM011
CM024 Counterpoint says edge AI-enabled wearables capture roughly 75% of a cumulative USD 1 trillion wearables revenue opportunity between 2025 and 2032. SM011
CM025 TrendForce expects the global automotive semiconductor market to grow from USD 67.7 billion in 2024 to USD 96.9 billion in 2029, with automotive logic processors growing faster at 8.6% CAGR. SM012
CM026 TrendForce says smart-cockpit and cockpit-ADAS converged architectures start commercializing in 2025 as AI-model complexity and sensor data raise in-vehicle compute demand. SM012
CM027 ModelBest’s homepage says it has deep cooperation with Geely, Changan, Volkswagen, and Huawei, with landings across cars, phones, PCs, and smart homes. SM001
CM028 ModelBest’s homepage says the company worked with Intel on AI PCs and with MediaTek on next-generation mobile SoCs. SM001
CM029 Gasgoo reports ModelBest’s automotive deployments include Changan Mazda’s EZ-60 and Geely’s Galaxy M9. SM004
CM030 36Kr says edge AI removes network latency, can work offline or in weak-network settings, and avoids repeated cloud bandwidth, compute, and power cost for each inference request. SM003
CM031 Grand View says on-device AI adoption is driven by real-time processing demand, privacy and security concerns, and the spread of specialized NPUs and AI chips. SM008
CM032 The 2023 generative-AI measures took effect on 2023-08-15 and apply to generative-AI services provided to the public in China. SM013, SM014
CM033 The same measures require providers to label generated image and video content under the deep-synthesis rules and to protect user input information and usage records. SM013, SM014
CM034 The same measures require security assessment and filing for generative-AI services with public-opinion or social-mobilization attributes. SM013, SM014
CM035 Linklaters and Inside Privacy both say China’s AI-generated-content labeling measures and supporting national standard were issued in March 2025 and took effect on 2025-09-01. SM017, SM018
CM036 The labeling regime requires explicit visible labels, implicit metadata labels, and distribution-platform processes to detect or propagate AI-generated-content labels. SM017, SM018
CM037 China’s 2025 AI+ action targets more than 70% penetration for new-generation smart terminals and agents by 2027 and more than 90% by 2030. SM015
CM038 China’s 2025 AI+ action explicitly prioritizes AI phones, AI computers, connected vehicles, wearables, smart homes, and intelligent-agent applications. SM015
CM039 SCIO and Xinhua say the May 2026 AI-agent implementation guidelines define AI agents and identify 19 typical application scenarios. SM016
CM040 CAC’s May 2026 filing announcement says 72 additional generative-AI services and 49 registered applications or functions were added in March-April 2026, taking cumulative totals to 868 and 530 by April 30, 2026. SM019
CM041 CNBC and Reuters report that on 2026-05-31 the U.S. Commerce Department moved to require licenses for advanced chips even when China-headquartered entities seek them outside China. SM020
CM042 CSIS argues export controls both disrupt China’s access to advanced semiconductors and intensify China’s drive toward domestic substitution and self-sufficiency. SM021
CM043 KrASIA reports Alibaba cut Qwen-Long pricing by 97% to RMB 0.5 per million tokens, undercutting ByteDance Doubao’s RMB 0.8 per million tokens. SM022
CM044 Digital Watch reports Baidu made Ernie Speed and Ernie Lite free for business users while ByteDance had already cut Doubao pricing far below the prior industry average. SM023
CM045 Counterpoint’s December 2025 forecast said China’s smartphone market would decline 5% year over year in 2026. SM024
CM046 Counterpoint’s April 2026 update expected China’s smartphone shipments to decline 9% in 2026. SM025
CM047 Counterpoint’s quarterly China market-share page shows Huawei at 20%, Apple at 19%, and OPPO at 17% in Q1 2026. SM026
CM048 ModelBest’s publicly disclosed route to market is more partnership-led and OEM-led than direct-consumer-app-led. SM001, SM004, SM026
CM049 High concentration in China smartphones implies strong OEM and channel bargaining power over smaller embedded-model suppliers. SM026
CM050 ModelBest’s attributable paid market share across AI phones, AI PCs, smart cockpits, and edge-enterprise deployments is not publicly disclosed in the reviewed sources.
CM051 No reviewed public source isolates a China-only AI PC revenue pool that can be confidently assigned to ModelBest’s served market.
CP001 ModelBest positions itself as an edge-AI company that puts large models on phones, AIPCs, smart cockpits, embodied robots, and wearables. SP001
CP002 ModelBest says MiniCPM models have extreme compute and memory efficiency, smaller parameter counts, faster inference, and flexible deployment. SP001
CP003 MiniCPM is framed in its technical paper as a resource-efficient small-language-model alternative to very large models, with scalable training methods rather than brute-force parameter growth. SP003
CP004 The main MiniCPM repo says MiniCPM5-1B is a dense 1B on-device model and that MiniCPM models are released under Apache-2.0. SP002
CP005 OpenBMB claims MiniCPM5-1B reaches a 42.57 average benchmark score versus 35.61 for strong open-source peers in the same size class, with strengths in tool use, code, and difficult reasoning. SP002
CP006 Microsoft markets Phi as a family of small language models created for AI directly on a device without cloud connectivity, and says Phi models are open source through the MIT License. SP005
CP007 Phi-4-mini-instruct is listed as a 3.8B-parameter lightweight open model with 128K context, function calling, multilingual support, and intended use in memory- or latency-constrained environments. SP006
CP008 Microsoft’s own Phi family guide recommends GPT models for more complex planning and orchestration, while using Phi-family models for lower-cost and edge-oriented use cases. SP007
CP009 Google says Gemma models are provided with open weights and permit responsible commercial use. SP009
CP010 Gemma 4 explicitly targets mobile and IoT deployment with small E2B and E4B variants and published mobile memory footprints as low as 0.84 GB for text-only use. SP008, SP009
CP011 Gemma 4 small variants combine on-device deployment claims with multimodal, function-calling, and 128K-context support, making Google a direct edge-model competitor rather than just a cloud incumbent. SP008, SP009
CP012 Apple says Apple Intelligence is integrated into iPhone, iPad, and Mac through on-device processing and that any app can use the on-device models offline at no cost per request. SP010
CP013 Apple pairs those local models with Private Cloud Compute for harder requests, giving it a vertically integrated hybrid edge-plus-cloud substitute inside its own device ecosystem. SP010
CP014 Mistral 3 introduces three small dense Ministral models in 3B, 8B, and 14B sizes for edge and local deployment, all under Apache 2.0. SP012
CP015 Mistral’s platform pitch stresses self-hosted, edge, and on-prem deployment with data staying inside the customer environment, plus an EU-hosted cloud option. SP011
CP016 Meta’s Llama 3.2 text models come in 1B and 3B sizes with 128K context and are intended for multilingual assistant use cases including mobile writing assistants, retrieval, and summarization. SP014
CP017 Meta says Llama models have been downloaded hundreds of millions of times and support thousands of community projects, giving Llama far more open-source distribution than ModelBest can currently match. SP013
CP018 Use of Llama 3.2 is governed by the Llama 3.2 Community License, so Meta is less permissive than MIT- or Apache-licensed edge-model competitors even while remaining open-access. SP014
CP019 Qwen3 open-weights dense models all the way down to 0.6B, with Qwen3-8B and larger dense models supporting 128K context and Apache 2.0 licensing. SP015
CP020 Qwen3-Coder-Next is an open-weight coding-agent model with 256K native context, 1M-context extension via Yarn, and support for 358 coding languages. SP016
CP021 Qwen’s code repositories mix openness with commercial constraints: the source code is Apache 2.0, but some larger legacy Qwen weights require a separate commercial agreement. SP017
CP022 Moonshot markets Kimi K2.6 as an open-source model focused on coding, long-horizon execution, and agent-swarm capabilities, available free to use with paid upgrades. SP018
CP023 Kimi API documents say K2.6 has 256K context, multimodal input, tool use, and full compatibility with OpenAI’s API format. SP019
CP024 MiniMax positions M3 as a native multimodal, 1M-context model with frontier coding and agentic performance, using a sparse-attention architecture for long contexts. SP020, SP021
CP025 MiniMax’s own pages emphasize packaging more than public per-token pricing: token-plan users inherit M3 performance improvements, but the retrieved pages do not publish a transparent standalone price card. SP020, SP021
CP026 Z.ai’s consumer surface brands itself as an advanced chatbot and agent powered by GLM-5.2, while GLM-5.1 docs position the underlying stack for complex coding and long-horizon tasks. SP022, SP023
CP027 GLM-5.1 says it can autonomously keep working for up to eight hours on a single task and exposes a chat-completions interface with reasoning mode, making it a direct substitute for long-running coding agents rather than a pure phone-size model. SP023
CP028 Baidu says ERNIE 5.1 reaches leading performance at only about 6% of comparable-model pretraining cost and ranked fourth globally, first among Chinese models, on Arena Search. SP024
CP029 ModelBest says MiniCPM has already landed at scale across automotive, phone, PC, and smart-home domains with named relationships including Geely, Changan, Volkswagen, and Huawei. SP001
CP030 ModelBest also highlights Intel, Baidu Cloud, MediaTek, Great Wall, and court-network deployments, suggesting a partnership-led go-to-market rather than a dominant consumer app surface. SP001
CP031 ModelBest’s trust posture is unusually tangible for an open-model startup: it cites a Beijing key laboratory, court-network deployment in the legal sector, and ASPICE L2 automotive process certification. SP001
CP032 OpenBMB claims MiniCPM-V 4.6 beats Gemma4-E2B-it, outperforms Qwen3.5-0.8B on most vision-language tasks, and surpasses Ministral 3 3B on the Artificial Analysis Intelligence Index while supporting iOS, Android, and HarmonyOS deployment. SP004
CP033 OpenBMB also claims MiniCPM-V 4.5 at 8B outperforms GPT-4o-latest, Gemini-2.0 Pro, and Qwen2.5-VL 72B, which is directionally positive but still vendor-authored rather than independently audited. SP004
CP034 MiniCPM4.1-8B is described as the first open-source reasoning LLM with trainable sparse attention, and OpenBMB says it delivers roughly 7x decoding speed improvement over Qwen3-8B on Jetson AGX Orin in long-text tasks. SP002
CP035 ModelBest’s own site says MiniCPM4.0-0.5B reaches same-class SOTA with int4 quantization and 600 token/s inference, while MiniCPM4.0-8B claims Qwen-3-8B-like performance and to exceed Gemma-3-12B at 22% training cost. SP001
CP036 Taken together, the MiniCPM line competes best when a buyer wants multilingual, multimodal, mobile-friendly deployment across fragmented hardware rather than the absolute largest reasoning model. SP001, SP002, SP004
CP037 Because Kimi and GLM expose OpenAI-compatible APIs and MiniCPM, Gemma, Mistral, Llama, and Qwen all support mainstream open-source inference stacks, model-layer switching costs are structurally low for many developers. SP002, SP009, SP012, SP014, SP019, SP023
CP038 Apple’s offline, no-cost-per-request on-device framework is a particularly strong incumbent substitute because it removes a major reason to license or integrate an external edge model inside the Apple ecosystem. SP010
CP039 AICPB’s April 2026 ranking shows just how concentrated Chinese AI-app distribution is: Doubao at 336.04M MAU, Qwen at 220.35M, Quark at 162.30M, DeepSeek at 138.98M, and Kimi at 25.33M. SP025
CP040 ModelBest is absent from AICPB’s top-50 China AI app ranking, which suggests its route to scale is OEM embedding and enterprise partnerships, not mass-consumer app pull. SP001, SP025
CP041 Open-source availability itself looks commoditized: Phi, Mistral, MiniCPM, Gemma, Qwen, Kimi, and MiniMax all present some form of open weights or broad developer access. SP002, SP005, SP009, SP012, SP015, SP018, SP020
CP042 ModelBest’s open-source strategy is most differentiated where buyers need cross-platform Chinese-edge deployment and are unwilling to rely on a single foreign cloud or operating-system vendor. SP001, SP002, SP004
CP043 That strategy is weaker where the buyer is already captured by an incumbent distribution surface such as Apple devices, Azure enterprise tooling, Meta’s community default, or Alibaba/Baidu consumer ecosystems. SP005, SP010, SP013, SP025
CP044 MiniMax M3 and Kimi K2.6 are less “phone-native” than MiniCPM, but they are still meaningful substitutes for local or self-hosted developers because they combine open access with long context and agentic coding workflows. SP018, SP019, SP020, SP021
CP045 Trust is also commoditizing in different forms: ModelBest cites OEM and court-grade delivery, Apple centers privacy, Mistral centers self-hosting, Microsoft centers safety post-training, and Baidu centers domestic cost-performance. SP001, SP005, SP010, SP011, SP024
CP046 ModelBest’s partnerships with Baidu Cloud, Intel, MediaTek, Great Wall, Geely, Changan, Volkswagen, and Huawei indicate real ecosystem traction but also imply dependence on larger counterparties with bargaining power of their own. SP001
CP047 Meta’s broad community adoption and Apple’s default OS-level integration show that in AI infrastructure, distribution can outweigh model openness or benchmark efficiency. SP010, SP013
CP048 MiniCPM’s clearest durable asset is not exclusive access to small-model technology, but a combination of efficient multimodality, cross-OS deployment, and China-specific partnership fit. SP001, SP002, SP004
CP049 Without partner attach-rate, renewal, or revenue data, the public record does not prove that ModelBest’s OEM relationships create high switching costs rather than reversible integration wins. SP001
CP050 Compared with Apple’s free on-device API surface, Microsoft’s enterprise distribution, and China’s top AI apps, ModelBest’s commercial surface appears much stronger in partner channels than in direct-user control. SP001, SP005, SP010, SP025
CP051 The open-source battlefield is crowded enough that even strong MiniCPM releases can be diluted by larger ecosystems around Llama, Gemma, Qwen, Phi, and Mistral. SP005, SP009, SP012, SP013, SP015
CP052 MiniCPM-V 4.6’s support for iOS, Android, and HarmonyOS plus open-sourced edge adaptation code is one of the strongest pieces of evidence that ModelBest is solving for fragmented device fleets rather than a single walled garden. SP004
CP053 The strongest public trust signals around ModelBest are regulatory or delivery-oriented—ASPICE, legal-sector deployment, and local lab backing—rather than dominant consumer distribution or proprietary platform control. SP001, SP025
CI001 ModelBest publicly positions itself around efficient on-device AI for phones, AIPC, intelligent cockpits, embodied robots, and wearables. SI001, SI002
CI002 ModelBest says the MiniCPM series can run on mainstream consumer electronics across different chip and system platforms with smaller parameter footprints and faster inference. SI001, SI002
CI003 ModelBest says its legal-focused model has been deployed on court intranets for professional judicial workflows. SI001
CI004 ModelBest maintains public Feishu documentation for MiniCPM and a separate deployment guide for MiniCPM-V 2.6. SI003, SI004
CI005 The MiniCPM raw README frames MiniCPM5-1B as an on-device, local-deployment model for resource-constrained scenarios. SI009
CI006 The MiniCPM-V-Apps README says the MiniCPM-V family can run fully on-device on iOS, Android, and HarmonyOS via llama.cpp. SI011
CI007 The official MiniCPM-o demo system recommends one NVIDIA GPU per worker-backend instance in Docker deployment. SI013
CI008 The official MiniCPM-o demo system reports roughly 21.5 GB of VRAM per initialized worker and about 0.5-0.9 seconds of omni full-duplex latency on A100 hardware. SI013
CI009 Tencent News cited Deloitte as estimating that inference would account for 66% of AI compute in 2026. SI024
CI010 The same Tencent infrastructure coverage says many online inference workloads on train-and-infer chips use only 5%-10% of their capacity, raising effective cost. SI024
CI011 36Kr says 2026 is the year Chinese AI chips begin crossing from inference into training deployment, underscoring that compute remains capital intensive. SI023
CI012 36Kr Research says China's large-model market reached RMB294.16 billion in 2024 and is expected to exceed RMB700 billion by 2026. SI022
CI013 36Kr Research says competition in China's large-model market is shifting from single-product battles to ecosystem building, industry enablement, and commercialization capability. SI022
CI014 No public API or token price card appears on the reviewed official homepage, Feishu docs, GitHub repos, or raw readmes as of 2026-07-02. SI001, SI003, SI004, SI005, SI006, SI009, SI010, SI011, SI012, SI013
CI015 The reviewed public technical surfaces emphasize deployment guides, demos, and open distribution rather than a public self-serve commercial checkout flow. SI003, SI004, SI009, SI011, SI012, SI013
CI016 OEM or device embedding is a likely revenue channel because public sources consistently center phones, cars, AIPC, and wearables as commercialization surfaces. SI001, SI014, SI019
CI017 Enterprise and on-prem deployment is a likely revenue channel because ModelBest markets legal workflows and private deployment materials rather than only public model access. SI001, SI003, SI004, SI021
CI018 Open-source-to-enterprise conversion, including support, tuning, and integration work, is a likely revenue channel because MiniCPM is widely distributed while commercial pricing stays private. SI002, SI005, SI007, SI008, SI009, SI011, SI012, SI013
CI019 Industrial and robotics-oriented commercialization is a likely revenue channel because April 2026 investor coverage explicitly links ModelBest to industrial automation and physical-world AI. SI018, SI020, SI026
CI020 Multiple independent April 2026 reports say ModelBest raised a new several-hundred-million-renminbi round led by Shenzhen Capital Group and Huichuan or Inovance-linked industrial capital, with Daohe, Guotai-related innovation capital, and Wuyuefeng following. SI014, SI015, SI016, SI017, SI018, SI019, SI020
CI021 Multiple independent April 2026 reports say the April financing was ModelBest's second round of 2026. SI014, SI017, SI018, SI019, SI020
CI022 Multiple independent April 2026 reports say ModelBest's cumulative first-quarter 2026 fundraising exceeded RMB1 billion. SI014, SI016, SI017, SI018, SI019, SI020
CI023 Multiple independent April 2026 reports say the latest round pushed ModelBest into the unicorn valuation threshold, effectively around a $1 billion floor. SI017, SI018, SI019, SI021
CI024 Independent April 2026 coverage says China Telecom led an earlier February 2026 financing round and acted as a strategic investor for cloud, network, and device-side collaboration. SI014, SI017, SI018, SI019, SI020
CI025 Eastmoney and Tencent coverage attribute Shenzhen Capital's investment thesis to ModelBest's ability to raise intelligence density under compute constraints and lower deployment or calling costs. SI017, SI018
CI026 Eastmoney and Tencent coverage say Huichuan or Inovance-linked capital saw ModelBest as a fit for industrial automation, robotics, and physical-world AI deployment. SI017, SI018
CI027 Eastmoney and Tencent coverage say China Telecom provided cloud, network, and terminal advantages that could widen ModelBest's reach in automotive, education, and other scenarios. SI017, SI018
CI028 Shenzhen Capital's official homepage presents the group as a hard-tech investor with a fund network spanning 28 provincial-level regions. SI025
CI029 Inovance's official homepage identifies the company as an industrial-technology and automation player, consistent with industrial-synergy narratives around Huichuan-linked capital. SI026
CI030 Sina/TechWeb coverage says MiniCPM downloads across GitHub and Hugging Face exceeded 24 million. SI014
CI031 The OpenBMB GitHub organization page shows active repositories updated through July 1-2, 2026, including MiniCPM-V Apps and MiniCPM-o demo projects. SI007
CI032 The archived OpenBMB Hugging Face organization page shows 18 collections, 11 Spaces, 158 models, and MiniCPM-V-4.6 with 802k downloads and 1.13k likes. SI008
CI033 ModelBest's homepage positions the company around AI-native legal workflows and agent-style applications, not just around raw model weights. SI001
CI034 ThePaper characterizes ModelBest as an industry outlier that avoided the compute war, price war, and C-end traffic war by staying focused on edge deployment. SI021
CI035 ThePaper says ModelBest reached April 2026 without the spotlight or giant financings associated with the highest-profile AI peers, implying a more resource-constrained operating path. SI021
CI036 The public sources reviewed here disclose no revenue, ARR, recognized revenue mix, or revenue-recognition policy for ModelBest. SI001, SI003, SI004, SI014, SI017, SI018, SI021, SI022
CI037 The public sources reviewed here disclose no cash balance, monthly burn, runway, or debt or project-finance obligations for ModelBest. SI001, SI014, SI017, SI018, SI020, SI021
CI038 The public sources reviewed here disclose no gross margin, CAC, payback, NRR, or customer-concentration schedule for ModelBest. SI001, SI003, SI004, SI014, SI017, SI018, SI021
CI039 Because no public price list or contract schedule exists, realized pricing for OEM, enterprise, and support contracts cannot be externally underwritten. SI001, SI014, SI018, SI021
CI040 On-device deployment can move part of inference economics onto customer hardware, but it does not eliminate model training, adaptation, and hosted-service costs. SI006, SI011, SI012, SI013, SI023, SI024
CI041 MiniCPM-o's official demo architecture still implies ongoing server-side cost through gateway, worker, backend, mounted-weight, and GPU requirements. SI013
CI042 ModelBest's 2026 capital stack spans telecom, state-backed venture, and industrial automation investors rather than pure financial VC alone. SI017, SI018, SI020, SI025, SI026
CI043 That investor mix likely reduces near-term financing fragility but increases pressure to prove deployment ROI in industrial and government-linked scenarios. SI017, SI018, SI020, SI021, SI025, SI026
CI044 ModelBest's open-source distribution intensifies monetization pressure because model access and deployment know-how are highly visible even if contract economics are not. SI002, SI005, SI007, SI008, SI009, SI021
CI045 ThePaper's description of roughly semiannual fundraising from April 2023 to April 2026 implies repeated external capital has been an important part of ModelBest's operating runway. SI021
CI046 The reviewed official and technical surfaces do not reveal whether OEM economics are license fees, revenue shares, bundled support contracts, or strategic cross-subsidies. SI001, SI003, SI004, SI009, SI011, SI013
CI047 The reviewed official and technical surfaces do not reveal whether enterprise deployments are recurring subscriptions, one-off project revenue, or usage-based contracts. SI001, SI003, SI004, SI013
CI048 The MIIT filing portal returned a 521 response during this review, so filing-based verification of company-specific registration details remained incomplete. SI027
CI049 No accessible official investor-side announcement or filing in the fetched set independently confirmed the latest round terms beyond finance-press reporting. SI025, SI026, SI027
CE001 ModelBest’s official homepage positions MiniCPM as an edge-first model family for phones, AIPC, intelligent cockpits, embodied robots, wearables, and legal workflows rather than as a single cloud chat endpoint. SE001
CE002 ModelBest says MiniCPM can run across mainstream consumer-electronics devices and across multiple chip and operating-system platforms. SE001
CE003 ModelBest’s homepage markets GUI Agent, intelligent cockpit, AI-native legal workflows, and an Internet-of-Agents vision as distinct application surfaces around the same model family. SE001
CE004 ModelBest says its legal-model workflow has been deployed on a court private network and serves professional judicial scenarios. SE001
CE005 The core MiniCPM release line documented in the repo spans MiniCPM-2B, MiniCPM-2B-128k, MiniCPM-MoE-8x2B, MiniCPM-1B, MiniCPM3-4B, MiniCPM4, MiniCPM4.1, and MiniCPM5-1B. SE007, SE014
CE006 The MiniCPM paper says the family’s 1.2B and 2.4B variants reached performance comparable to 7B-13B models while focusing on small-model efficiency. SE021, SE022
CE007 The MiniCPM repo says MiniCPM3-4B outperforms Phi-3.5-mini-instruct and GPT-3.5-Turbo-0125 while remaining comparable to several 7B-9B instruct models. SE014
CE008 ModelBest’s technical-blog surface explicitly frames BitCPM as its ultra-low-bit quantization track for MiniCPM models. SE001, SE014
CE009 MiniCPM-V 2.6 is an 8B multimodal model built on SigLip-400M and Qwen2-7B and supports single-image, multi-image, and video understanding. SE018
CE010 MiniCPM-V 2.6 can process images up to 1.8 million pixels while emitting only 640 visual tokens, which the model card says is 75% fewer than most peers. SE018
CE011 MiniCPM-V 2.6 lists llama.cpp support, 16 int4 or GGUF quantized sizes, vLLM support, fine-tuning paths, and local demo tooling. SE018
CE012 The MiniCPM-V 2.6 model card says the model can support real-time video understanding on end-side devices such as iPad. SE018
CE013 MiniCPM-o 2.6 is an 8B omnimodal model built on SigLip-400M, Whisper-medium-300M, ChatTTS-200M, and Qwen2.5-7B. SE019
CE014 MiniCPM-o 2.6 adds bilingual real-time speech conversation, configurable voices, end-to-end voice cloning, and multimodal live streaming. SE019
CE015 MiniCPM-o 2.6 also advertises llama.cpp support, multiple GGUF quantizations, vLLM support, and fine-tuning hooks. SE019
CE016 MiniCPM-V 4.5 is an 8B model built on Qwen3-8B and SigLIP2-400M. SE020, SE024
CE017 MiniCPM-V 4.5 introduces a unified 3D-Resampler that compresses six 448x448 video frames into 64 tokens, which the model card calls a 96x video-token compression rate. SE020, SE024
CE018 MiniCPM-V 4.5 advertises switchable fast and deep thinking modes plus official support across llama.cpp, vLLM, SGLang, and LLaMA-Factory. SE020
CE019 The MiniCPM-V 4.5 paper reports state-of-the-art sub-30B VideoMME performance while using 46.7% of the GPU memory cost and 8.7% of the inference time of Qwen2.5-VL 7B. SE024
CE020 The MiniCPM-o 4.5 paper says Omni-Flow aligns multimodal inputs and outputs on a shared temporal axis to enable full-duplex and proactive interaction. SE023
CE021 The MiniCPM-o 4.5 paper says the 9B model can perform real-time full-duplex omni-modal interaction on edge devices with less than 12GB RAM cost. SE023
CE022 The official MiniCPM-o surface says low-latency full-duplex live streaming can run locally on an Apple M4 Max system with 24GB RAM or on an Nvidia GPU with 12GB of memory. SE011
CE023 MiniCPM-V-Apps packages MiniCPM-V as fully offline multimodal chat across iOS, Android, and HarmonyOS and also bundles MiniCPM5-1B and VoxCPM2 as on-device components. SE012, SE016
CE024 The MiniCPM-V-Apps hardware guide lists roughly 5.4GB download and at least 8GB RAM for MiniCPM-V 2.6, roughly 1.6GB and at least 6GB RAM for MiniCPM-V 4.6, and roughly 0.5GB plus at least 4GB RAM for MiniCPM5-1B. SE016
CE025 The MiniCPM-V-Apps README says MiniCPM-V 2.6 has been validated on iPhone 15 Pro or 16 series devices and on recent iPads with M-series chips. SE016
CE026 The iOS app build path requires Xcode 26.1, an iOS 16.4 deployment target, and a locally built llama.xcframework from the llama.cpp-omni submodule. SE016
CE027 MiniCPM5-1B is a dense 1B on-device model aimed at local assistants, coding agents, tool-use workflows, and reasoning scenarios. SE014
CE028 MiniCPM5-1B ships official quickstarts for vLLM, SGLang, and Transformers, and its vLLM example requires version 0.21 or newer. SE014
CE029 The OpenBMB Hugging Face organization page shows 158 models and 11 spaces, indicating a distribution footprint larger than a single flagship repo. SE017
CE030 GitHub API metadata shows OpenBMB had 80 public repositories and 6,726 followers on the access date. SE006
CE031 GitHub API metadata shows the MiniCPM repo had 9,536 stars and 625 forks on the access date. SE008
CE032 GitHub API metadata shows the MiniCPM-V repo had 25,767 stars and 2,018 forks on the access date. SE010
CE033 GitHub API metadata shows the MiniCPM-V-Apps repo had 326 stars, 48 forks, and a 2026-07-01 push date on the access date. SE013
CE034 The MiniCPM and MiniCPM-V repositories are distributed under the Apache 2.0 license. SE008, SE010
CE035 The generic vLLM documentation says the Transformers modeling backend can load compatible Hugging Face models, but vision-language support there currently accepts only image inputs. SE025
CE036 ModelScope hosts a MiniCPM-V-4.6 GGUF page, showing a partner distribution surface beyond GitHub and Hugging Face even though the page is JavaScript-heavy. SE026, SE016
CE037 ModelBest’s latest product feed shows continued cadence beyond the 2.x and 3.x generations, including posts for MiniCPM-V 4.0, MiniCPM-V 4.5, and MiniCPM 4.1. SE003
CE038 ModelBest’s homepage says it became the first Chinese large-model company to pass ASPICE L2 and says MiniCPM has landed with automotive, smartphone, PC, and smart-home partners including Geely, Changan, Volkswagen, and Huawei. SE001
CE039 The MiniCPM-o documentation explicitly says full-duplex omni-modal capability still needs improvement and that omni-mode speech output can mispronounce characters. SE011
CE040 No public SOC 2, ISO 27001, trust-center, or uptime-status page was visible across the reviewed ModelBest and OpenBMB official surfaces. SE001, SE002, SE003, SE004
CE041 The strongest public MiniCPM packaging pattern is open-source distribution plus device-specific adaptation, not a unified closed commercial control plane. SE001, SE012, SE014, SE016, SE017, SE026
CE042 The public product stack spans base text models, multimodal models, omni voice models, and packaged mobile apps rather than a single checkpoint. SE001, SE014, SE015, SE016, SE019, SE020
CE043 ModelBest’s clearest product moat is unusually broad cross-platform edge deployment, especially for multimodal workloads, rather than a differentiated proprietary API or security-control layer. SE001, SE016, SE018, SE020, SE023
CE044 Trust and compliance disclosure lag capability breadth because public materials show open code, Apache licensing, ASPICE process maturity, and known-model caveats, but not a public security or content-governance control plane. SE001, SE008, SE010, SE011
CU001 Official ModelBest product surfaces position MiniCPM inside phones, AIPC devices, intelligent cabins, embodied robots, and wearable devices. SU001, SU002, SU003
CU002 ModelBest’s English site says MiniCPM has been widely acclaimed by the global open-source community and repeatedly topped GitHub and Hugging Face trending charts in 2024. SU002
CU003 The public evidence implies a buyer-user-payer split in which OEMs, telecoms, or enterprises likely pay for deployments while many end users experience MiniCPM indirectly through embedded products or community tools. SU001, SU003, SU005, SU011
CU004 ModelBest publicly claims its legal AI system is the first large model connected to a court case-handling workflow and that it has been deployed on a court private network. SU001, SU003
CU005 Gasgoo says ModelBest already has algorithm deployments across legal, automotive, and education sectors. SU017
CU006 CnTechPost and Gasgoo both report that MiniCPM has surpassed 24 million cumulative downloads across major platforms. SU016, SU017
CU007 The OpenBMB Hugging Face organization snapshot shows 18 collections, 11 spaces, and 158 models. SU011
CU008 The same Hugging Face snapshot shows about 322k downloads for MiniCPM5-1B and about 802k downloads for MiniCPM-V-4.6. SU011
CU009 The OpenBMB Hugging Face organization lists BitCPM-CANN variants, indicating a Huawei Ascend or CANN-oriented distribution surface around the ModelBest ecosystem. SU011
CU010 Sina Tech says MiniCPM-o 4.5 exceeded 250,000 Hugging Face downloads after its February 2026 launch. SU024
CU011 MiniCPM5-1B is explicitly documented as an on-device and local model for resource-constrained scenarios. SU005, SU006
CU012 The MiniCPM repo documents deployment cookbooks for llama.cpp, Ollama, LM Studio, MLX, and SGLang. SU005, SU006
CU013 The MiniCPM-V repo says the model family is adapted to SGLang, vLLM, llama.cpp, Ollama, and common mobile platforms. SU007, SU008
CU014 MiniCPM-V-Apps provides fully offline demos for iOS, Android, and HarmonyOS NEXT using llama.cpp. SU009, SU010
CU015 Intel published an OpenVINO optimization guide for MiniCPM-V-2 and its official documentation confirms CPU, GPU, and NPU inference support. SU014, SU015
CU016 Ollama hosts an official openbmb/minicpm5 entry, extending MiniCPM distribution into a mainstream local-LLM ecosystem. SU013
CU017 China Telecom led a 2026 financing round for ModelBest and planned deep business collaboration using cloud computing and computing-power networks. SU016, SU017
CU018 The China Telecom relationship is publicly framed as a channel for edge and complex-scenario expansion rather than as a disclosed revenue contract with unit economics. SU016, SU017
CU019 Gasgoo and Sina Finance both report that MiniCPM has been integrated into vehicles from Geely and Changan Mazda. SU017, SU026
CU020 The Geely Galaxy M9 is publicly named as carrying the MiniCPM multimodal model. SU017, SU026
CU021 The Changan Mazda EZ-60 is publicly described as the first mass-produced vehicle equipped with an edge-side model. SU017, SU026
CU022 Mazda’s own newsroom confirms the EZ-60 is a real Changan Mazda China-market EV program unveiled in April 2025. SU019
CU023 Sina Finance says ModelBest has deep cooperation with Geely, Changan, Volkswagen, and Huawei. SU026
CU024 ModelBest’s official product pages highlight intelligent cabins and mainstream automotive chip support but do not disclose OEM contract terms or unit economics. SU003, SU026
CU025 ModelBest names the legal workflow domain but not the specific court, contract size, or renewal status of the claimed court private-network deployment. SU001, SU003
CU026 OpenBMB, Tsinghua-linked collaboration, and the MiniCPM4 arXiv report show a real academic and research-user channel in addition to commercial deployment. SU004, SU020
CU027 MiniCPM-o 4.5 offers a public online demo, free API, and downloadable local installer package, widening self-serve trial usage. SU024, SU025
CU028 The MiniCPM-V-4.6 GGUF page says a public free API key was released and presents the model as the most edge-deployment-friendly version to date. SU012
CU029 No public source reviewed for this chapter discloses paying-customer count, ARR, NRR, GRR, contract length, or churn for any ModelBest customer segment. SU001, SU003, SU016, SU017, SU024, SU026
CU030 Public named commercial proof is concentrated in China Telecom plus a small number of auto programs, while much of the remaining traction evidence is ecosystem or community-based. SU016, SU017, SU026, SU011
CU031 Download and community metrics do not distinguish experimentation from paid production use. SU011, SU017, SU024
CU032 ModelBest’s open-source reach is global, but its named commercial proof in the fetched record is concentrated in China-centric channels. SU002, SU011, SU013, SU014, SU017, SU026
CU033 Developer adoption currently depends on third-party distribution channels such as GitHub, Hugging Face, Ollama, and OpenVINO rather than on a publicly visible direct-sales funnel. SU005, SU011, SU013, SU014
CU034 The visible customer journey starts with open-source discovery, proceeds through local evaluation and integration tooling, and only then surfaces a small set of named production-style deployments. SU005, SU007, SU009, SU011, SU014
CU035 Local deployment and privacy-preserving inference are the main reasons ModelBest can credibly sell into legal, telecom, automotive, and smart-terminal scenarios. SU001, SU003, SU014, SU017
CU036 Automotive concentration risk is elevated because only Geely Galaxy M9 and Changan Mazda EZ-60 are publicly named vehicle programs with product detail. SU017, SU019, SU026
CU037 Enterprise proof outside auto and telecom is thin because legal and education are mentioned publicly without named customers or renewal metrics. SU001, SU017
CU038 Sina’s MiniCPM-o 4.5 coverage says many users lose patience under weak half-duplex interaction and that multimodal rollout is hindered when experience quality is poor. SU024
CU039 Gasgoo and Pandaily frame recent capital raises as money to accelerate commercialization, implying monetization is still being built rather than already disclosed at scale. SU018, SU022
CU040 China Daily Brief says ModelBest’s Q1 2026 financing haul exceeded RMB1 billion, evidencing investor conviction rather than customer-revenue disclosure. SU023
CU041 The MiniCPM-V-Apps README specifically supports HarmonyOS NEXT alongside iOS and Android, deepening Huawei-adjacent device ecosystem reach. SU010
CU042 The raw MiniCPM README pairs deployment and fine-tuning cookbooks with agent skills, lowering developer onboarding friction. SU006
CU043 MiniCPM-V 4.6 is presented across repository and model-card surfaces as especially edge-deployment-friendly and officially supported by mainstream local-inference frameworks. SU008, SU012
CU044 The English site says MiniCPM became one of the most popular LLMs on Hugging Face in 2024 after repeatedly topping trending charts. SU002
CU045 OpenBMB’s site shows repeated MiniCPM base and multimodal launches through 2025, signaling ongoing community release cadence rather than a one-off model drop. SU004
CU046 Sina Finance says MiniCPM has scaled into cars, phones, PCs, and smart homes and names Geely, Changan, Volkswagen, and Huawei as deep partners, indicating visible device-channel concentration around a small group of brands. SU026
CR001 ModelBest publicly positions itself across phones, AI PCs, smart cockpits, embodied robots, wearables, and legal-workflow use cases rather than as a single-surface model vendor. SR001
CR002 ModelBests homepage and embedded news materials claim scale deployments across automobiles, phones, PCs, and smart homes with named companies including Geely, Changan, Volkswagen, and Huawei. SR001
CR003 ModelBest says its legal-AI offering has been deployed on court intranet infrastructure and serves professional judicial workflows. SR001
CR004 OpenBMB says the community was jointly initiated by Tsinghua NLP Lab and ModelBest, confirming the companys academic-research roots. SR002
CR005 MiniCPM5-1B is explicitly marketed for on-device, local deployment and other resource-constrained scenarios. SR003, SR005
CR006 MiniCPMs code is distributed under Apache License 2.0, which makes reuse and forking legally easier than a closed proprietary stack. SR004, SR005
CR007 MiniCPM-o 4.5 documentation highlights omnimodal full-duplex conversation and real-time speech cases, which expand compliance obligations beyond text outputs. SR006
CR008 Qichacha shows a broad shareholder and investor roster that includes China Telecom Investment, Beijing AI funds, Shenzhen Capital, Hubble-linked capital, Zhipu, and other strategic investors. SR007
CR009 Shenzhen Capital describes itself as mission-driven around discovering and backing major companies, indicating a state-backed investor profile rather than a passive financial sponsor. SR028
CR010 Inovances own positioning as an industrial-automation leader means Huichuan/Inovance-linked capital brings strategic industrial expectations, not only financial return expectations. SR029, SR026
CR011 Chinas Interim Measures require security assessment and algorithm filing for generative-AI services with public-opinion or social-mobilization characteristics and require providers to support regulator review of data sources and algorithms. SR008, SR013
CR012 The March 2025 AI-labeling rules define both explicit labels and implicit metadata labels for AI-generated synthetic content. SR009, SR010, SR015
CR013 The labeling rules require visible labels across text, audio, image, video, and virtual-scene outputs and require downloads or exports to preserve compliant explicit labels. SR009, SR010
CR014 The labeling rules require distribution platforms to propagate labeling or suspected-label warnings and to provide users with declaration functionality when posting AI-generated content. SR009, SR010
CR015 The labeling rules require user agreements to explain labeling methods and require at least six months of log retention when providers supply outputs without explicit labels. SR009, SR010, SR015
CR016 The AI-labeling measures took effect on 2025-09-01. SR009, SR010
CR017 CAC said 72 new generative-AI services were filed and 49 API-based applications or functions were newly registered in March to April 2026, for cumulative totals of 868 filed services and 530 registered applications or functions. SR011
CR018 CAC said launched generative-AI applications or functions should publicly disclose the model name and filing number or launch number they rely on. SR011
CR019 The amended Cybersecurity Law took effect on 2026-01-01, adds explicit AI-governance language, and links network operators personal-information handling to broader privacy-law obligations. SR012
CR020 The amended Cybersecurity Law raises penalties as high as RMB10 million for especially severe consequences and permits suspension, site or app closure, license revocation, or other shutdown-style remedies. SR012
CR021 GB/T 45654-2025 is a generative-AI service safety baseline that supports filing management and security assessment and took effect on 2025-11-01. SR014
CR022 The same TC260 standard says training-data sources with more than 5% illegal or bad information should not be used as training data. SR014
CR023 The same TC260 standard requires consent or another legal basis before using personal information in training data and says user agreements should warn users about intellectual-property risk in generated content. SR014
CR024 BISs May 31, 2026 guidance says a license is required for advanced-computing items when an entity or its ultimate parent is headquartered in Country Group D:5 or Macau, even if the entity itself is located outside those jurisdictions. SR016, SR017, SR019
CR025 BISs June 17, 2026 FAQ says the same guidance applies not only to 3A090.a ICs but also to 4A090 and related .z items, broadening the scope beyond a single chip category. SR017, SR018
CR026 CNBC reported that the May 2026 guidance followed a period when subsidiaries of Chinese AI firms in places such as Malaysia may have been receiving advanced chips, with one supply-chain source estimating volumes in the hundreds of thousands. SR019
CR027 Public analysis says Chinas AI compute response relies on large Huawei-chip clusters and cheap energy or subsidies, but requires more chips and much more power than Nvidia-based systems on some metrics. SR020, SR030
CR028 Li Dahai said the real bottleneck for on-device AI is ecosystem co-building with chips, memory, and bandwidth rather than model size alone. SR025
CR029 The same Tencent interview says ModelBest still needs deep cooperation with chip companies because domestic chip software ecosystems remain materially less complete than Nvidias environment. SR025
CR030 Public deployment proof is concentrated in automotive and device channels: Tencent names Geely Galaxy M9 mass-production cockpit deployment, while Gasgoo lists partnerships with Geely, Volkswagen, Changan, and GAC. SR025, SR027
CR031 Gasgoo says ModelBest claims deployments across automotive, smartphones, AI PCs, and smart homes, but the retained public record does not quantify paying unit economics or renewal terms. SR027
CR032 InfoQ says ModelBests 2026 first-quarter cumulative financing exceeded RMB1 billion and that a recent round was led by Shenzhen Capital and Huichuan-linked industrial capital. SR026
CR033 Caixins April 2024 profile framed commercialization as an open question even after a major financing round, with proceeds earmarked for talent, compute, data, and application landing. SR021
CR034 Caixins May 2025 follow-up said ModelBest had completed three rounds in just over a year and still did not disclose the rounds exact amount or valuation. SR022
CR035 Chinas model price war drastically compressed reference API rates: KrASIA said Alibaba cut Qwen-Long 97% to RMB0.5 per million tokens and Baidu made models free within hours. SR023, SR024
CR036 Digital Watch said the China LLM price war threatens profit margins as competition moves from cloud services into the model APIs themselves. SR024
CR037 KrASIA said hyperscalers can still justify rock-bottom model prices by selling complementary cloud services, but the same price war looks materially harder for smaller startups. SR023
CR038 MiniCPMs public GitHub repos, Apache license, and deployment recipes make ModelBests technical stack highly reproducible, which supports adoption but weakens secrecy-based pricing power. SR003, SR004, SR005, SR006
CR039 The fetched public surfaces reviewed for this chapter did not expose a visible ModelBest CAC filing number, public labeling statement, privacy policy, or user-agreement page, so public compliance disclosure remains unverified. SR001, SR011
CR040 ModelBests public materials place it in sensitive contexts including court-network legal workflows, smart cockpits, phones, and full-duplex voice or video interfaces, so compliance duties likely span personal data, safety, and content labeling simultaneously. SR001, SR006, SR014
CR041 QCC and public funding coverage show a mixed capital base spanning telecom, state, industrial, and peer-AI money, increasing governance complexity and the chance that investors carry strategic objectives beyond short-term financial returns. SR007, SR026, SR028
CR042 QCC lists Zhipu as a shareholder or investor, placing a direct China foundation-model competitor inside ModelBests wider capital structure. SR007
CR043 Public execution scope now extends beyond model R&D into hardware and system tooling: InfoQ says ModelBest plans products such as Pinea Pi and EdgeClaw Box. SR026
CR044 Tencent says low-latency and privacy-sensitive tasks fit on-device, but heavier reasoning remains cloud-side, implying that an edge-first company still carries hybrid-cloud economic exposure. SR025
CR045 Public evidence still does not disclose ModelBest revenue, public API pricing, production contract terms, or customer concentration, so monetization quality remains unverified despite clear financing and partner signals. SR021, SR022, SR026, SR027
CR046 Whether ModelBest has material public litigation, court judgments, or regulator-specific enforcement beyond the cited rulebooks remained unresolved after homepage, QCC, and targeted 2026 litigation/privacy query review.
CR047 Because ModelBest claims court-network legal deployment and TC260 requires lawful data-source and personal-information handling, legal-sector use cases heighten the diligence burden around case-data provenance, personal information, and output reliability. SR001, SR014
CR048 Residual risk remains high until ModelBest can evidence three things with product-specific proof: CAC filing or label compliance, contract-level monetization across named OEM channels, and compute resilience that survives export-control and domestic-toolchain friction. SR011, SR024, SR025, SR026
CR049 OpenBMBs stated goal of standardizing and popularizing big-model tooling helps ecosystem reach but also lowers entry barriers and increases community spillover risk for ModelBests commercial moat. SR002, SR003, SR005
CV001 Multiple April 2026 reports say the latest ModelBest financing pushed the company into the unicorn threshold, effectively a roughly $1 billion valuation floor. SV005, SV006, SV007, SV008
CV002 The reviewed public sources do not disclose ModelBest's exact April 2026 post-money valuation or investor-rights terms. SV005, SV006, SV007, SV008, SV030
CV003 April 2026 coverage says ModelBest's cumulative first-quarter 2026 financing exceeded RMB1 billion. SV006, SV007
CV004 The April 2026 round was widely reported as ModelBest's second financing of 2026 and was led by Shenzhen Capital Group and Huichuan-linked capital. SV005, SV006, SV007, SV030
CV005 A February 2026 round led by China Telecom gave ModelBest a telecom-backed strategic investor before the April financing. SV009
CV006 ThePaper described ModelBest as entering the large-model unicorn threshold at a 10亿美元级别 after the April 2026 round. SV008
CV007 ModelBest's official site positions MiniCPM across phones, AI PCs, smart cockpits, embodied robots, wearables, and legal-AI workflows. SV001
CV008 OpenBMB says the open-source community was jointly initiated by Tsinghua NLP Lab and ModelBest. SV002
CV009 The main MiniCPM GitHub repository frames MiniCPM5-1B as a dense 1B-class model for on-device and local deployment. SV003
CV010 The OpenBMB Hugging Face organization page showed 158 hosted models and 18 collections when fetched for this run. SV004
CV011 Several April 2026 articles say cumulative MiniCPM downloads on GitHub and Hugging Face exceeded 24 million. SV005, SV006, SV007, SV009
CV012 The 24 million download figure is company-reported through press coverage rather than independently audited platform telemetry. SV005, SV006, SV007, SV030
CV013 The reviewed public record discloses no ModelBest revenue, ARR, or customer-count anchor. SV001, SV005, SV006, SV007, SV030
CV014 The reviewed public record discloses no public pricing card or OEM and enterprise contract structure for ModelBest. SV001, SV003, SV030
CV015 CB Insights' public ModelBest page lists the 2026-04-07 round but masks valuation as $XXM and shows revenue as 0 FY undefined on the public page. SV030
CV016 Caixin framed ModelBest's 2024 financing story around the open commercialization question rather than around already-proven monetization. SV010
CV017 Caixin's May 2025 report said ModelBest had completed three financing rounds in a little over a year. SV011
CV018 Tencent's December 2025 profile argued that the automotive cockpit lane is crowded enough that ModelBest must prove commercialization rather than just tell a good story. SV012
CV019 Li Dahai said in June 2026 that edge-model progress is constrained mainly by co-building with chips, memory, bandwidth, and software ecosystems. SV013
CV020 The same June 2026 interview argued that low-latency, privacy-sensitive, and high-frequency tasks are better suited to local edge execution than to pure cloud routing. SV013
CV021 TechCrunch reported that Moonshot raised about $2 billion at a $20 billion valuation and that ARR topped $200 million in April 2026. SV014
CV022 CnTechPost reported that Moonshot's ARR surpassed $100 million in early March 2026 and that its valuation climbed to $18 billion while fundraising. SV015
CV023 Yicai said Zhipu AI closed its Hong Kong debut with a $7.4 billion market capitalization and disclosed H1 2025 revenue of CNY190.9 million with H1 net loss of CNY2.4 billion. SV016
CV024 Reuters' Yahoo Finance syndication said MiniMax raised HK$4.82 billion ($618.6 million) in its Hong Kong IPO. SV017
CV025 TechNode said MiniMax briefly topped $11.5 billion in market value on debut and remained in a high-investment phase with a $512 million loss in the first three quarters of 2025. SV018
CV026 TMTPost reported StepFun's first 2026 pre-IPO tranche at about $4 billion pre-money, a second tranche at $5-6 billion, and an expected cornerstone valuation around $10 billion. SV019
CV027 The Standard reported that StepFun completed a new $2.5 billion funding round and dismantled its red-chip structure to accelerate a Hong Kong IPO push. SV020
CV028 AFP via Tech Xplore reported that DeepSeek was valued at more than $50 billion in its first fundraising round. SV021
CV029 Moneycontrol reported that most DeepSeek investors faced a five-year lock-up and lacked voting rights, highlighting that headline valuation and security terms can diverge sharply. SV022
CV030 C3.ai's June 2026 SEC filing said FY2026 revenue was $250.3 million and GAAP gross margin was 31%. SV023, SV025
CV031 Stock Analysis showed C3.ai at about $1.41 billion market cap, 5.64x price-to-sales, and 3.57x EV/Sales on July 2, 2026. SV024
CV032 Stock Analysis showed Palantir at about $309.97 billion market cap, 59.25x price-to-sales, and 57.76x EV/Sales on July 2, 2026. SV026
CV033 Stock Analysis showed Palantir at $5.22 billion of trailing-twelve-month revenue on July 2, 2026. SV026, SV027
CV034 CB Insights said private AI companies raised $226 billion in Q1 2026 and that $100M+ mega-rounds accounted for 94% of total funding. SV028
CV035 CB Insights said the window for differentiation is narrowing for startups outside the frontier leaders as capital concentrates around the biggest model developers. SV028
CV036 KPMG said Hong Kong led the world in IPO funds raised in Q1 2026 with HK$109.9 billion across 40 completed IPOs. SV029
CV037 KPMG said Hong Kong had six specialist-technology listings and 366 active public IPO applications as of March 31, 2026. SV029
CV038 Reuters' MiniMax IPO coverage said Hong Kong had its strongest IPO year since 2021 in 2025, with $36.5 billion raised from 114 new listings. SV017
CV039 ModelBest's current valuation support appears to come more from strategic scarcity, edge-AI positioning, and investor coalition quality than from public monetization evidence. SV001, SV004, SV005, SV006, SV007, SV009, SV013, SV030
CV040 Relative to Moonshot, StepFun, and DeepSeek, ModelBest is much cheaper in headline dollars but more opaque on revenue and security terms. SV014, SV015, SV019, SV020, SV021, SV022, SV030
CV041 Because no public revenue base exists for ModelBest, scenario and milestone analysis are more appropriate than a precise discounted-cash-flow or single-multiple valuation call. SV013, SV014, SV024, SV026, SV030, SV031
CV042 A $1B-class mark can look fair if ModelBest converts downloads and named deployments into auditable OEM and enterprise revenue over the next 12 to 18 months. SV001, SV003, SV004, SV005, SV006, SV007, SV011, SV013
CV043 Without audited revenue, pricing, gross-margin, and preference disclosures, the same $1B-class mark looks stretched for a 2022-founded startup. SV013, SV014, SV024, SV026, SV030
CV044 A down-round, weak OEM monetization evidence, or a cooling China AI funding and listing window would compress ModelBest's current strategic premium quickly. SV010, SV012, SV013, SV019, SV028, SV029
CV045 The evidence supports research-more rather than buy at the current price: ModelBest looks like an investable company, but not yet like a price-insensitive security. SV002, SV013, SV029, SV030, SV031
CV046 At a $1 billion equity value, matching Palantir's 59.25x sales multiple would require about $16.9 million of annual revenue, while matching C3.ai's 5.64x would require about $177 million. SV024, SV026
CV047 Because ModelBest discloses no revenue, the market cannot tell from public evidence whether it is closer to the Palantir-like upper public band or the C3.ai-like lower public band. SV013, SV024, SV026
CV048 Open-source reach and partner visibility make ModelBest too real to dismiss, but not yet transparent enough to justify a buy call at its current headline mark. SV004, SV011, SV013, SV030, SV031
CV049 In a bull case where audited monetization emerges and China AI scarcity stays bid, ModelBest could reasonably support a roughly $1.3-1.8 billion valuation band. SV006, SV008, SV028, SV029
CV050 In a base case where deployment proof grows but disclosure remains partial, ModelBest looks closer to a roughly $0.9-1.1 billion valuation band. SV001, SV006, SV013, SV030
CV051 In a bear case where monetization disappoints or the market window cools, ModelBest can plausibly re-rate toward roughly $0.6-0.8 billion. SV012, SV013, SV028, SV029, SV030
CV052 A new financing below the current headline mark or a heavy preference stack would be a direct thesis-break for the present valuation case. SV029, SV030, SV031
来源
编号出版方标题引文
SO001 ModelBest 面壁智能 把大模型放到离用户最近的地方。
SO002 QCC 北京面壁智能科技有限责任公司
SO003 Baidu Baike 北京面壁智能科技有限责任公司 北京面壁智能科技有限责任公司于2022年08月12日成立。
SO004 THUNLP Zhiyuan Liu Zhiyuan Liu is an associate professor at the Department of Computer Science and Technology, Tsinghua University.
SO005 OpenBMB open-bmb about us OpenBMB开源社区由清华大学自然语言处理实验室和面壁智能共同支持发起。
SO006 GitHub OpenBMB organization
SO007 GitHub GitHub - OpenBMB/MiniCPM MiniCPM5-1B is a dense 1B Transformer built for on-device, local deployment, and resource-constrained scenarios.
SO008 GitHub GitHub - OpenBMB/MiniCPM-V MiniCPM-V and MiniCPM-o are multimodal LLM series designed for strong performance and efficient deployment on devices.
SO009 Hugging Face openbmb (OpenBMB) OpenBMB (Open Lab for Big Model Base) aims to build foundation models and systems towards AGI.
SO010 Tencent News / 科创板日报 大模型公司面壁智能完成数亿元融资 投后估值迈入独角兽门槛 公司一季度累计融资规模已超10亿元人民币,投后估值迈入独角兽门槛。
SO011 Tencent News / 机器之心 面壁智能完成新一轮融资,26年累计融资超10亿,跻身基模独角兽行列 面壁智能宣布完成新一轮数亿元人民币融资。本轮由深圳市创新投资集团(深创投)和汇川产投联合领投。
SO012 JRJ Finance 清华系独角兽狂奔!面壁智能融资破10亿,MiniCPM下载量2400万:高智能密度模型,正在批量上车长安、吉利 MiniCPM系列开源模型在GitHub、Hugging Face等平台累计下载量突破2400万。
SO013 InfoQ China 面壁智能官宣新一轮数亿元融资,国家队与产业资本同时下注
SO014 Economic Information Daily / Xinhua-affiliated 面壁智能获数亿元新融资 高效能赛道释放商业张力 近日,面壁智能宣布完成新一轮数亿元融资,由中国电信领投,中信金石、中信私募跟投。
SO015 Economic Information Daily 面壁智能成立三周年 公司CEO李大海发出全员信 过去三年,面壁已经快速成长为法律大模型的领先者和“端侧第一”的大模型公司。
SO016 36Kr “面壁智能”完成新一轮数亿元人民币融资
SO017 Gasgoo Seeds | ModelBest Completes New Financing Round of Hundreds of Millions of Yuan Official data shows cumulative downloads for the MiniCPM series have surpassed 24 million across platforms like GitHub and Hugging Face.
SO018 Caixin GPT革命|“清华系”AI公司面壁智能新融数亿元 商业化怎么跑? 4月11日,由知乎CTO李大海和清华计算机系长聘副教授刘知远联合创办的大模型公司面壁智能宣布,于近日完成了新一轮数亿元融资。
SO019 Caixin 大模型创业公司面壁智能获数亿元融资 一年多已融三轮 包括此次融资在内,过去一年面壁智能完成了三轮融资。
SO020 Tencent News 故事好讲,规模化难:面壁智能押注端侧AI“主战场” 需要指出的是,汽车智能座舱所处的赛道,高手如林。
SO021 Tencent News 面壁智能李大海谈端侧模型的瓶颈困局:模型已经够小了,下一步是生态共建 李大海认为,端侧模型当前最大的制约不在模型能力本身,而在于需要和芯片、内存和带宽等共建良性生态。
SO022 Baidu Baike 刘知远 刘知远,北京面壁智能科技有限责任公司联合创始人、首席科学家,清华大学计算机系副教授、博士生导师。
SO023 Baidu Baike 曾国洋 现任面壁智能联合创始人兼首席技术官(CTO)。
SO024 ThePaper / 光子星球 面壁智能,大模型“另类”生存法则 面壁智能的处境可以用“夹缝生存”和“逆流而上”来形容。
SO025 Tencent News 面壁智能开源MiniCPM-o 4.5:实现AI即时自由对话 2月4日,面壁智能正式开源其新一代全模态旗舰模型——MiniCPM-o 4.5。
SO026 Sina Tech / TechWeb 面壁智能完成新一轮数亿元融资,深创投和汇川产投联合领投 MiniCPM 系列开源模型在 GitHub、Hugging Face 等平台累计下载量已突破 2400 万。
SO027 Eastmoney / 科创板日报 syndication 大模型公司面壁智能完成数亿元融资 投后估值迈入独角兽门槛 MiniCPM 系列开源模型已囊括语言模型、全模态模型、多模态模型、语音模型,是国内除阿里以外唯一开源的端侧模型全家桶AI 厂商。
SM001 ModelBest 面壁智能
SM002 OpenBMB / ModelBest GitHub - OpenBMB/MiniCPM: MiniCPM5-1B: A SOTA 1B on-device LLM, small yet powerful.
SM003 36Kr / 创投日报 北京国资领投端侧AI公司
SM004 Gasgoo ModelBest secures fresh funding to accelerate on-device AI deployment
SM005 IDC The Future of Next-Gen AI Smartphones
SM006 Counterpoint Research GenAI Smartphone Share to Rise to 45% of Global Shipments in 2026
SM007 Counterpoint Research Over 25% of Laptop PCs Shipped in 2024 Were GenAI Capable
SM008 Grand View Research On-device AI Market Size And Share | Industry Report, 2030
SM009 BCC Research Global Edge AI Market Size, Share & Growth Trends Analysis
SM010 MarketsandMarkets Edge AI Chip Market Size, Share, Latest Trends & Growth Analysis, 2025-2030
SM011 Counterpoint Research 8 in 10 Wearables to Feature On-Device AI by 2032
SM012 TrendForce Accelerating Vehicle Electrification and Intelligence to Drive Automotive Semiconductor Market to Nearly US$100 Billion by 2029, Says TrendForce
SM013 Cyberspace Administration of China 生成式人工智能服务管理暂行办法
SM014 Ministry of Industry and Information Technology 生成式人工智能服务管理暂行办法
SM015 State Council of China 国务院关于深入实施“人工智能+”行动的意见
SM016 SCIO / Xinhua China unveils guidelines to regulate, boost innovative development of AI agents
SM017 Linklaters Tech Insights China: dual-track AIGC labelling and latest AI regulatory development
SM018 Inside Privacy China Releases New Labeling Requirements for AI-Generated Content
SM019 Cyberspace Administration of China 关于发布生成式人工智能服务已备案信息的公告(2026年3月至4月)
SM020 CNBC / Reuters U.S. takes step to halt Nvidia AI chip shipments to Chinese firms outside China
SM021 CSIS The Limits of Chip Export Controls in Meeting the China Challenge
SM022 KrASIA LLM prices hit rock bottom in China as Alibaba Cloud enters the fray
SM023 Digital Watch Observatory Price war escalates in China as Alibaba and Baidu cut AI costs
SM024 Counterpoint Research China's Smartphone Market Forecast to Drop 5% YoY in 2026
SM025 Counterpoint Research China Smartphone Market in Q1 2026: Huawei Share Highest in 5 years, Apple Fastest Growing
SM026 Counterpoint Research China Smartphone Market Share: Quarterly
SM027 TechNode GenAI smartphone shipments to surpass 400 million in 2025, making up 30% of global market
SM028 CNMO via Tencent News IDC发布2026中国手机市场十大洞察 AI手机占比将过半
SP001 ModelBest 面壁智能 MiniCPM ... 已在汽车、手机、PC 及智能家居等多个领域实现规模化落地,与吉利、长安、大众、华为等多家知名企业达成深度合作。
SP002 OpenBMB GitHub - OpenBMB/MiniCPM MiniCPM5-1B ... built for on-device, local deployment, and resource-constrained scenarios, reaching 1B-class open-source SOTA.
SP003 arXiv MiniCPM: Unveiling the Potential of Small Language Models with Scalable Training Strategies This scenario underscores the importance of exploring the potential of Small Language Models (SLMs) as a resource-efficient alternative.
SP004 OpenBMB GitHub - OpenBMB/MiniCPM-V MiniCPM-V 4.6 ... can be deployed across common mobile platforms, including iOS, Android and HarmonyOS.
SP005 Microsoft Azure Phi Open Models - Small Language Models | Microsoft Azure These models were created to give developers tools to implement AI directly on a device without the need for cloud connectivity. Phi models are open source through the MIT License.
SP006 Microsoft Foundry AI Model Catalog | Microsoft Foundry Models Phi-4-mini-instruct is a lightweight open model ... 3.8B parameters ... supports 128K token context length.
SP007 Microsoft PhiCookBook/md/01.Introduction/01/01.PhiFamily.md at main · microsoft/PhiCookBook Need a solution that can work on edge without internet ... Start with Phi-3 /3.5 -Vision or Phi-4-multimodal.
SP008 Google DeepMind Gemma 4 Maximum compute and memory efficiency ... A new level of intelligence for mobile and IoT devices.
SP009 Google AI for Developers Gemma 4 model overview | Google AI for Developers Gemma models are provided with open weights and permit responsible commercial use.
SP010 Apple Apple Intelligence and Siri Any app can tap into the on-device models that power Apple Intelligence, and the features you build work offline. And it’s all at no cost per request.
SP011 Mistral AI Frontier AI LLMs, assistants, agents, services | Mistral Deploy frontier AI in your environment ... on virtual cloud, edge, or on-premises. Your data stays within your walls.
SP012 Mistral AI Introducing Mistral 3 | Mistral AI For edge and local use cases, we release the Ministral 3 series ... 3B, 8B, and 14B parameters ... all under the Apache 2.0 license.
SP013 Meta llama-models/README.md at main · meta-llama/llama-models Llama models have been downloaded hundreds of millions of times, there are thousands of community projects built on Llama.
SP014 Meta llama-models/models/llama3_2/MODEL_CARD.md at main · meta-llama/llama-models The Llama 3.2 collection ... in 1B and 3B sizes ... Intended Use Cases ... mobile AI powered writing assistants ... quantized models ... on-device use-cases.
SP015 Qwen Team Qwen3: Think Deeper, Act Faster Six dense models are also open-weighted, including Qwen3-32B, Qwen3-14B, Qwen3-8B, Qwen3-4B, Qwen3-1.7B, and Qwen3-0.6B, under Apache 2.0 license.
SP016 Qwen Team GitHub - QwenLM/Qwen3-Coder Qwen3-Coder-Next ... Long-context Capabilities: with native support for 256K tokens, extendable up to 1M tokens using Yarn.
SP017 Qwen Team GitHub - QwenLM/Qwen The source code ... is licensed under the Apache 2.0 License ... Qwen-72B, Qwen-14B, and Qwen-7B are licensed under the Tongyi Qianwen LICENSE AGREEMENT.
SP018 Moonshot AI Kimi K2.6 | Leading Open-Source Model in Coding & Agent Kimi K2.6 is an open-source model featuring SOTA coding, long-horizon execution, and agent swarm capabilities.
SP019 Kimi API Platform Kimi K2.6 - Kimi API Platform Kimi API is fully compatible with OpenAI’s API format.
SP020 MiniMax MiniMax M3 - Coding & Agentic Frontier, 1M Context, Multimodal The first open-weight model with three frontier capabilities.
SP021 MiniMax GitHub - MiniMax-AI/MiniMax-M3 MiniMax-M3 is a native multimodal model with 1M context. It has ~428B parameters and ~23B activated parameters.
SP022 Z.ai Z.ai - Advanced AI Chatbot & Agent powered by GLM-5.2 Z.ai - Advanced AI Chatbot & Agent powered by GLM-5.2
SP023 Zhipu AI GLM-5.1 - 智谱AI开放文档 GLM-5.1 ... 能够在单次任务中持续、自主地工作长达 8 小时。
SP024 Baidu ERNIE Blog ERNIE 5.1 Officially Released! Topping Multiple Leaderboards — A Model That Writes Better and Understands You More ERNIE 5.1 ... using only about 6% of the pre-training cost of comparable models.
SP025 AICPB China AI Rankings by App MAU — Issue 21 (Apr 2026 Edition) China AI Rankings for App are based on App Monthly Active Users (MAU) in Apr 2026.
SP026 OpenAI Introducing ChatGPT We’ve trained a model called ChatGPT which interacts in a conversational way.
SI001 ModelBest 面壁智能 把大模型放到离用户最近的地方;做高效的端侧智能。
SI002 OpenBMB OpenBMB - 让大模型飞入千家万户
SI003 ModelBest Feishu 你好!面壁小钢炮MiniCPM
SI004 ModelBest Feishu MiniCPM-V 2.6 部署指南
SI005 GitHub GitHub - OpenBMB/MiniCPM: MiniCPM5-1B: A SOTA 1B on-device LLM, small yet powerful.
SI006 GitHub GitHub - OpenBMB/MiniCPM-V: A Pocket-Sized MLLM for Ultra-Efficient Image and Video Understanding on Your Phone
SI007 GitHub OpenBMB
SI008 Hugging Face / OpenBMB openbmb (OpenBMB) OpenBMB (Open Lab for Big Model Base) aims to build foundation models and systems towards AGI.
SI009 OpenBMB MiniCPM README
SI010 OpenBMB MiniCPM-V README
SI011 OpenBMB MiniCPM-V-Apps README
SI012 OpenBMB MiniCPM-o README
SI013 OpenBMB MiniCPM-o Demo README VRAM (per Worker, after initialization) ~21.5 GB.
SI014 新浪科技 / TechWeb 面壁智能完成新一轮数亿元融资,深创投和汇川产投联合领投 今年一季度面壁智能累计融资规模预计已超 10 亿元人民币。
SI015 网易 / 瑞财经 面壁智能完成新一轮数亿元融资,CEO李大海称AGI长跑已经进入中段
SI016 网易 / 新京报贝壳财经 面壁智能完成新一轮数亿元融资,深创投入场
SI017 东方财富网 大模型公司面壁智能完成数亿元融资 投后估值迈入独角兽门槛 投后估值迈入独角兽门槛,也正式跻身国内具备全栈自研能力的基座大模型独角兽企业行列。
SI018 腾讯新闻 / 科创板日报 大模型公司面壁智能完成数亿元融资 投后估值迈入独角兽门槛 深创投相关负责人表示,面壁智能在算力约束下持续提升模型智能密度……能够有效降低智能体的部署与调用成本。
SI019 腾讯新闻 / TechWeb 面壁智能完成新一轮融资,26年累计融资超10亿,跻身基模独角兽行列
SI020 新浪看点 面壁智能最新一轮融资的投资方分别是什么背景?
SI021 澎湃新闻 / 光子星球 面壁智能,大模型“另类”生存法则 没有打在身上的聚光灯、没有超高金额的融资,面壁智能咬紧“端侧”方向,接连躲过了算力战、价格战和C端流量战。
SI022 36氪研究院 36氪研究院 | 2025年中国大模型行业发展研究报告 2024年中国大模型市场规模已达294.16亿元,预计到2026年将突破700亿元。
SI023 36氪 2026,国产AI芯片,跨越天堑:从“推理”走向“训练”
SI024 腾讯新闻 从训练到推理:国产GPU重构AI算力成本与产业格局 2026年推理算力在整体AI计算中的占比将达到66%,首次超过训练算力。
SI025 深创投集团 深创投集团
SI026 汇川技术 汇川技术(INOVANCE) - 推进工业文明 共创美好生活
SI027 工业和信息化部 ICP备案管理系统
SE001 ModelBest ModelBest homepage 全球领先的轻量高性能大模型,可有效运行在日常生活中主流消费电子和各类终端上,覆盖不同的芯片和系统平台。
SE002 ModelBest Feishu MiniCPM-V 2.6 deployment guide
SE003 ModelBest Feishu ModelBest latest product feed
SE004 OpenBMB OpenBMB homepage
SE005 GitHub OpenBMB organization page
SE006 GitHub API OpenBMB organization metadata
SE007 GitHub OpenBMB/MiniCPM repository page
SE008 GitHub API OpenBMB/MiniCPM repository metadata
SE009 GitHub OpenBMB/MiniCPM-V repository page
SE010 GitHub API OpenBMB/MiniCPM-V repository metadata
SE011 GitHub OpenBMB/MiniCPM-o repository page Foundation Capability. The full-duplex omni-modality live streaminig capability still needs improvement. Unstable Speech Output in Omni Mode.
SE012 GitHub OpenBMB/MiniCPM-V-Apps repository page
SE013 GitHub API OpenBMB/MiniCPM-V-Apps repository metadata
SE014 GitHub Raw MiniCPM README
SE015 GitHub Raw MiniCPM-V README
SE016 GitHub Raw MiniCPM-V-Apps README
SE017 Hugging Face OpenBMB organization page
SE018 Hugging Face MiniCPM-V-2_6 model card
SE019 Hugging Face MiniCPM-o-2_6 model card
SE020 Hugging Face MiniCPM-V-4_5 model card
SE021 Papers with Code MiniCPM paper page
SE022 arXiv MiniCPM: Unveiling the Potential of Small Language Models with Scalable Training Strategies
SE023 arXiv MiniCPM-o 4.5: Towards Real-Time Full-Duplex Omni-Modal Interaction on Edge Devices
SE024 arXiv MiniCPM-V 4.5: Cooking Efficient MLLMs via Architecture, Data, and Training Paradigm Innovations
SE025 vLLM vLLM supported models documentation
SE026 ModelScope MiniCPM-V-4.6-gguf model page
SU001 ModelBest ModelBest homepage AI+法律=高质量法治和高水平司法全国首个接入法院办案流程的大模型。
SU002 ModelBest ModelBest English site MiniCPM InsidePhones InsideAIPC InsideIntelligent Cabins InsideEmbodied Robots InsideWearable Devices.
SU003 ModelBest ModelBest product page 高效手机 高效AIPC 高效智能座舱 高效具身机器人 高效可穿戴设备。
SU004 OpenBMB OpenBMB homepage
SU005 GitHub OpenBMB/MiniCPM repository MiniCPM5-1B ... built for on-device, local deployment, and resource-constrained scenarios.
SU006 OpenBMB MiniCPM README (raw)
SU007 GitHub OpenBMB/MiniCPM-V repository
SU008 OpenBMB MiniCPM-V README (raw)
SU009 GitHub OpenBMB/MiniCPM-V-Apps repository
SU010 OpenBMB MiniCPM-V-Apps README (raw) This demo runs the MiniCPM-V family of multimodal models fully on-device on iOS, Android, and HarmonyOS NEXT.
SU011 Hugging Face OpenBMB organization page View 11 Spaces models 158 ... openbmb/MiniCPM5-1B ... 322k ... openbmb/MiniCPM-V-4.6 ... 802k.
SU012 Hugging Face MiniCPM-V-4.6-gguf model page This repository hosts the GGUF (llama.cpp) quantized version of MiniCPM-V 4.6 ... public free API key ... most edge-deployment-friendly model to date.
SU013 Ollama openbmb/minicpm5 library page
SU014 Intel OpenVINO MiniCPM-V-2 model enabling with OpenVINO MiniCPM is an End-Side LLM developed by ModelBest Inc. and TsinghuaNLP.
SU015 Intel OpenVINO Supported Devices documentation
SU016 CnTechPost AI firm ModelBest raises multi-million yuan round led by China Telecom MiniCPM series ... has surpassed 24 million cumulative downloads across major platforms, achieving large-scale deployment in automotive, smartphones, AIPC, and smart home sectors.
SU017 Gasgoo Seeds | ModelBest Completes Hundreds of Millions of Yuan Financing, Led by China Telecom Its MiniCPM series of edge-side models are now integrated into vehicles from Geely and Changan Mazda.
SU018 Gasgoo ModelBest secures fresh funding to accelerate on-device AI deployment
SU019 Mazda Mazda Unveils MAZDA EZ-60 New Electric SUV at Auto Shanghai 2025
SU020 arXiv MiniCPM4: Ultra-Efficient LLMs on End Devices
SU021 Pandaily ModelBest Raises Funding, Enters USD 1 Billion+ Foundation Model Unicorn Tier
SU022 Pandaily ModelBest Raises Tens of Millions of Dollars to Advance Edge-Optimized Large Models
SU023 China Daily Brief The Edge of Intelligence: China’s ModelBest Becomes a Unicorn in the Race for On-Device AI
SU024 Sina Tech 断网可用!首款全双工全模态大模型技术报告发布,附一键安装包 大多数用户无法在与大模型产品的交互中获得良好的体验感,甚至由于交流的「时空割裂」逐渐失去耐心。长此以往,大模型在多模态场景的落地无疑大大受阻。
SU025 OpenBMB MiniCPM-o 4.5 Omni Demo Page
SU026 Sina Finance / Securities Times 面壁智能完成数亿元融资 投资方阵容多元化 MiniCPM面壁小钢炮端侧模型已在汽车、手机、PC及智能家居等多个领域实现规模化落地,与吉利、长安、大众、华为等多家知名企业达成深度合作。
SR001 ModelBest 面壁智能
SR002 OpenBMB OpenBMB About Us
SR003 GitHub / OpenBMB GitHub - OpenBMB/MiniCPM
SR004 OpenBMB MiniCPM LICENSE (Apache 2.0)
SR005 OpenBMB MiniCPM README
SR006 OpenBMB MiniCPM-o README
SR007 Qichacha 北京面壁智能科技有限责任公司
SR008 Cyberspace Administration of China 生成式人工智能服务管理暂行办法
SR009 Cyberspace Administration of China 关于印发《人工智能生成合成内容标识办法》的通知
SR010 State Council of China 关于印发《人工智能生成合成内容标识办法》的通知
SR011 Cyberspace Administration of China 关于发布生成式人工智能服务已备案信息的公告(2026年3月至4月)
SR012 State Council of China 全国人民代表大会常务委员会关于修改《中华人民共和国网络安全法》的决定
SR013 Ministry of Industry and Information Technology 生成式人工智能服务管理暂行办法
SR014 TC260 / State Administration for Market Regulation GB/T 45654-2025 网络安全技术 生成式人工智能服务安全基本要求
SR015 China Law Translate Measures for Labeling of AI-Generated Synthetic Content
SR016 Bureau of Industry and Security BIS Homepage / Special Issues: Guidance Regarding Enforcement of License Requirements for Advanced Computing Items for Entities Headquartered in Country Group D:5 and Macau
SR017 Bureau of Industry and Security Guidance Regarding Enforcement of License Requirements for Advanced Computing Items for Entities Headquartered in Country Group D:5 and Macau [May 31, 2026]
SR018 Bureau of Industry and Security Frequently Asked Questions about Guidance Regarding Enforcement of License Requirements for Advanced Computing Items for Entities Headquartered in Country Group D:5 and Macau [Updated June 17, 2026]
SR019 CNBC U.S. takes step to halt Nvidia AI chip shipments to Chinese firms outside China
SR020 CSIS The Limits of Chip Export Controls in Meeting the China Challenge
SR021 Caixin GPT革命|“清华系”AI公司面壁智能新融数亿元 商业化怎么跑?
SR022 Caixin 大模型创业公司面壁智能获数亿元融资 一年多已融三轮
SR023 KrASIA LLM prices hit rock bottom in China as Alibaba Cloud enters the fray
SR024 Digital Watch Observatory Price war escalates in China as Alibaba and Baidu cut AI costs
SR025 Tencent News 面壁智能李大海谈端侧模型的瓶颈困局:模型已经够小了,下一步是生态共建
SR026 InfoQ 面壁智能官宣新一轮数亿元融资,国家队与产业资本同时下注
SR027 Gasgoo Seeds | ModelBest Completes New Financing Round of Hundreds of Millions of Yuan
SR028 Shenzhen Capital Group 深创投集团
SR029 Inovance 汇川技术(INOVANCE) - 推进工业文明 共创美好生活
SR030 CNBC Chinas key weapons in its AI battle with the U.S. — massive Huawei chip clusters and cheap energy
SV001 ModelBest ModelBest official website 把大模型放到离用户最近的地方 / 做高效的端侧智能。
SV002 OpenBMB OpenBMB About Us OpenBMB开源社区由清华大学自然语言处理实验室和面壁智能共同支持发起。
SV003 GitHub / OpenBMB OpenBMB / MiniCPM repository MiniCPM5-1B ... a dense 1B model built for on-device and local deployment.
SV004 Hugging Face / OpenBMB OpenBMB organization page View 18 collections ... models 158.
SV005 Sina Finance ModelBest raises another several-hundred-million RMB and joins the base-model unicorn ranks 其MiniCPM系列开源模型在GitHub、Hugging Face等平台累计下载量已突破2400万。
SV006 QQ News / STAR Market Daily ModelBest completes several-hundred-million RMB financing; post-money enters unicorn threshold 公司一季度累计融资规模已超10亿元人民币,投后估值迈入独角兽门槛。
SV007 Eastmoney / STAR Market Daily ModelBest completes second 2026 financing and crosses unicorn threshold 这是面壁智能2026年完成的第二笔融资,公司一季度累计融资规模已超10亿元人民币,投后估值迈入独角兽门槛。
SV008 The Paper Why ModelBest can become a front-runner in China's second tier of large-model startups 截至今年4月最近的新一轮融资,面壁智能正式迈入大模型独角兽门槛(10亿美元级别)。
SV009 Xinhua / Economic Information Daily ModelBest raises hundreds of millions in February 2026 round led by China Telecom 其MiniCPM系列开源模型在GitHub、Hugging Face等全球主流平台的下载量已突破2400万。
SV010 Caixin GPT Revolution | Tsinghua-linked AI company ModelBest raises new round: how will commercialization run? 商业化怎么跑?
SV011 Caixin ModelBest announces another several-hundred-million RMB financing round 包括此次融资在内,过去一年面壁智能完成了三轮融资。
SV012 Tencent News How can ModelBest break out in the crowded automotive intelligent cockpit market? 故事好讲,规模化不易,面壁智能尚需尽快通过商业化落地成果来展示其实力。
SV013 QQ News How do edge models break compute limits through knowledge density? 端侧模型当前最大的制约不在模型能力本身,而在于需要和芯片、内存和带宽等共建良性生态。
SV014 TechCrunch China's Moonshot AI raises $2B at $20B valuation as demand for open-source AI skyrockets Moonshot AI ... has raised about $2 billion at a valuation of $20 billion ... annual recurring revenue topped $200 million in April.
SV015 CnTechPost Kimi maker Moonshot's annual recurring revenue tops $100 million after K2.5 launch Moonshot's ARR surpassed $100 million ... latest valuation has climbed to $18 billion.
SV016 Yicai Global Zhipu AI Soars in Hong Kong Stock Market Debut as Chinese Startup Becomes World's First LLM Firm to Go Public bringing its market capitalization to HKD57.5 billion (USD7.4 billion).
SV017 Reuters / Yahoo Finance China's AI startup MiniMax Group raises $619 million in Hong Kong IPO it raised HK$4.82 billion ($618.60 million) in its Hong Kong initial public offering.
SV018 TechNode MiHoYo-backed AI firm MiniMax jumps on Hong Kong debut, market value tops $11.5 billion briefly pushing the company’s market capitalisation above HK$90 billion ($11.5 billion).
SV019 TMTPost / NextFin StepFun to complete pre-IPO financing by April pre-investment valuation of approximately 4 billion dollars ... 5 billion to 6 billion U.S. dollars ... cornerstone valuation of around 10 billion dollars.
SV020 The Standard Stepfun finishes new US$2.5b funding round for HK IPO completed a new US$2.5 billion funding round and dismantled its red-chip structure, accelerating its push toward a Hong Kong initial public offering.
SV021 Tech Xplore / AFP DeepSeek valued at more than $50 bn after funding round: reports Investors have valued Chinese artificial intelligence startup DeepSeek at more than $50 billion.
SV022 Moneycontrol DeepSeek raises $7.4 billion at over $50 billion valuation, founder keeps control: Report Investors are subject to a five-year lock-up and will not have voting rights.
SV023 U.S. Securities and Exchange Commission C3 AI Announces Fiscal Fourth Quarter and Full Fiscal Year 2026 Results (EX-99.1) Total Revenue was $250.3 million ... Gross Profit ... 31% gross margin.
SV024 Stock Analysis C3.ai (AI) Statistics & Valuation market cap or net worth of $1.41 billion ... PS Ratio 5.64 ... EV / Sales 3.57.
SV025 Stock Analysis C3.ai (AI) Revenue 2019-2026 In the fiscal year ending April 30, 2026, C3.ai had annual revenue of $250.27M.
SV026 Stock Analysis Palantir Technologies (PLTR) Statistics & Valuation market cap or net worth of $309.97 billion ... PS Ratio 59.25 ... EV / Sales 57.76.
SV027 Stock Analysis Palantir Technologies (PLTR) Revenue 2018-2026 revenue in the last twelve months to $5.22B.
SV028 CB Insights State of AI Q1'26 Report Private AI companies raised $226B in Q1'26 ... $100M+ deals accounted for 94% of total funding.
SV029 KPMG Chinese Mainland and Hong Kong IPO Markets: 2026 Q1 review Hong Kong claimed the top spot globally in terms of funds raised ... HKD109.9B ... 40 completed IPOs.
SV030 CB Insights ModelBest Stock Price, Funding, Valuation, Revenue & Financial Statements 4/7/2026 ... Valuation $XXM ... Revenue 0 FY undefined.
SV031 Hong Kong Exchanges and Clearing Limited (HKEXnews) MiniMax Group Inc. Global Offering Prospectus MiniMax Group Inc. ... incorporated in the Cayman Islands with a weighted voting rights structure.