初创公司尽调
尽调报告 infrastructure / devtools series-a 2026-06-04

Liquid AI

源自 MIT、获 AMD 支持的液态神经网络先行者,Series A 估值 $2.3B

Liquid AI 拥有差异化的边缘部署技术和可信的战略资本背书,但 $2.3B 私募估值仍跑在公开商业披露之前。

封面要素

Series A 融资额 01
250 USD M [CO020]
最新公开估值 02
2300 USD M [CO036]
已披露累计融资 03
296.6 USD M [CO022]
产品栈 05
LFMs + LEAP + Apollo [CO011, CO026]
成立时间 06
2023 [CO001]

公司概况

Liquid AI 是 2023 年从 MIT CSAIL 孵化出来的公司,正在把液态神经网络及相关高效序列模型架构的研究脉络商业化。公司把自己放在以 Transformer 优先的实验室之外:它构建多模态 Liquid Foundation Models(LFMs),并配套 LEAP 这套定制与部署平台,主攻隐私、延迟和算力效率更关键的边缘、本地部署和混合环境。 公开证据显示,公司从隐身发布到战略合作推进很快。Liquid 在 2023 年 12 月宣布 $46.6M 种子轮,随后在 2024 年 12 月完成 AMD 领投的 $250M Series A。此后,公司围绕 AMD Ryzen 和 Ryzen AI 处理器的硬件优化、2026 年 3 月与 Insilico Medicine 合作开发私有药物发现模型、以及 2026 年 4 月面向车内量产部署的 Mercedes-Benz 合作,持续扩展商业叙事。核心待解问题不是科学是否成立,而是公司能否把合作伙伴牵头的验证转化为可复制、可披露的商业规模。

官网
www.liquid.ai
成立时间
2023-01-01
创始人
Ramin Hasani, Mathias Lechner, Alexander Amini, Daniela Rus
创立地点
Cambridge, Massachusetts
总部
Cambridge, Massachusetts
产品
Liquid 销售的产品栈围绕 Liquid Foundation Models、用于模型定制和部署的 LEAP,以及作为本地消费者演示界面的 Apollo 搭建。公开产品叙事强调高效多模态推理、可下载检查点、企业本地许可,以及跨 CPU、GPU 和 NPU 硬件部署,而不是依赖第一方托管 API。
客户
汽车、金融服务、电商、生物技术、消费电子以及其他对延迟、隐私和部署敏感的环境中的企业与开发者。
商业模式
通过 LEAP 以及本地 / 私有化访问,对 LFMs 做销售驱动的企业许可、定制和部署;面向小型开发者开放 / 可下载分发,但超过公开收入门槛后需协商商业许可。
阶段
series-a
融资情况
$46.6M 种子轮(2023 年 12 月)加上 AMD 领投的 $250M Series A(2024 年 12 月);一份创始人传记后来提到 $2.3B 估值。
[CO001, CO002, CO003, CO004, CO005, CO006, CO007, CO008]

执行摘要

主要优势

  • 液态神经网络架构差异化明确,并且确实来自 MIT CSAIL 的研究脉络。
  • AMD 参与融资并与硬件路线对齐,给这家早期模型公司带来罕见的强战略验证。
  • 产品栈已经覆盖 LFMs、LEAP 和 Apollo,面向边缘、本地和混合部署场景。

主要风险

  • 收入、客户数、留存和毛利率披露仍未出现在公开记录中。
  • 公开客户证明仍集中在少数具名标杆合作上。
  • 对一家生产规模和商业化只露出一部分的私营公司,$2.3B 估值显得激进。

未决问题

  • 公开资料没有收入、ARR、毛利率、烧钱速度或现金跑道披露。
  • 公开客户广度、续约行为和集中度仍无法量化。
  • 董事会 / 控制权格局和股权结构经济条款仍只露出一部分。
  • Mercedes、Insilico 等具名合作仍需要披露生产规模 KPI。

目录

Chapter 01

01公司概况

1.1 身份、产品栈与商业姿态

Liquid AI 更应被理解为一家高效基础模型公司,而不是 API 优先的前沿模型实验室。公司概况、模型、企业解决方案和定价页面都反复把业务放在高性能通用 AI 系统上:这些系统可在设备端、本地或混合环境中运行,适用于延迟、隐私、安全和算力效率比通用云规模更重要的场景。如今公开可见的产品栈也强化了这一定位:Liquid Foundation Models 是模型家族,LEAP 是定制与部署平台,Apollo 则是面向消费者的本地 playground。商业姿态同样有辨识度。Liquid 明确表示目前不运营自己的托管 API,而是把用户导向 playground 访问、OpenRouter、Hugging Face 下载,以及通过 LEAP 和直销完成企业定制。这意味着公司卖的不只是 token 消耗,而是模型架构、部署工具和企业支持。贯穿官方垂直页面的承诺一致:更小内存占用、更低延迟、本地数据控制,以及在 CPU、GPU 和 NPU 上部署,而不是依赖一个中心化 serving 栈。[CO001, CO002, CO003, CO008, CO009, CO010]

KPI 快照表
指标数值 / 状态日期 / 期间置信度缺口 / 备注
成立时间2023历史官方发布和多个数据库都指向 2023 年成立。
证据最充分的总部马萨诸塞州 Cambridge当前CB Insights、Built In 和 PitchBook 指向 Cambridge;其他来源更宽泛地提到 Boston 或 Brookline。
当前阶段Series A 轮 / 私营当前PitchBook、CB Insights 和 Tracxn 都把公司列为 AMD 领投轮之后的 Series A 阶段。
核心商业模式高效基础模型加部署平台当前官方概览、企业和定价页面指向模型授权、定制与部署,而不是单纯转售 API。
最新披露的新股融资2502024-12官方融资公告和多家媒体 / 数据库来源都指向一轮 US$250M Series A。
已披露累计融资296.6截至 2024-12由官方 US$46.6M 种子轮加官方 US$250M Series A 推导;Tracxn 将其四舍五入为 US$297M。
最新公开估值>20002024-12独立报道称超过 US$2B;Tracxn 显示正好 US$2B;一份创始人简介后来引用 US$2.3B。
员工数当前公开标记明显冲突:PitchBook 为 49 名员工,Hugging Face 为 81 名团队成员,Tracxn 为 121 名员工。
客户数没有留存的公开来源给出 Liquid 自身的标准客户数或部署数。
收入 / ARR没有留存的公开来源披露收入、ARR 或运行率。
托管 API没有自有托管 API当前官方定价页面称,Liquid 目前不提供自有托管 API。

空值表示公开披露不支持,而不是数值为零。融资数值以 USD million 计。估值行保留公开区间,因为各来源没有收敛到一个精确投后数。

[CO001, CO008, CO010, CO015, CO017, CO020]
FO002: 公司快照逻辑

Liquid 围绕边缘和本地部署的高效 AI,把源自研究的架构、部署工具、战略资本和隐私优先商业化拼在一起。

[CO003, CO010, CO011, CO020, CO024, CO026]

1.2 创始人、技术谱系与治理可见度

创始人故事证据扎实,治理可见度则没有同等扎实。官方发布材料、TechCrunch 和公司数据库都把 Liquid AI 标记为 2023 年 MIT 相关孵化公司,由 Ramin Hasani、Mathias Lechner、Alexander Amini 和 Daniela Rus 领导,其中 Hasani 任 CEO,Lechner 任 CTO,Amini 任首席科学官。创始叙事重要,因为 Liquid 商业化的不是通用开源模型,而是一条扎根于液态神经网络和 liquid time-constant networks 的特定研究脉络。arXiv 上的 LTC 论文和 Liquid 自身研究页面,把这家创业公司直接连到公司成立前已经发展的连续时间和状态空间工作。公开记录里的治理信息弱得多。Tracxn 给出六人董事会名单,但 Liquid 自身没有发布当前董事会或治理页面,留存公开来源也没有清楚描述 Series A 后的控制权。因此,投资人仍需追问投资人监督、董事会独立性,以及围绕创始科学家团队的关键人集中度。[CO002, CO003, CO004, CO005, CO006, CO007]

领导层与创始人表
人物职务背景 / 公开锚点创始人—市场匹配或职能覆盖关键人物依赖
Ramin Hasani联合创始人兼 CEO官方发布材料、TechCrunch 和 Tracxn 均确认 Hasani 为 CEO;TechCrunch 提到他此前在 Vanguard 和 MIT CSAIL 工作。把核心液态神经网络研究接到商业产品策略和企业叙事上。很高;他是融资、合作伙伴和产品发布中的主要公开发言人。
Mathias Lechner联合创始人兼 CTO官方发布材料和他的个人页面确认 Lechner 为 CTO;他的学术履历横跨 TU Wien、ISTA 和 MIT CSAIL。为产品栈提供架构深度、系统可信度和边缘效率重点。高;他对技术差异化很关键,并且偶尔披露其他地方没有重复的战略细节。
Alexander Amini联合创始人兼首席科学官官方发布材料、TechCrunch 和 Tracxn 确认 Amini 为 CSO 和创始科学家。负责科学可信度、多模态与研究深度,以及论文和产品主张之间的连接。高;科学路线图仍与创始研究团队紧密绑定。
Daniela Rus联合创始人;MIT CSAIL 主任官方发布材料和 TechCrunch 将 Rus 描述为联合创始人,也是这次从 MIT CSAIL 孵化成独立公司背后的重要人物。带来机构可信度、机器人学积累,以及投资人与合作伙伴的外部信任。中;战略可信度很高,但日常经营控制似乎由 Hasani、Lechner 和 Amini 掌握。

这是创始人与领导层视角,不是完整高管名单。最大的公开缺口仍是创始科学团队之外的正式治理结构。

[CO002, CO003, CO004, CO005, CO006, CO007]
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公开证据尽调问题
OSS Capital种子轮领投方锚定了 2023 年启动融资,仍是标准的早期支持者信号。官方第一性原理发布文章;TechCrunch;Tracxn 融资页面。确认当前持股、董事会权利,以及 Series A 中是否有按比例跟投。
PagsGroup / Stephen Pagliuca种子轮领投方和持续战略支持方传递高知名度金融赞助信号,并可能从创立之初就影响治理。官方发布文章;Tracxn 融资页面。确认 Pagliuca 或关联方在 Series A 后是否保留董事会或观察员权利。
AMDSeries A 领投方和战略硬件伙伴最显眼的后期资本提供方,也是商业相关的部署伙伴。官方融资博客;TechCrunch 2024 融资报道;AMD 官方新闻稿。澄清持股比例、排他性,以及硬件优化是否附带商业化承诺。
G42商业企业合作伙伴暗示美国创业公司渠道之外还有主权和本地 AI 需求。2025 年 6 月合作的官方新闻室摘要。要求披露交易经济性、区域排他性和收入贡献。
Shopify公开材料点名的战略合作伙伴说明商业用例相关性和潜在设计伙伴身份。官方新闻室摘要;创始人简介冲突注记;外部新闻报道。确认 Shopify 是投资人、客户、合作伙伴,还是三者兼具,以及已签约的量级。
开放权重开发者社区分发渠道,而非股权所有者Hugging Face、文档和社区材料显示,这条采用漏斗可能影响未来企业转化。官方社区和创业公司页面;Hugging Face 组织;文档。要求提供从社区到付费企业账户的下载、微调和转化指标。

这张图混合了股权利益相关方和经济上重要的商业渠道,因为 Liquid 的公开故事把资本和分发绑在一起。它不重建完整股权结构表、SAFE 或优先股堆叠。

[CO017, CO018, CO019, CO020, CO023, CO024]

1.3 融资历史、规模标记与仍缺什么

Liquid AI 的资本历史足以支撑本章判断,但运营规模不够透明。官方发布材料称,2023 年 12 月公司完成由 OSS Capital 和 PagsGroup 领投的 US$46.6 million 种子轮;官方 2024 年融资材料随后宣布 AMD 领投 US$250 million Series A。TechCrunch、Tech Funding News 和 Tracxn 大体印证了轮次规模,但估值标记方向一致、细节并不完全相同:独立报道称 Series A 对公司的估值超过 US$2 billion,Tracxn 四舍五入到 US$2 billion,Mathias Lechner 2025 年 7 月的个人传记则提到 US$2.3 billion 估值和更广的战略投资方阵容。合理推论是,Liquid 是一家从后期种子到 Series A 阶段、已达独角兽规模的公司,拥有足够资本推进产品化;但准确投后估值和投资人分配仍是尽调事项。公开规模指标明显更薄。PitchBook、Hugging Face 和 Tracxn 的员工数信号互相冲突,收入、ARR、客户数或部署量仍没有干净的公开披露。遗漏很关键:公司已经可见地商业化、招聘 GTM 和财务岗位、宣布重大合作,但从技术承诺到可复制企业收入的变现转化仍不透明。[CO017, CO018, CO019, CO020, CO021, CO022]

FO003: 资本与披露质量 KPI

Liquid 的融资标记相对清晰,但精确估值、员工数和商业披露质量在公开来源中仍不均衡。

估值和员工数按公开区间展示,因为保留来源相互冲突;收入披露是缺失,不是为零。

[CO020, CO021, CO022, CO023, CO036, CO037]

1.4 记录在案的里程碑、合作牵引与公开风险信号

里程碑记录显示,公司从研究源头叙事推进到垂直商业化的速度异常快。Liquid 在 2023 年 12 月带着种子轮走出隐身状态,2024 年 10 月 MIT 发布活动前后公开首批产品,随后在 2025 年和 2026 年初借 LEAP 和 Apollo、G42 商业合作、AMD 笔记本支持,以及与 Insilico Medicine 和 Mercedes-Benz 的具名合作继续加速。这些里程碑说明,公司试图在对部署敏感的行业里取胜:更小、更私有、可本地运行的模型有机会打败依赖云的默认方案。它们也显示 Liquid 同时扩展软件和硬件关系,而不是押注单一 GTM 渠道。留存材料中最主要的反向公开信号不是丑闻,而是产品成熟度:Constellation Research 2024 年的笔记称首批 LFMs 仍在打磨中,并在零样本代码、时效性信息和人类偏好优化上明显偏弱。2026 年又出现了一个更商业化的约束:Liquid 发布的开放许可对研究和小公司仍然慷慨,但要求收入超过 US$10 million 的企业购买商业许可。合起来看,公司有真实动能,但仍在把技术差异化转换为持久市场标准。[CO025, CO026, CO027, CO028, CO029, CO030]

里程碑表
日期事件类型金额 / 估值 / 状态参与方含义
2020-12-14Liquid Time-constant Networks 论文形成已接收的公开版本产品AAAI 2021 接收论文谱系作者:Hasani;Lechner;Amini;Rus;Grosu奠定后来创业公司投资逻辑的技术根基。
2023-12-06Liquid AI 走出隐身状态并宣布种子轮融资创立US$46.6M 种子轮参与方:Liquid AI;OSS Capital;PagsGroup把 MIT CSAIL 研究谱系变成一家已融资的独立公司。
2024-10-23围绕 MIT 活动的首次公开产品发布产品LFM 发布活动Liquid AI;MIT Kresge 参与者公司从隐身科学叙事进入公开产品类别创建。
2024-12-13宣布 Series A融资US$250M;估值约为或高于 US$2BAMD;Liquid AI为商业化提供资本和战略硬件协同。
2025-06-17宣布 G42 商业合作合作企业商业化合作G42;Liquid AI传递核心美国创业公司渠道之外的主权和私有 AI 需求信号。
2025-07-15LEAP 和 Apollo 上线产品开发者平台和消费者应用上线Liquid AI从模型供应商扩展为工具和边缘部署工作流提供方。
2025-08-18LEAP 增加 AMD Ryzen 和 Ryzen AI 支持合作原生支持 AMD 笔记本电脑Liquid AI;AMD加深端侧 AI 的硬件优化逻辑。
2025-11-13宣布 Shopify 合作合作面向商业用例的低于 20ms 基础模型Shopify;Liquid AI为商业部署场景提供标杆验证。
2026-03-03宣布 Insilico Medicine 科学合作合作覆盖药物发现任务的 2.6B 科学模型Liquid AI;Insilico Medicine显示公司正在纵向切入制药和私有科学基础设施。
2026-04-23宣布 Mercedes-Benz 嵌入式车载 AI 合作合作目标在 H2 2026 完成首次生产部署Mercedes-Benz;Liquid AI创造了从模型效率走向物理世界部署的最清晰公开路径之一。
2026-04-28LFM Open License 更新商业门槛反向收入超过 US$10M 后不再免费商用Liquid AI改善变现控制,但可能增加成长型创业公司和中端市场采用方的摩擦。

这是本章有日期的记录性时间线。日期采用留存材料中最清晰的公开发布标记,可能反映公告时间,而不是内部产品完成日期。

[CO017, CO020, CO025, CO026, CO027, CO028]
FO001: 公司里程碑时间线

2023 年末到 2026 年初,Liquid AI 从有研究根基的拆分创业公司,转成融资密集、以部署为中心的创业公司。

时间线采用保留来源的公告日期;发布时间可能滞后于内部开发或商业测试启动。

[CO025, CO026, CO027, CO028, CO029, CO030]

1.5 图表

Chapter 02

02市场分析

2.1 市场边界与替代集合

如果把完整基础模型市场定义为所有从托管 API 购买 token 的组织,Liquid AI 并不整齐地落在这个市场里。它自己的企业、解决方案、定价、创业公司和社区页面指向一个更窄、也更偏运营的范畴:在数据本地性、延迟、隐私、安全或硬件约束重要的地方部署高效 AI。汽车页面强调在受限硬件上运行的车内智能;电商页面聚焦搜索、智能体店铺前端工作流,以及相对零售利润率的成本;金融服务页面强调 PII 密集和合规敏感工作流;创业公司或社区页面则销售小模型、文档、工具和导师支持的直接访问。这意味着市场边界围绕可部署、可定制、具备边缘意识的基础模型,而不是通用云 LLM 访问。替代集合也因此改变。OpenAI、Google、Anthropic、Writer 和 xAI 是便利型买家的 API 优先替代;Cohere、Mistral、Microsoft Phi、Meta Llama、Qualcomm 以及其他设备端或自托管方案,则更直接地在部署控制上竞争。因此,Liquid 不做托管 API 的姿态会收窄一部分需求,却强化了它在私有和嵌入式工作流中的契合度。[CM001, CM002, CM003, CM004, CM005, CM006]

市场定义表
细分 / 类别纳入支出排除支出买方 / 付款方与 Liquid 的相关性
私有且可部署的基础模型工作流面向本地或混合 AI 的模型授权、定制、部署工具、支持和硬件适配工作没有部署或数据边界决策的纯消费者聊天机器人使用企业产品、IT、数据或运营预算对 Liquid 实际市场边界最接近的高层描述
汽车座舱与嵌入式 AI车载助手、推理、多模态交互和边缘推理层与车辆系统无关的通用云聊天OEM 软件和车型项目预算直接体现在 Liquid 的汽车定位和 Mercedes 合作中
电商搜索与智能体店面工作流搜索、推荐、购物车与结账编排、商家智能助手没有店面集成的通用营销文案生成数字商务、CX 和商品运营预算匹配 Liquid 的电商页面和商业伙伴叙事
私有基础设施上的金融服务 AI需要数据控制的反欺诈、支付、交易、客服和知识工作流没有合规负担的低敏感度通用生产力用途CIO、运营、风险和转型预算匹配 Liquid 的金融服务页面和主权 / 隐私定位
创业公司与开发者采用漏斗模型实验、微调、文档、导师支持和平台使用,可转化为付费部署没有商业意图的休闲爱好者使用创始人、工程或产品预算重要获客线索渠道,但不等同于企业收入
托管公共 API 支出来自 OpenAI 或 Google 等供应商的按 token 计费云模型访问本地和端侧部署经济性开发者或产品云支出预算更像相邻或替代市场,而不是 Liquid 当前最清晰的适配方向,因为 Liquid 没有运营自有托管 API

这张表定义的是 Liquid 的实际市场边界,而不是整个 AI 市场。纳入和排除支出是分析类别,不是经审计的收入池。

[CM001, CM002, CM003, CM004, CM005, CM006]
FM001: 市场规模测算视角

Liquid 的机会不是宽泛的边缘 AI 大盘,而是更小的一组私有、低延迟关键部署,且要匹配其产品姿态。

金字塔本身是概念性而非数字化的;它把市场边界从公开外层 TAM 估算收窄到 Liquid 更现实的可服务细分。

[CM001, CM002, CM008, CM021, CM039]

2.2 规模测算视角与需求边界

公开证据支持一个很大的机会边界,但不足以得出唯一权威 TAM。Deloitte 2026 年企业 AI 调查显示,需求正从试点走向规模化生产:2025 年员工访问量上升 50%,至少 40% 项目进入生产的公司占比将在六个月内翻倍,58% 受访公司已使用物理 AI,且 80% 预计两年内使用。这些采用信号重要,因为 Liquid 的产品故事绑定私有、本地和物理部署,而不只是网页聊天。市场规模报告支持同一方向的判断,但数量级分歧很大。Fortune Business Insights 和 Verified Market Reports 都把边缘 AI 市场估为 2026 年 US$47.59 billion、2034 年 US$385.89 billion;Stratistics MRC 则仅把边缘 AI 推理估为 2026 年 US$153.84 billion、2034 年 US$635.51 billion。这个差距太大,不能当作四舍五入噪音忽略;它说明这些发布方在衡量不同市场边界。尽调时,正确姿态是把公开估计当作外部边界,并把 Liquid 的可服务市场看作一个更窄子集,绑定受监管、延迟关键或设备受限的部署。[CM012, CM013, CM014, CM015, CM016, CM017]

TAM、SAM 与规模测算视角表
视角 / 发布方年份地理范围数值增长 / 信号方法论 / 局限
边缘 AI 市场 - Fortune Business Insights2026全球47.59到 2034 年 CAGR 为 29.9%广义边缘 AI 市场,包含多个组件和行业;仅用于外层 TAM 视角
边缘 AI 市场 - Verified Market Reports2026全球47.59到 2034 年 CAGR 为 29.9%聚合行业数据集和贸易分析;标题数字相近,但仍是广义边缘市场
边缘 AI 市场 - Fortune Business Insights2034全球385.89预测可用作长期上限,但不是 Liquid 专属 SAM
边缘 AI 推理市场 - Stratistics MRC2026全球153.84到 2034 年 CAGR 为 19.4%以推理为中心的定义明显大于广义边缘 AI 估计
边缘 AI 推理市场 - Stratistics MRC2034全球635.51预测如果推理密集定义占主导,显示上行边界
企业扩展信号 - Deloitte2026全球调查2x生产中项目占比 >=40% 的公司将在六个月内翻倍采用信号,不是美元口径 TAM;有助于证明需求时点,而不是市场规模

除 Deloitte 的 2x 生产扩展信号外,美元数值以 USD billions 计。表格有意保留发布方估计之间的矛盾,而不是强行给出一个虚假的单一 TAM 数字。

[CM012, CM013, CM014, CM015, CM016, CM017]
FM002: 市场估算区间

公开的 2026 年和 2034 年市场估算差异很大,取决于发布方衡量的是广义边缘 AI,还是专门的边缘 AI 推理。

中点只是用来可视化公开来源差距的分析占位符,不是独立来源的市场估算。

[CM015, CM016, CM017, CM041]

2.3 买方、用户、付款方与采用路径

Liquid 的公开垂直页面暗示了几种不同买方动作,而不是一种通用企业销售。在汽车领域,主要买方是想提升车内助手、同时守住延迟、隐私和硬件限制的 OEM 软件或信息娱乐团队;用户是驾驶员或乘客;付款方是车辆项目或软件平台预算。在电商领域,买方是数字产品、搜索、商品运营或客户体验团队;用户是购物者和品牌运营者;预算来自电商、增长或客户支持条线。在金融服务领域,潜在买方是 CIO、数据与 AI 平台、风控或运营团队;用户是欺诈、支付、服务和交易工作流里的员工与客户;付款逻辑落在转型、合规或运营预算。创业公司和开发者社区用户形成另一条路径:买方是技术创始人或产品工程师,付款方是风险资金支持的产品预算。所有路径都不是瞬时采用。Liquid 必须先赢下基准测试或概念验证,再完成硬件适配和模型定制,再通过安全或治理审查,最后才进入部署或支持合同。[CM021, CM022, CM023, CM029, CM030, CM031]

细分 / 买方地图
细分主要买方主要用户付款方 / 预算所有者采用触发因素销售路径
汽车 OEM软件定义汽车、信息娱乐或语音栈团队驾驶员和乘客车型项目或软件平台预算需要在受限硬件上运行私有、低延迟的车内智能长企业周期,包含硬件验证和量产集成
电商运营方数字产品、搜索、商品运营或 CX 团队购物者和商家运营人员商务、增长或支持预算需要改善搜索、推荐或智能体结账,同时不破坏利润率从店面 API 试点开始,再扩展到更广的工作流集成
金融服务企业CIO、AI 平台、风险或运营负责人高敏感度工作流中的员工和终端客户转型、合规或运营预算围绕受监管数据,需要隐私、延迟和本地控制先证明控制与治理,再推进生产推广
通用企业端点部署IT、安全或知识平台团队在笔记本电脑或端点上使用本地智能助手的员工工作场所生产力或端点预算需要离线韧性、低延迟或降低对云的依赖基准测试、端点试点、治理审查,然后托管部署
创业公司和开发者技术创始人、开发者或平台工程师内部产品团队和终端应用用户产品预算或风投资助的工程预算需要用小模型和更快本地部署做出差异化自下而上的实验,之后可能转化为付费平台或授权

买方和付款方不一定是同一个人;这张表根据 Liquid 自己的解决方案页面和相邻市场证据,映射各细分最可能的公开买方路径。

[CM006, CM021, CM029, CM030, CM031, CM032]
FM003: 买方 / 细分市场地图

Liquid 的买方细分沿两条关键轴线分化:隐私或主权敏感度,以及部署复杂度。

这些单元格是根据保留的公开证据做出的分析性定位,不是 Liquid 直接披露的客户分群数据。

[CM021, CM029, CM030, CM031, CM032, CM033]
FM004: 采用漏斗 / 价值链地图

Liquid 要抓住市场,必须把买方从对高效模型的好奇推进到受治理的生产部署。

漏斗是根据 Liquid 公开材料和相邻企业 AI 证据综合出的通用企业采用路径。

[CM034, CM035, CM036, CM038, CM040]

2.4 增长驱动与采用约束

证据指向 Liquid 所处品类的真实结构性顺风。公开市场报告反复提到低延迟、隐私、数据主权、IoT 普及、5G、专用边缘芯片,以及自主或物理系统这些增长驱动。Wevolver 2026 年边缘报告提供了更技术化的视角:有用的设备端模型越来越落在一个不大不小的黄金区间,从十亿以下到个位数十亿参数不等,这与 Liquid 对小型多模态模型、设备端 TTFT,以及跨笔记本、嵌入式设备和企业端点部署的强调一致。但同一批来源也指出了真实摩擦。Deloitte 把技能缺口称为 AI 集成最大障碍。Stratistics 标记了跨大量边缘设备部署和维护的复杂性。Liquid 自身的定价和许可结构又增加了一个约束:它没有开箱即用的托管 API,买方必须接受一定部署负担;一旦客户收入超过 US$10 million,开放许可就依赖商业许可。因此,当总部署经济性、隐私和韧性最重要时,Liquid 更有吸引力;当便利性和即时消耗主导采购决策时,它的吸引力会下降。[CM021, CM022, CM023, CM024, CM025, CM026]

增长驱动因素与约束表
因素类型证据时点含义 / 尽调问题
AI 试点向生产扩展驱动因素Deloitte 称,2025 年员工 AI 访问量增长 50%,高产出部署预计六个月内翻倍当前支持近期企业需求,但 Liquid 仍需证明自己能拿下生产预算
物理 AI 扩张驱动因素Deloitte 称,58% 的公司已经使用物理 AI,80% 预计两年内会采用当前至中期有利于 Liquid 的汽车和嵌入式定位
延迟、隐私与主权需求驱动因素Liquid 页面和 VMR 都强调本地控制与实时响应当前买家可能因此更偏好可部署模型,而不是托管 API
小模型效率趋势驱动因素Wevolver 和 Liquid 社区材料都指向小型、高效、端侧模型,将其视为 2026 年设计重心当前如果基准测试说法经得起客户评估,Liquid 会受益
AI 技能缺口约束因素Deloitte 将员工技能不足列为最大集成障碍当前可能拉长企业部署周期,并增加解决方案工程负担
部署与维护复杂约束因素Stratistics 将异构边缘设备的部署和维护列为威胁持续会抬高支持成本,并使实际 SAM 低于自上而下测算的 TAM
托管 API 便利性约束因素OpenAI、Google、Anthropic、Writer 和 xAI 提供更容易使用的托管路径当前Liquid 必须靠部署经济性和控制权取胜,而不是即时 API 便利性
收入超过 US$10M 后的商业许可门槛约束因素Liquid 的开放许可证在超过该门槛后不再免费商用当前可能让增长顺利的初创公司和中端市场软件买家采用时多一层摩擦
相互矛盾的市场报告约束因素2026 年公开估计从 US$47.59B 到 US$153.84B 不等当前后续估值工作需要敏感性分析,而不是只看一个标题式 TAM

该表混合了直接公开事实和分析含义。时间判断是方向性的,反映各项驱动或约束何时与 Liquid 的商业化推进最相关。

[CM012, CM013, CM014, CM015, CM016, CM017]

2.5 矛盾与尽调缺口

后续估值工作需要带着两个分析边界。第一,公开市场规模来源没有分离出 Liquid 的实际 SAM 或 SOM。宽口径边缘 AI 数字可以说明部署侧 AI 市场很大,但若没有内部管线、胜率或垂直行业组合数据,这些数字彼此太不一致,无法支撑单一自上而下收入模型。第二,公开来源没有展示 Liquid 把技术分发转化为付费企业使用的效率。创业公司和社区页面、文档以及开放许可姿态说明它有真实获客漏斗,但没有说明漏斗中有多少比例会变成付费部署、经常性收入或战略伙伴扩张。这个缺口重要,因为 Liquid 的替代集合既包括极其便利的托管 API,也包括部署更重的开放权重替代。缺少客户转化、交易规模和支持负担数据时,正确结论是 Liquid 看起来与强市场驱动高度契合,但实际捕获率仍是内部运营问题,而非公开事实。[CM015, CM016, CM017, CM024, CM027, CM039]

2.6 图表

Chapter 03

03竞争格局

3.1 格局与直接同业集合

Liquid AI 竞争的不是一个单体基础模型赛道。它自己的页面反复强调在算力受限环境中的高效部署,而竞争对手分成几个面向买方的集群:OpenAI、Anthropic、Google 这类封闭托管 API 领导者;Mistral、Cohere、AI21、Writer、Microsoft Phi 和 xAI 这类高效企业模型供应商;以及 Meta Llama、以 Qualcomm 为中心的设备工作流等硬件相邻或开放权重替代。关键在于,Liquid 最强的重叠不只是最大聊天界面,而是那些提供隐私、定制或企业控制的供应商。因此格局很宽,但仍能看清:Liquid 试图在买方更看重部署灵活性、低延迟和本地控制,而不是通用公共 API 无处不在的地方取胜。实践中,采购团队会同时把 Liquid 拿来和模型实验室、平台选择比较,而不只和某一个基准测试领导者比较。直接同业集合因此应由买方要完成的任务和部署约束定义,而不能只看品牌。[CP001, CP002, CP003, CP011, CP015, CP016]

竞争对手画像表
供应商类别客户 / 部署重点产品包装信号关键差异化相对 Liquid 的主要短板
OpenAI前沿托管 API开发者和企业构建者公开按 token 定价的 API公开多模态 API 覆盖最广成本结构偏云优先,对嵌入式私有部署着墨较少
Anthropic企业助手 / API知识工作与企业安全买家座席计划加 API 定价信任与安全品牌心智强,并有企业计划结构相比 Liquid,边缘设备切入点不明显
Google Gemini平台型既有厂商开发者和大型企业免费层加付费 API 和搜索费用搜索分发和 Google 平台触达云和平台捆绑,与嵌入式边缘 AI 的竞争方式不同
Meta Llama开放权重替代想要本地控制的构建者下载路径加自定义许可选项开放可得性和生态广度默认没有 Liquid 式企业部署服务
Mistral高效企业平台需要隐私和平台控制的企业Studio 平台 / 企业销售路径所有权、治理和部署灵活性边缘设备叙事差异化较弱
Cohere面向企业的私有 AI对安全和合规敏感的企业平台主导的企业销售私有化、可定制的企业定位设备原生叙事弱于 Liquid
Writer工作流软件既有玩家运行受治理工作流的受监管企业按座席收费的企业计划应用层工作流和治理深度对可下载模型部署着墨较少
Microsoft Phi + Foundry开放小模型加云平台以 Azure 为中心的企业和开发者开放模型加受治理平台将开放小模型与巨大分发能力结合买家可能得到平台锁定,而不是 Liquid 的专业化

这些行概括买家评估 Liquid 的高效企业 AI 时会遇到的主要实际替代品,并非涵盖市场上每一家模型供应商。

[CP001, CP011, CP015, CP016, CP019, CP021]
FP001: 竞争定位地图

方向性地图,展示 Liquid 相对于托管 API 领导者、企业工作流厂商和开放权重替代品的位置。

坐标轴取值是基于已审阅公开页面的序数型分析师评分,不是经审计的基准测试或客户调研数据。

[CP004, CP008, CP009, CP010, CP013, CP020]

3.2 能力与定价比较

最清晰的竞争分割出现在产品包装和定价透明度上。OpenAI、Anthropic 和 Google 都发布买方可以直接对标的 token 或工具定价,OpenAI 和 Google 也强调广泛多模态使用。相比之下,Liquid 明确表示不运行自己的托管 API;它的自助路径包括免费 playground、通过 OpenRouter 付费访问和模型下载,而 LEAP 与企业部署走销售。Meta、Microsoft 和 AI21 通过把可下载或开放模型与企业友好定位结合,进一步扩展替代集合。Writer、Cohere 和 Mistral 更侧重受治理的企业工作流,而不是纯消费者聊天规模。实际含义是,Liquid 在包装和部署架构上的竞争,与原始模型质量一样重要。评估总拥有成本的买方可能喜欢 Liquid 的设计,但他们仍能从托管竞争对手那里看到清晰得多的目录价。除非 Liquid 在销售周期中快速拿出定制基准测试或试点证据,否则这种不对称会拖慢评估。[CP004, CP006, CP007, CP008, CP009, CP010]

功能 / 能力矩阵
能力LiquidOpenAIGoogleMeta LlamaMicrosoft PhiWriter
多模态支持是——营销覆盖文本、视觉、音频、视频、信号是——GPT-4o 覆盖文本、音频、图像、视频是——Gemini 系列定位广泛所审材料主要定位为文本系列是——Phi 系列覆盖文本、音频和视觉工作流层比基础模型模态更突出
端侧 / 本地部署公司核心叙事所审页面未将其作为核心定位所审页面未将其作为核心定位是,可通过可下载权重实现是,靠开放小模型姿态实现不是主要切入点
私有 / 本地定制是——LEAP 和企业防火墙定位可从企业工具隐含推导,但不是主要信息有企业路径,但不是所审材料的主要信号用户自托管可实现可在 Azure 和开放模型工作流中实现是——企业治理和受控工作流
公共托管 API没有第一方托管 API所审来源未见第一方 Meta API经由 Azure 平台承载是应用平台,而非原始公共 API
可下载权重是,通过 Hugging Face所审来源未见所审来源未见所审来源未见
企业治理叙事有一些所审来源体现有限是,通过 Foundry
边缘硬件叙事是——聚焦 AMD 和嵌入式设备有限有限通过生态间接体现通过 Azure 生态间接体现有限

标为“有限”或“非核心”的单元格表示,所审公开页面强调了其他市场路径;并不证明该能力不存在。

[CP002, CP004, CP005, CP007, CP011, CP013]
定价 / 产品包装比较
供应商公开入门定价单位公开可见内容仍不透明的内容对 Liquid 的含义
Liquid没有第一方托管 API 价格N/A免费 Playground、付费 OpenRouter 路径、下载、销售主导的 LEAP实际合同价值和支持定价更多靠定制部署竞争,而不是公开 API 价目表
OpenAIGPT-5.5 输入 $5 / 输出 $30每 1M tokensAPI 经济性和工具定价透明企业折扣和混合交易条款为按 token 采购树立可见基准
Anthropic输入 $5 / 输出 $25,另有座席计划每 1M tokens / 每座席API 与企业座席混合模式大客户折扣和捆绑结构提供一个可与 Liquid 对标的信任导向商业模式
Google Gemini免费层加付费 token、缓存和搜索费用每 1M tokens 和每 1,000 次搜索生产使用有细粒度计量大企业定制条款显示竞争对手定价可以多透明
Writer企业版和入门版座席计划每座席 / 每计划受治理工作流定价模型和数据留存承诺定制扩展模块和服务在应用工作流层竞争,而不是比原始 token 成本
Meta Llama下载路径加自定义商业许可许可 / 自托管带门槛的开放权重访问下载后的完整商业经济性对想控制成本的买家,抬高替代风险
Mistral / Cohere所审页面显示以销售或平台为主导合同企业平台姿态所审语料未披露准确标价即使没有简单公开价目表,仍具可信竞争力

该表比较公开采购界面;不估计折扣、预留容量或合作伙伴转售条款后的实际定价。

[CP004, CP008, CP009, CP010, CP013, CP015]

3.3 分发、替代与多栖使用

分发是 Liquid 与前沿 API 实验室结构性差异最大的地方。OpenAI、Google、Microsoft 和 xAI 受益于公共 API、庞大既有生态或平台捆绑。Writer 和 Cohere 通过工作流软件和治理走可信企业路线。Liquid 看起来则在通过嵌入式设备、硬件和行业合作搭建分发:AMD 面向 PC 级设备端加速,Mercedes-Benz 面向车内智能。Qualcomm 和其他设备端栈也构成替代,因为它们可以承载许多更小或开放模型,而不需要 Liquid 的完整商业栈。这让买方多栖使用的可能性很高。客户可以用一个托管 API 处理通用工作负载,用开放权重处理私有任务,再在低延迟边缘部署重要的地方测试 Liquid。因此,Liquid 的护城河取决于能否让这些边缘部署实质性变得更好,而不只是“可以做到”。分发力量将来自可复制的合作伙伴牵头落地、参考部署,以及证明 Liquid 能在受限硬件上比通用替代更快集成或更便宜运行。[CP022, CP024, CP025, CP026, CP027, CP028]

分发与替代模式表
模式代表性竞争对手买家吸引力为何威胁 LiquidLiquid 仍有差异化之处
公共 API 平台OpenAI / Google / Anthropic采购快,开发者能立即采用让 Liquid 不提供托管 API 的姿态显得更慢、更不透明私有和嵌入式部署
开放权重自托管开放权重:Meta Llama / Microsoft Phi / AI21 Jamba控制权和本地部署买家无需购买 Liquid 服务,也能追求效率Liquid 仍可销售优化和特定领域部署支持
受治理企业工作流层Writer / Cohere合规、治理和工作流可重复性预算转向应用,而不是基础模型供应商模型部署本身卡住时,Liquid 仍可胜出
云平台捆绑Microsoft Foundry / Google 企业栈单一供应商采购和治理既有厂商可把 AI 捆进更大的平台支出Liquid 可在硬件和云选择之间保持厂商中立
贴近硬件的部署栈Qualcomm / AMD 生态端侧执行和 OEM 关系合作伙伴可以承载多个模型系列,不只 Liquid当模型在受限硬件上表现更好时,Liquid 有差异化
嵌入式垂直解决方案汽车或制药专业厂商采购看结果,而不是泛泛挑选模型如果垂直伙伴在别处标准化,可能降低对 Liquid 的需求Liquid 已在汽车类部署中展现早期证明

这些是替代采购模式,并非互斥竞争对手;许多买家可以同时采用不止一种。

[CP022, CP024, CP025, CP027, CP028, CP030]

3.4 护城河耐久性与替代风险

Liquid 的护城河在效率故事与私有或嵌入式部署要求绑定时最清晰,因为这些场景同时重视延迟、数据本地性和硬件约束。AMD 和 Mercedes 信号有帮助,因为它们是与公司叙事匹配的上市路径证明,而不是泛泛 AI 宣传。但替代风险仍真实存在。Meta 和 Microsoft 的开放权重替代扩大了买方选择,Mistral、Cohere、Writer 和 AI21 这类高效企业供应商可以打包相邻采购方案,公开基准测试生态仍让 OpenAI、Google 和其他 API 优先实验室更容易被快速评估。这意味着 Liquid 的差异化可信,但还没有结构性主导。它更像是边缘私有工作负载中的尖锐楔子,而不是横跨企业 AI 的赢家通吃护城河。因此,承销问题不是楔子是否存在,而是公司能否在大型供应商缩小部署差距之前,把楔子变成可复制的商业习惯。[CP037, CP038, CP039, CP040]

护城河耐久性 / 竞争风险登记表
Liquid 护城河主张主要威胁严重性威胁为何可信尽调要求
高效边缘私有部署开放小模型加硬件渠道Meta、Microsoft、AI21 和硬件生态都在创造相邻替代品要求提供客户证据,证明在目标设备上的性能或 TCO 明显更好
销售主导的企业定制Microsoft、Google 和 Writer 的平台捆绑中高既有厂商可以把治理和平台采购捆进更大的企业合同要求提供对云平台既有厂商的输赢数据
合作伙伴主导分发合作伙伴非独家AMD、Mercedes 和其他设备渠道可以支持多个模型供应商澄清独家性、优先合作伙伴权利和续约机制
开放模型友好度开放权重直接替代买家可能下载 Meta 或 Phi 模型,完全绕过 Liquid说明 Liquid 的模型质量或优化在哪里仍有差异
隐私 / 延迟切入点小型托管模型和本地模型快速进步对手正在快速提升多模态效率和本地运行时跟踪公开发布节奏和客户迁移行为
基准测试叙事公共榜单可见度集中在别处公共 API 和广泛基准覆盖让评估对手更简单提供面向买家的基准测试材料包和客户横评材料

严重性是基于所审语料的分析判断,应在管理层尽调中用输赢和使用证据重新检验。

[CP032, CP035, CP036, CP037, CP038, CP039]
FP002: 护城河 / 就绪度 KPI

基于已审阅公开记录,提炼 Liquid 竞争就绪度和护城河耐久性的紧凑指标。

这些指标只是公开信号代理;不能替代非公开客户队列、留存或赢单 / 输单数据。

[CP006, CP026, CP027, CP028, CP032, CP038]
Chapter 04

04财务情况

4.1 收入模型与包装架构

Liquid 的公开商业界面描述了一套形态可理解、实际经济性不透明的变现架构。公司明确表示不运营自己的托管 API,这立刻把它和按 token 计量的领导者区分开。公开自助路径包括免费 playground、付费 OpenRouter 访问、直接模型下载,以及由销售推动的 LEAP 和企业路径。许可又加了一根重要商业杠杆:小用户享有宽泛权利,但商业用户一旦跨过 $10 million 年收入门槛,就必须切换。合起来看,这意味着收入模型围绕企业部署、许可、支持和定制搭建,而不是大规模第一方 API 流量。机制清楚到足以分析,但实际 take rate、ASP 和收入结构仍无法从公开来源获得。[CI001, CI002, CI003, CI004, CI011, CI012]

收入流表
收入流机制单位当前公开状态收入质量判断尽调要求
Playground 探索免费发现和评估使用量明确免费,但有速率限制直接收入低,但漏斗信号有用要求提供 Playground 使用转向付费渠道的转化情况
OpenRouter 接入付费第三方托管访问通过合作伙伴产生的 token 使用量公开承认为付费路径无需自有 API 基础设施,也可能低摩擦变现要求说明扣除合作伙伴经济性后保留的收入份额
模型下载可直接下载的模型许可 / 部署可通过 Hugging Face 和文档公开获取能拉动广泛采用,但单靠下载直接变现较弱要求披露从下载用户转为企业合同的付费路径
LEAP 定制销售主导的定制与部署工具项目 / 合同大多数模型支持定制和部署若可复制,可撑起更高价值的企业销售要求披露平均合同金额、期限和服务组合
商业许可转化超过门槛后增购许可 / 合同用户年收入超过 $10M 门槛后触发若能执行,可能是强变现抓手要求披露因门槛触发的谈判数量和成交率

公开证据说明 Liquid 可以怎样收费,但看不出各收入流目前贡献多少收入。

[CI001, CI002, CI003, CI004, CI011, CI024]
定价 / 变现表
路径或可比项公开价格或合同线索标价与成交价买方获得内容未知项解读
Liquid 试用台免费,有速率限制标价上手试用转化率和使用上限经济性更像漏斗,不是收入中心
Liquid OpenRouter 路径付费后限额更高合作伙伴标价,不是 Liquid 实际成交价不由 Liquid 自营 API,也能获得托管访问收入分成、量折扣和实际抽成率证明存在一定自助变现,有参考价值
Liquid LEAP / 企业联系销售 / 定制仅成交价定制、部署、支持和私有基础设施选项ASP、最低承诺和服务负担可能是核心变现路径
OpenAIGPT-5.5 输入 $5 / 输出 $30标价透明的托管 API 访问企业折扣按 token 计价 AI 采购的基准
Anthropic输入 $5 / 输出 $25,另有席位套餐标价助手 + API 的混合打包大客户折扣展示信任导向的商业结构
Google Gemini免费层级 + token、缓存和搜索付费标价细颗粒度生产计量企业定制条款强化 Liquid 的透明度缺口

Liquid 的公开标价缺口本身就有财务意义,因为外部无法把实际变现与公开可比项做基准对照。

[CI002, CI012, CI013, CI014, CI015, CI016]
FI001: 收入模型桥

定性桥,展示 Liquid 公开产品路径如何把使用和部署兴趣转成收入。

公开信息只证明这些路径存在;定价兑现、转化率和毛利获取都未披露。

[CI001, CI002, CI003, CI004, CI024, CI027]

4.2 GTM 代理信号与可比定价压力

Liquid 最好的公开 GTM 信号不是客户数量披露或 ARR 表,而是企业合作证明,以及公司谈论部署的具体方式。AMD、Mercedes-Benz 和 Insilico 材料都强化了同一个模型:Liquid 通过启用通用托管 API 难以利落解决的私有或嵌入式 AI 工作负载切入。这在战略上令人鼓舞,但不足以证明销售效率。公开竞争者定价让对比更尖锐。OpenAI、Anthropic 和 Google 都发布详细价格卡;Writer 展示按席位计费的企业路径。这些可比对象给了买方透明锚点,而 Liquid 没有。因此,Liquid 可能从定制交易的灵活性中受益,但采购环节也背负更重证明负担,因为外部无法从公开记录对标它的实际经济性。汽车、电商、金融服务和创业公司等额外垂直页面强化了一个判断:管理层在围绕解决方案模板组织需求,而不只是模型访问。这有助于解释为什么公开 GTM 证据更像渠道建设,而不是典型自助 SaaS 规模指标。独立市场报告也支撑企业和边缘 AI 的更广预算背景;Liquid 卖进的是采用曲线,而不是成熟商品市场。[CI008, CI009, CI010, CI013, CI014, CI015]

单位经济性与 GTM 代理指标表
代理指标公开数值或状态置信度重要性含义尽调要求
AMD 渠道支持公开发布,并给出 Ryzen 和 Ryzen AI 支持的基准主张暗示企业用例更快落地可能降低技术采用阻力要求披露渠道来源销售管线和转化
Mercedes 合作多年期嵌入式车载智能协议显示企业愿意购买嵌入式部署是有用验证点,但不是收入披露要求披露合同规模、里程碑和预期量产爬坡
Insilico 合作私有基础设施上的科学模型合作显示通用聊天之外的垂直商业路径可能支撑更高价值的垂直部署要求披露定价模型和续约结构
企业市场叙事强调安全、成本高效、私有的 AI中高指向受监管或基础设施敏感型买方销售周期可能更长,但合同质量更好要求披露典型销售周期,以及各垂直领域赢单率
竞争对手透明价目表OpenAI、Anthropic 和 Google 发布标价给买方提供可直接比较的采购锚点Liquid 需要更有力地证明定制定价合理要求披露相对公开基准的折扣逻辑
公开客户和 ARR 披露已审阅语料中未找到卡住 CAC、回本周期、留存和规模效率测算核心 GTM 经济性公开不可知要求披露客户队列、ARR 和扩张数据
垂直解决方案页面汽车、电商、金融服务和创业公司页面暗示按解决方案分层和打包提升 GTM 可读性,但不提升收入可见度要求按垂直领域披露销售管线和 ARR
企业 AI 与边缘市场背景Deloitte 企业 AI 调研 + 边缘 AI 市场研究支撑伙伴驱动企业销售的需求背景提升漏斗顶端信心,不等于披露单位经济性要求管理层说明目标垂直 TAM 和预算捕获

这些只是代理指标,不应误认为 CAC、回本周期或 NRR 的直接披露。

[CI008, CI009, CI010, CI019, CI020, CI021]
FI002: 单位经济模型桥

从企业证明通向真正单位经济模型所缺指标的公开 GTM 桥。

桥图刻意止于未解节点,因为客户数量、CAC、回本期和毛利率都不公开。

[CI008, CI009, CI010, CI019, CI025, CI026]
FI004: 资本强度 / 定价透明度地图

矩阵对比 Liquid 与透明 AI 可比公司的变现和披露姿态。

单元格概括已审阅来源中的主导公开商业姿态;不是经审计的财务分类。

[CI012, CI013, CI015, CI016, CI017, CI029]

4.3 资本充足性与披露缺口

$250 million 融资显著改善了 Liquid 对外呈现的资本实力,独立报道给出超过 $2 billion 的公司估值,也说明投资人在承销一条有分量的商业化路径。但资本充足不等于资本清晰。被审阅的公开材料没有披露账上现金、烧钱速度、现金跑道、毛利率、客户数或类债务义务。因此,这轮融资更像信心信号,而不是真正的跑道计算。Microsoft 2025 年 10-K 在这里可作比较,不是因为 Microsoft 在规模上是同业,而是它展示了公开 AI 企业能够给出的收入、利润率和资本披露水平。按这个基准看,Liquid 仍高度不透明。正确结论是资产负债表标题偏正面,但承销文件在通常决定资本风险判断的每个指标上仍不完整。[CI005, CI006, CI007, CI021, CI022, CI023]

资本充足性表
项目公开数值或状态置信度重要性当前判断尽调要求
最新股权融资已宣布 $250M 融资最重要的公开资本锚点正面融资信号要求披露交割日期、净募资额和交割后账上现金
隐含估值据独立报道,> $2B显示投资者信心和定价水平支撑强投资者胃口要求披露投后估值、所有权稀释和优先股堆叠
手头现金未公开披露计算现金跑道必需Unknown要求披露月末现金和受限现金
月度烧钱未公开披露判断融资依赖性必需Unknown要求披露过去六个月 opex 和现金消耗桥
现金跑道(月)公开资料无法支持把融资转成生存能力指标Unknown要求披露基准、下行和招聘计划现金跑道情景
债务或已承诺基础设施义务已审阅语料中未公开披露中高可能实质改变资本强度Unknown要求披露算力承诺、供应商最低采购额和任何债务安排

融资新闻提升信心,但缺少现金、烧钱速度和义务披露,就无法计算现金跑道。

[CI005, CI006, CI021, CI030, CI031, CI040]
公开财务缺口表
缺失指标缺失为何重要对结论的影响具体尽调路径严重性
ARR 或收入运行率缺少判断经常性收入质量的规模锚点卡住估值与增长研判获取当前 ARR 桥接和上年对比阻断性
客户数和集中度无法判断平台覆盖广度或单一客户风险卡住 GTM 耐久性判断获取付费客户数和前 10 大客户收入集中度阻断性
毛利率无证据判断重部署模式更像软件还是服务卡住毛利率路径判断按路径获取托管、支持和交付毛利率阻断性
CAC、回本周期和销售周期无法判断企业销售动作是否高效限制对增长经济性的信心获取漏斗指标和队列回本分析重要
现金消耗和下一轮融资触发点无法判断融资依赖何时重现限制资本充足性判断获取董事会计划、下行现金跑道和下一次融资触发点阻断性
标杆客户经济性有案例研究外观,但经济性未披露限制对 ROI 和可复制性的信心要求披露具名客户成果,并绑定合同金额和续约重要

这些缺口反映的是公开数据限制,不一定是公司弱点;但对外部投资判断仍是决定性阻断项。

[CI021, CI023, CI030, CI032, CI035, CI042]
FI003: 财务估算区间

公开可支撑的数字锚点包括融资和估值,但收入和跑道期仍缺乏依据。

未获支撑的经营指标被刻意呈现为未知,而不是估算。

[CI003, CI005, CI006, CI030]

4.4 财务结论与尽调阻断项

公开记录支持一个平衡但约束清晰的财务判断。Liquid 似乎拥有连贯的企业部署收入机制、一些令人鼓舞的合作伙伴证明,以及继续建设的新资本。它的商业设计也在逻辑上区别于按 token 计量的前沿 API 和按席位收费的应用供应商。但完整承销所需的几乎每个指标仍是私有信息:ARR、客户集中度、烧钱速度、毛利率、CAC、回本周期、续约和下一轮触发条件。这意味着公开投资逻辑最多只能走到结构,不能走到业绩。尽调时,公司应被视为一家有前景但仍不透明的私营企业 AI 供应商:收入模型可读,经济性不可见。下一个决策点不应再靠叙事,而应取决于管理层是否愿意打开运营模型和资本计划。因此,谨慎投资人可以认可战略方向,同时仍拒绝只凭公开数据承销估值。剩余工作是管理层接触尽调,而不是继续解读列表页面。[CI026, CI033, CI035, CI036, CI041, CI042]

Chapter 05

05产品与技术

5.1 产品栈、模型家族与面向买方的界面

Liquid 已不只是一个研究故事;公开产品栈现在有三层可见产品。第一,公司提供 Liquid Foundation Models 本身,包括文本、视觉语言、音频和 nano 变体。第二,它提供 LEAP 作为帮助开发者定制和部署这些模型的工作流。第三,它推出 Apollo,作为消费者友好、无需云的移动界面,用设备端智能做演示,不要求企业合同。关于页面、模型页面、文档和社区页面都强化了这一产品栈形态。 模块地图还显示,Liquid 优化的是部署现实,而不只是基准测试营销。公开文档强调 GGUF、MLX 和 ONNX 包装,几个检查点可训练,并兼容 Transformers、llama.cpp、vLLM、MLX 和 Ollama 等广泛使用的运行时。这让开发者和平台团队更容易判断 Liquid 是封闭黑箱,还是一个可迁移模型家族。主要保留意见在于,面向买方的包装远比 LEAP 自身的商业采用清晰:公开网站解释了可以试什么、去哪里下载,但没有披露已有多少企业客户真正把平台标准化。[CE001, CE002, CE003, CE004, CE034, CE036]

产品模块 / 资产矩阵
模块 / 资产主要用户公开状态部署目标差异化尽调缺口
LFM2 / LFM2.5 文本模型开发者、企业 AI 团队公开模型库CPU、GPU、NPU;云端或本地小模型效率 + 多种打包格式未公开付费用户数或企业部署总量
视觉语言模型OEM、电商、企业构建者公开模型库边缘设备和云端多模态支持,面向边缘的占用更小未公开从基准表现到客户转化的数据
音频模型语音和对话产品团队公开模型库低延迟本地或混合部署主打响应式对话的 1.5B 音频文本模型未公开生产客户案例
LEAP 平台开发者和平台工程师商业平台任意 OS;从笔记本到企业端点统一定制和部署工作流内部机制和商业规模大多不透明
Apollo 应用消费者和评估者公开应用界面移动设备无云、本地化展示 Liquid 推理消费应用使用量不能证明企业变现
定制垂直解决方案企业买方和具名合作伙伴销售主导汽车、金融、电商、制药围绕工作流需求,把模型和部署打包大多数公开证据仍未具名或处早期

行里合并了模型家族、软件层和面向客户的界面。公开状态只反映公司在官网和文档披露的内容,不代表已知收入规模。

[CE002, CE004, CE010, CE013, CE014, CE017]
工作流 / 用例表
用户任务当前工作流痛点Liquid 触点公开收益主张局限
部署本地 AI 的平台工程师托管 API 带来延迟和数据治理摩擦LEAP + 本地 LFM一个工作流覆盖跨设备定制和部署无付费采用或支持负担的公开证据
交付车载助手的车企依赖云端会损害隐私和响应一致性汽车解决方案 + LEAP优先跑在边缘、利用现有车辆硬件的助手Mercedes 证据仍处量产前
运行敏感发现工作流的制药科学家使用云端可能暴露专有分子和实验Insilico 科学 LFM 部署私有基础设施上的药物发现模型已可用无公开客户规模或续约指标
零售或交易平台运营方搜索和商品智能体太慢或太贵电商定制模型意图原生搜索和智能体店面动作已发布案例研究未具名
反欺诈或支付团队实时评分必须够快且保持私密金融服务工作流解决方案公开案例研究中反欺诈处理速度提升 2x没有具名机构或生产环境标签
评估本地 AI 的移动用户或开发者试用本地推理通常要配置环境、整理模型Apollo / 试用台 / HF 下载无云演示路径和可下载模型检查点评估入口不能证明持久商业使用

本表把公开产品触点映射到客户工作任务。收益来自公司主张和公开案例摘要,不是独立结果审计。

[CE010, CE013, CE014, CE015, CE016, CE031]
FE001: 产品架构地图

从研究脉络到终端用户界面的公开可见栈。

[CE005, CE002, CE003, CE010, CE013]

5.2 架构、LEAP 部署路径与关键依赖

Liquid 的差异化故事先是架构,再是商业。公司反复把 LFMs 描述为液态神经网络和更广状态空间研究的延伸,使用第一性原理、动力系统、信号处理和数值线性代数的语言,而不只是 Transformer scaling。独立发布报道基本接受了这一框架,尤其是在内存效率和小模型性能上。公开研究页面也显示,公司在发布后继续交付架构论文;这一点重要,因为公司的商业可信度取决于延展真实研究边缘,而不是简单重贴一个高效推理包装器的标签。 在部署上,LEAP 是研究与收入之间的关键中间层。公开材料把 LEAP 描述为模型定制和部署的统一工作流,AMD 新闻稿加上 TechIntelPro 报道提供了最清楚的具体证据:Liquid 能把模型绑定到特定硬件目标。上行空间很强:AMD 笔记本支持、合作伙伴背书的 sub-100ms 响应声明,以及面向私有推理的云无关路径。风险在于,公开文档仍没有透露 LEAP 底层更深的编译器、调度器或编排内部机制。除了一篇报道把这套栈和 llama.cpp 联系起来,Liquid 要求买方信任一个大体封闭的部署层。[CE005, CE006, CE007, CE010, CE011, CE012]

技术 / 运营架构表
层 / 流程公开组件作用依赖风险
架构核心LFM / LNN / 状态空间谱系把 Liquid 与 transformer 优先的同行区分开内部研究优势和持续发表节奏营销强于外部验证的消融细节
模型分发HF + 文档 + 打包格式让模型可跨本地运行时迁移Hugging Face、模型打包、开源运行时分发广度不等于企业标准化
定制层LEAP把微调和部署绑进一个工作流Liquid 自有平台 + 硬件集成编译器、调度器或服务运营的公开细节很少
推理路径llama.cpp / 受支持运行时 / ONNX / MLX在笔记本、终端和本地栈上运行模型第三方运行时和硬件专项优化性能可移植性可能因设备和模型而异
硬件集成AMD,以及公开声称支持 Apple、Qualcomm、Cerebras、NVIDIA触达范围不再局限于单一芯片栈伙伴配合和优化质量对硬件伙伴的依赖可能影响交付节奏
客户验证层Mercedes / Insilico / 未具名案例研究证明架构已进入真实工作流有名称的伙伴和垂直行业合作证据仍较集中,且部分偏前瞻

架构行把公司声称的内部机制和最具体的公开部署信号放在一起。表格刻意拆开模型、平台和伙伴层,因为商业栈要靠三者共同支撑。

[CE005, CE003, CE010, CE012, CE011, CE030]
FE002: 客户工作流 / 运营流程

在 Liquid 公开工作流中,买方如何从模型评估走向设备原生部署。

[CE035, CE010, CE013, CE011, CE031]
FE003: 关键依赖地图

Liquid 的公开部署叙事依赖少数生态节点。

[CE011, CE003, CE008, CE030, CE031]

5.3 信任、隐私、许可与质量控制

Liquid 的信任姿态更多建立在部署控制和商业规则上,而不是成熟的公开合规界面上。最强的公开信任主张是运营性的:没有第一方托管 API、客户防火墙后的企业本地访问、Apollo 被营销为无云且离线,以及 AMD 背书的“零依赖云 API”表述。这些主张与公司的边缘和主权 AI 定位一致。实践中,它们最应吸引那些明确优化隐私、延迟和本地控制的买方,而不是想要无摩擦托管 API 的买方。 法律界面也异常重要。Liquid 的 LFM Open License 在许多方面宽松,但在 $10 million 收入门槛以上终止免费商业权利,并迫使更大公司进入协商商业许可。这可能是聪明的漏斗设计——小型开发者免费、企业变现——但也意味着产品采用部分受销售和法律谈判限制。最弱的区域是第三方信任证据。公开来源没有给出扎实的信任中心、审计报告、已发布事故历史,或除营销和许可之外的详细 LEAP 治理材料。对于一家卖进汽车、金融和受监管科学领域的公司,这个缺口很实质。[CE008, CE009, CE013, CE022, CE023, CE028]

信任 / 质量 / 合规表
控制项 / 约束公开状态范围证据缺口
无第一方托管 API已披露分发和隐私姿态定价页不能替代经审计的企业级控制
本地 / 企业内网访问已披露企业部署模型 FAQ 和定价页未看到公开 DPA、SOC 范围或信任中心资料包
Apollo 设计上离线运行已披露消费者移动体验Apollo 和社区页面消费者隐私声明不等同于企业治理
开放许可证的商业使用门槛已披露年收入 > $10M 的用户LFM Open License(开放许可)更大客户仍需私下谈判条款
发布时的预览 / 红队状态部分披露模型质量流程VentureBeat 发布报道无公开事故历史或模型风险指标
独立信任证据薄弱跨栈保障本章审阅的公开页面除 Liquid 自身说法外,公开审计线索不充分

公开信任证据最强的是部署控制和法律条款,而不是第三方审计透明度。「薄弱」指本次审阅的公开材料没有提供详细的第三方保障材料。

[CE008, CE009, CE013, CE022, CE028, CE029]
FE004: 产品成熟度 / 能力地图

公开证据在模型广度和部署定位上最强,在经审计信任细节和 LEAP 采用规模上最弱。

矩阵取值是基于公开证据密度的定性判断,不是内部 KPI。

[CE002, CE036, CE030, CE031, CE028]

5.4 路线图、真实世界证明与产品风险

最有说服力的公开产品证明不只是基准测试,而是一组靠近部署的合作伙伴关系和案例研究。Liquid 现在有三个不同的外部证明点。第一,Mercedes-Benz 给了 Liquid 一条旗舰级汽车量产路径,目标是在 2026 年下半年进入生产部署。第二,Insilico Medicine 给了 Liquid 一个药物发现中当前进行时的私有基础设施部署故事。第三,案例研究页面展示了汽车、欺诈、电商和合成视频生成等多个垂直场景的结果声明。合起来看,这些信号说明公司正试图从模型供应商转向行业特定解决方案伙伴。 但路线图仍比承销依据更容易看见。融资和发布材料显示,公司继续投入边缘和本地部署就绪能力,关于页面也宣传 2026 年 LFM2.5 发布节奏。然而,公开客户指标仍很薄,LEAP 商业采用不透明,一些最大胆的结果声明仍来自匿名案例研究。独立报道也捕捉到编码、精确数值和时效性信息上的真实技术保留意见。当前结论是,Liquid 已拼出可信产品栈,并有合作伙伴背书的部署证据,但还没有公开运营记录,足以完全降低企业级可靠性或平台采用风险。[CE018, CE019, CE020, CE021, CE030, CE031]

路线图 / 发布 / 开发阶段表
日期 / 阶段里程碑状态含义来源
2023 发布退出隐身模式,提出本地部署和私有基础设施愿景已完成商业平台意图从第一天就存在TechCrunch + first-principles 博客
2024-10首次公开发布 LFM(1.3B / 3.1B / 40.3B MoE)已完成Liquid 从研究主张转向公开模型厂商LFM 发布博客 + 独立报道
2024-12Series A 聚焦边缘 / 本地部署准备度,以及推理 + 微调栈已完成指向平台化,而不只是模型训练融资博客
2025-08LEAP 增加 AMD Ryzen / Ryzen AI 支持已完成形成有硬件支撑的具体部署路径AMD 新闻稿 + TechIntelPro
2025-12 至 2026LFM2 技术报告,加上可见的 2026 LFM2.5 发布节奏进行中说明模型家族仍在扩张研究页面 + 关于页面
2026 H2 目标Mercedes 首个量产部署路径待定最可见的旗舰部署仍在前方,尚未被证明Business Wire

发布阶段混合了已完成和待定里程碑。待定指已经公开宣布,但截至 2026-06-04 运行日期尚未展示为规模化量产。

[CE033, CE024, CE032, CE011, CE007, CE030]
Chapter 06

06客户情况

6.1 分层与公开客户界面的形态

Liquid 的公开客户界面在目标行业上很宽,在已披露账户上很窄。公司明确面向企业、汽车制造商、金融机构、电商运营者、医疗和生命科学团队、工业用户、创业公司和开发者营销。这个宽度重要,因为它说明 Liquid 销售的是灵活的部署和定制能力,而不是单一用途应用。潜在买方也因路径不同而变化:企业里的平台团队和 CTO,汽车里的 OEM 软件组织,金融服务里的风控和欺诈团队,电商里的商品运营或搜索团队,以及受监管制药工作流里的科学家。 但谁真正付费的公开证据比细分叙事薄得多。Liquid 没有披露客户数、付费账户数、按细分市场划分的收入结构,也没有像更成熟应用供应商那样给出具名企业 logo 墙。公司呈现的是混合界面:面向定制企业工作的直销页面、通过社区资源进行免费或低摩擦评估,以及少量具名或匿名部署故事。因此,细分叙事可作为 TAM 框架是可信的,但作为规模化采用证据仍不完整。[CU001, CU002, CU003, CU011, CU012, CU013]

客户分群表
分群买方 / 用户 / 付款方公开证据面部署方式收入 / 战略价值缺口
开发者社区构建者 / 评估者 / 未来买方文档、HF、playground、Apollo、黑客松自助评估和本地试验漏斗顶端触达和生态心智无公开转化或付费账户数据
企业定制解决方案CTO、平台负责人、AI 团队 / 内部用户 / 企业预算负责人企业页面 + 定价 + 案例研究销售主导的定制部署可能是最大 ACV 路径无公开客户数或合同指标
汽车 OEM车载软件团队 / 驾驶员 / OEM 项目Mercedes 加汽车页面嵌入式、端侧、多年期项目高可信度旗舰垂直领域公开证据仍处于量产前
医药与生命科学科研负责人 / 研究人员 / R&D 预算Insilico 合作私有基础设施专项部署强监管场景信号未披露续约或规模指标
电商 / 零售搜索、商品运营、运营团队 / 购物者 / 电商预算垂直页面 + 未具名案例研究定制模型加 API 集成证明工作流相关性无有名称的商户参考
金融服务风控 / 反欺诈团队 / 分析师 / 业务单元预算垂直页面 + 未具名欺诈案例研究私有或混合实时部署结果导向的节省叙事无有名称的银行或保险公司参考

分群反映公开定位和证据面,而不是已披露的收入拆分。「战略价值」是基于交易对手质量和可能合同规模得出的尽调判断。

[CU001, CU002, CU012, CU003, CU021]
客户增长 / 采用轨迹表
信号公开证据日期 / 状态显示什么缺失的分母
无公开客户数概览、定价、融资和案例研究均未披露当前公开报道对客户广度仍很稀疏付费账户数量未知
开发者分发触点HF 组织、文档、playground、Apollo、黑客松当前存在较大的漏斗顶端评估触点用户到客户的转化未知
有名称的旗舰汽车客户Mercedes 合作2026 目标蓝筹背书,但截至运行日期尚未量产数量、合同金额和里程碑完成情况未知
有名称的监管科学部署Insilico 合作现已可用表明当前可在私有基础设施上部署席位数、使用量或收入规模未知
匿名垂直场景成果汽车、欺诈、电商、视频案例研究当前证明多垂直领域相关性logo、续约率和量产标签未知
伙伴测试叙事融资博客 + 直播生态露出近期 / 持续Liquid 同时推进多个行业哪些测试转化为合同未知

这里的轨迹使用公开证据的密度和新近程度,而不是已披露的客户 KPI 序列。几乎每个公开采用信号都缺少分母。

[CU003, CU013, CU006, CU007, CU014]
FU001: 客户旅程图

Liquid 公开可见的客户路径从评估走到定制部署,之后只有少数案例能形成可见的旗舰证明。

[CU013, CU012, CU007, CU006, CU033]

6.2 具名客户证明、试点与生产、参考质量

具名公开证明集集中但真实。Mercedes-Benz 是最可见的旗舰关系:一项多年汽车合作,绑定北美 MBUX 代际,首个生产部署目标是 2026 年下半年。这很有意义,因为客户是一线蓝筹,用例在运营上要求很高,但它仍不是发布后的证明。Insilico Medicine 是更强的当前状态部署信号,因为双方称科学模型现在已经可在私有制药基础设施上使用。合起来看,这两个账户说明 Liquid 能赢下严肃对手方。 离开这两个具名证明后,参考质量迅速下降。Liquid 其余公开客户证据大多来自匿名案例研究短文——汽车、欺诈、电商、翻译和合成视频——结果数字有意思,但客户 logo、生产标签和续约证据都很薄。因此,本章可以支持“真实采用存在”的结论,但不能支持“生产采用已经广泛或可在许多账户中复制”的结论。[CU004, CU005, CU006, CU007, CU008, CU009]

有名称客户证据表
交易对手分群公开状态成果证据量产 vs 试点新鲜度 / 局限
Mercedes-Benz汽车 OEM已宣布多年期合作公开目标是在北美为 MBUX 嵌入车内智能量产前;首次部署目标为 2026 H2有名称的旗舰客户,但发布指标未公开
Insilico Medicine医药 / 生命科学已宣布战略合作LFM2-2.6B-MMAI 据称现已可在私有基础设施上使用当前专项部署有名称且仍在进行,但规模和商业条款未披露
未具名全球车企汽车仅有案例研究摘要在现有车载硬件上部署视觉语言模型,体积小 50%、速度快 10x可能是真实部署,具体阶段未标明成果证据有用,但没有客户名称
未具名反欺诈客户金融服务仅有案例研究摘要处理速度快 2x,估计每年节省 ~$230M可能是近似量产的工作流,具体阶段未标明无 logo、合同或续约细节

这张表有意保持不完整,因为 Liquid 的公开客户证据集中在两个有名称交易对手,以及若干匿名参考。每行至少由两个公开来源,或一个直接证据来源加相关垂直页面支撑。

[CU004, CU005, CU006, CU007, CU008, CU009]
FU002: 采用 / 部署漏斗

公开证明密度的示意指数,从广泛评估触达到当前已具名部署逐层收窄。

漏斗数值是证明密度指数,不是客户数。它们表示各阶段有多少公开证据,以最宽的漏斗顶部层为 100。

[CU013, CU021, CU004, CU007, CU006]
FU003: 客户证明矩阵

具名证明在交易对手质量上很强,但广度和留存可见度偏弱。

[CU004, CU006, CU007, CU008, CU009, CU034]

6.3 留存、耐久性与公开材料仍没展示什么

耐久性是 Liquid 公开客户故事中最弱的一环。没有披露 NRR、GRR、流失率、续约率、席位扩张曲线或客户满意度数据。最好的耐久性信号是结构性的,而非测量得出的:Mercedes 关系明确是多年期,Insilico 的私有基础设施设置说明工作流一旦集成,切换成本会更高。但这些仍是代理信号,不能替代队列或账户级证据,证明 Liquid 能在初始试点、调优或部署工作完成后继续留住客户。 外部市场证据让这个缺口更重要,而不是更不重要。Deloitte 2026 年企业 AI 报告称,生产使用在增长,但治理、数据、风险和人才准备仍滞后,只有少数组织已经报告 AI 带来收入提升。对 Liquid 来说,这意味着客户可能有兴趣,也具备技术试点能力,却仍会缓慢转化为持久生产收入。在 Liquid 披露续约、满意度或重复扩张之前,留存故事仍主要是从部署架构和对手方质量中推断出来的。[CU028, CU029, CU024, CU025, CU015, CU016]

留存 / 重复使用 / 满意度表
指标公开数值 / 状态信号来源含义尽调要求
NRR / GRR未披露公开网站审阅无账户扩张效率证据按分群索取客户级 NRR / GRR
流失 / logo 留存未披露公开网站审阅无法检验定制部署的耐久性索取前 10 大账户的流失和续约历史
有名称的续约未公开披露案例研究 + 媒体审阅没有清晰的重复客户叙事至少索取两个续约客户访谈
多年期合同代理指标Mercedes 被描述为多年期Business Wire现有最好的耐久性信号仍是前瞻性的索取里程碑计划和取消 / 扩张条款
社区重复使用HF 活动和社区露出可见,但不是变现HF + 社区页面显示兴趣,不代表收入耐久性索取评估者到付费的转化和活跃使用队列

本表区分结构性耐久代理指标和真正的留存指标。只有 Mercedes 关系提供了明确的多年期信号,但在公开证据中,即便它也仍处于量产前。

[CU028, CU029, CU025, CU034]
FU004: 留存 / 重复队列

按客户类型给出的示意性耐久度代理;Liquid 未披露实际留存队列。

仅为代理队列值。这些百分比按客户类型推断相对切换成本和可见度;它们不是公司披露的留存曲线,只应作为尽调框架。

[CU024, CU025, CU013, CU028]

6.4 扩张潜力、集中度与伙伴依赖

Liquid 确实有合理的先落地再扩张逻辑。Mercedes 明确为初始车内部署后的其他产品领域探索留下空间。公司营销覆盖的行业也多于它目前能用具名参考证明的行业,免费评估界面应能扩大漏斗顶部。但扩张仍由伙伴调解。Mercedes 控制汽车时间表和分发。AMD 和其他硬件伙伴塑造设备性能和部署经济性。开发者可以轻松试用模型,但公司看起来仍通过协商访问和定制解决方案工作来变现更大用例。 这带来两类集中风险。第一,公开参考集集中在极少数旗舰名称上,因此这些项目若延迟或令人失望,会不成比例地损伤 Liquid 的客户叙事。第二,渠道依赖很高,因为除直销和模型分发端点之外,看不到主要转售商或市场平台动作。这并不意味着客户故事弱;它意味着故事脆弱。Liquid 有足够证据证明真实市场拉力,但还没有足够多元的公开证据说明客户耐久性已经足够广泛。[CU019, CU032, CU030, CU033, CU036, CU037]

扩张和集中风险表
扩张驱动因素证据集中度 / 依赖风险影响尽调路径
Mercedes 后续产品领域Business Wire 称可能探索其他产品领域扩张由伙伴掌控,并取决于首次发布成功如果发布顺利,可加深汽车可信度;若延期,则可能消失索取联合路线图和成功里程碑
Insilico 监管科学适配度私有基础设施部署契合医药需求证据可能过度集中在一个专业垂直领域科学背书强,但分群宽度窄索取其他监管科学参考
开发者漏斗转向付费企业HF、文档、社区、Apollo 和 playground 降低试用门槛从兴趣到合同的转化率未知广泛认知未必转化为可持续收入索取漏斗转化和队列数据
直销定制路径定价和企业页面把大客户导向销售少数大单可能主导叙事和收入ACV 上行空间高,但有集中风险索取头部账户收入占比和 pipeline 组合
硬件 / 伙伴生态AMD 和旗舰交易对手塑造交付叙事少数伙伴延期或重新排序,可能很快削弱证据叙事和部署经济性对生态敏感索取伙伴依赖图和应急计划

扩张机会真实存在,但明显由伙伴居中撮合。最大的公开缺口是缺少分群层面的收入组合或集中度数据,无法量化下行风险。

[CU019, CU032, CU030, CU033, CU036, CU026]
Chapter 07

07风险

7.1 按严重性排序的顶层风险与传导路径

Liquid AI 的风险堆栈,与其说是科学彻底失败,不如说是差异化架构能否熬过从研究可信度到可复制企业变现的残酷转换。公开证据支持这是一家真实公司,拥有资本、技术深度和具名伙伴;但同一批公开材料也显示运营系统仍很早期:商业化通过许可、下载、伙伴渠道和定制来路由,而不是可见的第一方 API 业务;标杆部署证明集中在 AMD、Mercedes 和 Insilico;独立报道仍把 LFMs 描述为预览阶段,并在某些任务上有限。组合起来,就形成层层叠加的传导路径。如果伙伴试点延期、基准优势在生产中变窄,或客户对许可条款和渠道复杂度犹豫,收入可信度可能比技术叙事更快走弱。因此,正确的严重性视角是累积性的:集中度、就绪度和变现不透明会互相强化,而不是彼此孤立。[CR004, CR006, CR013, CR014, CR021, CR022]

FR001: 风险热力图

商业化若依赖集中的合作伙伴证明,或依赖仍有争议的生产就绪假设,残余严重性仍然最高。

[CR022, CR024, CR026, CR031, CR034, CR043]
FR002: 风险传导图

商业和监管风险通过转化、部署证明和估值支撑互相放大。

[CR013, CR026, CR031, CR034, CR035, CR043]

7.2 法律、监管与商业合同风险

Liquid 的公开法律界面有实质商业意义。LFM Open License 对研究和小企业慷慨,但用户一旦跨过 $10 million 年收入门槛就会变得受限。由于版权和专利授权以这一门槛为条件,公司实际上是在告诉成功采用者转入协商商业关系。这可以是有意设计的变现杠杆,但如果企业认为免费 / 开放渠道条件过多或太容易终止,也会成为采购风险。监管背景会抬高利害关系。EU AI Act 明确把 AI 视为安全、权利和经济风险义务来源,NIST 的 AI RMF 现在也有生成式 AI 和关键基础设施覆盖层。Liquid 卖进金融、汽车和生物技术等隐私与延迟敏感行业,因此即使产品价值主张仍强,公开治理界面偏薄也会在尽调中越来越重要。[CR031, CR032, CR033, CR034, CR035, CR040]

监管 / 法律风险登记表
规则 / 问题管辖区状态可能性严重性缓释措施残余敞口尽调路径
面向企业和行业部署的 EU AI Act 义务EU / EEA2026 年起适用,涉及 GPAI 和高风险影响架构效率和端侧定位可能降低部分敞口中高索取 EU 合规地图、技术文档包和行业特定符合性计划
LFM Open License 的商业使用门槛全球合同年收入超过 $10M 后免费使用终止该门槛为企业许可创造变现杠杆审阅付费许可转化条款、定价和争议历史
附条件的版权和专利授权全球合同权利受商业使用限制约束基于 Apache 的语言让小型用户更熟悉中高请法律顾问比较其可执行性、客户接受度与 Apache 或商业源代码规范的差异
自动终止加 AS-IS / 责任限制全球合同在公开许可证中明确列明中高标准企业合同可通过谈判绕开公开条款审阅已谈判的企业条款、赔偿安排和支持例外条款
关键基础设施和受监管行业的治理负担美国 / 全球企业买方NIST 和行业买方要求更明确的风险管理Liquid 的私有 / 本地部署定位契合主权和隐私需求中高索取模型治理材料、事件流程、审计轨迹和受监管部署案例

各行按实际尽调严重性排序,而不是按已知执法行动排序;公开资料展示的是法律设计选择和监管背景, 不是已裁判争议。

[CR031, CR032, CR033, CR034, CR035, CR040]

7.3 运营、伙伴与部署依赖风险

从运营上看,Liquid 正在承销一条艰难路径:异构边缘部署、定制企业集成,以及跨受监管行业的生产支持。公开材料显示了真实进展,尤其是 AMD 优化和 Mercedes 汽车路径,但也暴露了投资逻辑仍有多少依赖对手方和 Liquid 无法完全控制的生产里程碑。公司尚未提供第一方托管 API,这意味着公开变现路径要穿过许可、伙伴渠道、下载和定制部署工作。长期看,这可能创造更强锁定和利润率;短期看,它也提高实施负担,并减少容易观察的使用信号。独立评论进一步放大担忧,因为它指出 LFMs 在 2024 年仍在打磨中,并在几个任务类别上更弱。投资人因此应把 AMD、Mercedes 和 Insilico 证明视为有价值但集中的证据,而不应假定一个灯塔交易会自动外推为广泛企业可复制性。[CR013, CR014, CR015, CR016, CR021, CR022]

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风险敞口未解决缺口
基准测试优势未能转化为稳定的生产质量中——技术深度和合作伙伴测试确有支撑需要按部署模式拆分的客户级延迟、正常运行时间和质量指标
没有第一方托管 API,限制了可见的自助变现和遥测中高中低——当前模式更偏授权和合作伙伴路径中高需要产品线收入结构,以及合作伙伴渠道能高效扩张的证据
代码、数值或时效性任务上的已知短板会制造采用阻力中高中低——模型工作仍在推进,但独立评估已有保留意见需要最新基准测试包和 2024 年后的独立评估
跨异构硬件做边缘部署会抬高 QA 和支持负担中——AMD 路径存在,Liquid 也强调硬件感知调优中高需要非 AMD 和混合设备环境下的部署成功率
标杆试点在转入生产前延迟中——Mercedes 和 Insilico 提供了真实路径需要里程碑追踪、客户验收标准和合同约定的上线日期

运营评级关注商业化可靠性,而不是网络安全事件历史;公开证据在架构上更强,在现场表现上更弱。

[CR013, CR014, CR021, CR022, CR023, CR024]
合作伙伴 / 依赖风险登记表
依赖项交易对手 / 层级角色集中度失效情景严重性缓释措施剩余风险敞口
硬件优化AMD / Ryzen / Instinct 生态主要公开芯片加速合作伙伴客户硬件组合一旦分化,性能主张就会变窄Liquid 对外主打 CPU/NPU/GPU 可移植性和更广的硬件野心中高
汽车部署证明Mercedes-Benz最清晰的大规模边缘端生产部署路径试点或生产延迟会削弱物理 AI 投资逻辑多年合作关系和已明确的目标上市年份
垂直科学验证Insilico Medicine医药领域专用用例和基准测试合作伙伴成功若停留在狭窄场景,就无法外推到更广企业需求中高验证显示私有基础设施和真实任务效用
分发和访问渠道OpenRouter、试用平台、下载、第三方平台当前公开访问路径,替代第一方托管 API中高渠道经济性或用户体验仍是间接的中高企业授权和 LEAP 提供替代路径中高
竞争基准测试背景Artificial Analysis 跟踪的主要模型实验室性能、价格和部署预期的参照组竞争对手效率提升后,Liquid 会失去相对经济优势Liquid 靠边缘部署和定制化拉开差异

本表排序的是对可重复商业化验证最关键的依赖项,不是公司官网上可见的所有生态关系。

[CR014, CR024, CR025, CR026, CR027, CR028]
FR003: 依赖关系图

Liquid 公开可见的商业化图谱,仍然围绕少数芯片、渠道和灯塔客户关系展开。

[CR014, CR024, CR026, CR028, CR043]

7.4 人员、经济性与投资逻辑失效触发项

人员与经济性风险更关乎执行带宽,而不是标题上的资产负债表弱点。Liquid 对一家年轻公司来说资本充足,但公开来源仍显示规模快照混杂、深度研究和 GTM 岗位都在活跃招聘,董事会构成也只有第三方可见度。同时,公司正试图服务许多行业和部署模式,从创业公司和企业定制,到汽车和药物发现。Deloitte 和 MAPEGY 都显示,市场对 AI 和边缘推理的企业需求真实存在,但治理成熟度、物理 AI 就绪度和监管仍限制市场。这意味着 Liquid 竞争的品类中,买方需求在增长,实施摩擦也很高。若公司无法把合作伙伴牵头的证明转化为可复制企业部署,许可漏斗无法转成商业合同,或执行范围扩张超过组织稳定交付和支持产品的能力,投资逻辑就会失效。[CR008, CR009, CR010, CR011, CR012, CR036]

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
领导层 / 治理董事会构成和治理主要来自第三方追踪平台,而非公司披露中高创始研究团队经验丰富,Tracxn 披露了具名董事会成员索取董事会材料、委员会结构和投资者权利摘要
研究与工程招聘覆盖面显示,公司同时需要分布式训练、边缘推理、多模态工作和开发者关系能力新资金支撑招聘和研究连续性索取组织架构图、管理幅度和关键岗位空缺清单
商业化 / 解决方案公司一边面向多个行业销售,一边招聘解决方案架构、销售和市场岗位合作伙伴验证在多个垂直行业打开切入口按垂直行业审阅管线,并按客户类型核查实施能力
运营纪律文化明确推崇自主和低流程,能加快交付,也会拉扯可重复性中高白盒可解释性和绩效导向文化有助于问责要求说明事件复盘节奏、项目管理层和支持升级机制设计

组织结构、服务能力和正式治理纪律的公开能见度最低,执行风险也最高。

[CR009, CR010, CR011, CR012, CR030, CR044]
缓释措施与否决标准表
风险可监控触发项阈值 / 事件行动含义
Mercedes 商业化风险汽车部署里程碑到 2026 年底没有可信生产部署证据降低对物理 AI / 汽车上行空间的信心,并基于纯软件证据重新测算估值
授权转化风险商业授权采用管理层无法证明付费转化超过 $10M 门槛将开源到付费变现视为未验证,并下调收入质量假设
基准到生产缺口客户质量证据独立或客户证据显示核心生产任务质量偏弱下调护城河判断,新资金进入前要求更强的垂直行业验证
合作伙伴集中风险标杆客户广度到下一轮融资时,公开验证仍局限于 AMD、Mercedes 和 Insilico计入集中度折价,并假设企业采用更慢
执行摊子过大风险组织准备度对比垂直行业广度关键岗位仍空缺,行业野心却继续扩张承销增长前,要求明确排序、收窄 GTM 重点,并压实运营计划

这些触发项设计成未来尽调周期可观察的信号,而不是抽象风险表述。

[CR012, CR026, CR028, CR043, CR045]
Chapter 08

08估值

8.1 融资背景与价格纪律为何重要

Liquid AI 的融资信号真实存在:公司完成 $250 million 融资,公开追踪数据把这一轮的投后估值放在 $2 billion 左右,PitchBook 至少显示公司到 2025 年初已经开始产生收入。这足以把 Liquid 视为一家严肃公司,而不是一条实验室设想。但同一组公开材料也说明,为什么估值纪律很重要。Liquid 仍没有自有托管 API,公开材料没有披露收入、毛利率或留存,可见的商业化故事高度依赖授权、定制和合作伙伴主导的部署。换句话说,投资人看得到公司为什么值得关注,但还看不到通常支撑倍数法承销的经营指标。因此,合适的估值姿态要从价格敏感性出发:争论点不是 Liquid 有没有前景,而是今天的私募估值是否在更广泛客户转化和单位经济性可见之前,已经提前消化了太多显而易见的边缘 AI 上行空间。[CV001, CV002, CV003, CV004, CV005, CV006]

推荐摘要表
维度评估原因决策含义
推荐继续研究Liquid 具备战略吸引力,但公开材料仍缺少做出高信心入场判断所需的指标。保持跟踪,但在按当前估值承销上行空间前,需要更深入的尽调周期。
信心融资和同业锚点可见,但收入质量和转化数据仍未公开。将该推荐视为随证据变化,而不是最终结论。
风险评级竞争、商品化、治理和合作伙伴集中度都可能快速重置价值。采用里程碑触发项和下行阈值,而不是被动持有。
估值立场合理至偏高对于一家有差异化边缘 AI 定位的公司,~$2B 说得通;但和商业化证据更清楚的同业相比, 并不明显便宜。不要假设公司会自动长进本轮估值。
改善判断的因素更多证据或更好价格收入披露、利润率质量、合作伙伴转化和更清晰的股权结构表,都会提高承销信心。如果这些数据点够强,未来一轮融资或老股交易可能变得有吸引力。

本表刻意对价格敏感:它把公司质量、入场吸引力和披露质量拆开看。

[CV004, CV035, CV036, CV045, CV046, CV047]
FV001: 建议逻辑

建议质量不取决于是否喜欢这家公司,更取决于公开证据能否撑住当前价格。

[CV002, CV004, CV030, CV033, CV035, CV045]

8.2 可比公司组与市场结构检验

可比公司组给出的不是一个干净锚点,而是一个更细的结论。Writer 和 AI21 说明,面向企业的 AI 公司可以在 $1.4 billion 至 $1.9 billion 区间内生存,只要商业化叙事更明确。Cohere、Mistral 和 xAI 说明,当企业定位叠加更宽的平台广度、更大的算力规模,或更强的分发与公开验证时,这个类别可以拿到高得多的估值。与此同时,OpenAI 和 Google 的定价,加上 Artificial Analysis 和 Amadeus 的数据,说明核心模型能力正变得更容易横向比较,也更难单靠原始规模守住防线。这一点对 Liquid 很关键,因为它的战略差异化在高效部署,而不在压倒性的消费者分发或已披露收入规模。因此,可比公司组支持把 Liquid 放进高溢价企业 AI 的讨论里,但并不说明当前价格天然便宜。它的估值高过一部分企业同业,却还缺少最高梯队前沿实验室对应的公开证明。[CV009, CV010, CV011, CV012, CV013, CV014]

投资逻辑 / 反向逻辑表
视角当前观点改变观点的条件
架构投资逻辑在边缘、本地部署和主权 AI 场景中,Liquid 有一套差异化的高效部署叙事。如果独立客户证据显示该架构在生产中持续胜出,投资逻辑会明显增强。
商业化反向逻辑公司仍未公开收入、利润率和留存披露。更详细的 KPI 披露或更低入场价格会削弱这一担忧。
合作伙伴验证逻辑AMD 和 Mercedes 表明,这项技术可信度足以进入严肃的工业和设备场景。如果当前伙伴之外出现更广客户,验证就会从集中个案变成可重复模式。
商品化反向逻辑价值可能正从通用模型层迁向应用、数据和编排。如果 Liquid 证明自己实际掌握这些更高价值层,商品化威胁就会下降。
类别上行逻辑Cohere、Mistral 和 xAI 说明,只要分发或算力规模真实存在,私有市场估值可以被推得很高。Liquid 要可信地进入这一上行区间,需要更强的部署规模和市场能见度。
披露反向逻辑现有证据支持关注和继续尽调,但不足以给出明确买入推荐。管理层披露收入质量、股权结构条款和客户集中度,是评级上调的最短路径。

各行把上行情景和推动判断变化所需的具体缺失证据配对,而不是把信心当成静态变量。

[CV033, CV035, CV039, CV040, CV041, CV045]
可比估值表
可比公司使用指标倍数 / 估值 / 状态对 Liquid 的参考价值主要限制
Writer2024 年 11 月 Series C 轮 $1.9B 估值 企业 AI 平台,公开材料有客户和 ROI 表述。应用层重点和客户验证比 Liquid 更明确。
AI21 Labs2023 年 8 月 Series C 轮 $1.4B 估值 高效企业模型公司,强调自托管和安全部署姿态。轮次较早,产品组合 / 市场时点不同。
Cohere2026 年融资更新 据报道估值 $7B 主权企业 AI 参照,聚焦隐私和部署。规模显著更大,企业落地足迹更成熟。
Mistral AI2025 年 9 月 Series C 轮 11.7B€ 投后估值 显示市场能给企业 / 前沿模型平台多高估值。平台宽度和资本基础远大于 Liquid。
xAI2026 年 Series E 轮 融资 $20B;顶级算力叙事 说明分发加基础设施如何抬高估值天花板。消费者触达和算力规模不能与 Liquid 类比。
Liquid AI2024 年 12 月 Series A 轮 据报道投后估值 ~$2B 当前承销判断的锚点。公开来源没有披露收入、利润率或留存,无法验证该价格。

本清单覆盖本章使用的六个私募估值锚点。它们不是可互换的可比公司;参考价值和限制比标题估值更重要。

[CV002, CV004, CV015, CV018, CV020, CV022]
FV002: 估值敏感性

同行头部估值给出了 Liquid 被比较的区间,但并非所有同行都有同等商业化证明。

使用留存来源报道的已披露私募轮或头部融资锚点;币种按披露口径保留,不做汇率归一化。

[CV004, CV015, CV018, CV020, CV022, CV024]

8.3 投资逻辑、反向逻辑与情景承销

投资逻辑很直接:如果不想承担最大前沿实验室完整的资本开支和规模竞赛,Liquid 可能是投资高效主权 AI 的少数可信路径之一。公司有真实的架构故事,明确站在端侧和本地部署定位上,也有 AMD 与 Mercedes 这样的高价值合作伙伴验证。反向逻辑同样直接:基础模型价值可能正在离开通用模型层,公开定价压力真实存在,而 Liquid 还没有拿出足够的收入质量或部署转化数据,来支撑在当前估值下给出高确信度买入判断。这组张力决定了情景。乐观情景下,Mercedes 进入量产,Liquid 证明企业部署可以重复。基准情景下,Liquid 仍有战略吸引力,但只是慢慢长进一个今天大致合理的估值。悲观情景下,合作伙伴集中度和定价商品化把公司拉回较低梯队的企业 AI 可比公司。这个不对称性支持谨慎,而不是被动:继续跟踪公司,但在高价进入前要求更多证明。[CV025, CV026, CV027, CV028, CV029, CV030]

乐观 / 基准 / 悲观情景表
情景核心假设估值逻辑隐含价值概率信号
乐观Mercedes 进入生产,LEAP 成为可重复的企业边缘栈,管理层在合作伙伴胜利之外披露强劲商业转化。随着边缘部署证据拓宽,应用层防御力抵消商品化风险,估值框架可高于当前轮次、进入高溢价企业 AI 区间。$2.8B-$4.0B需要里程碑转化,还要证明收入质量好于当前公开记录。
基准Liquid 保持战略势能,但披露仍少,合作伙伴驱动的验证只会逐步扩展。最新一轮大致维持合理至偏高;公司是在长进估值,而不是明显跑到估值之上。$1.8B-$2.4B最符合当前公开验证和缺失的单位经济细节。
悲观合作伙伴里程碑延迟,商业化仍然狭窄,定价压力压缩独立模型经济性。当市场不再愿意为无差异模型敞口买单,公司估值会重估到更低的企业 AI 锚点。$1.0B-$1.6B触发因素包括里程碑落空、转化数据偏弱,或收入质量继续不透明。

情景价值是估值区间,不是预测;锚点来自公开私募轮次可比公司、类别结构,以及 Liquid 未披露经济性的具体不确定性。

[CV037, CV038, CV039, CV042, CV043, CV044]
投资逻辑破裂与否决触发项表
触发项阈值 / 事件对投资逻辑的传导行动含义
Mercedes 验证未能转化到 2026 年底没有可信的初始生产部署削弱最强的公开边缘商业化验证点。下调投资逻辑中的边缘溢价部分。
商业转化仍不透明管理层无法展示超出试点和下载的付费授权或部署转化变现故事会变成科学故事,而不是商业故事。转化证据出现前,持有或避免投入新资金。
类别商品化加速通用模型价格继续下跌,而 Liquid 缺少应用层验证即便效率更高,「又一家模型厂商」的价值也会缩水。将估值框架移向更低的企业 AI 锚点。
治理轨迹落后于部署野心受监管或自主部署扩张,却没有强验证和监督材料增加敏感行业的采购摩擦和下行风险。承销受监管市场上行空间前,先要求治理材料包。
披露缺口延续到下一次融资事件到下一轮仍看不到收入、利润率、留存或集中度使市场无法用基本面验证私有估值标记。基本面出现前,将任何上轮估值融资都视为情绪驱动。

否决触发项以里程碑为基础,未来尽调可以直接检验,而不是依赖叙事印象。

[CV026, CV030, CV034, CV035, CV039, CV044]
FV003: 估值 / 回报区间

情景区间的核心,是 Liquid 能否证明可重复部署,并披露足够经济性来守住当前估值标记。

[CV042, CV043, CV044]
FV004: 投资 KPI

Liquid 在战略差异化上得分较高,但公开商业化可见度和估值支撑偏弱。

[CV025, CV027, CV033, CV035, CV047, CV048]

8.4 哪些变化会改变判断,哪些因素会打破逻辑

Liquid 不必改变科学路线,也能提高确信度;它主要需要补上信息缺口。最有价值的尽调问题是:按产品和部署模式拆分的收入、毛利率质量、企业合同结构、客户集中度,以及合作伙伴主导的里程碑是否正在转化为可重复的生产使用。Mercedes 尤其重要,因为它是最清晰的近期待验证据,能证明 Liquid 的边缘 AI 逻辑可以在实验室基准之外发挥作用。反面就是投资逻辑失效清单。如果公司无法证明付费商业转化超过公开许可门槛,如果合作伙伴验证仍然狭窄,如果部署进入监管场景后治理和合规材料仍然单薄,或者市场继续让通用模型商品化的速度快过 Liquid 证明应用层防御力的速度,那么私募估值就应该下调。在管理层补齐这些缺失指标之前,正确答案不是永远回避,而是继续研究、谨慎定价,并坚持未来尽调围绕里程碑,而不是围绕故事。[CV033, CV034, CV035, CV036, CV042, CV044]

最终尽调要求表
主题缺失证据重要性负责人 / 尽调路径
收入质量按产品、渠道和部署模式拆分的 ARR决定这门生意配不配软件溢价,还是只有战略期权价值。管理层 KPI 包和董事会材料。
利润率和现金画像2024 年融资后的毛利率、托管 / 支持负担和现金消耗检验高效架构是否真的带来经济杠杆。财务尽调和经审计财务审阅。
客户集中度头部客户、管线结构和伙伴依赖判断公开标杆交易是否高估了可复制性。销售运营导出数据和客户推荐计划审阅。
Mercedes 里程碑转化按车型年拆分的试点、验证和量产就绪证据直接检验最强的边缘商业化证据点。与产品、OEM 合作伙伴和客户负责人做项目审阅。
许可变现超过公开免费使用门槛后的付费合同条款检验从开放使用到商业付费的漏斗是否跑得通。法律 / 定价审阅,加已签企业合同样本。
股权结构和优先权清算顺位、参与权和二级市场悬空筹码影响实际入场价格和下行保护,且独立于公司质地。公司律师和融资文件审阅。

这些是最小一组尽调请求,最可能直接改变建议或估值立场。

[CV035, CV036, CV042, CV045, CV046, CV047]

免责声明

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

证据索引

结论
编号陈述可信度来源
CO001 Liquid AI was founded in 2023. SO003, SO011, SO016, SO018, SO019
CO002 The best-supported founder set is Ramin Hasani, Mathias Lechner, Alexander Amini, and Daniela Rus. SO003, SO011, SO016
CO003 Liquid AI publicly presents itself as an MIT CSAIL spinout rooted in liquid-neural-network research. SO003, SO011, SO009
CO004 Ramin Hasani is publicly identified as Liquid AI’s CEO and co-founder. SO003, SO011, SO016
CO005 Mathias Lechner is publicly identified as Liquid AI’s CTO and co-founder. SO003, SO011, SO020
CO006 Alexander Amini is publicly identified as Liquid AI’s chief science officer and co-founder. SO003, SO011, SO016
CO007 Daniela Rus appears in retained sources as a co-founder whose institutional role adds MIT CSAIL credibility to Liquid AI. SO003, SO011
CO008 Cambridge, Massachusetts is the best-supported current headquarters marker in the retained public record. SO018, SO019, SO024
CO009 Liquid’s public operating footprint includes Boston and San Francisco roles in addition to its Cambridge-area identity. SO011, SO024, SO025
CO010 Liquid says it builds capable and efficient general-purpose AI systems for real-world environments where compute is constrained. SO001, SO008
CO011 Liquid’s product family includes Liquid Foundation Models, the LEAP deployment platform, and the Apollo local app. SO005, SO007, SO008
CO012 Liquid positions LFMs as multimodal models spanning text, vision, audio, and other sequential data. SO001, SO005
CO013 Liquid claims its models can be deployed across CPU, GPU, and NPU environments. SO005, SO008, SO028
CO014 Liquid Docs expose multiple downloadable model formats, including GGUF, MLX, and ONNX, alongside trainability guidance. SO023
CO015 Liquid explicitly says it does not currently offer a hosted API of its own. SO006, SO008
CO016 Liquid routes public and enterprise access through a mix of playground usage, OpenRouter, Hugging Face downloads, LEAP, and direct sales contact. SO006, SO008, SO022
CO017 Liquid announced a US$46.6 million seed round when it emerged from stealth in December 2023. SO003, SO011, SO017
CO018 OSS Capital and PagsGroup were the named lead investors on Liquid’s seed round. SO003, SO011, SO017
CO019 The disclosed seed syndicate also included Samsung Next, Automattic, Naval Ravikant, Tobias Lütke, Tom Preston-Werner, and Bob Young among others. SO003, SO011, SO017
CO020 Liquid announced a US$250 million Series A in December 2024 led by AMD. SO004, SO012, SO013, SO017
CO021 Independent coverage supports a public valuation marker above roughly US$2 billion for Liquid’s Series A. SO012, SO013
CO022 Adding the disclosed seed and Series A amounts implies about US$296.6 million of publicly disclosed primary capital. SO003, SO004
CO023 Tracxn rounds Liquid’s total raised capital to US$297 million and its valuation to US$2 billion. SO016, SO017
CO024 The AMD-led financing was paired with an explicit model-optimization and deployment relationship on AMD hardware. SO004, SO028
CO025 Liquid publicly scheduled its first major product-unveiling event for October 23, 2024 at MIT after teasing the launch on October 15, 2024. SO007, SO014
CO026 Liquid launched LEAP and Apollo in July 2025 to extend from model architecture into edge deployment workflows and a local consumer experience. SO007
CO027 Liquid announced a commercial partnership with G42 in June 2025. SO007
CO028 Liquid announced native Ryzen and Ryzen AI support for LEAP in August 2025. SO007, SO028
CO029 Liquid announced a multi-year Shopify partnership in November 2025 focused on sub-20 millisecond commerce experiences. SO007
CO030 Liquid and Insilico Medicine announced a March 2026 partnership centered on a 2.6B-parameter scientific model for drug discovery. SO007, SO026
CO031 Mercedes-Benz and Liquid AI announced an April 2026 embedded in-car AI partnership targeting first production deployment in the second half of 2026. SO007, SO027
CO032 Liquid publicly targets sectors including automotive, e-commerce, financial services, biotechnology, and consumer electronics. SO004, SO007, SO008, SO026
CO033 Liquid’s technical differentiation is explicitly tied to liquid neural networks, continuous-time learning systems, and state-space methods. SO009, SO021
CO034 Liquid’s research page shows the company continued publishing technical work into 2026. SO009, SO007
CO035 Constellation Research said early LFMs were still a work in progress and specifically weak at zero-shot code, precise numerical calculation, time-sensitive information, and human-preference optimization. SO015
CO036 Mathias Lechner’s July 2025 biography cites a US$2.3 billion valuation and names AMD, Shopify, Samsung, and G42 around the Series A, which is more expansive than the official December 2024 disclosures. SO020
CO037 Liquid’s public headcount is not canonical because PitchBook, Hugging Face, and Tracxn imply materially different team sizes. SO016, SO018, SO022
CO038 The retained public source pack does not disclose Liquid’s revenue, ARR, or customer count. SO006, SO007, SO019
CO039 Tracxn lists a six-person board that includes Joseph Jacks, Daniela Rus, Ramin Hasani, Stephen Pagliuca, Louis Hunt, and Mathias Lechner. SO016
CO040 Liquid does not publish enough public information to verify current cap-table control, investor rights, or secondary liquidity terms. SO016, SO017, SO018
CO041 Liquid’s 2026 open license allows free commercial use only for organizations below US$10 million in annual revenue. SO010
CO042 The same license structure suggests Liquid is combining open distribution with a paid enterprise-licensing path for larger commercial users. SO006, SO010, SO008
CO043 Current job postings span GTM, sales, finance, developer relations, localization, and research, indicating commercialization beyond a pure research lab. SO024, SO025
CO044 Liquid’s about page emphasizes a meritocratic, product-driven, and white-box-explainable internal culture. SO002
CM001 Liquid’s practical market is narrower than the full foundation-model category and is best framed around deployable, efficient AI for private, local, or hybrid workflows. SM001, SM002, SM009
CM002 Liquid’s enterprise pages consistently pitch on-device, cloud, or hybrid deployment where latency, privacy, security, and compute efficiency matter. SM001, SM002
CM003 Liquid’s automotive page centers the market on embedded in-vehicle assistants that must run under hardware, latency, and privacy constraints. SM003
CM004 Liquid’s ecommerce page centers the market on search, recommendations, checkout agents, and cost pressure relative to retail margins. SM004
CM005 Liquid’s financial-services page centers the market on fraud, payments, trading, and other privacy-sensitive workflows on or near private infrastructure. SM005
CM006 Liquid’s startup and community pages show a second market motion aimed at venture-backed builders and developers seeking small models, docs, tooling, and mentorship. SM006, SM007
CM007 Liquid’s pricing page says the company does not currently offer a hosted API of its own. SM009
CM008 Pure token-based hosted API spend is therefore best treated as an adjacent substitute market rather than Liquid’s clearest current fit. SM001, SM009, SM016, SM017
CM009 OpenAI, Google, Anthropic, Writer, and xAI all market hosted or SaaS-style access models that appeal to buyers prioritizing convenience and immediate consumption. SM016, SM017, SM018, SM019, SM020
CM010 Cohere and Mistral explicitly market private, VPC, self-hosted, dedicated, or hybrid enterprise deployment options. SM023, SM024
CM011 Microsoft Phi, Meta Llama, and Qualcomm materials show a parallel small-model and on-device ecosystem competing for local deployment use cases. SM021, SM022, SM025
CM012 Deloitte reports that worker access to AI rose 50% in 2025. SM010
CM013 Deloitte reports that the number of companies with at least 40% of AI projects in production is set to double in six months. SM010
CM014 Deloitte reports that 58% of companies already use physical AI in at least a limited way and that figure is set to reach 80% in two years. SM010
CM015 Fortune Business Insights and Verified Market Reports both place the global edge AI market at US$47.59 billion in 2026 and US$385.89 billion by 2034. SM012, SM013
CM016 Stratistics MRC places the global edge AI inference market at US$153.84 billion in 2026 and US$635.51 billion by 2034. SM014
CM017 The large gap between the US$47.59 billion and US$153.84 billion 2026 estimates shows that public reports are using materially different market boundaries. SM012, SM013, SM014
CM018 Fortune Business Insights says automotive accounted for 24.54% of the edge AI market in 2026. SM012
CM019 Fortune Business Insights says hardware accounted for 62.41% of the edge AI market in 2026. SM012
CM020 Verified Market Reports highlights automotive, enterprise robotics, drones, head-mounted displays, smart speakers, phones, PCs or tablets, and security cameras as primary edge AI applications. SM013
CM021 Wevolver says 2026 on-device language models increasingly cluster in a sub-billion to single-digit-billion parameter Goldilocks zone because of thermal, power, and memory limits. SM011
CM022 Liquid’s community page claims sub-100 millisecond time-to-first-token on AMD processors with zero dependency on cloud APIs. SM007
CM023 Liquid’s community page says Apollo runs entirely on-device with no internet or API calls. SM007
CM024 Liquid’s open license allows free commercial use only for organizations below US$10 million in annual revenue. SM008
CM025 That revenue threshold can create adoption friction for scaling startups and mid-market software companies that outgrow the free license. SM006, SM008
CM026 Meta Llama 3 uses a custom community license with additional commercial terms for organizations above 700 million monthly active users. SM022
CM027 The small-model market is therefore competitive but license-fragmented rather than fully permissive. SM008, SM022, SM027
CM028 Artificial Analysis evaluates competition across intelligence, openness, price, speed, and latency instead of one single accuracy metric. SM015
CM029 Buyer and budget ownership differ by workflow: automotive OEM software teams, ecommerce digital-product teams, financial-services transformation teams, and startup builders all have distinct adoption paths. SM003, SM004, SM005, SM006, SM007
CM030 Automotive buyers are best understood as OEM software or infotainment teams, with drivers and passengers as end users and vehicle-program budgets as payers. SM003, SM010
CM031 Ecommerce buyers are best understood as digital product, search, merchandising, or customer-experience teams, with shoppers and merchants as end users and commerce budgets as payers. SM004
CM032 Financial-services buyers are best understood as CIO, AI platform, risk, or operations leaders, with employees and customers as users and transformation or compliance budgets as payers. SM005, SM010
CM033 Startup and developer buyers are best understood as technical founders and product engineers, with product or venture budgets as payers. SM006, SM007
CM034 Liquid’s adoption path requires benchmarking, customization, hardware fit, governance review, and deployment support rather than a trivial API switch. SM001, SM002, SM007, SM009
CM035 Deloitte identifies the AI skills gap as the biggest barrier to integrating AI into existing workflows. SM010
CM036 Stratistics highlights complex deployment and maintenance across distributed edge devices as a threat to market adoption. SM014
CM037 Verified Market Reports stresses privacy, localized processing, and data sovereignty as core market drivers for edge AI. SM013
CM038 Hosted API substitutes can win when convenience and low deployment overhead matter more than data sovereignty or hardware fit. SM009, SM016, SM017, SM018, SM019, SM020
CM039 Liquid’s immediate serviceable market is narrower than the full public edge-AI TAM and is best framed around regulated or latency-critical deployments in automotive, commerce, finance, and enterprise endpoints. SM001, SM002, SM003, SM004, SM005
CM040 Public sources do not show Liquid’s actual paid-customer count, conversion rate from community usage, or revenue mix by vertical. SM006, SM007, SM008, SM009
CM041 Published market data is directional but not precise enough to defend a single TAM, SAM, or SOM number without internal pipeline evidence. SM012, SM013, SM014
CM042 Liquid’s no-hosted-API posture means it competes less on public token price and more on total deployment economics, privacy, and resilience. SM009, SM016, SM017, SM018, SM019
CP001 Liquid positions itself as an efficient general-purpose AI vendor for compute-constrained edge environments. SP001, SP002
CP002 Liquid says its LFMs are multimodal hybrid models built for agentic tasks, instruction following, data extraction, and RAG. SP002
CP003 Liquid documentation advertises 32K context across the library and 128K context for LFM2.5-8B-A1B. SP005
CP004 Liquid says it does not currently operate a hosted API of its own and instead offers a free playground, paid OpenRouter access, model downloads, and LEAP customization. SP003, SP002
CP005 Liquid pitches enterprise deployments as cost-efficient, secure, and available on device, in the cloud, or in hybrid form. SP004, SP001
CP006 Liquid’s open license ends free commercial use once a company reaches $10 million in annual revenue. SP006
CP007 OpenAI describes GPT-4o as a real-time multimodal model that accepts text, audio, image, and video inputs. SP007
CP008 OpenAI API pricing lists GPT-5.5 at $5.00 per million input tokens and $30.00 per million output tokens. SP008
CP009 Anthropic publishes seat-based plans and flagship model pricing with a $5-per-million-token input tier and a $25-per-million-token output tier. SP009
CP010 Google publishes Gemini API pricing with free usage tiers plus paid token, storage, and search-query fees. SP011
CP011 Google presents Gemini 3.5 as an active model family competing on capabilities and performance for general-purpose AI workloads. SP010
CP012 Meta markets Llama 3 as an openly available model family. SP012
CP013 The Meta Llama 3 model card says users must share contact information to access weights on Hugging Face. SP013
CP014 The same model card points users to a custom commercial license path and direct checkpoint download instructions. SP013
CP015 Mistral markets its studio platform around enterprise privacy, governance, and deployment ownership. SP014, SP015
CP016 Cohere positions itself as private, secure, and customizable enterprise AI rather than as a consumer chat product. SP016
CP017 Cohere describes Command A as a high-performance model built with minimal compute requirements. SP017
CP018 AI21 markets Jamba as an open enterprise model family that emphasizes speed, cost efficiency, and secure workflows. SP018
CP019 Writer positions itself around regulated-enterprise workflows, governance, and compliance rather than raw model openness. SP019, SP020
CP020 Writer’s pricing page says enterprise plans include seat types, zero data retention by default, and no model training on customer data by default. SP020
CP021 xAI markets Grok as a combined chat, search, reasoning, voice, image, and video product. SP021
CP022 xAI’s docs say realtime events require search tools to be enabled. SP022
CP023 Microsoft markets Phi as an open small-model family with multimodal variants covering text, audio, and vision. SP023
CP024 Azure Foundry is sold as an interoperable enterprise AI platform with centralized governance and security. SP024
CP025 Qualcomm AI Hub is a hardware-adjacent substitute route for deploying models on devices outside Liquid’s own stack. SP025, SP001
CP026 Artificial Analysis benchmarks model intelligence, openness, and blended price across many AI vendors on one price-quality frontier. SP026
CP027 Liquid says LEAP support on AMD Ryzen and Ryzen AI processors can deliver sub-100-millisecond responsiveness on device. SP027
CP028 Mercedes-Benz announced a multi-year partnership with Liquid to scale embedded in-car intelligence in North America. SP028
CP029 Liquid’s closest direct peer set is the efficient-enterprise model cluster that includes Mistral, Cohere, AI21, Writer, Microsoft Phi, and xAI rather than only the largest frontier API labs. SP015, SP016, SP018, SP019, SP023, SP021
CP030 OpenAI, Google, Anthropic, and xAI compete primarily on hosted-model breadth and public pricing transparency rather than Liquid’s embedded-deployment thesis. SP007, SP008, SP009, SP010, SP011, SP022
CP031 Meta Llama, Microsoft Phi, and Liquid all offer downloadable or local-deployment paths that raise substitution risk for buyers who prioritize control. SP003, SP005, SP013, SP023
CP032 Liquid’s lack of a first-party hosted API narrows self-serve monetization relative to OpenAI, Anthropic, Google, and xAI. SP003, SP008, SP009, SP011, SP022
CP033 Liquid’s license threshold can attract small developers while still forcing larger commercial users toward direct negotiation. SP006, SP004
CP034 Competitor distribution moats are structurally different across OpenAI APIs, Google enterprise platforms, Microsoft Azure, and Writer’s workflow software. SP008, SP010, SP011, SP024, SP020
CP035 Liquid’s own distribution moat is emerging through embedded-device and hardware partnerships rather than mass public developer traffic. SP004, SP027, SP028
CP036 Buyer multi-homing is likely to remain high because hosted APIs, open-weight models, and device-specific deployment tools can be mixed in one stack. SP003, SP013, SP023, SP025
CP037 Switching costs are strongest only after a buyer commits to LEAP customization, private deployment, or embedded-device optimization. SP002, SP004, SP027
CP038 Commoditization risk is elevated because efficient or open alternatives from Meta, Microsoft, AI21, Mistral, and Cohere are all available to enterprise buyers. SP013, SP015, SP017, SP018, SP023
CP039 Liquid’s moat is strongest where privacy, low latency, and hardware constraints matter more than raw public leaderboard visibility. SP001, SP004, SP027, SP028
CP040 Public benchmark visibility still favors vendors with public APIs and wide benchmark coverage more than private-deployment vendors like Liquid. SP026, SP003
CI001 Liquid says it does not currently offer a hosted API of its own. SI001
CI002 Liquid routes self-serve discovery through a free playground, paid OpenRouter access, direct downloads, and LEAP customization. SI001, SI005
CI003 Liquid’s license grants broad rights but ends free commercial use once a company reaches $10 million in annual revenue. SI003
CI004 The public monetization stack points to revenue from enterprise deployment, support, licensing, and customization rather than from a first-party token-metered API. SI001, SI002, SI003
CI005 Liquid announced a $250 million financing round to scale capable and efficient general-purpose AI. SI004, SI007
CI006 Independent coverage placed the financing at a valuation above $2 billion. SI007, SI008
CI007 PitchBook and Tracxn both profile Liquid as a private company founded in 2023 and now at the Series A stage. SI009, SI010
CI008 Liquid says the AMD-tuned LEAP path gives developers a direct route for on-device deployment. SI011, SI039
CI009 Mercedes-Benz announced a multi-year partnership with Liquid for embedded in-car intelligence in North America. SI012, SI033
CI010 The Insilico partnership frames Liquid’s architecture as usable on private pharmaceutical infrastructure. SI013, SI034
CI011 Liquid documentation shows a downloadable library of models with standard 32K context and one 128K variant. SI005, SI006
CI012 Liquid’s public pricing page does not disclose realized enterprise contract values, token rates, or support fees. SI001
CI013 OpenAI publishes explicit token prices that buyers can compare directly against Liquid’s opaque enterprise packaging. SI015, SI001
CI014 OpenAI lists GPT-5.5 at $5 per million input tokens and $30 per million output tokens. SI015
CI015 Anthropic publishes both seat plans and flagship model pricing including a $5-per-million-token input tier and a $25-per-million-token output tier. SI016
CI016 Google publishes Gemini API pricing with free usage, paid token fees, cache-storage fees, and search-query charges. SI017
CI017 Writer sells enterprise and starter plans around seat types instead of exposing a simple public token card. SI018, SI024
CI018 Liquid’s list-price opacity means outsiders can see the route to monetization without seeing actual ASP, discounting, or revenue mix. SI001, SI002, SI003
CI019 Liquid’s public GTM proxies are concentrated in enterprise vertical pages and partnership announcements rather than disclosed customer or ARR metrics. SI002, SI011, SI012, SI013, SI027, SI028, SI029, SI030, SI032, SI035, SI036
CI020 Liquid’s own messaging emphasizes cost-efficient, secure, and private AI for enterprises rather than a mass-consumer subscription business. SI002, SI022
CI021 Reviewed public sources do not disclose Liquid’s customer count, ARR, gross margin, or headcount. SI001, SI002, SI004, SI007, SI008, SI009, SI010
CI022 Microsoft’s 2025 10-K is an audited example of the revenue, margin, and capital disclosures that are absent from Liquid’s public record. SI021
CI023 Using public filings as comparators highlights how incomplete Liquid’s current public financial disclosure is for underwriting. SI021, SI001, SI004, SI007
CI024 Liquid’s revenue bridge likely starts with free exploration and downloads and moves into paid enterprise deployment, support, and licensing. SI001, SI002, SI003, SI005
CI025 Liquid’s route to market depends partly on hardware and enterprise partners rather than solely on direct API traffic. SI011, SI012, SI013
CI026 Because Liquid does not run a first-party hosted API, its gross margin profile likely depends more on software-services mix and partner deployment economics than on hyperscale inference revenue. SI001, SI002, SI011, SI012
CI027 The license threshold creates a natural upsell trigger once a project grows beyond research or small-company use. SI003, SI002
CI028 Comparator pricing from OpenAI, Anthropic, and Google gives buyers transparent anchors when Liquid pitches bespoke contracts. SI015, SI016, SI017, SI001
CI029 Transparent competitor rate cards create adverse pricing pressure because Liquid has fewer public reference points to justify premium contracts. SI015, SI016, SI017, SI020
CI030 Capital adequacy improved after the $250 million round, but runway still cannot be underwritten because cash and burn remain undisclosed. SI004, SI007, SI008
CI031 Funding dependence remains material because Liquid is still private, early in commercialization, and building deployment-heavy enterprise products. SI007, SI009, SI010, SI014, SI035, SI036, SI037, SI038, SI042, SI043
CI032 Private-company gaps remain on revenue mix, customer concentration, NRR, burn, gross margin, and sales efficiency. SI001, SI002, SI007, SI009, SI010
CI033 Liquid’s on-prem and device-oriented positioning could improve unit economics if customers shoulder more infrastructure footprint, but public evidence does not quantify the effect. SI002, SI006, SI011, SI012
CI034 Comparable AI vendors monetize through a mix of token metering, seat subscriptions, and custom enterprise deals, so Liquid should not be underwritten as a pure SaaS or pure API company. SI015, SI016, SI017, SI018, SI023
CI035 The financial verdict from public evidence is that the revenue mechanism is understandable but realized economics remain blocked on management disclosure. SI001, SI002, SI003, SI004, SI007
CI036 Public use-of-funds evidence supports spending on model development, scaling, deployment, and enterprise expansion rather than proving present revenue quality. SI004, SI007, SI041, SI042, SI043
CI037 Mercedes and Insilico prove some enterprise willingness to adopt Liquid’s architecture, but they do not reveal recurring revenue or contract profitability. SI012, SI013, SI031
CI038 Independent price-comparison services show a crowded price-performance environment across AI models. SI019, SI020
CI039 Liquid’s lack of first-party API pricing makes outside benchmarking of its token economics materially harder than benchmarking OpenAI, Anthropic, or Google. SI001, SI015, SI016, SI017, SI020
CI040 Microsoft’s filing demonstrates how important audited cash-flow and capital disclosures become in AI infrastructure businesses, underscoring the significance of Liquid’s undisclosed burn and capex posture. SI021, SI004
CI041 Liquid publicly markets automotive, ecommerce, financial-services, and startup solutions, indicating a vertical GTM thesis broader than a single generic AI landing page. SI028, SI029, SI030, SI032
CI042 Liquid maintains a case-studies surface that supports the existence of reference customers even though contract economics remain undisclosed. SI031
CI043 Deloitte’s 2026 enterprise AI survey supports the view that enterprise buyers are still expanding AI adoption budgets and governance efforts. SI035
CI044 Edge AI market reports support the commercial relevance of device-side inference and embedded deployment channels that Liquid targets. SI036, SI037, SI038, SI040
CI045 Liquid's LEAP platform now markets a packaged workflow to discover, specialize, and deploy models locally on supported devices in minutes. SI044
CE001 Liquid positions LFMs as efficient general-purpose multimodal systems built for smartphones, laptops, vehicles, embedded systems, and hybrid cloud-edge use. SE001, SE005
CE002 Liquid's public stack includes text models, vision-language models, an audio model, and nano models rather than a single text-only SKU. SE005, SE009
CE003 The docs say public LFMs ship in GGUF, MLX, and ONNX formats and can run through Transformers, llama.cpp, vLLM, MLX, Ollama, and LEAP. SE009, SE017
CE004 Liquid says LFM2.5 extends LFM2 with 28 trillion tokens of pretraining and a scaled reinforcement-learning pipeline across text, vision, and audio models. SE005, SE008
CE005 Liquid traces LFMs back to liquid-neural-network work rooted in dynamical systems, signal processing, and numerical linear algebra rather than pure transformer scaling. SE006, SE007, SE003, SE010
CE006 Liquid's research page claims earlier liquid-neural-network work established adaptability, causal and interpretable behavior, and efficient long-term dependency handling for sequential data. SE007, SE028
CE007 Liquid's public research surface continued through 2026 with new papers and listed an LFM2 Technical Report in December 2025, implying active post-launch architecture work. SE008, SE002
CE008 Liquid's pricing page says the company does not currently offer its own hosted API and instead routes access through a free playground, OpenRouter, direct downloads, and LEAP. SE016
CE009 Liquid says enterprises can purchase full local access to LFMs plus an on-prem customization stack for private, safety-critical, and latency-bound use cases behind the enterprise firewall. SE005, SE016
CE010 Liquid describes LEAP as a unified developer platform for customizing and deploying LFMs across any device and operating system in a single workflow. SE017, SE016
CE011 Liquid and AMD say the LEAP SDK now supports Ryzen and Ryzen AI processors with low-latency, memory-optimized models and zero dependency on cloud APIs. SE023, SE036
CE012 TechIntelPro reports that the AMD-tuned LEAP path is built on the open-source llama.cpp inference engine, offering one rare public clue about LEAP internals. SE036
CE013 Apollo is marketed as a low-latency, cloud-free mobile playground that runs entirely on the device with no internet, logging, or API calls. SE022, SE017
CE014 Liquid's automotive offer centers on edge-first small language models, multimodal assistants, and hybrid edge-cloud agentic architectures for in-vehicle functions. SE019, SE033
CE015 Liquid's ecommerce offer targets intent-native search, storefront-connected agents, and merchant copilots tied to catalog, checkout, and campaign data. SE020, SE025
CE016 Liquid's financial-services page pitches private, low-latency models for fraud, payments, trading, and other workflows constrained by sensitive on-prem data. SE021, SE025
CE017 Liquid's enterprise positioning emphasizes tailored, secure, cost-efficient AI deployed on device, in the cloud, or in hybrid mode rather than an off-the-shelf hosted service. SE018, SE001
CE018 Liquid's public case-studies page says an automotive vision-language deployment was 50% smaller, 10x faster, and deployed within one week on existing vehicle hardware. SE025
CE019 Liquid's case-studies page says a fraud-prevention deployment ran 2x faster and was tied to an estimated roughly $230 million in additional annual fraud detection. SE025
CE020 Liquid's ecommerce case-study summary says a compact model for product cataloging delivered 65% faster deployment with better accuracy and no higher infrastructure cost. SE025
CE021 Liquid's case-studies page says a locally deployed synthetic-video workflow cut cloud bottlenecks and reduced costs by 70 percent. SE025
CE022 Liquid's LFM Open License allows broad use but ends free commercial rights once a company exceeds $10 million in annual revenue. SE027
CE023 Liquid's license says derivative models can remain proprietary and are not subject to a copyleft release obligation. SE027
CE024 Liquid's first public LFM release centered on roughly 1.3B, 3.1B, and 40.3B-MoE model classes. SE006, SE029, SE031
CE025 Liquid publicly reports MMLU, GPQA, IFEval, IFBench, GSM8K, MGSM, and MMMLU benchmark scores for its LFM2 line across multiple sizes. SE005, SE006
CE026 Independent launch coverage broadly corroborated the public model lineup and Liquid's narrative that the smallest LFMs beat similarly sized transformer peers on several benchmarks. SE029, SE031, SE032
CE027 Independent coverage highlighted Liquid's claim that LFM-3B can deliver long-context performance with materially lower memory needs than comparable transformer models. SE029, SE030
CE028 Constellation Research noted LFMs were still weak on zero-shot code, precise numerical calculation, time-sensitive information, and human-preference optimization at launch. SE030
CE029 VentureBeat reported the first public LFMs were still in preview and that Liquid was explicitly inviting early feedback and red-teaming before broader rollout. SE029
CE030 Mercedes and Liquid say the in-car intelligence partnership has a clear path to first production deployment in the second half of 2026, which is material proof but still forward-looking. SE033, SE019
CE031 Liquid and Insilico say the LFM2-2.6B-MMAI checkpoint is available now and runs entirely on private pharmaceutical infrastructure. SE024, SE034, SE035
CE032 Liquid says the Series A capital will fund edge and on-prem product readiness, including inference and fine-tuning stacks, rather than only more research hiring. SE004, SE023
CE033 TechCrunch reported that Liquid planned from its 2023 launch to provide on-prem private AI infrastructure and a platform for customers to build their own models. SE028, SE016
CE034 Liquid's about page now presents LFMs, LEAP, Apollo, docs, and community resources as one coherent product stack rather than isolated research outputs. SE002, SE017
CE035 Liquid maintains an active Hugging Face organization with multiple LFM2.5 and vision-language checkpoints, giving the company a visible public developer distribution surface. SE015, SE017
CE036 Liquid's docs show several public models are trainable through TRL and available in formats tailored to local CPU, GPU, Apple Silicon, and ONNX deployment paths. SE009, SE015
CE037 Liquid's October 2024 launch event included panel appearances from AMD, Deloitte, Microsoft, Shopify, Samsung Next, and Capgemini representatives, signaling an ecosystem-led go-to-market posture. SE026, SE004
CE038 Liquid's January 2026 LFM2.5 launch says the 1.2B family extends pretraining from 10T to 28T tokens and adds Japanese, vision-language, and audio variants for on-device agents. SE037
CE039 Liquid's LFM2-1.2B Hugging Face page says LFM2 ships openly across 350M, 700M, 1.2B, and 2.6B checkpoints with faster training and CPU inference positioning for edge apps. SE038
CE040 Enterprise AI World reported LFM2 as purpose-built for local and edge use cases, citing 2x faster CPU decode and prefill than Qwen3 alongside stronger small-model performance. SE039
CE041 Liquid's official documentation repository says the docs cover open-weight LFMs and the LEAP SDK on laptops, mobile, and edge devices, reinforcing a real developer-platform packaging layer. SE040
CU001 Liquid publicly targets enterprise, startup, automotive, ecommerce, financial-services, healthcare, industrial, and developer-community segments rather than a single buyer class. SU001, SU002, SU007, SU008, SU009, SU005
CU002 Liquid's public vertical pages imply different buyers and users by segment, including enterprise AI teams, automaker software teams, pharma researchers, merchant operators, fraud leaders, and developers. SU006, SU007, SU008, SU009, SU005
CU003 Liquid does not publicly disclose total customer count, active enterprise count, or paid-account count across its overview, pricing, funding, or case-study pages. SU001, SU004, SU010, SU003
CU004 The clearest public named counterparties in Liquid's customer proof set are Mercedes-Benz and Insilico Medicine. SU016, SU015, SU017
CU005 Mercedes-Benz describes its arrangement with Liquid as a multi-year partnership tied to third- and fourth-generation MBUX models in North America. SU016
CU006 Mercedes says the collaboration targets first production deployment in the second half of 2026, so the public proof is still pre-production as of the run date. SU016, SU007
CU007 The Insilico collaboration is stronger near-term proof because the parties say the scientific model is available now on private pharmaceutical infrastructure. SU015, SU017, SU018
CU008 Liquid's case-studies page describes an unnamed global automaker that deployed a tailored vision-language model to current head units with faster in-vehicle response. SU003, SU007
CU009 Liquid's case-studies page also describes an unnamed fraud-prevention deployment with 2x faster processing and estimated nine-figure annual savings. SU003, SU009
CU010 Liquid's public proof set includes an unnamed ecommerce cataloging deployment with faster rollout and better accuracy, but no named merchant logo. SU003, SU008
CU011 Liquid says it does not run a hosted first-party API, which means the public customer funnel is split between free evaluation channels and bespoke enterprise engagements. SU004, SU011
CU012 Liquid's enterprise materials push visitors into sales-led custom-solution conversations instead of a transparent self-serve B2B SaaS motion. SU006, SU004
CU013 Liquid maintains a parallel evaluation funnel through docs, cookbooks, hackathons, browser play, Hugging Face, and Apollo that broadens reach without proving paid conversion. SU005, SU012, SU011
CU014 Liquid says it is testing AI products in market with key partners across sectors including consumer electronics, telecom, finance, ecommerce, and biotech, but it does not name the full partner set. SU010, SU019
CU015 Deloitte reports that only one in five companies has mature governance for autonomous AI agents, which is a real barrier to turning pilot interest into production rollouts. SU020
CU016 Deloitte says only 20 percent of organizations are already seeing AI-driven revenue growth even though far more expect it in future, implying many deployments are still pre-revenue proofs. SU020
CU017 Deloitte sees production use rising, but the report still frames many organizations as moving from pilot to scale rather than already scaled. SU020
CU018 Deloitte says companies feel more strategically ready for AI than operationally ready on infrastructure, data, risk, and talent, which could slow enterprise adoption of Liquid. SU020
CU019 Mercedes owns the vehicle OS, customer relationship, and production timetable, leaving Liquid dependent on the OEM partner for one of its highest-credibility public deployments. SU016, SU007
CU020 Liquid reserves customized support and full local access for sales-led enterprise engagements, reinforcing a high-touch motion with fewer, larger accounts. SU004, SU006
CU021 Liquid's public case-study surface is stronger on vertical outcomes than on named customer logos, which limits reference quality for underwriting. SU003, SU002
CU022 Liquid's about page references enterprise, startup, silicon, and ecosystem partners without surfacing readable partner names in the fetched text. SU002
CU023 Liquid's public vertical pages consistently describe custom solution design and deployment, suggesting implementation cycles are likely longer than instant API adoption. SU006, SU007, SU008, SU009
CU024 For regulated buyers like pharma, Liquid's private-infrastructure deployment model should raise switching costs after integration because data and workflows stay inside the customer environment. SU015, SU004, SU011
CU025 The Mercedes agreement is the clearest public durability signal because it is explicitly multi-year even though production is still pending. SU016
CU026 Liquid's $10 million revenue threshold can attract small developer projects first while forcing larger commercial users into negotiated licensing. SU013
CU027 Liquid uses free community surfaces, hackathons, and prizes to drive discovery and experimentation before enterprise monetization conversations. SU005
CU028 Liquid publishes no NRR, GRR, churn, active-seat, or renewal metrics for any customer cohort on the public materials reviewed. SU001, SU004, SU003, SU010
CU029 Liquid also does not publish named renewals, cohort curves, satisfaction scores, or public testimonials that would show repeat usage durability over time. SU003, SU002, SU005
CU030 As of the run date, Liquid's public named customer proof is concentrated in two counterparties while most other evidence remains anonymous or developer-oriented. SU016, SU015, SU003, SU012
CU031 Liquid's current proof mix combines one available-now specialist deployment with one future-dated flagship automotive deployment rather than a broad set of current production references. SU015, SU016
CU032 No public reseller or marketplace channel appears central to Liquid's motion; the company instead routes enterprise demand into direct sales and partner-specific work. SU004, SU006, SU002
CU033 Mercedes and Liquid say they may explore other areas of product development together, creating a credible but partner-controlled expansion path after initial launch. SU016
CU034 Liquid's Hugging Face activity demonstrates broad public interest and evaluation behavior, but likes and downloads are not equivalent to contracted revenue accounts. SU012, SU005
CU035 Liquid's anonymized case-study blurbs do not consistently label whether the underlying deployments are pilot, private beta, or scaled production. SU003
CU036 Because public credibility is tied to a small number of flagship deployments and hardware partners, any delay or reprioritization by those parties could materially weaken Liquid's customer narrative. SU016, SU015, SU014
CU037 Liquid targets more sectors publicly than it can currently prove with named references, creating a gap between addressable verticals and visible production proof. SU010, SU001, SU003
CU038 Liquid's community and models pages make browser and playground evaluation easy, which should help top-of-funnel adoption but offers no direct retention or monetization disclosure. SU005, SU011
CU039 Ashby postings for account-executive and developer-relations roles show Liquid is still investing in direct-sales and developer-enablement capacity, which supports funnel expansion but not disclosed customer scale. SU030
CU040 Liquid's Shopify release describes a multi-year partnership to bring sub-20ms foundation models into core commerce experiences, adding a named commerce counterparty to the public proof set. SU031
CU041 Liquid's G42 announcement says the partnership aims to deliver private, local, and efficient AI solutions to enterprises at scale, but public production metrics and customer counts remain undisclosed. SU032
CU042 The Alef Education announcement shows Liquid pursuing a named education customer globally, widening the visible vertical footprint beyond automotive and pharma. SU033
CU043 The Brilliant Labs partnership adds a named consumer-electronics design partner for vision-language deployment, broadening the public customer surface while remaining early-stage proof. SU034
CU044 Liquid's official Mercedes release reinforces that the flagship automotive relationship remains partner-controlled and tied to embedded in-car deployment milestones. SU035
CU045 Liquid's October 2024 first-products event announcement showed the company courting consumer electronics, telecom, finance, ecommerce, and biotech partners before it could publicly enumerate many named accounts. SU036
CR001 Liquid AI was founded in 2023 as an MIT spinout focused on liquid neural network-based foundation models. SR011, SR017, SR019
CR002 Public sources identify Daniela Rus, Ramin Hasani, Mathias Lechner, and Alexander Amini as the core founding team. SR011, SR017
CR003 TechCrunch reported Liquid's 2023 seed financing at $37.5 million and a $303 million post-money valuation. SR011
CR004 Liquid announced a $250 million Series A to scale LFMs, compute infrastructure, and edge/on-prem product readiness. SR002, SR012, SR018
CR005 Tracxn reports Liquid has raised about $297 million across two funding rounds. SR017, SR018
CR006 TechCrunch and Tech Funding News both reported the 2024 round valued Liquid AI at over $2 billion. SR012, SR003
CR007 Tracxn lists a $2 billion post-money valuation for the Dec. 13, 2024 Series A round. SR018
CR008 PitchBook's March 2025 snapshot described Liquid as generating revenue and showed 49 employees. SR019
CR009 Tracxn lists 121 employees as of April 2026, creating a materially different public scale snapshot from PitchBook's 2025 view. SR017
CR010 Tracxn identifies six current board members, but that governance detail is not directly disclosed in Liquid's own public materials reviewed here. SR017
CR011 Liquid's culture page emphasizes critical-path work, low process tolerance, autonomy, and comfort with ambiguity. SR001
CR012 Current Ashby postings show hiring across applied ML, distributed training, edge inference, developer relations, solutions architecture, finance, marketing, and sales. SR020
CR013 Liquid's official pricing page says the company does not currently offer a hosted API of its own. SR005
CR014 Liquid says users can access models through a free playground with rate limits, OpenRouter, direct downloads on Hugging Face, and LEAP customization/deployment. SR005, SR003
CR015 The official models page says enterprises can license or purchase full local access to LFMs and an on-prem customization stack. SR003
CR016 Liquid markets LFMs as deployable on CPU, NPU, and GPU hardware across on-device, cloud, and hybrid environments. SR003, SR008
CR017 Liquid's public model lineup spans text, vision-language, audio, and nano models. SR003
CR018 Liquid says LFM2.5 extends LFM2 with 28T tokens of pretraining and a scaled reinforcement-learning pipeline. SR003
CR019 Liquid's research page shows active publication cadence into 2026, including the LFM2 technical report and multiple 2026 papers. SR006
CR020 Liquid's research-lineage page claims the team has worked across liquid neural networks, state-space models, Hyena-family architectures, and beyond-transformer scaling. SR007
CR021 VentureBeat reported Liquid's first LFMs were released in preview form and in 2024 were not open source. SR013
CR022 Constellation wrote that Liquid's LFMs were a work in progress and weak on zero-shot code, precise numerical calculations, time-sensitive information, and human preference optimization. SR014
CR023 Constellation framed Liquid's efficiency advantage as promising but still early rather than conclusively proven at production scale. SR014
CR024 Liquid's AMD press release says the LEAP SDK delivers zero dependency on cloud APIs for AMD-supported deployments. SR009
CR025 The same AMD release claims sub-100 millisecond responsiveness for on-device AI on supported Ryzen hardware. SR009
CR026 Business Wire says the Mercedes partnership is multi-year and targets initial production deployment in the second half of 2026 in North America. SR015
CR027 Mercedes describes Liquid's role as complementing cloud LLM ecosystems by moving essential speech, language-understanding, and reasoning elements on board. SR015
CR028 Liquid and PR Newswire say the Insilico partnership produced a 2.6B-parameter model trained on roughly 120 billion pharmaceutical tokens across more than 200 tasks on private infrastructure. SR010, SR016
CR029 The Insilico proof point relies partly on internal or domain-specific benchmarks, which is useful commercialization evidence but not the same as broad horizontal enterprise adoption. SR010, SR016
CR030 Liquid's public GTM materials span consumer electronics, telecom, financial services, e-commerce, biotechnology, automotive, startups, and general enterprise use. SR002, SR008, SR010
CR031 The LFM Open License grants broad rights but free commercial use ends once a company exceeds $10 million in annual revenue. SR004
CR032 Liquid's license conditions its copyright and patent grants on that commercial-use limitation, making it materially more restrictive than Apache 2.0. SR004
CR033 The license also auto-terminates on non-compliance and disclaims warranties and liability, shifting commercial and legal risk to users. SR004
CR034 The EU AI Act establishes harmonized obligations for AI systems in the Union and explicitly addresses risks to safety, rights, and economic harm. SR024
CR035 NIST's AI RMF now includes a generative AI profile and an April 2026 concept note for critical infrastructure, signaling higher governance expectations for enterprise deployments. SR025
CR036 Deloitte's 2026 survey says worker access to AI rose by 50% in 2025 and that firms with at least 40% of projects in production are set to double within six months. SR022
CR037 Deloitte also reports only one in five companies has a mature governance model for autonomous AI agents. SR022
CR038 MAPEGY says edge AI demand is shifting inference away from centralized cloud-only architectures toward devices and distributed systems. SR023
CR039 MAPEGY estimates edge AI total addressable market at roughly $170 billion to $260 billion by the early 2030s with 21% to 30% CAGR through 2032. SR023
CR040 MAPEGY says regulatory pressure around privacy, secure AI systems, and the August 2026 EU AI Act is one driver of edge deployment. SR023, SR024
CR041 Research and Markets describes edge AI as spanning cloud and on-prem deployments across automotive, healthcare, consumer electronics, and smart-city use cases. SR026
CR042 Artificial Analysis' pricing and intelligence comparison page lists intense rivalry among Anthropic, Google, Mistral, NVIDIA, OpenAI, xAI, and others. SR021
CR043 Liquid's public commercialization proof remains concentrated in a small set of named partners rather than a broad disclosed customer base. SR009, SR010, SR015, SR016
CR044 The combination of a still-young organization, wide vertical ambition, partner concentration, and non-standard licensing creates a real execution stack rather than a single isolated risk. SR001, SR012, SR017, SR020, SR023
CR045 The fastest thesis-break events would be production slippage at Mercedes, failure to turn restricted free-license users into paid enterprise accounts, or evidence that benchmark wins do not translate into stable deployment quality. SR004, SR014, SR015
CR046 Liquid's docs say LFMs are available in multiple formats including GGUF, MLX, and ONNX for local and production deployment workflows. SR030
CR047 The automotive page frames the market need around hardware-limited in-vehicle assistants, privacy-sensitive local inference, and hybrid edge-cloud architectures. SR028
CR048 The financial-services page says key AI workflows remain latency-, privacy-, and compliance-heavy and are still trapped in pilots or legacy systems. SR029
CR049 Liquid's enterprise solutions page says it does not offer self-service enterprise packages and instead relies on custom pricing, downloads, and LEAP-based customization. SR027
CR050 The same solutions page claims LEAP covers model selection, customization, evaluation, and instant on-device testing in one workflow. SR027
CR051 Liquid Apollo is described as a low-latency, secure, cloud-free playground for local AI interaction. SR032
CR052 Liquid's October 2024 launch event publicly associated the company with speakers from Deloitte, Samsung Next, AMD, Shopify, Microsoft, and biotech partners, signaling broad ecosystem ambition early in commercialization. SR033
CR053 Liquid's case-studies page effectively acknowledges that current public customer examples are still selective by inviting domains not yet shown to contact the enterprise team directly. SR031
CR054 EurekAlert's copy of the Insilico announcement repeats the on-premise drug-discovery thesis and adds reference to more than 1,000 pharmaceutical benchmarks inside MMAI Gym for Science. SR034
CR055 A 2026 third-party feature on Liquid notes that porting non-transformer architectures into software and hardware stacks optimized for transformers requires substantial engineering effort. SR036
CR056 The same feature says future scaling will need to prove that liquid-network stability benefits hold at much larger parameter counts and operational complexity. SR036
CR057 The original MIT CSAIL coverage positioned liquid networks around robustness to noisy data, interpretability, and lower compute cost, linking the commercial story to an older research agenda rather than a purely recent marketing pivot. SR037
CR058 NIST's AI RMF Playbook frames deployment risk around concrete Govern, Map, Measure, and Manage practices, raising the governance bar for enterprise AI rollouts beyond benchmark quality alone. SR039
CR059 MIT CSAIL's 2026 coverage says liquid networks adapt their underlying equations to new data inputs, highlighting how much of Liquid's commercial story still depends on translating an ambitious research agenda into dependable production systems. SR040
CR060 Ramin Hasani's public site foregrounds publications and awards, underscoring how much of Liquid's visible technical identity remains concentrated around the founder-scientist brand. SR041
CR061 Analytics Insight says LEAP's AMD laptop path creates a unified software-and-hardware route for privacy-preserving real-time AI on PCs, reinforcing dependency on specific partner optimization layers. SR042
CV001 Liquid's official funding blog says the company raised $250 million to scale LFMs, compute infrastructure, and edge/on-prem product readiness. SV001
CV002 TechCrunch reported Liquid's 2024 Series A valued the company at over $2 billion. SV002, SV003
CV003 Tech Funding News also reported the round at $250 million with AMD as lead investor and valuation above $2 billion. SV003
CV004 Tracxn reports Liquid has raised about $297 million over two rounds and marks the 2024 Series A at a $2 billion post-money valuation. SV004, SV005
CV005 PitchBook's 2025 profile describes Liquid as generating revenue while showing the latest deal type as Series A and the latest deal amount as $250 million. SV006
CV006 Liquid's official pricing page says the company does not currently offer a hosted API of its own. SV007
CV007 Liquid instead routes public access through a rate-limited playground, OpenRouter, model downloads, and LEAP customization/deployment. SV007, SV008
CV008 The models page says enterprises can license full local access to LFMs and buy an on-prem customization stack. SV008
CV009 Liquid markets LFMs as deployable across CPU, NPU, and GPU hardware in on-device, cloud, and hybrid environments. SV008
CV010 Artificial Analysis tracks model intelligence and pricing across Anthropic, Google, Mistral, NVIDIA, OpenAI, xAI, and other leading labs. SV009
CV011 OpenAI's API pricing page lists GPT-5.5 at $5 input and $30 output per million tokens. SV010
CV012 OpenAI also lists GPT-5.4 mini at $0.75 input and $4.50 output per million tokens, showing a wide quality-price ladder inside one vendor. SV010
CV013 Google's Gemini pricing page offers free, paid, and enterprise tiers, indicating that major rivals can subsidize developer acquisition before charging for scale or security. SV011
CV014 Writer's pricing and homepage emphasize enterprise seats, zero data retention by default, governance, and measurable workflow outcomes rather than pure API access. SV012, SV013
CV015 Writer announced a $200 million Series C at a $1.9 billion valuation in November 2024. SV014, SV015
CV016 Writer said hundreds of large companies use its platform and cited an average 9x ROI in the Series C announcement. SV014
CV017 AI21's Jamba page positions the company around low-latency enterprise workflows, self-hosting, long context, and cost-efficient deployment. SV016
CV018 AI21 Labs announced a $155 million Series C in 2023 that brought total capital to $283 million at a $1.4 billion valuation. SV017
CV019 Cohere markets private, secure, customizable deployment in VPC, on-prem, or dedicated model-vault environments. SV018
CV020 Tech Funding News reported Cohere added $100 million and reached a $7 billion valuation after a prior $500 million round. SV019
CV021 Mistral's documentation presents a broad platform spanning models, APIs, agents, RAG, workflows, and enterprise workspace controls. SV020
CV022 Mistral announced a September 2025 Series C of 1.7 billion euros at an 11.7 billion euro post-money valuation. SV021
CV023 xAI's docs position Grok as a fast, high-end model family with search tooling and rapid model-alias updates. SV022
CV024 xAI announced a $20 billion Series E and said the financing would accelerate world-leading compute infrastructure buildout. SV023
CV025 Deloitte says worker access to AI rose by 50% in 2025 and the share of companies with at least 40% of projects in production is set to double within six months. SV024
CV026 Deloitte also says only one in five companies has a mature governance model for autonomous AI agents. SV024
CV027 MAPEGY estimates edge AI total addressable market at roughly $170 billion to $260 billion by the early 2030s with 21% to 30% CAGR through 2032. SV025
CV028 MAPEGY says inference is increasingly moving to edge devices while training remains concentrated in the cloud. SV025
CV029 Research and Markets shows edge AI deployment modes split across cloud and on-prem and highlights automotive, healthcare, and consumer electronics as major end sectors. SV026
CV030 Amadeus argues that price competition is pushing LLMs toward commoditization and shifting profit pools toward data, tooling, safety, and specialized edge silicon. SV027
CV031 The Amadeus analysis says defensibility will depend less on raw model scale and more on proprietary context, orchestration, trust, and efficient inference. SV027
CV032 The arXiv economics paper models foundation-model openness as a strategic competition variable and warns that some policy interventions can reduce welfare or investment. SV028
CV033 Liquid's most visible public commercialization proof is still partner-led: AMD for developer deployment, Mercedes for automotive production, and previously announced vertical custom work. SV029, SV030, SV001
CV034 Mercedes's April 2026 release sets an explicit target of initial production deployment in the second half of 2026, meaning a key proof point is still forward-looking rather than delivered. SV030
CV035 The public Liquid source pack does not disclose revenue, gross margin, net retention, customer concentration, or cap-table preference terms. SV001, SV004, SV006, SV007, SV008
CV036 Because those operating metrics are missing, a clean revenue-multiple valuation method is not supportable from public evidence alone. SV004, SV006, SV027
CV037 Writer's $1.9 billion valuation and AI21's $1.4 billion valuation suggest Liquid's roughly $2 billion mark is already in the upper tier of disclosed enterprise-AI peers with more public commercialization proof. SV002, SV004, SV014, SV017
CV038 Cohere, Mistral, and xAI show that far higher private valuations are possible in this category, but those companies also pair larger capital bases with clearer platform breadth, compute scale, or market visibility. SV019, SV021, SV023
CV039 OpenAI and Google pricing, together with Artificial Analysis and Amadeus, imply that standalone foundation-model economics face ongoing pricing pressure and value-chain compression. SV009, SV010, SV011, SV027
CV040 Liquid's differentiated architecture, edge/on-prem positioning, and partner proofs support a real investment thesis around efficient sovereign AI. SV001, SV008, SV025, SV029, SV030
CV041 The anti-thesis is that Liquid still has less public evidence on commercial scale and unit economics than the price already seems to require. SV004, SV006, SV014, SV017, SV027
CV042 The bull case depends on Mercedes reaching production, LEAP converting edge performance into repeatable enterprise deployments, and additional customer proofs emerging quickly. SV025, SV029, SV030
CV043 The base case is that Liquid remains strategically interesting but disclosure-light, leaving the latest valuation roughly fair-to-stretched rather than obviously attractive. SV004, SV006, SV027
CV044 The bear case is that commoditization, slower partner conversion, or regulatory friction push Liquid toward lower valuation anchors nearer Writer and AI21 than frontier-scale labs. SV017, SV027, SV028, SV030
CV045 A recommendation of research-more is more defensible than buy because public evidence does not yet support high-conviction upside from the current mark. SV004, SV006, SV027
CV046 Confidence should be medium because the financing and comparable anchors are visible, but the key unit-economics and deployment-conversion variables remain private. SV004, SV006, SV024
CV047 Risk rating should be high because pricing pressure, governance demands, partner concentration, and commercialization opacity can all transmit directly into value. SV024, SV027, SV030
CV048 Valuation stance is best described as fair-to-stretched rather than unsupported or obviously cheap. SV004, SV014, SV017, SV027
CV049 Liquid's own funding blog says the company wants to integrate products into mission-critical workflows across telecom, finance, e-commerce, biotech, and consumer electronics. SV001
CV050 Writer's funding release pairs valuation with named Fortune 500 customers and ROI claims, a higher level of public commercialization proof than Liquid currently exposes. SV014, SV015
CV051 Mistral's funding release emphasizes custom decentralized frontier AI solutions and high-performance compute infrastructure for strategic industries, illustrating a higher-scale enterprise model story. SV021
CV052 xAI's financing release ties valuation support to compute scale and broad product reach, reinforcing how much distribution and infrastructure matter in private AI valuations. SV023
CV053 Public AI vendors such as C3.ai expose periodic SEC filing trails that investors can use to inspect revenue and risk factors, a disclosure standard that private Liquid does not currently match. SV031
来源
编号出版方标题引文
SO001 Liquid AI Liquid AI: Build efficient general-purpose AI at every scale.
SO002 Liquid AI About Liquid AI
SO003 Liquid AI Liquid AI: A New Generation of AI Models from First Principles Founded by a quartet of MIT Computer Science and Artificial Intelligence Laboratory (CSAIL) AI and machine learning scientists – Ramin Hasani, Mathias Lechner, Alexander Amini, and Daniela Rus.
SO004 Liquid AI We raised $250M to scale capable and efficient general-purpose AI This funding will help us accelerate the development, scaling, and deployment of Liquid Foundation Models (LFMs).
SO005 Liquid AI Liquid Foundation Models | Liquid AI
SO006 Liquid AI Pricing | Liquid AI We do not currently offer a hosted API of our own.
SO007 Liquid AI Newsroom | Liquid AI
SO008 Liquid AI Enterprise Solutions | Liquid AI
SO009 Liquid AI From Liquid Neural Networks to Liquid Foundation Models | Liquid AI
SO010 Liquid AI LFM License | Liquid AI Commercial Use Threshold. Rights to use the model for commercial purposes end if your company annual revenue exceeds $10 million USD.
SO011 TechCrunch Liquid AI, a new MIT spinoff, wants to build an entirely new type of AI
SO012 TechCrunch Liquid AI just raised $250M to develop a more efficient type of AI model
SO013 Tech Funding News Liquid AI closes $250M, hits $2B valuation with AMD-led funding
SO014 VentureBeat MIT spinoff Liquid debuts non-transformer AI models and they are already state-of-the-art
SO015 Constellation Research Liquid AI launches non-transformer genAI models: Can it ease power crunch? LFMs are not good at zero-shot code tasks, precise numerical calculations, time-sensitive information and human preference optimization.
SO016 Tracxn Liquid AI
SO017 Tracxn Liquid AI funding and investors
SO018 PitchBook via Internet Archive Liquid AI 2025 Company Profile: Valuation, Funding & Investors | PitchBook
SO019 CB Insights Liquid AI - Products, Competitors, Financials, Employees, Headquarters Locations
SO020 Mathias Lechner Mathias Lechner - Personal Page
SO021 arXiv Liquid Time-constant Networks
SO022 Hugging Face LiquidAI (Liquid AI)
SO023 Liquid AI Docs Liquid Foundation Models - Liquid Docs
SO024 Built In Liquid AI Careers, Perks + Culture
SO025 Ashby Liquid AI Jobs
SO026 Liquid AI Liquid AI and Insilico Medicine Announce Strategic Partnership Delivering Lightweight Scientific Foundation Models for Drug Discovery | Liquid AI
SO027 Business Wire Mercedes-Benz and Liquid AI Partner to Scale Embedded In-Car Intelligence in North America
SO028 Liquid AI Liquid’s Edge AI Platform, LEAP, expands support to laptops with best-in-class performance on AMD Ryzen and Ryzen AI Processors | Liquid AI
SM001 Liquid AI Enterprise Solutions | Liquid AI
SM002 Liquid AI Enterprise Solutions | Liquid AI
SM003 Liquid AI Automotive | Liquid AI
SM004 Liquid AI Ecommerce | Liquid AI
SM005 Liquid AI Financial Services | Liquid AI
SM006 Liquid AI Start Up Solutions | Liquid AI
SM007 Liquid AI Developer Community | Liquid AI
SM008 Liquid AI LFM License | Liquid AI
SM009 Liquid AI Pricing | Liquid AI
SM010 Deloitte The State of AI in the Enterprise - 2026 AI report
SM011 Wevolver Introduction | The 2026 Edge AI Technology Report
SM012 Fortune Business Insights Edge AI Market Size, Share, Growth & Global Report [2034]
SM013 Verified Market Reports Global Edge AI Market Size, Share, Industry Growth & Forecast 2026-2034
SM014 Stratistics MRC Edge AI Inference Market CAGR, size, share, trends, growth, value, key players analysis | Stratistics MRC report
SM015 Artificial Analysis Comparison of AI Models across Intelligence, Performance, and Price
SM016 OpenAI OpenAI API Pricing
SM017 Google Gemini Developer API pricing
SM018 Anthropic Plans & Pricing | Claude by Anthropic
SM019 Writer WRITER plans
SM020 xAI Models | xAI Docs
SM021 Microsoft Azure Phi Open Models - Small Language Models | Microsoft Azure
SM022 Hugging Face / Meta meta-llama/Meta-Llama-3-70B · Hugging Face
SM023 Cohere Enterprise AI: Private, Secure, Customizable | Cohere
SM024 Mistral AI Mistral Studio - your AI production platform
SM025 Qualcomm Qualcomm AI Hub
SM026 Google DeepMind Gemini 3.5
SM027 MarkTechPost Liquid AI Introduces Liquid Foundation Models (LFMs): A 1B, 3B, and 40B Series of Generative AI Models
SP001 Liquid AI Liquid AI: Build efficient general-purpose AI at every scale.
SP002 Liquid AI Liquid Foundation Models | Liquid AI
SP003 Liquid AI Pricing | Liquid AI
SP004 Liquid AI Enterprise Solutions | Liquid AI
SP005 Liquid AI Docs Liquid Foundation Models - Liquid Docs
SP006 Liquid AI LFM License | Liquid AI
SP007 OpenAI Hello GPT-4o
SP008 OpenAI OpenAI API Pricing
SP009 Anthropic Plans & Pricing | Claude by Anthropic
SP010 Google DeepMind Gemini 3.5
SP011 Google Gemini Developer API pricing
SP012 Meta Introducing Meta Llama 3: The most capable openly available LLM to date
SP013 Hugging Face meta-llama/Meta-Llama-3-70B · Hugging Face
SP014 Mistral AI Mistral AI Documentation
SP015 Mistral AI Mistral Studio - your AI production platform
SP016 Cohere Enterprise AI: Private, Secure, Customizable | Cohere
SP017 Cohere Introducing Command A: Max performance, minimal compute | Cohere Blog
SP018 AI21 Jamba | AI21
SP019 Writer WRITER
SP020 Writer WRITER plans
SP021 xAI Grok — Truth-seeking AI Chatbot with Voice & Image Generation | xAI
SP022 xAI Models | xAI Docs
SP023 Microsoft Phi Open Models - Small Language Models | Microsoft Azure
SP024 Microsoft Microsoft Foundry | Microsoft Azure
SP025 Qualcomm Qualcomm AI Hub
SP026 Artificial Analysis Comparison of AI Models across Intelligence, Performance, and Price
SP027 Liquid AI Liquid’s Edge AI Platform, LEAP, expands support to laptops with best-in-class performance on AMD Ryzen™ and Ryzen AI™ Processors | Liquid AI
SP028 Business Wire Mercedes-Benz and Liquid AI Partner to Scale Embedded In-Car Intelligence in North America
SP029 TechCrunch Liquid AI, a new MIT spinoff, wants to build an entirely new type of AI
SI001 Liquid AI Pricing | Liquid AI
SI002 Liquid AI Enterprise Solutions | Liquid AI
SI003 Liquid AI LFM License | Liquid AI
SI004 Liquid AI We raised $250M to scale capable and efficient general-purpose AI | Liquid AI
SI005 Liquid AI Docs Liquid Foundation Models - Liquid Docs
SI006 Liquid AI Liquid Foundation Models | Liquid AI
SI007 TechCrunch Liquid AI just raised $250M to develop a more efficient type of AI model
SI008 Tech Funding News Liquid AI closes $250M, hits $2B valuation with AMD-led funding
SI009 PitchBook via Internet Archive Liquid AI 2025 Company Profile: Valuation, Funding & Investors | PitchBook
SI010 Tracxn Liquid AI
SI011 Liquid AI Liquid’s Edge AI Platform, LEAP, expands support to laptops with best-in-class performance on AMD Ryzen™ and Ryzen AI™ Processors | Liquid AI
SI012 Business Wire Mercedes-Benz and Liquid AI Partner to Scale Embedded In-Car Intelligence in North America
SI013 PR Newswire Insilico Medicine and Liquid AI Announce Strategic Partnership Delivering Lightweight Scientific Foundation Models for Drug Discovery
SI014 TechCrunch Liquid AI, a new MIT spinoff, wants to build an entirely new type of AI
SI015 OpenAI OpenAI API Pricing
SI016 Anthropic Plans & Pricing | Claude by Anthropic
SI017 Google Gemini Developer API pricing
SI018 Writer WRITER plans
SI019 Artificial Analysis Comparison of AI Models across Intelligence, Performance, and Price
SI020 PricePerToken LLM API Pricing 2026 - Compare 300+ AI Model Costs
SI021 U.S. Securities and Exchange Commission Microsoft Corporation Form 10-K for fiscal year ended June 30, 2025
SI022 Liquid AI Liquid AI: Build efficient general-purpose AI at every scale.
SI023 xAI Models | xAI Docs
SI024 Writer WRITER
SI025 Google DeepMind Gemini 3.5
SI026 Liquid AI About Liquid AI
SI027 Liquid AI Enterprise Solutions | Liquid AI
SI028 Liquid AI Automotive | Liquid AI
SI029 Liquid AI Ecommerce | Liquid AI
SI030 Liquid AI Financial Services | Liquid AI
SI031 Liquid AI Case Studies | Liquid AI
SI032 Liquid AI Start Up Solutions | Liquid AI
SI033 Mercedes-Benz USA Mercedes-Benz and Liquid AI Partner to Scale Embedded In-Car Intelligence in North America
SI034 EurekAlert Liquid AI and Insilico Medicine announce strategic partnership delivering lightweight scientific foundation models for drug discovery
SI035 Deloitte The State of AI in the Enterprise - 2026 AI report
SI036 Fortune Business Insights Edge AI Market Size, Share, Growth & Global Report [2034]
SI037 Verified Market Reports Global Edge AI Market Size, Share, Industry Growth & Forecast 2026-2034
SI038 Stratistics MRC Edge AI Inference Market CAGR, size, share, trends, growth, value, key players analysis
SI039 TechIntelPro Liquid AI’s LEAP Boosts Edge AI on AMD Ryzen Processors
SI040 Wevolver Introduction | The 2026 Edge AI Technology Report
SI041 MarkTechPost Liquid AI Introduces Liquid Foundation Models (LFMs): A 1B, 3B, and 40B Series of Generative AI Models
SI042 Impact Lab Liquid AI Unveils Groundbreaking Foundation Models, Challenging Transformer-Based AI
SI043 FinancialContent The Fluidity of Intelligence: How Liquid AI’s New Architecture is Ending the Transformer Monopoly
SI044 Liquid AI Liquid Edge AI Platform
SE001 Liquid AI Liquid AI: Build efficient general-purpose AI at every scale.
SE002 Liquid AI About Liquid AI
SE003 Liquid AI Liquid AI: A New Generation of AI Models from First Principles
SE004 Liquid AI We raised $250M to scale capable and efficient general-purpose AI
SE005 Liquid AI Liquid Foundation Models | Liquid AI
SE006 Liquid AI Liquid Foundation Models: Our First Series of Generative AI Models | Liquid AI
SE007 Liquid AI From Liquid Neural Networks to Liquid Foundation Models | Liquid AI
SE008 Liquid AI Research | Liquid AI
SE009 Liquid Docs Liquid Foundation Models - Liquid Docs
SE010 arXiv Liquid Time-constant Networks
SE011 Liquid AI Liquid AI Launches LEAP and Liquid Apollo: A New Era for Edge AI Deployment Begins | Liquid AI
SE012 Liquid AI Liquid AI Releases World’s Fastest and Best-Performing Open-Source Small Foundation Models | Liquid AI
SE013 Liquid AI Liquid unveils “Nanos”: Extremely small foundation models that match frontier-model quality—running directly on everyday devices | Liquid AI
SE014 Liquid AI Liquid AI to Unveil First Products Built on Liquid Foundation Models (LFMs) at Exclusive MIT Event | Liquid AI
SE015 Hugging Face LiquidAI (Liquid AI)
SE016 Liquid AI Pricing | Liquid AI
SE017 Liquid AI Developer Community | Liquid AI
SE018 Liquid AI Enterprise Solutions | Liquid AI
SE019 Liquid AI Automotive | Liquid AI
SE020 Liquid AI Ecommerce | Liquid AI
SE021 Liquid AI Financial Services | Liquid AI
SE022 Liquid AI Liquid Apollo
SE023 Liquid AI Liquid’s Edge AI Platform, LEAP, expands support to laptops with best-in-class performance on AMD Ryzen™ and Ryzen AI™ Processors | Liquid AI
SE024 Liquid AI Liquid AI and Insilico Medicine Announce Strategic Partnership Delivering Lightweight Scientific Foundation Models for Drug Discovery | Liquid AI
SE025 Liquid AI Case Studies | Liquid AI
SE026 Liquid AI Product Launch Livestream | October 23rd 2024
SE027 Liquid AI LFM License | Liquid AI
SE028 TechCrunch Liquid AI, a new MIT spinoff, wants to build an entirely new type of AI | TechCrunch
SE029 VentureBeat MIT spinoff Liquid debuts non-transformer AI models and they're already state-of-the-art
SE030 Constellation Research Liquid AI launches non-transformer genAI models: Can it ease the power crunch?
SE031 MarkTechPost Liquid AI Introduces Liquid Foundation Models (LFMs): A 1B, 3B, and 40B Series of Generative AI Models
SE032 Impact Lab Liquid AI Unveils Groundbreaking Foundation Models, Challenging Transformer-Based AI
SE033 Business Wire Mercedes-Benz and Liquid AI Partner to Scale Embedded In-Car Intelligence in North America
SE034 PR Newswire Insilico Medicine and Liquid AI Announce Strategic Partnership Delivering Lightweight Scientific Foundation Models for Drug Discovery
SE035 EurekAlert! Liquid AI and Insilico Medicine announce strategic partnership delivering lightweight scientific foundation models for drug discovery
SE036 TechIntelPro Liquid AI’s LEAP Boosts Edge AI on AMD Ryzen Processors
SE037 Liquid AI Introducing LFM2.5: The Next Generation of On-Device AI | Liquid AI
SE038 Hugging Face LiquidAI/LFM2-1.2B · Hugging Face
SE039 Enterprise AI World Liquid AI's Open Source, Small Foundation LFM2 Models Outperform and Outclass Competitors
SE040 GitHub GitHub - Liquid4All/docs: Liquid documentation
SU001 Liquid AI Liquid AI: Build efficient general-purpose AI at every scale.
SU002 Liquid AI About Liquid AI
SU003 Liquid AI Case Studies | Liquid AI
SU004 Liquid AI Pricing | Liquid AI
SU005 Liquid AI Developer Community | Liquid AI
SU006 Liquid AI Enterprise Solutions | Liquid AI
SU007 Liquid AI Automotive | Liquid AI
SU008 Liquid AI Ecommerce | Liquid AI
SU009 Liquid AI Financial Services | Liquid AI
SU010 Liquid AI We raised $250M to scale capable and efficient general-purpose AI
SU011 Liquid AI Liquid Foundation Models | Liquid AI
SU012 Hugging Face LiquidAI (Liquid AI)
SU013 Liquid AI LFM License | Liquid AI
SU014 Liquid AI Liquid’s Edge AI Platform, LEAP, expands support to laptops with best-in-class performance on AMD Ryzen™ and Ryzen AI™ Processors | Liquid AI
SU015 Liquid AI Liquid AI and Insilico Medicine Announce Strategic Partnership Delivering Lightweight Scientific Foundation Models for Drug Discovery | Liquid AI
SU016 Business Wire Mercedes-Benz and Liquid AI Partner to Scale Embedded In-Car Intelligence in North America
SU017 PR Newswire Insilico Medicine and Liquid AI Announce Strategic Partnership Delivering Lightweight Scientific Foundation Models for Drug Discovery
SU018 EurekAlert! Liquid AI and Insilico Medicine announce strategic partnership delivering lightweight scientific foundation models for drug discovery
SU019 Liquid AI Product Launch Livestream | October 23rd 2024
SU020 Deloitte The State of AI in the Enterprise - 2026 AI report
SU021 TechIntelPro Liquid AI’s LEAP Boosts Edge AI on AMD Ryzen Processors
SU022 Liquid AI Enterprise Solutions | Liquid AI
SU023 Liquid AI Start Up Solutions | Liquid AI
SU024 Liquid AI Newsroom | Liquid AI
SU025 Liquid AI Newsroom | Liquid AI (Japanese)
SU026 Liquid AI Careers | Liquid AI
SU027 TechCrunch Liquid AI just raised $250M to develop a more efficient type of AI model | TechCrunch
SU028 Built In Liquid AI Careers, Perks + Culture
SU029 Built In Liquid AI Jobs + Careers
SU030 Ashby Liquid AI Jobs
SU031 Liquid AI Liquid AI Announces Multi‑Year Partnership with Shopify to Bring Sub‑20ms Foundation Models to Core Commerce Experiences | Liquid AI
SU032 Liquid AI G42 and Liquid AI Partner to Deliver Private, Local and Efficient AI Solutions to Enterprises at Scale | Liquid AI
SU033 Liquid AI Alef Education Collaborates with Liquid AI to Advance AI in Education Globally | Liquid AI
SU034 Liquid AI Brilliant Labs Partners With Liquid AI to Bring Vision-Language Tech to Your Glasses | Liquid AI
SU035 Liquid AI Liquid AI and Mercedes-Benz partner to scale embedded in-car intelligence | Liquid AI
SU036 Liquid AI Liquid AI to Unveil First Products Built on Liquid Foundation Models (LFMs) at Exclusive MIT Event | Liquid AI
SR001 Liquid AI About Liquid AI
SR002 Liquid AI We raised $250M to scale capable and efficient general-purpose AI
SR003 Liquid AI Liquid Foundation Models | Liquid AI
SR004 Liquid AI LFM License | Liquid AI
SR005 Liquid AI Pricing | Liquid AI
SR006 Liquid AI Research | Liquid AI
SR007 Liquid AI From Liquid Neural Networks to Liquid Foundation Models
SR008 Liquid AI Enterprise Solutions | Liquid AI
SR009 Liquid AI Liquid’s Edge AI Platform, LEAP, expands support to laptops with best-in-class performance on AMD Ryzen™ and Ryzen AI™ Processors
SR010 Liquid AI Liquid AI and Insilico Medicine Announce Strategic Partnership Delivering Lightweight Scientific Foundation Models for Drug Discovery
SR011 TechCrunch Liquid AI, a new MIT spinoff, wants to build an entirely new type of AI
SR012 TechCrunch Liquid AI just raised $250M to develop a more efficient type of AI model
SR013 VentureBeat MIT spinoff Liquid debuts non-transformer AI models and they're already state-of-the-art
SR014 Constellation Research Liquid AI launches non-transformer genAI models: Can it ease power crunch?
SR015 Business Wire Mercedes-Benz and Liquid AI Partner to Scale Embedded In-Car Intelligence in North America
SR016 PR Newswire Insilico Medicine and Liquid AI Announce Strategic Partnership Delivering Lightweight Scientific Foundation Models for Drug Discovery
SR017 Tracxn Liquid AI
SR018 Tracxn Liquid AI - Funding and Investors
SR019 PitchBook Liquid AI 2025 Company Profile: Valuation, Funding & Investors | PitchBook
SR020 Ashby Liquid AI Jobs
SR021 Artificial Analysis Comparison of AI Models across Intelligence, Performance, and Price
SR022 Deloitte The State of AI in the Enterprise - 2026 AI report
SR023 MAPEGY 2026 Edge AI Technology Report: Trends, Signals & Strategic Insights
SR024 EUR-Lex Regulation (EU) 2024/1689 (Artificial Intelligence Act)
SR025 NIST AI Risk Management Framework
SR026 Research and Markets Edge AI Market Report 2026 - Research and Markets
SR027 Liquid AI Enterprise Solutions | Liquid AI
SR028 Liquid AI Automotive | Liquid AI
SR029 Liquid AI Financial Services | Liquid AI
SR030 Liquid AI Docs Liquid Foundation Models - Liquid Docs
SR031 Liquid AI Case Studies | Liquid AI
SR032 Liquid AI Liquid Apollo
SR033 Liquid AI Product Launch Livestream | October 23rd 2024
SR034 EurekAlert! Liquid AI and Insilico Medicine announce strategic partnership delivering lightweight scientific foundation models for drug discovery
SR035 Ramin Hasani Ramin Hasani's Official Website
SR036 FinancialContent / TokenRing AI The Fluidity of Intelligence: How Liquid AI’s New Architecture is Ending the Transformer Monopoly
SR037 MIT CSAIL via reader “Liquid” machine-learning system adapts to changing conditions
SR038 Liquid AI Mercedes-Benz and Liquid AI partner to scale embedded in-car intelligence in North America
SR039 NIST NIST AI RMF Playbook
SR040 MIT CSAIL via reader “Liquid” machine-learning system adapts to changing conditions
SR041 Ramin Hasani Ramin Hasani's Official Website
SR042 Analytics Insight Liquid’s Edge AI Platform, LEAP, Expands Support to Laptops with Best-in-Class Performance on AMD Ryzen™ and Ryzen AI™ Processors
SV001 Liquid AI We raised $250M to scale capable and efficient general-purpose AI
SV002 TechCrunch Liquid AI just raised $250M to develop a more efficient type of AI model
SV003 Tech Funding News Liquid AI closes $250M, hits $2B valuation with AMD-led funding
SV004 Tracxn Liquid AI - Funding and Investors
SV005 Tracxn Liquid AI
SV006 PitchBook Liquid AI 2025 Company Profile: Valuation, Funding & Investors | PitchBook
SV007 Liquid AI Pricing | Liquid AI
SV008 Liquid AI Liquid Foundation Models | Liquid AI
SV009 Artificial Analysis Comparison of AI Models across Intelligence, Performance, and Price
SV010 OpenAI OpenAI API Pricing
SV011 Google Gemini Developer API pricing
SV012 WRITER WRITER plans
SV013 WRITER WRITER
SV014 WRITER WRITER raises $200M Series C at $1.9B valuation to fuel leadership in agentic enterprise AI
SV015 Business Wire Writer Raises $200M Series C at $1.9B Valuation to Fuel Leadership in Agentic Enterprise AI
SV016 AI21 Jamba | AI21
SV017 AI21 AI21 Labs Announces Series C Funding Round at $1.4 Billion Valuation
SV018 Cohere Enterprise AI: Private, Secure, Customizable | Cohere
SV019 Tech Funding News Cohere raises $100M, boosting valuation to $7B and deepening AMD partnership
SV020 Mistral AI Mistral AI Documentation
SV021 Mistral AI Mistral AI raises 1.7B€ to accelerate technological progress with AI
SV022 xAI Models | xAI Docs
SV023 xAI xAI Raises $20B Series E
SV024 Deloitte The State of AI in the Enterprise - 2026 AI report
SV025 MAPEGY 2026 Edge AI Technology Report: Trends, Signals & Strategic Insights
SV026 Research and Markets Edge AI Market Report 2026 - Research and Markets
SV027 Amadeus Capital Partners Where Will LLM Value Flow After Commoditisation?
SV028 arXiv The Economics of AI Foundation Models: Openness, Competition, and Governance
SV029 Liquid AI Liquid’s Edge AI Platform, LEAP, expands support to laptops with best-in-class performance on AMD Ryzen™ and Ryzen AI™ Processors
SV030 Business Wire Mercedes-Benz and Liquid AI Partner to Scale Embedded In-Car Intelligence in North America
SV031 Securities and Exchange Commission EDGAR Entity Landing Page - C3.ai, Inc.