Mistral Unveils Enterprise AI Models to Challenge Cloud Rivals

Mistral发布企业级AI模型挑战云巨头 开放权重的多模态架构允许受监管行业在本地部署定制AI，从而降低成本并避免API绑定。 Pavan Kumar Reddy与Howard Cohen介绍了Mistral涵盖开放权重通用和专用AI模型的完整生态系统。他们详细阐述了Mistral Large 3等旗舰架构，以及OCR4、Leanstral 1.5、Robostral、Shieldstral和Voxtral语音系列等专用模型。Cohen解释了开源AI为企业定制与自主控制带来的战略优势，并展示了在Mistral全栈产品套件支持下，其在金融、工业工程和软件开发领域的实际应用。 Mistral 阐述其端到端全栈愿景，并展示不同参数量级的旗舰开放权重模型，包括 675B 总参数的 Mistral Large 3、企业级密集模型 Mistral Medium 3.5 以及适配边缘设备的 Mistral 3 系列。 (0:08) Mistral 推出针对特定高价值场景的专用模型矩阵，涵盖高准确率文档解析系统 OCR4、形式化数学证明智能体 Leanstral 1.5、物理具身导航模型 Robostral，以及可在单张 16 GB GPU 上运行的 3B 开源安全分类模型 Shieldstral。 (3:17) 团队强调开放权重的战略价值，企业不仅能借此彻底规避云端 API 供应商绑定，还可在保障数据主权的前提下利用私有数据进行定制化微调（RLFT）。 (8:49) Mistral 展示了覆盖全生命周期的全栈 AI 生态，包括 Mistral Work/Code 应用、Forge 微调平台、Studio 开发环境、底层 GPU 算力集群以及前线工程师支持服务。 (16:20)

Mistral AI positions open-weight architectures and an integrated tooling stack as a secure alternative to proprietary models, arguing that organizations require full operational control, customization, and data sovereignty across edge and cloud environments. Tiered Generalist Foundation Models (0:08) According to AI research scientist Pavan Kumar Reddy, Mistral balances compute performance and latency across operational tiers using multimodal open-weight models supporting text, vision, and speech in over 100 languages. At the top of this hierarchy is Mistral Large 3, an open-weight Mixture-of-Experts architecture featuring 675 billion total parameters and 41 billion active parameters designed for heavy-duty research, chatting, and coding agents. For standard enterprise deployments, Mistral Medium 3.5 provides a 128-billion-parameter dense model (activating 6 billion parameters) engineered to operate on a single node with low latency. For local execution and resource-constrained environments, the Mistral 3 Series offers three compact dense models optimized for edge devices. Specialized Systems for Verification and Physical AI (3:17) Beyond generalist LLMs, Mistral has developed targeted models for technical and embodied domains. Leanstral 1.5 functions as an open formal verification agent that translates mathematical hypotheses into Lean proofs, achieving an 87% pass rate on PutnamBench through test-time scaling compute. In robotics, Robostral maps RGB camera feeds and natural language instructions into dynamic navigation paths, reaching a 76.6% success rate across diverse platforms. For document intelligence, OCR4 processes documents across 170+ languages with bounding box segmentations and confidence scores, achieving a reported 72% win rate against competitors. Mistral also introduced Shieldstral, an Apache 2.0-licensed 3-billion-parameter guardrail model capable of evaluating multimodal queries against custom safety policies on a single 16 GB GPU. In speech, the Voxtral suite (Voxtral Chat, Transcribe, and TTS) targets real-time speech-to-speech interactions without rigid turn-taking constraints. Open Weights for Data Sovereignty (8:49) Howard Cohen, Head of Communications for North America, and Reddy argue that open-weight models provide practical commercial protections against API vendor lock-in, service deprecation, and regulatory non-compliance. Deploying models within private environments permits organizations to execute custom fine-tuning (SFT) and reinforcement learning (RLFT) on proprietary data without leaking intellectual property or violating regional data sovereignty rules. In financial services, institutions deploy open models on-premise to keep proprietary alpha and market strategies confidential rather than routing sensitive data through cloud APIs. In industrial engineering, Mistral applies physics-informed AI gained through its acquisition of MEI, while software enterprises deploy domain-tuned open weights to reduce the high per-seat software costs associated with third-party proprietary coding assistants. The End-to-End Enterprise Stack (16:20) Mistral complements its model releases with an infrastructure ecosystem covering the complete machine learning lifecycle. The Forge platform provides environments for supervised fine-tuning and reinforcement learning, Studio offers tools to build and deploy agentic workflows, and direct GPU clusters supply compute for training and inference workloads. For enterprise workflows, Mistral provides pre-built agent environments named Mistral Work and Mistral Code. To facilitate enterprise adoption, the company deploys forward-applied engineering teams that embed directly with client organizations to design and deploy custom automation pipelines. Mistral advocates for open weights as an enterprise safeguard against vendor dependency and data exposure; however, the practical execution of this strategy requires clients to navigate significant infrastructure complexity, often relying on specialized fine-tuning platforms or embedded engineering support to achieve production readiness.

Mistral 研究科学家 Pavan Kumar Reddy 与北美传播主管 Howard Cohen 阐述了公司的核心战略，即通过构建从底层算力到垂直应用的开放权重模型生态，使企业在保有数据主权与技术控制权的前提下，实现前沿人工智能的定制化落地。 分级覆盖的通用多模态模型矩阵 (0:08) Mistral 针对不同计算规模与延迟需求构建了结构清晰的通用多模态模型体系，支持超过 100 种语言的文本、视觉与语音处理。其中，旗舰级混合专家模型 Mistral Large 3 延续了 Mixtral 架构，总参数量达 6750 亿，激活参数量为 410 亿，主要面向复杂学术研究、通用对话及编程智能体开发。 面向企业级主流工作负载，Mistral Medium 3.5 采用稠密架构，总参数量 1280 亿，激活参数量仅 60 亿，旨在以极低延迟实现在单节点上的平滑部署并提供百亿级模型的性能支撑。此外，Mistral 3 系列则包含三款紧凑型轻量稠密模型，专门用于边缘设备与本地终端的离线高效计算。 聚焦高价值任务的专用模型布局 (3:17) 除通用语言模型外，Mistral 针对特定高难度场景推出了专用架构。在文档解析领域，第四代 OCR4 模型在 170 多种语言中对主流竞品保持了 72% 的胜率，支持边界框分割与置信度打分，便于引入人工校验。在数学与形式化验证方向，Leanstral 1.5 结合推理阶段计算扩展技术，能将数学假设转化为 Lean 语言证明，并在 PutnamBench 测试集上达到了 87% 的通过率。 在物理交互与安全治理方面，具备具身智能的 Robostral 可结合自然语言指令与 RGB 图像动态规划避障路径，跨机器人平台任务成功率为 76.6%；最新发布的 Shieldstral 则是基于 Apache 2.0 协议开源的 30 亿参数多模态安全防护模型，仅需单张 16GB 显存显卡即可运行，测试表现优于体积为其数倍的同类工具。此外，音频团队打造的 Voxtral 系列正从单向转录与合成拓展至支持自然交替对话的实时端到端语音系统。 开放权重的战略价值与行业实践 (8:49) Mistral 将开放模型权重视为区别于闭源生态的关键商业壁垒。演讲者指出，开放权重不仅能防止企业陷入商业 API 的厂商锁定或突发弃用风险，还允许机构在严格遵守区域数据主权的前提下，利用私有数据开展监督微调（SFT）与强化学习（RLFT），同时为机制可解释性研究提供了无需巨额计算门槛的科研载体。 这种架构在强监管行业呈现出显著落地优势：金融机构可以在本地私有环境中部署模型以保护量化阿尔法交易策略与敏感数据；工业工程领域通过整合专注于物理模拟与前沿 AI 融合的欧洲团队 MEI，加速了物理驱动智能体的开发；在软件工程场景下，直接针对私有代码库微调的开放模型帮助企业大幅削减了订阅商业编程助手的成本支出。 端到端的全栈企业交付生态 (16:20) 为支撑从底层研发到业务集成的全生命周期，Mistral 构建了闭环式产品矩阵。基础设施层提供训练与推理所需的原始 GPU 集群算力；定制与开发层则通过 Forge 平台提供监督微调与强化学习对齐工具，并在 Studio 环境中支持定制化智能体的编排与部署。 在面向最终业务的交互层，Mistral 推出了用于企业工作流的 Mistral Work 与专注软件开发的 Mistral Code 预置智能体环境。针对复杂集成场景，Mistral 组建了前线部署工程团队（Forward-Deployed Engineering），直接嵌入企业客户现场，联合攻坚量身定制的自动化系统实施方案。 Mistral 主张通过全栈开源模型破除闭源商业 API 对企业自主权的限制；不过，其工业物理智能体及具身智能方案的实际效果与大规模工程稳定性，仍需在严苛的现实复杂生产环境中接受进一步检验。

## Background 该视频由行业频道 Ai4 于 2026 年 8 月 23 日发布在 YouTube 平台上。内容简介指出，Mistral 致力于构建紧凑、高性能的模型并公开开源其主要成果，本次分享由 Mistral 的 AI 科学家 Pavan Kumar Reddy 与北美传播负责人 Howard Cohen 共同带来，重点介绍了团队在音频处理、OCR、机器人导航及推理模型等领域的研究进展，展示其如何赋能物理自主性并推动前沿 AI 的透明化与变革。 ## TL;DR Mistral 在演讲中全面阐述了“将前沿 AI 控制权交到企业手中”的全栈生态愿景，系统展示了涵盖 Mistral Large 3、Medium 3.5 和 Ministral 3 的开源多模态通用模型家族，以及第四代 OCR 模型、形式化数学证明智能体 Leanstral 1.5 和具身导航模型 Robostral。现场还重磅发布了轻量级安全护栏模型 Shieldstral 与 Voxtral 语音体系，确立语音为未来 AI 交互的核心界面。最后，双方对谈深入剖析了开源模型在保障金融数据主权、工业物理工程融合与大幅降本增效方面的核心价值，并推出了涵盖 Forge 微调定制、Studio 智能体编排、算力支持及 FDE 驻场交付的端到端企业落地闭环。 ## Atmosphere & style 现场弥漫着务实、硬核且极具极客底蕴的技术前沿气息。前半程科学家直奔主题，以密集的技术指标、架构演进与前沿评测数据展开紧凑演讲，节奏清晰利落；后半程转入双人对谈后，交流自然松弛却发人深省，从开源价值观的理论坚守切入到企业降本与数据主权的现实痛点。整场分享不仅传递出对开源重塑行业格局的坚定信念，更为关注大模型落地的听众带来了沉稳而兼具前瞻性的实操底气。 ## Key takeaways - [00:08](https://www.youtube.com/watch?v=ISBFhvg6zHA&t=8s) Mistral 的核心愿景是让前沿 AI 普及并让企业真正拥有完全的掌控权，构建从算力基础设施到上层应用的端到端全栈生态。 - [00:08](https://www.youtube.com/watch?v=ISBFhvg6zHA&t=8s) Mistral 布局了分层的开源通用模型矩阵，涵盖 675B MoE 旗舰 Mistral Large 3、128B 单节点主力 Mistral Medium 3.5 以及低延迟边缘端 Ministral 3 系列，均原生支持视觉多模态与百种语言。 - [03:17](https://www.youtube.com/watch?v=ISBFhvg6zHA&t=197s) Mistral OCR 4 创新性地引入边界框溯源与行内置信度机制，支持人机协同（Human-in-the-loop）高效复核，在文档解析对比中取得 72% 的胜率。 - [03:17](https://www.youtube.com/watch?v=ISBFhvg6zHA&t=197s) 形式化验证智能体 Leanstral 1.5 实现了 Lean 4 环境下的自主编译与反思迭代，在 PutnamBench 数学竞赛测试中 Pass 率达 87%；物理 AI 智能体 Robostral 仅凭单张 RGB 图像即可实现 76.6% 的高泛化导航成功率。 - [07:03](https://www.youtube.com/watch?v=ISBFhvg6zHA&t=423s) 全新开源的 3B 护栏模型 Shieldstral 能在单张 16GB 显存 GPU 上运行，多模态内容安全分类性能超越自身 7 倍体积的模型。 - [07:03](https://www.youtube.com/watch?v=ISBFhvg6zHA&t=423s) Voxtral 语音家族涵盖转写、TTS 与 Chat 套件，并正向原生 Speech-to-Speech 架构演进，以消除传统级联语音系统的生硬交互延迟。 - [08:49](https://www.youtube.com/watch?v=ISBFhvg6zHA&t=529s) 语音已成为与 Coding Agent 及协同智能体交互的终极界面，在工业控制、车载驾驶等双手受限场景和跨语言翻译中是不可或缺的准入门槛。 - [10:42](https://www.youtube.com/watch?v=ISBFhvg6zHA&t=642s) 坚持开源模型权重不仅能避免闭源供应商锁定，更能让企业根据任务复杂度精准选型、大幅削减推理成本，并通过 SFT 与 RL 微调深度定制专有模型。 - [12:40](https://www.youtube.com/watch?v=ISBFhvg6zHA&t=760s) 开源部署满足了金融机构保护核心交易策略与 Alpha 逻辑的数据主权诉求，并通过收购 MEI 团队将 AI 前沿技术向物理工程与制造领域深度推进。 - [16:02](https://www.youtube.com/watch?v=ISBFhvg6zHA&t=962s) Mistral 构建了覆盖裸金属 GPU 算力、Forge 微调平台、Studio 智能体编排工作台以及 FDE 前向工程专家团队的全栈交付体系，确保企业端到端落地。 ## Detailed walkthrough ### [00:08](https://www.youtube.com/watch?v=ISBFhvg6zHA&t=8s) Mistral 的愿景与通用模型家族 来自 Mistral 旧金山办公室的 AI 研究科学家兼语音研究团队负责人 Pavan Kumar Reddy 登台开场。他强调，Mistral 作为一家前沿 AI 公司，致力于构建全栈 AI 生态体系——从底层的计算基础设施到上层的 AI 应用，核心目标就是让企业能够真正拥有、定制并掌控自己的 AI 技术。一言以蔽之，Mistral 的愿景就是“将前沿 AI 交到每个人手中，并让他们真正拥有掌控权”（"put frontier AI in everyone's hand that they truly control"）。为了在成本与效率前沿上满足企业多样化的落地需求，Mistral 在不同量级上均布局了前沿通用模型： - **Mistral Large 3**：旗舰级最大的开放权重模型，采用混合专家架构（Mixture of Experts, MoE）。Mistral 早在 2023 年就凭借 Mixtral 开创了开源 MoE 的先河，而该模型拥有高达 6750 亿的总参数量（675B）以及 410 亿的激活参数量（41B active）。 - **Mistral Medium 3.5**：定位为企业级主力模型（Enterprise workhorse），采用 1280 亿参数（128B）的密集（dense）架构，单节点即可轻松部署。 - **Mistral 3 / Ministral 3 系列**：包含三款密集型端侧与紧凑模型，兼顾性能与极低推理延迟。以小巧体量爆发媲美千亿参数模型的实力，仅需约 60 亿激活参数即可实现极高推理速度。 所有通用模型均具备原生多模态视觉推理能力，支持超过 100 种语言，并且全部开源模型权重，彰显了 Mistral 对开源社区的坚定承诺。 ```mermaid flowchart TD accTitle: Mistral Generalist Model Portfolio subgraph S1["Mistral 前沿通用模型矩阵"] direction TB M1["Mistral Large 3<br/>675B 总参数 / 41B 激活<br/>旗舰级 MoE 模型"] M2["Mistral Medium 3.5<br/>128B 密集模型<br/>企业级单节点主力"] M3["Ministral 3 系列<br/>紧凑 / 端侧边缘部署<br/>低延迟与极致性价比"] end S1 --> S2["统一特性：原生多模态视觉 + 100+ 多语言支持 + 开源权重"] ``` --- ### [03:17](https://www.youtube.com/watch?v=ISBFhvg6zHA&t=197s) 面向专业场景的专用模型矩阵 除了通用大模型，Mistral 针对企业客户的垂直需求研发了多款专用模型。首先是第四代光学字符识别模型 **Mistral OCR 4**。该模型不仅能从文档扫描件中高精度提取文本，与同类最强竞品相比取得了 72% 的胜率，还具备创新的版面分割与溯源能力：为每个提取的文本片段提供精准的边界框（Bounding Box）和置信度得分（Confidence Score）。在对准确率要求极高的人机协同（Human-in-the-loop）敏感场景中，人工只需针对低置信度区域复核，即可实现近乎完美的解析效果。该模型支持 170 多种语言，且体积小巧，支持本地私有化部署。 ```mermaid flowchart LR accTitle: Mistral OCR and Specialized Pipeline subgraph OCR_Flow["Mistral OCR 4 工作流"] direction LR A["文档/扫描图像"] --> B["Mistral OCR 4"] B --> C["文本提取"] B --> D["边界框溯源"] B --> E["行内置信度得分"] E --> F{"置信度判断"} F -- 低置信度 --> G["人工介入复核"] F -- 高置信度 --> H["直接入库输出"] end ``` 在高级推理与具身智能领域，Mistral 推出了形式化验证智能体 **Leanstral 1.5** 与物理 AI 智能体 **Robostral**： - **Leanstral 1.5**：不同于单纯输出文字答案的语言模型，它能够接收数学命题并在 Lean 4 形式化编程语言中给出完整证明。它具备真正的 Agent 闭环行动力，能自主编辑代码仓库、执行编译指令并根据编译器报错不断迭代优化证明。在基于普特南数学竞赛真题的 PutnamBench 基准测试中，随着测试期计算量（Test-time Compute/Tokens）的扩展，其 Pass 率可飙升至 87%。 - **Robostral**：Mistral 在具身智能（Physical AI）领域的首次探索。该 Agent 仅凭车载/机器人 RGB 摄像头输入的单张图像，就能理解高层自然语言指令（例如“导航至会议室正前方的最近沙发”），实时生成路径规划、执行动作，并在遇到未见障碍物时动态重构路径，取得了 76.6% 的导航成功率，对不同机器人底盘、相机型号及未知障碍具有极强的泛化性。 --- ### [07:03](https://www.youtube.com/watch?v=ISBFhvg6zHA&t=423s) 新发布 Shieldstral 与 Voxtral 语音家族 在演讲现场，Pavan 隆重宣布正式发布全新护栏模型 **Shieldstral**。这是一款基于 Apache 2.0 协议开源的 3B 轻量级内容安全分类器。用户只需自定义内容安全策略，Shieldstral 即可判断用户输入的 Query 或文档是否违规并给出置信度。尽管参数量仅为 3B，能轻松运行在单张 16GB 显存的消费级 GPU 上，其实测性能却超越了数倍于其体积（7倍参数量）的大模型，并同样原生支持文本与图像多模态输入。 随后 Pavan 介绍了由他亲自领衔打造的 **Voxtral** 语音模型家族。过去一年中 Mistral 已陆续推出： - **Voxtral Chat**：接收音频与文本提示，直接生成文本回答、摘要、转写或触发工具调用（Tool Calls）。 - **Voxtral Transcribe 与 TTS**：作为语音智能体的基础构建模块，提供高精度转写与拟真语音合成。 - **下一代实时语音大模型**：当前团队正全力攻关原生的 Speech-to-Speech 架构，旨在打破传统级联语音管道生硬的轮次交互（Turn-taking）限制，实现真正自然、实时的拟人双向对话。 --- ### [08:49](https://www.youtube.com/watch?v=ISBFhvg6zHA&t=529s) 对话 Howard Cohen：为什么语音是 AI 的终极交互界面 北美传播负责人 Howard Cohen 随后登台与 Pavan 展开对谈。两人探讨了语音在下一代 AI 交互中的核心地位。Cohen 指出，开源 AI 正在行业内迎来属于自己的高光时刻，而语音无疑是 AI 最重要的未来界面。 Pavan 分析称，语音不仅是人际沟通最本能、底层的媒介，也是人机协同最高效的接口。随着 Coding Agent 和各类协同智能体的爆发，用户最习惯以自然的说话方式指派任务。此外，在车载驾驶、工业重型机械操控等严苛的“双手被占用”（Hands-free）场景下，语音指令是唯一的交互选择；在跨语言沟通中，实时语音翻译系统更能作为中介消除人类之间的理解障碍。语音交互已经从附加功能转变为与 AI Agent 协作的“准入门槛”（Table stakes）。 --- ### [10:42](https://www.youtube.com/watch?v=ISBFhvg6zHA&t=642s) 坚守开源根基与企业落地的务实价值 谈及 Mistral 的创立初衷，Cohen 提到公司由 Arthur Mensch、Timothée Lacroix 和 Guillaume Lample 创立之初，就将推动开源 AI 科学发展确立为核心价值观。Pavan 进一步阐述了开源对于 Mistral 和广大企业的双重意义： - **科研哲学与传承**：创始人均出身学术研究前沿，深信开放研究的价值。从最初以 Mistral 7B 树立同量级 SOTA 标杆开始，Mistral 始终坚持向全球社区开放核心模型权重。 - **企业自主可控（Control）**：企业客户不再受制于单一闭源 API 供应商可能随时断供或强制下线的风险，可以按照自身节奏掌控版本迭代和数据隐私。 - **架构效率与成本（Efficiency）**：企业无需为所有任务盲目调用昂贵庞大的闭源通用大模型，而是可以针对具体场景选用最适宜的轻量模型。 - **深度定制与信任（Customization & Trust）**：开源权重让企业能够通过监督微调（SFT）或强化学习微调（RL Fine-tuning）极大释放特定任务潜能，带来无可替代的透明度与安全信任。 ### [12:40](https://www.youtube.com/watch?v=ISBFhvg6zHA&t=760s) 开源 AI 的企业级价值：数据主权、物理工程与成本优化 开源模型不仅能够帮助企业满足严苛的本地监管法规、建立透明度与信任机制，同时也为学术界提供了至关重要的研究基石。在当前大模型研发极度依赖重资本投入的背景下，开源权重让研究人员得以在没有天价算力支持的情况下，深入探索机械可解释性（Mechanistic Interpretability）与后训练（Post-training）等前沿课题。从全球市场来看，开源 AI 正受到金融服务、高端制造等受监管行业的广泛青睐，美国与欧洲的头部企业正加速引入 Mistral 系列模型来构建专属的定制化智能系统。 在金融领域，大型银行、对冲基金和资产管理公司拥有极其敏感的投资策略与核心“超额收益”（Alpha）逻辑。这些机构极度不愿将私密数据通过通用商业 API 发送给第三方提供商，以免稀释自身好不容易建立起来的行业竞争优势；开源模型在本地化安全部署上的优势恰好契合了这一刚需。与此同时，Mistral 也在积极向工业工程领域拓展，不仅收购了专注于物理学与前沿 AI 交叉融合的欧洲创新团队 MEI，更坚信继代码智能体（Coding Agents）彻底颠覆软件工程之后，下一波生产力革命必将在工业工程领域全面爆发。 面对目前许多数字化原生企业为每个员工在编程智能体上耗费数十万美元的沉重财务负担，首席技术官（CTO）与企业高管的核心诉求是在不牺牲生产力的前提下大幅削减成本。开源生态让企业能够告别“无论面对什么钉子都盲目挥动最大号锤子”的低效做法，赋予了团队针对具体任务灵活选用最适模型、并随时无缝迁移至更高性价比架构的自主权。 - **学术研究赋能**：开源模型为研究界提供了低门槛的探索载体，推动机械可解释性与微调后训练等方向的理论突破。 - **金融数据主权**：支持对冲基金与大型银行私有化部署，避免通过公共 API 泄露敏感策略与核心商业 Alpha。 - **物理工程与前沿结合**：通过收购欧洲团队 MEI，将 AI 前沿技术与物理学建模相融合，探索工业制造场景中的自动化与效能跃迁。 - **精准选型控本**：打破对超大单一商业模型的依赖，让企业能够根据任务复杂度匹配最轻量高效的模型，大幅降低推理与订阅开销。 ### [16:02](https://www.youtube.com/watch?v=ISBFhvg6zHA&t=962s) Mistral 的全栈 AI 布局：从模型底座、Forge 定制到算力与工程落地 随着企业级合作的不断深化，合作伙伴的诉求早已超越了单纯的模型权重，延伸至上层应用构建、底层算力支撑以及端到端落地交付。Mistral 虽以研发前沿基础模型起家，但如今已构建起一套完整的“全栈 AI”（Full-Stack AI）产品矩阵，旨在全方位融入企业现有工作流，实现端到端的业务自动化与效能倍增。 在模型与应用层，Mistral 提供了针对办公协同与代码生成的智能体套件；在模型定制方面，Mistral 推出了 Forge 定制化平台。企业开发者可以借助 Forge 对开源权重模型执行监督微调（SFT）或强化学习微调（RL Finetuning）。在拥有垂直领域专业知识与特定上下文的企业场景中，经过 Forge 深度调优的紧凑型轻量模型，往往能在特定小众工作流中爆发出超越超大规模通用模型的惊人表现。为了让微调后的成果顺利转化，Mistral 还配备了 Studio 开发平台，供企业基于 Forge 产出的定制模型与各类前沿构件，快速编排、开发并部署企业级 AI 智能体与专属业务应用。 ```mermaid flowchart TD accTitle: Mistral全栈AI产品与服务体系架构 subgraph ComputeLayer["底层算力层 (Compute)"] A["裸金属 GPU 集群算力"] --> B["覆盖训练与异构/长尾推理工作负载"] end subgraph CustomizationLayer["模型与定制层 (Forge & Models)"] C["前沿开源权重模型"] --> D["Forge 定制平台"] D --> E["SFT 与强化学习微调 (RL)"] E --> F["垂直场景轻量定制模型"] end subgraph ApplicationLayer["应用与交付层 (Studio & FDE)"] F --> G["Studio 智能体与应用编排平台"] H["智能体套件 (Work / Coding Agents)"] --> G G --> I["前向部署工程团队 (FDE / Applied Science)"] I --> J["企业端到端业务自动化落地"] end ComputeLayer --> CustomizationLayer CustomizationLayer --> ApplicationLayer ``` 在底层基础设施与实施交付方面，Mistral 直接提供裸金属 GPU 集群算力（Compute），灵活承载训练与长尾推理等多样化异构算力需求。此外，针对复杂业务场景，Mistral 还组建了前向部署工程（Forward Deployed Engineering, FDE）与应用科学专家团队，工程师会直接深入企业一线业务环境，量身打造符合企业特定环境与流程的自动化 AI 系统。从基础模型、Forge 定制、Studio 智能体搭建，到底层算力供给与专家驻场交付，Mistral 的全栈 AI 体系正在为全球各地的企业与政府机构提供真正具备全流程控制权、高透明度与深远变革力的前沿智能化支持。 - **Forge 定制化平台**：支持基于 SFT 和 RL 的高效模型定制，凭借私有数据与业务上下文让小尺寸模型在专业场景实现“以小博大”。 - **Studio 智能体工作台**：无缝承接 Forge 产出的模型资产，提供开箱即用的企业级 AI 智能体编排与应用开发环境。 - **灵活的 Compute 算力支持**：提供裸金属 GPU 算力资源，全面兼顾异构业务场景下的高并发推理与模型训练负载。 - **前向部署工程（FDE）体系**：由 Mistral 应用科学家与工程师直接驻场或协同攻坚，确保定制化 AI 应用在真实业务场景中无缝交付落地。 ## Notable quotes > [00:08](https://www.youtube.com/watch?v=ISBFhvg6zHA&t=8s) put frontier AI in everyone's hand that they truly control — Pavan Kumar Reddy > [03:17](https://www.youtube.com/watch?v=ISBFhvg6zHA&t=197s) 为每个提取的文本片段提供精准的边界框和置信度得分，人工只需针对低置信度区域复核，即可实现近乎完美的解析效果 — Pavan Kumar Reddy > [03:17](https://www.youtube.com/watch?v=ISBFhvg6zHA&t=197s) 不同于单纯输出文字答案的语言模型，它能够接收数学命题并在 Lean 4 形式化编程语言中给出完整证明 — Pavan Kumar Reddy > [07:03](https://www.youtube.com/watch?v=ISBFhvg6zHA&t=423s) 实测性能却超越了数倍于其体积的大模型，并同样原生支持文本与图像多模态输入 — Pavan Kumar Reddy > [08:49](https://www.youtube.com/watch?v=ISBFhvg6zHA&t=529s) 开源 AI 正在行业内迎来属于自己的高光时刻，而语音无疑是 AI 最重要的未来界面 — Howard Cohen > [08:49](https://www.youtube.com/watch?v=ISBFhvg6zHA&t=529s) 语音交互已经从附加功能转变为与 AI Agent 协作的“准入门槛”（Table stakes） — Pavan Kumar Reddy > [10:42](https://www.youtube.com/watch?v=ISBFhvg6zHA&t=642s) 企业客户不再受制于单一闭源 API 供应商可能随时断供或强制下线的风险，可以按照自身节奏掌控版本迭代和数据隐私 — Pavan Kumar Reddy > [12:40](https://www.youtube.com/watch?v=ISBFhvg6zHA&t=760s) 继代码智能体彻底颠覆软件工程之后，下一波生产力革命必将在工业工程领域全面爆发 — Howard Cohen > [12:40](https://www.youtube.com/watch?v=ISBFhvg6zHA&t=760s) 告别“无论面对什么钉子都盲目挥动最大号锤子”的低效做法，赋予了团队针对具体任务灵活选用最适模型、并随时无缝迁移至更高性价比架构的自主权 — Pavan Kumar Reddy > [16:02](https://www.youtube.com/watch?v=ISBFhvg6zHA&t=962s) 在拥有垂直领域专业知识与特定上下文的企业场景中，经过 Forge 深度调优的紧凑型轻量模型，往往能在特定小众工作流中爆发出超越超大规模通用模型的惊人表现 — Pavan Kumar Reddy ## Mentioned Mistral, Pavan Kumar Reddy, Howard Cohen, Arthur Mensch, Timothée Lacroix, Guillaume Lample, Mistral Large 3, Mistral Medium 3.5, Mistral 3, Ministral 3, Mistral 7B, Mixtral, Mistral OCR 4, Leanstral 1.5, Lean 4, Robostral, Shieldstral, Voxtral, Voxtral Chat, Voxtral Transcribe, Voxtral TTS, MEI, Forge, Studio, PutnamBench, Apache 2.0 ## Video description Pavan Kumar Reddy - AI Scientist, Mistral Howard Cohen - Head of Communications, North America, Mistral Mistral builds compact, high-performance models and openly releases much of its work to advance the industry. Join Pavan Kumar Reddy, one of our first US-based AI scientists, as he discusses our breadth of research and innovation in audio processing, OCR, robotics navigation and reasoning models which is helping power physical autonomy and sophisticated machine intelligence across diverse environments. Find out more about how our lab is making frontier AI more transparent, transformative, and powerful for the global community. [Source](https://www.youtube.com/watch?v=ISBFhvg6zHA)

[0:08] Pavan Kumar Reddy: put frontier AI in everyone's hand that they truly control [3:17] Pavan Kumar Reddy: 为每个提取的文本片段提供精准的边界框和置信度得分，人工只需针对低置信度区域复核，即可实现近乎完美的解析效果 [3:17] Pavan Kumar Reddy: 不同于单纯输出文字答案的语言模型，它能够接收数学命题并在 Lean 4 形式化编程语言中给出完整证明 [7:03] Pavan Kumar Reddy: 实测性能却超越了数倍于其体积的大模型，并同样原生支持文本与图像多模态输入 [8:49] Howard Cohen: 开源 AI 正在行业内迎来属于自己的高光时刻，而语音无疑是 AI 最重要的未来界面 [8:49] Pavan Kumar Reddy: 语音交互已经从附加功能转变为与 AI Agent 协作的“准入门槛”（Table stakes） [10:42] Pavan Kumar Reddy: 企业客户不再受制于单一闭源 API 供应商可能随时断供或强制下线的风险，可以按照自身节奏掌控版本迭代和数据隐私 [12:40] Howard Cohen: 继代码智能体彻底颠覆软件工程之后，下一波生产力革命必将在工业工程领域全面爆发 [12:40] Pavan Kumar Reddy: 告别“无论面对什么钉子都盲目挥动最大号锤子”的低效做法，赋予了团队针对具体任务灵活选用最适模型、并随时无缝迁移至更高性价比架构的自主权 [16:02] Pavan Kumar Reddy: 在拥有垂直领域专业知识与特定上下文的企业场景中，经过 Forge 深度调优的紧凑型轻量模型，往往能在特定小众工作流中爆发出超越超大规模通用模型的惊人表现

[0:08] Pavan Kumar Reddy: putting frontier AI directly into everyone's hands with complete control and customization [0:08] Pavan Kumar Reddy: Mistral Large 3: [0:08] Pavan Kumar Reddy: The flagship open-weight Mixture-of-Experts (MoE) model, building on the architecture Mistral pioneered in 2023 with Mixtral. [3:17] Pavan Kumar Reddy: OCR4: [3:17] Pavan Kumar Reddy: Mistral's fourth-generation optical character recognition system achieves a 72% win rate over major industry competitors. [3:17] Pavan Kumar Reddy: Leanstral 1.5: [3:17] Pavan Kumar Reddy: Robostral: [3:17] Pavan Kumar Reddy: Shieldstral: [3:17] Pavan Kumar Reddy: Voxtral Family: [8:49] Howard Cohen: why open source is central to Mistral's philosophy and commercial momentum. [8:49] Howard Cohen: Organizations avoid API vendor lock-in or sudden service deprecation. [13:02] Howard Cohen: open-weight models tuned specifically for their codebases.

[0:08] Hello everyone. I'm Pawan. I'm a AI  research scientist uh at Mistral uh [0:13] based in I'm based in the San Francisco office.  I lead the audio uh research team. So, Mistral, [0:22] so Mistral is a frontier AI company and we build  open models and a full AI stack uh ranging from [0:32] compute infrastructure to AI applications uh that  enable enterprises and businesses to truly own, [0:39] customize and control their AI. Uh in a single  line, we want to put frontier AI in everyone's [0:45] hand that they truly control. uh to begin with  I'll start with our frontier model offerings. [0:52] These range from state-of-art small compact edge  models to Asgentech cloud models. Uh and they [1:00] have capabilities uh ranging from multimodel  which involves text, speech, vision and even [1:07] uh physical AI like robotics. uh to start with uh  the generalist model family. These are the models [1:15] that you might be most familiar with when you  think of uh LLMs and AI. These are the models that [1:20] you would use for conversational chat bots or deep  research agents or more recently the very popular [1:29] coding agent uh phenomena. And we know a single  model doesn't uh cover the use cases uh across the [1:39] cost. efficiency frontier uh for every use case  that you might have in an enterprise. So we have [1:46] a offering at each of the weight class starting  with Mistral large 3 which is our flagship largest [1:54] open weight model. It's a mixture of expert model  which we also pioneered all the way back in 2023 [2:00] with Mixtra uh the one of the first open mixture  of experts model. This model in particular is [2:08] pretty large. It's 675 billion total parameter and  a 45 billion 41 billion active parameter model. [2:16] Uh and mistral medium 3.5 it's our enterprise  workhorse. It's a 128 billion dense model. Uh [2:23] owing to its compact dense nature. It's easy  to deploy on a single node on enterprises [2:28] and small four trades uh performance uh with  efficiency. So it's uh it packs the punch of [2:36] a 100 billion uh parameter model but uh over  into its 6 billion active parameters it runs [2:42] really fast and very low latency. Uh last is  the minstral 3 series. Uh it's a family of [2:49] three models again and these are dense models and  they are meant for edge applications and use cases [2:57] uh which can be serviced by small dense models.  All of these models are multilingual. They support [3:04] 100 plus languages. All of them are multimodal  as well. Uh they support native vision input for [3:10] image reasoning and they're all open weights  uh showcasing our commitment to open source. [3:17] In addition to these generalist models, we also  have a lot a lot of specialized models which are [3:23] which have been purpose-built for uh different use  cases that we saw in our customers in enterprise. [3:31] Um I I will talk about a few of them today.  Starting with OCR4 uh which is our fourth [3:37] generation optical character recognition model.  So these models can take a image of a document or [3:44] a scan for instance and extract the text content  from it as you can see on the right. Uh this is a [3:51] state-of-the-art model and uh one of the best  uh in the industry with 72% win rate against [3:59] uh most of our closest competition. This model  in particular doesn't just extract the text, [4:04] it also does segmentation. So it provides you  the provenence for the text. So where the text [4:09] came from in the scan by providing the bounding  box uh associated with each text snippet that was [4:15] extracted. And in addition to this bonding box, it  also gives you inline confidence scores. The main [4:21] utility of these confidence scores is you can  improve the overall performance of the system [4:27] by even looping in humans uh by having them check  the places where the model is not confident in [4:34] and in many instances we saw uh almost perfect  u parsing performance with humans in the loop uh [4:42] which is very critical for sensitive applications.  For instance, uh this is multilingual. It supports [4:48] 170 plus languages and oming is compact nature.  It can run on your own infrastructure. Next is [4:54] link install 1.5. It's a formal verification open  agent. So what does this mean? So instead of just [5:00] providing you an answer uh this it can you can  take a mathematical hypothesis and provide you [5:08] a proof uh in lean for programming language. Uh  this model in particular is an agent. So it can [5:15] actually perform uh actions and iterate. Uh these  actions could include editing repositories or [5:21] running commands, getting feedback from a compiler  and improving on its own proof. On the right side, [5:28] um I show uh the strong test time scaling  capability of this model. Basically, if you give [5:35] it more compute in the form of tokens at runtime,  the performance of the model improves. uh on this [5:42] specific benchmark that's being highlighted,  the Patnum bench, which is a benchmark which [5:47] comprises uh problems taken from this very famous  Patnham mathematics competition. It reaches 87% [5:55] uh pass rate uh especially as you give it more  and more uh test time token and it's a pretty [6:01] small model to too. Uh next is Robostrol. It's our  first foray into a physical AI. It's an embodied [6:11] AI agent which can take navigation instructions  like the one shown on the right. Uh these could be [6:16] something like u navigate uh to the closest couch  uh just ahead of the conference room something [6:23] like that. And it can take that instruction uh  and it takes a single input in the form of the [6:29] RGB image uh from the camera and able to come up  with a navigation plan and uh execute actions and [6:36] as it encounters new obstacles it can u reform the  plan and um it and dynamically update the plan and [6:46] navigate to the uh desired destination. Uh it  it has a very strong performance reaches 76.6% [6:52] 6% success rate. Uh, and it's quite generalizable.  Works across a wide range of robots, wide range [6:57] of cameras, and even and uh can handle unseen  obstacles. Uh, handle unseen obstacles during [7:03] training. Uh, next up, I'm really happy to  announce a new model today. Uh, shield straw. [7:10] We're releasing it just today. It's a guardrail  model. Uh, essentially a 3B open weight safety [7:17] classifier which we're releasing under Apache  2. So, what does it do? It's a content policy [7:23] classifier. So you can define a content policy  in the form of what is acceptable for your use [7:27] case versus not. And given a query and a document  from a user for instance, it can tell you whether [7:34] it violates the policy or whether adheres to  the policy along with a confidence score. It [7:38] it actually punches well above its weight uh by  outperforming models which are seven times its [7:44] size. And given it's a 3B tiny model, it can run  on a single 16 GB GPU. It's also multimodel. So in [7:53] addition to your queries and documents being text,  it also supports uh image input. Uh I encourage [7:59] everyone to try it out if it fits your use case.  Finally, walkthrough family. So this is our speech [8:06] uh lang of models which actually I mostly  worked on in the last one year personally. [8:13] Uh so we released three models over the last one  year starting with vauil chat which takes a audio [8:19] input and text prompt and produce text answers  like summarization transcription and event tool [8:24] calls. And earlier in the year we released two  models the vaual transcribe and TTS which are uh [8:30] building blocks of a voice agent. Currently, we're  focused on the next generation of speech-to-pech [8:36] model which can enable natural conversation as  opposed to turn taking dynamics which is usually [8:41] uh reminiscent of the current cascaded uh set of  models. Uh to for the next section uh I'd like [8:49] to invite uh the head of communication for North  America Cohen onto the stage. So open source AI [8:57] it's having its moment in the industry today and  for good reason. Open source helps enterprises [9:03] have more control over their AI systems. Uh it  helps to ensure that we can deliver on this vision [9:09] of frontier AI in everyone's hands. So we'll be  diving into all of that today. But first let's [9:15] talk about audio and the team that you've been  building at Nestrol. So audio really is the future [9:22] interface of AI. Can you speak to to why you think  that's the case? Uh great question. I think uh [9:29] audio is a fundamental interface for human human  communication and increasingly for human machine [9:35] communication. Uh especially with adu uh advent  of AI agents like coding agents, co-work agents, [9:42] it's just natural to get work done in a modality  that people are already most comfortable in and [9:47] audio is just a a fundamental interface. uh it's  actually even more uh relevant in hands-free [9:54] context for instance in an incar assistant where  you're driving and don't want to type or in an [10:00] industrial setting where you're operating heavy  machinery for instance and want to ask questions [10:04] or issue commands for instance uh in addition to  it being a natural interface to machines. It also [10:10] is a nice accessibility mechanism to enable human  human communication. For instance, for instance, [10:16] if there are people who don't speak each other's  language or have struggle understanding each [10:21] other, then uh the a real-time speech translation  system could operate as an intermediary uh to aid [10:30] the communication between each other. Uh so there  it's just a uh barebones communication medium [10:37] and becoming increasingly table stakes for AI  interacting with AI agents. Yeah, audio AI really [10:42] is critical to being able to break down barriers  to communication all around the world. Let's [10:47] shift to open source. Uh so this is fundamental to  Mistrol. uh when mistrol was founded several years [10:55] ago by Arthur, Tim and Guom, it was very much to  take uh AI and the science and the direction of [11:02] open source. Uh and the reason for that was it  was a fundamental value. It was a belief in the [11:08] technology and in the benefit that it could bring  to the scientific community. Can you speak to [11:12] why open-source is so fundamental to the Mistral  journey? Definitely. uh as you rightly highlighted [11:20] uh Mistral's founders uh have research roots and  uh they have a strong belief in open research [11:27] um uh philosophy and which it started from our  uh uh first flagship model 7B uh which kind of [11:38] one of the early opensource pioneers uh and was  actually state-of-the-art for its weight when it [11:45] was released and we continued that uh uh culture  event to this day uh with all the generalist model [11:52] family that I kind of uh walked through earlier  in the presentation all of them are open weights [11:58] and we're committed to open weights in addition  to a philosophical thing it's all there is also [12:01] pragmatic angle to this where uh enterprises have  a lot of control you're not at the mercy of an API [12:07] provider to cut off access for whatever reason  and you can control your own transition from [12:12] one model to another model uh based on your  internal needs needs. In addition to control, [12:17] you also get efficiency from it. You pick the  right model for the task as opposed to one [12:22] super powerful journalist model for all tasks.  You can customize it. Um we we have showcased [12:28] again and again that if you fine-tune it or do re  reinforcement learn finetuning, you can improve [12:34] performance significantly. Uh it also gives you  a sense of trust and transparency. uh ability to [12:40] adhere to uh local regulation when you require  to have this trust and transparency and it's [12:46] also a nice artifact for research community with  uh developing these very large models being so [12:51] capital intensive. The artifact is really helpful  on pushing the frontier on things like mechanistic [12:56] interpretability or post- training research.  Now enterprises as you point out all around the [13:02] world are seeing the value of open-source AI. Um  we're seeing that across industries in particular [13:09] in regulated industries in financial services  in manufacturing all of which are are spaces [13:16] where we have quite a meaningful presence.  Can you speak to the value specifically for [13:21] enterprises and being able to build their own  custom intelligence um from open source? Yeah, [13:27] definitely. I think uh with this enterprises  are increasingly appreciating the benefits of [13:33] open source. some of this uh things that you also  mentioned around control uh uh having the ability [13:40] to trust the model and so forth and we are seeing  increasing adoption of Mistral family of models in [13:47] uh the US enterprises uh in addition to European  enterprises and rest of the world. To give you [13:53] a few examples, uh we can start with finance for  instance. There is a lot of sensitive information, [13:59] sensitive strategy information and uh what can  be described as an alpha for a specific financial [14:04] institution and uh they would like to keep it  uh pretty safe and not have it going through all [14:12] kinds of a uh API providers and also uh diffuse  their very hardearned uh advantage across the [14:21] industry and it's a natural fit. uh some of our  biggest customers today are like uh largest banks [14:27] uh hedge funds and asset managers for instance in  addition to finance we also seeing an uptick in [14:33] industrial engineering uh earlier this year  we acquired MEI uh any European based uh uh [14:41] group which develop um AI at the intersection  of the frontier technology today and physics [14:49] uh in service of industrial engineering and  similar to how coding agents are revolutionizing [14:55] uh software engineering. uh we strongly believe  the next wave is uh AI uh AI powered productivity [15:04] gains and in the context of industrial engineering  and uh talking about coding agents uh a lot of [15:10] digital native companies are increasingly adopting  coding agents uh some of them are paying as much [15:16] as hundreds of thousands of dollars per employee  uh which is a a lot of money and one of the top [15:22] of mind consideration for CTOs and executives  is u to reduce costs and so the the avenue for [15:32] reducing cost is you don't want to lose the  productivity gains but pick the right model [15:36] for your task and you don't want to be picking  the biggest hammer out there for every nail you [15:43] have. Uh so that's where open source also comes  in. So it allows you the flexibility of picking [15:49] the right model and also flexibility for moving to  a more efficient model when one comes along. So it [15:56] has uh advantages across the board, but these are  just a few instances of it. Yeah. And it's such an [16:02] interesting point that you make that we've seen  really in software uh today this AI revolution, [16:09] mass adoption, but really we're shifting to see  industrial engineering as a big bet where AI is [16:15] going to have quite a bit of of an impact. And  that's certainly an area of focus for ML. Um, [16:20] let's shift to the full stack because a lot of  the partners that we're working with, they're not [16:25] just looking at the models, but they also want the  application layer and they also want the compute. [16:31] Uh, and you need a partner that's able to provide  that full stack AI offering to really deliver that [16:37] level of control and innovation to the partners  we're working with. Can you speak to that full [16:42] stack AI approach and why that's so fundamental  to the Mistl way? Yeah. So we started as a model [16:48] company and we still uh are at the at the core  build frontier AI models. But models are just [16:55] a component in a broader stack. Uh at the end of  the day we want to provide value for enterprises [17:01] and hopefully uh have them get their productivity  gains and uh improvements across the workflows and [17:09] automate some of the workflows. So to that end  we have a full stack of AI uh offerings uh like [17:16] you highlighted uh for instance uh a few of them  covering this would be starting with Mistral W for [17:22] instance it's our core work and coding agents uh  with wbe work and wbe code and we have forge which [17:30] is our uh customization platform you can take  any AI model open weights AI model and do simple [17:38] fine-tuning or reinforcement learning finetuning  and adapt it to the way you want it. Like you can [17:43] take a small model, sometimes very small compact  models can surprise you by outperforming even the [17:49] largest models because it's a niche workflow and  you have the enterprise context and data for it. [17:55] Uh and then on the context of enterprise  context, we also have studio which allows [17:59] you to develop which is a platform for you to  develop AI agents and AI applications in your [18:04] enterprise using Mistral's offerings using the  models you might have customized and produced [18:10] the artifacts you produced from Forge. All of  that can be uh developed through our studio [18:16] offering. And finally, compute. uh compute is the  bare bones access to GPU clusters and these cover [18:26] the heterogenous use cases longtail use cases  not necessarily all long time all long tail they [18:32] can be training workloads inference workloads it  just gives you access to the barebones compute and [18:37] finally we also have the applied engineering  applied science forward deployed engineering [18:42] offering as well where uh our employees come work  in in the context of your enterprise and develop [18:48] an application which is tailor made for your  enterprise situation and automation and hopefully [18:55] give the gains that you're looking for from AI.  So you really it starts with that model uh it can [19:01] be an open model you can then take forge uh to  be able to customize it uh and then finally you [19:08] in order to really have control over your AI you  need to also have access to that compute and so [19:12] that's the full stack AI offering that mistrol is  really delivering to enterprises and governments [19:17] all around the world um at the end of the day  there are so many partners at this conference [19:21] who have been uh building with us on this journey  and we want to you know thank them for for being [19:26] on that journey with and we're excited to keep  building with many of you here at the event.