微软Foundry上线Mistral Large 4
Mistral Large 4 in Microsoft Foundry: Build sovereign AI within a consistent platform experience
What's launching
即将发布
Leading AI delivers more than model quality. For governments, critical infrastructure operators, financial institutions, healthcare organizations, and global enterprises, requirements for data control, deployment location, security, resilience, latency, and operational continuity are also crucial. Today, Mistral Large 4 (Le Chonk) reaches general availability from Mistral. As the first cloud provider to bring Mistral Large 4 latest in public preview to customers, Microsoft is making Mistral’s latest open-weight model available through Microsoft Foundry in the coming days. This gives customers access to Mistral's European-built model portfolio for reasoning, coding, cybersecurity, and agentic applications, with multilingual and multimodal support built in. Microsoft Foundry provides the consistent platform layer for evaluation, governance, and operations across all of these workloads.
领先的 AI 提供的不仅仅是模型质量。对于政府、关键基础设施运营商、金融机构、医疗组织和全球企业而言，对数据的控制、部署位置、安全性、韧性、延迟以及运营连续性同样至关重要。今天，Mistral Large 4 (Le Chonk) 从 Mistral 正式全面推出。作为首家将 Mistral Large 4 最新公开预览版带给客户的云提供商，微软将在未来几天通过 Microsoft Foundry 提供 Mistral 最新的开放权重模型。这使客户能够访问 Mistral 的欧洲构建模型组合，用于推理、编码、网络安全和智能体应用，并内置多语言和多模态支持。Microsoft Foundry 为所有这些工作负载提供了统一的评估、治理和操作平台层。
Also becoming generally available in Microsoft Foundry is Mistral OCR 4 (latest), moving from public preview to GA on the same day Microsoft makes available Mistral Large 4 latest in Public Preview. OCR 4 (latest) extends the Mistral portfolio in Foundry for document AI scenarios leveraging the latest OCR 4.1 weights internally, giving developers a purpose-built option for extracting and structuring information from documents alongside Mistral Large 4 latest for reasoning, coding, multimodal, and agentic applications.
同时，Mistral OCR 4 (latest) 也在 Microsoft Foundry 中实现全面可用，与微软在公开预览阶段提供 Mistral Large 4 latest 的同一天从公共预览转为 GA（一般可用性）。OCR 4 (latest) 扩展了 Foundry 中的 Mistral 产品组合，针对文档 AI 场景利用内部最新的 OCR 4.1 权重，为开发者提供了一种专为从文档中提取和结构化信息而设计的选项，可与 Mistral Large 4 latest 配合使用，以支持推理、编码、多模态和智能体应用。
“Mistral Large 4 represents exactly the kind of innovation enterprises want access to, with the openness and control they increasingly require. By bringing it to Blackbox through Microsoft Foundry, we are giving customers more choice without asking them to compromise on trust.”
“Mistral Large 4 代表了企业希望获得的那种创新，具备他们日益需要的开放性和控制权。通过 Microsoft Foundry 将其引入 Blackbox，我们为客户提供了更多选择，而不必要求他们在信任方面做出妥协。”
Robert Rizk, Co-Founder and CEO Blackbox AI
Robert Rizk，Blackbox AI 联合创始人兼首席执行官
*Blackbox AI, offering a secure E2EE AI inference and coding platform for enterprises, is among the early interested adopters deploying Mistral Large 4 through Microsoft Foundry.
*Blackbox AI 为企业提供安全的端到端加密（E2EE）AI 推理和编码平台，是首批有兴趣采用并通过 Microsoft Foundry 部署 Mistral Large 4 的客户之一。
European model choice for sovereign AI
主权 AI 的欧洲模型选择
According to Mistral, Mistral Large 4 is Mistral’s largest and most capable Mixture-of-Experts (MoE) model to date, meaning only a subset of the model's parameters activate for each input, improving efficiency. The model was trained in Mistral’s datacenters in Europe, with 1 trillion parameters and 52 billion active parameters. Its open weights add another important dimension for organizations pursuing sovereign AI: greater control over where the model runs, how it is operated and adapted, and the infrastructure and data boundaries around it.
据 Mistral 称，Mistral Large 4 是 Mistral 迄今为止最大、最强大的混合专家（MoE）模型，这意味着对于每个输入，仅激活模型参数的一部分，从而提升效率。该模型在 Mistral 位于欧洲的数据中心进行训练，拥有 1 万亿参数和 520 亿活跃参数。其开放权重为追求主权 AI 的组织增添了另一个重要维度：对模型运行位置、操作与适配方式以及围绕它的基础设施和数据边界拥有更大的控制权。
That matters because sovereignty is not a single deployment claim. It is a continuum of control across data, models, infrastructure, operations, and connectivity. Microsoft and Mistral are working together to expand access to Mistral models across Microsoft platforms and deployment environments, giving organizations more options as they design AI architectures around sovereignty, security, resilience, and operational requirements. In Microsoft Foundry, Mistral Large 4 adds a European-built, open-weight model to the choices organizations can evaluate as they design AI architectures around those requirements. Mistral Large 4 will be available in Sweden Central region for Global Standard and EU DataZone deployment, providing options for teams with data residency requirements.
这之所以重要，是因为主权并非单一的部署声明。它是一个涵盖数据、模型、基础设施、运营和连接性的控制连续体。微软与 Mistral 正携手合作，在微软平台和部署环境中扩大对 Mistral 模型的访问权限，使组织在设计围绕主权、安全性、弹性和运营需求的 AI 架构时拥有更多选择。在 Microsoft Foundry 中，Mistral Large 4 为组织在评估满足这些需求的 AI 架构时，提供了一个在欧洲构建的开放权重模型选项。Mistral Large 4 将在瑞典中部地区的 Global Standard 和 EU DataZone 部署中提供，为有数据驻留要求的团队提供选择。
Capabilities that matter for builders
对开发者至关重要的能力
- Cybersecurity reasoning and automation. Synthesize threat intelligence, security telemetry, vulnerability context, secure-coding guidance, and response procedures. High-impact actions should remain behind explicit permissions, approvals, and analyst review.
- Reasoning, coding, and tool use. Analyze complex instructions, generate and explain code, produce structured outputs, and coordinate approved tools across multi-step workflows. According to Mistral, Mistral Large 4 is a leading agentic coding model, reaching the highest open weights score on Cognition's private FrontierCode 2.0 benchmark, with 19.7%.
- Long-context and multimodal understanding. Work across large documents, repositories, policies, histories, and supported visual inputs for grounded research, document analysis, and operational support.
- Multilingual experiences. With native fluency in more than 160 languages, teams can use Mistral Large 4 to build localized applications and cross-border workflows while evaluating quality on the languages and specialized terminology that matter to the organization.
- 网络安全推理与自动化。综合威胁情报、安全遥测数据、漏洞上下文、安全编码指南和响应程序。高影响的操作应保留在明确的权限、审批和分析师审查之后。
- 推理、编码和工具使用。分析复杂指令，生成并解释代码，产生结构化输出，并在多步骤工作流中协调已批准的工具。据 Mistral 称，Mistral Large 4 是一款领先的代理式编码模型，在 Cognition 的私有 FrontierCode 2.0 基准测试中取得了最高开放权重分数，达到 19.7%。
- 长上下文和多模态理解。处理大型文档、代码库、策略、历史记录和支持的视觉输入，以支持基于事实的研究、文档分析和运营支持。
- 多语言体验。凭借对 160 多种语言的母语级流利度，团队可以使用 Mistral Large 4 构建本地化应用程序和跨境工作流，同时评估对组织重要的语言和专用术语的质量。
Why build with Mistral Large 4 in Microsoft Foundry
为何在 Microsoft Foundry 中使用 Mistral Large 4 进行构建
Platform consistency across models and environments
跨模型和环境的平台一致性
Microsoft Foundry helps teams avoid separate architectures for every model or environment. Developers can discover and compare models, evaluate them on representative data, deploy through consistent interfaces, and connect models to enterprise data and tools. Azure identity, network isolation, policy, secrets management, content safety, monitoring, logging, and governance provide controls around endpoints and applications. For sensitive agents, teams can apply least-privilege tool access, explicit action boundaries, human approval, and observable execution.
Microsoft Foundry 帮助团队避免为每个模型或环境构建独立的架构。开发人员可以查找并比较模型，在代表性数据上评估模型，通过一致的接口进行部署，并将模型连接到企业数据和工具。Azure 身份验证、网络隔离、策略、密钥管理、内容安全、监控、日志记录和治理为端点和应用程序提供控制措施。对于敏感型智能体，团队可以应用最小权限工具访问、明确的操作边界、人工审批以及可观察的执行过程。
Security and governance for sensitive workloads
针对敏感工作负载的安全与治理
This combination is especially relevant to public sector, critical infrastructure, manufacturing, industrial operations, healthcare, and financial services. These organizations need advanced AI, but adoption depends on regional alignment, regulatory compliance, operational resilience, and clear accountability. European model choice combined with Microsoft security, compliance, governance, and operational support provides a practical path from experimentation to production.
这种组合对于公共部门、关键基础设施、制造业、工业运营、医疗保健和金融服务尤其相关。这些组织需要先进的 AI，但采用程度取决于区域合规性、监管遵从性、运营韧性和明确的问责制。结合欧洲模型选择与 Microsoft 的安全性、合规性、治理和运营支持，提供了从实验到生产的可行路径。
For cybersecurity workloads, the operating model matters as much as model quality. Teams should define what data the model can access, which tools it may call, which actions require approval, how evidence is logged, and when the system must defer to an analyst. These controls turn model capability into a reviewable, auditable workflow rather than an autonomous black box.
对于网络安全工作负载，运营模式的重要性不亚于模型质量。团队应定义模型可以访问哪些数据、可以调用哪些工具、哪些操作需要批准、证据如何记录，以及系统在何时必须交由分析师处理。这些控制措施将模型能力转化为可审查、可审计的工作流程，而非自主的黑盒系统。
Get started
开始使用
Start by selecting the Mistral Large 4 model card in Model Catalog | Microsoft Foundry once live. Deploy the model in your Foundry project and test it with representative prompts, tools, and evaluation datasets. Compare quality, safety, latency, and cost against the requirements of your application before scaling to production.
一旦 Model Catalog | Microsoft Foundry 上线，首先选择 Mistral Large 4 模型卡片。在您的 Foundry 项目中部署该模型，并使用代表性的提示词、工具和评估数据集进行测试。在扩展到生产环境之前，根据应用程序的需求对比其质量、安全性、延迟和成本。
Useful resources
有用资源
- Mistral AI announcement: Mistral Large 4
- Microsoft Foundry Models overview
- Deploy Microsoft Foundry Models
- Mistral AI 公告：Mistral Large 4
- Microsoft Foundry 模型概述
- 部署 Microsoft Foundry 模型
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力