跳到主内容
@wquguru
精选80Microsoft Foundry 博客(RSS)产品发布/更新多源精选 ×7

微软发布MAI-Code-2等模型及Foundry平台多项更新

From model choice to your durable advantage: Build AI that keeps improving

原文
发到 X

Every organization now has access to the same frontier models. What sets your business apart is how you put those models to work with your data, your expertise, and your definition of success. In our conversations with customers, three priorities come up again and again: flexibility to choose the right model for each task, control to turn their own data, expertise and feedback into durable advantage, and confidence in the models and platform they build on. Their expertise should strengthen their business, not deepen their dependence on any one AI lab.

如今,每个组织都能访问相同的前沿模型。真正让你的业务脱颖而出的是,你如何利用这些模型结合你的数据、专业知识和对成功的定义来开展工作。在与客户的交流中,三个优先级反复出现:为每项任务选择合适模型的灵活性、将自身数据、专业知识和反馈转化为持久优势的控制力,以及对所构建模型和平台的信心。他们的专业知识应强化其业务,而不是加深其对任何一家 AI 实验室的依赖。

With more than 11,000 models from Microsoft, OpenAI, Anthropic, Meta, Mistral, the open-source community and others, Microsoft Foundry gives organizations the breadth to match models to their needs. Today at Microsoft Ignite, we’re announcing updates across Foundry Models that help teams make that choice pay off: new models from Microsoft AI (MAI) and partners, intelligent model routing, more ways to run open models, and new training options that turn your data and feedback into models built for your business.

微软 Foundry 提供了涵盖来自 Microsoft、OpenAI、Anthropic、Meta、Mistral、开源社区及其他来源的 11,000 多个模型的广泛选择,使组织能够根据需求匹配模型。今天在 Microsoft Ignite 大会上,我们宣布了对 Foundry Models 的一系列更新,帮助团队做出更有价值的选择:新增来自 Microsoft AI (MAI) 和合作伙伴的模型、智能模型路由、更多运行开源模型的方式,以及能将你的数据和反馈转化为专为业务打造的模型的新训练选项。

Make it fit: Start with what good looks like

量身定制:从明确优秀标准开始

The leaderboard is a starting point, but what matters is how a model performs on your workload. In Foundry, you define success first, including quality, latency, cost per task and data residency, then evaluate candidate models against your own data with built-in and custom evaluators. The model that clears your success bar wins, not the one at the top of a benchmark.

排行榜只是一个起点,重要的是模型在你的工作负载上的表现。在 Foundry 中,你先定义成功标准,包括质量、延迟、每任务成本和数据驻留要求,然后使用内置和自定义评估器,用你自己的数据对候选模型进行评估。通过你成功标准的模型胜出,而不是排行榜榜首的模型。

Different jobs call for a different balance of capability, speed and cost. Today’s new models give you more options to match each one.

不同的工作需要能力、速度和成本的不同平衡。今天发布的新模型为你提供了更多选项,以匹配每一项需求。

Coding and agentic software development

编码与智能体软件开发

Developers need AI that can take on their hardest engineering work while keeping everyday coding fast and cost-effective. MAI-Code-2 is built for demanding agentic coding work, such as multi-step tasks, complex refactors and agent-driven development. MAI-Code-2-Flash is tuned for high-volume everyday coding. Developers can choose the balance of performance and cost that fits each task instead of using one model for everything. Try both today in GitHub, and access Code-2-Flash through a limited access preview in Foundry.

开发者需要能够承担最艰巨工程任务的 AI,同时保持日常编码的快速和成本效益。MAI-Code-2 专为具有挑战性的智能体编码工作而设计,例如多步骤任务、复杂重构和智能体驱动的开发。MAI-Code-2-Flash 则针对高容量的日常编码进行了优化。开发者可以根据每项任务的需求,选择性能与成本的平衡点,而不是对所有任务都使用同一个模型。今天即可在 GitHub 上试用这两款模型,并通过 Foundry 中的有限访问预览版获取 Code-2-Flash。

High-volume decision-making

大规模决策制定

Teams need to make high-volume, structured decisions quickly and cost-effectively to keep AI applications moving. Alongside complex work such as multi-step coding, these focused tasks call for a different balance of speed, cost and capability, helping teams reserve larger models for the complex reasoning that needs them most.

团队需要快速、高效地做出大规模的结构化决策,以推动 AI 应用的持续进展。与多步编码等复杂工作不同,这些聚焦型任务需要在速度、成本和能力之间寻求不同的平衡,从而帮助团队将大型模型保留给最需要它们进行复杂推理的场景。

Microsoft-Decision-1.1 is now generally available. Built on Gemma, it delivers fast, cost-effective decisions across routing, classification, prioritization and verification, and now supports image inputs. Find it today in the Foundry model catalog.

Microsoft-Decision-1.1 现已全面可用。该模型基于 Gemma 构建,可在路由、分类、优先级排序和验证等环节提供快速且具成本效益的决策,并新增了对图像输入的支持。请立即在 Foundry 模型目录中查找它。

Creative media generation

创意媒体生成

Creative teams need to explore ideas quickly without losing the visual consistency that makes their brand recognizable. MAI-Image-2.7 and MAI-Image-2.7-Flash bring sharper editing and more consistent image generation, helping brands keep the same visual identity from the first asset to the last. Both are now in public preview in Foundry – find them in the Foundry model catalog.

创意团队需要快速探索创意,同时保持品牌辨识度所需的视觉一致性。MAI-Image-2.7 和 MAI-Image-2.7-Flash 带来了更精准的编辑功能和更一致的图像生成效果,帮助品牌从首个素材到最后一个素材始终保持统一的视觉形象。这两款模型目前已在 Foundry 中开启公开预览——请在 Foundry 模型目录中查找它们。

Through our new partnership with fal on Microsoft Foundry, we’re also bringing Kling v3 Pro, MiniMax H3 Max, MiniMax H3 Max Turbo, LTX-2.5 Pro and LTX-2.5 Fast to Foundry Models, with more to come. Developers get more options for multimodal applications that combine video, images and other AI capabilities. Find all of them in the Foundry model catalog.

通过我们与 fal 在 Microsoft Foundry 上的新合作伙伴关系,我们还引入了 Kling v3 Pro、MiniMax H3 Max、MiniMax H3 Max Turbo、LTX-2.5 Pro 和 LTX-2.5 Fast 至 Foundry Models,未来还将有更多模型加入。开发者将获得更多选择,用于构建结合视频、图像及其他 AI 能力的多模态应用。所有模型均可在 Foundry 模型目录中找到。

Cybersecurity and vulnerability remediation

网络安全与漏洞修复

Security teams need to move from identifying a potential vulnerability to confirming the risk and fixing the code. They need tools that support that full workflow while keeping the time and cost of each investigation manageable.

安全团队需要从识别潜在漏洞转向确认风险并修复代码。他们需要能够支持整个工作流程的工具,同时确保每次调查所耗费的时间和成本处于可控范围。

MAI-Cyber-2-Flash is purpose-built for cybersecurity. Its new capabilities span code vulnerability discovery, validation and remediation. Because it was trained inside the Codename MDASH harness, it helps security teams find, confirm and fix vulnerabilities faster and more cost-effectively. It will be available through Codename MDASH.

MAI-Cyber-2-Flash 专为网络安全打造。其新功能涵盖代码漏洞的发现、验证与修复。由于它在 Codename MDASH 环境中训练而成,它能帮助安全团队更快、更具成本效益地发现、确认并修复漏洞。该模型将通过 Codename MDASH 提供。

Frontier and efficient models from Meta

来自 Meta 的前沿与高效模型

Meta’s Muse Wonder, Muse Spark 1.3 and Muse Ember are coming soon to Microsoft Foundry. They give teams frontier, balanced and high-efficiency options to match each step of an agentic workflow, from long-horizon planning and complex coding to fast, cost-efficient execution at scale. Find them in the Foundry model catalog.

Meta 的 Muse Wonder、Muse Spark 1.3 和 Muse Ember 即将登陆 Microsoft Foundry。它们为团队提供了前沿、均衡和高效率的选项,以匹配智能体工作流的每个步骤,从长期规划和复杂编码,到大规模下的快速、高成本效益执行。请在 Foundry 模型目录中查找它们。

Make it better: Let Foundry choose the model for each request

优化体验:让 Foundry 为每次请求自动选择最合适的模型

Not every request needs the most capable model. For teams that would rather focus on building than pick a model for every task, model router in Foundry Models matches each prompt to the right model in real time, so simple requests use smaller, faster models and complex ones get frontier reasoning. Today we’re making intelligent selection easier to adopt:

并非每个请求都需要最强大的模型。对于更专注于构建而非为每项任务挑选模型的团队,Foundry Models 中的模型路由器会实时将每个提示词匹配到合适的模型,从而使简单请求使用更小、更快的模型,而复杂请求则获得前沿推理能力。今天,我们让智能选择更容易采用:

  • Automatic model selection by default. In supported Foundry projects, automatic model selection is on by default, so developers can start building without first deploying the model router.
  • More models. Model router now chooses from 44 models from OpenAI, Anthropic, xAI, DeepSeek, Meta and others.
  • Model router in AI Gateway. AI Gateway customers can turn on model router to add intelligent model selection to their existing setup.
  • 默认自动选择模型。在支持的 Foundry 项目中,默认开启自动模型选择,因此开发人员无需先部署模型路由器即可开始构建。
  • 更多模型。模型路由器现在可从 OpenAI、Anthropic、xAI、DeepSeek、Meta 等公司的 44 款模型中进行选择。
  • AI Gateway 中的模型路由器。AI Gateway 客户可以启用模型路由器,为其现有设置添加智能模型选择功能。

Make it scale: Run and improve models on your terms

实现规模化:按您的条件运行和改进模型

Choosing the model is just the start. Getting value in production means running models efficiently and improving them over time against your business goals.

选择模型只是起点。在生产环境中创造价值意味着高效运行模型,并根据业务目标随时间推移持续改进它们。

More choices for running open models: Baseten joins Microsoft Foundry

运行开源模型的更多选择:Baseten 加入 Microsoft Foundry

Open models are moving fast, and Foundry is keeping pace. Baseten on Microsoft Foundry, now in public preview, brings Baseten-optimized open models such as Kimi-K3, GLM 5.3, GLM-5.3-Flash and DeepSeek V4.1-Flash-1017 into Foundry as a Microsoft first-party service, tuned for speed, throughput and cost. Start with serverless pay-as-you-go, globally or in the US Data Zone, and scale with provisioned throughput starting in December. Baseten joins Fireworks, already generally available on Foundry, so you can evaluate, deploy and govern open models that fit each workload in the same Foundry workflow you use for agents, all on your existing Azure agreement with Microsoft support.

开源模型发展迅速,Foundry 正与之同步。目前处于公共预览阶段的 Baseten on Microsoft Foundry,将 Kimi-K3、GLM 5.3、GLM-5.3-Flash 和 DeepSeek V4.1-Flash-1017 等经 Baseten 优化的开源模型作为微软第一方服务引入 Foundry,针对速度、吞吐量和成本进行了调优。可从全球或美国数据区的按需付费无服务器模式起步,并于 12 月起通过预置吞吐量进行扩展。Baseten 加入了已在 Foundry 上全面可用的 Fireworks,使您能够在用于代理的同一 Foundry 工作流中评估、部署和管理适合每种工作负载的开源模型,所有操作均在您现有的与微软的合作协议框架下进行,并获得微软支持。

Train your models in Microsoft Foundry

在 Microsoft Foundry 中训练您的模型

Anyone can call a general-purpose model. Your data and expertise are your edge, and training in Foundry turns them into a model built for your business, with as much control as you want.

任何人都可以调用通用模型。您的数据和专业知识是您的竞争优势,而在 Foundry 中进行训练可将其转化为专为您的业务打造的模型,并提供您所需的控制力。

With Managed Training (previously managed fine-tuning), bring your data and choose a method such as supervised or reinforcement fine-tuning, and Foundry runs the job. Want to own the loop? Programmed Training, now in public preview, lets you write your training recipe in Python, from sampling to scoring to weight updates, while Foundry runs it on serverless GPUs. Start with a cookbook of ready-made recipes and open-weight models such as gpt-oss-120b and Qwen3.8-27B. There are no GPU clusters to manage, and trained models can be evaluated and deployed in the same Foundry workflow you use for agents.

通过托管训练(原称托管微调),您可以提供数据并选择监督微调或强化学习微调等方法,Foundry 将自动运行任务。希望完全掌控流程?编程式训练现已进入公开预览阶段,允许您使用 Python 编写从采样、评分到权重更新的全流程训练配方,而 Foundry 将在无服务器 GPU 上执行该任务。您可以从现成的配方库以及 gpt-oss-120b 和 Qwen3.8-27B 等开源权重模型开始。无需管理 GPU 集群,且训练完成的模型可在您用于智能体的同一 Foundry 工作流中进行评估和部署。

Customers are already seeing results. Heidi, an AI care partner for clinicians, used Programmed Training to post-train an open model for clinical notes that matches a widely used frontier model on its clinical quality metrics, at an estimated 87% lower cost per note. Docusign cut AI processing costs by 90% with fine-tuned models in Foundry.

客户已初见成效。Heidi 是一款面向临床医生的 AI 护理助手,它利用编程式训练对开源模型进行后训练,使其在临床质量指标上媲美广泛使用的领先模型,同时每份病历的处理成本预计降低 87%。Docusign 通过在 Foundry 中使用微调模型,将 AI 处理成本降低了 90%。

Run open and custom models on dedicated infrastructure: Managed Compute is now generally available

在专用基础设施上运行开源和自定义模型:托管计算服务现已正式全面可用

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

进入量化体系 →