Meta Muse Spark Open 模型上线 Microsoft
Muse Spark Open is available in Microsoft Foundry Models
Building useful coding agents takes more than generating code. Agents need to analyze an existing codebase, work across files and dependencies, use tools reliably, and keep making progress as a task evolves.
构建实用的编码代理不仅需要生成代码。代理需要分析现有代码库，跨文件和依赖项工作，可靠地使用工具，并在任务演进过程中持续取得进展。
Muse Spark Open, a coding-optimized open-weight model from Meta, is available today in Microsoft Foundry Models through Hugging Face collection. Trained specifically for agentic coding workflows, Muse Spark Open combines a focus on first-attempt accuracy with reliable tool calling and deep codebase understanding, giving developers another option for building agents around the repositories they work in every day.
Muse Spark Open 是 Meta 推出的一款针对编码优化的开源权重模型，今日已通过 Hugging Face 集合在 Microsoft Foundry Models 中提供。该模型专为代理式编码工作流训练，将首次尝试的准确性、可靠的工具调用以及对深层代码库的理解相结合，为开发者围绕其日常使用的仓库构建代理提供了另一种选择。
Built for the way developers work
专为开发者的工作方式而构建
Real development work spans debugging sessions, multi-file refactors, and changes that reach across an entire project. Muse Spark Open is trained on these workflows, with capabilities across code generation, debugging, and codebase understanding.
实际的开发工作涵盖调试会话、多文件重构以及影响整个项目的变更。Muse Spark Open 针对这些工作流进行了训练，具备代码生成、调试和代码库理解等方面的能力。
A 1M-token context window allows developers to bring extensive project context into a task, including legacy code, complex dependencies, and large collections of files. That broader context helps the model connect a local change to the surrounding codebase and reduces the need to split work into disconnected pieces. For teams building coding agents, this supports workflows such as investigating a bug across services, understanding an unfamiliar repository, or coordinating changes across multiple components. Learn more about this model.
100万 token 的上下文窗口允许开发者将广泛的项目上下文带入任务中，包括遗留代码、复杂依赖项和大量文件集合。更广阔的上下文有助于模型将局部更改与周围代码库联系起来，并减少将工作拆分为不连贯部分的需求。对于构建编码代理的团队而言，这支持了诸如跨服务调查错误、理解陌生仓库或协调多个组件间变更等工作流。了解更多关于此模型的信息。
Designed to sustain progress across complex tasks
旨在在复杂任务中维持进展
Muse Spark Open is built to plan, act, observe results, and adjust its approach over extended workflows. Its training encompasses long-running tasks that require multiple steps, tools, and interactions across websites. With an appropriate agentic harness, Muse Spark Open can spawn and coordinate subagents for work that can be divided into separate tasks. It also supports web browsing, including multimodal browsing, and computer use across desktop, mobile, and browser environments. Separate commentary lets Muse Spark Open outputs a planned action before taking action, giving developers visibility into what the agent is doing and why as work progresses.
Muse Spark Open 被设计用于规划、行动、观察结果并在长期工作流中调整其方法。其训练涵盖了需要多步骤、工具和跨网站交互的长时间运行任务。借助适当的代理框架（agentic harness），Muse Spark Open 可以生成并协调子代理来处理可划分为独立任务的工作。它还支持网页浏览（包括多模态浏览）以及在桌面、移动和浏览器环境中的计算机使用。单独的评论功能让 Muse Spark Open 在行动之前输出计划的操作，使开发者能够在工作进展过程中了解代理正在做什么以及原因。
Understand more of the context behind the code
更深入地理解代码背后的上下文
Software development draws on more than source files. Requirements arrive in documents, interfaces appear in screenshots, and issues may be captured in recordings. Muse Spark Open combines image and document understanding with audio and audio-video perception. As a reasoning model, it can work across what it sees, hears, and reads, creating opportunities for agents to connect these inputs with the code and tasks they support.
软件开发不仅仅依赖源代码文件。需求以文档形式呈现，接口出现在截图中，问题可能被录制在音视频文件中。Muse Spark Open 结合了图像与文档理解能力，以及音频和音视频感知能力。作为推理模型，它能够跨模态处理所见、所听和所读的内容，为智能体将这些输入与其支持的代码和任务建立连接创造了机会。
Open weights, more developer choice
开放权重，赋予开发者更多选择
Muse Spark Open is shared under the permissive Muse Model License with narrow commercial terms. Its open weights give developers flexibility to pursue cloud and on-premises deployments, as well as fine-tuning for specialized needs, subject to the license.
Muse Spark Open 在宽松的美洲狮模型许可（Muse Model License）下发布，仅包含有限的商业条款。其开放权重使开发者能够灵活选择云端和本地部署方案，并根据许可证规定进行针对特定需求的微调。
Availability through Hugging Face in Microsoft Foundry brings that choice into Foundry Models. Hugging Face models use a managed compute deployment path with dedicated GPU-backed inference endpoints, helping teams move from model discovery to integration with their applications.
通过 Microsoft Foundry 中的 Hugging Face 提供该模型选项，使其纳入 Foundry Models。Hugging Face 模型采用托管计算部署路径，配备专用的 GPU 支持推理端点，帮助团队从模型发现快速过渡到将其集成到应用程序中。
Get started in Microsoft Foundry
在 Microsoft Foundry 中开始使用
Explore Muse Spark Open through the Hugging Face collection in Microsoft Foundry today and evaluate it against your own coding workflows. Start with the repositories, debugging tasks, and multi-file changes that matter to your team, then assess how the model performs within your agent’s tools and harness. Try it today Model Catalog | Microsoft Foundry.
今天即可通过 Microsoft Foundry 中的 Hugging Face 集合探索 Muse Spark Open，并结合您自身的编码工作流对其进行评估。从对您的团队至关重要的代码仓库、调试任务和多文件变更入手，然后评估该模型在其智能体工具和环境中的表现。立即试用：模型目录 | Microsoft Foundry。
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