Muse Wonder

本文为原文前 6,000 字符的节选翻译，完整内容请查看原文。

Muse Wonder

A frontier model that leads on personal and professional agents

今天，我们很高兴推出 Muse Wonder 系列，包括 Muse Wonder 和 Muse Wonder Full。这是我们能力最强的模型，在个人和专业代理任务中提供了最先进的性能。它们共同标志着我们在扩展阶梯上迈出了下一步，也是迈向个人超级智能的重要飞跃。Muse Wonder 结合了“主动记忆”（Active Memory）和“主动专注”（Active Focus），能够追踪对您重要的事情，并为实现您的目标而勤奋工作。

Muse Wonder is available today through the Meta Model API and Muse Code, and Muse Wonder Full is available today for Muse Code subscribers.

Muse Wonder 即将为 Muse 中的体验提供支持，并将首先向订阅者推出。Muse Wonder 还将支持工件生成以及需要更高智能的流程。未来，我们将使用 Muse Wonder 开发专门的模型，并将其提供给所有使用 Muse 的用户。

Fourteen months ago, we began a ground-up overhaul of our artificial intelligence efforts. In that time, we’ve gone from rebuilding our stack from first principles to pushing the frontier.

十四个月前，我们开始对人工智能工作进行彻底的全面改革。在此期间，我们从基于基本原则重建技术栈，发展到推动行业前沿。

When we introduced Muse Spark, we outlined our approach to predictable, rigorous scaling. Today’s release builds on that work across pretraining and reinforcement learning, delivering our most capable models yet. Trained at our gigawatt-scale Prometheus data center, Muse Wonder and Muse Wonder Full bring together the work of teams across Meta to deliver frontier capabilities for complex tasks at work and in everyday life — another step toward bringing personal superintelligence to the billions of people who use our products.

当我们推出 Muse Spark 时，我们概述了可预测且严谨的扩展方法。今天的发布建立在预训练和强化学习工作的基础上，交付了我们迄今为止能力最强的模型。Muse Wonder 和 Muse Wonder Full 在我们吉瓦级的 Prometheus 数据中心进行训练，汇集了 Meta 各团队的成果，为工作和日常生活中的复杂任务提供了前沿能力——这是向使用我们产品的数十亿人提供个人超级智能的又一步。

The Muse Wonder series achieves state-of-the-art results across key benchmarks for personal and professional agents, memory, and coding.

Muse Wonder 系列在个人和专业代理、记忆以及编码的关键基准测试中取得了最先进的成果。

The full results below compare Muse Wonder and Muse Wonder Full with other frontier models across a broad evaluation suite.

以下完整结果对比了 Muse Wonder、Muse Wonder Full 与其他前沿模型在广泛评估套件中的表现。

For more details about our evaluations, see our report.

有关我们评估的更多详细信息，请参阅我们的报告。

The results above provide a broad view of the models’ performance. We now take a closer look at four areas central to their work on real-world tasks: domain expertise, Active Memory, Active Focus, and frontier coding.

以上结果提供了模型性能的概览。现在，我们深入探讨其在现实任务中至关重要的四个领域：领域专业知识、主动记忆、主动专注和前沿编码。

Muse Wonder and Muse Wonder Full apply specialized knowledge to problems people encounter at work and in daily life, from understanding a lease to making sense of financial or health information. Our evaluations include tasks in law, finance, and health that test this ability. Harvey’s Legal Agent Benchmark (LAB), for example, requires models to work with source documents and follow detailed instructions to produce legal work products.

Muse Wonder 和 Muse Wonder Full 将专业知识应用于人们在工作和日常生活中遇到的问题，从理解租赁合同到解读财务或健康信息。我们的评估包括法律、金融和健康领域的任务，以测试这种能力。例如，Harvey 的法律代理基准（LAB）要求模型处理源文档并遵循详细说明，以产出法律工作成果。

One of the most critical steps towards useful personal and professional agents is reliable memory. Models need to remember what matters and recognize when earlier information becomes relevant. We call this Active Memory: carrying important context about you across conversations and actively retrieving the details relevant to the task at hand.

实现实用个人和专业代理最关键的步骤之一是可靠的记忆。模型需要记住重要事项，并识别早期信息何时变得相关。我们称之为“主动记忆”：在对话中携带关于您的重要背景信息，并主动检索与当前任务相关的细节。

Using that context to complete a task also requires sustained effort. Active Focus describes how Muse Wonder continues working toward a challenging goal: trying alternative approaches, working through blockers, and adapting to new information. Along the way, it asks for clarification when needed and checks in before taking important actions. We evaluate focus through FocusBench and sustained execution through SWE-Marathon, which tests work on long-running software projects.

利用这些背景信息完成任务还需要持续的努力。“主动专注”描述了 Muse Wonder 如何持续致力于实现具有挑战性的目标：尝试替代方案、克服障碍并适应新信息。在此过程中，它会在需要时寻求澄清，并在采取重要行动前进行确认。我们通过 FocusBench 评估专注力，并通过 SWE-Marathon 评估持续执行能力，后者测试了在长期软件项目上的工作表现。

Coding brings these capabilities together. Building software requires tracking requirements, making changes across a codebase, and working through repeated cycles of testing and revision. Muse Wonder and Muse Wonder Full deliver frontier coding performance, supporting projects from small applications to complex software. We trained both models on diverse, long-running tasks to help them sustain progress across this process.

编码将这些能力整合在一起。构建软件需要跟踪需求、在代码库中进行更改，并经历反复的测试和修订周期。Muse Wonder 和 Muse Wonder Full 提供了前沿的编码性能，支持从小型应用程序到复杂软件的各类项目。我们在多样化、长期的任务上训练了这两个模型，以帮助它们在整个过程中保持进展。

Muse Wonder builds a playable 3D kitchen game from customizable image inputs.

Muse Wonder 通过可定制的图像输入构建了一个可玩的 3D 厨房游戏。

Two models often have similar benchmark scores, yet one seems to handle a much wider range of unfamiliar tasks. That’s what complaints of “benchmaxxing” and praise for “big model smell” both get at: a breadth of capability that popular benchmarks in a few domains can’t fully capture.

两个模型通常具有相似的基准测试分数，但其中一个似乎能处理更广泛的陌生任务。这就是对“基准测试刷分”（benchmaxxing）的抱怨和对“大模型气息”（big model smell）的赞美所指向的：一种在少数领域流行的基准测试无法完全捕捉到的能力广度。

To quantify this mismatch, we investigate how performance on tasks targeted during post-training relates to performance on unfamiliar tasks as we scale test-time compute. The plot below compares Muse Spark 1.3 and Muse Wonder Full, with each point denoting performance at a distinct reasoning effort. The horizontal axis shows the model’s score on a composite representative of the tasks we focused on during post-training (“in-distribution”). The vertical axis shows the average score on a set of tasks we did not target during post-training (“out-of-distribution”).

为了量化这种差异，我们研究了随着测试时计算量的扩展，训练后针对的任务表现与陌生任务表现之间的关系。下图比较了 Muse Spark 1.3 和 Muse Wonder Full，每个点代表不同推理努力下的性能。横轴显示模型在训练后重点关注的任务组合（“分布内”）上的得分。纵轴显示模型在训练后未针对的任务集（“分布外”）上的平均得分。

Both of our models improve similarly on out-of-distribution tasks as we scale test-time compute. But given its in-distribution performance, Muse Wonder Full performs even better on unfamiliar tasks than the Muse Spark 1.3 trend would predict. This upward shift is because Muse Wonder and Muse Wonder Full build on larger, stronger base models: a stronger base yields post-training gains that carry over more effectively to unfamiliar tasks. That versatility especially matters for personal agents, which need to handle many combinations of goals, tools, and constraints — not all of which we can anticipate during training.

随着测试时计算量的扩展，我们的两个模型在分布外任务上的表现都有类似的提升。但考虑到其分布内性能，Muse Wonder Full 在陌生任务上的表现甚至超过了 Muse Spark 1.3 趋势的预测。这种向上偏移是因为 Muse Wonder 和 Muse Wonder Full 构建在更大、更强的基础模型之上：更强的基础带来了训练后的收益，这些收益能更有效地迁移到陌生任务中。这种多功能性对于个人代理尤为重要，因为它们需要处理目标、工具和约束的多种组合——而我们无法在训练期间预见所有这些情况。

Working directly in a browser and across apps, Muse Wonder builds a running plan (Map data © OpenStreetMap contributors) based on personal preferences and circumstances.

Muse Wonder designs and orders custom swag based on real photos taken from a personal Instagram profile.

Muse Wonder 根据从个人 Instagram 个人资料中获取的真实照片设计并订购定制周边。

Muse Wonder prepares a custom wedding seating chart based on a list of complex matching requirements.

Muse Wonder 根据一系列复杂的匹配要求准备定制的婚礼座位表。

Delivering personal superintelligence to everyone requires a significant investment in alignment and safety. A personal agent tak

为每个人提供个人超级智能需要对对齐和安全进行重大投入。一个个人代理任务