Meta发布Muse Wonder AI模型，具备主动记忆功能
Meta Platforms于10月13日宣布推出其最新前沿人工智能模型Muse Wonder系列，该系列在AI智能体独立执行能力方面有重大改进，并具备主动记忆功能。
原文深度解读
本文报道Meta Platforms于10月13日发布其最新前沿人工智能模型Muse Wonder系列，该系列分为Muse Wonder和Muse Wonder Full两个版本，均仅通过每月50美元或200美元的付费订阅提供。Muse Wonder可通过SDK或按量付费API供外部开发者使用，Muse Wonder Full则仅限Meta内部专用环境，更适合长周期复杂任务。文章重点介绍了Active Memory和Active Focus两项功能，并引用Artificial Analysis的评测数据说明其在多项基准测试中的表现、token效率与定价。该文章与电子元器件供应链的关联在于：前沿AI模型训练与推理依赖大规模数据中心算力，可能间接影响AI加速芯片、高带宽内存、服务器及网络设备等上游元器件的需求结构与供给节奏。
- Meta于10月13日发布Muse Wonder系列，包含Muse Wonder和Muse Wonder Full两个版本，二者区别在于是否向外部开发者开放以及是否用于Meta自有工作环境，且均仅通过每月50美元或200美元的付费订阅提供。
- Muse Wonder可通过SDK或按量付费API直接供开发者访问并集成或定制；Muse Wonder Full没有面向外部的独立API或SDK，仅限Meta专用环境，更适合涉及复杂、延长工作序列的长周期任务。
- Meta将高性能AI模型拆分为外部开发者版本和自用服务版本，文章认为该策略旨在通过直接销售模型创收，同时强化其专有AI生产力工具，并在企业AI市场与OpenAI、Anthropic PBC和Google LLC竞争；Meta此前主要通过开源模型家族Llama扩展生态。
- Muse Wonder系列的核心功能为Active Memory和Active Focus：前者主动保留此前对话信息和重要用户上下文并在需要时检索以保持任务连续性，后者在困难任务遇到障碍时尝试替代方法以帮助模型完成任务。
- Artificial Analysis的评测显示，Muse Wonder Full在CritPt科学推理测试和GDP.pdf专业文档工作基准中排名第一，Terminal-Bench 4.0得分从上一版本的33%大幅升至57%；在Intelligence Index最高推理设置下，Muse Wonder得52分、Muse Wonder Full得56分，后者与Anthropic的Claude Sonnet 5.5并列第二，且较六周前发布的Muse Spark 1.3的48分显著提升。
- AA称Meta两款新模型输出token用量比Muse Spark少约10%，Muse Wonder Full输入token用量比Muse Wonder少16%，因而更智能高效；Muse Wonder定价为每百万输入token 4美元、每百万输出token 20美元，处于市场较高水平。Meta表示两款模型均在其吉瓦级Prometheus数据中心通过预训练和强化学习开发，并在编码、复杂长周期任务执行和记忆管理等关键AI智能体基准上达到SOTA，同时加强了对抗性提示攻击防御以及针对网络攻击和CBRN威胁的防滥用保障。
供应链影响
- Meta在吉瓦级Prometheus数据中心开发并运行Muse Wonder系列，可能意味着其对AI服务器、加速芯片、高带宽内存及数据中心网络设备存在持续需求，但具体采购规模取决于模型训练与推理的实际部署节奏。
- Muse Wonder Full仅限Meta专用环境且不提供外部API或SDK，可能使相关推理负载集中在Meta自有数据中心，从而影响其对自建算力基础设施与配套元器件的需求结构。
- 两款模型输出token用量较Muse Spark减少约10%、Muse Wonder Full输入token用量较Muse Wonder减少16%，若被广泛采用，可能降低单位任务所需推理算力，进而间接影响AI加速芯片与内存的需求增速，但实际影响取决于用户调用量与部署规模。
- Muse Wonder通过SDK和按量付费API向外部开发者开放，可能扩大第三方应用集成，从而带动分散式推理算力需求及相关服务器、网络与存储元器件的需求，但取决于开发者采用速度。
- Meta在企业AI市场与OpenAI、Anthropic和Google竞争并推出付费订阅，可能推动行业整体推理算力投入，从而对上游电子元器件供应链形成潜在拉动，但具体幅度取决于各厂商模型部署与商业化进展。
系统解读，仅作为判断线索，不构成备货、出货、涨价或投资建议。
Meta Platforms于10月13日宣布，已推出其最新前沿人工智能（AI）模型Muse Wonder系列，该系列在AI智能体独立处理复杂任务的能力方面有重大改进。
该系列包含两个版本：Muse Wonder和Muse Wonder Full。两款模型的区别在于是否向外部开发者开放以及是否在Meta自有工作环境中使用。二者均仅通过每月50美元或200美元的付费订阅提供。
Muse Wonder可由开发者通过软件开发工具包（SDK）或按量付费应用程序编程接口（API）直接访问，使其能够将该模型集成到应用中或针对各种工作流进行定制。相比之下，Muse Wonder Full没有面向外部访问的独立API或SDK，仅限Meta专用环境使用。它也更适合涉及复杂、延长工作序列的长周期任务。
Meta将其高性能AI模型拆分为面向外部开发者的版本和自用服务版本，这一决定值得注意。该策略似乎旨在通过直接销售模型创收，同时强化其专有AI生产力工具。Meta正寻求在企业AI市场扩大变现机会，在该市场它与OpenAI、Anthropic PBC和Google LLC竞争。迄今为止，该公司主要通过其开源AI模型家族Llama扩展生态。
Muse Wonder系列的突出功能是Active Memory和Active Focus。Active Memory主动保留此前对话中的信息和重要用户上下文，并在需要时检索，以保持任务之间的连续性。Active Focus在困难任务遇到障碍时尝试替代方法，帮助模型完成任务并实现目标。
在独立AI评估机构Artificial Analysis（AA）的基准测试中，Muse Wonder Full在任务表现方面获得高分。它在科学推理测试CritPt和专业文档工作基准GDP.pdf中排名第一。其Terminal-Bench 4.0得分也从上一版本的33%大幅升至57%。
Meta Muse Wonder在Artificial Analysis智能指数上的得分。（来源：Artificial Analysis）
在衡量模型整体智能的AA智能指数上，Muse Wonder在最高推理设置下得分为52，Muse Wonder Full得分为56。Muse Wonder的52分使其接近Google近期发布的Gemini 4 Argon（得分为53）和OpenAI的GPT-6.1 Sol。Muse Wonder Full以56分与Anthropic的Claude Sonnet 5.5在评测时并列第二。
Meta Muse Wonder Full以56分在Artificial Analysis智能指数上并列第二。该模型不通过API提供，仅可通过Muse Code的付费月度订阅访问。（来源：Artificial Analysis）
这些得分也较六周前发布的上一代Muse Spark 1.3在最大推理设置下记录的48分有大幅提升。
AA表示，Meta的两款最新模型输出token用量比Muse Spark少约10%。Muse Wonder Full的输入token用量也比Muse Wonder少16%，据该评估机构称，这使其“更智能、更高效”。Muse Wonder定价为每百万输入token 4美元、每百万输出token 20美元，处于市场较高水平。
Meta Muse Wonder Full在Terminal-Bench 4.0上的基准表现，该测试衡量模型在没有图形用户界面的终端环境中完成软件和系统任务的能力。（来源：Artificial Analysis）
Muse Code是Meta由Muse Wonder Full驱动的编码工具，其表现与使用Sonnet 5.5的Claude Code相当。AA的编码智能指数综合了三项基准的结果：评估实际软件工程能力的DeepSWE v1.1、Terminal-Bench 4.0，以及衡量代码理解能力的SWE-Atlas-QnA。（来源：Artificial Analysis）
Meta表示，两款模型均在其吉瓦级Prometheus数据中心通过预训练和强化学习开发。该公司称，它们在关键AI智能体基准上达到了最先进（SOTA）性能，包括编码、执行复杂长周期任务和记忆管理。
Meta还强调了其最新AI模型中纳入的安全措施。该公司表示，已加强对对抗性提示攻击的防御，并实施了防止涉及网络攻击以及化学、生物、放射和核（CBRN）威胁滥用的保障措施。
Meta Platforms announced Oct. 13 that it has unveiled its latest frontier artificial intelligence (AI) models, the Muse Wonder series, featuring major improvements in AI agents' ability to independently handle complex tasks.
The series comprises two versions, Muse Wonder and Muse Wonder Full. The models are differentiated by their availability to external developers and their use within Meta's proprietary work environment. Both are available only through paid subscriptions priced at $50 or $200 a month.
Muse Wonder can be accessed directly by developers through a software development kit (SDK) or a pay-as-you-go application programming interface (API), allowing them to integrate the model into applications or customize it for various workflows. Muse Wonder Full, by contrast, has no standalone API or SDK for external access and is restricted to Meta's dedicated environment. It is also better suited to long-horizon tasks involving complex, extended sequences of work.
Meta's decision to separate its high-performance AI models into versions for external developers and its own services is notable. The strategy appears designed to generate revenue from direct model sales while strengthening its proprietary AI productivity tools. Meta is seeking to expand its monetization opportunities in the enterprise AI market, where it competes with OpenAI, Anthropic PBC, and Google LLC. The company has until now focused on expanding its ecosystem through Llama, its open AI model family.
The standout features of the Muse Wonder series are Active Memory and Active Focus. Active Memory proactively retains information from previous conversations and important user context, retrieving it when needed to maintain continuity across tasks. Active Focus attempts alternative approaches when obstacles arise during challenging assignments, helping the model complete tasks and achieve its objectives.
Muse Wonder Full received strong marks for task performance in benchmarks from Artificial Analysis (AA), an independent AI evaluation organization. It ranked first in CritPt, a scientific reasoning test, and GDP.pdf, a benchmark for professional document work. Its Terminal-Bench 4.0 score also rose sharply to 57% from 33% for the previous version.
Meta Muse Wonder scores on Artificial Analysis' Intelligence Index. (Source : Artificial Analysis)
On AA's Intelligence Index, which measures overall model intelligence, Muse Wonder scored 52 and Muse Wonder Full scored 56 at the highest reasoning setting. Muse Wonder's score of 52 puts it near Google's recently released Gemini 4 Argon, which scored 53, and OpenAI's GPT-6.1 Sol. Muse Wonder Full, with a score of 56, tied Anthropic's Claude Sonnet 5.5 for second place at the time of the evaluation.
Meta Muse Wonder Full ranked joint second with a score of 56 on Artificial Analysis' Intelligence Index. The model is not available through an API and can be accessed only through a paid monthly subscription to Muse Code. (Source : Artificial Analysis)
The scores also marked a substantial improvement from the 48 recorded by the previous Muse Spark 1.3 model at the maximum reasoning setting, released six weeks earlier.
AA said both of Meta's latest models use about 10% fewer output tokens than Muse Spark. Muse Wonder Full also uses 16% fewer input tokens than Muse Wonder, making it "smarter and more efficient," according to the evaluator. Muse Wonder is priced at $4 per million input tokens and $20 per million output tokens, putting it at the higher end of the market.
Meta Muse Wonder Full's benchmark performance on Terminal-Bench 4.0, which measures a model's ability to complete software and systems tasks in a terminal environment without a graphical user interface. (Source : Artificial Analysis)
Muse Code, Meta’s coding tool powered by Muse Wonder Full, demonstrated performance comparable to Claude Code, which uses Sonnet 5.5. AA’s Coding Intelligence Index combines results from three benchmarks: DeepSWE v1.1, which evaluates practical software engineering capabilities; Terminal-Bench 4.0; and SWE-Atlas-QnA, which measures code comprehension. (Source : Artificial Analysis)
Meta said both models were developed through pretraining and reinforcement learning at its gigawatt-scale Prometheus data center. The company said they achieved state-of-the-art (SOTA) performance across key AI agent benchmarks, including coding, execution of complex long-horizon tasks, and memory management.
Meta also highlighted safety measures incorporated into its latest AI models. The company said it had strengthened defenses against adversarial prompt attacks and implemented safeguards to prevent misuse involving cyberattacks and chemical, biological, radiological, and nuclear (CBRN) threats.