Muse Wonder Puts Meta Back at the AI Frontier

When Meta CEO Mark Zuckerberg rebooted his company’s artificial intelligence efforts last year, it seemed to be a very expensive gamble. Meta’s latest AI models, Muse Wonder and Muse Wonder Full, seem to show that the bet is finally paying off by putting the company back at the frontier of AI development.

当 Meta 首席执行官马克·扎克伯格去年重启公司的人工智能业务时，这看起来是一场代价高昂的赌博。Meta 最新的 AI 模型 Muse Wonder 和 Muse Wonder Full 似乎表明，这一赌注终于得到了回报，使该公司重新回到了 AI 开发的前沿。

Muse Wonder Full boasts frontier capabilities, Meta says, meaning that it is among the most capable AI models in the world. “Together, they mark the next step on our scaling ladder and a major leap toward personal superintelligence,” the company said in a blog post today.

Meta 表示，Muse Wonder Full 拥有前沿能力，这意味着它是世界上最强大的 AI 模型之一。该公司在今天的一篇博客文章中称：“它们共同标志着我们扩展阶梯上的下一步，也是迈向个人超级智能的重大飞跃。”

The Muse Wonder models will be available through an API and through Muse Code, a tool for writing software with AI. Meta says the best version of the model will power features across its platforms, including its Muse personal agent.

Muse Wonder 模型将通过 API 和 Muse Code（一种利用 AI 编写软件的工具）提供。Meta 表示，该模型的最佳版本将为其平台上的各项功能提供支持，包括其 Muse 个人代理。

Zuckerberg overhauled his company’s AI research last year by sinking billions into the recently formed Meta Superintelligence Labs, and hiring the former Scale AI CEO Alexandr Wang to lead it. This effort has sometimes been rocky, with some researchers not sticking around for long.

扎克伯格去年对公司的 AI 研究进行了彻底改革，向新成立的 Meta 超级智能实验室投入了数十亿美元，并聘请了前 Scale AI 首席执行官亚历山大·王（Alexandr Wang）来领导该实验室。这一过程有时并不顺利，一些研究人员并没有长期留任。

Meta was once a trailblazer in AI. Its Llama series of models were the first frontier open-weight offerings from a big lab. Llama, which could be downloaded and run on one’s own hardware, was widely adopted by researchers and startups.

Meta 曾是 AI 领域的先驱。其 Llama 系列模型是大型实验室推出的首批前沿开放权重产品。Llama 可以下载并在个人硬件上运行，因此被研究人员和初创公司广泛采用。

But the company fell behind after delivering several disappointing versions of the model, and as Anthropic, OpenAI, and Google rapidly introduced more capable models.

但在发布了几个令人失望的模型版本，以及 Anthropic、OpenAI 和 Google 迅速推出更强大的模型后，该公司落后了。

Meta announced the first model from its new lab, Muse Spark, in April of this year. The company has also released a small open-weight model called Muse Glimmer, and says it plans to release larger open-weight models in the future. It did not disclose details of how new Muse Wonder models were built or trained.

Meta 于今年 4 月宣布了其新实验室的首个模型 Muse Spark。该公司还发布了一个名为 Muse Glimmer 的小型开放权重模型，并表示计划在未来发布更大的开放权重模型。它没有披露 Muse Wonder 新模型的构建或训练细节。

Muse Wonder was built to compete with the very best, however. Meta says the largest version of the new model outperforms Anthropic’s second-best model, Claude Fable 5.1, and ranks just below its best, Claude Opus 5.5, on the Artificial Analysis Intelligence Index, a widely used aggregated score of model capabilities.

然而，Muse Wonder 的构建初衷是为了与最顶尖的模型竞争。Meta 表示，该新模型的最大版本在“人工智能分析指数”（Artificial Analysis Intelligence Index，一个广泛使用的模型能力综合评分）上，表现优于 Anthropic 的次优模型 Claude Fable 5.1，仅次于其最强模型 Claude Opus 5.5。

Meta says the model outperforms OpenAI’s best model, GPT-6 Astra, and both Claude Fable 5.1 and Claude Opus 5.5 on some specialized benchmarks, including ones that measure a model’s ability to work with legal documents and to build software. Benchmarks are not always a reliable measure of real-world performance, but this is still a significant step forward for Meta’s models.

Meta 表示，该模型在一些专业基准测试中优于 OpenAI 的最强模型 GPT-6 Astra，以及 Claude Fable 5.1 和 Claude Opus 5.5，其中包括衡量模型处理法律文件和构建软件能力的测试。基准测试并不总是衡量现实世界性能的可靠指标，但这对于 Meta 的模型来说仍然是重大进步。

Muse Wonder also includes two novel features designed to help make agents more effective and less inclined to misbehave. The new features, called Active Memory and Active Focus, are designed to keep the model locked into a stated goal and prevent it from becoming confused. This can be a challenge when AI performs complex research or builds a large software project, for example, and can lead to expensive mistakes or compounded errors.

Muse Wonder 还包含两项新颖功能，旨在帮助代理更有效且不易出现异常行为。这些名为“主动记忆”（Active Memory）和“主动聚焦”（Active Focus）的新功能，旨在使模型锁定既定目标，防止其陷入混乱。例如，当 AI 进行复杂研究或构建大型软件项目时，这可能是一个挑战，并可能导致昂贵的错误或复合性失误。

Meta says it has put significant effort into training Muse Wonder to refuse to generate dangerous outputs, such as recipes for building bioweapons or help with hacking. It says the company scores well on several benchmarks designed to measure how resilient models are to those prompts.

Meta 表示，它在训练 Muse Wonder 拒绝生成危险输出（如制造生物武器的配方或提供黑客攻击帮助）方面投入了大量精力。该公司称，在旨在衡量模型对这些提示的抵御能力的几项基准测试中，其得分表现良好。

A series of recent incidents involving rogue AI agents from Anthropic, OpenAI, and others also shows how easily powerful AI models can stray when given open-ended commands. Meta did not disclose whether it has seen agents powered by Muse Wonder misbehave in testing or what it has done to mitigate this issue specifically.

最近发生的一系列涉及 Anthropic、OpenAI 等公司“流氓”AI 代理的事件也表明，当被赋予开放式指令时，强大的 AI 模型很容易偏离轨道。Meta 没有披露在测试中是否观察到由 Muse Wonder 驱动的代理出现异常行为，也没有说明其为缓解这一问题采取了哪些具体措施。

Zuckerberg said in a post on X last month that the company had delayed the release of its Muse models to work on safety. Meta also revealed a framework designed to ensure that it scales its models’ capabilities safely in a blog post earlier this month.

扎克伯格上个月在 X 上发帖称，公司推迟了 Muse 模型的发布，以致力于安全性工作。Meta 在本月初的一篇博客文章中还披露了一个旨在确保其安全扩展模型能力的框架。

Muse Wonder may add more fuel to an already furious race to develop more capable AI, however. It certainly suggests that Meta is once again competitive in the increasingly high-stakes AI race.

然而，Muse Wonder 可能会为本已激烈的 AI 开发竞赛火上浇油。这无疑表明，Meta 在这场竞争日益激烈、赌注不断加码的 AI 竞赛中再次具备了竞争力。