Meta 发布首个前沿模型 Muse Wonder

Meta 超级智能实验室周一发布了一对新的业界领先模型：Muse Wonder 和 Muse Wonder Full。这是该公司首次尝试推出前沿模型。

与旨在以低延迟运行的轻量级模型 Muse Spark 不同，Wonder 系列模型规模更大，据称能力也更强，拥有更高的基准测试分数，并能在编程等复杂任务中取得更好的结果。

根据 Meta 的测试结果（见模型卡），Wonder 和 Wonder Full 的基准测试成绩可与 Fable 5.1 和 GPT-6 Astra 相当。

“它们共同标志着我们规模化能力阶梯上的下一步，也是迈向个人超级智能的一次重大飞跃。”Meta 在介绍这些模型的博客文章中写道。

这一规模也反映在定价上：通过 Meta Model API 访问时，Wonder 的价格为每百万输入 token 4 美元、每百万输出 token 20 美元，成本大致相当于 GPT-6 Sol 或 Opus 5.5。

值得注意的是，Wonder Full 将仅向 Muse Code 订阅用户提供，不会通过 API 提供服务。这两款模型也都将用于驱动 Muse agent。

延续 Meta 对开发能够长期执行任务的个人助手的兴趣，该公司尤其强调新模型保留和综合用户所提供信息的能力，Meta 将这一特性称为“主动记忆”（active memory）。

为实现更好的信息回忆能力，Meta 着重提到了MemoryBench 基准测试。该测试旨在挑战持续学习和记忆构建框架。根据该公司的测试，Muse Wonder 目前在 MemoryBench 上的得分高于 Fable、Opus 和 Astra。

“我们投入了大量精力，以确保你的模型能够在数天、数周、数月乃至更长时间里，记住对你重要的事情。”研究科学家 Kavya Kopparapu 在随发布一同推出的视频中表示。

这是自 Scale AI 创始人 Alexandr Wang 于去年 8 月重组 Meta 的 AI 工作以来，该公司首次发布前沿模型。此后，Meta 发布了轻量级模型 Muse Spark 和Muse agent，但一直避免推出 OpenAI 和 Anthropic 所擅长的那类业界领先模型。

根据文章，这些模型在 Meta 位于俄亥俄州中部的 Prometheus 数据中心完成训练；该数据中心于今年早些时候上线。

Meta’s Superintelligence Lab released a pair of new state-of-the-art models on Monday, Muse Wonder and Muse Wonder Full, representing the company’s first attempt at a frontier model.

Where Muse Spark aimed to be a lightweight model that could perform at low latency, the Wonder models are larger and said to be more capable, with higher benchmark scores and better results on complex tasks like coding.

According to Meta’s testing (presented in the model card), Wonder and Wonder Full achieve comparable benchmark scores to Fable 5.1 and GPT-6 Astra.

“Together, they mark the next step on our scaling ladder and a major leap toward personal superintelligence,” Meta’s blog post about the models reads.

That scale is reflected in the pricing: Accessed through the Meta Model API, Wonder will cost $4 per million input tokens and $20 per million output tokens, making it roughly equivalent in cost to GPT-6 Sol or Opus 5.5.

Notably, Wonder Full will only be available to Muse Code subscribers and will not be served through the API. Both models will also be used to power the Muse agent.

In keeping with Meta’s interest in developing personal assistants that can perform tasks over a long period, the company placed particular emphasis on the new models’ ability to retain and synthesize information given by the user, a property Meta calls “active memory.”

To achieve better recall, Meta highlighted the MemoryBench benchmark, which challenges continual learning and memory construction frameworks. According to the company’s testing, Muse Wonder now scores higher on MemoryBench than Fable, Opus, or Astra.

“We put a tremendous amount of effort into making sure your model can remember what’s important to you over days, weeks, months, and beyond,” said research scientist Kavya Kopparapu in a video accompanying the launch.

The new models are the first frontier model release since Scale AI founder Alexandr Wang restructured Meta’s AI efforts in August of last year. Since then, Meta has released its lightweight Muse Spark model and Muse agent but has avoided the kind of state-of-the-art model releases OpenAI and Anthropic are known for.

According to the post, the models were trained at Meta’s Prometheus data center in central Ohio, which came online earlier this year.