11-19 更新 实时
首页/ 科技公司/ 正文

Meta 发布 Muse Ember 轻量级智能体执行模型

科技公司 来源:TechCrunch AI 2026-11-10 02:00 0 阅读 机器翻译
Meta 发布 Muse Ember 轻量级智能体执行模型

Meta Superintelligence Lab 于周一发布了一款名为 Muse Ember 的新 AI 模型,这是该实验室迄今最轻量、最具成本效益的 AI 模型。Meta 在发布中表示,Ember“专为复杂推理执行以及智能体整日重复的工作而设计:工具调用、编码子任务、路由、信息抽取和长文档解析”。

Meta 尤其强调了 Ember 的每任务成本,并借助 Artificial Analysis 的综合基准测试得分来表明,Ember 比 Anthropic 的 Haiku 和 OpenAI 的 Luna 等竞争对手更便宜、更智能。

在每百万输入 Token 10 美分的价格下,Ember 比 Haiku 和 Luna 等前沿竞争对手便宜得多。不仅如此,它在成本上还能与 DeepSeek V4.1 Flash 和 GLM 5.3 Flash 等开源竞争对手相抗衡。

鉴于其重点是大规模场景下的低成本与高准确率,Ember 的设计初衷是与 Meta 的 Muse Wonder 等更大模型搭配,组成智能体运行框架。以客服智能体为例,你可能会看到较重的模型为如何处理某个特定问题制定规则,然后再由较轻量的模型把规则应用到数万个案例中。正如 Meta 在公告中所说:“Muse Wonder 负责规划,Muse Ember 负责处理海量事务。”

考虑到用量庞大的使用场景,Meta 还训练 Ember 在文本、图像和音频等不同类型的媒体上都能良好工作。结果是,该模型在可处理的输入类型上出奇地灵活。正如 Meta 在公告中所说:“智能体可以读取截图、图表和扫描文档,也可以直接处理语音备忘、会议录音和视频,无需另用模型。”

Ember 是 Meta 旗下 Muse 品牌发布的第四款模型,此前依次为中等规模的 Spark(4 月发布)、Spark 的开源孪生模型 Glimmer(8 月发布),以及前沿模型 Wonder(10 月发布)。旗舰消费级 AI 产品 Muse Agent 于 9 月发布,取得了出乎意料的成功,Spaces 和视频聊天等新功能也于本月早些时候陆续上线。

当您通过我们文章中的链接购买时,我们可能会赚取少量佣金。这不会影响我们编辑的独立性。

Russell Brandom 自 2012 年起一直报道科技行业,重点关注平台政策与新兴技术。他此前曾在 The Verge 和 Rest of World 工作,并为 Wired、The Awl 和 MIT 的 Technology Review 撰稿。可通过 [email protected] 或 Signal 412-401-5489 与他联系。

正在加载下一篇文章

译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。

查看英文原文(Meta releases Muse Ember, a new lightweight model for agent execution)

Meta Superintelligence Lab released a new AI model on Monday called Muse Ember, the lab’s lightest and most cost-efficient AI model yet. In its release, Meta said Ember was “designed for complex reasoning execution and the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing.”

Meta particularly emphasized Ember’s cost-per-task, using Artificial Analysis’s composite benchmark score to show Ember as cheaper and smarter than competitors like Anthropic’s Haiku and OpenAI’s Luna.

Image Credits:Meta Superintelligence Labs

At 10 cents per million input tokens, Ember is significantly cheaper than frontier competitors like Haiku and Luna. Not only that, it’s cost-competitive with open-source competitors like DeepSeek V4.1 Flash and GLM 5.3 Flash.

Given the emphasis on low cost and accuracy at scale, Ember is designed to pair with a larger model like Meta’s Muse Wonder in an agentic harness. In the case of a customer-support agent, for instance, you might see a heavier model establishing rules for how to handle a particular issue, then sending a lighter-weight model to apply the rules across tens of thousands of cases. As Meta puts it in the announcement, “Muse Wonder plans, Muse Ember handles the volume.”

With volume-heavy use cases in mind, Meta also trained Ember to work well across different types of media, including text, images, and audio. The result makes the model surprisingly flexible in what kind of inputs it can handle. As Meta puts it in the announcement, “agents can read screenshots, charts, and scanned documents, or work from voice memos, meeting recordings, and video without a separate model.”

Ember is the fourth model released under Meta’s Muse brand, following the mid-weight Spark (released in April), Spark’s open-source twin Glimmer (released in August), and the Wonder frontier model (released in October). The flagship consumer AI product, the Muse Agent, was released in September to unexpected success, with new features like Spaces and video chat rolling out earlier this month.

When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.

Russell Brandom

Russell Brandom has been covering the tech industry since 2012, with a focus on platform policy and emerging technologies. He previously worked at The Verge and Rest of World, and has written for Wired, The Awl and MIT’s Technology Review. He can be reached at [email protected] or on Signal at 412-401-5489.

Loading the next article

原文为英文,正文由自动抓取并保留原文。

本文由采集器自动同步

原文发布于 TechCrunch AI(Meta releases Muse Ember, a new lightweight model for agent execution),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。

阅读原文 ↗