Meta超级智能实验室于周一推出了其迄今最轻量且最具成本效益的AI模型Muse Ember,该模型专为复杂推理执行及日常代理任务设计 1。Ember旨在处理工具调用、编码子任务、路由、信息提取以及长文档解析等高频重复性工作,强调在保持高准确性的同时实现低每任务成本 12。作为Meta旗下Muse品牌发布的第四个模型,它延续了此前Spark、Glimmer和Wonder的产品线 1。
在定价与性能方面,Muse Ember输入令牌价格为每百万10美分,输出令牌为每百万50美分,缓存输入仅为每百万0.4美分 2。Meta声称,与Anthropic的Claude Haiku 5.5相比,Ember能以约三分之一的成本匹配最佳分数;相较于OpenAI的Luna,其在所有设置下均表现更优 12。此外,该模型还与开源竞品DeepSeek V4.1 Flash和GLM 5.3 Flash具备成本竞争力 1。
技术规格上,Muse Ember提供100万令牌的上下文窗口,支持最高512K的输出令牌,并兼容Responses API 2。该模型原生支持文本、图像、音频及视频等多模态输入,能够端到端执行业务工作流 12。架构策略上,Meta将Ember定位为“处理数量”的角色,通常与负责规划的更大模型Muse Wonder配合使用 1。目前,该模型已上线Meta Model API,并在Google Cloud和Oracle Cloud提供服务 2。安全性评估显示,Ember在Meta超级智能扩展框架下低于高风险阈值,并具备抵抗隐藏指令等行为控制能力 2。
Meta Superintelligence Lab launched Muse Ember on Monday, describing it as the lab’s lightest and most cost-efficient AI model to date 12. The model is designed for complex reasoning execution and routine agentic tasks, such as tool calling, coding subtasks, routing, extraction, and long document parsing 12. As the fourth release under Meta’s Muse brand—following Spark, Glimmer, and Wonder—Ember aims to handle high-volume workloads with low per-task costs and high accuracy 1.
The company outlines an architectural strategy where the larger Muse Wonder model handles planning while Muse Ember manages volume, summarized by the phrase "Muse Wonder plans, Muse Ember handles the volume" 1. Pricing is set at $0.10 per million input tokens and $0.50 per million output tokens, with cached inputs costing $0.004 per million tokens 2. Meta claims Ember outperforms Anthropic’s Haiku and OpenAI’s Luna in both cost and performance, stating it matches Haiku’s best scores at approximately one-third of the cost and surpasses all Luna configurations 12. The model also remains cost-competitive against open-source alternatives like DeepSeek V4.1 Flash and GLM 5.3 Flash 1.
Technically, Muse Ember supports a 1-million-token context window and can generate up to 512,000 output tokens via the Responses API 2. It features native multimodal capabilities for text, image, audio, and video understanding, enabling end-to-end business workflow execution 2. Trained from Muse Wonder Full, the model is available through the Meta Model API and on Google Cloud and Oracle Cloud infrastructure 2. Security evaluations conducted under Meta’s superintelligence scaling framework indicate the model remains below high-risk thresholds, including resistance to hidden instruction behaviors 2.
评论
还没有评论,欢迎留下第一条。