According to the official announcement, Meta Super Intelligence Lab has launched Muse Ember, a lightweight model optimized for complex reasoning execution and high-frequency agent tasks, including tool invocation, coding subtasks, task routing, information extraction, and long-document parsing. Ember's input pricing is $0.1 per million tokens, with Meta stating its cost is lower than leading models such as Anthropic’s Haiku and OpenAI’s Luna, and it can compete with open-source models like DeepSeek V4.1 Flash and GLM 5.3 Flash.

The model can be used in conjunction with the larger model Muse Wonder, where Wonder handles planning and Ember executes tasks in bulk. Ember also supports text, image, and audio inputs, enabling direct processing of screenshots, charts, scanned documents, voice memos, meeting recordings, and videos without requiring additional models. Muse Ember is the fourth model released in the Meta Muse series, following the prior launches of Spark, the open-source model Glimmer, and the flagship model Wonder.