Meta Muse Ember explained: the small Muse model built for the work agents repeat all day

Muse Ember is a small, cheap language model from Meta. Meta describes it as "our smallest and most efficient model", designed for "the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing."

It sits at the bottom of Meta's Muse family. Meta's own description of the family is: Muse Wonder for the hardest problems, Muse Spark for coding and agent workflows, and Muse Ember for repeatable coding and agent tasks at scale. Meta also says Ember is the first model it trained using Muse Wonder Full.

Ember reads text, images, audio and video natively, so an agent can look at a screenshot, a chart or a scanned page, or listen to a voice memo, without calling a second model.

Meta says Ember is available as muse-ember-1.0 in Muse Code and through the Meta Model API, which supports the Responses API format. Meta also lists the Muse models on Google Cloud and Oracle Cloud, and you can try Ember in the browser at dev.meta.ai.

The pattern Meta suggests is to pair it with a bigger model. The large model reads the problem and makes a plan; Ember carries out the many small steps, such as calling a tool, pulling fields out of a document or deciding where a request should go. Most of the tokens in an agent are spent on those small steps, so that is where a cheap model saves the most.

Read next: Model routing and cascades: sending each request to the right model

Small models are where the cost of an agent is decided. An agent may make dozens of model calls for one user request, and if each call goes to a frontier model the bill grows with every step. A model that scores one point below Muse Spark 1.3 on the Intelligence Index, at about a seventeenth of the cost per task by Meta's numbers, changes which steps are worth sending to a big model at all.

Two cautions. Meta's benchmark table includes its own internal tests, and its chart notes that Ember's costs there are estimates at proposed pricing. Before you move real traffic, run Ember on a sample of your own tasks and compare the answers and the bill with what you use today.

Read next: The AI Engineering course: agents, routing and cost, from the ground up