Meta’s Muse Ember is a lightweight AI model aimed at work that agents need to repeat at scale. Meta says it can take on tasks such as tool calls, coding subtasks, routing, extraction and parsing long documents, with an emphasis on keeping the cost of each task low.
A smaller model for repeated tasks
Meta describes Ember as its lightest and most cost-efficient model so far. The company highlighted its cost per task using Artificial Analysis’s composite benchmark score, presenting Ember as cheaper and smarter than Anthropic’s Haiku and OpenAI’s Luna.
Ember costs 10 cents per million input tokens, according to the source article. That price is also described as competitive with open-source models DeepSeek V4.1 Flash and GLM 5.3 Flash.
The model’s intended role is narrower than handling every part of an agent’s work. A larger model can establish instructions for a situation, while Ember applies those instructions repeatedly. In a customer-support example, a larger model could set rules for handling an issue and Ember could apply them across tens of thousands of cases.
Meta summarized the arrangement this way: “Muse Wonder plans, Muse Ember handles the volume.” The pairing reflects a division of labor: the larger model sets direction, and the lightweight model takes on recurring execution.
Designed for different kinds of input
Meta says it trained Ember to work with text, images and audio, broadening the material an agent can use. The company says agents can read screenshots, charts and scanned documents, or work from voice memos, meeting recordings and video without a separate model.
That range could let an agent handle tasks that draw on several kinds of information. For example, a workflow could involve extracting information from a scanned document or interpreting a chart, as well as working with spoken recordings. The article does not describe specific deployments, but positions this flexibility as useful for volume-heavy use cases.
Combined with its low cost and focus on accuracy at scale, multimodal input is part of Ember’s pitch as an execution model. It is intended to carry out repeated work while fitting into an agent system that may also use a larger model.
The Muse lineup keeps growing
Ember is the fourth model under Meta’s Muse brand. It follows Spark, described as a mid-weight model and released in April; Glimmer, Spark’s open-source twin, released in August; and the Wonder frontier model, released in October.
The wider Muse offering also includes the Muse Agent, Meta’s flagship consumer AI product. It was released in September and had unexpected success, while features including Spaces and video chat rolled out earlier this month.
With Ember, Meta is adding a model aimed at the repetitive execution side of agent workflows. Its core proposition is that a larger model can plan how to handle a task, then a cheaper, lighter model can apply that plan across a large volume of work.