Meta Superintelligence Lab released Muse Ember on November 9, 2026, to handle repetitive tasks for AI agents. Meta designed the model for tasks including tool calls, coding subtasks and processing long documents.
A smaller model for high-volume work
Ember is the lab’s lightest model to date. It is designed to work in an AI agent system alongside a larger model such as Muse Wonder, which can plan a task while Ember handles a large volume of individual cases.
In customer support, for example, a larger model could set rules for resolving an issue, with Ember then applying them to tens of thousands of cases. Meta also designed the model for routing and data extraction.
Text, images and meeting recordings
Meta trained Ember to work with text, images and audio. According to the company, agents can use it to read screenshots, charts and scanned documents, or work with voice notes, meeting recordings and video without a separate model.
Ember is the fourth model released under the Muse brand. It follows the mid-sized Spark, its open-source twin Glimmer and the powerful Wonder model.
Comparison with competing models
Using Artificial Analysis’s composite score, Meta says Ember outperforms Anthropic’s Haiku and OpenAI’s Luna. Input costs 10 cents per million tokens, and Meta says Ember is cheaper than Haiku and Luna while also being cost-competitive with the open-source DeepSeek V4.1 Flash and GLM 5.3 Flash models.
