Meta releases Muse Ember, a new lightweight model for agent execution

Meta Superintelligence Lab on Monday released Muse Ember, describing it as the lab's lightest and most cost-efficient model to date. The company positioned it specifically for what agents handle in bulk: tool calls, coding subtasks, routing, extraction, and long document parsing. Rather than competing for frontier reasoning tasks, Ember is aimed at complex reasoning execution — the repetitive layer of agentic workloads where per-token price compounds quickly.

Ember is built to run alongside a larger model inside an agentic harness — in Meta's own example, Muse Wonder. The company illustrated the split with a customer-support scenario: a heavier model sets the rules for handling a given issue, then a lightweight model applies those rules across tens of thousands of cases. "Muse Wonder plans, Muse Ember handles the volume," the announcement reads.

Ember was also trained across text, images, and audio, which Meta says lets a single model handle screenshots, charts, and scanned documents as well as voice memos, meeting recordings, and video — without routing to a separate specialist model. That breadth matters for the volume-heavy use cases the model targets, where maintaining parallel models for each input type would erode the cost advantage Ember is designed to deliver.

Ember becomes the fourth release under the Muse brand, following the mid-weight Spark in April, Spark's open-source twin Glimmer in August, and the Wonder frontier model in October. It arrives behind the flagship consumer product Muse Agent, launched in September, which Meta said found unexpected success and which gained Spaces and video chat earlier this month. Open questions going forward: whether the Haiku and Luna cost claims hold up outside Artificial Analysis's composite scoring, whether Ember gets an open-source twin like Glimmer did, and how the Wonder-Ember pairing performs once builders put it through real production traffic.