Meta released Muse Ember through its Model API for repeated work such as extraction, tool calls, and routing. An agent can use Ember for those steps while assigning the overall plan to Spark or Wonder.
Muse Models
Also known as: Meta Muse Models · Muse Spark · Muse Spark Open · Muse Glimmer · Muse Wonder · Muse Ember
Plain English
Muse is Meta’s family of models for assistants and agents. It includes models served through Meta’s apps and API, as well as downloadable releases developers can run on their own hardware.
Technical Overview
The family spans multimodal reasoning models such as Muse Spark and Wonder, the smaller Ember model for tool calls and subtasks, and open-weight variants such as Glimmer and Spark Open. Spark Open uses sparse expert routing. These models can power the Muse personal assistant, Muse Code, and third-party agents, depending on access and tool support.
Chronological Entries (6)
Meta released Muse Spark Open under the Muse Model License. The multimodal expert model has 816 billion parameters and activates 42 billion per token. Developers can download the weights or use endpoints from providers such as Baseten and Together AI.
Meta introduced Muse Wonder and Wonder Full for work that spans many tool calls and decisions. The public launch made Wonder available through the Meta Model API and Muse Code, with Full available to Muse Code subscribers.
Meta released Muse Glimmer under Apache 2.0. The 30-billion-parameter model was distilled from Muse Spark for local uses such as document analysis and coding.
Meta released Muse Spark 1.1 in the public preview of Meta Model API. Developers could call the multimodal reasoning model from their own applications, with support for a million-token context window.
Meta Superintelligence Labs introduced Muse Spark as the first model in its new series. It began powering Meta’s assistant in the Meta AI app and on meta.ai, applications that Meta had previously launched with Llama 4.