Definition
Muse models are Meta Superintelligence Labs' models sold or released under the Muse brand. Spark is the general multimodal reasoning line; Glimmer supplies downloadable weights for local use. Other Muse names identify specialized generation or transcription models.
The Muse personal agent and Muse Code are applications that use these models. Their harnesses add tools, stored state, and permissions.
Origin and attribution
Meta introduced Muse Spark on April 8, 2026, calling it the first Muse model. The company credits Meta Superintelligence Labs. This model introduction predates the September launch of the Muse personal agent.
Meta released Spark 1.3 on September 2, 2026, with a focus on sustained agentic workflows and coding. Record the version when evaluating performance or reproducing a result.
Families, modalities, and access
The October 7 catalog lists Spark 1.1, 1.2, and 1.3 separately. Spark produces text and accepts multiple input modalities. Its documentation warns that audio understanding in 1.3 is not fully supported, despite audio appearing in the catalog table. A broad multimodal label does not establish equal quality for every modality.
Muse Glimmer is a 30B dense model distilled from Spark, with text and image inputs and text output. Meta releases its weights under Apache 2.0 for self-hosting. That access claim applies to Glimmer; it does not make all Muse models downloadable under the same terms.
Muse Image generates and edits images. Muse Voice Transcribe returns text transcripts from audio and does not synthesize speech. Meta's July 7 announcement also previewed Muse Video. A preview announcement and an available API model are different release states.
Limits and deployment boundaries
Spark's Standard and Contributor API tiers have different data-use terms. Meta describes Standard data as excluded from training and Contributor access as discounted in exchange for training permission. A model-family name alone cannot resolve the privacy terms of a request.
Local Glimmer inference gives the operator control over the serving environment. Connected tools, telemetry, or a hosted wrapper can still send information elsewhere. Inspect the full deployment before calling an agent local or private.
Meta presents Muse as part of its personal-superintelligence program. That positioning does not establish that a model satisfies a general superintelligence threshold. Likewise, demonstrations and provider benchmarks need their exact model version, reasoning settings, harness, and evaluation conditions.
Distinguish it from nearby terms
- Muse is also the personal agent's product name. A Spark API call does not create that agent's memory, cloud computer, or permission system.
- Muse Code is a coding application; Muse Spark is its model line.
- Open-weight describes parameter access. It applies to the specified Glimmer release without determining the access terms of Spark or media models.
- Multimodal describes supported kinds of input or output. Spark's text output, Image's image output, and Voice Transcribe's transcripts have different contracts.
Check your understanding
A deployment report says only that it uses Muse. Which model family, version, serving location, data-use tier, output contract, and application permissions would you need before judging its privacy and capabilities?