Meta introduced Muse Spark, the first public release in its new Muse family, moving away from Llama’s open-source approach to closed development.

Model and positioning

Muse Spark is presented as a successor to earlier in-house models, with Meta emphasizing more advanced capabilities and controlled distribution.

The company describes the Muse family as a separate initiative from Llama, reflecting a deliberate change in licensing and deployment strategy.

Leadership and development

Meta has assigned Alexander Wang, formerly of Scale AI, to lead the Muse team after being recruited by Mark Zuckerberg to advance AGI-related work.

Muse Spark is the team’s first public release and indicates Meta’s intention to centralize development and oversight of its next-generation models.

Capabilities and access

The model is positioned for multimodal understanding and enhanced generation, but Meta’s public disclosures provide limited technical benchmarks and performance details.

Unlike Llama, Muse Spark is not offered as open-source; access appears to be managed by Meta through its own channels and partnerships.

What to expect

Coverage of Muse Spark outlines its differences from previous releases and mentions practical testing options, although precise access paths depend on Meta’s rollout policy.

• Multimodal: intended support for combined text and other input types in user scenarios.
• Advanced reasoning: framed as improved inference for complex tasks compared with earlier models.
• Integration: aimed at smoother incorporation into Meta products under controlled terms.

By keeping development closed, Meta positions Muse Spark for managed experimentation and closer oversight, distinguishing this family from the open-source trajectory of Llama.