Muse Spark
Era 5 · Generative AI, reasoning, and agents (2022–2026) › Frontier models
- Lineage
- Neurons and networks
- Kind
- Implementation
Who
Meta Superintelligence Labs
What happened
Meta Superintelligence Labs releases Muse Spark, its first model: a multimodal reasoning model that now powers the Meta AI assistant.
Why it mattered
The first model from the lab around which Meta reorganized its research. It replaces Llama in Meta's products.
The full story
Since 2023 Meta's language models had been the Llama models. In June 2025 the company reorganized its research: it announced a $14.3 billion investment in Scale AI, and on the 30th Mark Zuckerberg created Meta Superintelligence Labs, with Alexandr Wang as chief AI officer.
The new lab's first model arrived on April 8, 2026. Muse Spark, the first in the Muse family, is a natively multimodal reasoning model, with tool use, visual chain of thought, and coordination of multiple agents. Its Contemplating mode sets several agents reasoning in parallel, and according to Meta it reaches 58 percent on Humanity's Last Exam.
Meta emphasized efficiency: it claimed the model matched the capabilities of Llama 4 Maverick, its previous model, with more than ten times less compute, thanks to improvements in architecture, optimization, and data curation.
Muse Spark began powering the Meta AI assistant, in the app and on meta.ai, and was offered to some users through a private API preview. Meta presented it as the first step toward what it calls "personal superintelligence," with larger models in development.
Sources
Scholarship and context
How to cite
“Muse Spark.” The History of AI, revised November 13, 2026. https://thehistoryofai.com/en/milestone/muse-spark
This entry was last reviewed on November 13, 2026.