Model
Muse
Meta Superintelligence Labs' proprietary model family, launched with Muse Spark in April 2026 as the successor to the open-weight Llama line.
Muse is the model family developed by Meta Superintelligence Labs (MSL), the AI organization Meta formed in June 2025. Its first release, Muse Spark, launched on 8 April 2026 and was the first model built from the ground up by the team led by Chief AI Officer Alexandr Wang. Reported under the code name Avocado during development, it was produced over roughly nine months and now powers the Meta AI assistant.
Muse Spark marks a change in Meta’s model strategy. Meta’s previous flagship family, Llama, was released with open weights from February 2023 through Llama 4 in April 2025 and became widely used by developers. Muse Spark shipped as a closed, proprietary model. Meta said it planned a version under an open license, but at launch it did not publish weights, parameter counts, architecture details or training data, and commentators described the move as a pivot away from the open-weights approach. Meta described Spark as a major upgrade over Llama 4 and said it “significantly narrows the performance gap” with rival labs, while an executive said it was not a new state of the art.
The model accepts text, image and speech input and produces text. It offers several reasoning modes, from instant to extended, a shopping mode, tool use and multi-agent orchestration. It is free to use on meta.ai and in the Meta AI app, with rollout planned across WhatsApp, Instagram, Facebook, Messenger and Ray-Ban Meta glasses, and an API preview for selected partners. Meta has said it plans larger models on the same redesigned architecture.
Muse is the model line through which Meta pursues its stated vision of “personal superintelligence for everyone.” Meta has not claimed that any Muse model is a Super Intelligence.
For current details see meta.ai.