According to a Hugging Face blog post, Meta has released Muse Spark Open, an open-weight multimodal model built on a mixture-of-experts architecture. The post frames it as a frontier model, suggesting it targets high-end capability while remaining openly available to developers and researchers.
The mixture-of-experts design is notable because it can route different inputs to specialized subnetworks, potentially improving efficiency relative to dense models of similar scale. By releasing the weights openly, Meta is positioning the model as a resource for the broader AI community rather than a product locked inside its own ecosystem.
Hugging Face’s welcome post emphasizes the significance of hosting such a model on its platform. Since this article draws on a single source, there are no conflicting accounts to compare, but the blog’s tone makes clear that the release is seen as a meaningful step for open multimodal AI.