Meta Releases Muse Spark Open: A Frontier-Class, Open-Weight Multimodal MoE Model
Meta has introduced Muse Spark Open, an open-weight frontier model built on a multimodal mixture-of-experts (MoE) architecture. The release signals Meta's continued push to make high-capability AI models broadly accessible to researchers, developers, and the open-source community.
What Is Muse Spark Open?
Muse Spark Open is a multimodal model, meaning it processes and reasons across multiple input types rather than being limited to text alone. Its underlying mixture-of-experts design activates only a subset of the model's parameters for any given task, enabling stronger performance at a lower inference cost compared to dense architectures of similar scale.
As an open-weight release, the model's weights are made publicly available, allowing the community to inspect, fine-tune, and deploy it on their own infrastructure โ a contrast to fully closed, API-only frontier models.
Model Assets and Access
The associated model assets are hosted on the Hugging Face Hub under the repository:
- Repository:
merve/meta-muse-spark-open-assets - Last updated: October 15, 2026
- Status: Available for viewing and download
Developers and researchers can access the assets directly through the Hub to begin exploring the model's capabilities.
Why This Release Matters in 2026
In the current landscape, the gap between proprietary and open-weight frontier models continues to narrow. Releases like Muse Spark Open reinforce two key trends:
- Multimodality as the default โ Leading 2026 models increasingly handle images, text, and other modalities within a unified architecture.
- Efficient scaling via MoE โ Mixture-of-experts designs have become the standard approach for achieving frontier-level quality without proportional compute costs.
By open-sourcing a frontier multimodal MoE model, Meta lowers the barrier for organizations that want to build on top-tier AI capabilities without relying on closed APIs.
Getting Started
To explore Muse Spark Open, visit the model assets repository on Hugging Face, review the documentation, and evaluate the model against your own tasks. As with any frontier model, consider the compute requirements, licensing terms, and intended use guidelines before deployment.
Stay tuned for further benchmarks, fine-tuning recipes, and community tooling as the ecosystem around Muse Spark Open develops.