Meta has released Muse Spark Open, an open-weight AI model featuring 816 billion total parameters and 42 billion active parameters. The model utilizes a mixture-of-experts architecture and includes a 1-million-token context window. It is natively multimodal, supporting text, images, video, and audio processing.
The model is optimized to run locally on two Apple M5 Ultra Mac Studios using the Apple MLX framework. Each Mac Studio supports up to 512GB of unified memory, creating a combined 1TB memory pool. Meta developed custom Helion GPU kernels to enhance performance across various hardware, including chips from AMD, Intel, and Nvidia.
Developers can access the model weights via platforms such as Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models. Inference services are provided by Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI.
The release is subject to the Muse Model License, which imposes specific restrictions on commercial use. Meta conducted safety evaluations using its Superintelligence Scaling Framework prior to the public release of the model weights.
