Meta and Apple join forces for advanced artificial intelligence

Muse Spark Open is based on a multimodal architecture and includes 816 billion parameters, of which 42 billion are active during operation. The model supports a context window with one million tokens and is primarily intended for programming, complex multi-step tasks and working with large datasets. It takes advantage of the latest version of Apple's open source MLX framework, which is optimized for the unified memory architecture of Apple Silicon processors.

One of the biggest advantages of this design is the efficient use of shared memory. The integrated graphics processor can access up to 512GB of memory on a single device, so two Mac Studio computers can run the entire model together as a single system with up to one terabyte of memory and up to 1.2TB/s of bandwidth. This approach allows for large models to be run without the need for dedicated server-side graphics cards.

Meta demonstrated the system's capabilities with the Muse Code tool, where the model independently performed more complex tasks, such as fixing the code of open source projects and editing videos in the DaVinci Resolve environment. Meta sees this as an important step towards more advanced AI agents that can independently perform complex workflows for longer periods of time.

The Muse Spark Open model is now available for download via the Hugging Face platform.