Meta launches 816-billion-parameter Muse Spark Open AI model that runs on Apple M5 Ultra Macs

Meta has released Muse Spark Open, an 816-billion-parameter open-weight artificial intelligence model that can run locally on Apple’s latest M5 Ultra Mac Studio computers, marking an unusual collaboration between the two technology companies.

The model has been optimised to run across two Apple Mac Studios using Apple’s MLX machine learning framework. Meta CEO Mark Zuckerberg highlighted the collaboration on X, writing that “Muse Spark Open – weights releasing today – runs great on new Mac Studios thanks to work with @johnternus and Apple.”

Apple CEO John Ternus responded to Zuckerberg, saying it was “amazing to run a model of Muse Spark Open’s caliber and scale this efficiently, right from your desk.”

Muse Spark Open has 816 billion total parameters, of which 42 billion are active at a time. It also has a 1-million-token context window and is natively multimodal, meaning it can process text, images, video and audio. Meta says the model has been built for coding and agentic work, including long-horizon tasks, complex software engineering and multi-step reasoning.

How Muse Spark Open runs on Apple Mac Studio

The Apple connection is significant because a model of this size would normally require data-centre infrastructure. Meta says Muse Spark Open can run on a desktop using two Apple M5 Ultra Mac Studios, with each machine configured with up to 512GB of unified memory. Together, the two machines provide 1TB of memory.

Apple’s MLX framework allows the model to be distributed across the two Macs, with the machines connected using high-speed communication. Apple has also published technical details on running Muse Spark Open across multiple Mac Studios.

Meta describes Muse Spark Open as its “most capable US open-weight model to date”. Open-weight models make their trained parameters available to developers, allowing them to download, run and adapt the models rather than accessing them only through a cloud-based service.

The model uses a mixture-of-experts architecture, activating only 42 billion parameters for each token rather than all 816 billion at once. This helps make the model more practical to run despite its overall size.

Meta says Muse Spark Open has been tested with Muse Code, its AI coding terminal, on software engineering tasks, including making bug fixes, adding features and completing a full build of an open-source Mastodon client. The company also tested it with DaVinci Resolve, where the model was able to review footage and identify and crop sections of video.

Apple’s testing found that a cluster of four M5 Ultra Mac Studios running Muse Spark Open could generate more than 36 tokens per second. The setup demonstrates how Apple’s unified memory architecture and MLX framework can be used to run large AI models locally.

Meta is making Muse Spark Open available through API platforms including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. The model weights are also available through platforms including Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models.