Why Meta and Apple Teamed Up to Run a Massive AI Model on Mac

A model that normally needs a data center can now, in principle, sit on a desk. Meta released the weights of Muse Spark Open on October 15, calling it the most capable open-weight model in the United States, and said it worked with Apple so the model can run locally on Apple silicon.

That superlative is Meta's own, based on evaluation results it published, and outside testing has yet to confirm it.

The headline numbers are large. Meta says the model has 816 billion total parameters, 42 billion of them active at any moment, a 1 million token context window, and native support for text, images, video and audio. It is built for coding and agentic work such as long, multi-step software tasks. A few details show where the Mac fits in:

• It runs on two Apple M5 Ultra Mac Studios using Apple's MLX framework, with each machine offering up to 512GB of unified memory, per Unite.AI's summary of Meta's post
• Its mixture-of-experts design activates only 42 billion parameters per token, which reduces the computation needed for each response
• The weights also work with inference tools including vLLM, SGLang, llama.cpp and Hugging Face Transformers, with optimizations for AMD, Intel and Nvidia hardware as well

The caveats matter as much as the demo. A local run needs at least 1 terabyte of memory, according to Engadget, which means two top-specification Mac Studios rather than a typical laptop or desktop. Neither Meta nor the outlets covering it have published what such a setup costs.

The practical audience is therefore developers, researchers and companies that want frontier-scale AI without sending data to a cloud provider. For them, the appeal is control over privacy, cost and customization, and the real test will be how Muse Spark Open performs against rival models when independent developers run it on their own workloads.