Overview
- Apple and Meta announced the collaboration Thursday that lets developers download Muse Spark Open’s weights and run the model locally on linked M5 Ultra Mac Studios.
- Apple’s MLX framework splits the mixture‑of‑experts model across machines and pools nearly one terabyte of unified memory so two to four Mac Studios can behave like a single system.
- Muse Spark Open is an open‑weight, multimodal model with about 816 billion parameters and a 1 million‑token context window, and Meta says it is tuned for agentic tasks like coding and long workflows.
- Running the model today requires multiple top‑tier M5 Ultra Mac Studios with very large RAM configurations that cost at least $10,000 per machine, so the setup targets professionals and enterprises rather than typical consumers.
- Apple published performance benchmarks and technical docs for MLX, and the move underscores Meta’s open‑weights strategy and Apple’s chip strengths while raising questions about safety, privacy, enterprise controls and how this could change the balance between local and cloud AI services.