Meta and Apple team up on October 15 to make Muse Spark Open run better on Macs
Meta and Apple jointly announced on October 15 an effort to optimise the open-weights Muse Spark Open model for Mac, using Apple MLX and Metal acceleration so a single M3 Ultra with 512GB of unified memory can load the full model and reach around 18 tokens/s inference.
Meta and Apple jointly announced on October 15 an effort to optimise Meta's newly open-sourced Muse Spark Open multimodal large model for running on Mac. The optimisation uses Apple MLX machine learning framework and Metal graphics API; a single Mac M3 Ultra with 512 GB of unified memory can load the 2-bit quantised version of Muse Spark Open and run inference at around 18 tokens/s; the MLX community simultaneously released 3-bit and 4-bit oQ quantised weights for the M2 and M4 series Macs. The cooperation covers shared kernel-level optimisations, Metal built-in debugging features and the maintenance of quantisation checkpoints in the Hugging Face MLX community; official demos of Muse Spark Open on Apple Silicon include a browser agent (based on OpenEnv and Hugging Face Sandbox), a voice browser assistant (combining audio, image and reasoning to autonomously search products on Amazon), and a 3D asset generator (combining llama.cpp with a Pi coding agent and Three.js to iteratively generate 3D assets). This is the first time Meta and Apple have officially cooperated on running an open-source large model locally on Mac, contrasting with Microsoft's tight historical integration with OpenAI and Apple's historical integration with OpenAI, and reflects Meta's strategy of using the open ecosystem to reach consumer endpoints.