Meta has released the weights for Muse Spark Open, a new open model that it calls the most capable American open-weight model so far. It has also done something few people expected: worked closely with Apple to make sure the model runs well on a Mac.
The numbers are large. Muse Spark Open has 816 billion parameters in total, but only 42 billion are active for any given token. That mixture-of-experts design keeps it from overwhelming the hardware running it. It handles images as well as text, and its context window is one million tokens. Meta is aiming it at coding and agentic work, meaning long tasks with many steps where the model plans, acts and fixes its own mistakes. Those are Meta’s claims for now. Independent benchmarks and developer testing will show how it really compares with closed models from OpenAI, Anthropic and Google, and with strong open-weight rivals from China.
The Apple connection is the more surprising part. Meta tuned the model for the Mac Studio with M5 Ultra using MLX, Apple’s open-source machine learning framework built for Apple silicon’s unified memory. A single M5 Ultra Mac Studio can give its GPU up to 512GB of that memory, so two linked machines can hold the whole model in about a terabyte of shared RAM. Meta says the chip’s 1.2TB/s of memory bandwidth is fast enough because each token only needs a small part of the model. That’s the same reason high-memory Macs have become popular with people running large models locally.
To show what this looks like in practice, Meta says it used its Muse Code tool to make changes to Ice Cubes, an open-source Mastodon client for iOS, including bug fixes, new features and full builds, entirely on the device. It also used the model to control DaVinci Resolve, going through footage frame by frame to find highlights and reframe them as vertical video. Apple has backed the effort with a post on its Machine Learning Research site about running models locally with MLX.
The partnership is notable because Apple and Meta have been on bad terms for years, most visibly after Apple’s App Tracking Transparency feature hurt Meta’s ad business. Meta AI chief Alexandr Wang told Axios the two companies have been working together for several months, and credited former Apple executive Daniel Gross with helping bring them together.
There is a practical limit to who can use this. Running a frontier-scale model on your own desk requires two top-spec Mac Studios, which few hobbyists can afford. For businesses that want to keep sensitive code and data off the cloud, though, Apple and Meta are making a strong argument for private AI that runs locally. The model is available now on Hugging Face.
