Meta has released Muse Spark Open, which it calls the most capable open-weight model in the United States. It carries 816 billion parameters and a 1 million token context window, and it is natively multimodal.
The detail that makes it unusual is the target: it was designed to run locally on Apple hardware, built in partnership with Apple. The catch is memory. Running it needs at least 1 terabyte of RAM, and Meta says it performs best on a Mac Studio with the M5 Ultra chip.
Meta paired the model with Muse Code, its AI coding terminal, and describes using it for agentic work including bug fixes, new features and a full build on a Mastodon client. It also says the model was used to operate DaVinci Resolve, where it reviewed and cropped footage.
The model is available now through several API platforms as well as locally.
Why it matters
- A terabyte of RAM is not a laptop, but it is a single machine in a room rather than a datacenter contract. Local frontier-class inference is becoming an option for teams that need it.
- A 1 million token context window running on your own hardware changes what you can hand a model — an entire repository, not a selection of files.
- Open weights mean you can inspect and modify the thing you depend on, which closed APIs do not offer.