Meta Unveils Muse Spark Open: AI Model Runs Locally on Apple Mac Studio

What Muse Spark Open is

Start with size. Muse Spark Open carries 816 billion parameters, and 42 billion of them are active at any given time. The model is natively multimodal, so it was built from the start to handle more than plain text. It also has a context window of 1 million tokens. Meta says the model is tuned for agentic work. That covers coding, long-horizon workflows and multi-step reasoning. The weights are available through several major platforms, including Hugging Face and Amazon SageMaker.

Where the Mac comes in

Running an 816 billion parameter model usually calls for a data center. Muse Spark Open can instead run on two Apple M5 Ultra Mac Studios linked together. The setup relies on Apple's MLX framework, a software toolkit for machine learning on Apple chips. MLX lets the two machines behave like a single computer. Together they pool nearly 1 terabyte of unified memory.

Building out the Muse ecosystem

Muse Spark Open isn't arriving alone. It extends Meta's broader Muse AI agent lineup, which already includes a native Mac app. Meta launched its Muse personal AI agent on September 8, 2026, roughly five weeks before Spark Open. That agent performs tasks across native macOS apps, operating within permissions the user controls.

What this means for Apple, Meta and the AI race

For Apple, the partnership is a strong advertisement for its silicon. Running a model of this size on two desktop machines demonstrates on-device AI at frontier scale. For Meta, the move reinforces its bet on open weights. Rivals that keep their best models locked behind cloud APIs force customers to send data out and pay per use. Meta is offering a different deal: take the model and run it yourself.

\1\1\1