Skip to content

Meta releases Muse Spark Open, the "most capable" open model in the US: it runs on a Mac, if you have at least a terabyte of RAM

1 min read
Share
Meta releases Muse Spark Open, the "most capable" open model in the US: it runs on a Mac, if you have at least a terabyte of RAM

Meta has released a new AI model, Muse Spark Open, and immediately slapped a label on it: the "most capable" open-weights model in the United States. The verdict comes from the company itself, based on its own tests in which the model, naturally, beats other popular open models. Just three days ago Meta unveiled Muse Wonder, its first top-tier model. Now comes the version with weights open to everyone.

On paper the numbers are big: 816 billion parameters and a context window of 1 million tokens, enough for the model to hold very long documents or a large chunk of a codebase in its head at once. It is multimodal from the ground up and built for complex programming and so-called agentic tasks, where the AI carries out a series of steps on its own rather than just answering questions.

The headline is where it runs: locally, on Apple computers. This is where the small print begins. To run it on your own machine you need at least 1 terabyte of RAM, and it performs best on a Mac Studio with the M5 Ultra chip. That's not the laptop you write your term paper on. "Runs on a Mac" sounds democratic - until you read the spec sheet.

Meta has paired the model with Muse Code, its AI coding terminal. In one test they set it loose on a client for the Mastodon social network: the model fixed bugs, added new features and built the entire app from start to finish. In another test it worked in the video editing software DaVinci Resolve, where it reviewed and cut footage on its own.

The model is already available through several API platforms. Why is Meta giving away something others keep under lock and key? Open weights are the fastest route to an ecosystem: developers build on your model, not someone else's. For a developer in Skopje or Belgrade, an open model means independence from someone else's cloud and from a monthly subscription in dollars - but only if they have a terabyte of memory. Is openness that requires that kind of hardware really openness for everyone?