Meta's Muse Spark Open brings an 816-billion-parameter AI model to the Mac

Meta has launched what it calls the "most capable" open-weight model in the United States — and it designed the thing to run on a Mac. The model, named Muse Spark Open, packs 816 billion parameters and a 1 million token context window, according to evaluation results Meta released on Thursday. That combination of raw scale and local, on-device availability puts Meta in a rare position in a market where the largest open-weight releases typically demand data-center hardware.

The claim of superiority is Meta's own. The company says its evaluation results show Muse Spark Open outperforming other popular open-weight models. Meta did not frame this as a marginal win; it staked the launch on the phrase "most capable," a label that invites immediate comparison with the other heavyweight open-weight contenders the industry has been tracking all year.

What exactly did Meta release?

The headline numbers are straightforward. Muse Spark Open is an 816-billion-parameter model. It handles a context window of 1 million tokens. It is natively multimodal, and Meta built it for two kinds of work in particular: complex coding tasks and agentic labor — AI that takes actions across software rather than just answering questions.

Those last two details matter more than they might first appear. Coding and agentic workflows are exactly where enterprises are spending money on AI right now, and they're the categories where closed, proprietary models from the largest labs have held their strongest foothold. Meta is aiming this release squarely at that gap.

The model is available now through a number of different API platforms, the company said.

Can a Mac actually run this?

Here's the catch, and it's a big one. Running Muse Spark Open locally requires at least 1 terabyte of RAM, per Engadget. That's not a configuration most people own. It's not even a configuration most professionals own.

Meta says the model runs best on the Mac Studio equipped with the M5 Ultra chip. Apple's top-tier desktop machines are the realistic floor here, and even then, buyers need to push the memory configuration to the top of the range to hit the requirement. So the "runs on a Mac" pitch is technically true and practically narrow.

Still, the framing is significant. A model of this scale being designed for local execution on consumer-adjacent hardware — however expensive — signals where Meta thinks the boundary between cloud and desktop AI will sit in the near term. If the memory requirement drops over successive releases, the desktop-AI story gets much broader.

What has the model actually done?

Meta offered two concrete demonstrations of Muse Spark Open at work, and both involved pairing it with Muse Code, the company's AI coding terminal.

In the first, Meta deployed the model to various agentic workflows, including coding fixes for a Mastodon client. The work included "bug fixes, new features, [and] full build," Meta said. That quote is worth pausing on. A full build — taking a codebase from broken or incomplete to compiled and working — is the kind of multi-step, sustained task that separates agentic systems from chatbots that suggest a snippet and stop.

In the second, Meta put Muse Spark Open in charge of DaVinci Resolve, the high-end video editing software. The model was capable of reviewing and cropping footage, the company revealed. That's a different register of capability entirely: visual understanding plus operational control of professional creative software, executed on the same desktop the model runs on.

Both demonstrations share a theme. They're not benchmarks on a chart. They're the model doing jobs inside real software, which is precisely the test enterprise buyers care about.

Why does open-weight matter here?

The open-weight question is the strategic core of this launch. Open-weight models differ from closed, proprietary models in ways that carry direct commercial consequences, and Meta's release lands on the advantages.

• Cost efficiency. Open-weight models can be significantly cheaper to operate than closed alternatives, since users aren't paying per-token rates to a single provider.
• Transparency. Users can inspect what the model actually is, which closed-source models don't allow.
• Flexibility. Organizations can deploy, modify and integrate the model on their own terms rather than a vendor's.

Those three properties explain why Meta keeps investing here. Every capable open-weight release applies pricing and control pressure to the closed-model business. A company that can run its coding agent locally, on hardware it owns, with a model it can inspect, has an alternative to renting intelligence by the token from a competitor.

What's the so-what?

The tension in this launch is the RAM. An 816-billion-parameter model with a 1-million-token context window that "runs on a Mac" but needs a terabyte of memory is, for now, a statement about what's possible at the top of Apple's hardware line — not a mass-market shift. The Mac Studio with the M5 Ultra is the machine that makes the claim true.

But the direction is the story. Meta has tied its most capable open-weight model to local execution, demonstrated it inside real workflows — a Mastodon client taken to "full build," footage reviewed and cropped in DaVinci Resolve — and made it available through multiple API platforms on day one. If memory requirements fall as they historically do, Thursday's release will read as the moment desktop-class frontier AI stopped being hypothetical.