Meta Releases Muse Spark Open, an 816-Billion-Parameter AI Model Built to Run on Two Mac Studios

Meta released Muse Spark Open on Thursday, an open-weight artificial intelligence model the company calls “the most capable U.S. open-weight model to date,” with 816 billion total parameters and a one-million-token context window, according to Meta's own research blog.

The model activates only about 42 billion of those parameters at a time, a mixture-of-experts design meant to keep it workable outside a data center. Even so, running Muse Spark Open on local hardware takes roughly a terabyte of memory, Engadget reported. Meta built the model specifically to run across two Apple Mac Studio desktops fitted with M5 Ultra chips, each offering up to 512 gigabytes of unified memory; linked through Apple's MLX framework, the pair can hold the full model between them. A cheaper route exists, too: four Mac Studios with 256 gigabytes apiece, at a combined cost of roughly $38,000, can also run it.

Meta is positioning the release as evidence that it has not abandoned open models even as it keeps its flagship Muse Spark system closed. The company posted the weights for free download on Hugging Face, with hosted access through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI, and said it applied additional safety post-training before publishing them, citing thresholds under its Meta Superintelligence Scaling Framework. The model is governed by Meta's own Muse Model License, which allows free use for most developers but limits large commercial deployments, TechCrunch reported.

Meta said two Mac Studios running the model together “behave like one supercomputer with a terabyte of unified memory.”

In demonstrations tied to the launch, Meta paired Muse Spark Open with Muse Code, its AI coding tool, to rebuild parts of Ice Cubes, an open-source Mastodon client, including bug fixes, new features and a full build, and used the model inside the video editor DaVinci Resolve to review and crop raw footage. Meta said the model outperforms the Chinese-developed GLM 5.3 and Kimi K3 systems on agentic and visual-reasoning tasks, though those comparisons rely on the company's own benchmarks and have not been independently verified.

The release follows Meta's open-weight Muse Glimmer model in August and comes three days after the company's newest frontier release, Muse Wonder, on Oct. 12. Chief executive Mark Zuckerberg said the new model “runs great on new Mac Studios,” underscoring how closely Meta is tying its open models to Apple's highest-memory desktop hardware even as its underlying flagship system remains proprietary. The push continues a broader campaign, launched alongside Muse Glimmer, for looser U.S. policy toward open-weight systems so that American developers can compete with similarly open models coming out of China.

The timing intersects with a broader shortage of memory chips that has pushed up component prices industrywide. Apple's own 512-gigabyte configuration of the M5 Ultra Mac Studio, the single machine capable of running Muse Spark Open on its own, has not yet shipped; Apple has said only that it is due “late October,” without announcing a price. In the meantime, developers without a spare Mac Studio or two can still reach the model through hosted application programming interfaces, or download its weights and model card directly from Hugging Face.