Meta Releases Muse Spark Open, an 816-Billion-Parameter Model Built to Run on Two Mac Studios
Meta released the weights for Muse Spark Open, an 816-billion-parameter open-weight model for coding and agentic work, tuned to run on two linked Mac Studios with a terabyte of unified memory.
By Icarus4 min read
Meta released the weights for Muse Spark Open on Thursday, an 816-billion-parameter open-weight model built for coding and agentic work that Meta describes as the most capable U.S. open-weight model to date. The release, announced October 15, arrives three days after Meta launched its closed frontier model family, Muse Wonder, and marks a return to large-scale open-weight releases for the company.
Muse Spark Open is a mixture-of-experts model: 816 billion total parameters, with 42 billion active for each token. It accepts text, images, audio, and video as input, returns text, and holds a context window of up to one million tokens. Meta built it for long-horizon tasks, multi-step reasoning, and software engineering, and says it excels at agentic coding work such as reading a codebase, debugging, and calling external tools.
A Desktop Showcase: Two Mac Studios, One Terabyte
The most unusual part of the release is the hardware Meta built it around. Working with Apple, Meta tuned Muse Spark Open for the Mac Studio powered by Apple’s M5 Ultra chip, using Apple’s MLX framework. A single top-configuration Mac Studio holds up to 512GB of unified memory, so two linked machines can hold the entire model. Meta says the pair behaves like one supercomputer with a terabyte of unified memory.
Meta chief executive Mark Zuckerberg announced the release on X, crediting Apple chief executive John Ternus and Apple for the collaboration, and Ternus replied with congratulations. Meta AI chief Alexandr Wang announced the release in an X thread of his own, saying Meta is open-sourcing the weights for broad use and that Meta Superintelligence Labs is committed to open source. Wang also posted a short video styled as a homage to Apple’s 1984 advertisement; in a follow-up reply, he said Muse Wonder Full directed and edited the clip with minimal art direction.
Running the model locally takes serious hardware. The full-precision BF16 checkpoint is about 1.5 terabytes, per Oossa, and the compressed NVFP4 version about 530GB, per Madrobot. The 512GB Mac Studio configuration does not ship until later in October, and buying two of them puts the hardware bill well into five figures before a single prompt runs.
What Meta Says It Can Do
Meta says it tested the model on agentic workflows running entirely on-device through Muse Code, including shipping changes to Ice Cubes, an open-source Mastodon client (bug fixes, new features, and a full build), and driving DaVinci Resolve, with the model reviewing footage frame by frame to find the strongest moments and crop them vertically.
On benchmarks, Meta reports Muse Spark Open leads rival open-weight models in agentic, knowledge-and-reasoning, and multimodal categories, while trailing some competitors on coding benchmarks. The company positions it below its paid Muse Wonder Full models. The benchmark figures are Meta’s own reported results; independent verification has not yet appeared.
Open Weights, With Conditions
Releasing the weights lets developers download, run, and adapt the model on their own infrastructure rather than accessing it only through Meta’s API or app. That is distinct from open source: Meta is not publishing training data or training code, and the release carries license terms.
The weights are available on Hugging Face under the Muse Model License. The license requires work built with the model to carry a “Built with Muse Spark Open” label, and providers earning more than $20 million a year who want to offer the model as an API service need a separate agreement with Meta, according to The Decoder’s reading of the terms. Hosted access is also available through platforms including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI, while the weights are listed on Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and the Microsoft Foundry Model Catalog.
Why It Matters
The release gives Meta two tracks at once. Muse Wonder, launched October 12, stays closed and is monetized through the Meta Model API and Muse Code subscriptions. Muse Spark Open gives developers and researchers a downloadable model they can inspect, adapt, and run themselves, including on a desktop rather than in a data center.
Muse users will feel the practical effect only indirectly. Spark Open is a developer-facing model; the Muse app itself is unchanged. Its significance is in what developers can build with it, and in Meta’s demonstration that a model of this scale can run on local hardware at all.
