Muse Spark Open: Meta's open model that runs on two Mac Studios
Muse Spark Open, Meta's open-weights model: 816B parameters, 1M context and how to run it on two Mac Studios.
Muse Spark Open is Meta's open-weights model that, according to Meta's own post, can run on two Mac Studios with 512 GB of memory each. It has 816 billion parameters in total, activates about 42 billion per token and handles a 1 million token context. Meta published the weights on October 15, 2026.
What Muse Spark Open is and why it matters
It is a mixture-of-experts model: although it adds up to 816B parameters, it only activates about 42B for each token, which makes it more manageable than the headline number suggests. It accepts text, image, video and audio as input, and Meta positions it for agent work such as coding or customer support.
Three ways to run it
- On your desk. As a hardware example, Meta points to two Mac Studios with M5 Ultra and 512 GB of unified memory each, or four with 256 GB, using MLX.
- By downloading it. The weights are on Hugging Face, in the repository meta-models/Muse-Spark-Open-816B-A42B. The compressed NVFP4 version weighs about 530 GB and the full-precision one about 1.6 TB.
- By renting it. It is available through providers such as OpenRouter, Together AI and Fireworks AI, among others. This is the only option that needs no hardware of your own.
What Meta does not say
Meta's post publishes no benchmarks, so there is no way to compare its quality with other models from that source alone. We also did not review the terms of the Muse Model License here: read them before using the model in a product. And the post gives no prices, so it cannot be said that renting is cheaper than buying hardware.
How to apply it in your business?
Instead of buying hardware, you can start by renting the model through an API and paying only for what you use. If you later want your data to stay inside your office, that is one reason to evaluate running it on your own Macs. Before deciding, read the license and test the model on a real task from your business. For more on Meta's AI lineup, read the full breakdown of Muse Wonder.
Frequently asked questions
How much memory do I need to run Muse Spark Open on a Mac?
The NVFP4 version weighs about 530 GB, which is why Meta proposes two 512 GB Mac Studios or four 256 GB ones. There are GGUF versions down to about 2 bits that need roughly 256 to 300 GB of combined memory.
Is Muse Spark Open open source?
Meta published the model's weights, meaning they can be downloaded. It is distributed under the Muse Model License, so read its terms before using it commercially.
Where can I download it?
On Hugging Face, in the repository meta-models/Muse-Spark-Open-816B-A42B. Meta also lists it in its developer documentation.
Can I use it without buying Macs?
Yes: it can be rented through an API on OpenRouter, Together AI, Fireworks AI and other providers Meta mentions.
Conclusion
Muse Spark Open puts an 816B-parameter model within reach of desktop machines with a lot of memory, or of a rented API. What remains to be confirmed is its quality against other models and the exact terms of its license.