Meta released the weights for Muse Spark Open on October 15, giving developers access to its multimodal agent model. But Meta’s featured Mac setup pairs two Mac Studio computers, each with 512 gigabytes of unified memory.

“Open weights” means developers can access and deploy the model. It does not mean most desktop computers can run it locally.

The model is available to download or use through hosted inference providers. For developers without high-memory machines, hosted services offer a way to try it without buying the hardware.

Open weights, heavyweight hardware

Muse Spark Open has 816 billion parameters in total, with 42 billion active for each token. Its context window extends to 1,048,576 tokens. Meta says it can process text, images, audio and video.

The active-parameter figure is not the whole hardware story. Meta’s documentation says serving the model requires loading all its weights. Its NVFP4 checkpoint is about 537 gigabytes, while the full-precision BF16 version is about 1.65 terabytes.

  • Meta’s featured setup uses two Mac Studios with 512 gigabytes of unified memory each.
  • An alternative uses four Mac Studios with 256 gigabytes each.
  • The NVFP4 checkpoint is about 537 gigabytes; the BF16 version is about 1.65 terabytes.

That is a very different proposition from installing an ordinary desktop app. Downloading the weights is one step; running the model locally requires substantial memory.

Hosted access and license terms

Meta lists Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI among the hosted providers. It also says the weights are available through Hugging Face and several cloud model catalogs.

Meta released the weights under its Muse Model License. The company describes the license as having limited commercial terms and allowing many developers to use the model for free. That is not unrestricted commercial use, so developers should check the terms before building a product around it.

What changes for developers

The release gives teams the option to self-host, while hosted access remains available for those without the required hardware. The choice is practical: supply the machines to run the model, or use a provider that does.

Meta’s announcement and developer guide explain the release and availability. Axios also reported on Meta and Apple’s work to bring the model to Apple’s high-end Macs.