Meta released its 816-billion-parameter Muse Spark Open model on 15 October after working with Apple to run it locally on connected M5 Ultra Mac Studios, Axios reported. Axios described the model as similar to Muse Spark 1.3.
Key points
- Meta and Apple adapted the model to run across two or four connected M5 Ultra Mac Studios using Apple’s MLX framework.
- Meta puts the model at 42 billion active parameters per token; MacStories reports 40 billion.
- The weights are free to download under a licence that permits commercial use.
- A 256GB Mac Studio configuration with 2TB of storage costs more than $11,000, according to MacStories.
Meta puts Muse Spark Open on Macs
Meta’s AI chief Alexandr Wang said the companies had worked together for several months. He said Daniel Gross, a former Apple executive now at Meta, played a substantial part in bringing them together. The work gives Meta a way to distribute its model for use on Apple hardware without requiring developers to send their work to a cloud service.
Muse Spark Open is available through Hugging Face as a free download with a licence allowing commercial use. Local use gives developers another deployment option, while availability through cloud providers and developer tools offers a route that does not require them to acquire the Mac Studios used in the demonstration.
Meta calls Muse Spark Open the most capable US-developed open-weight model available, Axios reported. Nvidia released updated Nemotron models on 13 October, while Reflection AI and France’s Mistral have also released open-source models. Meta’s commercial-use licence and Mac optimisation put its release alongside those efforts to reach developers who want access to model weights.
The collaboration sits alongside a history of public disagreements between Apple and Meta over privacy and their business models. Apple’s privacy changes in 2021 affected Facebook’s advertising business. The companies still compete in devices and computing, even as each has an interest in this release: Meta in distributing its model, and Apple in demonstrating what its desktop hardware can run.
MLX spreads the model across Mac Studios
Meta says the model has 42 billion active parameters per token and a 1-million-token context window, 9to5Mac reported. MacStories reported 40 billion active parameters per token. The 816 billion figure refers to the model’s total parameters.
Meta used Apple’s MLX framework to optimise Muse Spark Open for the M5 Ultra Mac Studio. MLX is designed for the unified memory architecture in Apple silicon. The M5 Ultra’s integrated GPU can use up to 512 GB of unified memory in each Mac Studio, allowing two machines to hold the model, 9to5Mac reported.
Meta says it tested agentic work entirely on the machines, including changes to the open-source Mastodon client Ice Cubes and video editing tasks in DaVinci Resolve. Those examples describe the companies’ own tests rather than a customer deployment. The arrangement demonstrated by Apple and Meta uses two or four connected Mac Studios.
Apple’s local-computing work also comes as Meta expands where people can use its assistant. AI Affairs previously reported that Meta launched Muse on iPad a month after its mobile debut. The Mac Studio work concerns access to model weights for developers, rather than an announcement that Apple will use Muse Spark Open in Apple Intelligence.
A 493GB model meets Mac Studio costs
The 4-bit version of Muse Spark Open requires 493GB of memory. Apple’s support for distributing work over Thunderbolt 5 lets the model run across multiple M5 Ultra Mac Studios. Some Apple testing used a 512GB configuration, and running the model requires at least two such machines.
Hardware is the direct cost for a developer choosing the demonstrated local setup. A Mac Studio with 256GB of memory and 2TB of storage costs more than $11,000. Apple has not said what it will charge for the 512GB configuration, according to the publication.
Apple is slated to release the 512GB M5 Ultra Mac Studio at the end of 2026.
