Meta releases 816B-parameter Muse Spark Open for coding and agent work

Apple and Meta, once fierce competitors, have found an unlikely common ground to work together. The two tech giants have joined forces to enable Meta’s Muse Spark Open model to run locally on Apple’s powerful M5 Ultra Mac Studios. Announced on October 15, this collaboration allows developers to download, customize, and utilize the open-weight model, though it requires two connected Apple desktops to operate at its full potential.

As Meta’s AI chief Alexandr Wang put it to Axios, “We’re excited for mommy and daddy to be playing nice again.” Despite a strained history, with Meta’s CEO Mark Zuckerberg once criticizing Apple on Joe Rogan’s podcast for its lack of innovation and restrictive policies, the two companies now demonstrate how Apple’s silicon can perform.

Muse Spark Open builds upon Meta’s Muse Spark model family, introduced in April, and is hailed as the most capable U.S. open-weight model available based on Meta’s own evaluations. The open-weight model means developers can download and run the model’s learned parameters without relying on Meta’s apps or cloud services, subject to the model’s license.

In tests, Meta showcased Muse Code, powered by the new model, making changes to the Mastodon app Ice Cubes and operating video editor DaVinci Resolve, among other tasks. However, there’s a significant caveat to this desktop partnership. The 816-billion-parameter model requires two M5 Ultra Mac Studios with up to 512GB of unified memory each, running through Apple’s MLX machine-learning framework.

Therefore, standard MacBooks are not equipped to handle this model. This collaboration follows the release of Meta’s smaller model, Muse Glimmer, designed to run on laptops or PCs. Both models highlight the appeal of running AI on locally controlled hardware, enabling processing without sending each prompt to a hosted model. This arrangement also provides Apple and Meta with a workaround to a longstanding issue: privacy.

Meta assures that it has no access to developers’ local use of the model, making this collaboration a means to conduct AI work on Apple hardware without compromising user privacy. Muse Spark Open is now available on GitHub, along with the MLX Adapter needed to run it.