Zuckerberg thanks Apple CEO after Meta releases Muse Spark Open weights

Mark Zuckerberg publicly thanked Apple CEO John Ternus on X after Meta published the weights for its Muse Spark Open model, and Ternus replied with congratulations and praise for Mac Studio’s on-device AI capability.

Meta described Muse Spark Open as an open-weight, mixture-of-experts model with roughly 816 billion parameters and said it is optimised to run on Apple’s Mac Studio equipped with the M5 Ultra chip.

Meta’s technical description, reflected across the reporting, says the model activates about 42 billion parameters at a time and can run locally when spread across two Mac Studio machines, each configured with up to 512GB of unified memory — a total of 1TB — allowing the full weights to be held and operated on desktop hardware.

Meta also made the weights available for download on platforms including Hugging Face, and one report states the licence allows commercial use with some limits. The companies’ exchange drew notice because it marks a publicly cordial moment between Meta and Apple after years of commercially fraught interactions described in the sources.

Reports recall that Apple’s earlier privacy changes and App Tracking Transparency affected Meta’s advertising business, and that past comments from Apple executives had provoked sharp public responses from Facebook and Meta leadership; the sources present the Zuckerberg–Ternus exchange as a contrast to that history.

Technical and cost details provided in the coverage show practical constraints for running Muse Spark Open on consumer hardware.

Reporters cite Meta’s recommendation that two Mac Studios with M5 Ultra chips and large unified memory configurations are required to operate the model locally; one account gives a retail price benchmark for an M5 Ultra Mac Studio with 256GB of memory and notes the model needs at least two machines.

Meta AI chief Alexandr Wang is quoted in the sources saying Meta lacks control or visibility into users’ Apple hardware setups used to run the model. Outlets vary in emphasis: some foreground the headline exchange between Zuckerberg and Apple’s chief, while others highlight the technical architecture and the claim that the model can run on desktop Apple hardware.

All of the supplied reporting agrees that Meta published the Muse Spark Open weights and described the model as able to run on Mac Studio hardware with M5 Ultra chips, and that Zuckerberg publicly thanked John Ternus with a reciprocal congratulatory reply from Ternus.

The release and the public thanks signal that Meta is making a large-scale model available as open weights and that Apple’s latest desktop silicon is being presented as capable of hosting such models locally.

What is confirmed in the reporting is the publication of the weights, the model specifications Meta released, the stated hardware configuration for local operation and the public exchange between Zuckerberg and Ternus; licensing terms and cost estimates for required hardware are reported but differ in detail across outlets.

This episode illustrates how large AI model releases are being coupled in the sources with attention to on-device performance and distribution choices.