Recently, Meta released the open-weight AI model Muse Spark Open and announced a partnership with Apple to optimize the model's local deployment for Mac Studio. The Muse Spark Open model features a total of 816 billion parameters, employs a sparse activation design, and activates approximately 42 billion parameters during each inference, supporting a context window of up to 1 million tokens, with a focus on programming, complex reasoning, and AI agent tasks.
Using two Mac Studio devices equipped with M5 Ultra chips and 512GB of unified memory each, users can run this large-scale model in a desktop environment without relying entirely on cloud AI services.
This collaboration is particularly noteworthy. Apple and Meta have previously had significant disagreements over privacy policies, platform rules, and application ecosystems, yet they are now jointly promoting the operation of open-weight models on Apple hardware. Meta CEO Mark Zuckerberg expressed gratitude to Apple CEO John Ternus for participating in the collaboration, while Ternus stated that Muse Spark Open can run efficiently on desktop devices, showcasing the potential of Mac Studio for local AI computation.
Meta and Apple's collaboration advances local deployment of AI models
Apple Muse Spark Open is the open-weight version of Meta's Muse model series. Unlike closed models that can only be accessed via remote services, open-weight models allow users to obtain model weights and deploy them independently under the terms of the license, providing developers, researchers, and enterprises with greater control over their operating environments.
On the hardware front, Meta indicated that the model has been optimized for Mac Studio and can leverage the unified memory architecture of Apple chips. Traditional discrete graphics card systems typically require model weights to be placed in video memory, while Apple's unified memory architecture allows the CPU and GPU to share the same memory pool. Two Mac Studio devices can operate collaboratively through the appropriate software framework, giving the system approximately 1TB of unified memory. Each device is equipped with an M5 Ultra chip that offers a maximum memory bandwidth of 1.2TB/s.
Apple's MLX machine learning framework supports model execution on Apple Silicon, and Meta has made corresponding adjustments for Muse Spark Open.
Item Specifications Processor M5 Ultra RAM 512GB unified memory Memory bandwidth up to 1.2TB/s
Disclosure: This article contains links to products from partner merchants. If you purchase through these links, TechRitual may receive a commission, but this does not affect product evaluations or recommendations. For more details, please see our privacy policy.