Meta and Apple Collaborate to Run Muse Spark Open, an 816B-Parameter Model, on Mac Studio

In an unusual show of cooperation between two tech giants, Meta and Apple jointly announced on October 15 that the Muse Spark Open AI model—an 816-billion-parameter multimodal model—can now run on two Apple M5 Ultra Mac Studios using Apple's MLX framework. The announcement marks the first time a model of this scale is capable of running on local desktop hardware, a development with significant implications for engineers, researchers, and startups building on open-weight AI.

Meta CEO Mark Zuckerberg confirmed the collaboration on X, stating that the model "runs great on new Mac Studios thanks to work with [Apple CEO] John Ternus and Apple." Ternus replied: "Amazing to run a model of Muse Spark Open’s caliber and scale this efficiently, right from your desk. Mac Studio keeps pushing the boundaries of on-device AI!" Meta’s chief AI officer, Alexandr Wang, told Axios that the companies had been collaborating over several months, adding, "We’re excited for mommy and daddy to be playing nice again."

The Muse Spark Open model is described by Meta as a "natively multimodal model" with 816 billion total parameters that "excels at long-horizon tasks, complex software engineering, and multi-step reasoning." Meta claims it is the "most capable US open-weight model to date." Apple published a technical post detailing how MLX—an array framework for machine learning on Apple Silicon—supports Muse Spark Open along with two other open models.

This release follows a busy autumn for Meta's AI division. On September 2, Meta pushed out Muse Spark 1.3, a minor update to the model family. Then, on September 8, the company announced Muse, a "personal AI agent" that can handle tasks like online shopping, sending emails, and planning trips—all in the background, even after the user closes the app. The Muse agent rolls out for free on iOS, Android, and web in the US, with paid subscriptions planned but not yet detailed. Meta did not specify usage limits for free users or the exact pricing of future tiers.

The Muse Spark Open model is not explicitly tied to the Muse agent product, but it likely provides the underlying capability for such agents. For developers, the ability to run an open 816B-parameter model on two Mac Studios (which together cost roughly $20,000) opens up possibilities for fine-tuning, inference, and experimentation without relying on cloud GPU clusters. This could lower the barrier to entry for smaller teams and academic researchers.

However, caveats remain. Meta's claim of "most capable US open-weight model" is self-reported and has not been independently verified on standard benchmarks compared to models like Llama 3.1 405B or Mistral Large. The hardware requirement—two M5 Ultra Mac Studios—is far from a typical desktop setup, and the real-world performance for tasks like coding or reasoning has not been publicly benchmarked. Additionally, the Muse agent’s free tier limits and pricing remain undisclosed, leaving open questions about how accessible the consumer product will actually be.

For the technical audience, the key takeaway is clear: Meta and Apple are betting that on-device, open-weight models at hundreds of billions of parameters are viable. Whether that translates into practical advantages over cloud-based competitors like OpenAI’s GPT-5 or Anthropic’s Claude 4 will depend on how well the software stack (MLX) and hardware (Mac Studio) handle the workload in daily use. Independent benchmarks and third-party developer reports will be crucial in the coming weeks.