Apple and Meta Collaborate to Run 816B-Parameter Muse Spark Open Locally on Mac Studios

Key points

• Meta released Muse Spark Open, described as the most powerful U.S.-developed open-weight AI model, with 816 billion parameters and 42 billion active parameters in its mixture-of-experts design.
• The model runs locally on clusters of two to four Mac Studios with M5 Ultra chips, utilizing Apple’s MLX framework and RDMA over Thunderbolt for distributed inference.
• Performance benchmarks show the model generating over 36 tokens per second on a four-machine cluster, exceeding average human reading speed, while consuming only about 300 watts of power per machine.
• The collaboration was facilitated by Daniel Gross, a former Apple executive now at Meta, and represents a détente between Apple and Meta, whose leaders previously clashed over privacy and advertising.
• Meta retains no control or visibility over user data when the model runs locally, addressing historical privacy disputes between the two companies.
• A Mac Studio with the required M5 Ultra chip and 256GB of memory costs at least $10,799, with a 512GB version releasing later this month.

Background

This collaboration follows recent tensions between Apple and Meta regarding AI agent privacy. In early October 2026, Apple introduced new controls for macOS Full Disk Access to prevent AI agents from accessing sensitive data without explicit consent, following a controversy where Meta’s Muse agent allegedly accessed user messages. Additionally, Apple had recently partnered with Google to integrate Gemini models into Apple Intelligence, while Meta’s Muse AI app had reached the top of the Apple U.S. App Store in September 2026, driving a 12% stock surge for Meta.

How outlets are covering it

Axios emphasizes the symbolic nature of the Apple-Meta détente, highlighting the 'strange bedfellows' dynamic in the AI race, while noting that the companies still compete in other areas. CNBC focuses on the technical and business implications, noting that the collaboration challenges OpenAI and Anthropic and signals Apple’s growing strength in AI chips, while also mentioning that Apple’s AI strategy has faced investor criticism. Apple’s technical documentation provides detailed benchmarks, showing that the M5 Ultra’s GPU Neural Accelerators and high memory bandwidth enable efficient local inference, with specific performance metrics for time to first token and decode speed. Meta’s announcement frames the model as a frontier capability for agents, emphasizing its open-weight nature and commercial license.

Why it matters

The Apple-Meta collaboration demonstrates how the AI race is reshaping traditional Silicon Valley rivalries, forcing competitors to cooperate when interests align. It provides a proof of concept for running large, powerful AI models locally on consumer hardware, potentially reducing reliance on cloud services and proprietary models from companies like Nvidia. This could shift the economics of AI by giving companies more control and alternatives, while also addressing privacy concerns through local execution. However, it raises concerns about the potential misuse of powerful open-weight models by bad actors.

What to watch

Observers will watch whether this collaboration marks the beginning of a broader relationship between Apple and Meta, such as Apple incorporating Meta’s models into Apple Intelligence. The availability of a more expensive Mac Studio version with 512GB of RAM later this month will also be key, as will the adoption of Muse Spark Open by developers and its impact on the competitive landscape for AI models.