Running Meta’s Muse Spark Open on Mac Studio: Performance Benchmarks

Muse Spark Open Performance Benchmarks

Muse Spark Open is an open-weight model announced by Meta on October 15, 2026, which the company describes as the most capable model of its kind in the United States. It features 816 billion total parameters with 42 billion active parameters and a 1 million-token context window.

The model is designed specifically for coding and agentic work. Meta reports that it excels at software engineering, multi-step reasoning, and extended tasks.

Running Muse Spark Open on Mac Studio: Real-World Performance

This partnership marks the first public collaboration between Apple and Meta since approximately 2017. The relationship has historically been defined by friction, including public criticism between leadership and disagreements over Apple’s 2021 privacy changes. Because the companies frequently argue rather than work together, the joint effort on Muse Spark Open is considered a surprise. Meta AI chief Alexandr Wang commented on the situation, stating, "We're excited for mommy and daddy to be playing nice again".

The collaboration lasted for months and involved former Apple executive Daniel Gross, who played a key role in bringing agent capabilities to the desktop. Meta CEO Mark Zuckerberg thanked Apple CEO John Ternus for the work, noting that the model runs effectively on new Mac Studios. Ternus responded by congratulating the team and highlighting the efficiency of running a model of this caliber directly from a desk.

This work targets the M5 Ultra Mac Studios. While Apple has not yet released the 512 GB Mac Studio, it is expected to arrive later this month. The hardware maintains the same 7.7-inch square, 3.7-inch tall aluminum design used for four generations. This development coincides with reports that Apple has shelved its plans to launch an AI pendant.

Technical Challenges of Running Local AI on Apple Hardware

You can run the 816-billion-parameter Muse Spark Open model by linking multiple Mac Studio units together. According to macdailynews.com, this configuration requires between two and four Mac Studio machines equipped with M5 Ultra chips and high amounts of unified memory. Meta worked with Apple to bring this model to the desktop, utilizing the latest MLX framework for the process.

The core of this performance is Apple’s unified memory architecture. Each M5 Ultra chip allows the integrated GPU to access up to 512 GB of unified memory per Mac Studio. By combining two of these machines, you create a system that acts like a single supercomputer with one terabyte of unified memory, which is enough to hold the entire model. The chip’s 1.2 TB/s memory bandwidth is fast enough to read the small slice of data required for each token during inference.

To manage these multiple machines, Apple uses its MLX framework to optimize performance for distributed inference. As detailed in the Apple Machine Learning Research paper, the system relies on local and distributed Large LLM inference using MLX and RDMA. This setup is designed to handle the mixture-of-experts architecture utilized by the Muse Spark Open model. While the Mac Studio was updated in March 2025 to include M4 Max and M3 Ultra chips with features like Thunderbolt 5 at 120Gb/s, this new collaboration focuses on the capabilities of the M5 Ultra hardware.

The Significance of the Meta and Apple AI Partnership

Building a workstation for this model requires a significant financial investment. According to MacDailyNews, a Mac Studio configuration with an M5 Ultra chip and at least 256GB of memory starts at $10,799. WorldProgramming reports that 256GB Mac Studio models are priced over $10,000.

If you are considering lower-tier hardware, be aware that Apple increased prices for its machines in June 2026. The entry-level Mac Studio M4 Max saw a price increase from $1,999 to $2,499.

What are the performance benefits for developers?

The hardware architecture provides specific benefits for intensive development tasks. The M5 Ultra chip supports up to 512GB of memory and provides a data transfer speed of 1.2TB/s.

According to a review by Macworld’s Roman Loyola, the most notable upgrade in the new Mac Studio is significantly faster storage speeds rather than the chip itself.

Accessing Muse Spark Open Without Local Hardware

You do not need local hardware to use Muse Spark Open, as the model is available through several cloud platforms. You can access the model via APIs from Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI.

Alternatively, you can download the model weights from Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and Microsoft Foundry Models. The model is free to download and includes a license that allows for limited commercial use.

The Future Roadmap for Meta’s AI Hardware Integration

The company purchased hundreds of Macs earlier in 2026 for its own internal AI experimentation.

According to Fast Company, Meta and Apple share a common interest in the potential of open-weight AI. Mark Zuckerberg also announced a new handheld AI companion called the Muse Charm, a puck-sized device featuring an interactive avatar, during the company's annual Connect event.