One Model, Two Computers: Meta Open-Sources Muse Spark Open, and the Entry Fee Is 1TB of Memory

One model, and it takes two computers just to hold it. Meta's newly open-sourced Muse Spark Open comes in at 800 billion parameters, which the company describes as "the most capable American open-weight model available today." But the real story isn't that self-assessment—it's the sheer scale of hardware needed to actually run it.

According to Meta, Muse Spark Open is the first model optimized specifically for Mac Studios equipped with the M5 Ultra chip. The key lies in Apple's unified memory architecture: the M5 Ultra's built-in GPU can tap into up to 512GB of memory on a single Mac Studio, and when you link two together, you effectively get a supercomputer with 1TB of unified memory. Meta made this point explicitly, stating that "two Mac Studios operate like a single supercomputer with 1TB of unified memory."

Mark Zuckerberg confirmed on X that this release involved collaboration with Apple's engineering team, specifically crediting Apple executive John Ternus: "Muse Spark Open—weights dropping today—runs incredibly well on the new Mac Studios, thanks to working with John Ternus and the team at Apple."

Apple has yet to announce pricing or an exact release date for the 512GB Mac Studio, only saying it will launch in "late October." If you instead opted for a setup of four 256GB M5 Ultra Mac Studios to run this model, the total cost would land around $38,000. This level of compute clearly isn't something the average developer—or so-called "vibe coder"—would consider. But for enterprises already burning through massive budgets on token costs, buying a few Mac Studios might actually turn out to be the more economical option—assuming Muse Spark Open's real-world performance lives up to the numbers Meta has published.

Benchmarks released by Meta show the model outperforming two Chinese models, GLM 5.3 and Kimi K3, on agentic tasks and visual reasoning. Meta also reiterated its stance that the U.S. should not cede ground to Chinese labs in the race for open-source AI leadership.

The weights for Muse Spark Open are now available for download on Hugging Face.