Meta's Muse Spark Open needs 1TB of RAM to run on your Mac

By Sophie Lindqvist

Meta has released Muse Spark Open, an 800-billion-parameter open-weight model that the company calls "the most capable US open-weight model to date." Running it locally requires roughly 1TB of RAM.

The weights went live on October 15, 2026, and are available to download now from Hugging Face. Meta built the model for Apple Silicon, making it the first model optimized to run locally on Mac Studio machines equipped with Apple's M5 Ultra chip.

Why does this matter? Meta has said it believes the US should not allow Chinese labs to lead on open-source AI. Muse Spark Open is the company's concrete move in that fight, and Meta says its benchmarks show the model beating two leading Chinese models, GLM 5.3 and Kimi K3, on agentic tasks and visual reasoning. Open-weight models are the battleground where US and Chinese labs currently compete for developers, and Meta just raised the hardware stakes considerably.

How do two Mac Studios become one supercomputer?

The mechanism here is Apple's memory design. Most computers split memory between the CPU and the GPU. Apple's unified memory architecture lets both share the same pool, and that changes the math for running huge models.

Meta explains the setup plainly: "Apple's unified memory architecture lets the M5 Ultra's integrated GPU use up to 512 GB per Mac Studio, so two machines can hold an entire model." The company adds: "The result is two Mac Studios behave like one supercomputer with a terabyte of unified memory."

So the 800-billion-parameter model splits across two machines, each contributing 512GB. That's the entire trick — no data center, no cloud rental, just two high-end desktops stitched together.

Mark Zuckerberg announced the release directly, crediting cooperation with Apple: "Muse Spark Open — weights releasing today — runs great on new Mac Studios thanks to work with @johntus and Apple." (The post, dated October 15, 2026, names Apple executive John Ternus.)

The takeaway: unified memory turns consumer-adjacent hardware into a viable host for frontier-scale models, provided you can afford the memory.

What does it actually cost?

Apple has not announced specific pricing or a ship date for the 512GB Mac Studio. The company has only said it will be available in "late October."

There is a firmer number for the alternative configuration. Meta says Muse Spark Open can also run on four 256GB M5 Ultra Mac Studios, which will cost roughly $38,000.

That figure puts local inference out of reach for most individual developers. But the economics look different for organizations already spending heavily on API tokens. If Meta's performance claims hold up, a one-time $38,000 hardware purchase could undercut a recurring cloud bill — the model is open-weight, so there are no per-token charges.

The takeaway: this is a model priced for teams, not hobbyists.

How does it compare on benchmarks?

Meta's own benchmarks are the only evidence available so far, and they compare Muse Spark Open against GLM 5.3 and Kimi K3. The company reports it outperforms both on:

• Agentic tasks — work where the model plans and executes multi-step actions
• Visual reasoning — interpreting and reasoning about images

These are self-reported numbers, so treat them as a starting claim rather than a settled verdict. Independent evaluations will determine whether "the most capable US open-weight model to date" holds up against third-party testing.

The takeaway: Meta claims the crown among US open-weight models, but the claim rests on its own benchmark suite.

What's the bigger picture?

Muse Spark Open signals two shifts at once. First, Meta is doubling down on open weights as a strategic position against Chinese labs, matching capability claims directly against GLM and Kimi. Second, the Apple Silicon optimization shows a new route to local AI: instead of Nvidia data center GPUs, frontier-scale inference can now run on unified-memory consumer workstations — if you buy several of them.

The download is live on Hugging Face now. Whether the "two Mac Studios as one supercomputer" setup becomes a practical standard or a niche flex depends on independent benchmarks and on Apple's still-unannounced pricing for the 512GB Mac Studio arriving in late October.