The India Decade

Meta’s Muse Spark Open runs on Apple’s M5 Ultra Mac Studios

Meta has open‑sourced its 816‑billion‑parameter Muse Spark Open model and, together with Apple, made it run locally on two to four M5 Ultra‑equipped Mac Studios.

By The India Decade

Mac Studio with peripherals in an Apple Store setup.
Mac Studio with peripherals in an Apple Store setup. · “Mac Studio with Studio Display, Magic Keyboard, and Magic Trackpad in Apple Store” by Dancingpolishcow (CC BY-SA 4.0) via Wikimedia Commons

Meta releases Muse Spark Open for on‑device use

Meta announced on October 15 that its new Muse Spark Open model – 816 billion parameters in total, with about 42 billion active per token – is now available for download on Hugging Face. The company describes it as “the most capable U.S. open‑weight model to date.”

Apple’s hardware makes it possible

Apple’s latest Mac Studio, powered by the M5 Ultra chip, provides up to 512 GB of unified memory per machine. Meta says two such machines can share a full terabyte of memory, effectively acting as a single super‑computer. The MLX framework distributes the model across two to four Mac Studios, allowing the entire weight set to stay on‑device.

What the setup costs

A base‑configuration M5 Ultra Mac Studio with 256 GB of memory starts at US$10,799. Running the full model across two machines therefore costs a little over $20,000, while a four‑machine configuration can exceed $40,000. In Canada the same base unit is listed at CAD 7,799, putting a two‑machine rig at roughly CAD 15,600 before taxes.

Performance claims

Meta’s benchmarks claim Muse Spark Open outperforms leading Chinese models (GLM 5.3 and Kimi K3) on agentic tasks and visual reasoning. In a X post, Mark Zuckerberg wrote, “Muse Spark Open — weights releasing today — runs great on new Mac Studios thanks to work with @johnternus and Apple.” Apple’s chief John Ternus replied, “Congrats to the team! Amazing to run a model of Muse Spark Open’s caliber and scale this efficiently, right from your desk.”

Why the partnership matters

For Apple, the demo validates the M5 Ultra’s unified‑memory architecture and its MLX framework as a platform for frontier AI. For Meta, offering an open‑weight model that can run locally gives developers a cloud‑free alternative and puts the company’s work on hardware that many creators already own.

Open‑weight vs. closed models

Meta also unveiled Muse Wonder, a more capable but closed model available only through its API. Muse Spark Open is the version developers can freely download, modify and run, subject to Meta’s licensing terms.

Looking ahead

Both companies said the collaboration does not signal a broader partnership, but the move highlights a growing trend toward on‑device AI that sidesteps cloud‑centric compute. Analysts expect more developers to experiment with large models locally, especially where data privacy and latency are paramount.

Key numbers

WhatFigureSource
Model size816 billion total parameters, ~42 billion active per tokenMeta press release
Context window1 million tokensMeta documentation
Base Mac Studio price (US)$10,799Apple product listing
Minimum two‑machine cost (US)just over $20,000Gadget Review analysis
Maximum four‑machine cost (US)over $40,000The Tech Portal report
Unified memory per M5 Ultra Mac Studioup to 512 GB (two machines give 1 TB)Meta statement
Weight download locationHugging Face repositoryMeta announcement

In this story

NameRole in the story
Hugging FaceHost for model weight download
Mac StudioApple desktop used for running the model
Apple Inc.Provider of M5 Ultra Mac Studio hardware
Meta PlatformsDeveloper of Muse Spark Open
M5 UltraApple silicon chip powering Mac Studio

Topics

The day’s reporting, once each morning. No advertising.

More on this story