Meta's Muse Spark Open Runs on Macs, if You Mean Two M5 Ultra Mac Studios

Meta released the weights for Muse Spark Open on October 15, an 816-billion-parameter model that uses 42 billion parameters per token and accepts a context window of up to 1 million tokens, and it demonstrated the model running on two Apple M5 Ultra Mac Studios. The weights are genuinely downloadable, so this is a real open-weight release. But "runs on Macs" describes Meta's two-machine setup with roughly a terabyte of unified memory, built around a 512GB Mac Studio that is not shipping yet. For a normal home setup, nothing about what you can run has changed.

What was actually announced

Meta's announcement post calls Muse Spark Open "the most capable US open-weight model to date." That is Meta's own claim, and the benchmark comparisons in the post are Meta's tables. The model card on Hugging Face describes a mixture-of-experts transformer with 256 experts, 8 of them active for each token, and a 1M-token context at inference. VentureBeat reports that it takes text, images, video, and audio as input and produces text.

On hardware, the model card says a self-hosted run needs a GPU setup with at least 700GB of memory for the NVFP4 version and 2TB for full BF16 precision. Separately, it says the model is optimized for two M5 Ultra Mac Studios running Apple's MLX framework. Meta's post describes the pair as acting like one machine with about a terabyte of unified memory, each Mac contributing up to 512GB. Apple's Mac Studio page confirms the M5 Ultra tops out at 512GB with 1.2TB/s of memory bandwidth, and says the 512GB option is coming late October. It lists no price for it.

On benchmarks, Meta's own table has Muse Spark Open ahead of GLM 5.3, Kimi K3, and Qwen3.8 on JobBench (61.8) and MCP-Atlas (89.2), and behind them on several others, including Terminal-Bench 2.1 (76.4 against Kimi K3's 85.0). Its score of 39 on the Artificial Analysis Intelligence Index, as listed by Meta, sits below all three of those models. VentureBeat noted that an independent Artificial Analysis assessment of Spark Open was not available at publication, so treat all of these as Meta's figures for now.

The license is a custom one, the Muse Model License. It allows running, modifying, fine-tuning, and redistributing the model, with conditions: products built on it must display a "Built with Muse Spark Open" notice, and model-as-a-service providers above $20 million in revenue over any 12-month period must notify Meta and sign a separate agreement. There is also a usage policy, and the model is not intended for people under 18. The Hugging Face download is gated: you must accept the terms, and VentureBeat reports that the form asks for your full legal name, date of birth, and organization, handled under Meta's privacy policy.

What it means for you

The detail that matters most is that only 42 billion of the model's parameters are active for any one token. That makes generation faster than a dense model of the same total size would be, but every expert still has to sit in memory. The "active parameters" number does nothing to shrink the 700GB-plus requirement, which is the same reason the model finder judges whether something fits by file size and uses active parameters only for its speed estimate.

The file sizes are where the real picture is. A community GGUF conversion from ggml-org (not an official Meta release) lists 1.64 TB at full precision, 872 GB at 8-bit, 470 GB at 4-bit, and 304 GB at 2-bit, and says the 2-bit version fits in 512GB of unified memory. Unsloth's run-locally guide gives memory needs of 512GB for 4-bit, 346-410GB for 3-bit, 250-316GB for 2-bit, and 180-212GB for 1-bit, without saying whether those are measured or estimated. Neither source gives a tokens-per-second figure for any hardware, and I did not find one anywhere else, so how fast this actually generates on a Mac Studio is unknown right now.

My read, and this is analysis rather than reported fact: the heavily compressed versions are what make a single Mac Studio plausible, and compression costs quality, so Meta's two-machine setup is the one that shows the model at its best. On price, Apple has not announced one for the 512GB configuration. Our earlier Mac Studio coverage walks through the lineup, and VentureBeat puts a 256GB M5 Ultra at $10,799.

Does this change what this site recommends?

No. The first path is the same: install Ollama, pull a small model, and go, as covered in choosing your first model. For nearly everyone reading this, an 816-billion-parameter model is a headline, not a download. The models in the finder are the ones sized for machines people actually own, and the reasons to run AI locally do not depend on the biggest model in the world fitting on your desk.

If you are considering a Mac Studio specifically because of this release, I would wait. Independent evaluations have not landed, there is no published speed number, and the 512GB configuration has not shipped. The buying side is covered in the Mac Studio pick.

What to actually do now

All optional, and none of it urgent:

• Check what fits the machine you already have. The model finder takes your RAM and graphics card and returns models that fit, with the exact command to pull each one.
• If you already own a 256GB or 512GB Mac and want to try it, read the license first, expect to use one of the smaller quantizations, and treat every quality claim as unverified until independent numbers exist.
• If you are weighing a purchase on this release alone, wait for the 512GB Mac Studio to ship and for independent speed and quality tests before spending anything.