Can You Run Muse Spark Open on a Mac? Memory, Speed and Cost

Yes, with enough memory. Meta released Muse Spark Open on 15 October 2026: 816B parameters, of which 42B are active for each token, a 1M token context, and open weights on Hugging Face. The 4 bit version needs about 493 GB, and Apple ran it on two M5 Ultra Mac Studios with 512 GB each at 25 tokens per second. One Mac Studio with 256 GB holds the 2 bit version. Below: memory for each version, which Mac fits, measured speeds, and what it costs to try. Figures checked 16 October 2026.

How much memory Muse Spark Open needs

The full BF16 weights are over 1.6 TB, so nobody runs the model unquantized on a desk. Unsloth publishes GGUF versions down to 1 bit; the ranges below are the total memory each one needs, from the Unsloth guide. Apple measured the NVFP4 version it tested at about 493 GB.

Longer context needs more memory on top of these figures. If you want the same arithmetic for any other model, the Mac LLM calculator runs it for every chip and memory size.

Which Mac runs it, and how fast

Apple tested the model with MLX on M5 Ultra Mac Studios linked by RDMA over Thunderbolt 5, at the 4 bit NVFP4 size. Generation speeds from Apple’s Machine Learning Research post of 15 October 2026, with an 8,192 token context:

An 816B model runs at that speed because it is a mixture of experts: each new token reads only the 42B active parameters, about one twentieth of the weights. Apple published no single machine figures for this model and none for the M3 Ultra. Across the models it ran on both chips, the M5 Ultra generated tokens 1.6 times faster than the M3 Ultra and returned the first token up to 4.6 times sooner. The 512 GB M5 Ultra is not on sale yet; MacStories reports it is due at the end of the year.

Why a Mac, and not NVIDIA or AMD, for this model

For a model this size, memory capacity decides before speed does. The 4 bit version is about 493 GB. An RTX 5090 holds 32 GB, roughly one fifteenth of that. A DGX Spark or a Ryzen AI Max (Strix Halo) box holds 128 GB, about a quarter. One Mac Studio with 512 GB holds almost all of it, and two hold it with room for long context. Data center GPUs do the job too, at data center prices. Our Mac vs NVIDIA vs AMD comparison runs one speed model across all of them.

How to run it on a Mac

• MLX. Meta ships an MLX adapter for Mac Studio that pulls the weights from Hugging Face. Splitting the model over several Macs uses the MLX RDMA backend over Thunderbolt 5, which needs macOS 26.2 and MLX 0.30.1 or later; Apple used two cables per link.
• llama.cpp and Unsloth. The GGUF versions run on Metal, which llama.cpp enables by default on a Mac. The Unsloth command for the 2 bit version: unsloth run --model unsloth/Muse-Spark-Open-816B-A42B-GGUF:UD-IQ2_XXS
• Memory limit. On a 256 GB Mac running the 2 bit version, raise the wired memory limit with sudo sysctl iogpu.wired_limit_mb=N, where N is larger than the model in megabytes and smaller than the memory of the machine, as the MLX LM documentation describes. Keep the context short.

What it costs to try

Apple has not announced the price of the 512 GB M5 Ultra, and Apple used two of them. To try the model before buying, the shortest route is a machine that holds the 2 bit version: we rent the M3 Ultra Mac Studio with 256 GB from $530/mo billed annually or $857/mo billed monthly, a fixed price per machine with no per token fees.

Rent an M3 Ultra Mac Studio with 256 GB

License

The weights come under the Muse Model License, which Meta describes as limited commercial terms that let many developers use the model for free. Read it before you build a product on the model.

Sources

• Meta: Introducing Muse Spark Open, 15 October 2026
• Apple Machine Learning Research: local and distributed LLM inference with MLX and RDMA on Mac Studio, 15 October 2026
• Unsloth: Muse Spark Open, run locally guide
• MacStories on the release and the 512 GB M5 Ultra, 15 October 2026
• MLX LM documentation, large models