Will it fit?

On October 15, Meta released the weights of Muse Spark Open. It has 816 billion parameters. Meta’s guide runs it on two Mac Studios with M5 Ultra chips. How does a model that big fit on a desk?

A model is a long list of numbers. Each number takes some bits, and eight bits make a byte. So the memory a model needs is:

bytes = parameters × bits per weight ÷ 8

Count the weights, then the bits

Muse Spark Open has 816 billion weights. Every one sits in memory, even when a token does not use it.

At 16 bits (BF16), the formula gives 816B × 16 ÷ 8 = 1,632 GB. The published BF16 files are 1,644 GB. That is close.

At 4 bits it is less simple. NVFP4 stores each weight in 4 bits, plus one 8-bit scale for every 16 weights. That is 4.5 bits per weight. The formula gives 459 GB. The published files are 533 GB, or about 5.2 bits per weight. The model card does not say why the files are bigger.

Leave room

File sizes are in GB (10⁹ bytes). Memory sizes are in GiB (2³⁰ bytes). So 533 GB of files is 496 GiB.

One Mac Studio with M5 Ultra has up to 512 GB of memory, which is 512 GiB. That leaves 16 GiB for the system and the context. Meta’s guide says: “The model doesn’t fit in one Mac’s memory.” Two Macs give 1,024 GiB, so 528 GiB is left.

Try other models and machines below.

Why it is not slow

To write one token, the chip reads the weights it uses from memory. A dense model uses all of them. Muse Spark Open is a mixture of experts. It has 256 experts and sends each token to 8 of them. Shared parts, such as attention, run for every token. In total, 42 billion of the 816 billion parameters are active per token.

Our estimate: 42B weights at the file’s average 5.2 bits is about 27 GB per token. The M5 Ultra reads memory at 1.2 TB/s. So the ceiling is about 44 tokens per second. If all 816 billion weights were active, the chip would read all 533 GB per token: about 2 tokens per second.

This is a ceiling, not a benchmark. It assumes the active weights are stored at the file’s average 5.2 bits. It ignores the context cache, compute time and the Thunderbolt link. Splitting the work across Macs can change it: Meta says four Macs serve fastest. Real speed is lower. We did not measure it.

Total parameters decide if it fits. Active parameters set the speed limit.

Use it on the next model

• Size any model before you download it. At 4.5 bits, an 8B model is about 4.5 GB and a 70B model is about 39 GB. Then add room for the context.
• Check the real file. For Muse Spark Open NVFP4, the formula was 14% below the published size.
• Check speed claims. Divide memory bandwidth by the bytes read per token. If a claim beats that number, ask how.

The numbers, with sources

Bits per weight is the file size in bits divided by the parameter count. NVFP4 adds one scale per 16 weights, so expect 4.5. The Muse file averages 5.2, and the model card does not say why. GiB is 2³⁰ bytes, the unit memory sizes use. Hugging Face shows GB (10⁹), so 533 GB is 496 GiB.

Item	Size	Bits / weight	Source
Muse Spark Open, BF16	1,531 GiB	16.1	Hugging Face
Muse Spark Open, NVFP4	496 GiB	5.2	Hugging Face
Muse Spark Open, parameters	816B / 42B active	–	Meta
Meta’s stated minimum GPU memory	700 GB / 2 TB	–	Model card
Meta’s Mac setup	2 × 512 GB or 4 × 256 GB	–	Meta guide
Reflection Beam, BF16	933.81 GiB	16.0	Reflection
Reflection Beam, FP8	468.72 GiB	8.0	Reflection
Reflection Beam, NVFP4	273 GiB	4.7	Reflection
FLUX 3 Image Klein, BF16	11.2 GB	16.0	Black Forest Labs
FLUX 3 Image Klein, FP8	5.6 GB	8.0	Black Forest Labs
Mac Studio, M5 Ultra	512 GB, 1.2 TB/s	–	Apple
NVFP4 format	4 bits + 1 scale / 16	4.5	NVIDIA

Every claim and calculation is in the fact-check ledger.