Can You Run Muse Spark Open Locally? Memory at Every Quantization

Meta's Muse Spark Open has 816 billion parameters, and you need roughly 260 to 300 GB of memory to run even its 2-bit version locally. It was the most upvoted story on Hacker News on October 15, 2026, and "runs on Macs" was in the headlines. That is true, but only for Macs with 512 GB of unified memory. This guide shows how much memory each version needs, why a model that runs only 42 billion parameters per token still needs all 816 billion in memory, and what to check before you download anything.

What Meta released

According to the official model card, Muse Spark Open was released on October 15, 2026. It is an open-weight model under the Muse Model License, which you must accept on Hugging Face before downloading the weights. The key facts:

The expert counts come from Hugging Face's launch post. "Open-weight" means you can download and run the weights yourself under the license terms. It is not the same as open source; read the license before building a product on it.

How much memory does Muse Spark Open need?

The rule of thumb for any model is parameters × bits per weight ÷ 8, in gigabytes when the parameters are in billions. Here is what that gives for 816 billion parameters at common precisions, before any context:

These are estimates from the formula, and the published figures line up with them:

• 2-bit: Hugging Face measured about 304 GB in use for a 2-bit build on a Mac with an M3 Ultra and 512 GB, running at about 18 tokens per second (launch post). Unsloth lists about 256 GB for its 2-bit version.
• 3-bit: the launch post says a 3-bit MLX version needs more than 256 GB of RAM.
• BF16: the launch post puts the original checkpoint at about 1.5 TB on disk.
• Server GPUs: Meta's model card says self-hosting needs at least 700 GB of VRAM for the NVFP4 checkpoint and 2 TB for BF16. Those figures include headroom for long contexts and serving many users at once, so they run higher than the bare weights.

You can redo any of these numbers, with your own context length and memory, in the LLM VRAM Calculator. The "816B MoE · 42B active" preset loads this model's size.

Can I run Muse Spark Open on my Mac or PC?

Compare your memory with the table above, keeping 10 to 20% free for the operating system and the context.

For most people the honest answer is no. That is not a flaw in Meta's release: frontier-sized models have always needed data-center hardware, and open weights make them available to anyone who has it.

Why 42 billion active parameters still need 816 billion in memory

Muse Spark Open is a mixture-of-experts (MoE) model. For each token, a small router picks 8 of its 256 expert networks, so only about 42 billion parameters do work on that token. That keeps it fast: per token, it reads about as much data as a 42B dense model.

The catch is that the next token may go to a completely different set of experts. Every expert has to stay loaded and ready, so memory follows the total parameter count, while speed follows the active count. The Hugging Face launch post makes the same point. What is mixture of experts? explains the design in more detail.

How long context changes the math

The million-token context window is the other memory trap. Every token in the context stores attention keys and values in the KV cache, which grows linearly with context length. In Hugging Face's tests, a server with eight RTX Pro 6000 cards ran the model with a 131,072-token context, while eight H200 cards had room for the full 1,048,576 tokens (launch post).

On a home machine, keep the context short, 8,000 to 32,000 tokens, unless you have memory to spare. In the calculator, open Model shape and enter the layer and head counts from the model's config.json to see the KV cache for any context length.

What to check before you download

1. Your memory. Look up your GPU's VRAM or your Mac's unified memory (Apple menu → About This Mac) and compare it with the table.
2. The quantization. Community builds come in GGUF (for llama.cpp-based apps) and MLX (for Apple silicon) formats at different bit widths. Pick the largest one that leaves headroom.
3. Your disk. The download is about the size of the weights: about 240 to 260 GB for 2-bit builds, more for larger ones. Make sure you have the space, and expect a long download.
4. The checksum. Hugging Face shows a SHA-256 hash for each large file. After a multi-hour download, check it with File Hash Multi or shasum -a 256 before loading the model.
5. The license. Accept the Muse Model License on the official repository. Third-party builds are still covered by it.

If your hardware falls short, a smaller open-weight model at Q4_K_M is the practical choice. An 8B model needs about 5 GB for its weights and a 70B model about 42 GB, and the calculator shows where your machine's limit is. The Data Storage Converter helps when a model page lists sizes in GiB rather than GB.

Common questions

Is Muse Spark Open free? The weights are free to download after you accept the Muse Model License on Hugging Face. Running them costs what your hardware costs; read the license for the terms on commercial use.

Can Muse Spark Open run on a 128 GB Mac? No. Even the 2-bit builds need about 256 to 304 GB, so a 128 GB machine has less than half the memory required.

What is the smallest version of Muse Spark Open? The smallest builds listed at launch are 2-bit quantizations, about 256 GB according to Unsloth. Every version keeps all 816 billion parameters; quantization only stores them in fewer bits.

Why is a mixture-of-experts model faster than a dense model of the same size? Each token only passes through 8 of the 256 experts, about 42 billion parameters, so the hardware reads far less data per token. All experts still have to be in memory.

How do I calculate VRAM for any LLM? Multiply the parameters in billions by the bits per weight and divide by 8 to get gigabytes for the weights, then add the KV cache for your context and about 10% overhead. The LLM VRAM Calculator does this for every quantization at once.