Can You Run Muse Spark Open at Home? The Real Hardware It Takes

On October 15, 2026, Meta did something unusual for a frontier AI lab: it gave away the weights to an 816-billion-parameter model. Muse Spark Open — the open-weight sibling of the Muse Spark model family behind Meta's AI products — is free to download from Hugging Face right now. The catch, and it is a big one, is that downloading it and running it are two very different things.

So: can you actually run Muse Spark Open at home? Here is the real hardware math, the three ways to use the model, and what each one costs in practice.

What you are downloading: the specs

Muse Spark Open is a mixture-of-experts (MoE) model: 816 billion total parameters, of which only 42 billion are "active" for each token (256 experts per layer, 8 active per token). That design keeps compute reasonable — but here is the part people miss: all 256 experts still have to sit in memory, so deployment needs the memory footprint of the full 816B model, not the 42B active slice.

The headline specs from Meta's model card:

• Parameters: 816B total / 42B active per token (Hugging Face lists it as 816B-A42B)
• Context window: 1 million tokens
• Input: text, images, video, and audio in; text out (natively multimodal, with a separate 1.8B-parameter perception encoder)
• Built for: coding and long-horizon agentic work — multi-step tasks, tool use, fixing real bugs
• Checkpoints: full-precision BF16 (~1.6TB) and compressed NVFP4 (~530GB)
• License: the custom Muse Model License — free to download with limited commercial terms, but not an OSI open-source license like Apache 2.0. Check the license text before any commercial deployment.

Meta calls it "the most capable US open-weight model to date" and publishes its own benchmark results on the model card — including a top score of 89.2 on MCP-Atlas. Those are Meta's tests; independent results will tell us more about everyday coding performance, where it currently trails models like Kimi K3 and Qwen 3.8 on some coding benchmarks.

Route 1: Run it locally (the two-Mac-Studio setup)

This is the headline demo, and it is genuinely remarkable — with a catch. Meta worked directly with Apple (a notable thaw in a frosty relationship) to optimize the model for Apple's MLX framework. The reference configuration:

• Two Mac Studios with the M5 Ultra chip, 512GB of unified memory each, linked over Thunderbolt 5 — behaving, in Meta's words, like "one supercomputer with a terabyte of unified memory."
• Apple's own ML research post gives the memory math: in native BF16 the model needs more than 1.6TB for parameters alone; in 4-bit NVFP4 about 493GB, plus working memory — which is why it cannot fit on one machine and needs two or more.
• Apple reports the 4-bit NVFP4 build generating around 15–25 tokens per second on two machines, and over 36 tokens per second on four.

Now the practical problems. First, the 512GB Mac Studio was not shipping yet — Apple announced the M5 Ultra Mac Studio on August 25 at a starting price of $5,499, with the 512GB memory option "coming in late October," and as of mid-October the US store configurator still stopped at 256GB. The buyable version of Meta's recipe was therefore the four-machine 256GB configuration, which independent coverage priced at roughly $38,000 (an estimate — check current Apple pricing).

Second, community members have produced smaller quantized builds (down to ~2-bit) that reportedly run on a single 256GB machine — but those are heavily compressed versions, closer to a floor than to the experience Meta demonstrates. For almost everyone reading this, local inference is a spectator sport.

Route 2: Download the weights (free, for builders)

If you are a developer or a team that wants control of your data, the weights themselves cost nothing. They are available on Hugging Face (gated — manual approval required), plus Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models. The license allows limited commercial use, with two strings attached worth knowing: products built on it must display "Built with Muse Spark Open," and inference companies above $20M in revenue need a separate agreement with Meta (per independent readings of the license — verify the current text yourself).

Meta also released a smaller sibling, Muse Glimmer, back in August under the permissive Apache 2.0 license — the genuinely "just use it" option if Spark Open's license or size gives you pause.

Route 3: Use a hosted API (the practical choice)

For everyone else, day-zero launch partners offer Muse Spark Open as a hosted API: Together AI, Baseten, Fireworks AI, Groq, Modal, Nebius, and OpenRouter, with both serverless and dedicated options at Together AI. Meta's launch materials do not publish per-token prices — rates depend on each provider — but this is the only route where you get the model's agentic coding capabilities without buying a server rack.

This is also, not coincidentally, how most people will actually experience Meta's agent technology: not by self-hosting weights, but through the Muse personal agent app itself (iPhone, Android, Mac, web at muse.ai, and inside WhatsApp), where the models do the work on Meta's cloud computers. If you want the agent rather than the weights, that is the product to set up — currently US and Canada, 18+, with a free tier. A guided Muse account registration assistance covers signup for personal use, within the platform's terms.

The bottom line

Muse Spark Open being free to download is real and significant — open weights at this capability level, with an Apple-optimized local path, is genuinely new. But "open" never meant "runs on your laptop." Until memory gets dramatically cheaper or smaller, the honest summary is: download it if you build, API it if you use, and admire the two-Mac-Studio demo from a respectful distance.

FAQ

What hardware does Muse Spark Open need?
Meta's reference setup is two Apple Mac Studios with M5 Ultra chips and 512GB of unified memory each (about 1TB total), running the 4-bit NVFP4 checkpoint through Apple's MLX framework. The model card asks for at least ~700GB of GPU memory for NVFP4 or ~2TB for full-precision BF16. Community 2-bit quantized builds reportedly run on a single 256GB machine.

Is Muse Spark Open free?
The weights are free to download under Meta's custom Muse Model License, which allows limited commercial use — but it is not an OSI open-source license. Products built on it must display "Built with Muse Spark Open," and large inference providers need a separate agreement. Check the license text before commercial use.

Where can I download Muse Spark Open?
From Hugging Face (gated, requires manual approval), Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models. Hosted APIs are available from Together AI, Baseten, Fireworks AI, Groq, Modal, Nebius, and OpenRouter.

Can Muse Spark Open run on a single Mac?
Not at useful quality with official checkpoints — Apple's own testing shows even the compressed NVFP4 build needs two or more Mac Studios once working memory is included. Heavily compressed community builds (~2-bit) can run on a 256GB machine, but with reduced quality.

How is Muse Spark Open different from Muse Wonder?
Both were announced days apart in October 2026. Wonder is Meta's closed frontier model (API-only, coming to the Muse app for subscribers). Spark Open is the open-weight release — downloadable, self-hostable, and aimed at developers building coding and agent applications.