Meta's Muse Spark Open fits on two Mac Studios

Open weights at this size used to mean renting a rack. Meta says this one fits on two desks.

Meta released Muse Spark Open on Oct. 15. It is a mixture-of-experts model with 816 billion total parameters and 42 billion active per token, and a 1 million token context window. It takes text, image, audio and video in and puts text out, according to Hugging Face. Meta calls it "the most capable US open-weight model to date." That is Meta's claim, and so far it comes without a scorecard.

Two Mac Studios is the number that matters

Meta says the model runs on two Mac Studios with M5 Ultra chips using Apple's MLX framework, and calls it the first model in its class to do so. Each machine's GPU can use up to 512 GB of unified memory, so the pair works like one computer with about a terabyte.

Two machines.

Hugging Face's launch post gives the smaller setups. A 3-bit MLX quant used about 344 GB on a 512 GB M3 Ultra and generated about 21 tokens a second. A 2-bit llama.cpp quant used 304 GB and ran about 18 tokens a second. Even the 3-bit build needs more than 256 GB of RAM, Hugging Face says.

Both posts pitch it at coding and agent work. Meta says it tested on-device workflows with its Muse Code tool, including changes to the open-source Mastodon client Ice Cubes and driving DaVinci Resolve end to end.

The weights are everywhere, the terms are not

Yes, another giant open model. This one is different because it is already where you would want it. Meta lists API access on Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI, and weight downloads on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models.

Read the license before you build on it. Meta ships the model under the Muse Model License, which it describes as having "limited commercial terms" that let many developers use it for free. Hugging Face's post does not state a license at all.

The launch posts leave out two more things. Neither publishes benchmark scores: Meta links an evaluation methodology report and says a full evaluation report is "available soon." And reasoning cannot be switched off. Hugging Face says effort runs from low to xhigh, with xhigh the default. You pay for that in tokens and latency on every simple call.

Earlier this week TopFive covered Muse Wonder, which shipped with no benchmarks or prices. Meta is repeating the pattern.

What to do with it this week

If you run models locally, this is the open-weight model to try at the top end. One 512 GB Mac and a 3-bit quant gets you a working copy, by Hugging Face's numbers. Two M5 Ultras get you the full model, by Meta's.

If you call models through an API, add it to the list of coding models you test against what you use now, on your own tasks. Meta has not published its numbers, so yours are the only ones that count.

If you plan to ship a product on it, get the license read first. "Limited commercial terms" usually ends in a threshold.

What to watch

Watch for Meta's full evaluation report and the license text. Then watch the independent coding results on OpenRouter and Hugging Face as developers run it.

Until then, test it yourself. Then decide.