Muse Spark Open: is it free, and can you run it on a Mac?

Muse Spark Open is free to download and use under Meta's Muse Model License, with a few strings attached, and hosted versions cost about $1.25 per million input tokens and $4.25 per million output tokens. You can run it on a Mac, but the full model needs two 512GB M5 Ultra Mac Studios linked together. A compressed version fits on one machine with 256GB or more of memory.

Meta released the weights on October 15. @finkd, Mark Zuckerberg's account, announced it with a line aimed at Apple fans: it "runs great on new Mac Studios thanks to work with @johnternus and Apple." So Apple worked on the Mac support with Meta directly.

What Muse Spark Open is

According to Meta's announcement, it is a mixture-of-experts model with 816 billion total parameters, of which 42 billion are active for any given token. It has a 1 million token context window, reads text, images, video and audio, and answers in text. Meta built it for coding and agent work: long multi-step tasks, software engineering, tool use. It is not Muse Wonder, the closed model Meta launched earlier this week. Meta says Wonder is the stronger of the two. Spark Open is the one you can download.

Is Muse Spark Open free?

The weights cost nothing. The model card on Hugging Face spells out the terms:

• You have to show "Built with Muse Spark Open" somewhere in your product, docs or settings.
• Companies with more than $20 million in annual revenue need a separate deal with Meta before they sell the model as an API to others. Internal use, and building it into your own product, aren't covered by that limit.
• Downloads are gated. The access form asks for your full legal name, date of birth and company name, and you can't edit it once it's submitted.
• Meta's usage policy applies, and it rules out things like unlicensed medical or legal advice.

For a small business that just wants to use it, that's free in practice. If you'd rather not host anything, Meta lists Together AI, Fireworks, Baseten, Groq, OpenRouter and others as API providers.

How good is it?

@ArtificialAnlys gave it 39 on its Intelligence Index, the highest of any open-weights model from a US lab. Chinese labs still lead, though. GLM-5.3 scores six points higher overall, and Artificial Analysis says that whole gap comes from agentic tests, where Muse Spark Open placed last of the nine top open models. It is also efficient. It used about 40,000 output tokens per test task, fewer than any of the other eight, and works out to roughly $0.84 per task at launch prices.

Meta's own table is more mixed. On MCP-Atlas, which tests how well a model uses tools through connectors, it scores 89.2, ahead of Kimi K3 and GLM 5.3 in Meta's table. On Terminal-Bench 2.1, a coding-in-the-terminal test, it trails both by seven points or more.

Can you run Muse Spark Open on a Mac?

Yes, if you have a lot of memory. At full precision the weights alone take about 1.6TB. Meta's own demo ran the model on two M5 Ultra Mac Studios with 512GB each, linked through Apple's MLX framework, so the pair acts like one machine with a terabyte of memory. Meta used that setup to fix bugs in an open-source iPhone app and review video footage in DaVinci Resolve without touching a data center.

For a single machine there's the compressed route. @UnslothAI posted "Dynamic 2-bit" builds that run on a Mac Studio, and its guide says one of those 2-bit builds fits in 256GB of RAM. Compression costs some quality, and Unsloth notes that longer conversations need extra memory on top of that.

My read

This is mostly a big deal for developers and for Apple. For most owners, the API price matters more than the Mac demo. At $1.25 and $4.25 per million tokens, Muse Spark Open costs more than the cheapest open models, and it isn't the strongest open model either. Its pitch is a US-built model with weights you can keep, tuned for tool use.

That matters if a client asks where their data goes. A law office or a dental practice could eventually run a model like this on hardware it owns, with nothing sent to a vendor. The hardware isn't cheap, but you can buy it at retail.

What it means for a business

The model is the easy part. The work is deciding which tasks to hand it, which data it can see, and who checks its output before it reaches a customer. If you want help sorting that out, New Face Design's free process audit goes through your week with you and points out the tasks worth automating, along with the ones that should stay on your own machines.