Can You Run Muse Spark Open Locally? AI Hype vs Reality

Only on a computer with hundreds of gigabytes of memory. Muse Spark Open needs about 493 GB in the 4-bit format Apple ran it in, and even its smallest 1-bit version needs 180 GB. My MacBook Pro has 16. Meta released it on Thursday as “the most capable US open-weight model to date” and “the first model in its class to run on your desktop”. This post is about that desktop and the US in that sentence. And it’s about why I think AI launches like this one keep promising more than they deliver.

Can you run Muse Spark Open locally?

On Thursday I asked Claude Code in my terminal the obvious question: “can i run muse spark open on this mac”. The answer was a pretty firm no!! This Mac is a 2021 MacBook Pro with an M1 Pro chip and 16 GB of memory.

Muse Spark Open runs locally on Macs only if you join 2 Mac Studios with 512 GB of memory each, or 4 with 256 GB each. That’s from Meta’s own guide for developers. Apple’s machine learning team wrote up how they did it with MLX (Apple’s open source framework for running models on its own chips). They’re refreshingly blunt about it: “the model cannot fit on a single machine.”

And the 512 GB Mac Studio isn’t even on sale yet!! VentureBeat checked: “Apple’s 512GB-memory configuration doesn’t arrive until late October.” So the desktop in Meta’s headline is one nobody can buy, at a price nobody can tell you. What you can buy today is 4 of the 256 GB models:

Price before tax

$43,196

Tokens a second

37

A token is roughly a word and that’s faster than anyone can read. It’s also more than twice the price of the $18,000 AI server everyone on X was arguing about this week. I worked out that one needs about 9 months to cost less than 10 seats of a $200 cloud plan. And it’s the cheap one.

Simon Willison, who tries new models the day they come out, couldn’t run it either. On the launch thread he wrote “I had to run them via OpenRouter - I don’t have access to a machine that can handle this size of model.”

Meta’s small model, Muse Glimmer, is the one meant for computers like mine. LM Studio says even that one needs “at least 26 GB of RAM”. So no luck there either!

How much RAM do you need to run Muse Spark Open?

Muse Spark Open needs about 493 GB of memory at 4-bit, more than 1.6 TB at full precision and 180 to 212 GB for the smallest 1-bit version. It’s big because it’s built big: an open-weight large language model (LLM) with 816 billion parameters, 42 billion of them active for each token, and a context window of 1 million tokens. In a mixture of experts (MoE) model like this one, every layer holds 256 small expert networks and picks 8 of them for each token (Hugging Face’s write-up has the details). That’s how a model this big still answers at a usable speed: most of it sits idle on any given word. Idle experts still take up memory though (all 256 of them, in every layer).

Full precision, BF16

1,600 GB

4-bit, NVFP4

493 GB

Smallest, 1-bit

180 GB

My MacBook Pro

16 GB

Full precision is BF16 (brain floating point, the 16-bit format the model was trained in), where Apple says the weights alone take “more than 1.6TB of RAM”. Quantization shrinks that by storing each number in fewer bits. In NVIDIA’s 4-bit floating point format (NVFP4) the model takes “about 493GB” and Apple’s team still had to split it across 2 machines because of the memory it needs while it’s working. The 1-bit version comes from Unsloth, who squeeze models down so they fit on smaller machines. It’s still more than 11× my laptop.

Is Muse Spark Open really open source?

Muse Spark Open is open weights and not open source. You can download its weights for free under Meta’s own Muse Model License. To download it you give Meta your “full legal name, date of birth, and full organization name”. Anything you build with it has to say “Built with Muse Spark Open”. A company making more than $20 million a year that offers it as a service has to “notify Meta and enter into a separate written agreement” (all of it is on the model card). Open weights means you get the finished model to run yourself (the data and the code that made it stay with Meta).

Meta can ship an open source license when it wants to. Muse Glimmer, the small one, is under Apache 2.0. And Muse Spark Open is still more open than the model it’s named after: in 2024 Mark Zuckerberg wrote “Meta is committed to open source AI.” In April this year Muse Spark became its first frontier model without open weights.

Is Muse Spark Open really that good?

Muse Spark Open is very good for an open-weight model. Meta’s claim only covers open-weight models from American companies (a much smaller field). Artificial Analysis ranks it #8 of 117 open-weight models in its size class, with 39 on its Intelligence Index (the median for that class is 19 so it’s good!). Meta’s own model card puts it behind Kimi K3 from the Chinese lab Moonshot AI on both of the coding benchmarks it reports:

Terminal-Bench 2.1, Kimi K3

85.0

Terminal-Bench 2.1, Muse Spark Open

76.4

DeepSWE 1.1, Kimi K3

68

DeepSWE 1.1, Muse Spark Open

60

VentureBeat looked at the same table and called it “a contender with specific strengths, rather than an overall open-weight leader.” That’s fair! It’s just not what “the most capable US open-weight model” sounds like at a glance. Alexandr Wang, Meta’s chief AI officer, even ended his launch post with “america! 🇺🇲”.

Meta has been here before. When it launched Llama 4 in April 2025, the announcement led with “an experimental chat version scoring ELO of 1417 on LMArena”. LMArena said afterwards that “Meta should have made it clearer that ‘Llama-4-Maverick-03-26-Experimental’ was a customized model to optimize for human preference.” The biggest Llama 4, Behemoth, was “still training” at launch. A month later Axios reported, citing the Wall Street Journal, that Meta had delayed it over “concerns that it may not be enough of an advance.”

Is AI overhyped?

A lot of AI launches are. The tools work, but launch posts keep promising more than the product does on the day you get it. I said this on stage at React fwdays in 2024, about my then brand new iPhone 16 Pro Max that still didn’t have Apple Intelligence in Europe:

there’s devices like the R1, the Rabbit R1 that’s claiming to do a bunch of things and then I get it and it doesn’t do that. I even feel that with Apple. I was promised Apple Intelligence…

A few seconds later I said “there’s a lot of overpromise, underdeliver and just AI AI AI everywhere and it’s really exhausting.” Here’s that bit of the talk:

Apple said in March 2025 that “It’s going to take us longer than we thought to deliver on these features”. When I finally used Apple’s on-device model in my own app this year, it refused a voice note of a normal conversation with my wife as “unsafe”. Humane raised more than $230 million for its AI Pin, sold its assets to HP for $116 million and switched every Pin off on February 28, 2025. A couple of weeks ago the US government started calling AI “Super Intelligence” without changing its definition.

Marques Brownlee reviewed the Rabbit R1 and the AI Pin in 2024. In his Rabbit R1 review he called AI products “the apex of this horrible trend where the thing that you get at the beginning is like borderline nonfunctional compared to all the promises”. His advice near the end is the best test I know for a launch like Thursday’s: “buy the product based on what it is today and not what it’s promised to be in the future.”

Applied to Muse Spark Open, his test says it’s a very good open model that runs on hardware almost nobody has today (and through an API for everyone else). That’s a fine thing to ship! It’s just a different thing from a model that runs on your desktop.

People believe the promise even after it doesn’t land. In a 2025 study by Model Evaluation and Threat Research (METR), experienced open source developers expected AI to speed them up by 24%. They took 19% longer with it. Afterwards they “still believed AI had sped them up by 20%”!! METR has since marked those results “out of date” and its 2026 update says developers are likely faster with AI tools now. Good! I still keep thinking about the gap between how fast it felt and how fast it was.

AI should be like salt

Right after the Rabbit R1 bit in that talk I said “AI features should be like salt”. You “sprinkle them on your products and you can create true beauty and flavor”, but if you drown a product in it “people get sick”. Muse Spark Open is a lot of good engineering, from 256 experts in every layer to the MLX work that makes 4 Macs behave like one computer. I just wish Meta had launched it the way Apple’s engineers wrote it up, as a great model that doesn’t fit on one machine.