Running Meta's Muse Spark Open locally costs at least $21,598 in Mac Studios
Apple and Meta have enabled Meta's 816-billion-parameter Muse Spark Open model to run locally on clustered Mac Studios. CNBC reported that each 256GB M5 Ultra machine costs at least $10,799, so a two-machine cluster comes to $21,598 or more.
The Investor · Invest desk
Key figures: 512GB of pooled memory gives about 0.63 bytes per parameter; 1,024GB of pooled memory gives about 1.25 bytes per parameter; Muse Spark Open has 816 billion parameters.
| Figure | Value | Note | Claim |
|---|---|---|---|
| Memory per parameter, 512GB pooled, about | 0.63 bytes | 16 | |
| 1,024GB pooled, about | 1.25 bytes | 17 | |
| Model parameters | 816 billion | 1 |
What happened
- Meta AI chief Alexandr Wang said in an interview that many powerful open-weight models are too big for most personal computers, but Mac Studios offer enough memory to make Muse Spark Open practical.
- Developers can download a version from Hugging Face with no extra configuration for Apple silicon, and Apple published benchmarks and MLX documentation on splitting models across machines.
- According to Cryptobriefing, 42 billion of the model's parameters are active at a time, and its context window is 1 million tokens.
- Apple and Meta have had no public partnership since 2017; relations hit bottom after Apple's 2021 privacy changes, which Meta estimated in early 2022 would cut its revenue that year by $10 billion.
Why it matters
- What it costs cost A four-machine cluster costs at least $43,196 at the 256GB price, and the companies' promise of savings over cloud use was not set against any cloud bill.
- Who has to decide decision A team sizing a cluster must pick between more of today's 256GB machines and fewer of the 512GB version, which costs more than the 256GB model and is due later this month.
- What it limits constraint The weights carry a license that allows only limited commercial use, per Meta, and come from a company that has since centered its AI work on proprietary models, so a team building on them depends on Meta's later choices.
- Who's exposed exposure CNBC says the pairing could challenge OpenAI and Anthropic and could signal Apple's strength in AI chips, a market Nvidia has dominated so far.
Memory sets the size of the cluster, and the two sources describe it differently. Two 256GB machines hold 512GB [20], or 0.63 bytes for each of the model's 816 billion parameters, about five bits [16]. That is a ceiling: it assumes the whole 512GB holds weights, and anything the operating system or the model's context needs comes out of the same pool, so the precision of the downloadable version would decide whether two such machines are enough. Cryptobriefing describes two linked machines pooling nearly 1 terabyte [19]. Two 512GB machines would pool 1,024GB [21], or 1.25 bytes per parameter, twice the room [17].
Only 42 billion of the 816 billion parameters are active at a time, about 5.1% [18]. A cluster sized for memory fits that profile: roughly 95% of the weights wait on any one token [18].
On the alliance, we think the evidence supports a narrow version. Apple and Meta worked together for months [9] on one model and one chip family, and representatives of both declined to say whether it is the start of a deeper relationship [8]. CNBC calls the announcement a sign that the AI boom has shifted historical alliances in Silicon Valley [25]. Wang, the only executive quoted on the history, said, "Most of the animosity predates my tenure at the company, but obviously there's been a long history between Meta and Apple" [11]. He pointed to Daniel Gross, an AI researcher on his team and a former Apple director, as a reason for the collaboration [12]. A shared hire and a shared model is a thin base for the larger claim.
The counter-thesis is that this is the first of several joint releases, and it is cheap to test. A second Meta model tuned for Apple silicon, or a commercial agreement between the companies, would make this the start of something larger; a next release tuned first for other hardware would confirm the narrow view.
Local running also changes how the spend arrives. Cryptobriefing describes cloud-API vendors as forcing customers to send data out and pay per use [23]. A cluster is bought once, and the sale goes to Apple. The companies' claim that local running saves money [22] fails for any team whose per-use bill over the machines' working life would come to less than $21,598 [14].
What to watch
- Whether the 512GB Mac Studio ships later this month and at what price, which fixes the cost of a two-machine cluster.
- Whether Meta tunes a second model for Apple silicon, or the two companies announce a commercial agreement.
- Independent tests from developers pulling the weights from Hugging Face and running them on two-machine and four-machine clusters.
Information, not investment advice. This story is general information about companies and markets, drawn from cited reporting. It is not a recommendation to buy, sell or hold any security; check the cited sources before acting.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence42
- Adoption
- Insufficient
- Hype gap+30
- Incentives65
- Confidence55
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Muse Spark Open is an 816-billion-parameter open-weight model using a mixture-of-experts architecture; Meta says it is the most capable open-weight model released under a license that allows limited commercial use.
- [2]
Customers will be able to pair between two and four Mac Studio PCs to use Muse Spark to code, analyze video or do other agentic work.
- [3]
A Mac Studio with the required M5 Ultra chip and 256GB of memory costs at least $10,799.
- [4]
A more expensive Mac Studio version with 512GB of RAM, costing more than the 256GB version, will be released later this month.
- [5]
Meta AI chief Alexandr Wang said in an interview that many powerful open-weight models are too big to run on most personal computers, but Apple's Mac Studios make enough memory available that it is practical to run Muse Spark Open.
- [6]
Developers can download a version of Muse Spark Open from sites like Hugging Face without extra configuration for Apple's silicon; the model has been tuned to run faster on Apple hardware, and Apple revealed new benchmarks and published technical documentation about its MLX framework's ability to split open-source models across multiple machines.
- [7]
Muse Spark Open has 42 billion parameters active at any given time and a context window of 1 million tokens.
- [8]
Representatives from Apple and Meta declined to say whether the Muse Spark Open collaboration was the start of a deeper relationship.
- [9]
The announcement followed months of collaboration between Apple and Meta.
- [10]
Apple and Meta have not struck any public partnership since 2017, when Apple removed system-level Facebook integration from the iPhone; the relationship hit rock bottom after Apple's 2021 privacy changes, and Meta estimated in early 2022 that the change would cut its revenue that year by $10 billion.
- [11]
Wang said, "Most of the animosity predates my tenure at the company, but obviously there's been a long history between Meta and Apple."
- [12]
Wang pointed to Daniel Gross, an AI researcher on his team and a former Apple director, as a reason for the collaboration.
- [13]
Meta once centered its AI strategy on its open-weight Llama family but has since focused on proprietary models under Wang, who joined when Meta invested more than $14 billion into his former startup Scale AI last year.
- [14]
Two Mac Studios at the 256GB price cost at least $21,598.
- [15]
Four Mac Studios at the 256GB price cost at least $43,196.
- [16]
512GB across 816 billion parameters is about 0.63 bytes per parameter, roughly five bits.
- [17]
1,024GB across 816 billion parameters is about 1.25 bytes per parameter, roughly twice the 0.63 bytes available on two 256GB machines.
- [18]
42 billion active parameters is about 5.1% of 816 billion, leaving roughly 95% inactive for any one token.
- [19]
Two Apple M5 Ultra Mac Studios linked through MLX pool nearly 1 terabyte of unified memory.
- [20]
Two 256GB Mac Studios hold 512GB of memory in total.
- [21]
Two 512GB Mac Studios would pool 1,024GB of memory.
- [22]
The model runs locally instead of in the cloud, adding a layer of security and saving users money, the companies said.
- [23]
Cryptobriefing says rivals that keep their best models behind cloud APIs force customers to send data out and pay per use.
- [24]
CNBC says the collaboration could pose a challenge to OpenAI and Anthropic and could signal Apple's growing strength in AI chips, a market Nvidia has dominated so far.
- [25]
CNBC calls the announcement a sign that the AI boom has shifted historical alliances in Silicon Valley.
Get this ledger inside ChatGPT. Ask about any of our stories and the answer brings its figure, each claim’s verdict and the sources. Free, no account.
Sources
8 independent publishers whose own reporting we read for this story.
- 9to5mac.comMeta launches new Muse Spark Open model, optimized to run on Mac Studio
1 article · October 15, 2026
- cnbc.comApple and Meta team up to make Muse AI model run on Mac computers
1 article · October 15, 2026
- cryptobriefing.comMeta’s Muse Spark Open AI model can run locally on Apple Mac Studio
1 article · October 15, 2026
- engadget.comMeta built a huge open-weight AI model for Macs, but you'll need 1TB of RAM to run it
1 article · October 15, 2026
- macrumors.comMeta and Apple Bring Muse Spark Open AI Model to Mac Studio
1 article · October 15, 2026
- macstories.netMeta Releases Muse Spark Open, an Over 800 Billion Parameter Model That Can Run Locally on Multiple M5 Ultra Mac Studios - MacStories
1 article · October 15, 2026
- qz.comMeta and Apple are partnering to run an open-weight AI model on Mac Studio computers
1 article · October 15, 2026
- Meta releases Muse Spark Open, saying it is the most capable US open-weight model, with 816B total and 42B active parameters and a 1M-token context window
research.meta.ai
1 article · October 14, 2026
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.