To run Meta's newest open model at home, you need two Mac Studios
One Tuesday item: Stratechery on Meta's open-weight Muse Spark Open and the Apple deal that runs it on two Mac Studios, with what that setup actually costs.
One item landed in your Gmail reading queue since yesterday's edition, and it comes from Stratechery alone. Stratechery's Tuesday Update covers Meta's new open-weight model, the Apple collaboration that makes it run on a desktop, and a third section titled "Thin and Thick" 1. Lenny's Newsletter published nothing new in the window; its newest item is still Monday's How I AI bundle, which yesterday's edition covered 2.
This update is subscriber-only, where Monday's was readable in full. The public page gives one summary line and three section titles, and both the text and the audio sit behind Stratechery Plus 1. The facts below come from Meta's, Apple's, Hugging Face's and CNBC's own accounts of the release and the deal.
One decision sits behind the item: whether the read is worth your time. A second sits inside the claim itself, in whether "runs on your desktop" changes anything you would do this quarter.
Open weights, on someone's desk
"Muse Spark Open, Meta + Apple, Thin and Thick" — Stratechery, Ben Thompson
Published October 20, 2026, 18:00 Shanghai. Three sections. Subscriber-only. 1
Stratechery's summary frames it as a surprise: "Meta has a new open source model, and to everyone's surprise, they are partnering with Apple to make sure you can run it locally" 1. The three sections behind that line are the model, the Meta and Apple partnership, and an argument about thin and thick 3.
- The model is huge, downloadable, and weaker than the one Meta kept. Muse Spark Open carries 816 billion parameters with 42 billion active per token, and accepts text, images, audio and video as input 4. The context window runs to a million tokens, and the full-precision checkpoint is about 1.5 terabytes on disk 5. Meta calls it "the most capable US open-weight model to date" and releases it under the Muse Model License, whose terms let many developers use it free 4. CNBC reports that it trails Muse Wonder, the proprietary model Meta released three days earlier. Meta's AI chief Alexandr Wang put it plainly: "These aren't the most capable models that we've developed or released" 6.
- Apple contributed memory, and the hardware bill is the claim's real test. The weights shipped alongside an Apple research post showing MLX splitting a model too large for one machine across several Mac Studios, over RDMA on Thunderbolt 5 7. Four M5 Ultra Mac Studios sustain more than 36 tokens per second at 8,192 tokens of context, faster than most people read 7. In four-bit precision the model's memory footprint drops to about 493GB, and in full precision the parameters alone need more than 1.6TB, so one desktop cannot hold it 7. CNBC puts the cheapest Mac Studio with the required M5 Ultra chip and 256GB of memory at $10,799, with a 512GB version due later this month 6. Meta describes two of those machines as the minimum, which puts the entry bill above $21,000.
- The partnership is the newer fact, and Apple now buys models from several rivals. Apple and Meta last partnered on iPhone extensions more than a decade ago, and struck nothing public after 2017, when Apple removed system-level Facebook integration 6. In early 2022 Meta estimated that Apple's app-tracking change would cut its revenue that year by $10 billion 6. The thaw arrives under John Ternus, Apple's chief executive since September 1, and sits beside Apple's OpenAI deal for ChatGPT in Siri in 2024 and its Gemini deal for Apple Intelligence earlier this year 6. Zuckerberg announced this one with a picture of the Meta agent working on a Mac, crediting "work with @johnternus and Apple"; Ternus replied that the Mac Studio "keeps pushing the boundaries of on-device AI" 6.
The third section, "Thin and Thick", is where the argument lives, and it is the part that stays behind the paywall 3. What the public record settles is the setup it argues from: a model given away for download, and hardware that runs past $21,000 before it generates a single token at home 6.
Where to start
A subscriber with twenty minutes should start with "Thin and Thick" and read the model section last, since the specifications are all in Meta's own release 4. A non-subscriber gets the summary line and the three titles from the public page, and the numbers above are the substance behind them 1.
Read it if you are working out how much of your AI stack can sit on hardware you already own, or if you track Apple's push to sell its chips as an AI platform. Skip it if you want benchmark comparisons against the proprietary models; those live in Meta's evaluation table and Apple's throughput charts 47.
The queue refills when Stratechery posts Wednesday's Update and Lenny's Newsletter publishes next.
References
- 1Muse Spark Open, Meta + Apple, Thin and Thick — Stratechery
stratechery.com
- 2Lenny's Newsletter — newest items on the public feed
lennysnewsletter.com
- 3
- 4Introducing Muse Spark Open — Meta AI Research
research.meta.ai
- 5Welcome Muse Spark Open — Hugging Face
huggingface.co
- 6
- 7
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.