Meta's new open-weight AI model needs a Mac Studio and a terabyte of RAM to run
The release turns Apple's professional hardware into inference infrastructure for a rival's AI strategy — and sets up a local-versus-cloud cost decision for buyers.
Published September 9, 2026 · Updated October 16, 2026
Why it's worth posting
Muse Spark Open is reported as a model release, but the consequence is where the story lives. The model requires at least a terabyte of RAM and runs best on a Mac Studio with the M5 Ultra chip, which means Apple's silicon and memory engineering are now being evaluated — and effectively depended on — by Meta, a company whose priorities Apple does not control. That creates a concrete decision for buyers rather than a hypothetical one. Enterprises and evaluators can provision M5 Ultra hardware at scale for local inference, or pay Meta's cloud API rates of $4 per million input tokens and $20 per million output tokens. With Muse already past 1 million daily active users and live connector integrations for tools like Figma, GitHub, and Notion following the October 7 iPad launch, the adoption pressure is present now, not forecast.
The strongest reason to post is the reframe: a Meta model release quietly makes Apple's professional Mac line a candidate inference appliance for a competitor. Because Muse Spark Open runs best on the Mac Studio with the M5 Ultra chip and demands at least a terabyte of RAM, Apple's hardware roadmap is now being pressure-tested by decisions it did not make and does not control.
There is also a technical first worth noting honestly. At 816 billion parameters with a 1 million token context window and native multimodal capability — demonstrated reviewing and cropping footage in DaVinci Resolve — the model is designed to run on a single workstation rather than a data center. That extends Meta's broader Muse push, which crossed 6.6 million installs between its September 8 mobile launch and the October 7 iPad release, in a new direction aimed at developers and prosumers who want inference they control.
The open question is whether local-first deployment is a genuine strategy or a demonstration. The real test is cost: whether running the model on owned hardware meaningfully undercuts the hosted API tier. And the limits matter — the Muse Wonder Full tier is reserved for Muse Code subscribers and is not served through any API, so the local-everything story is not complete.
Angles to take
Treat it as an Apple story, not a Meta story: a rival's model release turns the Mac Studio and M5 Ultra into de facto inference hardware, putting Apple's memory-ceiling roadmap under pressure it did not create.
Write this post →Lay out the buyer's binary for late 2026 — provision terabyte-class M5 Ultra hardware for local inference, or pay Meta's cloud rates of $4 per million input and $20 per million output tokens — and ask which actually wins on cost.
Write this post →Frame the technical first plainly: an 816-billion-parameter, million-token, multimodal model built to run on one workstation rather than a data center, and what that opens up for developers who want inference they control.
Write this post →Note the catch that undercuts the local-everything pitch: the Muse Wonder Full tier is locked to Muse Code subscribers and served through no API, so the most capable layer isn't something you can run yourself.
Write this post →