Picture this. A Mac Studio sits on your desk, priced near the top of what a personal computer can cost. You download an open-weight model advertised as built for Macs, hit run, and the system tells you there isn’t enough memory.
It sounds like a joke, but it is roughly what Meta’s latest release looks like in practice. On October 15, Meta launched Muse Spark Open. TechCrunch reported that Meta calls it the most capable open-weight model in the United States, and that it was designed to run on Apple computers. The same report, citing Engadget, adds the catch: running it locally takes at least 1 terabyte of RAM.
Runs on a Mac is not the same as runs on your Mac
Start with the numbers as TechCrunch relays them. Muse Spark Open has 816 billion parameters and a context window of 1 million tokens. It is natively multimodal and aimed at complex coding and agent-style work. Meta says its evaluation results beat other popular open-weight models. Those results come from Meta itself, and I haven’t seen independent replication yet, so hold them loosely.
The memory figure follows from the parameter count. Every parameter has to live somewhere in memory, and 816 billion of them need hundreds of gigabytes even with aggressive compression. The report says the model runs best on a Mac Studio with the M5 Ultra chip.
So “built for Mac” is best read as “tuned for Apple silicon’s unified memory.” On a Mac, the CPU and GPU share one memory pool, so a machine with a lot of RAM can hold models that a typical graphics card cannot. That is a real advantage for local inference. It also comes with a price tag, because a machine with terabytes of memory is not something most developers buy on a whim.
Who is this model for
With a bar this high, who is Meta aiming at?
Two details in the report help. First, Meta paired the model with Muse Code, its AI coding terminal, and had it work on a Mastodon client: bug fixes, new features, and a full build. Second, it used the model to run DaVinci Resolve, the professional video editor, where it reviewed and cropped footage. Neither demo is a chat. In both, the model does the work itself.
That points to the pitch: open-weight models can act as desktop labor. The more practical line for most readers is the last one in the article. The model is already available through several API platforms. Most people will call it like any other hosted model, and running it locally will be an option for teams with top-end workstations, research groups, and companies willing to pay to keep data in the building.
Does the open-weight advantage still hold
TechCrunch notes that open-weight models can be cheaper than closed ones and offer more transparency and flexibility. That is true in general, but it needs a discount here.
Cheaper is not obvious next to a 1TB memory requirement. A machine that can hold this model may cost more than what many teams spend on API calls in a year. The case for local hosting is strongest when call volume is very high or when data must stay on site, because then the one-time hardware cost spreads thin. Transparency and flexibility do hold up. With the weights in hand you can fine-tune, audit, and avoid depending on one cloud vendor’s pricing.
The trade-off is easy to sum up for people building products. If your app makes modest daily calls, an API is the simplest route. If you work in a regulated field, it is worth running the numbers on local deployment. If you just want to tinker on a laptop, this model isn’t for you, and a smaller distilled or quantized version from the community will be a better fit. The report doesn’t say whether Meta plans a smaller version, so I won’t guess.
Why unified memory matters
Local inference comes down to one hardware detail. On a typical desktop PC, the model has to fit in the graphics card’s memory, which on consumer cards is somewhere between a dozen and a few dozen gigabytes. Once a model is bigger than that, data shuffles between video memory and main memory, and speed drops sharply.
Apple’s chips take a different route. The GPU can use the machine’s whole memory, so the RAM you buy is the RAM the model can use. That is why many people run local models on Macs: not because they are the fastest, but because they can hold the model at all. Meta’s choice to target the Mac follows from that. It is aiming at a single desktop machine instead of a rack of accelerators. At 816 billion parameters, though, that picture only works at the very top configuration.
One more point on context. The model claims a 1 million token window, and longer context costs extra memory at inference time. That is my inference from how these systems generally work, not a figure from the report, but it suggests 1TB is a floor and not a comfortable target.
What to watch next
Three things are worth tracking over the coming weeks. Independent benchmarks come first, since Meta’s own numbers need outside checking. Next is pricing and speed on the API platforms, which will decide the real experience for most users. The third is whether the community produces quantized builds that fit in 128GB or 256GB of memory, because that is what would actually lower the barrier.
Until those land, the fair reading is that Muse Spark Open is a significant release and also very far from something everyone can run on their own Mac.
Related reading
- The Real Lock-In Was Never the Model: AI Memory Can Now Move With You
- The Real Cost of Monday.com in 2026: 5 Alternatives That Wont Drain Your Budget
- The Real Cost of Scaling AI Agents in 2026: What the ROI Numbers Don’t Tell You
- Tracecat Alternatives: Best Open Source Security Automation Platforms 2026
- Authentik Alternatives: Choosing an Open Source SSO for Small Teams (2026)
Stay updated with our latest AI insights
