View the chart data
| M5 Ultra machines | Memory per machine | Total memory |
|---|---|---|
| 2 | 512 GB | 1,024 GB |
| 4 | 256 GB | 1,024 GB |
Meta's new AI model can run on two high-end Apple desktop computers linked together. Released October 15, Muse Spark Open gives businesses another way to use powerful AI: download the model and run it themselves, instead of relying entirely on a remote service. [1] [2]
For Meta Platforms (Nasdaq: META), that puts its technology into more developers' hands. For Apple (Nasdaq: AAPL), it gives the Mac Studio a bigger job in the AI race. The opportunity is selling computers that can do work customers might otherwise send to the cloud. [3] [6] [7]
More than a chatbot on a screen
Meta says it tested the model doing real software and video work entirely on the devices. Its examples include fixing bugs and adding features to Ice Cubes, an open-source social app, and using DaVinci Resolve to review footage and prepare vertical video. These are company demonstrations, rather than evidence of widespread customer use. [1]
The downloadable part matters. Developers receive the model's learned settings, known as weights, so they can run and adapt it within their own systems. That gives them a choice about where the work happens, rather than simply opening another chatbot website. [4]
“freedom to run advanced models wherever, including at home.”
Meta research announcement · October 15, 2026 [1]
Apple gets a reason to sell bigger desktops
This is demanding work for a personal computer. Meta documents two Mac Studios with M5 Ultra chips and 512 gigabytes of shared memory each, or four with 256 gigabytes each. The machines need fast direct connections. An ordinary laptop is not the setup being offered. [2]
Apple's October 15 research explains why one machine is insufficient. Even after compression shrinks the model, running it needs extra memory for calculations. Apple's MLX software divides the work among linked Macs. The company has published tests using both two and four machines. [3]
Local operation could appeal to teams that want more control over their work and data. But owning the machines also means paying for them, maintaining the software and managing security. Neither company's announcement establishes that this route is cheaper for every business. [3]
Giving away access still has to pay off
Meta's wider spending makes that business question substantial. In July, it forecast company-wide 2026 capital spending of $130 billion to $145 billion, including principal payments on equipment leases. That is a forecast for the whole company, not the cost of Muse Spark Open. [5]
Downloads alone do not show how much money Meta will earn. The model also comes with license conditions, including a separate agreement for some commercial services that sell access to models. Its model card recommends additional safeguards suited to how it is used. [4]
The immediate change is tangible: a large AI model can now do useful tasks on linked desktops. The next business test is whether customers use those machines for everyday work often enough to justify buying and running them.