You may have seen the headlines by now.
“Apple and Meta bury the hatchet — for AI“
“Apple and Meta team up to make Muse AI model run on Mac computers“
“Meta and Apple team up to make Muse AI models run better on Macs“
“Meta launches an open-weight AI model that runs on Macs“
None of these are exactly false, but you’d be forgiven for thinking that Meta’s popular new Muse agent will now run locally on your Mac. It will not. There’s a Muse app for Mac and iOS (and the web) but the actual agent itself and all the models it uses run on Meta’s servers in the cloud.
What these stories are referring to is the new open-weights model that Meta just released called Muse Spark Open, compared to Meta’s current generation of AI stuff, which is simply called “Muse.” Meta says Spark Open has the most capable open-weights model so far (from the U.S.). However, it’s not a competitor for the frontier closed models such as GPT-6 Astra or Claude Opus 5.5; that would be Muse Wonder and Muse Wonder Full, which are much bigger models that run only in the cloud.
Muse Spark is of the same generation, but it’s a smaller, lightweight, fast-response model made to be comparatively inexpensive. And yes, it will run natively on a Mac! Well, two Macs. But not your Macs.
Apple has a blog post about optimizing Muse Spark Open and two other models to run natively on Apple Silicon Macs. It’s interesting stuff, but it’s worth noting that the other two models (Mistral Medium 3.5 and Qwen 3.8) are the only ones that can run on a single Mac Studio.
Muse Spark Open would require a whopping 1.6 terabytes of RAM in its native precision (BF16) just for the 800+ billion parameters alone. So Apple worked with Meta to optimize and tune it, and if you run it at a quantized 4-bit precision (NVFP4) you can get the memory requirement all the way down to… 493GB. That’s just for the model. There’s other overhead involved in actually running it and for macOS itself, so a single Mac Studio with 512GB of RAM can’t run it. I’ll remind you that the 512GB Mac Studio with M5 Ultra is not even out yet, and is expected to cost at least $15,000 and probably quite a bit more.
You need two Mac Studios with 512GB of RAM to run Muse Spark Open, connected via their Thunderbolt ports and running the new MLX framework for distributed computing over multiple Macs. You might get away with three Mac Studios with 256GB, Apple doesn’t say.
So yeah, you can run the latest open-weights model from Meta “right from your desk” as Apple says, provided your desk has over $30,000 of Mac Studio hardware on it.
No, you are not able to run the Muse agent locally on your Mac. You are not able to run the latest and best frontier LLM models from Meta (Muse Wonder) on your Mac. You can only run the latest and best open weights model on at least two very expensive and high-end Macs. For all but a small subset of researchers, you’d be much better off paying to run it on a cloud server.
If there’s a story here for regular consumers, it’s this: Apple is very serious about being the go-to hardware for AI researchers running local models, and that might potentially pay off in the form of well-optimized AI software meant for everyday people with Apple Silicon.
