Meta's Muse Spark Open is an 816B open-weight model built for Macs

Meta launched Muse Spark Open on Thursday, an open-weight model with 816 billion parameters and a 1 million token context window. Meta positions it as the most capable open-weight model in the United States, and the detail that made it news is the hardware it targets: Apple silicon. According to TechCrunch, the model was designed from the start to run locally on Macs.

The Mac claim is real, with a very large asterisk

Running an 816B parameter model on a consumer machine is a memory problem before it is a compute problem, and the requirement is steep. Per Engadget, cited by TechCrunch, Muse Spark Open needs at least 1 terabyte of RAM to run locally. That puts it beyond every Mac except the top of the line, and TechCrunch notes the model runs best on a Mac Studio with the M5 Ultra chip.

So "it runs on a Mac" is true in the way that "it runs on the most expensive Mac with the largest memory option" is true. The local inference story is a capability demo for Apple's unified memory architecture, not an offer most teams can act on this quarter.

What Meta says it is good at

The model is natively multimodal, and Meta built it for complex coding tasks and agentic work. The launch story comes with two deployment examples:

• Paired with Muse Code, Meta's AI coding terminal, the model worked through a Mastodon client, handling bug fixes, new features and a full build.
• In a second workflow, it operated DaVinci Resolve, the professional video editing software, where TechCrunch reports the model reviewed and cropped footage.

Meta also published evaluation results claiming it outperforms other popular open-weight models. Those are Meta's numbers on its own benchmarks, so treat them the way you'd treat any vendor table: directionally interesting, not independently verified.

What the open weights actually change for you

Open-weight releases carry the usual advantages over closed models: you can inspect and fine-tune the weights, you can host the model on your own infrastructure when data has to stay put, and your per-token cost depends on your own hardware and provider choices instead of a single API price list.

Muse Spark Open is available now through a number of API platforms, which is the part that matters for most readers: you can route real workloads to it without buying a single Mac.

The takeaway

Two signals, one purchase. The signal: Meta is optimizing a flagship-class open model for unified-memory consumer hardware, which tells you where the frontier open-weight curve is heading in terms of context and multimodality at this parameter scale. The purchase: skip the Mac for now. Run it through an API or on your existing GPUs, measure it on your own workloads, and revisit local inference when the 1TB RAM requirement stops being a joke.