Meta launches Muse Spark Open, an 816-billion-parameter open-weight model built to run on Macs

Meta has released a new open-weight AI model that it calls the most capable of its kind in the United States. According to TechCrunch, the model, named Muse Spark Open, is notable for a second reason: it was built with Apple's Mac hardware in mind and is designed to run locally on those machines.

A very large model with a very large window

Muse Spark Open weighs in at 816 billion parameters and supports a context window of 1 million tokens, TechCrunch reports. Meta says benchmark evaluations it published alongside the launch show the model outperforming other widely used open-weight models, though those results come from Meta itself rather than an independent party. The model is natively multimodal and is aimed at demanding coding work as well as agentic tasks, where software carries out multi-step actions on a user's behalf.

The hardware catch

Running a model of this size on a desktop is not trivial. TechCrunch, citing Engadget, reports that local operation requires a machine with at least 1 terabyte of RAM. The best-performing configuration, per the same reporting, is Apple's Mac Studio fitted with the M5 Ultra chip, a workstation-class machine aimed at professionals. In other words, while the model is technically local-first, the hardware needed to run it comfortably sits well outside most consumers' budgets.

For those who can muster the memory, TechCrunch suggests the model may serve as a serious workhorse for programming and other desktop tasks.

Demonstrations focused on coding and creative software

Meta paired Muse Spark Open with Muse Code, its AI coding terminal, and tested it across several agentic workflows. According to TechCrunch, one demonstration had the model working on a Mastodon client, where it handled bug fixes, added new features and completed a full build. Another deployment put the model to work inside DaVinci Resolve, the high-end video editing suite, where it was shown reviewing and cropping footage.

Taken together, the demonstrations underline Meta's positioning of the model: less a chatbot and more an engine for autonomous desktop agents that operate real applications end to end.

Distribution and the open-weight argument

The model is available now through a number of API platforms, TechCrunch reports. As with Meta's earlier open-weight releases, the pitch rests on the advantages such models hold over closed, proprietary systems. Open weights can be significantly more cost-efficient to operate, and they offer a degree of transparency and flexibility that closed models do not, since users can inspect, adapt and self-host them.

Why it matters

Two things make this launch significant. First, it pushes the ceiling for open-weight models, giving developers and researchers access to frontier-scale capabilities without depending on a proprietary API. Second, the Mac-first design signals where Meta believes serious AI work is heading: onto local hardware. Running a model of this class on a desktop rather than in a data centre carries real benefits for privacy, latency and cost, even if the 1 terabyte RAM requirement means very few people will actually do it today. If workstation memory capacities keep climbing, locally run models at this scale could become practical for a much broader group of developers, and Meta is positioning itself to be the supplier when that happens.