Meta has released Muse Spark Open, an open-weight AI model designed for coding, automated computer workflows and local operation on Apple computers. The model has 816 billion parameters and a context window of 1 million tokens, placing substantial demands on the hardware needed to run it outside a hosted service.
Meta’s evaluation results put the model ahead of other popular open-weight systems. Those results are company-published benchmarks rather than independent confirmation of its relative performance. The release combines native multimodal capabilities with an emphasis on complex coding and agentic tasks, in which a model takes multiple steps and uses software tools to complete an assignment.
Mac compatibility is a central feature, but it does not make Muse Spark Open a practical download for a typical laptop. TechCrunch, citing Engadget, reported that local operation requires at least 1 terabyte of RAM and that the model is best suited to a Mac Studio equipped with an M5 Ultra chip. That memory requirement sharply narrows the hardware capable of running it locally.
Memory is a major constraint for large-model inference because the computer must accommodate model weights alongside the working data generated during processing. A long context window can add further demands. The advertised 1 million-token capacity describes how much material the model can accept within a context, not a guarantee that every workload will run quickly or efficiently on a particular machine.
Meta demonstrated the model with Muse Code, its AI coding terminal, on software-development workflows. One example involved changes to a Mastodon client, covering bug repairs, additional features and building the application. Such tasks extend beyond generating an isolated code snippet: they can require navigating an existing project, changing related files and checking whether the resulting software builds.
The company also described using Muse Spark Open with DaVinci Resolve, the professional video-editing application. In that demonstration, the model reviewed footage and cropped it. The examples show the range of desktop workflows Meta is targeting, although they do not establish how reliably the model handles unfamiliar projects or how much human supervision those tasks require.
Making model weights available gives developers deployment options that are generally unavailable with API-only proprietary systems. They can potentially operate the model on their own infrastructure, adapt it for particular uses and retain more control over how data is processed. Open weights do not, by themselves, mean that every part of a model’s development is public; training data, training code and other details may remain undisclosed.
Local deployment can also reduce dependence on a remote inference provider, but its economics depend on hardware costs, utilization and operating requirements. For a model of this scale, those considerations are substantial. Meta is also making Muse Spark Open available through several API platforms, offering a hosted route for developers who do not have the memory capacity to run it themselves.
Sources: TechCrunch AI