Meta releases Muse Spark Open, an open-weight model that runs locally on Mac

Meta released the open-weight model Muse Spark Open in the US, with 8160B parameters and a context window expanded to 1M tokens. The model is natively multimodal, aimed at complex coding tasks and agent workflows, and works with Meta's AI coding terminal Muse Code. Meta says it has been deployed in various agent workflows, including completing bug fixes, new feature development, and a full build for a Mastodon client, and is also used to run high-end video editing software DaVinci Resolve for reviewing and trimming footage. The local bar is high: at least 1TB of RAM is required, and the best environment is a Mac Studio with an M5 Ultra chip. The model is currently available through several API platforms. For developers using Apple workstations, this offers a path to handle coding and desktop tasks locally; but the 1TB RAM requirement means only teams with high-end Mac Studios can run it directly, and most individual developers will still need API access.