Meta has released Muse Spark Open, an 816 billion parameter open-weight model the company calls the most capable of its kind in the United States. Built together with Apple, the model is designed to run locally on Apple silicon, with the weights landing Thursday alongside evaluation results Meta says beat other popular open-weight releases. It combines a one million token context window with native multimodality.
The model targets complex coding and agentic work rather than casual chat. Meta says it tested Muse Spark Open with Muse Code, its AI coding terminal, and ran it against real jobs including bug fixes, new features and a full build for a Mastodon client, plus reviewing and cropping footage inside DaVinci Resolve. Local inference is not a laptop play: Meta points to a Mac Studio with the M5 Ultra chip, and Engadget reports the model needs at least one terabyte of RAM.
For founders the release matters less as a desktop toy than as a signal. Meta is betting that open weights plus high-end local hardware can compete with closed frontier APIs on coding and agents, while the model is also served through several API platforms for teams that will not buy the hardware. The catch is that the bar for local inference keeps climbing, so the near-term winners may be startups selling orchestration, memory tooling and agent infrastructure around models this large.