Meta Unveils Muse Spark Open: 816B-Parameter Model Optimized for Mac Workstations

Meta has released Muse Spark Open, an 816-billion-parameter open-weight model with a 1-million-token context window, positioning it as the most capable U.S.-origin open model to date. The model is natively multimodal and engineered for complex coding tasks and agentic workflows, with native support for Apple's Mac platform — specifically the Mac Studio equipped with the M5 Ultra chip.

Hardware requirements and local deployment

Running Muse Spark Open locally demands at least 1 TB of unified memory, placing it firmly in high-end workstation territory. Meta paired the model with Muse Code, its AI coding terminal, and demonstrated agentic capabilities including autonomous bug fixes, feature additions, and full builds for a Mastodon client. The model also showed competence in professional video editing workflows, reviewing and cropping footage within DaVinci Resolve.

Open-weight distribution brings cost efficiency, transparency, and flexibility that closed-source alternatives cannot match. Organizations can fine-tune, audit, and deploy the model on their own infrastructure without vendor lock-in. Meta is making Muse Spark Open available immediately through multiple API platforms for teams that prefer managed inference over local deployment.

Why it matters for GPU / AI infrastructure: The 1 TB memory requirement underscores a growing trend — frontier open models are outpacing commodity GPU memory capacities. For cloud providers and on-prem teams, this signals demand for larger-memory nodes (HBM3e, unified memory architectures) and efficient model-parallel serving stacks. AiGpu's high-memory GPU instances are purpose-built for exactly this class of workload.