Silicon Local: Meta Releases 816B Parameter Muse Spark Open Designed to Run on High-End Macs

Meta launched Muse Spark Open, claiming the title of the most capable open-weight artificial intelligence model built in the United States. The model sets itself apart from standard open releases by featuring direct design optimizations for Apple silicon, allowing high-performance Mac hardware to process massive open-weight workloads locally. Packed with 816 billion parameters and featuring a 1 million token context window, Meta’s latest release aims to outperform competing open-weight systems across complex coding benchmarks, spatial reasoning tests, and agentic workflows.

Muse Spark Open features native multimodal capabilities built specifically for heavy software development and automated desktop tasks. Meta integrated the system with Muse Code, its specialized AI coding terminal, deploying the model to execute multi-step software fixes. In internal deployment tests, the model handled full software builds, implemented new user features, and fixed complex coding bugs across a client application for the open-source social network Mastodon.

Beyond text generation and software engineering, Meta tested the model on complex video processing workflows. Engineers connected Muse Spark Open to DaVinci Resolve, a professional post-production video editing platform, allowing the model to review raw video clips, select target frames, and crop footage automatically. Handling non-text media directly inside creative desktop software demonstrates how open models are moving past simple chat windows into active desktop automation.

Running an 816 billion parameter model locally requires serious hardware power. Meta noted that running the system locally requires at least 1 terabyte of system RAM. The model runs best on Apple’s Mac Studio desktop powered by the M5 Ultra chip. Apple’s unified memory architecture lets integrated graphics processors share massive system RAM pools, allowing two high-end Mac Studio machines to function like a single desktop supercomputer capable of housing the entire model in local memory.

Deploying open-weight models offers distinct operational advantages compared to closed corporate cloud APIs. Open-weight models let engineering teams inspect internal software weights, modify core code, run offline inference, and avoid paying variable token fees to cloud vendors. Businesses running local hardware protect sensitive codebases, ensure data privacy, and eliminate latency bottlenecks associated with sending private data to external cloud servers.

Meta released Muse Spark Open across several major API hosting platforms, giving developers immediate cloud access while hardware enthusiasts set up local desktop installations. Bridging high-capacity open-weight models with consumer workstation hardware represents a major step toward decentralized local computing. As desktop hardware becomes more powerful, running massive models on local workstations will allow developers and creative teams to build private, automated workflows without relying on centralized corporate infrastructure.