Meta Opens Muse Spark Model to the Public

Meta Platforms Inc. announced that it has made the weights for its Muse Spark AI model open‑source. The model contains 816 billion parameters, but only 42 billion are active for any single token. It supports mixed‑modal prompts that can include text, images, audio or video, and the context window can hold up to one million tokens. Meta said the model meets the safety thresholds in its Superintelligence Scaling Framework, and a full evaluation report is forthcoming. The release builds on Muse Spark 1.2, which shipped in August, and adds post‑training safety so it can be run without Meta’s own guardrails.

The open‑weight version was optimized for Apple Silicon. Using the latest MLX framework, two Mac Studio machines with M5 Ultra chips can share 1 TB of unified memory and run the full model as one system. The M5 Ultra’s 512 GB GPU memory and 1.2 TB/s bandwidth allow the model to read the small slice each token requires. Tests on‑device included shipping iOS changes to the open‑source Mastodon client Ice Cubes and running DaVinci Resolve end‑to‑end, with the model selecting and cropping the best shots.