Overview
- Meta released Muse Spark Open on Thursday and described it as the most capable U.S. open-weight model to date.
- The model is a native multimodal system with 816 billion total parameters, 42 billion active parameters, and a 1 million‑token context window built for code generation, agent workflows, and long-range multi-step reasoning.
- Meta published the full model weights on major hubs and enabled API access through multiple providers while licensing the weights under the Muse Model License, which restricts commercial use for some actors.
- Meta and Apple confirmed a joint technical demo showing Muse Spark Open running locally via Apple’s MLX framework on two M5 Ultra Mac Studio units with 512GB unified memory each, and that high-memory Mac Studio configuration is not yet widely available and is expensive.
- Meta says it strengthened safety alignment after training and provided optimizations for major inference engines and chips to ease deployment, and outside researchers will likely seek independent verification of the model’s benchmarks and safety claims.