Meta Superintelligence Labs has released Muse Spark Open, an 816B-parameter Mixture-of-Experts open-weight model that activates 42B parameters per token and features a 1M-token context window. The natively multimodal model accepts text, image, video, and audio inputs while returning text, positioning itself as a capable US open-weight option built for coding and long-horizon agentic workflows. Meta claims it is the most capable US open-weight model to date, demonstrating strong performance in agentic tool use and knowledge reliability, such as scoring 89.2 on MCP-Atlas.
The model architecture includes 62 layers, a hidden dimension of 8192, and grouped-query attention with 128 query heads and 16 KV heads, alongside a separate 1.8B-parameter Perception Encoder for visual processing. Developers can deploy the NVFP4 build on hardware such as two Apple M5 Ultra Mac Studios using MLX tensor parallelism, four Blackwell Ultra GPUs, or eight H200s, while Unsloth Dynamic GGUFs allow execution in roughly 300 GB of combined memory. Day-one inference engines and hosted APIs are already live across multiple platforms, supported by a specialized Muse Model License that permits redistribution and fine-tuning under specific attribution conditions.
This release matters because it provides developers and enterprises with a powerful, multimodal open-weight model that pushes the boundaries of local and data-center deployment scaling.