Meta Releases Muse Spark Open: A Powerful 816B Parameter Multimodal Model
Meta has released Muse Spark Open, an 816B parameter open-weight multimodal model featuring a 42B active parameter Mixture of Experts architecture and a 1M token context window. It is specifically optimized for coding, agentic tasks, and long-horizon reasoning across text, image, video, and audio modalities. As one of the most capable open-weight models to date, Muse Spark Open represents a major milestone for open-source artificial intelligence development by providing developers with enterprise-grade capabilities. Its widespread availability across multiple leading inference engines and hardware platforms empowers the broader ecosystem to build advanced autonomous agents and complex software engineering tools. The model ships with custom Helion GPU kernels in SGLang for significant speedups and supports major platforms including Apple Silicon via MLX. Model weights are released under the Muse Model License on platforms like Hugging Face, Amazon SageMaker AI, and Google Cloud Model Garden.
## BACKGROUND
Mixture of Experts (MoE) is a machine learning architecture where a large total number of parameters is split into smaller 'experts,' but only a subset of these experts is actively used for any given token to reduce computational costs. Active parameters determine the inference cost and speed, while total parameters represent the model's overall capacity to store complex information. Long-horizon reasoning refers to an AI's ability to execute multi-step planning and maintain coherent decision-making over extended, complex tasks without losing track of long-term goals.