Summary
Meta’s Hugging Face collection describes Muse Spark as an 816B-parameter model with 42B active parameters. It lists BF16 and NVFP4 weight variants.
Meta’s Hugging Face collection for Muse Spark Open lists a model described as having 816 billion parameters, of which 42 billion are active. The collection includes two image-text-to-text model repositories, one named for BF16 weights and another for NVFP4 weights.
What the collection lists
The repositories are named Muse-Spark-Open-816B-A42B and Muse-Spark-Open-816B-A42B-NVFP4. BF16 is a 16-bit floating-point format; NVFP4 is a 4-bit floating-point format. Lower-bit formats can represent weights using fewer bits, though their practical memory and performance effects depend on the model implementation and the hardware running it.
The collection distinguishes the model’s total parameter count from its active parameter count. Those figures describe different aspects of model scale: the total counts parameters in the model, while the active figure refers to the parameters used during computation. The listing does not provide enough architectural detail to explain how Muse Spark selects or uses active parameters.
There is also a mismatch in the page’s figures: the collection overview says 816B parameters, while the individual repository listings display 484B and 822B. The page does not explain the difference, so the 816B figure is best treated as the collection’s stated summary rather than a reconciled count.
Parameter totals alone do not determine the memory or hardware needed to run a model. The collection provides no RAM or platform requirement, so its listing does not establish a Mac-specific requirement.