Meta released Muse Spark Open through its Hugging Face organisation on 15 October. It is a mixture-of-experts model with roughly 816bn total parameters, of which about 42bn are active for any given token. Each of its 62 layers holds 256 experts with eight selected per token, and the experts account for about 97% of the parameters. A perception encoder of about 1.8bn parameters handles images and video, with a separate audio encoder, and the context window runs to roughly 1.05m tokens. Output is text only.
This is an open-weights release in a family Meta has otherwise kept closed, which makes the licence the most important missing detail. The announcement does not state one, and without it there is no way to tell whether these weights can be used commercially. Meta calls the model the most capable US open-weights model to date but publishes no benchmark results alongside that claim, so it cannot be checked.
Practically, the BF16 checkpoint runs to about 1.5TB, which puts it beyond most single machines. Weights are available in BF16 and NVFP4 and run through transformers, vLLM, SGLang, llama.cpp and MLX. The release post is on Hugging Face.