What was released
Hugging Face's October 15 technical introduction describes Meta's Muse Spark Open: roughly 816 billion total parameters, about 42 billion active per token, and a context window up to one million tokens. It accepts text, images, audio, and video and produces text.
Read the memory and license details
The technical guide says all experts remain in memory in its described deployment; selecting only some experts for each token does not shrink the entire checkpoint to 42 billion parameters. It describes roughly 1.5 TB of BF16 weight files, alongside a lower-precision NVFP4 version. File size is not total runtime memory: the KV cache, which stores information used when processing earlier tokens, adds another requirement. Meta's gated repository uses the custom Muse Model License. Open weights do not imply unrestricted open source; review the access conditions and terms for your intended use. We have not downloaded or tested Spark Open.
Our take
Editorial analysis: start a local-model plan with the task and available memory, then check the exact file format, supported software, and context length. A model's active parameter count is a computation clue, not a promise that it fits your computer. A smaller model that handles your examples reliably can be a more useful starting point than a much larger download.