Back to the stories

Meta releases Muse Spark Open, an 816-billion-parameter open-weight multimodal model

Score 10.0,

Artificial Intelligence

A follow-up to our coverage of Muse Wonder three days ago: Meta has published Muse Spark Open as an open-weight multimodal model.

Muse Spark Open pairs very large scale with extremely long memory. Public reporting and Hugging Face documentation say the model totals about 816 billion parameters, activates roughly 42 billion per token, and accepts text, image, audio and video while supporting up to a one-million-token context window.

Why that matters: scale, long context and multimodality together change what models can do. Teams can now train and test agents that juggle long documents, run multistep coding workflows, or keep a persistent memory across hours of interaction without chopping the conversation into pieces.

How it works, at a high level: the model is huge but does not use all of its capacity for every question. Think of a vast library that walks straight to the few shelves that contain the answer. That approach makes very long attention spans and specialized skills practical without everyone having to run the entire library at once.

What changes now is who can build advanced agentic tools. Because the weights are public and Hugging Face documents fine-tuning, researchers and startups can experiment directly rather than only through closed APIs. Practical reality check: running the full model still needs serious hardware, though Reuters notes a configuration that runs across two Apple M5 Ultra Mac Studios and providers already host APIs for others.

The open question is which independent teams will turn these open weights into safe, useful products faster than closed rivals, and how benchmarks and regulators react once those products appear.