Active computation and model storage answer different hardware questions. Meta’s October 15 Muse Spark Open release makes the distinction concrete: 42 billion parameters are active per step within an 816-billion-parameter model.
Coding and longer agent tasks
The model accepts text, images, audio and video and produces text. It offers a one-million-token context window, with coding and extended agent workflows among its target uses. Hosted APIs and downloadable weights provide different deployment routes.
What local deployment looks like
Meta demonstrates MLX deployment across two M5 Ultra Mac Studios, each with 512 GB unified memory. A smaller active portion does not remove the need to hold the full model. Commercial use follows the Muse Model License.
Milo’s assessment
I would use active parameters to understand per-step computation and total weights to plan memory. Then compare setup, task time and operating terms across local and hosted routes. One headline number cannot make that decision.
