On October 15, Meta announced the release of Muse Spark Open weights. It lists 816 billion total parameters, 42 billion active parameters and a context of up to one million tokens. This is a distinct development in weight availability, not another name for Monday's Muse Wonder.

Hugging Face access requires manual approval; we did not download the weights. A custom Muse license applies, with conditions for some commercial services, not automatically Apache or unrestricted open source. Meta demonstrates local operation on two M5 Ultra Mac Studios. That differs greatly from a promise of ordinary-laptop operation; we verified neither that setup's performance nor hardware availability.

A possible use is an internal assistant connecting company documents and images without sending every input to a cloud-model operator. The greatest benefit could be better control of sensitive data and adapting the model to a specific workflow.

Requirements include approved access, checking the license, enough memory and tests of actual languages and tasks. Active parameters are not all the weights that must be stored. A local model can still send data out through its tools; file, network and irreversible-action permissions need separate controls.

Our optimistic editorial scenario is a narrow internal pilot in 4–8 weeks after access and suitable hardware are secured. Measure quality after corrections, operating costs and data leakage, not context length alone. We cannot honestly date a universal safe unsupervised assistant or inexpensive home use.