Meta releases Muse Spark Open for dual-Mac desktops
Local inference comes with substantial memory needs and conditional licensing.
Meta's October 15 release brings Muse Spark Open to two M5 Ultra Mac Studios through MLX. Each machine can expose up to 512 GB of unified memory to its GPU. Meta demonstrated changes to the Ice Cubes Mastodon client and video review/cropping in DaVinci Resolve.
The model card describes 816 billion total parameters, 42 billion active per token and a million-token context. It accepts text, images, audio and video, returning text. The BF16 version requires 2 TB of GPU memory; NVFP4 requires 700 GB.
The commercial boundary
The Muse licence requires attribution and compliance with its usage policy. Prior notification and a separate agreement apply to model-as-a-service providers above $20 million in aggregate gross revenue, including affiliates, over twelve consecutive months. Internal use and non-MaaS products are excluded from this revenue clause.
Desktop here means a pair of unusually memory-rich machines, with the launch demonstrations supplied by Meta itself. Meta says it met release thresholds; a full evaluation report was still forthcoming.
Open weights give deployment control, while the licence distinguishes internal use from providing access to a model.
Sources(1)
Other coverage of this story(2)
- Meta releases Muse Spark Open: an 816B-parameter multimodal model built for coding agentsThe Guardian Artificial Intelligence
- Meta returns to open weights with Muse Spark Open—but the license has conditionsThe Decoder
