Meta releases Muse Spark Open with local deployment support

Meta has released Muse Spark Open, an 816-billion-parameter open-weight model for businesses seeking to run and customize AI on their own infrastructure. Its mixture-of-experts architecture activates 42 billion parameters per token. It accepts text, images, video and audio, supports up to 1 million tokens of context, and produces text.

Meta demonstrates local operation using paired M5 Ultra Mac Studios, each supporting up to 512GB of unified memory. However, that memory configuration does not arrive until late October, and pricing remains unannounced. The demonstrations include coding and video-editing workflows, but Meta provides neither measured token-generation speeds nor a complete hardware cost for the setup.

Meta calls Spark Open its most capable U.S. open-weight model, rather than the global leader. Its published results show strengths in JobBench and MCP-Atlas tool-use evaluations, while several Chinese competitors perform better on coding and broader measures. Meta reports an Artificial Analysis Intelligence Index score of 39, but a separate public assessment of the released model could not be verified at publication.

The custom license permits modification, fine-tuning and commercial deployment, with conditions. Model-serving businesses whose licensees and affiliates generate more than $20 million in aggregate gross revenue during any consecutive 12 months must notify Meta and secure a separate agreement. Internal deployments and products that are not model-as-a-service offerings are exempt from that provision. Attribution and usage restrictions still apply.