Models
Meta opens the weights: Muse Spark Open goes public
816 billion parameters with 42 billion active, a million tokens of context and a label requirement: Meta's answer to the open-weight field is big, multimodal and free with conditions.
Meta published the weights of Muse Spark Open to Hugging Face on Wednesday, its most capable open-weight release and its clearest answer yet to Qwen, GLM and Kimi. The model totals 816 billion parameters with 42 billion active per token, handles text, image, video and audio, and reads up to one million tokens of context.
In Meta's published scorecard it leads the open field, including Qwen 3.8, GLM 5.3 and Kimi K3, across agentic, knowledge and multimodal categories, while trailing the company's paid Muse Wonder Full models on coding. The model is optimized for Apple's M5 Ultra and, per Meta, runs locally on two Mac Studio machines with 512 GB of unified memory each.
The license is free to use with one visible string attached: deployments must carry a "Built with Muse Spark Open" label. Providers earning more than $20 million a year serving the model as an API need a separate agreement with Meta, a line that puts most cloud vendors on notice.