Meta has released Muse Spark Open, an open-weight artificial intelligence model with 816 billion parameters, according to AI news coverage on October 16. The multimodal model carries a one-million-token context window and is designed for coding and agentic tasks, with Meta positioning it for local deployment on high-memory Apple silicon.
The scale is the story. At 816 billion parameters, Muse Spark Open is among the largest open-weight releases ever, and its release caps a week in which American labs shipped open models in rapid succession — Reflection AI also released its 501-billion-parameter Beam under an Apache 2.0 licence this week.
The catch is the hardware. TechCrunch reports that running the model locally requires at least 1 terabyte of RAM, putting it within reach of serious enthusiasts rather than ordinary laptops — though Apple demonstrated the full model running on two Mac Studios at more than 36 tokens per second, pitching local frontier inference as a realistic use case.
Meta's release widens the range of open models organisations can evaluate for private deployment. The debate now moves to the practical questions: the true cost of running such a model, the support burden, and whether revised licence terms — which drew backlash from business customers — change the calculus for enterprise adopters.
