META JUST RELEASED AN 816B OPEN-WEIGHT AI MODEL THAT RUNS ON TWO MAC STUDIOS @Meta has released Muse Spark Open, its new AI model built for complex coding, autonomous agents, and long-running engineering tasks. The unusual part is that developers can download the weights and run a model of this scale on local hardware instead of relying entirely on cloud APIs. Here's what Meta packed into it: 816B total parameters 42B active parameters per token 1M-token context window Native text, image, video, and audio input Support for AMD, Apple, Intel, and NVIDIA hardware The architecture uses Mixture of Experts. Instead of activating all 816 billion parameters for every token, the model uses roughly 42 billion, reducing the computation required during inference. That makes its desktop deployment especially interesting. Meta worked with Apple to optimize Muse Spark Open for two M5 Ultra Mac Studios using the MLX framework. Together, the machines provide up to 1TB of unified memory. The company also demonstrated actual agentic workflows running locally: Fixing bugs and implementing features in an iOS application Completing software builds through Muse Code Controlling DaVinci Resolve to analyze footage Identifying strong video moments and creating vertical crops The model is available through Hugging Face and API providers including Groq, Together AI, Fireworks, and OpenRouter. But there's an important distinction. Open-weight doesn't mean unrestricted open source. Meta's Muse Model License includes commercial limitations, and running an 816B model locally still requires expensive hardware. What interests me most is what this could mean for AI-native operations. Marketing teams are increasingly building agents that process customer conversations, analyze competitors, classify content, and coordinate production workflows. Today, many of those systems depend on external model providers for every step. More capable open-weight models create another option: running sensitive research, internal knowledge processing, and specialized agents on infrastructure the organization controls. That doesn't automatically make local AI cheaper or better. Deployment costs, inference speed, maintenance, and model quality still determine whether it makes sense. But the competitive landscape is changing. Access to powerful models is gradually becoming less dependent on a handful of closed API providers. And as model access becomes more flexible, the advantage shifts toward the teams that can build reliable workflows, connect proprietary data, and turn AI capabilities into useful systems.