Meta introduces Muse Spark Open. This 816-billion-parameter AI model can run locally on two Apple Mac Studio computers. Meta states that Muse Spark Open is designed for software development and autonomous AI agents. According to the company, the new model is currently the most powerful American open-weight model.
Muse Spark Open is based on Muse Spark 1.2, which was released in August. It uses a mixture-of-experts architecture. With this architecture, it activates only 42 billion parameters per token. Thanks to its one-million-token context window, it can process extensive codebases, documents, and complex instructions.
Meta states that the model performs well at agentic tasks. On the JobBench benchmark, it scored 61.8, compared to 31.1 for Inkling from Thinking Machines Lab. Muse Spark Open also topped the MCP-Atlas benchmark for tool usage, scoring 89.2.
Despite these results, Chinese models still lead. Kimi K3 scored 85 on Terminal-Bench 2.1, compared to 76.4 for Meta’s model. These results are based on Meta’s own evaluations.
AI agents run locally
A notable feature is the model’s optimization for Apple Silicon. Using Apple’s MLX framework, Muse Spark Open can run on two Mac Studio systems equipped with M5 Ultra processors. Each machine supports up to 512 GB of unified memory. This provides sufficient combined capacity to run the entire model.
Meta demonstrated the setup with its Muse Code agent, which modified an open-source iOS application, fixed bugs, and implemented new features. Another demonstration showed how the model controlled DaVinci Resolve to analyze and edit video footage.
On-premises deployment can benefit organizations that want to keep sensitive data and source code out of external cloud services. However, the required hardware entails a significant investment.
Hugging Face, Amazon SageMaker, Google Cloud, and Microsoft Foundry offer the new model. Developers can also access the model through various API providers.