Meta has released Muse Spark Open, a native multimodal large model specifically designed for code and agent tasks. This model boasts 816 billion parameters, with 42 billion activated parameters and a context length of 1M, optimized for Mac devices. Leveraging Apple's MLX framework, it is tailored to the unified memory architecture of Apple chips, enabling local operation with just two Apple M5 Ultra Mac Studios equipped with 512GB of memory each. Despite the previously lukewarm relationship between Meta and Apple, the two companies have been collaborating for several months. Muse Spark Open not only supports multiple mainstream inference engines but is also optimized for chips from various brands, allowing developers to customize using the PyTorch framework. Although its comprehensive capabilities are not the strongest among globally open-weight models, it excels in certain Agent evaluation metrics, indicating that the value of open-source models should be assessed in conjunction with practical Agent workflows.