Muse Spark Open

You can now run a highly capable, natively multimodal AI model with a million-token context window directly on your desktop Mac.

Meta AI Research has released Muse Spark Open, an 816 billion parameter (42 billion active) open-weight, natively multimodal large language model. It features a 1 million token context window and is designed for coding and agentic tasks, including long-horizon and multi-step reasoning. The model is optimized to run on Apple M5 Ultra Mac Studios using the MLX framework, requiring two machines to achieve a terabyte of unified memory. Meta emphasizes its commitment to open access for advanced AI, believing this enables broader innovation and personal superintelligence. Muse Spark Open is available through various API platforms and model hosting services, with model weights downloadable on major cloud providers and Hugging Face. The model is released under the Muse Model License, which includes limited commercial terms, and has undergone safety post-training evaluations.

The release of Muse Spark Open signifies a significant step towards democratizing access to state-of-the-art AI, particularly its on-device capabilities for complex tasks. Its optimization for Apple Silicon, leveraging the M5 Ultra chip's substantial memory and bandwidth, showcases a pathway for powerful AI to operate outside of cloud-based infrastructure. This development could accelerate the adoption of advanced AI tools for developers and power users who prioritize privacy, offline functionality, or lower latency.

By offering an open-weight model with advanced agentic and coding features, Meta is fostering an ecosystem where customization and further innovation are encouraged. The limited commercial terms of its license suggest a balance between widespread accessibility and controlled commercialization, potentially spurring new applications and research. Developers will want to watch how this model's performance compares to proprietary offerings and how effectively it can be fine-tuned for specialized use cases across various industries.