Meta releases Muse Spark Open, saying it is the most capable US open-weight model, with 816B total and 42B active parameters and a 1M-token context window (Meta AI Research)

Meta has unveiled Muse Spark Open, a cutting-edge US open-weight model boasting 816B total parameters and a substantial 42B active parameters, complemented by an expansive 1M-token context window. This model, designed with agentic capabilities, marks a significant leap in performance and is now available for download on leading platforms such as Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and Microsoft Foundry Models.

Tailored for coding and complex tasks requiring multi-step reasoning, Muse Spark Open runs seamlessly on desktops using Apple M5 Ultra Mac Studios equipped with the latest MLX framework. The model's impressive specifications, including the utilization of two Mac Studios, provide users with a terabyte of unified memory, enabling operations akin to a supercomputer on a single desktop.

This configuration allows Muse Spark Open to process information at unprecedented speeds, with optimizations for various silicon architectures, including AMD, Apple, Intel, and NVIDIA.

Developers can further tailor the model using the PyTorch framework, enhancing its functionality to suit specific needs. Meta's commitment to accessibility and safety is evident in the model's release under the Muse Model License, which permits free use for many developers and includes additional safety post-training measures. Extensive evaluations confirm that Muse Spark Open adheres to stringent safety standards across multiple risk categories, with detailed results available in the model card and a forthcoming evaluation report.

The model's performance has been demonstrated through real-world applications, including automated updates to open-source software and advanced video editing tasks. These demonstrations underscore the model's versatility and its potential to transform how users interact with technology on a personal level. To access Muse Spark Open and explore its capabilities, visit the Hugging Face collection at https://huggingface.co/collections/meta-models/muse-spark-open.