Meta makes 'Muse Spark' available for download and local operation

Meta has released the weights of its 'Muse Spark Open' model for download and local operation, a model with 816 billion parameters and a million-token context window, designed for programming and multi-step tasks. The release is based on 'Muse Spark 1.2' with additional safety training, following the smaller 'Muse Glimmer' model. Tests showed superiority in some tasks and lag in others compared to competing models.

Meta has made the weights of its new 'Muse Spark Open' model available for download and local operation, expanding options for developers who want to run and modify AI systems outside the company's services. The model, which Meta describes as the most capable among American open-weight models, has 816 billion parameters, of which 42 billion are activated per token processed, and a context window of one million tokens. It is designed for programming and multi-step tasks, and has been optimized to run on two Apple Mac Studio computers equipped with the M5 Ultra processor using the MLX framework. According to Silicon Angle, the release is based on 'Muse Spark 1.2' with additional safety training suitable for use outside the controls imposed by Meta's services. It follows the smaller 'Muse Glimmer' model, while Meta's more powerful 'Muse Wonder' models remain available only through its software services. Comparisons presented by the site, citing Meta's test table, showed varying results depending on the task: the model scored 76.4 in a programming test, compared to 85 for the 'Kimi K3' model, but advanced in another test for tool selection and usage. Hugging Face platform noted that the model accepts text, image, and audio input.