Meta releases open weights for 816B parameter Muse Spark model

Meta Platforms released the open weights for its Muse Spark artificial intelligence model on October 15, 2026. SiliconANGLE and Hugging Face both reported that the release, named Muse Spark Open, marks an open-weight transition for a model that had been proprietary since its debut in April. Meta described the system as the most capable open-weight model to come out of a United States laboratory to date. [2][3][4]

The model features a mixture-of-experts architecture containing 816 billion total parameters, with roughly 42 billion active parameters per token, according to SiliconANGLE and Hugging Face. Hugging Face reported that the system supports a context window of up to one million tokens and handles text, image, audio, and video inputs to generate text. Meta trained the model specifically for agentic coding tasks, including codebase analysis, debugging, and tool usage. [3][4]

Meta and Apple collaborated on the release to optimize desktop performance. The Verge reported that Meta designed the model as the first in its class capable of running locally on a desktop setup using two Apple M5 Ultra Mac Studios and the latest MLX framework. The two technology companies publicly supported the joint launch despite their usual marketplace rivalry. [2]

Meta Platforms is based in Menlo Park, California, and operates several major communication platforms and advertising networks. The release of Muse Spark Open follows Meta's earlier deployment of its Muse Glimmer architecture, expanding its mixture-of-experts design to a much larger scale. [4][6]