Meta Releases Muse Spark Open: 816-Billion-Parameter Open-Weight Model Built for Coding and SI Agents
Meta on October 15 released Muse Spark Open, an open-weight super intelligence model that the company calls its most capable U.S. open-weight model to date; Reuters reported the release the same day.
According to Meta, Muse Spark Open has 816 billion total parameters, with about 42 billion active per inference, a context window of up to one million tokens, and a mixture-of-experts architecture. It is natively multimodal — taking text, image, audio and video as input and producing text — and is aimed at coding and SI agent tasks. The weights are available for download through channels including Hugging Face, with API access offered by several providers; Amazon Web Services announced the same day that the model is available on its SageMaker JumpStart platform.
Meta said the model can run locally on two Mac Studio computers powered by Apple's M5 Ultra chips. In an October 15 post, Meta CEO Mark Zuckerberg noted the collaboration with Apple executive John Ternus and Apple's MLX framework team.
In August, Meta had released Muse Glimmer, a smaller open-weight model. Muse Spark Open is the largest version in the series so far.
Published by GSM · USA (original Chinese edition)
Photo: 图：Meta 位于门洛帕克的总部入口标牌（1 Hacker Way，资料图）。Nokia621 摄，CC BY-SA 4.0，维基共享资源 / Photo: Entrance sign at Meta's headquarters in Menlo Park, 1 Hacker Way (file photo). Nokia621, CC BY-SA 4.0, via Wikimedia Commons
