Meta has announced Muse Spark Open, a new open-weight model featuring 816 billion parameters. According to reporting from The Decoder, the model handles text, images, video, and audio inputs while supporting up to one million tokens of context. Performance evaluations indicate it sits roughly on par with leading open-weight Chinese AI models on standard benchmarks. Meta has optimized the system to run locally on Apple’s M5 Ultra hardware.
The Open-Weight Landscape
The release positions Meta directly against major competitors in the open-weight space, including models from developers like Qwen, GLM, and Kimi. Open-weight models allow researchers and developers to inspect, modify, and run software locally, shifting control away from purely closed, API-dependent cloud services. Optimizing a model of this scale for local hardware such as Apple silicon indicates a continued push toward decentralized execution.
Unanswered Questions
Despite the initial announcement, several key details remain unavailable. The documentation and disclosure from Meta did not specify precise training dataset compositions, exact hardware deployment requirements beyond the M5 Ultra optimization, or a definitive timeline for broader developer availability. Future updates from Meta will likely clarify how developers can access and implement the model.
As the open-weight race intensifies, the performance and practical utility of Muse Spark Open will depend heavily on real-world adoption and independent testing.
Lead image: Nicolas Foster / Pexels
