Meta Launches Muse Spark Open: An AI Model with 816B Parameters
Meta’s Superintelligence Labs has unveiled what it refers to as the most advanced open-weight AI model ever developed by a US laboratory. Launched on October 15, 2026, the Muse Spark Open boasts an impressive 816 billion total parameters. However, thanks to its Mixture-of-Experts (MoE) architecture, only 42 billion parameters are active per token, ensuring manageable inference costs.
This model supports an enormous 1-million-token context window and can accept text, images, audio, and video as input. This makes it inherently multimodal from its inception. It was specifically designed for coding and agentic workloads, including long-horizon tasks, complex software engineering, and multi-step reasoning.
In a remarkable feat, Meta has confirmed that the Muse Spark Open can be operated locally on just two Apple M5 Ultra Mac Studio machines using the MLX framework. For those who do not possess the necessary hardware, the model can be accessed via hosted inference on seven cloud platforms, including Groq, Fireworks AI, and Together AI. The weights are available for download on Hugging Face under the Muse Model License, which allows free use for many developers, including fine-tuning and creating derivative works.
The launch, which observers have described as more akin to a coordinated product rollout than a research release, was simultaneously activated across seven inference providers and four weight-distribution channels.
Source: Unite.AI – Meta’s 816B-Parameter Muse Spark Open Weights Are Now Public