Meta has released the Muse Spark Open AI model, which features 816 billion parameters. The model is designed for coding, complex reasoning, and autonomous agent tasks. It utilizes a Mixture-of-Experts architecture, where only 42 billion parameters are active per token, and supports a 1 million token context window.
The model is natively multimodal, accepting text, images, video, and audio inputs. Meta claims it is the most capable open-weight model from a US-based company to date. It is available for download on Hugging Face and has been optimized for chips from AMD, Intel, and Nvidia.
To run the model locally, Meta collaborated with Apple to optimize it for Mac Studio computers equipped with M5 Ultra processors. A setup of two Mac Studio units, each with 512 GB of unified memory, allows the model to run without cloud servers. The connection between devices is managed via the MLX framework and RDMA over Thunderbolt 5 technology.
In tests using 4-bit quantization, the model requires approximately 493 GB of memory for parameters. With a tensor-parallel scheme, two Mac Studio units generated about 25 tokens per second, while four units reached 36.7 tokens per second. These configurations require significant hardware investment, with a four-unit setup estimated to cost around $38 000.
The model is released under the Muse Model License, which permits commercial use with specific restrictions. Companies with annual revenue exceeding $20 million that offer API access to the model must secure a separate agreement with Meta. The project was developed by the Meta Superintelligence Labs division over several months.
