Meta has released the weights of the AI model Muse Spark Open with 816 billion parameters. The development was carried out by the Meta Superintelligence Labs division. The model is designed for programming, complex reasoning, and autonomous operation of AI agents. It is built on the MoE architecture: out of 816 billion parameters, only 42 billion are used for processing each token. The context window is 1 million tokens, and the model accepts text, images, audio, and video. Meta calls Muse Spark Open the most powerful American neural network with open weights to date.
For local deployment, Meta, in collaboration with Apple, optimized the model for Mac Studio with M5 Ultra processors. Each device is equipped with unified memory up to 512 GB with a bandwidth of 1.2 TB/s. According to the developers, two computers in tandem can accommodate the model's parameters entirely and operate without cloud servers. The MLX framework and RDMA over Thunderbolt 5 technology were used to connect the devices.
In tests, two Mac Studios with a tensor distribution scheme delivered about 25 tokens per second, and with a pipeline scheme, about 15.7. With four devices, the speed increased to 36.7 tokens per second with a context of 8192 tokens. Four-bit quantization was used, reducing the memory volume for parameters to approximately 493 GB.
For deployment, two Mac Studios with 512 GB of memory or four devices with 256 GB are required. According to Engadget, the second option will cost approximately $38,000. The 512 GB configuration has not yet been released: Apple promised to launch it at the end of October.
Muse Spark Open is available on Hugging Face under the Muse Model License. It allows commercial use with conditions: companies with revenue over $20 million per year providing access via API need a separate agreement with Meta.
This material is prepared solely for informational purposes and does not constitute financial advice or a recommendation.
Related articles
