Meta Opens Muse Spark Weights for Local Mac Studio Clusters

Meta released Muse Spark Open’s weights Thursday, Oct. 15, and partnered with Apple to run the model across Mac Studio clusters, giving developers a local deployment option that needs multiple high-memory computers.

The local deployment uses Apple’s MLX framework to distribute inference across connected computers. The setup links Macs over Thunderbolt 5 and RDMA, allowing the model to run on developers’ hardware instead of through a hosted service.

The model has 816 billion parameters, activates 42 billion per token and supports a 1 million-token context window. It accepts text, images, video and audio. The model is designed for coding and other tasks that involve an AI agent completing multiple steps.

The hardware setup calls for either two M5 Ultra Mac Studios with 512GB of unified memory each, or four systems with 256GB apiece. Both configurations provide 1TB of combined memory and use the NVFP4 version of the model. The 512GB unified-memory Mac Studio configuration was scheduled to arrive in late October 2026.

A decode test on four M5 Ultra Mac Studios returned 36.73 tokens per second. The test used an 8,192-token context and generated 128 tokens; the model’s full context window is 1 million tokens.

Meta listed weights for download on Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face and Microsoft Foundry Models. Hosted API access was listed through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI.

The Muse Model License requires products or services using the model to display a prominent “Built with Muse Spark Open” label. A licensee offering the model or its derivatives through a model-as-a-service product must notify Meta and sign a separate agreement before offering it if the licensee and its affiliates have aggregate gross revenue above $20 million over any consecutive 12-month period.

The release follows Meta’s Monday, Oct. 12, launch of the Muse Wonder models through the Meta Model API and Muse Code. Muse Spark Open adds a downloadable-weights option for local use.