Meta releases Muse Spark Open for deployment on own hardware
The multimodal model has downloadable weights and is optimized for two M5 Ultra Mac Studios. Local deployment requires substantial memory, and the license includes specific commercial conditions.
Summary
Running advanced models on owned infrastructure has gained another option with Meta’s Muse Spark Open, released on October 15. The company made the weights available for a multimodal model aimed at coding and agents that carry out multistep tasks.
The model has 816 billion total parameters, with 42 billion active per token, and a one-million-token context window. It accepts text, images, audio and video and generates text. This release is distinct from Muse Wonder, which Meta had already introduced.
In practice
Meta optimized deployment with MLX, a machine-learning library for Apple hardware. According to its announcement, two M5 Ultra Mac Studios, each with up to 512 GB of unified memory, can hold the model. Its demonstrations include changes to an iOS app and video editing in DaVinci Resolve performed locally. These are Meta’s demonstrations rather than independent productivity measurements.
The model card and license on Hugging Face show the infrastructure requirements: at least 700 GB of graphics memory for NVFP4 and 2 TB for BF16. The model is also available through hosted API providers, without requiring a self-managed deployment.
For a team, the choice is between paying for a service and taking responsibility for installation, operation and capacity. Access to weights enables customization and control of the environment, but does not remove maintenance costs or make the model suitable for an ordinary computer.
Context
The Muse Model License permits many commercial uses but includes attribution obligations and specific conditions for businesses offering models as a service. Downloadable weights should therefore not be confused with unrestricted use.
Meta recommends additional safeguards tailored to each application, particularly when an agent can act through tools. Local deployment changes where the model runs; permissions, network access and data handling still depend on the surrounding system.
Why it matters
- More models with available weights broaden deployment and customization options.
- Memory requirements remain a barrier to local access at this scale.
- License terms, operating costs and agent permissions belong in comparisons with hosted services.