Muse Spark Open 816B model runs on dual Mac Studios with 42B active parameters
Meta has published open-weight model files optimised for local execution on Apple silicon, enabling on-device agentic workflows.
Key points
- Muse Spark Open features 816 billion total parameters with 42 billion active.
- The model requires two Mac Studio units with M5 Ultra chips to run locally.
- Meta states the release represents the most capable US open-weight model.
Meta has released the weights for Muse Spark Open, a large multimodal artificial intelligence model designed to run locally on Apple Mac Studio computers. The company describes the release as the most capable open-weight model available in the United States. This announcement marks a significant shift in hardware requirements for running advanced generative AI systems outside of cloud data centres.
Root cause
The technical foundation for this deployment relies on Apple’s unified memory architecture within the M5 Ultra chip. This design allows the integrated graphics processor to access up to 512 gigabytes of memory per machine. According to Meta, this high bandwidth and capacity enable two separate Mac Studios to function collectively as a single system with a terabyte of unified memory.
View the original post on X
Attack path
This is not a security vulnerability but a software release. The model is built for coding and agentic work, excelling at long-horizon tasks and multi-step reasoning. Meta tested the system by shipping real code changes to an open-source Mastodon client and driving video editing software. The model reviews footage frame by frame to identify and crop vertical video segments automatically.
Affected versions
The model is optimised specifically for the Mac Studio powered by the M5 Ultra processor. It leverages the latest version of MLX, an open-source array framework created by Apple. The hardware configuration requires two machines, each with 512 gigabytes of memory, to hold the entire model. Apple has not yet released the 512GB Mac Studio, though it is expected later this month.
Mitigation
There is no security patch required as this is a new software release. The model weights are available for download through Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models. Users can also access the model via API platforms including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. The license allows limited commercial use.
Why it matters
Teams usually evaluate local AI deployment based on data privacy and latency requirements. Running models on-premises prevents sensitive code or media from leaving the organisation’s network perimeter. Security operations should verify that local hardware meets the specified memory and bandwidth thresholds before attempting deployment. Check whether the unified memory architecture supports the required model size without performance degradation. Ensure that the local environment has the necessary compute resources to handle multi-step reasoning tasks efficiently.
Background: Generative AI
Generative AI creates new content, code, or media by learning statistical patterns from vast datasets. Traditional AI predicts outcomes or classifies inputs based on defined rules or supervised learning. The former introduces risks of hallucination and data leakage; the latter offers predictable, auditable decision-making paths for security operations.
Read the full guide: Generative AI vs Traditional AI: Operational Differences and Risk Profiles
What to do and how to stay safe: Meta
- Verify that your hardware inventory includes the specific M5 Ultra Mac Studio units with 512GB memory.
- Confirm that the latest version of the MLX framework is installed and compatible with your operating system.
- Review the model’s license terms to ensure your intended commercial use falls within the permitted scope.
- Monitor system resources during initial testing to ensure the dual-machine setup maintains stable performance.
Step-by-step guide: Password Policies Examples That Reduce Password Fatigue
General security guidance from the Vector Bulletin newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
What are the hardware requirements for running Muse Spark Open locally?
You need two Mac Studio units equipped with M5 Ultra chips and 512GB of memory each.
Is the model available for commercial use?
Yes, the model is free to download and includes a license that allows limited commercial use.
Which companies are involved in this release?
Meta developed the model and worked with Apple to optimise it for Mac Studio hardware.
