Meta has opened public access to the weights of its newest multimodal AI model, Muse Spark Open. A key feature of the release is the targeted optimization of the system for Apple desktop computers powered by M5 Ultra chips: for the first time in its class, the model can run locally on desktop machines rather than exclusively in the cloud. Access to the weights is already available on several major platforms, making the system accessible to developers and researchers worldwide.

Architecture and Model Parameters

Muse Spark Open is a multimodal model with a total of 816 billion parameters, of which 42 billion are actively engaged during each generation step. This "sparse" approach combines the expressiveness of a large model with reasonable computational resource requirements. The model supports a 1-million-token context window, enabling work with extremely long documents, codebases, and video streams without losing coherence.

Local Execution on Apple Silicon

Optimization is achieved through the Apple MLX framework, enabling execution on Apple Silicon chips. Full local operation of the system requires at least 1 terabyte of unified memory. As an optimal configuration, the developers recommend a setup of two Mac Studio computers powered by M5 Ultra processors: the unified memory (up to 512 GB per device) and a throughput of 1.2 TB/s allow the two workstations to function as a single supercomputer.

NVIDIA and AMD Support and Distribution Channels

In addition to local execution on Apple Silicon chips, the model features specially optimized compute kernels to accelerate performance on NVIDIA and AMD graphics processors. This expands the range of hardware platforms where Muse Spark Open can be deployed. At launch, the model is available on Hugging Face, Google Cloud Model Garden, Microsoft Foundry, and OpenRouter.

Real-World Autonomy Tests

During real-world scenario testing, Meta demonstrated the high autonomy of Muse Spark Open when paired with the Muse Code terminal. The AI agent completely and independently edited the code of the open-source iOS Mastodon client called Ice Cubes: fixing bugs, adding new features, and building a fully functional app. Furthermore, the model was utilized for autonomous tasks in the DaVinci Resolve video editing software—the AI analyzed the video footage frame by frame, highlighted the best clips, and performed automatic framing and editing for a vertical format.

Why Meta is Releasing the Weights

According to Meta representatives, releasing open-weight models allows developers to reduce deployment costs, ensures a high level of transparency, and provides the freedom to customize the system for specific needs using the PyTorch framework. Thus, the company is betting on the ecosystem surrounding the open model rather than just closed cloud services.