Meta has officially released the weights for Muse Spark Open, which the company describes as its most capable US open-weight model to date. The model is a natively multimodal system capable of processing text, image, video, and audio inputs. It features a 1-million-token context window and is specifically engineered for 'agentic' workflows—tasks that require long-horizon planning, multi-step reasoning, and the ability to use external tools to correct errors. Examples of these workflows include shipping real code changes to open-source projects or performing end-to-end video editing by reviewing footage frame-by-frame.
This image serves as the primary visual representation of the Muse Spark Open model weights released by Meta.
Technically, Muse Spark Open is built on a sparse mixture-of-experts (MoE) architecture. While the model contains a total of 816 billion parameters, it only activates approximately 42 billion parameters for each token generated. This design allows the model to maintain high-level reasoning capabilities while remaining computationally efficient enough to run on desktop hardware. Meta has optimized the model for various inference engines, including SGLang, vLLM, and Apple’s MLX framework. Notably, the model can run on two Apple M5 Ultra Mac Studios by leveraging unified memory and tensor parallelism; the two machines act as a single unit where each holds half the weights and performs half the computation, with intermediate results traveling between them over high-bandwidth connections.
While the model is released under the Muse Model License, it includes specific commercial and safety constraints. The license allows for free use by many developers but requires a separate agreement with Meta if a company provides the model as a service (Model as a Service) and exceeds $20 million in annual gross revenue. Additionally, the model includes safety post-training to meet Meta's Superintelligence Scaling Framework thresholds. Developers can further customize the model using PyTorch’s TorchTitan or fine-tune it using various managed APIs and frameworks like Hugging Face TRL and NVIDIA NeMo AutoModel. The weights are currently available for download on major platforms including Hugging Face, Amazon SageMaker, and Google Cloud Model Garden.