Meta has released Muse Spark Open weights and described a local deployment using two Apple M5 Ultra Mac Studios. The Meta AI announcement positions the model for coding, long-running agent tasks and multimodal work.
Understand the size of the model
Muse Spark Open has 816 billion total parameters and 42 billion active parameters. Meta also lists a one-million-token context window.
The difference between total and active parameters matters when interpreting the deployment story. Sparse activation changes the work performed for a token, while storing the model still creates a substantial memory requirement. The announcement's desktop example should not be read as a claim that the full model fits an ordinary laptop.
Read the local demonstration in context
Meta describes running the model across two Mac Studios using MLX. The announcement says each M5 Ultra system can provide up to 512GB of unified memory, allowing the pair to hold the model.
The publisher reports on-device demonstrations involving software changes and video editing. Those are Meta's examples of the system's behavior, rather than independently reproduced results in this article.
For someone evaluating local deployment, the relevant next check is the exact memory, inference-engine and model-format configuration needed for their own hardware and workload.
Check hosted routes and model terms
The announcement lists hosted inference partners and downloadable weights. It also names inference engines including SGLang, vLLM, TensorRT-LLM, MLX and llama.cpp, with support and optimization depending on the platform.
The weights use the Muse Model License with limited commercial terms. “Open-weight” therefore should not be treated as a statement that every commercial use has unrestricted permission.
Meta says it conducted safety evaluations under its own scaling framework and points readers to model documentation. Compare the model's actual task behavior, access route and applicable terms before deciding whether a local or hosted setup makes sense.