Meta releases Muse Spark Open with limited commercial license and Mac deployment support
Meta released the weights for Muse Spark Open on October 15. The company describes a natively multimodal model with 816 billion total parameters, 42 billion active parameters and a one-million-token context window, aimed at coding and agentic work.
Meta says the model can run locally using two Mac Studio computers with M5 Ultra and the latest MLX framework. Its described configuration uses up to 512 GB of unified memory per computer. This is a specific high-memory, two-machine setup, not a claim that the model runs on any ordinary Mac.
Weights are listed through Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. Meta also names hosted API providers including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Distribution and hosted access are different deployment routes with their own requirements.
The release names SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, MLX, llama.cpp and Unsloth among supported inference engines. Meta reports performance improvements for selected kernels and systems; those are vendor measurements, not a guarantee of an equivalent end-to-end speed gain on every workload.
The weights use the Muse Model License with limited commercial terms. Meta says many developers can use the model for free, but open weights do not mean unrestricted commercial rights or a free hosted API. The announcement also describes safety post-training and framework-based evaluations, rather than proving every deployment safe.
What to check before adoption: the actual license, chosen model distribution, runtime, hardware and connected tools. Local model execution can keep that inference on the selected machines, but an application may still send information through external services. The release alone does not establish an entire agent workflow as private or reliable.