Meta AI has published the open weights for its flagship Muse Spark Open model on Hugging Face. It's a multimodal MoE system with 816 billion total parameters, activating 42 billion per token (816B-A42B across 256 experts, routing 8 experts per token).
Built for autonomous coding and agentic loops, the model features a 1-million-token context window—big enough to hold entire repositories and their documentation at once.
Key details:
- Local deployment: Apple MLX and NVFP4 support let you run it locally on two linked Mac Studios with 192 GB+ of unified memory.
- Data privacy: frontier-class performance running entirely inside your own perimeter.
- Broad compatibility: native support in vLLM, SGLang, TensorRT-LLM, and llama.cpp.
We previously examined the Meta Muse Wonder family and open-weight releases—now researchers have full independence from the cloud.
