Overview
- Muse Spark Open, which Meta released Oct. 15, 2026, makes full model weights available for download and hosting and is also offered through multiple API and cloud partners including Hugging Face, Amazon SageMaker, Google Cloud Model Garden, Baseten and Together AI.
- The model is a mixture‑of‑experts architecture with 816 billion total parameters, 42 billion active per token, and a 1,000,000‑token context window designed for agentic workflows, long‑horizon reasoning, and complex software engineering tasks.
- Meta worked with Apple to optimize an MLX adapter so the model can run locally across linked M5 Ultra Mac Studios by pooling unified memory, but the smallest practical desktop setup needs multiple high‑memory Macs and still yields modest throughput of roughly 20 tokens per second on two M5 Ultras and about 36 tokens per second on four in reported tests.
- The weights are released under the Muse Model License, which requires companies with more than $20 million in annual revenue to notify Meta and negotiate terms before offering the model as a service, and Meta says the model passed additional post‑training safety checks under its Superintelligence Scaling Framework.
- Making a frontier open‑weight model downloadable broadens options for self‑hosting and research but keeps use concentrated among well‑funded teams because of steep local hardware costs, and public tests on platforms like Arena.ai and independent benchmarks will be key to judging real‑world performance and safety.