Overview
- Meta published Muse Spark Open on Thursday, October 15, releasing the model weights on hubs like Hugging Face and making hosted access available through multiple API and cloud partners.
- The model uses a mixture‑of‑experts architecture with 816 billion total parameters, about 42 billion active at a time, and a 1,000,000‑token context window designed for long‑horizon and agentic coding tasks.
- Meta and Apple optimized the model for Apple’s MLX framework so two to four M5 Ultra Mac Studios can run it locally by pooling unified memory, with measured generation speeds near 20–25 tokens per second on two Macs and about 36–37 on four.
- Meta published Muse Spark Open under a Muse Model License that requires larger businesses to notify Meta and possibly sign agreements if they exceed $20 million in revenue and that also mandates product attribution, while Meta says the model passed additional post‑training safety checks under its Superintelligence Scaling Framework.
- Practical limits remain: the on‑device setup needs high‑memory, costly Mac Studio configurations and independent benchmarks show Muse Spark Open leads on agentic/tool use but trails some top Chinese open models on certain coding tests, shaping who can adopt and productize the model.