Overview
- Meta published Muse Spark Open weights on Oct. 15, making an 816‑billion‑parameter model available that activates 42 billion parameters and supports a 1 million‑token context window.
- Apple updated its MLX framework so the model can be split across two to four M5 Ultra Mac Studios, with Apple and Meta reporting generation speeds of roughly 25 tokens per second on two Macs and about 37 on four in 4‑bit tests.
- The release includes broad distribution through Hugging Face and major cloud catalogs plus APIs from Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI, and engineering optimizations such as Helion GPU kernels for NVIDIA and AMD.
- Meta is shipping the weights under a Muse Model License that requires attribution, bars certain sensitive uses, and forces companies with more than $20 million in revenue to notify Meta and sign a separate agreement before offering the model as a service.
- Meta says the model had extra safety post‑training and met its internal Superintelligence Scaling Framework thresholds, but the steep hardware cost and high memory needs mean local use is mainly practical for developers, researchers and enterprises and may shift demand toward expensive Mac Studio configurations and hosted API access.