Overview
- Meta published Muse Spark Open weights on Oct. 15 and made them available for download on Hugging Face, Amazon SageMaker, Google Cloud Model Garden, and Microsoft Foundry, plus hosted APIs from multiple vendors.
- The model uses a mixture-of-experts design with 816 billion total parameters, 42 billion active per token, and a 1 million‑token context window, and Meta positions it for coding, long-horizon tasks, and multi-step agentic workflows.
- Apple and Meta adapted Apple’s MLX framework so Muse Spark Open can run across two to four M5 Ultra Mac Studios that pool up to about 1 terabyte of unified memory, but the required 512 GB or multi‑machine setups make local use costly.
- Meta says it applied additional post‑training safety checks and that the model met thresholds in its Superintelligence Scaling Framework, while its internal benchmarks show agentic and tool-use strengths but lag behind some Chinese open models on pure coding tests.
- The release offers enterprises and researchers a way to run and customize a frontier open-weight model locally, yet licensing rules (including notice and agreement requirements for large commercial operators and a required ‘Built with Muse Spark Open’ label) plus high hardware costs make cloud-hosted access the more practical route for most users.