Meta Opens Muse Spark. Hosting It Takes More.
Muse Spark Open expands model control, while its hardware and license terms make self-hosting a serious operating decision.
Meta has released the weights for Muse Spark Open, a sparse mixture-of-experts model aimed at coding, tool use, and long-running agent tasks. It has 816 billion total parameters, activates 42 billion per token, accepts text, image, video, and audio, and supports a one-million-token context window. Gated weights are on Hugging Face, with hosted inference from third parties.
Tweet
The weights create deployment choice, not cheap ownership. Meta’s model card lists at least 700 GB of VRAM for NVFP4 and 2 TB for BF16. Compare hosted inference with a controlled self-hosted test on one representative workflow. Count hardware utilization, latency, platform staffing, data residency, upgrades, and the safety layer around tool calls.
The license may rule out some operating models. It requires “Built with Muse Spark Open” attribution. Model-as-a-service providers whose company group exceeds $20 million in revenue over any 12 months need a separate agreement with Meta. The clause exempts internal use and ordinary product features. Meta recommends human confirmation for irreversible agent actions. Reproduce the tasks that matter because published comparisons combine different test setups and reporting sources.
