Meta publishes weights for Muse Spark Open
Meta's description of the model as the most capable US open-weight model, and its benchmark results, are Meta's own and have not been independently rerun. The weights are released under Meta's Muse Model License, not a standard open-source licence; it adds conditions for companies whose model-hosting services exceed $20 million in revenue. Meta recommends deploying the model inside a broader system with additional guardrails.
What was claimed?
Meta released Muse Spark Open with downloadable weights on Hugging Face. Its model card describes a natively multimodal model built for coding and agent work, with 816 billion total and 42 billion active parameters and a 1-million-token context window. Weights come in BF16 and NVFP4 versions, and the card says this version received additional safety post-training for use without system guardrails.
What does the evidence support?
Documented availability, not performance proof
Three days after Meta released Muse Wonder through its own API and Muse Code, this release lets outside developers download, run and host a large Meta model themselves. Running it locally takes substantial hardware: the card lists at least 700GB of GPU memory for NVFP4 and 2TB for BF16.
What are the limits?
Meta's description of the model as the most capable US open-weight model, and its benchmark results, are Meta's own and have not been independently rerun. The weights are released under Meta's Muse Model License, not a standard open-source licence; it adds conditions for companies whose model-hosting services exceed $20 million in revenue. Meta recommends deploying the model inside a broader system with additional guardrails.
Claim or evidence date: 2026-10-15. Last verified: .
Colors classify evidence; they do not rank truthfulness. Explore hype versus reality →