Meta releases 816-billion-parameter Muse Spark Open as open weights
Meta published downloadable weights for Muse Spark Open, with a 1 million token context window, while keeping its most capable Muse Wonder models closed.
What happened
Meta published the weights for Muse Spark Open on 15 October, calling it the most capable US open-weight model to date. Weights are available on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry. The release came days after Meta's Superintelligence Lab launched its closed Muse Wonder models.
Specs
The model has 816 billion total parameters, 42 billion active per token, and a 1 million token context window. It accepts text, images, video and audio and outputs text. Full BF16 weights take about 1.5-2 TB. Meta's benchmark claims have not been independently verified.
Background
Open weights mean the trained parameters can be downloaded and run or fine-tuned locally, but training data and code are not shared. A mixture-of-experts model activates only relevant sub-networks per input, so it runs at the cost of a much smaller model.
The licence
The Muse Model License allows free use, fine-tuning, derivatives, distillation and distribution, with attribution and a separate agreement for model-as-a-service businesses above $20 million in revenue over 12 months. It is not OSI-approved.
Why Meta runs two tracks
Meta long championed open models through Llama, but after forming its Superintelligence Lab it began keeping its strongest models closed: Muse Wonder is paid via API and Muse Wonder Full is limited to Muse Code subscribers. Spark Open is strong enough to keep developers in Meta's ecosystem without giving away its most advanced technology, a hybrid approach that is spreading across the industry.
Why it matters
Open-weight leadership had shifted to Chinese labs such as DeepSeek, Qwen and Kimi. Meta's release brings a strong US option for banks, governments and others that cannot send data outside.
What comes next
Independent benchmark results are due in the coming weeks, and the community is expected to produce smaller quantised versions that run on modest hardware.
What it means for your business
Try a cloud-hosted version before buying hardware, consider self-hosted models for data that must not leave your environment, and read the licence if you plan to serve the model to customers.