Clay Foundation Model & Earth Embeddings

Clay is an open Earth-observation foundation model built by the Clay initiative, now a program of Renaissance Philanthropy and originally a fiscally sponsored project of the non-profit Radiant Earth Foundation. The model is a vision transformer trained as a self-supervised masked autoencoder on stacks of satellite imagery, so it learns a general representation of places on Earth from inputs such as Sentinel-2, Sentinel-1 and digital elevation data, together with location and time. Run on your own scenes, it converts raw imagery into compact embeddings you can use for search, classification, change detection and similarity without task-specific labels.

You access Clay in two ways. The model code lives on GitHub and the trained weights on the Hugging Face Hub, both freely downloadable, so you can generate embeddings for any location and date yourself. Separately, the maintainers publish ready-made embedding products on Source Cooperative, distributed per tile as cloud-hosted GeoParquet that you can query or pull in bulk, which saves you running the model over large areas.

For RAG and AI work Clay is an embedding backbone rather than a text corpus: it is best for building geospatial retrieval and similarity systems, feeding a vector index of places, or as a labelled-data-light starting point for downstream Earth-observation tasks. The embeddings slot into the same pipelines you would use for text vectors.

The licence position is split and worth reading before you rely on it. The code and weights are Apache-2.0 and the documentation is CC-BY, all permitting commercial use with attribution. The embedding products vary: some are CC BY 4.0 and some ODC-BY, and embeddings derived from Major TOM inputs can inherit a CC BY-SA share-alike obligation, so a derivative may have to carry the same licence. Check each product's own terms. Clay sits alongside Google's AlphaEarth Foundations embeddings and the Major TOM archive we also list, as the community-driven, fully open member of that group.