How to run Clay foundation model embeddings
How it is run today: Self-host open weights or process code
Clay turns multi-sensor satellite chips (Sentinel-2, Landsat, NAIP, Sentinel-1, MODIS, and others) into 1024-dimensional embeddings for a place and time. Those vectors are meant for classification, change detection, and other downstream heads.
Who offers it: the Clay Foundation, as open weights and code. There is no public hosted Clay inference API. You run or deploy the weights yourself.
If a customer asked us to deliver this, we would host the weights. The docs recommend a substantial GPU for wall-to-wall tile work. Exact VRAM for a given batch or chip size is unknown without benchmarking; the encoder checkpoint is on the order of ~1 GB on Hugging Face. We will not claim a specific card until that is measured. We do not host GPUs today.
License on code and weights is Apache, per the Hugging Face model card. The customer supplies georeferenced chips plus sensor band metadata (wavelengths, GSD, and optional time / lat / lon).
We did not build Clay. We do not own the science. We do not guarantee accuracy.
Upstream
Also in the catalog : Clay Foundation Model (v1.5).
Reach out
If this is the job you need, say so. We will not claim we already host it. We will say whether we pass through, stand it up, or cannot do it honestly.