Clay Foundation

license

Apache-2.0(OSI)

weights

open(Apache-2.0 checkpoint)

data

components-listed(Landsat 8/9, Sentinel-2 L2A, Sentinel-1 RTC, NAIP, LINZ and MODIS named one by one with bands and resolution

code

open(trainer, data module, configs, and model implementation released)

The full training machinery - trainer, data module and configs - is public, and the training-data page names each imagery source in turn rather than describing the corpus in prose. The assembled 33.8TB sample is not redistributed, which keeps Clay at open weights rather than fully open.

• https://clay-foundation.github.io/model/release-notes/data_sampling.html recorded 2026-10-13

  The training-data page names each imagery source in turn with its coverage and ground resolution - Landsat 8/9 at 30m, Sentinel-2 L2A at 10m, Sentinel-1 RTC at 10m, NAIP under 1m, LINZ under 0.5m and MODIS at 500m - and gives the assembled scale as 33.8TB across about 70 million chips, with the sampling index and stacchip processing code published but no download for the assembled corpus.

• https://raw.githubusercontent.com/Clay-foundation/model/main/README.md recorded 2026-10-13

  States that code and trained weights use Apache, and documents trainer.py, ClayDataModule, and configuration-driven training.

Downloads of the declared checkpoint on Hugging Face are the only adoption signal available for Clay.

• https://huggingface.co/api/models/made-with-clay/Clay recorded 2026-09-10

  downloads: 107; gated: false; private: false.

Clay documents two readable attributes, a configurable training path and sensor-, time- and location-aware inputs, one fewer than Prithvi-EO documents; capabilities the release page omits are not inferred from the architecture.

• https://raw.githubusercontent.com/Clay-foundation/model/main/README.md recorded 2026-09-10

  Documents a configurable training path and sensor, time, location, and band metadata consumed by the model interface.