Ai2 has added custom embedding exports to OlmoEarth Studio, its platform for building Earth-observation models.
Embeddings are compact numerical representations that models create from data. In this case, they represent Earth-observation inputs and can be used for downstream tasks such as similarity search, segmentation, clustering, and unsupervised exploration. Ai2 says the source code and model weights are available publicly alongside the research paper, which lets researchers inspect how the embeddings are generated.
The update matters because geospatial AI work often requires teams to turn large volumes of imagery and related data into reusable features before building specialized applications. Exportable embeddings can be a cheaper starting point than training a new model for every task.
The release does not replace the need for domain validation, especially in scientific or policy-sensitive uses of satellite and Earth-observation data. Its practical value is in making foundation-model representations easier to move from a studio workflow into separate analysis pipelines.