Ai2 releases OlmoEarth v1.3 satellite embeddings, trained with 53% of prior compute

Ai2, the Allen Institute for AI, has released OlmoEarth v1.3 along with precomputed annual satellite embeddings covering the whole planet. Each 10-by-10-meter patch of ground gets a numerical fingerprint built from a year of radar and optical imagery, and anyone can explore them in a browser without signing in.

The efficiency numbers are the company's own. Ai2 says v1.3 took 53% of the computing time that v1.2 needed to train and 18% of what v1 required, with carbon emissions at 50% and 22% of those earlier versions. Until now, using the model meant fine-tuning it for each task, which takes hours of GPU time and a lot of stored imagery. In the new explorer, Ai2 says a first pass on an idea takes seconds.

The example Ai2 leans on is Peru's Madre de Dios, where illegal gold mining cleared more than 1,200 hectares of forest around the Tambopata reserve. Comparing embeddings across years, without training a new model, shows 34 square kilometers of change between 2016 and 2020 and 9 square kilometers between 2020 and 2024. That is a useful check on a government crackdown, though it is one illustrative case rather than a measured benchmark.

The obvious users are conservation groups, agricultural analysts and land-use researchers who lack the budget for large compute. Ai2 also positions the embeddings as a cheap screen: teams can decide whether a task is worth the cost of fine-tuning before they spend it.