Ai2 releases OlmoEarth v1.3 satellite embeddings for whole planet

OlmoEarth v1.3 arrives as a cheaper-to-train successor to earlier versions, cutting compute time and carbon emissions while remaining a strong fine-tuning model that Ai2 says keeps v1.2's performance valuable for research and partner labs. The release also shifts how the model is used: previously, the main path was fine-tuning in OlmoEarth Studio and running a trained model over satellite data, which required GPU access and the ability to download and store imagery.

The new Embeddings Explorer changes that workflow by letting users compare pixels across years without model training. Ai2 illustrates this with Peru's Madre de Dios region, where illegal gold mining around La Pampa peaked in 2017 and 2018 and cleared more than 1,200 hectares of forest in the buffer zone of the Tambopata National Reserve. Comparing embeddings from 2016 to 2020 shows 34 square kilometers of significant change in the southern mining belt, and the same comparison from 2020 to 2024 shows that area falls to 9 square kilometers, making the 2019 shutdown by Operation Mercurio visible in the data.

For teams, the practical effect is a faster first pass: a test that once meant an hours-long fine-tuning run can now take seconds, and the embeddings can help decide whether a task deserves compute time and where the model falls short. Ai2 positions this as a way for anyone to turn OlmoEarth into real work, from tracking change over time to finding similar places.