Ai2 releases OlmoEarth v1.3 with planet-scale satellite embeddings

Ai2 released OlmoEarth v1.3 on Oct. 22, pairing precomputed annual satellite embeddings with a browser-based explorer. The package lets users begin map analysis without first running the full model or processing the underlying satellite archive themselves.

Ai2 says the explorer requires no account or setup. Users can switch among years, inspect Sentinel-2 imagery, search for similar locations, label points to train a small classifier and export results as a GeoTIFF. The organization says an initial analysis can run locally without a GPU.

Each covered 10-meter pixel receives one 128-dimensional vector per calendar year, generated from Sentinel-1, Sentinel-2 and Landsat 8/9 observations. The first 64 dimensions can serve as a smaller standalone embedding. The public dataset is anonymously readable without an account, API key or signed request. It also includes provenance tables identifying the source scenes used for each tile.

The release focuses on reusable satellite representations rather than AI reasoning with executable geospatial tools, addressing a different part of the map-analysis workflow.

Ai2 describes the archive as roughly 150 terabytes per year. Coverage for 2023 through 2025 was global within the dataset’s defined bounds at launch, while data for 2017 through 2022 covered selected countries and was still being backfilled. Those bounds run from 59.5 degrees south to 83.6 degrees north, excluding Antarctica and a narrow strip around the antimeridian. The release therefore does not literally cover every location.

Ai2 also published model weights, pretraining code and pretraining data. The model card describes a 128-million-parameter ViT-Base, a 14.3-million-parameter Perceiver and a 0.1-million-parameter linear student that projects 768-dimensional tokens into the 128-dimensional annual vectors.

Ai2 says the 128-dimensional embeddings placed first in 20 of 27 evaluations against AlphaEarth Foundations and Tessera v2, while the 64-dimensional version led 18 of 27. The research review did not identify an independent reproduction of those results.

The artifacts use the OlmoEarth Artifact License. It permits use, modification and distribution, but imposes attribution and downstream-license requirements and restricts military, defense, surveillance, policing and extractive-industry uses. Those field-of-use limits mean this is not an unrestricted open-source release.