Ai2 Releases OlmoEarth v1.3 With Global Coverage Excluding Antarctica

The Allen Institute for Artificial Intelligence (Ai2) released OlmoEarth v1.3 on Oct. 22, pairing an updated Earth-observation model with precomputed annual satellite embeddings and a browser-based explorer. The release makes reusable data available for mapping and change analysis at 10-meter resolution, avoiding a fresh run of the full model over satellite imagery for every project.

An embedding is a numerical representation of observed data. OlmoEarth creates one vector for each 10-meter pixel, with either 64 or 128 dimensions. The vectors draw on observations from Sentinel-1, Sentinel-2 and Landsat. Reusing those precomputed representations lets projects work with the same satellite-derived data across tasks such as mapping land cover or examining changes over time.

Annual embeddings cover 2023, 2024 and 2025 across the stated global extent, which runs from 59.5 degrees south to 83.6 degrees north. Antarctica is outside the mapped area. Coverage for 2017 through 2022 includes selected countries.

The browser explorer can search for places with similar features and cluster areas selected by users. It can also export analyses as Python scripts, allowing users to reproduce the work in code. Together, the precomputed data and browser tools offer ways to inspect locations and organize them for downstream analysis.

Pretraining OlmoEarth v1.3 Base used about 539 H100 GPU hours. The comparable pretraining figures are about 1,012 hours for OlmoEarth v1.2 Base and about 2,989 hours for OlmoEarth v1 Base. The v1.3 estimate also lists 161 kilowatt-hours of electricity and 0.07 metric tons of carbon dioxide equivalent; the corresponding emissions estimates were 0.14 metric tons for v1.2 Base and 0.32 metric tons for v1 Base. Those emissions estimates exclude hardware manufacturing and transportation.

Earlier OlmoEarth versions were principally used through model fine-tuning in OlmoEarth Studio. Version 1.3 adds precomputed embeddings that can be reused for multiple analyses, while its explorer supports similarity searches, clustering and script exports.