OlmoEarth v1.3 launches with global precomputed satellite embeddings and no-login explorer

By AI123

On October 23, 2026, OlmoEarth v1.3 was released, a satellite embedding model that converts a year of radar (Sentinel-1) and multispectral (Sentinel-2 and Landsat) time series into 128-dimensional vectors for every 10-meter pixel on Earth. The first 64 dimensions are also available as a standalone lightweight embedding.

The release includes globally precomputed annual embeddings at 10m resolution in 64- and 128-dimensional forms, along with a browser-based tool called the OlmoEarth Embeddings Explorer. Users can browse the false-color planet map, compare years, find similar places, train small classifiers, and export GeoTIFF results without signing in.

In evaluations against AlphaEarth Foundations (AEF) and Tessera v2 across 27 tasks on six continents, the 128-dimensional embeddings scored highest on 20 evals and ranked highest overall; the 64-dimensional version led on 18 evals. The team also introduced PLANTEUR, a new crop-type benchmark for tropical farming.

The v1.3 training run used 53% of the compute time of v1.2 and 18% of v1, with 50% and 22% of the carbon emissions, while matching v1.2's fine-tuning performance. Model weights, pretraining data, and code are fully open; reproducing the final run would cost about $2,720 in rented H100 time.

The archive holds embeddings for 2023–2025 globally and for select countries from 2017–2022, with each 20-kilometer tile including provenance tables linking back to source Sentinel and Landsat imagery.

Explore topicsModelsOlmoEarth

Source

View all