Hugging Face releases OlmoEarth v1.3, precomputed global satellite embeddings and an interactive explorer

Hugging Face published OlmoEarth v1.3, a state-of-the-art satellite embedding model that produces 128-dimensional vectors for every 10-meter pixel on Earth using radar and multispectral satellite time series. A smaller 64-dimensional version is also provided.

The release includes precomputed annual embeddings for the whole planet at 10m resolution, covering land and adjacent coastal waters, with 2025, 2024, and 2023 available globally and 2017–2022 covering select countries.

A new browser-based tool, the OlmoEarth Embeddings Explorer, lets users pan and compare embeddings across years, find similar locations, train lightweight classifiers on labeled points, and export results as GeoTIFFs without an account or GPU.

OlmoEarth v1.3’s embeddings were evaluated against AlphaEarth Foundations and Tessera v2 across 27 benchmarks spanning six continents; the 128-dimensional embeddings ranked highest on 20 of 27 tasks and the 64-dimensional version led on 18 of 27.

The model and resources are fully open: model weights, pretraining data, pretraining code, and a technical report are publicly available, and the team reports a 53% reduction in compute time and an 18% reduction in carbon emissions compared to the prior version.

The embeddings enable lightweight downstream tasks such as mapping smallholder crops in Kenya, detecting brick kilns near Dhaka, assessing fire damage in Indonesia, and tracking gold mining in Peru’s Madre de Dios region.