Ai2 releases global OlmoEarth v1.3 satellite embeddings

Ai2 released precomputed annual embeddings for the whole planet at 10-meter resolution, in 64- and 128-dimensional versions.[1] It also released a browser-based Embeddings Explorer for exploring and customizing maps based on those embeddings.[1] Ai2 says the embeddings and Explorer let people begin tracking change over time or finding similar places without signing in or needing advanced technical knowledge.[2] The Explorer offers a way to inspect the mapped data, while the precomputed files provide the model outputs for use in further work.[1]

OlmoEarth v1.3 processes a year of Sentinel-1 radar and Sentinel-2 and Landsat multispectral satellite time series.[1] It produces one 128-dimensional vector for every 10-meter pixel on Earth; the first 64 dimensions make up the smaller embedding.[1] Ai2 describes embeddings as compressed satellite data converted into reusable, ready-to-use features that can be applied cheaply across many tasks.[2] The model card identifies OlmoEarth-v1_3-Base as a 128-million-parameter ViT-Base foundation model for remote-sensing tasks.[4]

At the initial release on October 22, 2026, the 2023, 2024 and 2025 embeddings were global; 2017–2022 covered selected countries and were being backfilled.[3] Coverage excludes Antarctica and extends from 59.5° south to 83.6° north.[3] Access is public and anonymous, with no account, API key or request signing required.[3] Ai2 also says that GPU access and the ability to download and store imagery are required.[2] Teams should confirm the coverage of their target year and location before using the dataset for a project.[3]

Ai2 says it evaluated OlmoEarth v1.3 against AlphaEarth Foundations and Tessera v2 across 27 evaluations spanning six continents and multiple tasks.[1] The 128-dimensional embeddings ranked highest on 20 of 27 evaluations and highest overall, while the 64-dimensional version led on 18 of 27 despite being half the size.[1] Ai2 also says the embeddings generalize best to places a classifier has never seen and introduces PLANTEUR, a crop-type benchmark for tropical farming.[1] The model weights, pretraining data and pretraining code are fully open, according to Ai2.[1]