Ai2 Releases OlmoEarth V1.3 With Global 10M Satellite Embeddings

The Allen Institute for AI released OlmoEarth v1.3 on October 22, 2026, an open satellite embedding model accompanied by precomputed annual embeddings for the whole planet at 10-meter resolution and a browser-based OlmoEarth Embeddings Explorer that requires no sign-in.

The release includes the OlmoEarth v1.3 Base model, precomputed global embeddings in 128- and 64-dimensional versions, the Explorer, and a technical report, all under the OlmoEarth Artifact License used for OlmoEarth v1 and v1.2. Previously, using OlmoEarth meant fine-tuning in OlmoEarth Studio, which required GPU access and local imagery storage; embedding models instead compress satellite data into ready-to-use features reusable across many tasks.

According to its model card, OlmoEarth-v1_3-Base is a ViT-Base foundation model with 128 million parameters for remote sensing over Sentinel-1, Sentinel-2, and Landsat images or image time series. It also includes a 14.3-million-parameter Perceiver and a 0.1-million-parameter linear student projecting 768-dimensional tokens to 128 dimensions; the card states that this model, student included, generated OlmoEarth’s global annualized embeddings.

How the Model Produces Embeddings

OlmoEarth v1.3 reads a year of Sentinel-1 radar and Sentinel-2 and Landsat multispectral time series and produces one 128-dimensional vector for every 10-meter pixel on Earth; the first 64 dimensions also work on their own as a smaller embedding. A vision transformer reads the satellite data, an Attention Pooler pools the output tokens into one vector per patch of land, and a linear layer reduces that vector to 128 numbers. The technical report describes a Matryoshka-style distillation whose student is a single linear layer trained jointly with the backbone in the same forward pass, which the report says makes the distillation essentially free.

The report states that running the vision transformer and Attention Pooler at a 20-meter encoder resolution with 10-meter output is 7.7 times cheaper than running the transformer at 10-meter resolution: 308 versus 2,378 GMACs per window. Training used the same data and recipe as OlmoEarth v1.2 — Sentinel-1, Sentinel-2, and Landsat-8 observations with map targets such as WorldCover, WorldCereal, SRTM, and OpenStreetMap — with the AdamW optimizer at a batch size of 512 for 667,200 steps.

Benchmark Results and the PLANTEUR Dataset

Ai2 reports evaluating the v1.3 embeddings against AlphaEarth Foundations and Tessera v2 across 27 evaluations covering six continents; the 128-dimensional embeddings scored highest on 20 of 27, and the 64-dimensional version led on 18 of 27. The technical report states that the 128-dimensional embeddings outperformed both baselines on 7 of 8 AEF datasets (the exception was Ethiopia, where Tessera v2 outperformed OlmoEarth v1.3) and slightly outperformed Tessera v2 on the PASTIS crop-type dataset, 58.6 to 58.3 mIoU.

With the release, Ai2 introduced PLANTEUR, a crop-type benchmark for tropical farming built from labels in Overseas France: Martinique, Réunion, French Guiana, Guadeloupe, and Mayotte. Its leave-one-island-out setup tests performance on one island when classifiers are trained on the others, and Ai2 says the v1.3 embeddings generalize best to places a classifier has never seen. The report records 24.9 mIoU for the 128-dimensional embeddings on that leave-one-island-out measure, versus 19.1 for Tessera v2 and 15.8 for AlphaEarth Foundations.

Ai2 also describes v1.3 as a strong fine-tuning model that matches v1.2’s performance across research and partner labels; the report gives v1.3 a mean fine-tuning rank of 4.34 across all evaluated models and tasks, behind OlmoEarth v1.2 Base at 4.02 and ahead of OlmoEarth v1 Base at 4.74.

Training Cost and Environmental Impact

Ai2 describes OlmoEarth v1.3 as its cheapest OlmoEarth model to train, requiring 53% of the computing time of v1.2 and 18% of v1, with 50% and 22% of the carbon emissions, respectively. Ai2 estimates that reproducing pretraining of OlmoEarth v1.3 Base would cost roughly $2,720 in rented H100 time for the final run.

The report’s environmental impact table lists OlmoEarth v1.3 Base pretraining at 539 H100 GPU hours, 161 kWh, and 0.07 tCO2eq, compared with 1,012 GPU hours and 0.14 tCO2eq for v1.2 Base and 2,989 GPU hours and 0.32 tCO2eq for v1 Base. Computing the embeddings themselves is far larger: the report estimates one year of global embeddings at approximately 14,000 H100 hours, roughly 26 times the pretraining cost.

Global Archive, Explorer, and Licensing

The archive behind the Explorer holds embeddings at 10-meter resolution, roughly 150 TB per year, covering land and adjacent coastal waters. At launch, 2025, 2024, and 2023 are global, while 2017 through 2022 cover select countries, with more coverage in progress. The report puts the full 2017-to-2025 dataset at well over a petabyte, with coverage from 59.5 degrees south to 83.6 degrees north and excluding Antarctica, and states that Searchlight inference removed tiling artifacts in the released embeddings.

The embeddings are hosted on Source Cooperative as a Zarr store readable directly from object storage with any Zarr client, with no API key or tile server required. Open provenance tables released as GeoParquet list the Sentinel-1, Sentinel-2, and Landsat scenes behind each 20-kilometer tile and year; for 2025, the report records about 450,000 windows, 21 million mosaic entries, and 730,000 scenes.

The Explorer renders the embeddings as a false-color map of the planet. Users can flip between years, pull up the underlying Sentinel-2 imagery, click a location to find similar places, train a small classifier from a handful of labeled points, and export the result as a GeoTIFF, with no account required. Ai2’s announcement illustrates the approach with illegal gold mining in Peru’s Madre de Dios region: the post states that mining around La Pampa peaked in 2017 and 2018 and that Peru’s Operation Mercurio shut most of it down in February 2019, and that embedding comparisons between 2017–2020 and 2020–2025 show the earlier spread and continued change in the old mining zone. The post adds that Peru sent navy patrols back into the reserve in August 2026 after prosecutors found mining had crossed into Tambopata itself.

For work that needs a specific time window, OlmoEarth Studio runs the same model over user-chosen dates, modalities, and resolution, and a tutorial shows how to generate embeddings on a user’s own compute. The technical report states that the OlmoEarth Artifact License restricts use for military, defense-related, and extractive-industry applications; the model card adds that the model is intended for use in accordance with Ai2’s Responsible Use Guidelines.