Ai2 releases OlmoEarth v1.3 satellite embedding model with open global 10m embeddings

Ai2 released OlmoEarth v1.3 on October 22, 2026, a satellite embedding model that converts a year of radar and multispectral imagery into a 128-dimensional vector for every 10-meter pixel on Earth. Ai2 also published precomputed global embeddings and a browser-based tool, the OlmoEarth Embeddings Explorer. Model weights, pretraining data and pretraining code are all open.

What shipped

• OlmoEarth v1.3 Base, the model itself.
• Precomputed annual embeddings at 10m resolution, in 128-dimension and 64-dimension versions, hosted on Source Cooperative. The first 64 dimensions of the 128-dimension vector also work standalone.
• OlmoEarth Embeddings Explorer, which needs no sign-in. Users can pan across years, find places similar to a selected point, or place labeled points to train a small classifier and export the result as a GeoTIFF.
• A technical report and the PLANTEUR benchmark, a new crop-type benchmark for tropical farming.

How it works

The model reads a year of Sentinel-1 radar plus Sentinel-2 and Landsat multispectral time series. A vision transformer processes the data, an Attention Pooler produces one vector per patch of land, and a linear layer reduces it to 128 numbers. Parameter count, context window and pricing are not disclosed in the source material. The weights are released openly under the OlmoEarth Artifact License, the same license as v1 and v1.2.

Benchmark claims

Ai2 compared v1.3 against AlphaEarth Foundations (AEF) and Tessera v2 across 27 evaluations covering six continents. These results come from Ai2's own evaluation, and the full methodology is in the technical report.

• The 128-dimension embeddings score highest on 20 of 27 evaluations and rank highest overall, according to Ai2.
• The 64-dimension version leads on 18 of 27.
• Ai2 says the embeddings generalize best to places a classifier has not seen, based on PLANTEUR.

Ai2 also says v1.3 matches v1.2's fine-tuning performance across research and partner labels.

Training cost

Ai2 says v1.3 is its cheapest OlmoEarth model to train:

• 53% of v1.2's compute time, with 50% of the carbon emissions.
• 18% of v1's compute time, with 22% of the carbon emissions.
• Reproducing the final Base run would cost roughly $2,720 in rented H100 time, per Ai2.

Coverage and data

The archive runs to roughly 150 TB per year and covers land and adjacent coastal waters. 2023, 2024 and 2025 are global at launch. 2017 to 2022 cover select countries, with more promised. Each 20-kilometer tile and year has open provenance tables listing the Sentinel-1, Sentinel-2 and Landsat scenes the model read, linked to public catalogs.

Ai2 demonstrates change detection in Peru's Madre de Dios region by comparing per-pixel vectors across years with no model training. The 2017–2020 comparison shows mining spreading southeast, and the 2020–2025 comparison shows continued change in the older mining zone to the west.

What this means

The practical shift is in cost structure. Until now, using OlmoEarth meant fine-tuning in OlmoEarth Studio, which needs GPU access and imagery downloads. Precomputed global embeddings reduce many mapping tasks to training a small classifier on a laptop, which widens access for conservation groups, agronomists and researchers without GPU budgets.

The comparison with AlphaEarth Foundations and Tessera v2 matters because open weights, data and code are less common among global embedding products. That openness lets outside groups verify the 20-of-27 claim, and the provenance tables let users trace any pixel to its source scenes. Independent replication of the benchmark results is the key open question.

The coverage gap is the main limit. Only 2023–2025 are global, and earlier years are partial, so long-term change analysis is restricted to select countries for now.