OlmoEarth v1.3: Satellite Embeddings for the Entire Planet

OlmoEarth v1.3: Satellite Embeddings for the Entire Planet

Published October 22, 2026

Ai2 has released OlmoEarth v1.3, a satellite embedding model designed to represent landscapes across the planet. The release includes precomputed annual embeddings at 10-meter resolution in 64- and 128-dimensional formats, along with the browser-based OlmoEarth Embeddings Explorer.

The model and related resources include the OlmoEarth v1.3 Base model, technical report, precomputed data, pretraining code, and the Embeddings Explorer.

Compute once, reuse across tasks

An embedding is a numerical representation of information learned by a model. OlmoEarth v1.3 processes a year of radar data from Sentinel-1 and multispectral data from Sentinel-2 and Landsat. It produces one 128-dimensional vector for every 10-meter pixel on Earth. The first 64 dimensions can also be used independently as a smaller embedding.

Because the embeddings have already been computed for the planet, they can be reused for tasks such as crop mapping, land-cover classification, and change analysis. These tasks can often be reduced to training a small classifier on top of the stored features and running it locally.

The model uses a vision transformer to process satellite time series. An Attention Pooler combines the information into one vector for each land patch, and a linear layer reduces the result to 128 values.

Evaluation results

Ai2 evaluated OlmoEarth v1.3 against AlphaEarth Foundations (AEF) and Tessera v2 across 27 evaluations covering six continents and multiple tasks.

• The 128-dimensional embeddings achieved the highest score on 20 of 27 evaluations and ranked highest overall.
• The 64-dimensional embeddings ranked first on 18 of 27 evaluations.
• The evaluation introduced PLANTEUR, a crop-type benchmark for tropical farming.
• OlmoEarth generalized best to locations that were not represented in classifier training data.

The model weights, pretraining data, and pretraining code are available openly. A small classifier built on the embeddings can be used for applications such as identifying smallholder crops in Kenya, detecting brick kilns west of Dhaka, assessing fire damage in Indonesia, and monitoring gold mining in the Peruvian Amazon.

Compared with earlier OlmoEarth versions, v1.3 required less training compute. It used 53% of the computing time required by v1.2 and 18% of the time required by v1. The associated carbon emissions were 50% and 22% of those versions, respectively. OlmoEarth v1.3 also matched v1.2's performance as a fine-tuning model across research and partner labels.

The reported rented H100 pretraining cost for the final OlmoEarth v1.3 Base run was approximately $2,720, compared with $5,100 for OlmoEarth v1.2 Base and $15,060 for OlmoEarth v1 Base.

Comparing nine years of change

OlmoEarth embeddings can be compared across years to identify changes without training a new model for every analysis.

In Peru's Madre de Dios region, illegal gold mining around La Pampa peaked in 2017 and 2018. Amazon Conservation's MAAP monitoring program reported that mining cleared more than 1,200 hectares of forest in the buffer zone of the Tambopata National Reserve. Peru's Operation Mercurio shut down most of the activity in February 2019.

Comparing embeddings from 2017 and 2020 highlights the expansion of mining toward the southeast. A comparison between 2020 and 2025 shows a different pattern: activity has shifted toward the older mining zone to the west, where the embeddings also capture land recovery and renewed disturbance.

According to MAAP, pockets of mining continued in the northwest, while the former mining area continued to change as vegetation recovered. In August 2026, the Associated Press reported that Peru sent navy patrols back into the reserve after prosecutors found that mining had extended into Tambopata itself.

The OlmoEarth Embeddings Explorer

The OlmoEarth Embeddings Explorer presents the embeddings as a false-color map. Users can pan across the planet, switch between available years, and view the underlying Sentinel-2 imagery for comparison.

The tool also supports several forms of interactive analysis:

• Find locations with similar embeddings by selecting a place on the map.
• Add labeled points and train a small classifier directly in the browser.
• Export classification results as a GeoTIFF.

The first analysis pass can take seconds and does not require a GPU or an account. The tool can help determine whether a task requires fine-tuning and identify areas where the model may not perform adequately.

The archive stores embeddings at 10-meter resolution, with approximately 150 TB of data per year covering land and adjacent coastal waters. At launch, global coverage is available for 2023, 2024, and 2025. Selected countries are covered for 2017 through 2022, with additional coverage planned.

The data is available in both 128-dimensional and 64-dimensional formats. Each embedding includes provenance information. For every 20-kilometer tile and year, provenance tables identify the Sentinel-1, Sentinel-2, and Landsat scenes used by the model and link to the public catalogs containing those scenes.

Using OlmoEarth v1.3

The release includes the OlmoEarth v1.3 Base model, precomputed global embeddings, the Embeddings Explorer, and the technical report.

For projects requiring a specific date range, sensor combination, or resolution, OlmoEarth Studio can run the model over selected inputs. A tutorial explains how to generate embeddings on local compute.

The tools are released under the OlmoEarth Artifact License, the same open license used for v1 and v1.2.