Ai2 releases satellite AI that lets users build maps without a GPU

Building a crop or land-cover map no longer has to start with downloading satellite imagery and running a model on a GPU. Ai2 released OlmoEarth v1.3 on October 22, 2026, alongside precomputed satellite data and a browser tool that lets users label examples, train a small classifier and export a map without an account.

One satellite summary, many mapping tasks

The model reads a year of Sentinel-1 radar observations and multispectral imagery from Sentinel-2 and Landsat. For each 10-meter pixel, it produces an embedding: a numerical summary of what it learned from those observations. The full summary contains 128 numbers; its first 64 can also serve as a smaller, standalone version.

Previously, the main OlmoEarth workflow involved fine-tuning a model for each task, then running it over satellite imagery to produce a map. That required GPU access and enough storage for the imagery. Precomputed embeddings let different mapping tasks reuse features that Ai2 has already calculated.

A user can train a lightweight classifier—a model that assigns categories—on top of those features and run it locally. Ai2 gives examples including classifying smallholder crops in Kenya, detecting brick kilns near Dhaka and assessing fire damage in Indonesia. Comparing embeddings across years can also highlight landscape changes without training another model.

A browser front end for a large archive

The new Embeddings Explorer displays the data as a false-color map. Users can move between years, click a location to find similar places, or add labeled points to train a classifier. The Explorer also lets users view the underlying Sentinel-2 imagery to check the map against the satellite observations.

Results can be exported as GeoTIFF files. Ai2 says a first pass in the Explorer takes seconds, compared with hours for a fine-tuning run. For work requiring a particular time window, sensor combination or resolution, OlmoEarth Studio can run the same model with those choices instead.

The archive covers land and adjacent coastal waters at 10-meter resolution, with roughly 150 terabytes of embeddings per year. At launch, 2023, 2024 and 2025 have global coverage. Data for 2017 through 2022 covers selected countries, so the length of the available history depends on the place being studied.

Ai2 also publishes provenance tables for every 20-kilometer tile and year. These list the Sentinel-1, Sentinel-2 and Landsat scenes used, with links to their public catalogs. That provides a route from an embedding back to the imagery behind it, rather than leaving users with only the model’s numerical output.

20 of 27Evaluations led by the full embedding

Ai2 says its 128-dimensional embeddings scored highest in 20 of 27 evaluations against AlphaEarth Foundations and Tessera v2. The tests covered six continents and multiple mapping tasks.

Smaller representations, lower training requirements

The smaller, 64-dimensional version led 18 evaluations in Ai2’s tests, offering a half-size representation with a different scorecard. The release also introduces PLANTEUR, a benchmark for classifying crops in tropical farming regions. Ai2 says its embeddings performed best when classifiers were tested in places they had not seen during training.

Ai2 reports that training v1.3 took 53% of v1.2’s computing time and produced 50% of its carbon emissions. It estimates that reproducing the final Base-model pretraining run would cost roughly $2,720 in rented H100 GPU time. That figure applies to the final run, not the entire development effort.

The release includes model weights, pretraining data, pretraining code and a technical report, alongside the precomputed embeddings and Explorer. Ai2 distributes the artifacts under the OlmoEarth Artifact License, the same license used for v1 and v1.2. A tutorial also explains how to generate embeddings on a user’s own computing resources.