AINEWS
OlmoEarth v1.
Ai2 released OlmoEarth v1.3 today as open weights, with precomputed 10-meter embeddings for the whole planet on Source Cooperative, and a browser-based Embeddings Explorer that needs no sign-in or GPU. A small team or Global South researcher can start using it before lunch.
A satellite embedding is a year of raw orbital data compressed into a short list of numbers. OlmoEarth v1.3 reads a full year of Sentinel-1 radar plus Sentinel-2 and Landsat multispectral imagery, runs it through a vision transformer, and condenses every 10-meter patch of land into a 128-dimensional vector: 128 numbers that capture what the satellite saw at that spot over the year. The first 64 numbers also work on their own as a smaller embedding for teams with tighter memory budgets.
The expensive part is precomputing those vectors. Ai2 has already done it for the entire planet at 10-meter resolution and posted the result on Source Cooperative, around 150 TB per year. The 2023, 2024, and 2025 archives are global; 2017 to 2022 cover selected countries. Every 20-kilometer tile ships with a provenance table listing the Sentinel-1, Sentinel-2, and Landsat scenes behind it. For downstream users, the trade-off changes: a laptop is enough to train a small classifier on top of those vectors, and a user can export the result as a standard GeoTIFF.
The access surface has four doors. Model weights, pretraining data, and pretraining code sit on Hugging Face under the OlmoEarth Artifact License, the same open license as v1 and v1.2. The precomputed tiles are on Source Cooperative. OlmoEarth Studio runs the same model over user-chosen dates, modalities, and resolutions for anyone with a GPU. The Embeddings Explorer, a browser app with no account or setup, lets a user click a place to find the most similar places on Earth or drop a few labeled points and train a small classifier on the spot.
The headline evaluation: Ai2's own suite ranked OlmoEarth v1.3 against DeepMind's AlphaEarth Foundations and Tessera v2 across 27 evaluations spanning six continents. The 128-dimensional version scored highest on 20 of those 27, and the 64-dimensional version, at half the vector size, still led on 18. Ai2 ran the suite themselves; no independent reproduction is published alongside it, so the 20-of-27 ranking reflects Ai2's own benchmark, not a community-agreed leaderboard outcome.
The strongest check on the release cuts the other way. PLANTEUR, a new crop-type benchmark Ai2 also published today, extends the older PASTIS dataset to six tropical regions, including Réunion, Guadeloupe, Martinique, Guyane, Mayotte, and Saint-Martin, with 2,750 aligned time series and 4,506 square kilometers of parcel-derived labels across seven crop classes including sugarcane, banana, pineapple, and tropical tubers. The PLANTEUR authors write that every method they evaluated, OlmoEarth included, scores substantially lower on PLANTEUR than on PASTIS, and every model loses roughly half its accuracy when tested on an island it never saw during training. Tropical generalization is a real open problem. The dataset card and code repo are both public.
The cost story is also concrete. Compared with v1.2, OlmoEarth v1.3 used 53% of the compute time and 50% of the carbon emissions; compared with v1, 18% of the time and 22% of the carbon. The "$2,720 of rented H100 time" figure in the source is scoped to the final pre-training run, not the total project cost.
The release includes a worked visualization: illegal gold mining around La Pampa in Peru's Madre de Dios region peaked in 2017 and 2018, clearing more than 1,200 hectares of forest in the buffer zone of the Tambopata National Reserve, according to Amazon Conservation's MAAP monitoring program. Peru's February 2019 Operation Mercurio shut most of it down. The Associated Press reported in August 2026 that Peru sent navy patrols back into the reserve after prosecutors found mining had crossed into Tambopata itself. The embeddings surface that history on a map; they do not validate their own accuracy against it.
The 20-of-27 ranking is Ai2's own; whether the access delta holds up against AlphaEarth Foundations and Tessera v2 on third-party benchmarks is the open question the PLANTEUR results are already pointing at.