OlmoEarth v1.3: AI2 Releases State-of-the-Art Satellite Embeddings for the Whole Planet

OlmoEarth v1.3 is a state-of-the-art satellite embedding model from AI2 that turns a year of Sentinel-1, Sentinel-2 and Landsat time series into one 128-dimensional vector per 10-meter pixel, and AI2 has precomputed those embeddings for the entire planet and released them openly. The release matters because it converts global Earth observation from a GPU-heavy fine-tuning problem into a lightweight classification problem that anyone can run from a few labels, with a no-sign-in browser tool for exploring the results.

What OlmoEarth v1.3 actually is

OlmoEarth v1.3 is a satellite embedding model that reads a year of radar (Sentinel-1) and multispectral (Sentinel-2 and Landsat) satellite time series and produces one 128-dimensional vector for every 10-meter pixel on Earth. The first 64 dimensions also work on their own as a smaller, half-size embedding.

AI2 released the model alongside four supporting artifacts:

• Precomputed, annual embeddings for the whole planet at 10-meter resolution, published on Source Cooperative.
• The OlmoEarth Embeddings Explorer, a browser-based tool for exploring and customizing maps from those embeddings.
• A technical report describing results and methodology.
• An interactive story walking through worked examples.

The materials are linked from the announcement: model weights, technical report, data, code, embeddings explorer, and interactive story. The model weights, the pretraining data, and the pretraining code are all fully open, published under the OlmoEarth Artifact License, the same open license used for OlmoEarth v1 and v1.2.

Compute once, reuse everywhere: why embeddings change the workflow

Until OlmoEarth v1.3, the primary way to use OlmoEarth was fine-tuning in OlmoEarth Studio: training a model for each task and running it over satellite imagery to make a map. That workflow requires GPU access and the ability to download and store the imagery.

Embedding models replace that with a reuse pattern. An embedding compresses satellite data into ready-to-use features that can be reused cheaply for many tasks. Because AI2 has already computed the embeddings globally, a crop map, a land-cover map, or a change analysis becomes a lightweight modeling problem: train a small classifier on top and run it locally.

In practice, a small classifier trained on OlmoEarth embeddings is enough to map what is on the ground. AI2 names classifying smallholder crops in Kenya, detecting brick kilns west of Dhaka, assessing fire damage in Indonesia, and tracking gold mining in the Peruvian Amazon as examples.

Architecture in three steps

The embedding pipeline has three stages:

1. A vision transformer reads the satellite data.
2. An Attention Pooler pools it into one vector per patch of land.
3. A linear layer reduces it to 128 numbers.

The result is one vector per 10-meter pixel, with a 64-dimensional truncation available as a lighter alternative.

Benchmark results: 20 of 27 evaluations won

OlmoEarth v1.3's embeddings were evaluated against AlphaEarth Foundations (AEF) and Tessera v2 across 27 evaluations covering six continents and many different tasks. The 128-dimensional embeddings score highest on 20 of 27 evals and rank highest overall. The 64-dimensional version, at half the size, still leads on 18 of 27 evals.

As part of this evaluation, AI2 introduces PLANTEUR, a new crop-type benchmark for tropical farming. The embeddings generalize best to places a classifier has never seen. A separate PLANTEUR technical report is available.

OlmoEarth v1.3 also remains a strong fine-tuning model, matching v1.2's performance across research and partner labels.

Training efficiency: the cheapest OlmoEarth generation so far

OlmoEarth v1.3 was the cheapest OlmoEarth model to train. It took 53% of the computing time of version v1.2 and 18% of the computing time of v1, with 50% and 22% of the carbon emissions respectively.

AI2 frames the model as building on the same foundation as earlier OlmoEarth versions, tuned so the embeddings stay lightweight without giving up accuracy across tasks, with that efficiency work carrying into future versions.

On cost, AI2 states that reproducing the pretraining of OlmoEarth v1.3 Base would cost roughly $2,720 of rented H100 time for the final run. For comparison, the announcement lists $15,060 for OlmoEarth v1 Base and $5,100 for v1.2 Base.

Case study: nine years of illegal gold mining in La Pampa, Peru

A worked example shows what embeddings can do without any model training. In Peru's Madre de Dios region, illegal gold mining around La Pampa 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. In February 2019, Peru's Operation Mercurio shut most of it down.

The embeddings help answer whether it stayed down. Comparing each pixel's vector across years requires no model training. The 2017-to-2020 pair lights up where mining was spreading to the southeast, and the crackdown is visible in the embeddings. Running the same comparison from 2020 to 2025 shows the signal shifting: that corner has gone quiet, and the change now sits in the older mining zone to the west, part regrowth and part renewed digging.

The second map is not dark, which is itself informative. According to MAAP, pockets in the northwest kept operating, and the old mining zone keeps changing as the land recovers. The embeddings register that too. In August 2026, the Associated Press reported that Peru sent navy patrols back into the reserve after prosecutors found mining had crossed into Tambopata itself.

The OlmoEarth Embeddings Explorer

The OlmoEarth Embeddings Explorer is an interactive analysis tool built from the embeddings themselves, rendered as a false-color map of the planet. Its capabilities, as described by AI2:

• Pan anywhere and flip between years to see how landscapes have changed.
• Pull up the underlying Sentinel-2 imagery to check the map against the real thing.
• Click a place to find others that look similar.
• Drop a few labeled points, train a small classifier on the spot, and export the result as a GeoTIFF.
• No account and no setup required.

The workflow comparison matters for practitioners: testing an idea against OlmoEarth has often meant a fine-tuning run on the order of hours, while a first pass in the Explorer takes seconds. Entire regions can be analyzed from a handful of labels with no GPU. When a task does call for fine-tuning, the embeddings can help decide whether it deserves the compute and where the model falls short.

The global embeddings archive: coverage and provenance

The archive behind the Explorer holds embeddings at 10-meter resolution, roughly 150 TB per year, covering land and adjacent coastal waters.

At launch:

• 2025, 2024, and 2023 are global.
• 2017 to 2022 cover select countries, with more coverage on the way.

Embeddings can be retrieved in the full 128-dimensional version or the lighter 64-dimensional version. Every embedding also comes with a record of where it came from: for each 20-kilometer tile and year, provenance tables list the Sentinel-1, Sentinel-2, and Landsat scenes the model read, with links to the public catalogs that host them, so any pixel can be traced back to the imagery behind it.

What ships in the release and how to start

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

For work that needs a specific time window, OlmoEarth Studio runs the same model over the dates, modalities, and resolution the user chooses, and a tutorial in the allenai/rslearn repository shows how to generate embeddings on your own compute. The tools are designed to work together: explore in the Embeddings Explorer, fine-tune in Studio when a task calls for it, and take the map from there.