AllenAI releases OlmoEarth v1.3 with precomputed global embeddings

AllenAI has released OlmoEarth v1.3, a satellite embedding model with precomputed annual data for the entire planet. The update includes 10-meter resolution vectors in 64 and 128 dimensions and a browser-based tool called the OlmoEarth Embeddings Explorer. Users can track change over time or find similar locations without signing up or needing technical expertise.

Previously, working with OlmoEarth required fine-tuning in OlmoEarth Studio. That process demanded GPU access and the ability to download and store satellite imagery. Embedding models compress satellite data into features ready for reuse, removing the need for heavy training for every new task.

The team aims to give users more ways to explore planetary data. They have computed these embeddings for the whole world and shared them via Source Cooperative. This turns tasks like crop mapping or change analysis into lightweight problems. Users train a small classifier on top of the precomputed data and run it locally.

Compute once, reuse everywhere

An embedding captures what the model learned from millions of images. 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. The first 64 dimensions function independently as a smaller embedding.

Foundation models and global embeddings cost a lot to produce. To ensure access, AllenAI computed these for the entire planet and shared them with the community. A crop map or land-cover map becomes a lightweight modeling problem. Users train a small classifier and run it locally.

The process involves a vision transformer reading satellite data. An Attention Pooler reduces this to one vector per patch of land. A linear layer then reduces it to 128 numbers.

State-of-the-art embeddings

AllenAI evaluated OlmoEarth v1.3 against AlphaEarth Foundations and Tessera v2 across 27 evaluations covering six continents. The 128-dimensional embeddings scored highest on 20 of 27 evaluations and rank highest overall. The 64-dimensional version leads on 18 of 27 evaluations.

The team introduced PLANTEUR, a new crop-type benchmark for tropical farming. The embeddings generalise best to places a classifier has never seen. The technical report for PLANTEUR is available online. Everything behind the model is open: weights, pretraining data, and code.

In practice, a small classifier trained on these embeddings maps what is on the ground. Users can classify smallholder crops in Kenya, detect brick kilns west of Dhaka, assess fire damage in Indonesia, or track gold mining in the Peruvian Amazon.

OlmoEarth v1.3 was the cheapest model to train in the series. It took 53% of the computing time of v1.2 and 18% of v1. Carbon emissions were 50% and 22% of the previous versions respectively. It matches v1.2 performance across research and partner labels. The model builds on earlier foundations, tuned so embeddings stay lightweight without losing accuracy. This efficiency work continues into future versions.

Full results and methodology are in the technical report. Reproducing the pre-training for OlmoEarth v1.3 Base would cost roughly $2,720 in rented H100 time for the final run. The weights, data, and code are open.

One place, nine years

The visualisation shows pixel embedding changes between 2017 and 2020, and 2020 and 2025. Magenta indicates the largest change. The left map shows mining spreading southeast with a large patch at the lower right. The right map shows that corner quiet. The change now sits in the older mining zone to the west. Part is regrowth and part is renewed digging. Users can explore this in the OlmoEarth Embeddings Explorer.

In Peru’s Madre de Dios region, illegal gold mining around La Pampa peaked in 2017 and 2018. It cleared more than 1,200 hectares of forest in the buffer zone of the Tambopata National Reserve. Amazon Conservation’s MAAP monitoring program reported this. In February 2019, Peru’s Operation Mercurio shut most of it down. The question has been whether the operations stayed down.

The embeddings help answer that. Comparing each pixel’s vector across years requires no model training. The 2017 to 2020 pair lights up where mining spread to the southeast. Running the same comparison from 2020 to 2025 shifts the signal. The crackdown is visible in the embeddings.

The second map shows more. It is not dark. MAAP reported pockets in the northwest kept operating. The old mining zone keeps changing as the land recovers. The embeddings register this too. The reprieve was brief and pressure is building again. In August 2026, the Associated Press reported Peru sent navy patrols back into the reserve after prosecutors found mining had crossed into Tambopata itself.

The interactive story opens on this stretch of Madre de Dios and walks through other examples from the region.

The OlmoEarth Embeddings Explorer

The OlmoEarth Embeddings Explorer is an interactive analysis tool built from the embeddings. It renders them as a false-color map of the planet. Users can pan anywhere and flip between years to see how landscapes change. They can pull up underlying Sentinel-2 imagery to check the map against reality. Clicking a place finds others that look similar. Dropping labeled points trains a small classifier on the spot. The result exports as a GeoTIFF. No account or setup is required.

Testing an idea against OlmoEarth often meant a fine-tuning run taking hours. In the explorer, a first pass takes seconds. Users can analyse entire regions from a handful of labels without a GPU. When a task calls for fine-tuning, the embeddings help decide if it deserves the compute and where the model falls short.

The archive holds embeddings at 10-meter resolution. This is roughly 150 TB per year, covering land and adjacent coastal waters. At launch in 2025, data for 2025, 2024, and 2023 is global. Coverage from 2017 to 2022 covers select countries. More coverage is coming. Users can retrieve embeddings in the full 128-dimensional version or the lighter 64-dimensional version. Every embedding includes a record of its source. Open provenance tables list the Sentinel-1, Sentinel-2, and Landsat scenes the model read for each 20-kilometer tile and year. Links to public catalogs allow tracing any pixel back to the imagery.

Get started

This release includes the OlmoEarth v1.3 Base model, precomputed global embeddings, the OlmoEarth Embeddings Explorer, and the technical report. If work needs a specific time window, OlmoEarth Studio runs the model over chosen dates, modalities, and resolution. A tutorial shows how to generate embeddings on own compute. Tools work together: explore in the Embeddings Explorer, fine-tune in Studio when needed, and take the map from there. Everything is published under the OlmoEarth Artifact License, the same open license as v1 and v1.2.

Users can pull embeddings for a place they care about, ask questions, and show what they find.