OlmoEarth v1.3: State-of-the-art embeddings for the whole planet
Allen Institute (AI2) ● Covered by 3 sources
AI2 released OlmoEarth v1.3 and global satellite embeddings for Earth. It’s built to spot change fast, with no sign-in and no heavy setup.
Based on reporting by Allen Institute (AI2) — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
AI2 has released OlmoEarth v1.3, a satellite embedding model it calls state of the art. Alongside the model, the institute is also publishing precomputed annual embeddings for the whole planet, plus a browser tool called the OlmoEarth Embeddings Explorer. The pitch is simple: people should be able to use the system immediately, without GPU access, sign-ins, or specialized setup.
That matters because the old path was heavier. If you wanted to use OlmoEarth through fine-tuning in OlmoEarth Studio, you had to train a model for each job and run it over satellite imagery to produce a map. Embedding models take a different route. They compress satellite data into reusable features, so a crop map, land-cover map, or change analysis can become a much lighter task on top of the same base representation.
OlmoEarth v1.3 reads a year of radar and multispectral satellite time series and turns each 10-meter pixel on Earth into a 128-dimensional vector. A 64-dimensional version also stands on its own. AI2 says it has now computed those embeddings for the entire planet and shared them on Source Cooperative, so users can train a small classifier locally instead of starting from raw imagery.
The institute says its 128-dimensional embeddings beat AlphaEarth Foundations and Tessera v2 across 27 evaluations covering six continents. They scored highest on 20 of those tests overall. The 64-dimensional version, which is half the size, led on 18 of 27. AI2 also introduced PLANTEUR, a crop-type benchmark for tropical farming, and says the embeddings generalize best to places a classifier has never seen. The whole stack is open: model weights, pretraining data, and pretraining code.
AI2 is also making a broader efficiency claim. Compared with earlier OlmoEarth versions, v1.3 took 53% of the compute time of v1.2 and 18% of v1, with 50% and 22% of the carbon emissions. It still matches v1.2 on research and partner labels when used as a fine-tuning model. That combination of cheaper training and strong reuse is the real story here: the model is less about one giant map job and more about making planetary analysis cheap enough to repeat.
My take — AI-written commentary, not fact-checked reporting
The smart move here is the boring one: ship the embeddings, not just the demo. Too many earth-observation projects stay trapped in the “look what the model can do” phase, which is a nice way to spend compute and a terrible way to help anyone. Open weights plus precomputed globals is how this stuff stops being a lab exhibit and starts acting like infrastructure.
Read more about this at: Allen Institute (AI2)