For two years, gold miners moved through the rainforest along the edge of Peru’s Tambopata National Reserve, clearing more than 1,200 hectares around a settlement called La Pampa, according to Amazon Conservation’s MAAP monitoring program. In February 2019, the country’s government launched a crackdown to shut down the mining operations. Since then, conservationists have asked whether the miners really left.
Researchers at the Allen Institute for AI (Ai2) found an answer using OlmoEarth v1.3, the newest version of its satellite AI model. They had already run the model over imagery of the entire planet, producing a yearly summary of what every 10-meter square of land looked like. For La Pampa, the team compared those summaries from one year to the next and marked areas that had changed. Between 2016 and 2020, as mining operations pushed east through the area, the marked area covered roughly 13 square miles, a little more than half the size of Manhattan. But between 2020 and 2024, it shrank to about 3.5 square miles. Some of that remaining change comes from operations in the northwest that never stopped, according to MAAP, and some comes from the forest reclaiming the old mining zone.
But what makes the result notable is how little work it took. Ai2 said it never taught the model what a gold mine looks like or trained anything new. Instead, the answer came from comparing summaries the AI lab already computed and is now sharing. With OlmoEarth v1.3, Ai2 is publishing those yearly summaries for the entire planet, which it calls “embeddings,” along with a free browser tool anyone can use to explore Earth, all without needing an account or specialized hardware.
What’s Changed with OlmoEarth v1.3
First released in 2025, OlmoEarth is a platform with an open multimodal foundation model pre-trained on approximately 10 terabytes of global satellite and sensor data. It’s designed to give developers and researchers a better understanding of what’s happening to Earth, something traditionally limited to well-resourced organizations. Beyond the model itself, OlmoEarth initially included a studio for creating custom Earth intelligence models, a browser-based app for exploring model-generated maps, and a workflow engine for scheduling reproducible compute jobs.
But with the new version of OlmoEarth, how people use it has changed. “We’re excited to give users more ways to explore and engage with planetary artifacts in whatever way suits them best,” Ai2 wrote in a blog post.
It starts with the embeddings. The AI lab has already done the hard part by precomputing the world, saving developers and researchers the cost of running the AI model over the imagery. Now, for every 10-meter square of land on Earth, OlmoEarth v1.3 condenses a year of imagery from the Sentinel-1 radar satellite and the Sentinel-2 and Landsat optical satellites into a list of 128 numbers describing that spot. Put simply, the model studies a year of satellite photos of every house-sized patch of the planet and boils them down into a short description a computer can use to recognize what’s there and spot when it changes.
Ai2 also offers a lighter version that keeps only the first 64 numbers in each list. It uses half the storage space and downloads half the data. But while it sacrifices some accuracy, the organization reported that it still outperformed both rival products it tested against on most benchmarks.
In the spirit of openness, Ai2 is sharing the embeddings with the broader research community by placing them on Source Cooperative. “From there, 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,” Ai2 said.
The OlmoEarth Embeddings Explorer is also new. It’s an open analysis tool built on OlmoEarth’s embeddings that creates a “false-color” map of Earth. Users can move the map around in this browser tool and flip between years to create a manually powered pseudo-timelapse showing how landscapes have changed. No account or setup is required. The archive behind the Explorer tool holds about 150 TB of embeddings for each year of coverage. At launch, that includes 2025 for the entire planet and 2024 for the contiguous United States, and Ai2 plans to add more years going back to 2017.
Taking on Google’s AlphaEarth
Ai2 described OlmoEarth v1.3 as its “cheapest model” to train, requiring just over half (53 percent) of the computing time of v1.2 and less than a fifth (18 percent) of v1. It also produced half the carbon emissions of v1.2 and a fifth of v1’s—50 percent and 22 percent, respectively. The organization reported OlmoEarth v1.3 as being a “strong” fine-tuning model, with performance matching its predecessor across research and partner labels.
According to Ai2, reproducing the final training run for OlmoEarth v1.3 Base would cost roughly $2,720 using rented H100s and take less than three days. By comparison, v1.2 Base would cost $5,100 and take more than five days, and v1 Base would cost more than five times that of v1.3, and the run would take more than 15 days.
When evaluated against external rivals, Ai2 said OlmoEarth v1.3’s 128-dimensional embeddings outperformed both Google’s AlphaEarth Foundations and TESSERA v2, another open satellite embedding model. Ai2 reported that OlmoEarth v1.3 scored highest on 20 of 27 evaluations and ranked first overall. The lighter 64-dimensional version led on 18 of 27.
The most revealing result came from a benchmark Ai2 designed and published for this release. Called PLANTEUR, it measures how well a model can identify crops on tropical farms across France’s overseas territories, such as Martinique, Reunion, and French Guiana. Ai2 trained a classifier on some islands and then tested it on one it had never seen, mirroring the real-world challenge of mapping regions with little labeled data. OlmoEarth scored about 30 percent higher than TESSERA v2 on PLANTEUR and nearly 60 percent higher than AlphaEarth.
Ai2 believes that with the OlmoEarth v1.3 release, it provides researchers with an opportunity to “investigate geospatial foundation modeling” and “it becomes easier for anyone to make high-quality maps without running the model at all.”
OlmoEarth v1.3 is available now, from the base model and precomputed global embeddings to the Embeddings Explorer tool and the technical report. Ai2 said that everything is published under the OlmoEarth Artifact License, the same open license that v1 and v1.2 fall under. This means the technology can be widely used, but restrictions apply if used for military, defense, or “extractive industry” applications.
You can view the embeddings here. Ai2 has placed OlmoEarth v1.3’s training code on GitHub and both the pre-trained model and dataset on Hugging Face.
Discover more from Ken Yeung
Subscribe to get the latest posts sent to your email.
