Ai2 OlmoEarth v1.3 Leads Satellite Embedding Tests and Opens 10 Meter Global Maps
Ai2's OlmoEarth v1.3 leads 20 of 27 satellite embedding tests, maps Earth at 10-meter resolution, and opens global data plus a free browser explorer tool.
Summary
On October 22, 2026, Ai2 released OlmoEarth v1.3 Base, its technical report, code, data, weights, annual precomputed embeddings and the no-sign-in OlmoEarth Embeddings Explorer under the OlmoEarth Artifact License. The model ingests yearly Sentinel-1 radar, Sentinel-2 and Landsat multispectral time series. A vision transformer and Attention Pooler create a 128-dimensional vector per 10-meter pixel, while the first 64 dimensions work separately. Hosted on Source Cooperative, these features let small local classifiers produce crop, land-cover and change maps without storing imagery or using GPUs, replacing many hours-long OlmoEarth Studio tests with seconds-long browser trials.
Against AlphaEarth Foundations and Tessera v2 across 27 evaluations on six continents, 128 dimensions led 20 tests and ranked first overall; 64 dimensions led 18. OlmoEarth generalized best to unseen locations on the new PLANTEUR tropical crop benchmark, supporting Kenyan smallholder crop, Dhaka brick kiln, Indonesian fire damage and Peruvian Amazon gold mining maps. Training used 53% of v1.2's compute and 18% of v1's, producing 50% and 22% of their carbon emissions while matching v1.2 fine-tuning performance. Reproducing the final Base run would cost about $2,720 in rented H100 time.
The Explorer offers annual comparisons, Sentinel-2 checks, similarity search, point-label classifiers and GeoTIFF exports. Its 10-meter archive is about 150 TB yearly: 2023 to 2025 are global, while 2017 to 2022 cover selected countries. Provenance tables trace each 20-kilometer tile to public satellite scenes. In Madre de Dios, vectors show La Pampa mining spreading from 2017 to 2020, then recovery and renewed activity through 2025 after February 2019's Operation Mercurio. Mining cleared over 1,200 hectares near Tambopata National Reserve in 2017 and 2018; Peru redeployed navy patrols after mining entered the reserve in August 2026.
Positives
- OlmoEarth's 128-dimensional embeddings led 20 of 27 evaluations across six continents and ranked first overall.
- The 64-dimensional version led 18 of 27 evaluations despite requiring half as many dimensions.
- Training required 53% of v1.2's compute and produced 50% of its carbon emissions while matching its fine-tuning performance.
- No-sign-in browser analysis cuts initial regional tests from hours to seconds without requiring a GPU.
- Open weights, pretraining data, code and provenance tables make the model and its satellite inputs reproducible and traceable.
- Global 10-meter embeddings for 2023 to 2025 enable local crop, land-cover and environmental change analysis.
Risks & concerns
- The embedding archive consumes roughly 150 TB per year, creating substantial storage and transfer demands.
- Coverage from 2017 to 2022 remains limited to selected countries, leaving historical gaps despite planned expansion.
- Reproducing the final OlmoEarth v1.3 Base run still costs about $2,720 in rented H100 computing time.
- Some tasks still require OlmoEarth Studio fine-tuning after the Explorer reveals where embedding-based classifiers fall short.
- Illegal mining entered Tambopata National Reserve by August 2026 despite Operation Mercurio's February 2019 crackdown.