OlmoEarth v1.1: A more efficient family of Earth observation models

Hugging Face and Allen AI have introduced OlmoEarth v1.1, an updated family of transformer-based Earth observation models designed to reduce the high compute costs associated with processing satellite imagery. Following the release of OlmoEarth v1 in November 2025, partners applied the technology to various tasks such as tracking mangrove change, classifying drivers of forest loss, and producing country-scale crop-type maps. Because compute represents the highest cost over the model lifecycle across data export, preprocessing, inference, and post-processing, OlmoEarth v1.1 was built to lower these expenses. The new model family cuts compute costs by up to three times while maintaining the performance of OlmoEarth v1 on research benchmarks and partner tasks. To achieve this efficiency without suffering performance drops like a 10 percentage point decrease on m-eurosat kNN, the developers modified their pre-training regimen to collapse multiple resolutions into a single token for Sentinel-2 inputs, reducing token counts multiplicatively. Trained on the same dataset as the original version to isolate methodological changes, OlmoEarth v1.1 is available in Base, Tiny, and Nano sizes, offering speedups during fine-tuning and inference for teams running the technology.