OlmoEarth v1.1 Cuts Compute Costs by 3x for Satellite AI

The Allen Institute for AI released OlmoEarth v1.1, a new family of transformer-based models for remote sensing that reduces compute costs by up to three times while maintaining the performance of the original OlmoEarth v1. The efficiency gain comes from redesigning how satellite imagery is tokenized: instead of creating separate tokens for each resolution (10m, 20m, 60m) per timestep, the new model merges all resolutions into a single token per patch per timestep. This cuts token counts by a factor of three, addressing the quadratic scaling of compute with sequence length. The original OlmoEarth v1 was released in November 2025 and has been used for tasks like tracking mangrove change, classifying drivers of forest loss, and producing country-scale crop-type maps. OlmoEarth v1.1 is trained on the same dataset as v1, isolating the effect of methodological changes. The model family includes Base, Tiny, and Nano sizes, with weights and training code available. Some performance regressions are noted in the technical report.