OlmoEarth v1.1 cuts compute costs by 3x for geospatial analysis
olmoearth v1.1 reduces compute costs by up to 3x while maintaining benchmark performance.
allen institute for ai released olmoearth v1.1, a family of efficient vision transformer models designed for processing satellite imagery and geospatial data. the new models achieve up to a three-fold reduction in compute costs by collapsing sentinel-2 spectral bands into single tokens, significantly reducing the sequence length required for transformer processing. this architectural change necessitated adjustments to the pre-training regimen to preserve accuracy on benchmarks like m-eurosat knn without losing performance. the efficiency gains enable users to scale deployments to continental or global levels with lower financial and computational overhead.