OlmoEarth v1.1: 3x More Efficient Earth Observation Models

OlmoEarth v1.

1 is a new family of Earth observation models from the Allen Institute for AI that cuts compute costs by up to 3x while maintaining the performance of the original OlmoEarth v1.

The key technical innovation lies in redesigning how satellite imagery tokens are constructed: instead of creating separate tokens for each spatial resolution band (10m, 20m, 60m) at every timestep, the new approach collapses resolutions into a single token per patch.

This reduces token sequence length multiplicatively, yielding significant compute savings because transformer costs scale quadratically with sequence length.

However, naive merging caused a 10 percentage point drop on m-eurosat kNN benchmarks, so the team modified the pre-training regimen to preserve cross-band relationships.