OlmoEarth v1.1: A more efficient family of models

A new family of models, OlmoEarth v1.1, has been released for remote sensing applications, cutting compute costs by up to 3x while preserving the performance of the original version. The key innovation lies in a more efficient tokenization strategy for processing satellite imagery, which reduces the token sequence length. By modifying the pre-training process, the model can merge data from different resolutions into a single token without the performance degradation that would normally occur. This makes planet-scale analysis more affordable and advances the understanding of pretraining for remote sensing models.