OlmoEarth v1.1: A more efficient family of models

Allen Institute (AI2) · 1 month ago

OlmoEarth released v1.2 with rotary positional embeddings (RoPE) to reduce artifacts in satellite image processing, following a v1.1 release that reduced compute costs by up to 3x by consolidating multi-resolution tokens into single tokens per patch. The key improvement in v1.2 comes from replacing absolute positional embeddings with RoPE, which encodes patch location without introducing embedding artifacts. These efficiency gains enable organizations to process larger geographic areas and refresh planet-scale maps more frequently at lower cost.