Allen AI releases OlmoEarth v1.1, a new family of transformer-based remote sensing models that cuts compute costs by up to 3x compared to v1 while maintaining similar performance. The key innovation is collapsing multi-resolution Sentinel-2 tokens into a single token per patch, reducing sequence length threefold. Naive token merging caused significant performance drops, so the team modified their pre-training regimen to compensate. The models are trained on the same dataset as v1, allowing researchers to isolate the effect of architectural and methodological changes. Base, Tiny, and Nano model weights and training code are publicly available.

5m read timeFrom huggingface.co
Post cover image
Table of contents
Increasing efficiency by decreasing sequence lengthsDesigning the tokenFor developersFor researchersGet started

Sort: