OlmoEarth v1.1 Cuts Compute by 3x

AllenAI released OlmoEarth v1.1 on Hugging Face. It's a model family for remote sensing (satellite imagery) that keeps the same dataset as v1 while changing the token design and pretraining procedure to cut compute costs by up to 3x. Base, Tiny, and Nano models, along with weights, training code, and a tech report, are provided together.

The core is a design that reduces token sequence length. Sentinel-2 inputs multiply per-resolution patches by per-timestep tokens, causing the token count to grow rapidly, but merging resolution tokens into one reduces this to 3x fewer tokens. This change lowers costs across pretraining, fine-tuning, and inference all at once.

Simple merging caused a 10 percentage point drop on m-eurosat kNN. To address this, the pretraining procedure was modified, and performance on key benchmarks and partner tasks was maintained at v1 levels. However, regressions were also observed on some tasks.

• Use cases: mangrove change tracking, forest loss cause classification, country-scale crop mapping
• Research significance: enables isolating the effect of methodology changes on the same dataset