OlmoEarth v1.1 Cuts Compute Costs by Up to 3× While Maintaining Performance

Allen Institute for AI has released OlmoEarth v1.1, a new family of transformer‑based Earth‑observation models that reduce compute costs by up to three times compared to the original v1 while preserving performance on research benchmarks and partner tasks. The key efficiency gains come from collapsing Sentinel‑2 resolution bands into a single token per timestep, which reduces token sequence length and thus the quadratic MAC cost of transformer attention. OlmoEarth v1.1 was trained on the same dataset as v1, isolating methodological changes and allowing researchers to study the impact of token design on performance. The family includes Base, Tiny, and Nano variants, all released under the Apache 2.0 license, and supports up to 8192‑token context thanks to ModernBERT‑style pre‑training. The new models maintain similar accuracy to v1 on tasks such as mangrove change detection, forest‑loss driver classification, and country‑scale crop‑type mapping, while offering significant speedups during fine‑tuning and inference. Developers can switch to v1.1 to achieve faster, cheaper satellite‑image processing, enabling more frequent, planet‑scale map refreshes for partners and internal projects. The release also includes training code and weights for all variants, allowing teams to benchmark inference speed and adjust tokenization to use a single token per resolution. By providing a more efficient backbone, OlmoEarth v1.1 expands the practical deployment horizon for remote‑sensing applications across national, continental, and global scales.