source: hugging face blog: olmoearth v1.1: a more efficient family of models
level: technical
olmoearth v1.1 is a family of transformer models for processing satellite imagery. it builds on the original olmoearth release from november 2025. the main goal is to reduce compute costs during inference and fine-tuning. the models achieve this by decreasing the token sequence length. compute costs in transformers scale quadratically with sequence length, so even small reductions help. the new models use up to three times fewer multiply-accumulate operations than the previous version.
the key change is how the model handles sentinel-2 satellite bands. earlier versions created separate tokens for each spatial resolution, leading to more tokens. olmoearth v1.1 merges these into a single token per patch. this cuts the token count by a factor of three. naive merging caused performance drops, so the team adjusted the pre-training method. the new approach maintains performance on benchmarks like m-eurosat knn. the models are trained on the same dataset as v1, isolating the effect of the token design.
the model family includes base, tiny, and nano sizes. users can pick a size that fits their compute budget. the reduced cost makes large-scale mapping more accessible. partners have used olmoearth for tasks like tracking mangrove change and classifying forest loss. the new version supports faster, cheaper deployment. some tasks may see regressions, so testing is recommended. the code and weights are available on hugging face and github.
why it matters: cheaper inference enables more frequent and larger-scale satellite image analysis for environmental monitoring and other applications.
source: hugging face blog: olmoearth v1.1: a more efficient family of models