The Allen Institute for AI released OlmoEarth v1.

1, a new family of Earth observation models that reduces compute costs by up to 3x while matching v1’s performance on research benchmarks and partner tasks.

The key efficiency gain comes from redesigning the token representation for Sentinel-2 imagery.

In v1, satellite data was split into spatial patches, with each patch producing a separate token per timestep per resolution (10m, 20m, 60m), yielding 6 tokens per patch for 2 timesteps. v1.

1 collapses all three resolutions into a single token per patch per timestep, cutting the token count by a factor of 3.

Since compute costs in transformers scale quadratically with sequence length, this reduction translates directly into cheaper and faster inference.

Naively merging tokens caused a 10 percentage point drop on m-eurosat kNN, a standard remote sensing benchmark.

To recover performance, the team modified the pre-training regimen—details are in the accompanying technical report. The model family includes Base, Tiny, and Nano sizes. For developers, v1.

1 provides a drop-in replacement with a significant speedup during fine-tuning and inference, though the team notes some regressions documented in the report. For researchers, OlmoEarth v1.

1 uses the same training dataset as v1, isolating the effect of methodological changes and making it a cleaner object of study.

The models, tech report, and training code are released on Hugging Face and GitHub.