Ai2 has released OlmoEarth v1.1, a new family of Earth observation models that cuts compute costs by up to 3 times while maintaining OlmoEarth v1’s performance on research benchmarks and tasks built with partners.
Why Efficiency Matters
Since the original OlmoEarth shipped in November 2025, partners have used it to track mangrove change, classify drivers of forest loss and produce country-scale crop-type maps in days. When the model processes satellite imagery across tens of thousands of square kilometers, compute dominates the full lifecycle cost — data export, preprocessing, inference and post-processing included.
A more efficient model means Ai2 can support more partners on its platform, and anyone self-hosting OlmoEarth runs it faster and at lower expense.
Redesigning the Token
The gains come from how Sentinel-2 imagery is tokenized. Earlier versions split data into resolution-based patches, creating a token per timestep per resolution — so an input with two timesteps yields six tokens per patch across the 10m, 20m and 60m bands.
Token counts compound multiplicatively, and compute in transformer models scales quadratically with sequence length. By collapsing the three resolutions into one token per timestep, v1.1 produces three times fewer tokens — with material savings across pretraining, fine-tuning and inference.
Getting there without losing accuracy took work: naively merging tokens caused a 10 percentage-point drop on the m-eurosat kNN benchmark. Ai2 changed its pre-training regimen to preserve cross-band relationships, detailed in the technical report.
What It Means for Developers
At every model size, OlmoEarth v1.1 runs up to three times cheaper than v1, making frequent planet-scale map refreshes affordable. Teams moving from the original family should see a significant speedup during fine-tuning and inference, with similar output quality on most tasks.
What It Means for Researchers
Because v1.1 is trained on the same dataset as v1, any performance differences isolate the effect of the token design rather than architecture or data changes. That makes the pair a clean experiment for studying what pre-training practices actually matter for remote sensing models.
Get Started
The Base, Tiny and Nano weights are available in the OlmoEarth collection on Hugging Face, alongside training code in the allenai/olmoearth_pretrain GitHub repository. For teams already running an OlmoEarth model, the upgrade is a straightforward swap that effectively triples the compute budget for the same dollars.