Smarter tokens just cut satellite AI costs by 3x without losing accuracy

Ai4earth has released OlmoEarth v1.1, a family of Earth observation transformer models that cut compute costs by up to three times while matching the benchmark performance of its predecessor. The efficiency gains come from a redesigned tokenization approach that collapses multiple Sentinel-2 spectral bands into a single token per patch, trimming the sequence length that drives transformer compute costs. For teams building large-scale geospatial pipelines, this kind of infrastructure fix can matter more than any single model benchmark, see how AI search pipelines are getting similar plumbing treatments.

The original OlmoEarth (v1) launched in November 2025. Partners have since used it for tasks ranging from tracking mangrove change to classifying forest loss drivers to producing country-scale crop-type maps. But for large-area deployments, spanning tens to hundreds of thousands of square kilometers, compute is the highest cost in the full lifecycle of data export, preprocessing, inference, and post-processing. A smaller memory footprint means more partners can run on the OlmoEarth platform, and teams running their own infrastructure save significantly. As open models find their way into regulated industries, the ability to self-host at lower cost becomes a strategic advantage, as AI2's partners in finance and healthcare have already shown.

“Over the full lifecycle of running OlmoEarth, data export, preprocessing, inference, and post-processing, compute is by far the highest cost,” the team wrote in today's announcement. “A more efficient model means we can support more partners on the OlmoEarth Platform, and that anyone running OlmoEarth on their own can leverage this technology faster and at lower expense.”

How sequence length drives cost

Transformer models scale quadratically in compute with token sequence length, so even small reductions can cut cost. The key design question for remote sensing transformers: what makes an efficient token? The answer isn't obvious, recent work on dataset distillation for high-resolution imagery shows that naive compression often hurts more than it helps.

With Sentinel-2 imagery, the model ingests a tensor of height, width, temporal depth (T), and 12 spectral channels. The v1 approach split each patch into distinct tokens per timestep per resolution, 6 tokens per patch for a 2-timestep input (2 timesteps × 3 resolutions: 10m, 20m, and 60m). That yields (H/p × W/p × T × 3) tokens total for a single input.

This per-resolution token is standard practice: Galileo and SatMAE both use it, and SatMAE showed significantly better results than a merged-token baseline. But it inflates token counts multiplicatively. Naively merging all bands into a single token per timestep reduced token count by three times but caused a 10 percentage point drop on the m-eurosat kNN benchmark, a widely used remote sensing evaluation. The challenge of preserving semantic detail while merging visual representations is a recurring theme, see how VIQ tackles the same tension in visual quantization.

“We hypothesize that separating Sentinel-2 bands into different tokens makes it easier for OlmoEarth to model important cross-band relationships,” the team explained. To merge bands without sacrificing accuracy, they modified the pre-training regimen, details are in the accompanying technical report.

What changed in v1.1

The result is a model family that does more with less. At every size, Base, Tiny, and Nano, OlmoEarth v1.1 runs at roughly one-third the compute cost of the original, while matching performance on a mix of research benchmarks and partner-constructed tasks.

“It provides similar performance to OlmoEarth v1 while requiring one third of the compute, though we have seen some regressions,” the team noted, pointing users to the technical report for a full breakdown. For most use cases, the team expects a significant speedup during both fine-tuning and inference.

For researchers, the v1.1 release offers a cleaner experimental isolation: since both model families are trained on the same dataset, any performance shifts isolate the effect of methodological changes. That's a rare controlled variable in the often messy landscape of remote sensing pretraining. Having the right evaluation tools to measure such shifts is critical, AI2's own Olmo-Eval workbench provides a framework for exactly this kind of apples-to-apples comparison.

OlmoEarth v1.1 weights and training code are available now under the same open-source license as v1, including Base, Tiny, and Nano model variants. For teams that want full transparency into how such models are built, AI2's open release of every checkpoint and decision for Olmo 3 sets the standard for reproducibility in this space.