Ai2's OlmoEarth v1.1 cuts compute 3x with a token redesign

Ai2 released OlmoEarth v1.1 on May 19, a refresh of its open foundation model family for satellite imagery. The headline claim is a 3x reduction in compute cost for both inference and fine-tuning, with no real loss on the research benchmarks or partner-built tasks that v1.0 was measured against.

The trick is in how the model tokenizes Sentinel-2 inputs. The original OlmoEarth produced one token per resolution per patch per timestep, which for Sentinel-2's three resolutions and two timesteps meant six tokens per image patch. v1.1 collapses all three resolutions into a single token per patch per timestep, cutting the token sequence by 3x. Because attention cost grows quadratically with sequence length, the same architecture now costs much less to run end to end, and the team's multiply-accumulate measurements match that math.

Ai2 had to retrain the model with a modified objective to keep accuracy from regressing more than ten percentage points on benchmarks, so the gains are not free, but the trade is favorable. All three sizes (Nano, Tiny, Base) are on Hugging Face under their open license, with the technical report and pretraining code published alongside.

Why it matters

Earth-observation teams running country-scale maps were often bottlenecked on compute when refreshing predictions; cutting per-image cost by 3x means more frequent updates on the same hardware budget, and makes routine fine-tuning on local datasets cheaper for groups outside well-funded labs.