Ai2 cuts satellite imagery AI costs by 3x with a smarter token trick

The Allen Institute for AI (Ai2) is out with OlmoEarth v1.1, a revamped family of remote sensing models that slashes compute costs by up to three times while matching the performance of the original OlmoEarth v1. The update, laid out in a technical report, homes in on how satellite imagery tokens are generated, tackling the quadratic scaling headache that plagues transformer-based models.

Efficiency gains through token redesign

When you're processing satellite imagery across tens to hundreds of thousands of square kilometers, compute is by far the biggest line item in the entire lifecycle, from data export and preprocessing through inference and post-processing. Because the OlmoEarth models are transformer-based, compute costs grow quadratically with token sequence length. That means even small reductions in token count can make a big dent in expenses.

For Sentinel-2 imagery, a common input, a single image is represented as a tensor with height (H), width (W), temporal dimension (T), and 12 spectral channels. The original OlmoEarth v1 split this data into spatial patches of size p x p, then created a token per timestep per resolution (10m, 20m, and 60m). This meant a Sentinel-2 input with 2 timesteps yielded 6 tokens per patch (2 timesteps × 3 resolutions).

OlmoEarth v1.1 collapses these three resolution-specific tokens into a single token per patch per timestep, reducing token counts by a factor of three. “Because token counts compound multiplicatively, collapsing resolutions into a single token produces three times fewer tokens and material savings across pretraining, fine-tuning, and inference,” the Ai2 team wrote.

That wasn't a trivial fix, though. Naively combining tokens led to significant performance drops, including a 10 percentage point plunge on the m-eurosat kNN benchmark. The researchers suspect that separating Sentinel-2 bands into different tokens made it easier for the model to learn important cross-band relationships. To get around that, they tweaked the pre-training regimen, changes detailed in the accompanying paper.

Developer and researcher impact

For developers, OlmoEarth v1.1 runs up to three times cheaper than v1, making frequent, planet-scale map refreshes far more affordable. The new model family comes in Base, Tiny, and Nano sizes. Ai2 notes that while OlmoEarth v1.1 delivers similar performance to v1 at one third the compute, some regressions have been spotted, the technical report has the full breakdown.

For researchers, the update offers a clean before-and-after comparison. “We train OlmoEarth v1.1 on the same dataset as OlmoEarth v1, so any differences between the two isolate the effect of methodological changes,” the team wrote. That gives researchers a clear window into how tokenization and pre-training changes affect remote sensing performance.

Background and mission

Ai2 released OlmoEarth v1 in November 2025. Since then, partners have used the model to track mangrove change, classify drivers of forest loss, and produce country-scale crop-type maps in days. Each release brings the institute closer to its mission: putting state-of-the-art AI in the hands of organizations working to protect the planet.

The OlmoEarth v1.1 weights and training code are now available on the project website.