OlmoEarth v1.1: A more efficient family of Earth observation models
Hugging Face ● Covered by 2 sources
Allen AI dropped OlmoEarth v1.1, a satellite-image AI that runs up to 3x cheaper than the original. Same performance, way less compute, so more groups can run planet-scale maps without burning cash.
Based on reporting by Hugging Face — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Allen AI's OlmoEarth family made a name for itself tracking mangroves, spotting deforestation drivers, and churning out country-scale crop maps in days rather than months. But scaling that kind of analysis to hundreds of thousands of square kilometers runs into a blunt problem: compute is expensive, and it's the biggest line item in the whole pipeline, from data export through post-processing. OlmoEarth v1.1 is Allen AI's answer, a retooled model family that cuts compute costs by up to three times while holding onto roughly the same accuracy on the benchmarks and partner tasks the team uses to judge these things.
The fix comes down to how the model turns satellite pixels into tokens. OlmoEarth processes Sentinel-2 imagery, which arrives as a tensor with height, width, time, and 12 spectral channels split across three resolutions (10m, 20m, 60m). The original approach assigned a separate token for every timestep and every resolution, so a two-timestep image produced six tokens per patch. That's a common trick — Galileo and SatMAE do the same, and SatMAE's own results show it helps. But since transformer compute scales quadratically with sequence length, those extra tokens add up fast across pretraining, fine-tuning, and inference.
Allen AI's team tried collapsing the three resolution tokens into one, following an approach closer to what CROMA does. Done naively, that move tanked performance — a 10 percentage point drop on the m-eurosat kNN benchmark, which is not a rounding error. Their theory is that keeping bands in separate tokens lets the model learn relationships across spectral channels more easily, and merging them blindly throws that signal away. So instead of abandoning the idea, they reworked the pretraining regimen to make single-token resolution merging viable without the accuracy hit, a change they detail in the accompanying technical report.
The payoff is a model family — Base, Tiny, and Nano — that runs at roughly a third of the original's compute cost at every size, with some regressions the team is upfront about rather than glossing over. Because v1.1 trained on the exact same dataset as v1, any performance differences trace back cleanly to the architecture and training changes rather than data drift, which is a rarer courtesy in remote-sensing model research than it should be. Weights and training code are already up on Hugging Face and GitHub for anyone running OlmoEarth to swap in.
My take — AI-written commentary, not fact-checked reporting
Cutting compute 3x while isolating the change to a single, clean variable (same dataset, different token design) is the kind of unglamorous engineering that actually moves a field forward, and it's refreshing that Allen AI admits to regressions instead of burying them in an appendix. The real story here isn't the benchmark numbers, it's that cheaper inference means smaller NGOs and local governments can actually run planet-scale monitoring themselves instead of renting it from a cloud giant, which is exactly the kind of open-model win the EU's sovereignty crowd should be paying attention to.
Read more about this at: Hugging Face