OlmoEarth v1.1 processes satellite maps three times faster
Original: OlmoEarth v1.1: A more efficient family of models
Source: huggingface.co ↗
Who: Posted on the Hugging Face blog by Kyle Wiggers on behalf of Allen AI (Ai2), a nonprofit AI research lab; the underlying model family was built by the Ai2 OlmoEarth team.
What's new: Ai2 has released OlmoEarth v1.1, an updated family of satellite-imagery AI models that cuts compute costs by up to 3x compared to the November 2025 original, while preserving roughly the same accuracy across a range of remote-sensing tasks — mapping crops, tracking mangrove loss, classifying forest-change drivers — at national to global scales.
How it works: OlmoEarth is built on the . Satellite images from are sliced into spatial patches, and each patch was previously converted into multiple — one per timestep per resolution band, yielding six tokens per patch for a two-timestep Sentinel-2 input. The key change in v1.1 is collapsing those three per-resolution tokens into a single token per patch per timestep, reducing token count by a factor of three. Because compute in transformers scales quadratically with , this three-fold reduction in tokens translates directly into the claimed 3x cost reduction. Naively merging the tokens caused a roughly 10 percentage-point accuracy drop on , so the team modified the process — the specifics of which are detailed in the technical report — to recover that lost accuracy.
The numbers: The v1.1 family ships in Base, Tiny, and Nano sizes, giving teams flexibility to match model size to their compute budget. At every size, inference and fine-tuning run up to 3x cheaper than v1. The team reports that performance is roughly preserved across their benchmark mix, with some regressions on specific tasks documented in the technical report. Weights and training code are publicly available.
Why it matters: Because the team held the training dataset constant between v1 and v1.1, any performance difference between the two versions isolates the effect of the tokenization and pre-training changes alone — a cleaner scientific comparison than most model updates offer. Practically, a 3x cost reduction means organizations with limited budgets can refresh planetary-scale maps far more frequently, and researchers can iterate faster without needing large cloud compute allocations.