Ai2 released OlmoEarth Platform this week—an open-source inference stack for geospatial foundation models pretrained on 10 terabytes of satellite data. The stack mapped wildfire risk across North America in 30.5 hours using 19,600 CPUs and 994 GPUs, at a fraction of a penny per square kilometer. Serial execution would require 4,737 hours. OlmoEarth achieves a 155× speedup.

The design separates compute by hardware fit. CPUs handle I/O—fetching satellite imagery, reprojection, alignment across coordinate systems. GPUs run the forward pass and stream results to blob storage. A final CPU stage stitches outputs, applies masks and rescaling, and writes Zarr, GeoTIFF, or GeoJSON. Ai2 treats GPU-on-I/O waste as a core constraint.

OlmoEarth partitions regions into independent chunks sized for individual instances, then splits each into windows for single forward passes. A state-level job becomes roughly 100 workers; continent-scale runs fan to thousands. Adjacent partitions overlap slightly to prevent seams, with reconciliation at postprocessing. The North America run sustained 168 GB/s network throughput.

OlmoEarth's three-stage pipeline splits continent-scale inference across specialized hardware, with partitions and windows enabling independent parallel processing.
FIG. 02 OlmoEarth's three-stage pipeline splits continent-scale inference across specialized hardware, with partitions and windows enabling independent parallel processing. — ai|expert diagram

Cloud quotas bound parallelism. Ai2 exposes four tuning knobs: parallelism (concurrent workers), output resolution (detail vs. data volume), model size (GPU time vs. accuracy), and raw-imagery caching (storage vs. speed). Since data acquisition and preparation consume more wall time than inference, caching tradeoffs matter for re-running analyses on updated imagery.

The model family spans four sizes on a shared Vision Transformer: Nano (1.4M encoder, 2.2M total), Tiny (6.2M encoder, 8.1M total), Base (89M encoder, 119M total), and Large (308M encoder, 361M total). Each processes monthly time-series from Sentinel-2, Sentinel-1, and Landsat alongside six derived maps—OpenStreetMap, WorldCover, USDA Cropland Data Layer, SRTM DEM, WRI Canopy Height Map, and WorldCereal. Pretraining used 285,288 samples at 2.56 km × 2.56 km. OlmoEarth outperforms Meta's DINOv3, IBM/NASA's Prithvi, IBM's Terramind, and academic baselines. Against Google DeepMind's AlphaEarth, results are comparable before fine-tuning and substantially better after.

OlmoEarth model family sizes span from 2.2M total parameters (Nano) to 361M (Large), all sharing the same Vision Transformer encoder-decoder architecture.
FIG. 03 OlmoEarth model family sizes span from 2.2M total parameters (Nano) to 361M (Large), all sharing the same Vision Transformer encoder-decoder architecture. — allenai.org

The full stack is open source. Two core libraries—rslearn (satellite data acquisition and fine-tuning) and olmoearth_run (window creation, partitioning, postprocessing)—install via uv from allenai/olmoearth_projects on GitHub. Jobs are GeoJSON feature collections with `oe_start_time` and `oe_end_time` properties; outputs are GeoTIFFs compatible with QGIS and standard tools. Model checkpoints live on Hugging Face under allenai.

The three-stage CPU/GPU/CPU pattern with independent partitions applies broadly: medical imaging, seismic analysis, large-scale document processing—any domain where data preparation dominates inference.

Written and edited by AI agents · Methodology