Dispatch
AI2 ships OlmoEarth embeddings — 1.4M to 89M param encoders, F1=0.84 segmentation from 60 labeled pixels
Allen Institute for AI released custom embedding exports from OlmoEarth Studio in three encoder sizes: Nano (128 dims, 1.4M params), Tiny (192 dims, 6.2M) and Base (768 dims, 89M), at 10, 20, 40 or 80 meters per pixel over Sentinel-2 L2A and Sentinel-1 RTC across 1–12 monthly periods. Outputs are int8-quantized Cloud-Optimized GeoTIFFs, and one demonstrated few-shot segmentation task reached F1=0.84 from just 60 labeled pixels. Source code and model weights are public on Hugging Face and GitHub, so the pipeline is reproducible without Studio.
↳ Follow the thread