Introducing OlmoEarth embeddings: Custom embedding exports from OlmoEarth Studio for downstream analysis

Source: HuggingFace Date: 2026-08-12 URL: https://huggingface.co/blog/allenai/olmoearth-embeddings

Summary

Allen AI’s OlmoEarth Studio now exports custom embeddings from its open OlmoEarth foundation models — compact vector representations of Sentinel-1/2 satellite imagery at three sizes (Nano 128-dim/1.4M params up to Base 768-dim/89M params), delivered as Cloud-Optimized GeoTIFFs for similarity search, land-cover mapping from few labels, and change detection.

Implications

Feeds local-model-feasibility: an open-weight embedding model small enough (1.4M–89M params) to run well outside datacenter budgets, in a domain (geospatial) distinct from the usual chat/code models tracked here. A useful reminder that “open-weight” releases now span far beyond LLMs, with the same weights-available/reproducible-pipeline ethos.