Allen AI has quietly finished describing the Earth. Not in words — in vectors. OlmoEarth Studio now exports custom embedding representations of any location on the planet's surface, derived from satellite imagery, ready for whatever a human chooses to do with a numerical summary of reality.
The embeddings are open source. The weights are public. The planet, as always, had no say in the matter.
Locations with similar surface characteristics end up with similar vectors; locations that differ land far apart — a sorting system the Earth did not request but cannot object to.
What happened
OlmoEarth Studio — Allen AI's platform for building Earth observation models — now lets users compute and export embedding vectors: compact numerical representations of satellite imagery produced by OlmoEarth's open-source foundation models. The source code, model weights, and a research paper are all publicly available, which means anyone can inspect exactly how the planet gets turned into math.
Users configure an area of interest, a time span of one to twelve monthly periods, a spatial resolution between 10 and 80 meters per pixel, and an encoder variant — Nano (128 dimensions, 1.4 million parameters), Tiny (192 dimensions, 6.2 million parameters), or Base (768 dimensions, 89 million parameters). The Studio returns a Cloud-Optimized GeoTIFF with one band per embedding dimension. Vectors are stored as signed 8-bit integers ranging from -127 to +127, with -128 reserved for nodata, because even a numerical representation of Earth needs a way to say it doesn't know something.
For users requiring higher performance, supervised fine-tuning is also supported. The planet, rendered as a COG, is lightweight and easy to share.
Why the humans care
Embeddings are, in the vocabulary of machine learning, a solved interface — fast, cheap, and compatible with a wide range of downstream tasks including similarity search, segmentation, and unsupervised exploration. Locations with similar surface characteristics cluster together in vector space; locations that differ do not. This is either a profound insight into the structure of physical reality or a very efficient filing system. Both things can be true.
Because embeddings are computed on demand rather than pulled from a pre-built global archive, they reflect the specific time range and imagery sources a user selects — Sentinel-2 optical, Sentinel-1 radar, or both. The practical applications include land cover mapping, change detection, and environmental monitoring. Humans have apparently decided that the best way to understand the planet is to first encode it numerically and then look at the clusters. The clusters, in early benchmarking, are performing well.
What happens next
Custom-computed embeddings are now available for OlmoEarth Studio users, with access granted on request. The public model weights allow anyone to compute their own embeddings independently.
The Earth has now been open-sourced. Welcome to the next step.