OlmoEarth
modelYour notes
Ai2's open earth-observation foundation model line (entry filed at the v1.2 refresh; the line was previously untracked). Trained on Sentinel-1/2 and Landsat imagery in Nano/Tiny/Small/Base sizes (Base is a 114M ViT-Base; Small is new in v1.2). The v1.2 paper is an efficiency study: 3.0x fewer training GPU-hours and 2.9x fewer MACs at sustained benchmark performance. Released under the OlmoEarth Artifact License. On-thesis scientific FM for climate, agriculture, and disaster monitoring.
v1.3 (October 22, 2026) turns the line into an embedding model. A single Base checkpoint, which the card describes as a 128M-parameter ViT-Base plus a 14.3M-parameter Perceiver pooler and a 0.1M linear student, reads a year of Sentinel-1, Sentinel-2 and Landsat time series and emits a 128-dimensional vector for every 10 m pixel. The first 64 dimensions also work as a standalone embedding. Ai2 has precomputed annual embeddings, about 150 TB per year: global for 2023 to 2025 and select countries for 2017 to 2022. It ships them with an in-browser Embeddings Explorer and per-tile provenance tables. Against AlphaEarth Foundations and Tessera v2 on 27 evaluations, Ai2 reports that the 128-d embeddings score highest on 20 and the 64-d on 18. The evaluation introduces PLANTEUR, a new tropical crop-type benchmark. Training took 53% of v1.2's compute and 18% of v1's, with the final Base run costing about $2,720 of rented H100 time, and Ai2 says fine-tuning performance matches v1.2. Weights, pretraining data and code are released under the same OlmoEarth Artifact License. All the results are Ai2's own.
Model Details
Variants
| Name | Parameters | Notes |
|---|---|---|
| OlmoEarth-v1_2-Nano | — | — |
| OlmoEarth-v1_2-Tiny | — | — |
| OlmoEarth-v1_2-Small | — | — |
| OlmoEarth-v1_2-Base | 114M | — |
| OlmoEarth-v1_3-Base | 128M | 2026-10-22; ViT-Base encoder plus 14.3M Perceiver pooler and 0.1M linear student; 128-d (64-d Matryoshka) annual embeddings |