OlmoEarth v1.3
Ai2 released OlmoEarth v1.3, an open satellite embedding model, with precomputed annual 10m embeddings for the whole planet and a browser-based Embeddings Explorer.
The model reads a year of Sentinel-1, Sentinel-2 and Landsat time series and outputs one 128-dimensional vector per 10-meter pixel, and the first 64 dimensions work as a smaller embedding. Earlier OlmoEarth use mostly meant fine-tuning in OlmoEarth Studio, while embeddings can now be reused with a small classifier. Training took 53% of the compute of v1.2 and 18% of v1.
- Date
- Thursday 22 October 2026
- Lab
- Ai2
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| Evals where 128-dim embeddings score highest (vs AlphaEarth Foundations and Tessera v2) | 20 of 27 64-dim version leads on 18 of 27 | company |
| Training compute relative to v1.2 | 53% 18% of v1; carbon emissions 50% and 22% | company |
| Estimated cost to pre-train v1.3 Base final run | $2,720 rented H100 time | company |
Released under the OlmoEarth Artifact License; 2023 to 2025 embeddings are global and 2017 to 2022 cover select countries.
Sources
This record was partly confirmed: some claims could not be checked on 22 October 2026. How we check