OlmoEarth v1.3: State-of-the-art embeddings for the whole planet

Hugging Face has pushed out OlmoEarth v1.3, a satellite embedding model it says is state of the art. The release is bigger than a model drop. It also includes precomputed annual embeddings for the whole planet, plus a browser tool called the OlmoEarth Embeddings Explorer that lets people inspect and customize maps without signing in or wrangling GPU hardware.

The pitch here is simple: compute once, reuse everywhere. OlmoEarth v1.3 digests a year of Sentinel-1 radar, Sentinel-2, and Landsat data and turns each 10-meter pixel into a 128-number vector. A smaller 64-dimensional version is available too. Hugging Face has also put the global embeddings on Source Cooperative, so users can build lighter tasks on top of them instead of training a fresh model from scratch every time.

On its own evaluation, the new model comes out strong. Across 27 tests spanning six continents, the 128-dimensional embeddings ranked highest on 20 and the 64-dimensional version led on 18. The company also says it introduced a new crop benchmark, PLANTEUR, for tropical farming, and that the embeddings generalize best when a classifier faces places it has never seen before. The whole stack is open: weights, pretraining data, and pretraining code.

There’s also a clear efficiency story. Compared with previous OlmoEarth versions, v1.3 took 53% of the compute time of v1.2 and 18% of v1, with 50% and 22% of the carbon emissions. Hugging Face says reproducing the final pretraining run would cost roughly $2,720 in rented H100 time. And even though the model is built to stay lightweight, it still matches v1.2 on research and partner labels.

The Explorer is where the release becomes something people can actually poke at. You can flip between years, compare change, inspect the underlying Sentinel imagery, click for similar places, or label a few points and export a GeoTIFF. The archive behind it is huge: about 150 TB per year at 10-meter resolution, with global coverage for 2025, 2024, and 2023, plus select-country coverage from 2017 to 2022. Each tile also comes with provenance tables tying pixels back to the scenes that produced them.