Ai2's OlmoEarth v1.3 final training run cost about $2,720, down from $15,060 for v1
Ai2's new satellite embedding model trains on 18% of v1's compute and led 20 of 27 benchmark tasks against Google's AlphaEarth and Tessera v2.
OddBrief EditorialAI-assisted, human-reviewed
AIKey facts
- Who
- Allen Institute for AI (Ai2), released October 22
- Cost
- final run about $2,720 of H100 time vs $15,060 for v1
- Output
- a 128-dimension vector per 10 m pixel, about 150 TB per year
- Benchmarks
- best on 20 of 27 tasks in Ai2's own evaluation
The Allen Institute for AI (Ai2) released OlmoEarth v1.3 on October 22, a model that turns a year of satellite imagery into one 128-number vector for every 10-meter patch of land on Earth. Ai2 estimates its final training run cost about $2,720 in rented H100 time, compared with about $15,060 for v1 and $5,100 for v1.2.
Ai2 says fine-tuning results match v1.2, so the headline change is efficiency. The company also published ready-made global embeddings that anyone can download.
Same accuracy, a fraction of the compute
According to Ai2, v1.3 used about 18% of the compute of v1 and 53% of v1.2, with 22% and 50% of their carbon emissions. The $2,720 figure covers reproducing the final run, not the full cost of developing the model.
The model reads Sentinel-1 radar plus Sentinel-2 and Landsat multispectral time series. Its output vectors have 128 dimensions, and the first 64 can be used alone as a smaller version.
In Ai2's own evaluation across 27 tasks on six continents, the 128-dimension embeddings scored best on 20 of them against Google DeepMind's AlphaEarth Foundations and Tessera v2. The 64-dimension cut led on 18. Ai2 also introduced a tropical crop-type benchmark called PLANTEUR, where it says its embeddings generalized best to places they had not seen.
A planet of precomputed vectors
The precomputed archive runs to roughly 150 terabytes per year. Coverage is global for 2023, 2024 and 2025, while 2017 to 2022 covers select countries so far. For every 20 km tile and year, Ai2 lists exactly which satellite scenes went into it.
A browser tool, the Embeddings Explorer, lets users find similar places or train a small classifier from a few labeled points, with no sign-in. Ai2 says this kind of analysis can take seconds where fine-tuning takes hours. The model, data and code are released under the OlmoEarth Artifact License.
The fine print
The benchmark wins come from Ai2's own technical report, not an independent test. The Explorer is pitched as a first pass, and Ai2 suggests fine-tuning in OlmoEarth Studio when accuracy matters. How far the historical archive will extend beyond the select countries for 2017 to 2022 has not been announced.
Sources
- OlmoEarth v1.3: State-of-the-art embeddings for the whole planetAi2 on Hugging Faceprimary source