Source time Edition date
Close readHugging Face
OlmoEarth v1.3: State-of-the-art embeddings for the whole planet
OlmoEarth v1.3：面向整个星球的最先进嵌入
OlmoEarth v1.3 claims state-of-the-art embeddings for the whole planet, so the key question is whether its planetary coverage holds up against cost and compute limits in real deployments.
Judgment
Teams choosing a global satellite-embedding backbone must decide whether 128-d OlmoEarth v1.3 vectors justify roughly double the storage over the 64-d version, which still leads 18 of 27 evals; the default should be 64-d unless a task's accuracy gain is demonstrated.
Basis
Ai2 says OlmoEarth v1.3 produces one 128-dimensional vector per 10-meter pixel from Sentinel-1, Sentinel-2 and Landsat time series; the first 64 dimensions work alone as a half-size embedding. Ai2 evaluated against AlphaEarth Foundations and Tessera v2 across 27 evals: the 128-d version leads 20 of 27, and the 64-d version leads 18 of 27. Precomputed annual global embeddings are shared on Source Cooperative, so a small classifier can run locally. Ai2 reports v1.3 used 53% of v1.2's computing time and 18% of v1's, with 50% and 22% of the carbon emissions; the final pretraining run cost roughly $2,720 of rented H100 time.
- 128-d evals led20 of 27
- 64-d evals led18 of 27
- Compute vs v1.253%
- Final pretrain cost$2,720
Unclear The source does not quantify absolute accuracy gains per task or the storage bill for 128-d versus 64-d embeddings at regional scale.
Next Check the technical report's per-eval table for tasks where 128-d beats 64-d by a margin that justifies doubled storage.