The Allen Institute for AI (Ai2) has released version 1.3 of OlmoEarth, its family of foundation models built for geospatial data analysis. These models generate embeddings from satellite imagery and other Earth observation sources, which can then power downstream applications such as crop monitoring, environmental tracking, or disaster mapping.
OlmoEarth fits into a broader trend of domain-specific foundation models, where the goal is not to generate text or images but to produce compact, reusable representations of complex data — in this case, time series from multispectral satellite sensors. These embeddings can then be fed into lighter downstream models for specific tasks, avoiding the need to reprocess raw imagery each time.
Ai2 presents this latest version as delivering state-of-the-art results across several Earth observation benchmarks, though precise technical details — architecture, training data, and quantified comparisons — remain to be documented beyond the initial announcement. The project follows Ai2's usual approach of releasing its research models openly, in contrast to many industry players that keep such tools proprietary.
Geospatial AI remains a niche compared to general-purpose large models, but it is drawing growing interest from government agencies, environmental organizations, and agricultural companies, for whom access to capable open models offers a meaningful cost advantage over existing proprietary solutions.