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

Allen AI Releases Open OlmoEarth v1.3 for Planetary Analysis

Allen AI (Ai2) has released OlmoEarth v1.3, a new state-of-the-art satellite embedding model, making its weights, pretraining data, and code fully open. The release is accompanied by precomputed annual embeddings for the entire planet and a new browser-based tool, the OlmoEarth Embeddings Explorer. This approach significantly lowers the barrier for planetary-scale data analysis, allowing researchers, analysts, and developers to perform tasks like change detection or land classification without requiring direct GPU access or needing to process terabytes of raw satellite imagery.

Technical Specifications and Performance

The OlmoEarth v1.3 model is a vision transformer that processes a time series of radar (Sentinel-1) and multispectral (Sentinel-2, Landsat) data to produce a single vector for every 10-meter pixel on Earth. In evaluations against models like AlphaEarth Foundations (AEF) and Tessera v2 across 27 different tasks, OlmoEarth v1.3's 128-dimensional embeddings achieved the highest score in 20 evaluations. Notably, the model is also more efficient, with pre-training requiring only 53% of the compute time of its predecessor and costing approximately $2,720 in H100 rental time for a final run.

• Model: OlmoEarth v1.3 Vision Transformer
• Input Data: Sentinel-1, Sentinel-2, and Landsat time series
• Output Embeddings: 128-dimensional and 64-dimensional vectors at 10m resolution
• Availability: Open model weights, code, data, and precomputed global embeddings on Source Cooperative
• Benchmark: Performance validated using 27 evaluations, including the new PLANTEUR crop-type benchmark

By pre-computing and distributing these embeddings, Allen AI shifts the workflow for many common geospatial tasks. Instead of requiring users to fine-tune a large model for each specific application, they can now train small, lightweight classifiers on top of the embeddings. This drastically reduces the computational cost and technical expertise needed for projects such as mapping smallholder crops, tracking illegal mining activity, or assessing fire damage. The Embeddings Explorer further accelerates this process, enabling users to test ideas and generate regional maps in seconds from a handful of labeled points directly in their browser.

By providing precomputed, planetary-scale embeddings as a public data artifact, Allen AI is abstracting away the most compute-intensive step of geospatial AI. This shifts the economic and operational focus from model training to lightweight application development, effectively commoditizing the foundational layer of planetary monitoring and enabling a wider range of organizations to build sophisticated analytical tools.