OlmoEarth v1.1: A more efficient family of models

OlmoEarth v1.1 is a more efficient family of transformer-based models for processing satellite imagery.

• •Reduces compute costs by up to 3x while maintaining performance on remote sensing benchmarks
• •Token sequence length optimization is critical: compute costs scale quadratically with sequence length
• •Novel token design for Sentinel-2 data: creates 6 tokens per patch (2 timesteps × 3 resolutions)
• •Merges multi-resolution tokens without performance loss through modified pre-training approach
• •Available in Base, Tiny, and Nano model sizes with weights and code on HuggingFace and GitHub

This summary was automatically generated by AI based on the original article and may not be fully accurate.