OlmoEarth v1.1: A more efficient family of Earth observation models

Source: HuggingFace Date: 2026-05-19 URL: https://huggingface.co/blog/allenai/olmoearth-v1-1

Summary

Allen AI’s OlmoEarth v1.1 is a refreshed family of transformer-based Earth observation models (base, tiny, nano) that achieve up to 3× compute reduction relative to v1 while preserving benchmark performance. The efficiency gain comes from collapsing Sentinel-2’s per-resolution tokens into unified tokens, reducing quadratic attention costs across pretraining, fine-tuning, and inference. Applications include mangrove tracking, forest-loss classification, and country-scale crop mapping at planetary scale.

Implications

• Open-weight ecosystem. AllenAI continues its pattern of releasing capable, efficiency-tuned open models with reproducible training details. A 3× compute reduction makes satellite monitoring affordable at organizational scale without frontier-model cloud budgets.
• Specialized model releases. The token-collapsing technique is transferable beyond Earth observation — any multiresolution sensor modality (medical imaging, industrial vision) faces the same quadratic scaling problem. The published fix is a reference implementation for the pattern; the naive approach caused a 10-point benchmark drop before correction.
• Agent grounding thread. Planet-scale environmental monitoring feeds directly into agentic systems handling climate analysis, supply chain risk assessment, and regulatory compliance. As these models become cheaper to run, they become viable as background sensors in multi-agent pipelines rather than one-off research tools.