OlmoEarth v1.1: A more efficient family of models

2026-05-18 · Source: Ai2 Blog · Field: Science & Research — Environmental Science & Earth Systems, Research Methodology & Innovation · Depth: Advanced, medium

Summary

OlmoEarth v1.1, released May 19, 2026, is a family of transformer-based models from Ai2 designed to process satellite imagery with up to 3x greater compute efficiency than its predecessor, OlmoEarth v1, while maintaining performance. This efficiency is achieved by reducing token sequence lengths, a critical factor as compute costs scale quadratically with sequence length. Specifically, v1.1 modifies the token design for Sentinel-2 imagery, collapsing multiple resolution-based tokens per patch into a single token, thereby reducing the total token count by three times. This change required adjustments to the pre-training regimen to prevent significant performance degradation, such as a 10 ppt drop on the m-eurosat kNN benchmark. An update, OlmoEarth v1.2, released June 29, further enhances the models by incorporating rotary positional embeddings (RoPE), which reduces embedding artifacts and provides a small performance boost across Nano, Tiny, Small, and Base model sizes.

Key takeaway

For Machine Learning Engineers deploying satellite imagery analysis, OlmoEarth v1.1 offers a critical compute cost reduction of up to 3x without sacrificing performance. If you are currently using OlmoEarth v1, migrating to v1.1 can significantly lower operational expenses for planet-scale map refreshes. Furthermore, consider OlmoEarth v1.2 for cleaner embeddings and a small performance boost, especially if you are on Tiny or Nano models, where moving to Small is recommended.

Key insights

OlmoEarth v1.1 significantly cuts compute costs for satellite imagery processing by optimizing token design and positional embeddings.

Principles

• Compute costs scale quadratically with token sequence length.
• Positional embeddings can introduce artifacts if not carefully designed.
• Token design significantly impacts transformer model efficiency.

Method

OlmoEarth v1.1 reduces token sequence length by merging resolution-based tokens into a single token per patch, requiring pre-training regimen modification. OlmoEarth v1.2 uses RoPE for artifact-free positional encoding.

In practice

• Use OlmoEarth v1.1 for 3x cheaper satellite imagery processing.
• Consider OlmoEarth v1.2 for cleaner embeddings and performance.
• Migrate from Tiny/Nano to Small for improved efficiency.

Topics

• OlmoEarth Models
• Satellite Imagery Processing
• Transformer Architecture
• Positional Embeddings
• Compute Cost Optimization
• Remote Sensing Benchmarks

Code references

• allenai/olmoearth_pretrain

Best for: Computer Vision Engineer, AI Engineer, AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• OlmoEarth v1.1: A more efficient family of Earth observation models — OlmoEarth v1.1 achieves 3x compute cost reduction by optimizing transformer token sequence length without performance loss.
• Introducing OlmoEarth embeddings: Custom embedding exports from OlmoEarth Studio for downstream analysis — OlmoEarth embeddings offer a cost-effective, label-efficient way to analyze Earth observation data for diverse downstream tasks.
• OLMo Earth — OLMoEarth democratizes Earth observation AI through open-source models and a user-friendly platform.
• LALE: Lightweight-Transformer Architecture for Land-Cover Estimation — LALE efficiently segments remote sensing imagery by combining ConvMixers for local detail and transformers for global context in a bifurcated encoder.
• Learning Where to Embed: Noise-Aware Positional Embedding for Query Retrieval in Small-Object Detection — Selectively embedding positional information in foreground regions improves small-object detection and reduces model complexity.
• 🌍 OlmoEarth v1.1: 3x cheaper to run than v1 with the same SOTA performance, fully open — OlmoEarth v1.1 offers 3x cost reduction for weather forecasting with maintained SOTA performance.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Ai2 Blog.