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

2026-05-15 · Source: Hugging Face - Blog · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Remote Sensing & Earth Observation AI · Depth: Advanced, short

Summary

AllenAI has released OlmoEarth v1.1, an updated family of Earth observation models designed for greater efficiency. Building on OlmoEarth v1, released in November 2025, this new version reduces compute costs by up to 3x while preserving performance across various research benchmarks and partner tasks. The primary innovation involves decreasing token sequence lengths in the transformer-based architecture. For Sentinel-2 imagery, the models now consolidate resolution-based patches into a single token per timestep, effectively tripling the token count reduction. This modification required adjustments to the pre-training regimen to prevent performance degradation. OlmoEarth v1.1 aims to make planet-scale map refreshes more affordable and faster for organizations tracking environmental changes.

Key takeaway

For Machine Learning Engineers deploying Earth observation models, OlmoEarth v1.1 offers a critical efficiency upgrade. If you are currently using OlmoEarth v1, migrating to v1.1 can cut your compute costs by up to three times for inference and fine-tuning, while largely maintaining performance. You should test v1.1 for your specific tasks to confirm compatibility and realize significant speedups, making frequent, large-scale map refreshes more affordable.

Key insights

OlmoEarth v1.1 achieves 3x compute cost reduction by optimizing transformer token sequence length without performance loss.

Principles

• Compute costs scale quadratically with token sequence length.
• Model size and token sequence length control transformer efficiency.
• Pre-training regimen modifications can mitigate performance drops from token merging.

Method

The method involves collapsing multiple resolution-based tokens (e.g., 10m, 20m, 60m for Sentinel-2) into a single token per patch and timestep, reducing token count by three times. This required modifying the pre-training regimen.

In practice

• Use OlmoEarth v1.1 for 3x cheaper inference.
• Explore Base, Tiny, Nano models for compute budgets.
• Test v1.1 for fine-tuning and inference speedup.

Topics

• OlmoEarth v1.1
• Earth Observation
• Transformer Models
• Compute Efficiency
• Remote Sensing
• Satellite Imagery

Code references

• allenai/olmoearth_pretrain

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

Related on AIssential

• 🌍 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.
• Efficient Remote Sensing Instance Segmentation with Linear-Time State Space Distilled Visual Foundation Models — Distilling knowledge into linear-time state space models significantly boosts remote sensing instance segmentation efficiency.
• Adapting Prithvi-EO for Fallow Detection for Food-Water Nexus: ViT-Adapter Necks and Parameter-Efficient Backbone tuning of Geospatial Foundation Model — Parameter-efficient tuning and adapter necks enable geospatial foundation models to detect multi-scale features effectively.
• 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.
• Query-efficient model evaluation using cached responses — Leveraging cached model responses via Data Kernel Perspective Space enables query-efficient benchmark performance prediction.
• Choosing the Right Model is Hard. Maintaining Accuracy is Harder. — Continuous monitoring and iterative refinement of open-source LLMs in production can enhance accuracy and efficiency.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Hugging Face - Blog.