OlmoEarth v1.1: A more efficient family of models

At a glance:

• OlmoEarth v1.1 reduces compute costs by up to 3x compared to v1 while maintaining performance on remote sensing tasks
• The efficiency gain comes from decreasing token sequence lengths by merging multi-resolution satellite imagery data
• Available in Base, Tiny, and Nano sizes for different computational needs and budgets

What is OlmoEarth v1.1

AllenAI has released OlmoEarth v1.1, a new family of transformer-based models designed specifically for processing satellite imagery. This update to the original OlmoEarth v1, launched in November 2025, represents a significant advancement in efficiency for remote sensing applications. The models have been applied across a wide range of tasks, from tracking mangrove change to classifying drivers of forest loss and producing country-scale crop-type maps in days, with deployments scaling to national, continental, and global areas.

The primary motivation behind OlmoEarth v1.1 is addressing the high computational costs associated with processing satellite imagery at scale. When OlmoEarth processes satellite imagery to make predictions across tens to hundreds of thousands of square kilometers, efficiency becomes a critical factor that determines what's possible. Over the full lifecycle of running OlmoEarth – including data export, preprocessing, inference, and post-processing – compute costs represent by far the highest expense. By making the models more efficient, AllenAI aims to support more partners on the OlmoEarth Platform and enable anyone running OlmoEarth independently to leverage this technology faster and at lower expense.

How Efficiency is Achieved

The key improvement in OlmoEarth v1.1 is its ability to perform the same tasks with up to three times less computational resources. This efficiency gain is achieved by decreasing sequence lengths in the transformer-based models. In transformer architectures, compute costs scale quadratically with token sequence length, meaning even small reductions can lead to significant cost savings. The models use MACs (multiply-accumulate operations) as a metric to estimate computation needed for one forward pass, with lower MACs generally meaning cheaper, faster inference.

To understand how sequence length reduction works, it's important to consider how satellite imagery is processed. The models first convert the imagery into a sequence of tokens that the transformer can ingest. For Sentinel-2 imagery, which has dimensions of [H, W, T, D=12] (height, width, timesteps, and 12 channels), the traditional approach splits the data into resolution-based patches. Each patch creates multiple tokens per timestep per resolution, with Sentinel-2 typically having three resolutions (10m, 20m, and 60m). This results in a substantial number of tokens that must be processed.

Token Design Improvements

A significant innovation in OlmoEarth v1.1 is its redesigned token approach for processing Sentinel-2 imagery. Previously, models like Galileo and SatMAE used a unique token per resolution, which means creating separate tokens for the 10m, 20m, and 60m resolutions. While this approach has shown good results, it also generates a large number of tokens that compound multiplicatively across pretraining, fine-tuning, and inference stages.

The OlmoEarth team explored an alternative approach used by CROMA, which collapses all resolutions into a single token regardless of resolution. This method produces three times fewer tokens than the multi-resolution approach, leading to substantial computational savings. However, naively combining tokens in this way historically caused significant performance drops, including a 10 percentage point drop on the m-eurosat kNN benchmark, a common evaluation task for remote sensing models.

To overcome this performance challenge while maintaining the efficiency benefits, the AllenAI team modified their pre-training regimen. They hypothesized that separating Sentinel-2 bands into different tokens makes it easier for OlmoEarth to model important cross-band relationships, which is crucial for accurate remote sensing analysis. By carefully adjusting their pre-training approach, they were able to merge tokens without the typical performance degradation, achieving the efficiency gains while maintaining model effectiveness.

Benefits for Developers

For developers and organizations using OlmoEarth, the v1.1 release offers immediate practical benefits. At every model size, OlmoEarth v1.1 runs up to three times cheaper than the original OlmoEarth v1, making frequent, planet-scale map refreshes more affordable for every team running the platform. This cost reduction enables more frequent updates to environmental monitoring systems, which is critical for applications like disaster response, climate change tracking, and agricultural management.

The upgrade path is straightforward for existing users. If you're currently using a model from the original OlmoEarth family, the recommendation is to try OlmoEarth v1.1. It provides similar performance to OlmoEarth v1 while requiring only one third of the computational resources. While the team has observed some regressions in specific cases (detailed in their technical report), for most applications, users should see significant speedups during both fine-tuning and inference phases. This efficiency improvement means that organizations can process larger areas of satellite imagery or update their analyses more frequently without proportional increases in computational costs.

Benefits for Researchers

For the research community, OlmoEarth v1.1 presents an interesting case study in model efficiency and design. Pretrained remote sensing models have many degrees of freedom, including architecture choices, datasets, and pre-training algorithms, which makes it challenging to isolate the effects of specific changes. By training OlmoEarth v1.1 on the same dataset as OlmoEarth v1, the AllenAI team has created a controlled comparison where any differences between the two versions can be attributed solely to methodological changes.

This controlled approach advances scientific understanding of pre-training principles for remote sensing models. Researchers can now study the effects of token design choices and sequence length optimization without confounding factors from different datasets or architectures. The technical report accompanying the release provides detailed analysis of these methodological changes, offering valuable insights for the broader machine learning community working with geospatial data. As remote sensing becomes increasingly important for climate science, environmental monitoring, and sustainable development, such methodological advances contribute to the field's overall progress.

Getting Started with OlmoEarth v1.1

For those interested in implementing OlmoEarth v1.1, AllenAI has made the model weights and training code publicly available. The release includes weights for three model sizes: Base, Tiny, and Nano, allowing users to select the appropriate model based on their computational resources and accuracy requirements. The Base model offers the highest performance but requires more computational resources, while the Tiny and Nano models provide lighter alternatives that can run on more modest hardware or in resource-constrained environments.

The code repository contains comprehensive documentation and examples to help developers get started quickly. For organizations looking to leverage OlmoEarth for environmental monitoring, the efficiency gains in v1.1 make it more accessible than ever before. Whether you're working on deforestation tracking, agricultural monitoring, disaster response, or climate change research, OlmoEarth v1.1 provides a powerful toolset that balances performance with computational efficiency. The models are particularly well-suited for applications requiring analysis at national, continental, or global scales, where computational costs have traditionally been a limiting factor.