OlmoEarth v1.1: A More Efficient Family of Earth Observation Models

The cost of intelligence is increasingly measured in compute. As we scale AI to monitor the planet, the bottleneck is no longer just the availability of data, but the expense of processing it.

The release of OlmoEarth v1.1 introduces a new family of models designed to address this exact friction. Following the release of OlmoEarth (v1) in November 2025, the technology has been applied to tasks ranging from tracking mangrove change to producing country-scale crop-type maps in days. However, as these deployments scale to national, continental, and global areas, the economic reality of transformer-based architectures becomes unavoidable.

In the lifecycle of running these models—encompassing data export, preprocessing, inference, and post-processing—compute remains the highest cost. The architecture of OlmoEarth v1.1 focuses on cutting these compute costs by up to 3x. This is achieved by targeting the primary levers of transformer efficiency: model size and token sequence length. Because compute costs scale quadratically with the token sequence length, even small reductions can meaningfully cut the cost of running the model.

The strategic importance of this shift cannot be overstated. For organizations and communities working to protect people and our planet, efficiency dictates the scope of what is possible. When processing satellite imagery across tens to hundreds of thousands of square kilometers, a more efficient model allows for more partners to be supported on the OlmoEarth Platform. It also ensures that those running the technology independently can leverage it faster and at lower expense.

By reducing MACs—the multiply-accumulate operations required for a model forward pass—the v1.1 family moves toward cheaper and faster inference. This is not merely a technical optimization; it is a move toward democratizing high-fidelity Earth observation.

As the industry moves toward larger-scale deployments, we must ask: are we building models that are merely powerful, or are we building models that are economically sustainable for global-scale environmental stewardship?