OlmoEarth v1.1 slashes compute by up to three times

OlmoEarth v1.1 slashes satellite AI compute by up to three times, opening the door for bigger, more ambitious Earth monitoring deployments. The new family of models aims to preserve OlmoEarth v1’s performance on a wide range of benchmarks and partner tasks while dramatically cutting the compute bill across the full lifecycle of satellite data work. This is not a minor tweak; it changes what teams can run in-house and what scale they can reach on rented infrastructure, all without sacrificing accuracy.

The efficiency gains come from two levers, both centered on how the model ingests image data. OlmoEarth remains a transformer based system, but it processes imagery by converting it into a sequence of tokens, and the efficiency move hinges on decreasing those sequence lengths while also refining how the tokens are designed. In practical terms, shorter sequences mean fewer computations per prediction, while smarter tokens preserve essential information.

Compute is the scarcest resource across the satellite data lifecycle, from data export and preprocessing to inference and post processing. A leaner model reduces the bottleneck so partners can run larger areas and more frequent updates, whether in a government program, a regional alliance, or a global monitoring initiative. The result is a path to wider adoption, with organizations able to deploy OlmoEarth at national, continental, and even global scales while keeping costs manageable.

The paper underscores concrete use cases that illustrate the practicality of the improvement: tracking mangrove change, classifying drivers of forest loss, and producing country scale crop-type maps in days rather than weeks. These aren’t niche tasks; they are core workloads for environmental monitoring programs and policy planning, now within reach for more teams and budgets.

Since OlmoEarth v1 landed in November 2025, the v1.1 release builds on that momentum with a focus on efficiency without sacrificing what practitioners care about: accuracy and reliability across tasks. The emphasis on longer deployments and broader coverage signals a shift from laboratory benchmarks to real world impact, where every saved compute dollar can translate into more geographic coverage or refresh cycles.

Think of the shift as upgrading a powertrain without changing the chassis. You keep the same performance envelope, but you get it with far less fuel and heat. For teams pulling satellite streams into dashboards, that means faster iterations, lower operational costs, and the possibility to run more scenarios in parallel, which is crucial for risk assessment, early warning, and policy experiments.

Ablation studies in the technical report confirm the core claim: you can trim compute with sequence-aware tokenization and still preserve outcomes on a mix of research benchmarks and tasks constructed with partners. The evaluation metrics indicate the gains are real in practice, not just in theory, providing a credible reason for operators to revisit their pipelines this quarter.

For product teams and startups watching the quarter calendar, the takeaway is sharp: cheaper, faster, scalable satellite AI is closer to the edge of production than ever before. Expect more pilot programs to migrate to OlmoEarth v1.1 as partners test larger regions and more frequent updates, and watch for real world reports on how the new token design handles diverse environments and sensor configurations.