Processing satellite images at continental or global scale requires much more than a good model. It also means controlling the cost of exporting data, preprocessing it, running inferences, and turning the results into useful maps. The Allen Institute for AI presents OlmoEarth v1.1, a new family of models that reduces the required compute by up to three times while generally maintaining the performance of the previous version.
A more efficient model for observing Earth
OlmoEarth v1 was launched in November 2025 and has since been used for tasks such as tracking changes in mangroves, classifying the causes of forest loss, and creating crop maps at the national level.
The new version aims to make these applications more accessible. According to its creators, OlmoEarth v1.1 maintains performance similar to OlmoEarth v1 across different research benchmarks and tasks developed with partner organizations, while reducing compute costs by up to 3x.
Why? When a system analyzes tens or hundreds of thousands of square kilometers, even a small improvement per image can become significant savings when working at planetary scale.
The role of tokens in satellite images
OlmoEarth uses an architecture based on transformers, the same family of models that powers a large portion of today’s AI systems. To process remote-sensing images, the model converts visual data into a sequence of units called tokens.
In this context, a token is not necessarily a word. It can represent a fragment of an image, a spectral band, or an observation captured at a specific moment. The way these tokens are built directly affects inference performance and cost.
In Sentinel-2 images, the data can be organized as a tensor with height, width, time, and 12 spectral channels. OlmoEarth v1 divided the image into spatial patches and generated separate tokens for each moment and resolution.
For example, with two points in time and three resolutions, each patch produced six tokens: two moments multiplied by the 10-, 20-, and 60-meter resolutions.
In general terms, a Sentinel-2 input with dimensions [H, W, T, D=12] generated approximately H/p x W/p x T x 3 tokens, where p represents the patch size.
Fewer tokens, lower inference costs
Transformers process the relationships between all the tokens in a sequence. That is why computational cost grows quadratically as the sequence gets longer. Reducing the number of tokens can have a much greater effect than simply removing a small portion of the data.
OlmoEarth v1.1 combines, in a single token, information that was previously separated by resolution. This decision reduces the number of tokens used for Sentinel-2 by approximately three times and generates savings during pretraining, fine-tuning, and inference.
The challenge was avoiding a loss in quality caused by the compression. A direct combination led to significant performance drops, including a 10-percentage-point reduction in m-eurosat kNN, a common benchmark for evaluating remote-sensing models.
The researchers suggest that separating bands by resolution helps the model learn important relationships between them. To compensate for the information lost through merging, they modified the pretraining setup. The details are presented in their technical report.
A family for different compute budgets
OlmoEarth v1.1 is not a single model. The family includes Base, Tiny, and Nano versions, so each team can choose a balance between capability, speed, and budget.
This strategy is especially relevant for organizations without access to large GPU infrastructures. A smaller model may be enough for a specific task, while a larger version may be more suitable when working with complex maps or large volumes of data.
The metric used to estimate the work required for one pass through the model is MACs, short for multiply-accumulate operations. In simple terms, fewer MACs usually mean faster and less expensive inference, although final performance also depends on the hardware and the complete processing workflow.
OlmoEarth v1.1 delivers performance similar to OlmoEarth v1 on many tasks, but may require close to one-third of the compute.
An easier experiment to interpret
Remote-sensing models involve many variables: architecture, dataset, size, pretraining strategy, and fine-tuning procedure. When performance changes, it is not always easy to know which of these factors was responsible.
To reduce that uncertainty, the team trained OlmoEarth v1.1 with the same dataset used for OlmoEarth v1. This means the differences between both versions make it easier to study the effect of the methodological changes, especially the way tokens are built and the new pretraining setup.
This makes the release more than a speed update. It also serves as an experiment in how to design AI models for satellite data without continuously increasing execution costs.
What it means for people who use satellite images
The reduction in compute can make frequent maps of forests, crops, coastlines, and other ecosystems more feasible. An organization could update its results more regularly or analyze larger areas without multiplying its infrastructure budget.
Even so, the recommendation is to test the new version for each use case. The Allen Institute for AI warns that there are some performance regressions and refers readers to its technical report for details. In machine learning, an average improvement does not guarantee better results for every dataset.
The pretrained weights and training code for OlmoEarth v1.1 are already available, including the Base, Tiny, and Nano variants. For researchers, governments, companies, and environmental organizations, the change presents a concrete possibility: observe more territory, with more frequent updates and a smaller compute bill.